PHD_THESIS_VAN-ANDEL.

PHD_THESIS_VAN-ANDEL.
ANTICIPATORY WATER MANAGEMENT
USING ENSEMBLE WEATHER FORECASTS
FOR CRITICAL EVENTS
Anticipatory Water Management
Using ensemble weather forecasts for critical events
DISSERTATION
Submitted in fulfilment of the requirements of
the Board for Doctorates of Delft University of Technology
and of
the Academic Board of the UNESCO-IHE Institute for Water Education
for the Degree of DOCTOR
to be defended in public
on Tuesday, November 3, 2009 at 15:00 hours
in Delft, The Netherlands
by
Schalk Jan van ANDEL
born in Amsterdam, The Netherlands
Master of Science in Quantitative and Integrated Water Management with
Distinction, Wageningen University, The Netherlands
This dissertation has been approved by the supervisor
Prof. dr. R. K. Price
Members of the Awarding Committee:
Chairman
Rector Magnificus, TU Delft
Prof. dr. ir. A.E. Mynett
Vice-Chairman, UNESCO-IHE
Prof. dr. R.K. Price
Supervisor, UNESCO-IHE / TU Delft
Prof. drs. ir. J.K. Vrijling
TU Delft, The Netherlands
Prof. dr. ir. A.W. Heemink TU Delft, The Netherlands
Prof. dr. ir. E. Schultz
UNESCO-IHE, The Netherlands
Dr. J.C. Schaake
NWS / NOAA, USA
Dr. ir. A.H. Lobbrecht
UNESCO-IHE, The Netherlands
Prof. ir. F.H.L.R. Clemens TU Delft, The Netherlands, reserve member
CRC Press/Balkema is an imprint of the Taylor & Francis Group, an informa business
© 2009, Schalk Jan van Andel
All rights reserved. No part of this publication or the information contained herein
may be reproduced, stored in a retrieval system, or transmitted in any form or by any
means, electronic, mechanical, by photocopying, recording or otherwise, without
written prior permission from the publisher.
Although care is taken to ensure the integrity and quality of this publication and the
information therein, no responsibility is assumed by the publishers nor the author for
any damage to property or persons as a result of operation or use of this publication
and/or the information contained herein.
Published by:
CRC Press/Balkema
PO Box 447, 2300 AK Leiden, The Netherlands
e-mail: [email protected]
www.crcpress.com – www.taylorandfrancis.co.uk – www.balkema.nl
ISBN 978-0-415-57380-1 (Taylor & Francis Group)
v
Foreword
When in 2002, I learned of the existence of UNESCO-IHE, by that time still
named 'IHE-Delft', I realised that this institute for water education
encompassed all my professional interests and even some of my stronger
personal interests (if this division between personal and professional
interests really exists). UNESCO-IHE is entirely devoted to the aquatic
environment, with emphasis on capacity building in less privileged
countries. Perhaps its greatest merit is the 'peace building' through the cross
national and cross continental friendships between around 200 new MSc
students coming from all over the world every year. Walking through the
corridors, and having lunch in the canteen, together with all these water
professionals from such diverse background, is an honour and pleasure for
me every day. I was convinced by that time, through a short period of
introduction to different professional environments, that the combination of
education and research is for me the most attractive way of spending one's
professional life. Add to this my personal interest in water sports, such as
windsurfing and wave-surfing, and it should be clear why UNESCO-IHE
stands out as the most attractive working environment I can think of.
Two years later, in 2004, I learned from Dr. Lobbrecht, who is director of
HydroLogic BV and also is a faculty member of the Hydroinformatics Core
at UNESCO-IHE, what a challenge the highly variable weather conditions in
the Netherlands are for the Water Boards, which are responsible for the daily
operation of the regional water systems. Research was needed into the use of
weather forecasts in decision support systems for operational water
management. This research topic, involving Real-Time Control with its
combination of water management and ICT, is one of the research areas of
the Hydroinformatics and Knowledge Management Department at
UNESCO-IHE. With the importance of flood and drought management, and
the increasing availability and uptake of ICT in developing countries, the
research was internationally relevant.
vii
Acknowledgements
The research was funded by the Delft Cluster programme, the Principal
Water-board of Rijnland and UNESCO-IHE.
The data for this research was kindly provided by the Royal Netherlands
Meteorological Institute (KNMI), the Principal Water-board of Rijnland,
National Meteorological Agency (NMA) of Ethiopia, Ministry of Water
Resources Ethiopia, and Bahirdar metrological office. The HydroNET
software for processing the meteorological data was kindly provided by
HydroLogic BV.
I would like to thank Professor Roland Price for his enthusiasm for the topic,
the ideas for exploration, the discussions on the content, the continuing
efforts in correcting my English texts, the sharing of thoughts on
philosophical, societal and theological issues, and for his personal way of
supporting and guiding me. Professor Price's personal involvement with,
care for, and support and guidance of students is an example for me in my
further professional and private life.
With Arnold Lobbrecht I am working, not only on the current research, but
also on many other project activities within UNESCO-IHE. Because of these
always-ongoing busy activities, and because of the pleasant, informal way of
working together, there has been little opportunity to express my
appreciation and gratitude towards Arnold. Hence, I take the opportunity to
do so here. I think Arnold is too modest about his vast knowledge and
experience in water management and hydroinformatics.
This research would not have been possible without the countless inputs
from professionals in the water management and meteorological fields. I
would like to thank Frans van Kruiningen for his guidance throughout the
research, for his insights into the Rijnland water system, and for his support
in preparing journal papers. Also, my thanks go to thank René van der Zwan
for continuing the cooperation between Rijnland and UNESCO-IHE after
Frans joined the Principal Water-board of Delfland. Robert Mureau is the
Ensemble expert at the Royal Netherlands Meteorological Institute (KNMI).
My thanks go to Robert for the many discussions we had, for helping me to
understand the ensemble prediction systems, and for his thorough reviews of
journal papers. Robert Mureau is currently working with MeteoConsult. In
addition, I appreciate the support of Kees Kok (KNMI) in understanding
some of the verification techniques used in weather forecasting, and the
continuing cooperation. Thank you also to Sander Loos and Timmy
Knippers for their support with the HydroNET software, and to Jantine
viii
Bokhorst for her support in processing the ensemble forecast data.
I have had the pleasure to work with MSc students on their thesis research
work. In particular, I am grateful to Kibreab Amare Assefa with whom
Chapter 5 about the Upper Blue Nile case study was prepared. His research
work was funded by the WaterMill project.
I would like to thank Wilmer Barreto for providing me with the NSGAII
software he developed. The software proved invaluable for my research.
Then, I would like to express my appreciation to Jan Luijendijk for
introducing me to the Hydroinformatics and Knowledge Management
Department, and for supporting me in finding the right balance between PhD
work and the many other interesting activities available within the
department. Thank you also to Professor Dimitri Solomatine for his support
as the Head of the Hydroinformatics Core, for his input on the optimisation
problems in this research, and for his friendly and humorous cooperation. I
would like to thank my UNESCO-IHE colleagues Yasir Abas Mohamed, for
introducing me to case study opportunities in the Upper Blue Nile, and
Shreedhar Maskey, for his input concerning the decision-making challenges
in flood early warning.
To Professor Michael Abbott, I am also very grateful for the many
discussions on socio-economic and philosophical issues, and for introducing
me to the Water Knowledge Initiative.
I would like to thank Andreja Jonoski for inspiring me in the work at
UNESCO-IHE, on hydroinformatics and in general.
Thank you to Ioana Popescu, for the fine cooperation on other ongoing
projects within the hydroinformatics group, and in particular for supporting
me during the finalisation of this thesis.
I enjoy working together with my hydroinformatics colleagues: Professor
Arthur Mynett, Professor Guy Alaerts, Zoran Vojinovic, Ann van Griensven,
Biswa Bhattacharya, Yunqing Xuan, Carel Keuls, Judith Kaspersma,
Giuliano Di Baldassarre, Jos Bult, and Gerda de Gijsel. In addition, I would
like to thank my PhD colleagues and in particular Carlos Velez, Leonardo
Alfonso, and Gerald Corzo, who have been working in associated fields of
forecasting and control. Next to the scientific efforts, I have to recognise the
not always successful sports team efforts with them. Thank you to all staff
and participants of UNESCO-IHE for providing a great working experience.
Finally, I would like to thank the members of the doctoral examination
committee for evaluating this thesis.
ix
Summary
Most of today's inland surface-water systems are integrally connected to
developments in human society. These systems depend on good day-to-day
water management. Under normal operational conditions they present few
problems. Critical conditions may cause problems, such as floods and
droughts. These problems can be classified as having too much water, too
little water, or water of poor quality. We try to minimise the frequency and
extent of the damage due to critical events by water management. Strategic
water management is concerned with catchment land use, spatial planning
and water system design, while operational water management is concerned
with the daily management of a given water system.
A large group of critical events are caused by meteorological extremes.
Often operational water managers are informed too late about upcoming
events to respond to them in an optimal manner. The lead-time provided by
monitoring systems and hydrological predictions is not enough. Therefore,
weather forecasting can be used, e.g. as input to the hydrological models, to
expand the forecast horizon in water management. This is called
Anticipatory Water Management (AWM). It allows water managers to take
anticipatory actions to reduce the damage of critical events. An example of
an anticipatory action is the lowering of a reservoir water level as part of
flood control.
Similar to chess players when anticipating the moves of their opponents and
planning their own counter moves, water managers can improve the
performance of their systems, the more they are able to anticipate the
upcoming events.
Hydro-meteorological forecasts are not always accurate. There is some
degree of uncertainty whether the forecasted events will really occur. In
particular, weather forecasts have a high degree of uncertainty, because the
atmosphere is a chaotic system, in which small disturbances can grow
rapidly to influence large-scale events. Anticipatory actions may, therefore,
not be taken in time, or taken unnecessarily. Because of possible adverse
effects of anticipatory actions, like the shortage of water for supply in the
case of lowering of a reservoir water level for flood control, the uncertainty
of the forecasts and associated risks of applying Anticipatory Water
Management have to be assessed.
Ensemble Prediction Systems (EPS) have been developed to assess the
dynamic uncertainty of weather forecasts. For each forecast the probability
distribution is estimated by re-running the numerical prediction model with
x
different initial conditions. This takes into account our limitations to measure
or estimate the initial atmospheric state accurately at a high spatial and
temporal resolution. The forecasted probability distribution allows water
managers to make risk-based decisions. Much research focuses on providing
reliable hydro-meteorological ensemble predictions. Increasingly water
authorities and companies are making use of these predictions. This research
focuses on the improvement of the end-use of ensemble prediction systems
in Anticipatory Water Management.
A framework for developing Anticipatory Water Management strategies is
proposed. Firstly, in this framework emphasis is given to the availability of
hydroinformatics tools that allow flexible and realistic simulation of
controlled water systems. Using these simulation models, the current watermanagement strategy can be emulated, and compared with alternative,
anticipatory, strategies.
Secondly, it is emphasised that water authorities should themselves verify the
performance of the hydro-meteorological forecasts local to their catchment.
Generalised performance indicators, established on a regional or global scale
as they are provided by meteorological institutes, do not provide sufficient
information for local water management. The verification should be
customised for the intended end-use of the Anticipatory Water Management.
This means, for example, that the verification should focus on surface
precipitation or rainfall-runoff modelling in applications to flood control. In
addition, the verification should not be based on a fixed time interval, such
as a day, but should establish for each event (e.g. an intense rainfall episode
continuing for several days) whether or not it was predicted. The verification
should be done using continuous time series and simulation, not only on the
basis of a sample of critical events such as has been the practice in water
management until recently. Only with continuous simulation can the full
consequences of applying Anticipatory Water Management, including risks
of false alarms (a forecast of a critical event while no critical event occurs)
during normal conditions, be assessed. For this verification analysis,
archives of water system data, meteorological data, and weather forecasts are
needed. If an archive of weather forecasts is not available, than these
weather forecasts need to be prepared. This can be done by re-running the
numerical weather prediction models for the verification analysis period.
This is called re-forecasting or hindcasting.
The water system model, together with the meteorological hindcast as input,
allows "what-if" analyses for long periods. The analyses show water
managers what would have happened if they had used the weather forecasts
in their operational water management during the previous so many years.
This already gives an indication of the effectiveness of the Anticipatory
Water Management (AWM) in reducing the negative impact of critical
events. For many water authorities, however, this will not be enough
xi
information to decide whether to adopt AWM. In most cases the (economic)
efficiency should also be assessed. While general efficiency analyses are
often performed using cost-loss ratios, these are not applicable to AWM
because water management is highly dynamic. Each event is different and so
are the cost-loss ratios.
Therefore, thirdly, a dynamic cost-model related to water system states,
reflecting all the efficiency requirements, should be prepared by the water
authority. Then the continuous simulation of water management can be
translated into a time series of costs. The total damage costs of critical events
and their development over the years can be assessed, and compared for
different forecasting products and for different anticipatory management
strategies.
If a forecasting product has been selected and the effectiveness and
efficiency of AWM strategies are such that the water authority would like to
adopt it, then as an extra step, an optimisation of the AWM strategy can be
performed. The objectives of the optimisation in most cases would be to
minimise the damage of critical events, and at the same time minimise the
total damage. The standard risk-based approach, minimising the expected
risk for every decision time step, may not be reliable, because it assumes the
use of perfect probabilistic forecasts, while these are in reality not available.
To take these uncertainties into account the strategy with the minimum costs
over a long period (years) has to be found. This multi-year optimisation
problem, in which per day several ensemble predictions are available and the
best management strategy for the entire period needs to be defined, cannot
be captured in an analytical optimisation model. Therefore, global
optimisation methods with smart search methods, like evolutionary
approaches, are used. Importantly, these search methods can be used with
multiple objectives to provide a range of alternative strategies. This leaves
the freedom to the water authority to select the optimal water management
strategy depending on their perception of the importance of the different
objectives.
The framework for developing Anticipatory Water Management strategies
was applied to two case studies in flood early warning and control. One case
study concerned a land-reclamation area in the Netherlands, Rijnland, and
the other a tributary to Lake Tana in the catchment of Upper-Blue Nile,
Ethiopia. The ensemble precipitation forecasts from the ECMWF Ensemble
Prediction System and the NCEP Global Forecasting System (the frozen
version for re-forecasting ensembles) were used. The ECMWF EPS is
already received operationally by the Water Board in the Dutch case study.
The NCEP GFS is freely available through the Internet and is therefore a
very interesting research and operational tool for countries where
investments for hydro-meteorological forecasting systems are not yet readily
available.
xii
For the Rijnland case study, effective warnings were obtained for most of the
critical events in the analysis period. The optimisation of the Anticipatory
Water Management strategy resulted in a 30% reduction of the estimated
total costs, and a reduction of 35% of the flood damage costs over a 8-year
period. This shows clearly that Anticipatory Water Management outperforms
the traditional use of re-active operational water management. In the
Netherlands, ECMWF EPS forecasts can be used to expand management
horizons to three or more days. The fact that the optimal decision rules differ
from the ones currently used by the Rijnland Water Board confirms the need
for the Water Boards to perform hindcast analyses to improve their
anticipatory management strategies.
The case study of the Blue Nile shows that freely available weather forecasts
and hydrological modelling software can be used for research into prediction
systems and Anticipatory Water Management strategies. For this particular
case study, a warning could be obtained for a maximum of 60% of the peaks
in the simulated reference streamflow (above flood threshold). The
forecasting system needs further improvement before operational use is
considered. For this improvement, bias correction and downscaling methods
that are the focus of current international research efforts into Hydrological
Ensemble Prediction Systems should be used. These methods to produce
reliable probabilistic forecasts, with as small a predictive uncertainty as
possible, will also be used in ongoing research to increase still further the
efficiency of AWM for Water Boards in the Netherlands.
The backbone of developing successful and reliable AWM strategies is the
verification analysis with continuous simulation spanning multiple years.
Archives of weather forecasts of multiple years are necessary, because of the
low frequency of critical events. These archives are generally not available
because weather forecasting systems are continuously updated. Therefore,
there is a strong need to prepare hindcast archives for new products. Because
preparing these hindcasts interferes with the operational tasks of the
meteorological institutes, the task of hindcasting should be relegated to
separate, dedicated institutes. This would give a credible contribution to the
practical use of weather forecast products.
This is important for water management worldwide, because it is clear that
the performance of weather forecasts is such that water authorities cannot
afford not to use this available information. This applies not only to the flood
management case studies presented in this thesis, but for many more
applications, such as drought management, and for many more types of
water systems. Therefore, scientists and engineers are called on to join in an
effort to expose and to cover the complete scope of Anticipatory Water
Management, and to maximise the use of hydro-meteorological forecasts in
operational water management.
xiii
Content
FOREWORD ........................................................................................................... V
ACKNOWLEDGEMENTS ................................................................................. VII
SUMMARY.............................................................................................................IX
1
INTRODUCTION......................................................................................... 17
1.1
1.1.1
1.1.2
1.1.3
1.1.4
1.1.5
1.1.6
1.2
1.3
1.4
2
BACKGROUND ........................................................................................ 17
Hydroinformatics and Integrated Water Resources Management.... 17
Management of extreme events......................................................... 18
Operational water management ....................................................... 19
Benefits of increased forecast horizon.............................................. 20
Use of weather forecasts................................................................... 21
Ensemble forecasts ........................................................................... 22
ANTICIPATORY WATER MANAGEMENT.................................................. 23
HYPOTHESES AND OBJECTIVES ............................................................... 25
READER.................................................................................................. 26
ANTICIPATORY WATER MANAGEMENT .......................................... 27
2.1
2.2
2.2.1
2.2.2
2.2.3
2.2.4
2.2.5
2.2.6
2.3
2.3.1
2.3.2
2.3.3
2.3.4
2.3.5
2.4
2.4.1
2.4.2
2.4.3
2.4.4
2.5
2.5.1
2.5.2
2.5.3
2.5.4
2.5.5
2.6
INTRODUCTION....................................................................................... 27
OPERATIONAL WATER MANAGEMENT .................................................... 27
Definition .......................................................................................... 27
Components of operational water management ............................... 27
Water system control ........................................................................ 29
Reservoirs and polders ..................................................................... 31
Flood early warning and control...................................................... 32
Challenges in operational water management ................................. 33
WEATHER FORECASTING AND ENSEMBLE PREDICTIONS ......................... 34
Monitoring systems........................................................................... 34
From hand-drawn weather maps to numerical prediction ............... 36
From deterministic to probabilistic forecasts................................... 37
Ensemble Prediction Systems ........................................................... 38
Challenges in using weather forecasts for water management......... 40
MODELLING CONTROLLED WATER SYSTEMS .......................................... 41
Definitions ........................................................................................ 41
Model components ............................................................................ 42
Water system state prediction ........................................................... 43
Challenges in modelling controlled water systems........................... 44
DECISION MAKING WITH UNCERTAINTY ................................................. 46
Uncertainty ....................................................................................... 46
Risk ................................................................................................... 47
Threshold-based decision rules for Ensemble Prediction Systems... 48
Cost-benefit analysis......................................................................... 49
Decision Support Systems for Anticipatory Water Management ...... 50
KNOWLEDGE GAPS AND HYPOTHESES .................................................... 51
xiv
3
FRAMEWORK FOR DEVELOPING ANTICIPATORY WATER
MANAGEMENT (AWM)...................................................................................... 55
3.1
3.2
3.2.1
3.2.2
3.3
3.3.1
3.3.2
3.3.3
3.4
3.4.1
3.4.2
3.5
3.5.1
3.5.2
3.5.3
3.6
3.6.1
3.6.2
3.6.3
3.7
3.7.1
3.7.2
3.7.3
3.7.4
3.8
3.8.1
3.8.2
3.9
4
INTRODUCTION....................................................................................... 55
ESTABLISHING THE NEED AND POTENTIAL FOR AWM............................ 55
For which events is AWM needed..................................................... 55
Potential for anticipatory management action ................................. 60
VERIFICATION ANALYSIS........................................................................ 63
Product selection: time scales, spatial scales................................... 63
Continuous simulation of the real-time AWM forecasting system .... 63
Event based verification of a range of decision rules for AWM ....... 65
MODELLING CONTROLLED WATER SYSTEMS .......................................... 67
Input data based on end-use of model .............................................. 68
Framework for modelling controlled water systems......................... 68
STRATEGIES FOR ANTICIPATORY WATER MANAGEMENT ........................ 69
Rule-based ........................................................................................ 70
Pre-processing of ensemble forecasts to deterministic forecast ....... 71
Risk-based......................................................................................... 71
COST-BENEFIT OF SELECTED AWM STRATEGIES.................................... 73
Dynamic cost-benefit analysis .......................................................... 73
Sources of damage............................................................................ 74
Anticipatory Water Management modelling ..................................... 74
OPTIMISATION OF ANTICIPATORY WATER MANAGEMENT ..................... 75
Objectives ......................................................................................... 76
Parameterisation of AWM strategies................................................ 76
Optimisation using perfect forecasts ................................................ 77
Optimisation with actual forecasts ................................................... 77
DECISION MAKING FOR POLICY ADOPTION OF AWM.............................. 78
What-if analysis ................................................................................ 78
Re-analysis era ................................................................................. 79
FRAMEWORK FOR DEVELOPING ANTICIPATORY WATER MANAGEMENT 79
CASE STUDY 1 - RIJNLAND WATER SYSTEM ................................... 81
4.1
4.2
4.3
4.4
4.4.1
4.4.2
4.4.3
4.4.4
4.4.5
4.4.6
4.4.7
4.4.8
4.5
4.5.1
4.5.2
4.5.3
4.5.4
4.5.5
INTRODUCTION....................................................................................... 81
PROBLEM DESCRIPTION .......................................................................... 83
DATA ..................................................................................................... 84
WATER SYSTEM CONTROL MODEL .......................................................... 86
Model structure................................................................................. 86
Control strategy ................................................................................ 88
Model calibration ............................................................................. 89
Model validation............................................................................... 90
Visualise what is not known and explain .......................................... 92
Modelling the unknown phenomena ................................................. 95
Final model results ........................................................................... 97
Discussion....................................................................................... 102
ENSEMBLE FORECASTS VERIFICATION .................................................. 103
Precipitation ensemble forecasts archive ....................................... 103
Water level hindcasts...................................................................... 103
Event based verification for water managers ................................. 104
Precipitation and water level thresholds ........................................ 105
Presently used precipitation threshold for anticipatory pumping .. 105
xv
4.5.6
4.5.7
4.5.8
4.6
4.7
4.7.1
4.7.2
4.8
4.8.1
4.8.2
4.9
5
CASE STUDY 2 - UPPER BLUE NILE ................................................... 127
5.1
5.2
5.3
5.3.1
5.3.2
5.3.3
5.4
5.4.1
5.4.2
5.5
5.5.1
5.5.2
5.5.3
5.5.4
5.5.5
5.5.6
5.5.7
5.6
5.7
6
3-Day accumulated precipitation threshold for selected events ..... 108
5-Day accumulated precipitation threshold for selected events ..... 109
Discussion....................................................................................... 111
ANTICIPATORY WATER MANAGEMENT STRATEGY DEVELOPMENT ....... 114
COST-BENEFIT OF SELECTED AWM STRATEGIES.................................. 116
Water level - damage function ........................................................ 116
Inter-comparison of costs for selected strategies ........................... 118
OPTIMISATION OF ANTICIPATORY WATER MANAGEMENT STRATEGY . 120
Optimisation with perfect forecasts ................................................ 120
Optimisation with actual forecasts ................................................. 122
ADOPTION OF AWM IN OPERATIONAL MANAGEMENT POLICY ............. 125
INTRODUCTION..................................................................................... 127
PROBLEM DESCRIPTION ........................................................................ 127
DATA ................................................................................................... 128
Geographical data .......................................................................... 128
Meteorological data........................................................................ 128
Streamflow data .............................................................................. 130
HYDROLOGICAL MODEL ....................................................................... 131
Model set-up ................................................................................... 131
Calibration and validation.............................................................. 132
ENSEMBLE FORECASTS VERIFICATION .................................................. 135
Event selection ................................................................................ 135
Ensemble precipitation hindcasts ................................................... 138
Ensemble streamflow hindcasts ...................................................... 138
Verification analysis ....................................................................... 138
Statistical verification..................................................................... 139
Comparison by visual inspection.................................................... 140
Flood early warning verification .................................................... 142
ANTICIPATORY MANAGEMENT STRATEGY DEVELOPMENT ................... 144
ADOPTION OF AWM IN OPERATIONAL MANAGEMENT POLICY ............. 145
CONCLUSIONS AND RECOMMENDATIONS .................................... 147
6.1
6.2
6.3
6.4
6.5
CONTRIBUTIONS TO ANTICIPATORY WATER MANAGEMENT ................ 147
DISCUSSION OF THE HYPOTHESES ......................................................... 149
CONCLUSIONS ...................................................................................... 151
RECOMMENDATIONS FOR MANAGEMENT PRACTICE ............................. 152
RECOMMENDATIONS FOR FURTHER RESEARCH .................................... 153
REFERENCES ..................................................................................................... 159
LIST OF FIGURES.............................................................................................. 167
ABOUT THE AUTHOR ...................................................................................... 173
SAMENVATTING ............................................................................................... 177
17
1
Introduction
1.1 Background
This research has been carried out under the auspices of UNESCO-IHE.
UNESCO-IHE is a post-graduate educational and research institute entirely
devoted to the aquatic environment. It is concerned with water resources
management challenges worldwide, in particular, the challenges faced in less
privileged countries. Every year, 200 water professionals from all over the
word arrive in Delft, the Netherlands, to study for their MSc degree at
UNESCO-IHE.
One of the most confronting new experiences these students report when
arriving in the Netherlands, is the cold, rainy, and highly variable weather
(the academic year at UNESCO-IHE starts in October). These same
changing weather conditions also form a challenge for the Water Boards that
are responsible for the daily operation of the regional water systems. At the
start of this research in 2004, it had become apparent that although many
water boards had installed, or were installing, Decision Support Systems
(DSSs) with real-time weather monitoring and forecasting data and
hydrological simulation models to anticipate better the changing weather
conditions, many questions remained on how this wealth of information
could then be used best in practice. How could Water Boards assess the
quality of the weather information? How could they deal with errors in the
data, and with uncertainties in the meteorological and hydrological
forecasts? How should they make decisions to take anticipatory actions?
These same questions are just as relevant to any part of the world where
extreme rainfall events or prolonged periods of limited rain may cause floods
and droughts or agrevate water quality problems. The DSSs and hydrometeorological forecasting tools are becoming readily available to
developing countries. Therefore, the research fitted the mission of
UNESCO-IHE in general and the objectives of the Hydroinformatics group
at the institute in particular.
1.1.1 Hydroinformatics and Integrated Water
Resources Management
Hydroinformatics (Abbott, 1991) is the science of information and
communication technologies in integrated water resources management.
Integrated water resources problems are complex, and ICT, including
18
Anticipatory Water Management
computer simulation models and computer presentation tools, helps water
experts in their analysis. The use of this digitised, virtual world (Price, 2008)
is often preferred over physical experiments in the real world, because risks
associated with physical experiments in relation to water resources are too
high, and time and budgets too limited. Measuring what is happening to the
water resources is a prerequisite for informed management of these
resources. Ongoing developments in real-time monitoring, both in ground
station telemetry and remote sensing, and communication of the monitored
data into fast and easily accessible data bases, have greatly enhanced the upto-date information about the state of the water resources to be managed.
Also, water experts are not the only people who need the help of ICT to
analyse water resources. Just as important, and indeed still increasingly
important, is the communication of information about the water resources
and their management, to policy makers and the public. It is here that we
realise that hydroinformatics is a socio-technology (Abbott, 1999).
Developments in society, such as the full integration of the internet and
mobile telephony worldwide, influence how ICT and Integrated Water
Resources Management can best be combined. Hydroinformatics, in its turn,
influences the way integrated water resources management is performed,
increases the number of people involved and concerned, and as such
influences society as a whole. Hydroinformatics, through the application of
ICT, strives to make integrated water resources management available to
even the least privileged societies.
1.1.2 Management of extreme events
Extreme events often (temporarily) unbalance the management of water
resources. Many places on earth face extreme events, like floods and
droughts, with devastating effects. In the Netherlands the most recent river
floods occurred in 1993 and 1995, when 240,000 people and one million
animals were evacuated. Economic losses amounted to more then 100
million US dollars (Moll et al., 1996; Boetzelaer and Schultz, 2005).
Whereas in the past, local problems due to extreme events may have been
analysed in isolation, today, with the growing insights into the hydrological
cycle and the interdependence of the different components of the natural and
anthropogenic systems, the need for integrated water resources management
becomes more important. The ability to analyse the water system at bigger
spatial (catchments) and temporal (seasons, years) scales has improved
considerably over recent years. It enables the water community to look for
management practices that benefit both drought and flood management on
the local and catchment scale. Part of these management practices is
operational water management. This document presents a study of the
mitigation of the negative impacts of extreme events by enhancing
operational water management. Because there will always be events that are
Introduction
19
too extreme to be managed in any way, in this thesis we refer to critical
events, to indicate the group of events that do permit mitigation actions.
1.1.3 Operational water management
First we have to understand the scope of operational water management.
What would we do if we were responsible for today's operation of the Dutch
Delta Works, or of the Dutch pumping stations? Is there a storm surge?
Should we close the barriers? Will there be a rain storm tomorrow? Should
we activate the pumps? This dissertation deals with such operational water
management questions, and contributes to the enhancement of this
management. It does not however, deal with the design of water systems and
structural changes.
To stay focussed on operational water management, we first address some
definitions; starting with the term itself.
Operational water management is the set of day-by-day
decisions and subsequent actions that interact with the
water system.
In this dissertation, a system, according to Oxfords dictionary, refers to "a
group of related things or parts working together".
A water system thus becomes a set of water bodies with
their conveyance and regulating structures that work
together through natural and artificial processes.
Most of today's water systems affect people, and are affected by people.
People are interdependent on water systems through their roles as
beneficiaries, extractors of water, and, last but not least, managers. It should
be noted that through the involvement of people, the whole will not behave
in a systematic manner, and as a consequence the word "construct" is
preferred over "system" (Abbott, 2005). However, because the use of the
word "construct" is unfamiliar to most people, the word system is used in
this dissertation. As water managers, people interact with the water bodies
through the design and operation of structures, and with people themselves
through, for example, water supply systems, consumption regulations, and,
more incidental interactions, such as evacuation in times of flooding.
The objective of operational water management is to
maximise the benefits of water bodies for society.
According to developments in sustainable and integrated water resources
management, it is now commonly accepted that the maximisation of benefits
20
Anticipatory Water Management
should be done in such a way that today's people, as well as future
generations, benefit from the water resources. Through the objective of
making optimal use of available water and water systems, operational water
management inherently contributes to the benefits of water, taking acount
both of today's climate, as well as a future, a changed climate.
Operational water management has gone through a long history of
development. An overview of this development, with focus on the
Netherlands, was given by Lobbrecht (1997, pp 4-13). In the area of the
Netherlands active water management began around 800 AD with digging
ditches to divert water. The developments afterwards, up to the 20th century,
are mainly characterised by the expansion of available structures to manage
water. First dikes, then dams, windmills, series of windmills, and finally,
electrically powered structures became available. Examples in the
Netherlands include controllable or regulating structures such as the
Oosterschelde Storm Surge barrier and the Maeslandt Barrier to protect the
coastal areas, and the hundreds of diesel and electric pumping stations to
manage the land-reclamation areas (called "polders").
In the 20th century the technical advances in regulating structures levelled
out, and the focus of technological research in water management moved
towards the methodologies and means to operate all the structures
efficiently. These developments profited strongly from developments in
Hydroinformatics in general, and Real-Time Control (RTC) in particular.
The present state-of-the-art is the incorporation of developments in RTC
with monitoring networks, communication systems, data bases, hydrological
modelling software and decision support tools in complete hydroinformatic
systems for maximising the benefits of operational water management.
Yet, with all these technological means in place, events still occur that cause
much damage to the water system. It is the starting point of this dissertation
that further reduction of these damages, within the constraints of the system's
capacity, should be sought by enhancing the use of predictions of future
states of the water system. It is by increasing the forecast and decision
horizons, that water managers, like chess players who are able to think
several turns ahead, can further increase the efficiency of their water
systems.
1.1.4 Benefits of increased forecast horizon
Increasing the forecast horizon can be achieved, not only by using real-time
monitored meteorological variables as input to hydrological simulations, but
also by using forecasted meteorological variables as input to the models.
Monitoring proved to be insufficient to provide the required forecast horizon
Introduction
21
to take management actions for part of the critical events. As a consequence,
hydrological simulation models were developed and applied. These in turn
do not always provide the required forecast horizon, hence the need for
inclusion of forecasted meteorological inputs.
With the modelling studies, uncertainties in the predictions became apparent.
The hydrological systems and certainly the meteorological systems that drive
hydrological critical events are chaotic systems, in the sence that small
changes in the present state lead to large changes in the future state of the
system. Inherently, small errors in the initial conditions of a model lead to
big deviations in the predicted states on the one hand and the actual states on
the other. In response to this problem, ensemble modelling techniques have
been developed, and their further enhancement and use in practice is today at
the forefront of scientific research (Schaake et al., 2006).
1.1.5 Use of weather forecasts
Developments in numerical atmosphere modelling and atmosphere remote
sensing have resulted in a readily available suite of meteorological products
for water professionals. National weather services offer model output time
series and images directly, e.g. through File Transfer Protocol (FTP), to
water management agencies that can automatically process and forward
these time series as input to hydrological models for decision support. Next
to this increase in real-time availability of meteorological data, another
important development is the use of re-analysis and hindcasting. Hindcasting
means that when a new meteorological product becomes available a data set
is prepared of what would have been the results of the product if it had been
used for the past so many years. This data set can be used to compare a new
product with the old products and to train the use of the new product.
The need or significance to include weather forecasts in the preparation of
water system predictions differs per application and type of water system
(Figure 1.1). For management actions that need little time to become
effective, like the control of weirs and gates in irrigation canals, not much
lead-time is required and therefore the significance of weather forecasts is
less. Early warning and evacuation measures take a long time to become
effective, but for large rivers long forecast horizons can be achieved using
upstream measurements and river simulation models, without using weather
forecasts. Therefore, also in this case the significance of weather forecasts is
limited (Figure 1.1). For flood control measures that need a long time to
become effective, like lowering reservoir levels, in fast responding
catchments and in catchments where flooding problems follow directly from
extreme rainfall events (pluvial flooding), the significance of the use of
weather forecasts is very large. Also for drought management, where
seasonal forecasts are needed, the significance of long-range weather
22
Anticipatory Water Management
Significance of
meteorological forecasts
forecasts (rainfall and temperature) is large. In Figure 1.1 a number of
application areas for the use of weather forecasts in operational water
management have been tentatively positioned according to their required
lead-time and the significance of the meteorological forecast.
Pluvial flood control
Drought management
Flash floods warning
Hydropower
Long term
Urban drainage
Water supply
River flood control
Irrigation
Hydropower
short term
River flood early warning
and evacuation
Required lead time
Figure 1.1 Significance of meteorological forecasts for operational water
management applications
1.1.6 Ensemble forecasts
Ensemble forecasts are forecasts that contain a number of alternative
predictions for the same forecast period. One such prediction is called an
ensemble member. The differences between individual members can be the
result of differences in expert opinions, in atmosphere simulation models
used, or in the initial conditions used for the models, depending of the kind
of ensemble system that was used. In any case, the differences in the
ensemble members provide information about the uncertainty of the
particular forecast. If all the members are more or less the same, the forecast
has a measure of certainty about what is going to happen. If, on the other
hand, the members show large differences, it means that the forecast is
highly uncertain.
Ensemble forecasting to provide this real-time estimates of forecast
uncertainty has become common practice for the bigger international
meteorological organisations, such as the European Centre for Medium
range Weather Forecasting (ECMWF) and the National Centre of
Environmental Predictions (NCEP). The ensemble forecasts have become
available to more and more countries at lower costs or free of charge.
These ensemble weather forecasts can be used to expand the decision
horizon of operational water management to anticipate what may happen.
Introduction
23
1.2 Anticipatory Water Management
Anticipatory Water Management (AWM) is defined as daily operational
water management that pro-actively takes into account expected future
conditions and events. "Future events" refers to events that are not yet
measurable within the catchment. Therefore, weather forecasts must be
applied to prepare the predictions. Examples of Anticipatory Water
Management actions are the lowering of reservoir levels for flood control
and maintaining reservoir levels in anticipation of droughts.
We apply anticipatory management ourselves in every day life. If we expect
the train to be delayed, we take an earlier train to be on time. If we expect
the weather to be warm we dress accordingly. If Johan Cruijf expects the
Dutch soccer team to win, we put our money on the team. In each example
we make an estimate about the credibility, assess the risks and make a
decision. Johan Cruijf is regarded as the expert on soccer and thus we put a
lot of faith in his forecasts and take the risk of losing money.
Also in professional fields, predictions of all kinds are used in management.
Economic and market forecasts are used in organisation management and
product development. Demographic development models and climate
change predictions are used in land use planning, and meteorological
forecasts are used for natural hazard warnings in agriculture, transport
(aviation, shipping, road traffic), defence and healthcare (NHS, 2002).
Applications of meteorological forecasts mostly concern short-term forecast
horizons. If tonight's temperatures are forecasted to be below zero, farmers
protect their crops against freezing. Storm warnings for shipping and
aviation only apply to the present day, and action in many cases is taken only
when the forecasted event is already taking place.
The reason that mid-term and even short-term weather forecasts are in many
cases not decisive in daily management is that these forecasts are considered
not to be accurate enough. As a consequence their weight in the decision
making process is often very small (or the forecasts are not used at all), and
the role that forecasts should play in the decision process is not formalised. It
is left to the judgement of experts and managers what to do with them. In the
case of the farmer, wetting his crops to protect them from freezing, this is
not a problem, because the costs of the action are low. If the weather forecast
turns out to be wrong, the economic damage to the farmer is little. If the crop
is lost due to frost, the damage is huge, so the farmer's risk analysis leads to
a clear decision to protect his crops.
24
Anticipatory Water Management
With daily water management, especially when concerning critical events, a
risk analysis of the use of an uncertain forecast is more complicated. Costs
of pro-active management actions like an evacuation and controlled flooding
are high and thus a decision is delayed as long as possible. Lead-times of
management decisions and response times of actions (like an evacuation)
play a crucial role. In order for pro-active management to be effective the
decision has to be made before the actual critical event is due within the
response time of the management action.
Despite these difficulties, there are currently many reasons to put effort into
improving Anticipatory Water Management:
- Global annual loss of life and socio-economic costs due to extreme
events are still very high, and even increasing;
- Next to long term (structural) prevention and mitigation strategies,
there is a need to do what we can with the water systems that we
have now;
- Ever increasing human pressure on natural resources and growing
economic constraints call for optimal use of available water systems,
before large scale and expensive structural changes are made
(WB21, 2000, p. 51);
- New meteorological observations and forecasts have been (further)
developed, such as radar, satellite and, very importantly, ensemble
forecasts;
- Qualitative and quantitative water-system response models have
improved. Recent hydrological science has put effort in uncertainty
analyses of these models, enabling risk analyses and, therefore,
better informed decision making;
- Anticipatory management actions for critical events are available,
like lowering storage reservoir levels for flood control and
maintaining water levels to prevent droughts;
- In many different fields scientific and practical progress has been
made on risk analyses and decision-making.
The basic process of Anticipatory Water Management distinguishes four
steps. First, the present state of the water system has to be determined, and at
the same time meteorological forecasts of atmospheric variables, such as
precipitation and temperature, have to be acquired. Then the response of the
water system to the atmospheric variables can be predicted, resulting in a
forecast of the state of the water system. On the basis of all the information
acquired, operational management decisions have to be taken, taking into
account the uncertainty of the forecasts and the risk of each management
option. If management actions are required, these need to be implemented.
This basic concept of Anticipatory Water Management is illustrated in
Figure 1.2.
Introduction
Assess present state of the system
25
Forecast the atmospheric variables
Take a management decision and implement
management actions
Predict the response of the water system
Figure 1.2 Basic representation of the process of Anticipatory Water Management
1.3 Hypotheses and objectives
The main incentive and objective of this dissertation is to:
Improve the use of weather forecasts in operational
water management.
The present ways of using weather forecasts in operational water
management are analysed, challenges in using weather forecasts identified
and methods to meet these challenges are presented. These methods will
contribute to more effective pro-active operational measures, such that
operational practice can add to its real-time function a focus on Anticipatory
Water Management.
The main hypotheses that are proposed in this dissertation to support the
main objective are:
The use of ensemble precipitation forecasts to decide on
anticipatory control actions, in preference to re-active
control, can reduce the damage costs over a long period
of time.
Long-term simulation of the complete Anticipatory
Water Management strategy for a historic time series
enables an optimisation of the strategy.
Firstly, these hypotheses reveal a focus on recent developments in ensemble
forecasting in operational meteorology. Verification methods from
meteorological sciences have been applied and adjusted to analyse these
ensemble weather forecasts and water level forecasts derived from the
ensembles. Secondly, the focus of the methodologies is on the capitalisation
of enhancements in hydroinformatic systems, such that extensive computer
simulation of the operational water management strategies can be performed
to develop and evaluate novel strategies.
26
Anticipatory Water Management
Next to these methodological objectives and hypotheses, this dissertation
provides an assessment of the present day potential and limitations of
Anticipatory Water Management for two case studies: an extensive case
study of the Rijnland water system in the Netherlands, and a case study in
the Upper Blue Nile region in Ethiopia. Both these case studies concern the
application of Anticipatory Water Management in flood management. In
both these case studies the consequences of extending the forecast horizon
from a maximum of 1-day, to 3-days or more are analysed.
1.4 Reader
In Chapter 2 a literature review is presented to identify knowledge gaps and
hypotheses (listed in Section 2.6). In Chapter 3 theoretical and
methodological concepts are explored to develop a framework for enhancing
Anticipatory Water Management. In Chapter 4 and 5 this framework is
applied to the two case studies. The dissertation concludes with a discussion
on the hypotheses, and conclusions and recommendations for application of
and research on Anticipatory Water Management.
27
2
Anticipatory Water Management
2.1 Introduction
In this chapter the different aspects of Anticipatory Water Management are
reviewed. The introduction to operational water management is elaborated
further. Weather forecasting, water system modelling, and decision making
under uncertainty, are discussed in detail as key issues in Anticipatory Water
Management.
2.2 Operational water management
2.2.1 Definition
Operational water management has been defined in chapter 1 as the set of
day-by-day decisions and subsequent actions that interact with the water
system. Operational water management concerns the daily control of water
systems. It is performed to try and prevent water from threatening human life
and to optimise its use for functions we consider important. Operational
management is not concerned with the development of policy guidelines and
the structural design of water systems. In the first place, monitoring and
issuing early warnings in the event of calamity threats are very important
tasks. Secondly, in most modern systems, management involves operating
several regulating structures to minimise the frequency of calamities while
the requirements of stakeholders are adequately (or optimally) met.
2.2.2 Components of operational water management
Operational water management consists of many components and tasks. The
basic elements are:
- Structure and facilities
- Monitoring
- Objectives
- Management
The structure and facilities are, e.g. the river beds, canals, and embankments
that contain or convey the water body, and the regulating structures, such as
pumping stations and weirs. The monitoring includes everything that
(contributes to) the provision of up-to-date information on the state of the
water system. The word 'objectives' refers to the requirements of people, the
ecology, and the water body itself. Present day operational water
management has to take into account the requirements for these other
28
Anticipatory Water Management
beneficiaries (ecology and water systems as such) as well. Even remote
water systems, like rivers where no people are living, are nowadays partly
managed to preserve the right conditions for the ecosystems which the rivers
are part of. Without requirements that are not continuously met by selfregulation of the water body, there is no need and no direction for
operational water management. When concerning critical events,
requirements are, for example, the maximum frequency of damage
occurring, or the acceptable total damage over a certain period.
Management is done by the people responsible for making the daily set of
decisions and the actions concerning the operation of the water system. Note
that the decision makers and those taking subsequent action are grouped
here, because theoretically it could be just one person doing all the work. In
reality in almost all water systems the tasks of managing and implementation
are divided between several people, and implementation is often done
through automatically controlled structures. In the case of flood warning and
evacuation, for example, there is usually a team responsible for the decision
of issuing the evacuation order, while other people are responsible for the
communication and execution of this order. In the end, even the evacuees
themselves play an important role in making the evacuation effective. On the
other end, there is the example of an automatic weir. Such a weir is
controlled automatically (control actions are received from automatic
functions, called controllers) and implements the regulating action, every 5
minutes, say, without interference of a manager. The task of the operational
manager here is to check whether the settings of this automatic process are
still effective, or need to be adjusted.
Next to elements of operational water management, tasks of operational
water management can also be defined. These are:
- Development of a management strategy
- Daily management, operation and control
- Maintenance
- Evaluation of the operational management
With daily management we refer to the operational decisions that have to be
taken on a regular basis. The frequency is at least daily for almost all
systems, but often can be hourly or even every minute. So the term "daily"
does not exclude other frequencies here, but is used because it clearly refers
to the regularity of the decisions to be taken, as opposed to the determination
of the management strategy.
The present research focuses on the management strategy and the daily
management. It does not focus on maintenance, because for the case studies
the existing water systems, including their design, regulating structures,
monitoring networks and requirements are considered as given. Within the
daily management the focus is on the information and decision support
Anticipatory Water Management
29
systems that can be used to optimise the decision making. The process of
actually taking the necessary actions after the decision is made is also
considered, to ensure that the proposed strategies are realistic.
In operational management there are many differences between management
of a mainly natural system, such as flood warning and evacuation for large
rivers, and management of strongly controlled water systems, like an urban
water distribution network. The operational management of the latter is
referred to as Water System Control. The major difference is the presence of
regulating structures.
2.2.3 Water system control
In Chapter 1 the shift in development of water system control in the
Netherlands from developments of control structures (starting from 1200
AD) to the development of control strategies (in the 20th century) has
already been described. Most research and development has been done on
these control strategies and is still being performed within the research
community of Real-Time Control.
In the present day, the term "Real Time Control" is widely used. At first
sight it seems like a self-evident term for something that has already existed
for a long time. Operational water managers always make their decisions in
real time. However, the "time" in "Real Time Control" does not reflect the
absolute time of action (past, present or future), but the time origin of the
data used and the moment the decision is made relative to the event of
concern. Schilling (1990) defined real time control as "a synonym for the
manipulation of a process during its evolution". Traditionally control
strategies are pre-set following historic system analyses. Sophisticated time
series analyses and state-of-the-art, physically based scenario models will
continue to be used to define the boundaries and guidelines for water system
control. Examples are heights of river dikes, upper and lower boundary
reservoir levels, and water allocation schemes. In contrast, real time control
uses system state, prediction and user demand data that are as up-to-date as
possible, and then seeks the optimal control strategy. Inherently the control
strategy is continuously updated. Real time control seeks the best control
strategy for the given moment within the constraints set by historically based
boundaries.
The desirability to control in real time, or better "as soon as possible", has
resulted in developments in several areas. Measuring devices have been
modernised. Telemetric networks, radar and satellites have dramatically
increased the data availability and acquisition speed. Data assimilation has
become a very important research area. It focuses on state-of-the-art data
processing techniques, optimising their use in computational models.
30
Anticipatory Water Management
Despite the development of the computational power of computers, the time
taken to solve the optimisation problem for RTC using these computational
water system models can become the limiting factor. In response much
research has successfully adopted machine-learning techniques to replicate
hydrological modelling components using an order of magnitude less
computational time (Lobbrecht et al., 2002; Bhattacharya et al., 2003;
Lobbrecht and Solomatine, 1999).
Feedforward-feedbackward control
Real-time control has moved from simple manual control to complex multiobjective automatic control. When the system is controlled based on
measurements of the target variable it is called feed-backward control. The
advantage is that the effect of the control action (e.g. pumping) is monitored.
When the system is controlled based on measurements of disturbances (e.g.
rainfall) and the modelled effect on the target variable, it is called feedforward control. Because the target value itself is not measured, combined
feed backward-feed forward control is usually preferred in water
management.
Model Predictive Control
Originally, in feedforward-feedbackward control the control for each time
step is aimed at minimising the deviation from the target variable value by
counter balancing the predicted effect of a measured disturbance. In a
drainage system, for example, the predicted run-off for the coming hour is
equalled by a control response of a pumping station by a discharge of the
same amount, to keep the water level unchanged. This works fine for many
cases and is still used in many controlled water systems. However, in cases
when the control capacity is not enough to compensate the effect of the
disturbance, the target variable value cannot be maintained. Then an
intended deviation from the target variable value in opposite direction may
be desired to limit the maximum deviation over several control time steps.
To perform this automatically a control algorithm (controller) is needed that
takes into account constraints of the water system, and is able to perform an
optimisation on the basis of objective functions to assess whether, when and
to what extent intended deviation from the target variable value is desirable.
A widely known controller that has been developed to accommodate these
requirements is Model Predictive Control (Overloop, 2006; Weijs et al.,
2007).
Global control
In many cases, when facing extreme hydrological loads, not all of the water
systems storage capacity is used at the moment of failure. Especially in
sewer-system engineering, global (or central) control systems have been
used and have shown that flood management can be improved by reallocating water to other sub-systems before the most critical moment. To
achieve this, in global control, several structures, controlling different
Anticipatory Water Management
31
sections of the water system, are operated in such a way that the water is
optimally distributed within and discharged from the system. This is an
enhancement compared to local control, where a structure is operated based
only on the system state in its direct vicinity.
Dynamic control
The adjustment of the control of the water system with changing spatial and
temporal requirements is called dynamic water management or dynamic
water system control (Lobbrecht, 1997). An example is taking into account
the seasonal change in the requirements and risks of the agricultural sector.
Regardless the type of control, effectiveness depends on the available
storage or the throughflow capacity in the water system. Reservoirs and
polder are examples of systems that typically have a lot of storage.
2.2.4 Reservoirs and polders
In rivers, irrigation systems, and drainage systems the regulation often
involves reservoir control. Large river reservoirs traditionally have been
subject of intensive optimal control studies and practices. This is mainly due
to the high economic value of the reservoirs. The physical properties of the
reservoir are well defined such as the geographical boundaries, water volume
and water level. Control is usually straight forward using sluices to adjust the
reservoir water level. Determining the optimal real-time operation, however,
is far from easy. Most reservoirs have multiple functions such as
hydropower generation, water supply (irrigation, drinking) and safety against
flooding (multipurpose reservoirs). Therefore multi-criteria or multiobjective functions have to be satisfied to arrive at the best operation. The
most difficult problem, however, is the prediction of the future state of the
reservoir. This involves meteorological and hydrological calculations
throughout the entire upstream catchment and the determination of user
demand.
Due to developments towards integrated water management in which the
catchment hydrology plays a central role, research has expanded to include
multi-reservoir operations (Huang et Yuan, 2004). The focus of these
authors is on drought early warning and the practical use of real-time
control.
In irrigation and drainage, the system of canals together forms a special kind
of reservoir. In the Netherlands these regional irrigation and drainage
systems have received their own terminology. "Polder" is a Dutch word
referring to a low-lying land area that has been reclaimed from the water
(lakes or seas). Water is drained from the land and pumped into the main
conveyance channel, which is called a "Boezem". The excess water from this
32
Anticipatory Water Management
reservoir is discharged by pumps, weirs or sluices to rivers or the sea.
Schultz, 1992, provides an historic overview of design and water
management of the Dutch polders.
Polder management has made a shift, and is still in transition, towards
integrated water management. Water systems are considered in their entirety
and interests of different stakeholders are being taken into account
(Bhattacharya et al., 2003). There is also more emphasis on seeking
economic and optimal tailor-made solutions (Schultz, 1992; Wandee, 2005),
whereas previously, average-based, policy guideline solutions were
sufficient. This results in multi-objective or multi-criteria approaches in
which various interests are balanced while minimizing damage costs
(Lobbrecht et al., 2002).
In reservoir and polder management flood early warning and control often
plays an important role.
2.2.5 Flood early warning and control
The present research has chosen applications in flood management because
of the high relevance for the Netherlands and many developing countries,
and because of the strong developments in Quantitative Precipitation
Forecasting for the short to medium range (up to 15 days).
Therefore the focus is on rainfall induced flooding in fresh water systems.
The flooding occurs because the hydrological load exceeds the capacity of
the water system. The capacity is a combination of the discharge capacity
and the storage capacity. Depending on whether and which structures are
present, the discharge capacity consists of river or channel flow capacity and
discharge structure capacity (e.g. pump or sluice), and the storage capacity
consists of the "in bank" storage and the (emergency) storage basins and
reservoirs.
Effective flood management integrates structural and non-structural
measures, and long-term planning and operational preparedness (Price,
2006). Possible measures to reduce the negative consequences of flooding
are to move human and economic activities out of the flood prone areas
permanently or through early warning and evacuation. Non-structural
measures to reduce the frequency and magnitude of this kind of flooding,
consist of changes in land use that increase the water retention capacity of
upstream parts in the catchment, for example, by planting forests to prevent
erosion and subsequent surface runoff. Structural measures to reduce the
frequency of this kind of flooding are to increase the discharge capacity of
the control structures or the storage of the water system.
Anticipatory Water Management
33
All these measures take a long time before they become effective, and are
very expensive and spatially demanding. In today's world generally time,
money and space are scarce. Hence it is important to also optimise water
management within the current water system capacity. Maximising the use
of the current water-system capacity is even more important when flooding
problems are urgent or when structural measures are simply not sufficient to
meet safety standards.
To further improve flood early warning and control, decisions have to be
made on the basis of weather forecasts. This is defined as Anticipatory
Water Management (AWM). Where the application concerns the operation
of regulating structures it is called anticipatory water system control, or just
anticipatory control.
2.2.6 Challenges in operational water management
Most of today's operational water management is still re-active or includes
only hydrological predictions in feed-forward or model predictive types of
management. Yet, at the same time, many water authorities and enterprises
already receive and use weather forecasts and their number is increasing
rapidly. While this expert based experience is already there, publications on
these experiences are limited. Especially questions remain about how to deal
with the high level of uncertainty when dealing with weather forecasts
(Lobbrecht, 1997; Overloop, 2006). Therefore, there is a need on research to
the end-use of weather forecasts in operational water management, which
can result in a framework to develop and evaluate Anticipatory Water
Management strategies.
This research is application oriented. Application oriented research involves
not only academia, but also practitioners. The first group relies on theorems
and scientific research, while the latter group is used to adopting a systems'
approach in which confidence in a new system is built up over time through
experience and feed back loops. In water resources engineering, and in this
research in particular, such confidence is translated in the trust
hydroinformaticians put in the modelling of processes, while operational
managers trust in running extensive system tests. Practically oriented
research should accommodate both requirements. Therefore Section 2.2
discussed the different operational water management practices, while
Section 2.3, weather forecasting, and Section 2.4, water system modelling,
address the processes and models that govern the systems' operation. In
Section 2.5, decision making, we discuss how the academic and the system
approach can be brought together to foster change. In particular re-analysis
and verification as empowered by modelling systems are elaborated
throughout the remainder of the dissertation as the vehicles for bridging the
gap between theory and practice.
34
Anticipatory Water Management
2.3 Weather forecasting and ensemble predictions
Currently the traditionally distinct disciplines of meteorology and hydrology
are moving towards each other. Meteorologists try to incorporate the needs
of hydrologists in their research programmes and hydrologists look for ways
to optimise the potential of meteorological data (Lobbrecht and Loos, 2004;
Lobbrecht et al., 2003; WMO, 2004).
Synoptic meteorology and climatology deal with the weather on small and
large (averaged) time scales respectively. For operational water management
the actual and the expected near future (maximum one year ahead, if we do
not consider planning of groundwater extractions and such like) states of the
water system are most important. Therefore the focus of this section is on
synoptic meteorology, e.g. weather measurements and weather forecasts.
2.3.1 Monitoring systems
The backbone of meteorological science is still the global monitoring
network of ground stations, ships, data buoys and radio sondes, (almost)
directly measuring state variables such as temperature and wind. Many other
variables and measurement techniques require translation of the measured
parameter into the state variable, e.g. from satellite derived cloud-top
temperature to precipitation. Together with the interpolation from point or
local measurements to a 4-dimensional description of the state of the
atmosphere, this translation is called weather analysis.
In addition to the classical measuring devices such as thermometers for
temperature and tipping buckets for precipitation, the last decades have seen
the development of modern techniques like automated weather stations,
radar and remote sensing from satellites. The great advantage of the latter
two over local measurements is their spatial coverage. Satellites can have
global coverage, thus filling significant gaps in the observational network.
Weather satellites
The first satellites were launched in the early 1960's to make visual and
infrared images of the earth. Polar satellites orbit Earth in a north-to-south
direction at relatively low altitudes and obtain images of the entire globe in
12 hours. Geo-stationary satellites remain at a fixed point above Earth at the
expense of a greater distance and therefore of detail on the images. Most
applications still involve visual tracking of weather systems and potential
rain storms, but intensive research is going on to enhance the quantitative
analyses of satellite information. Wind speed can be estimated from cloud
movement and temperature at different heights as deduced from radiation
measurements of specific wavelengths.
Anticipatory Water Management
35
Weather radar and now-casting
The use of RAdio Detection And Ranging (RADAR) for weather analyses
has increased considerably. Conventional radar sends out electromagnetic
waves and detects the reflected proportion and delay. These can be
interpreted as a measure of the rainfall intensity and the distance of the storm
respectively. Doppler radars also measure differences in frequency of the
sent and reflected pulse. An increase of frequency indicates movement of the
storm towards the radar. Two or more Doppler radars can therefore
determine the direction and speed of storms. This technique is used for
tornado tracking in the United States. When applied near real time it is often
called now-casting (Lutgens and Tarbuck, 2001, p. 293, p. 329).
Rain radar provides users with the spatial variability of precipitation on a
resolution that is almost never met by ground station networks. These
ground stations can be used for calibration. The Royal Netherlands
Meteorological Institute (KNMI) operates two Doppler radars and provides
calibrated radar precipitation sums every 24 and 3 hours and non-calibrated
sums every 5 minutes, both on a 2.5 km and a 1 km grid.
The most recent technical development is the positioning of a weather radar
on board of a satellite. The Tropical Rainfall Measurement Mission
(TRMM) is the most famous example. Japan and the United States work
together on this project. The Satellite covers the tropical band (35N to 35S)
and the project provides several rainfall analysis products, such as real-time
3-hourly precipitation estimates (NASA, 2008). Another feature of this
project marks a very important development in meteorology, namely that the
products and research results from this satellite based weather radar are
freely accessible through the Internet. The increasing availability of
meteorological data at no or low cost is a land mark development, that
benefits science and society in general, and developing countries in
particular (Akhtar et al., 2009).
Data availability and assimilation
With the growing measuring network and techniques, the available data has
increased tremendously. Fortunately techniques of gathering, storing,
processing and presenting this data have also been improved. Automated
weather stations communicate their data to a central computer based system
(telemetry), without intervention of human observers (Lobbrecht and Loos,
2004, Lobbrecht et al., 2003). Data assimilation techniques are used to
model accurate initial fields for numerical weather prediction (Falkovich et
al., 2000).
The present research uses rain radar and satellite derived radar indirectly
through the use of numerical prediction models, but also directly in the
hydrological modelling for the case studies. Ground based rain radar from
the Royal Netherlands Meteorological Institute (KNMI) is used for the
36
Anticipatory Water Management
Rijnland case study (section 4.3) and TRMM data is used for the Upper Blue
Nile case study (section 5.3.2). For both case studies ground station
measurements are used as well.
2.3.2 From hand-drawn weather maps to numerical
prediction
A prediction of the state of the atmosphere at a certain time and certain place
can be based on observations and knowledge of past events and physical
relationships. This is called a weather forecast. The state variables that are of
most concern to water managers are wind direction and speed (for open
water levels), evaporation (water balance) and most of all, precipitation.
Therefore Quantitative Precipitation Forecasting (QPF) is a lively scientific
discipline on its own. Precipitation typically results from either small-scale
convective weather systems or large-scale systems (fronts). The first is more
difficult to forecast, because it concerns local events of high intensity
precipitation.
Mostly three major groups of forecasts are distinguished based on their leadtime. The classification is not universal but in many cases up to 2-days ahead
forecasts are referred to as short-term forecasts. Medium-range forecasts
concern lead-times up to 10 days ahead, and long-range forecasts concern
monthly and seasonal forecasts (Persson and Grazzini, 2007).
Meteorological science has seen the development of forecasting
methodologies from synoptic forecasts, via deterministic numerical weather
prediction (NWP) to probabilistic forecasts.
Synoptic forecasting
Traditionally the collected data are presented on maps. These synoptic
weather maps, containing amongst other features isobars and wind direction
and speed, are then used by meteorologists to extrapolate future conditions.
Over the years, as experience had grown, the analogue forecasting method
became the most important. Forecasts were made by comparing current
patterns with patterns from the past. From this experience, rules of thumb
were developed that still serve an important role in short term forecasting.
The expertise of individual forecasters is of decisive importance for the
quality of synoptic forecasts.
Numerical weather prediction
In numerical weather prediction processes and dynamics of the atmosphere
are mimicked using physical laws. The gas law and the hydrostatic law
describe the static relation between variables (diagnostic), and the equations
of continuity, motion and the first law of thermodynamics describe the
dynamic changes (prognostic). More specific processes are described using
Anticipatory Water Management
37
parameterisation schemes, such as a cloud scheme. The idea of numerical
weather prediction had already been posed in 1904 by Vilhelm Bjerknes, but
only in the 1960's were the theory sufficiently developed and the
computational power (computers) advanced enough to come up with the first
general circulation models. To solve the equations the atmosphere is
discretised in vertical layers and a horizontal grid. Together with the
required time step, suitable numerical schemes have to be chosen to process
the forecast (Persson and Grazzini, 2007; Lutgens and Tarbuck, 2001).
Based on the spatial coverage of the models a distinction could be made
between Global (Circulation) Models and Limited Area Models or Local
models. Operational global models are hosted, for example, by the US
National Centres for Environmental Prediction (NCEP) and the European
Centre for Medium-Range Weather Forecasts (ECMWF). The ECMWF
model has a spatial resolution of approximately 25 km and a time step of 15
minutes. The model output time step is 6 hours and the forecast horizon, or
lead-time, is 15 days.
National weather services develop their Limited Area Models (LAM). The
boundary conditions for these models are mostly provided by the global
models. The LAM is then nested in the global model. The Royal Netherlands
Meteorological Institute uses the HIRLAM model, which has a spatial
resolution of 11 km, an output time step of 1 hour and a lead-time of 1 day.
For certain applications, such as flood forecasting in mountainous areas,
these resolutions are still not sufficient. Therefore downscaling techniques
are being developed to try and bridge this gap (Ferraris et al., 2003).
2.3.3 From deterministic to probabilistic forecasts
The models described in Section 2.3.2 produce a single valued output for
atmospheric variables (i.e. precipitation depth) at a certain place or area and
at a certain moment in the future. This is called deterministic forecasting. In
reality these forecasts are uncertain and therefore will often contain errors.
Because of these errors, and perhaps to a greater extent because of the lack
of knowledge and communication of the uncertainty in the predictions, the
general public and end-users put little trust in the forecasts. Moreover, for
proper decision-making, the uncertainty has to be known to carry out a risk
analysis.
Already in 1965 precipitation probability forecasts were issued. Probability
refers to the chance that an event will occur and is represented as a number
between 0 and 1 or as a percentage (Lutgens and Tarbuck, 2001). The
probability is mostly estimated by expert forecasters, based on the above
described synoptic maps, deterministic model output and tracking tools like
radar and satellite images (Atger, 2001). Interpretation can be very difficult
38
Anticipatory Water Management
and presentation by the forecasting bureaus misleading. In the Netherlands
for instance a common way of presenting precipitation forecasts is by giving
a precipitation depth in mm and a probability. Many people do not know that
the precipitation depth is the deterministic output of HIRLAM and the
probability refers to the chance that any amount of precipitation will fall at
any point in the area during the forecast time period. The probability is
therefore not related to the precipitation depth.
To overcome these difficulties methods have been developed that are all
based on a comparison of several model outputs for the same forecast. If the
model outputs are similar it is expected that the probability that the
forecasted event will occur is high. If the model outputs are very different
from each other the probability is low. Apparently the particular weather
system is difficult to predict by the models. Examples are the use of different
LAMs and running the same model using different initial conditions or
parameterisations (Lobbrecht et al., 2003; Atger, 2001). The latter is called
the Ensemble Prediction System (EPS) and accounts for the chaotic
behaviour of the atmosphere ("butterfly effect"). Chaotic phenomena occur
in deterministic systems that are very sensitive to initial conditions. Because
of the chaotic behaviour, small errors in the determined initial state of the
system will result in large errors in the forecasts of the future states of that
system.
2.3.4 Ensemble Prediction Systems
Because of the sensitivity of the atmosphere to initial conditions, the present
state of the atmosphere needs to be known up to a high spatial resolution to
make accurate forecasts. Because it is impossible to monitor the whole
atmosphere with required accuracy and spatial resolution, weather forecasts
are often faced with much uncertainty.
The initial conditions are assessed as accurately as possible on the basis of
the global monitoring network and the interpolation by an atmospheric
model. The monitoring data is interpolated for the whole globe in the three
spatial dimensions and in time, and for all the atmospheric variables needed
to initialise the atmospheric model. The atmospheric model is run at a high
resolution to provide the deterministic ("best guess") forecast.
In Ensemble Prediction Systems the probability distribution of the future
atmospheric state is estimated by running the physically based atmospheric
model repeatedly (e.g. 51 times at the ECMWF), each time with a different
initial state of the atmosphere. The first run is with the initial conditions as
they are assessed for the deterministic run. The spatial resolution of the
model used for the ensemble prediction is often lower to reduce
computational costs. For the other runs, to make up the probability
Anticipatory Water Management
39
distribution, the original initial conditions need to be perturbed (changed) in
such a way that the resulting forecasted states of the atmosphere by each
individual run are equally likely to occur. Together the forecasted states
should form a reliable estimate of the probability distribution of what the
actual state of the atmosphere will become every time step of the prediction.
The method for preparing the perturbations is different for different
institutes. It is often based on statistical knowledge and assumptions of
distributions of monitoring and initial state errors, and most importantly, on
dynamic estimates of analysis error and error growth. At the ECMWF, for
example, samples of possible initial state errors are fed to a simplified, lower
resolution, atmospheric model for a limited forecast horizon (36 hours), to
find iteratively for which perturbations the development of the atmosphere,
in different locations, is changing the most. These paths of rapid changing
atmospheric states are called Singular Vectors, and hence, the approach used
by ECMWF for its EPS is called the Singular Vector approach (Mureau et
al., 1993; Molteni and Buizza, 1996). NCEP uses a method called Breeding.
This method is based on consecutive initial state analyses. The biggest
deviations between two consecutive initial state analyses indicate areas
(locations) of large analysis errors and/or fast growth of analysis errors
(Molteni and Buizza, 1996). By choosing the perturbations, such that the
potential rapid changes in different locations and directions are covered
globally, it is aimed to estimate the (full) probability distribution of future
atmospheric states.
The perturbed initial conditions are fed to the global circulation model used
to make the final ensemble prediction. Each of the forecasts is called an
ensemble member. The result is a number of time series for each of the
surface cells of the global model grid, for all atmospheric variables of the
model. The medium range EPS of ECMWF and NCEP run from 0 to 15 days
ahead. In addition to the perturbations in the initial conditions, also different
parameterisations of the atmospheric model are used to account for the
model uncertainty. The EPS systems for the medium range from Canada, the
USA, and Europe, have been compared in Buizza et al. (2005).
Meteorological institutes like NCEP, ECMWF and the Meteorological
Service of Canada (MSC), provide these ensemble forecasts on an
operational basis to national and regional meteorological and hydrological
organizations (Buizza et al., 2005). These provide Water Boards with access
to consistently generated, near real-time, uncertainty information of weather
forecasts and the means to generate ensemble based hydrological forecasts.
This permits risk based decision making, which has been shown to be more
cost effective compared to decisions based only on a deterministic (or single)
forecast (Roulin, 2007).
40
Anticipatory Water Management
In the Netherlands, the Water Boards are increasingly incorporating the EPS
forecasts of the ECMWF into their Decision Support Systems (DSS). In a
cooperative project, the Royal Netherlands Meteorological Institute (KNMI)
provides warnings for critical events in probabilistic form to several Water
Boards (Kok and Vogelezang, 2006, personal communication). The quality
of these forecasts and warnings for the Water Boards is still under
investigation. On European scale the ECMWF EPS forecasts are used for
flood forecasting and early warning for the main rivers in the European
Flood Alert System (Werner et al., 2005; Thielen et al., 2009) Long-term
verification analysis is needed to develop and test decision rules and control
strategies when using the EPS forecasts (Franz et al., 2005).
So far, most studies with ensemble weather forecasts in water management
applications have not been concerned with water-system control, but have
focused instead on flood forecasting and early warning. Until now, these
studies have mostly been performed on single (flood) events (Roo et al.,
2003; Bálint et al., 2005; Hlavcova et al., 2005). Only a few studies have
carried out a verification analysis for flood forecasting, based on ensemble
precipitation forecasts, over a long period of time (e.g. Roulin and
Vannitsem, 2005). The present dissertation presents long period verification
analyses for water-system control and flood early warning for a water system
in the Netherlands and a sub-catchment of the Blue-Nile in Ethiopia.
Next to the ability to "forecast forecast accuracy", a second important benefit
of EPS is the increased ability to forecast critical events. Because of the
perturbations in the initial conditions, a greater part of the possible spectrum
is represented. Therefore, extreme precipitation events will be forecasted
more frequent, at least by some of the members.
State of the art research meteorologists aims at verifying the probability of
the EPS forecasts of extreme events (Sattler and Feddersen, 2003; Legg and
Mylne, 2003; Atger, 2001). These research efforts to provide reliable
ensemble meteorological predictions are discussed in combination with
related research on hydrological predictions in Section 2.4.3.
2.3.5 Challenges in using weather forecasts for water
management
Institutional legacy and confidence
Rayner et al. (2005) have identified several challenges for water managers in
using weather forecasts. Institutional limitations consist of a legacy of many
years of comparatively successful operational strategies, and of supporting
policies and regulations. There is a natural reluctance to change. Even if
official policy is changed, the lack of personnel who have been educated in
Anticipatory Water Management
41
and are experienced with the new operational strategy limits and delays a
shift in practice. These hurdles would however be overcome if (most of) the
policy makers and operators were convinced of the effectiveness of weather
forecasts and water-system control models. Rayner et al. (2005) have also
identified that the policy makers and operators in general remain to be
convinced. It does not matter if this is due to unjustified conservatism or
justified recognition of the limitations of the modelling systems, because in
either case there is still a need for customized analysis methods for longterm verification of forecasting systems and decision rules for anticipatory
control actions.
Long-term verification analysis is needed for two reasons. First, verification
analysis is required in order to assess the quality of the forecasts for the
particular water system at hand. If this is satisfactory, the analysis builds the
confidence of the operational water managers to use the forecasts in their
decision-making. Second, verification analysis is also needed to develop and
test decision rules and control strategies, given the forecasts. In particular,
the verification analyses permit decision rules to be simulated so that water
managers can see the effect of potential management strategies.
Handling uncertainty: probability forecasts
Another challenge is the analysis, communication and handling of the
uncertainty of the weather forecasts. Krzysztofowicz (2001) describes the
danger of providing deterministic, single forecasts to decision makers. If
such a forecast is considered by the decision maker as representing the
"truth", it could lead to disaster. On the other hand, if the decision maker
realizes that the forecast is uncertain, but he has no information about the
degree of uncertainty, the decision maker may choose to ignore the forecast
and delay the decision until measurements come in.
Ensemble prediction systems provide the necessary information on forecast
uncertainty. The challenge addressed in the present dissertation is to handle
the estimated uncertainty as presented by EPS in decision making for AWM.
Simulation models of controlled water systems (Section 2.4) are important
decision support tools to transform the ensemble weather forecasts into
predictions of water system state, and to analyse the effect of different
management strategies.
2.4 Modelling controlled water systems
2.4.1 Definitions
Water systems are continuously becoming more controlled and less natural.
Therefore this section addresses the modelling of (partly) controlled water
systems and the differences with modelling natural water systems. The focus
42
Anticipatory Water Management
in this research is on instantiating a reliable model of a particular water
system (Chu and Steinman, 2009), not on hydrodynamic and numerical
challenges for computational modelling software (Holly and Merkley, 1993;
Clemmens et al., 2005).
Controlled water systems are systems in which the target variables, or the
system state, is determined to a large extent by control structures, and not by
hydrological processes alone. Only when the hydrological processes go
beyond the control capacity of the system, is the system state mainly
determined by hydrological processes. An irrigation system is a typical
example of a controlled water system. A river is an example of a natural
system. A hydropower reservoir is an intermediate example, where the
influence of the natural forced river inflow is often large compared to the
controllability provided by the discharge structures in the dam. In this
research we refer to the first type, the "fully" controlled water systems, like
irrigation systems and land-reclamation systems (polders). Many studies
focus either on modelling the natural processes, e.g. river flow forecasting
problems, or on modelling (optimising) the control strategy, e.g. water
allocation problems. The water system and its control strategy need
integrated (conjunctive) modelling (Belaineh et al., 1999; Park et al., 2007).
2.4.2 Model components
Models of controlled water systems generaly consist of three main
components:
- Rainfall-runoff from contributing catchments;
- Hydrodynamics of flow in conveyance system;
- Control structures and their operation.
The rainfall-runoff part is included, because except for separated (urban)
water distribution systems, all controlled water systems still have rainfall as
input to their system. The water becomes controllable, only after the rainfallrunoff process has taken place and the hydrological load is being
concentrated in the conveyance system, consisting of conduits like sewer
pipes and irrigation or drainage canals. The hydrodynamics model the
pressures, water levels and discharge in these conduits. In controlled water
systems, the up- and downstream boundary conditions are governed by the
control structures, as compared to natural boundary conditions like sea-level
in natural systems.
The control structures, like pumps, weirs and gates, usually transport water
from an up-steam conduit to a downstream conduit or boundary. The
discharge through the control structures depends on the discharge capacity of
the structure, the operational strategy applied (on-off), the up-stream conduit
Anticipatory Water Management
43
boundary condition and in some cases the down-stream conduit boundary
condition (e.g. head-dependent pumping stations). Next to these three
components, increasingly, biological and chemical model components are
added and integrated with the hydrodynamics to requirements for
(ecological) water quality control (Nestler et al. 2005).
2.4.3 Water system state prediction
The physical processes that govern the state of the atmosphere and water
quantity take place at short time and space distances, such that generally
Eulerian modelling frameworks (Nestler et al., 2005) are used in this study.
In the meteorological forecasting the exception is the application of tracking
models for the prediction of path and travel velocity of depressions or storms
on the basis of radar or satellite images.
Mostly the target variables are discharge or water level. Discharge is a target
variable in an irrigation system, to meet the required volume of irrigation
water. Water level is usually the target variable in drainage systems in landreclamation areas, where water level control is needed to prevent flooding,
drought and soil subsidence. Water quality has long been a secondary target
variable in irrigations systems, where flushing is needed to prevent
salinisation of the soil. Water quality control is becoming more important,
because of ongoing urbanisation and subsequent increasing pressure on
waste water treatment plants and receiving waters.
The water system state predictions in this research concern the hydrometeorological ensemble forecasts. Next to the uncertainty in the
precipitation forecast as expressed by the meteorological EPS, other sources
of uncertainty in hydro-meteorological forecasts are: parameter uncertainty,
model structure uncertainty, initial state (measured, data assimilation)
uncertainty, and statistical uncertainty (Maskey, 2004; De Vriend, 2002).
The aim of ongoing research in hydro-meteorological ensemble forecasting,
is to develop methods such that the forecast uncertainty is well represented
(e.g. GLUE method: Beven and Freer, 2001), while at the same time
uncertainty is minimised as much as possible in a Bayesian approach
(Krzysztofowicz, 2002; Todini, 1999). A scheme, generally adopted in the
international Hydrological Ensemble Prediction Experiment (HEPEX), of a
general hydro-meteorological EPS (Schaake et al., 2007) that encompasses
these two characteristics is presented in a simplified manner in Figure 2.1.
The meteorological pre-processor stands for all analyses that make the
ensemble weather forecasts more suitable for use as input to a hydrological
model. More suitable means, for example, reliable probability forecasts and
sufficient spatial resolution. Despite the efforts of meteorological institutes
in preparing the ensemble forecasts, for particular variables at particular
44
Anticipatory Water Management
locations systematic errors often occur. On the basis of archived local
monitoring data and archived forecasts or hindcasts, bias correction can be
performed (e.g. Clark et al., 2004). For processing the weather forecast
output to the required resolution, downscaling techniques need to be applied.
Downscaling techniques can be divided in statistical downscaling, dynamic
downscaling, and analogue methods (Hamill, 2009). Statistical downscaling
uses similar bias correction techniques relating high resolution monitored
data of the target variable with the forecasted data. Dynamic downscaling is
done by nesting finer resolution atmospheric models in the atmospheric
model used for the ensemble prediction system. Analogues refers to preprocessing weather forecasts on the basis of comparison with historically
similar forecast and actual state characteristics. The hydrologic postprocessor performs tasks similar to the meteorological pre-processor, but
now on the basis of historic and real-time data of target water system
variables (e.g. Olsson and Lindstrom, 2008).
In this dissertation focus is on the end-use of ensemble prediction. This step
requires the interpretation of the ensemble prediction for decision-making. In
Section 2.5 it is suggested to expand the hydro-meteorological EPS scheme
of Figure 2.1 with, what could be defined as a "Decision support preprocessor".
Next to the sources of meteorological and hydrological uncertainty, in
applications of Anticipatory control also the uncertainty in control actions
and system response needs to be taken into account. The control uncertainty
is discussed in Sections 2.4.4.
Meteorological ensemble prediction
Meteorological pre-processor
Deterministic hydrological model
Hydrological post-processor
Hydro-meteorological ensemble prediction
Figure 2.1 Elements of a hydro-meteorological ensemble prediction system
2.4.4 Challenges in modelling controlled water
systems
A challenge of modelling controlled water systems is that the modelling
problem often consists of a high degree of freedom. For example, a model of
Anticipatory Water Management
45
a controlled water system with water level as the target variable can be
considered, e.g. a drainage canal. The water level in the water conduits is to
be controlled within a pre-defined range by a down-stream pumping station.
Then, the water level can easily be modelled by including a pumping station
with switch on and -off levels according to the target control range. Even
without the correct rainfall input and discharge capacity of the pumping
station, the model would produce a fair reproduction of the actual water
level, because the modelled pump would simply adjust its pumping
frequency and duration to keep the water level within the control range. The
target variable, water level, is then modelled accurately, but the operation of
the pumping station, and the volume that goes through the system might be
totally different from reality. This illustrates the extra degree of freedom that
is created by the inclusion of control structures in models. The risk of
producing the right output for the wrong reasons when modelling controlled
water systems, therefore, does not only refer to the risk of over-calibration
but also to the risk of providing nonsense system characteristics and input to
the model without noticing directly in the model results. Likewise, the actual
control structures in the actual water system also give a high degree of
freedom to the operators on how to manage the water levels. This can make
the water levels highly unpredictable, because operators may depart from
operational routine at any moment. This makes it more difficult to model
controlled systems accurately for long time series, compared to natural
systems.
An advantage of modelling controlled water systems is that in many cases
the data availability is better than in natural water systems. Regulating
structures usually perform discharge measurement at the same time, and in
addition up- and downstream measurement stations are often installed. The
cross-section is usually known for pipe-networks, canals, and wellmaintained ditches, with higher accuracy than the cross-sections for natural
streams. Although for the underground pipe-networks this is only true when
proper reporting during construction has been performed.
These particular issues in the modelling of controlled water systems are used
in the formulation of a modelling framework in Section 3.4.2.
Another challenge, particular for ensemble hydro-meteorological forecasting
is the time constraints involved in applying all ensemble weather forecast
input to the hydrological model. This is sometimes solved through
aggregating the ensembles by taking percentiles. However, this leads to time
series that have never been predicted, and as such may not make any sense
deterministically. Selection of a number of ensemble members always runs
the risk of missing extremes. Application of fast models or high performance
computing would allow inclusion of all the ensemble information.
46
Anticipatory Water Management
2.5 Decision making with uncertainty
2.5.1 Uncertainty
The different sources of uncertainty in weather forecasts and water system
state prediction have been mentioned in Sections 2.3 and 2.4. The problem
for decision making of the uncertainty of forecasts is two-fold. First in an
absolute sense, both the weather forecast and the system response modelling
cannot be fully accurate, especially when the uncertainty in the weather
forecasts is high. Secondly, the uncertainty is difficult to assess beforehand.
A decision maker would benefit from knowing the probability that the
forecasted event will actually occur. Only then can he make a cost-benefit
based risk analysis of taking or not taking anticipatory measures.
In water management decisions, three forms of inaccuracy are important. If
we want to optimise water management with the advent of an extreme event,
the event has to be forecasted correctly in terms of:
- location
- timing
- magnitude
If the event occurs in a different location, outside the water management
system of our interest, anticipatory measures are in vain. If the event arrives
sooner than expected, anticipatory measures will not have enough time to
become effective and safety is at stake. If the event arrives later the negative
consequences depend on the degree of optimisation of the system. It could
be argued that the more a system is optimised, the greater the damage when
the system fails. An emergency retention basin, for instance, needs a perfect
forecast of the timing of the flood wave. If the flood wave arrives later than
expected, the filling of the retention basin starts too early and by the time the
peak of the wave arrives, the basin is already full and flooding will occur.
Examples of forecast magnitude are the precipitation depth, hydrological
load or water levels. Thus, the sensitivity to particular forms of inaccuracies
depends on the anticipatory management action.
Unfortunately the uncertainty of forecasts is neither constant in time nor
uniform in space. Weather above large flat land areas can be forecasted
better than weather in mountainous areas. The trajectory of large frontal
systems is sometimes easier to predict than the inception of convective
storms. Wet season forecasts might generally be more accurate than dry
season forecasts. The same applies to water system models. Some models
may be tuned to perform well during average hydrological conditions, others
on critical events. A model of a system with one straight concrete lined canal
and a weir maybe more accurate than a model of a complex network of
canals with peat embankments. Therefore, generalised performance indices
on a regional or continental scale, as provided by meteorological institutes,
Anticipatory Water Management
47
give the water authorities only little information. Water authorities should
perform verification analysis for their own water system, and customised for
the type of Anticipatory Water Management to be applied. In this way the
water authority can train (calibrate) its interpretation of the ensemble hydrometeorological predictions for decision support in AWM (Decision support
pre-processor, Figure 2.2).
For assessing the behaviour of the Ensemble Prediction Systems to improve
decision-making, verification data is needed. In many cases verification data
will be limited, since the safety level is already high and thus extreme events
of interest seldom occur. On top of this the meteorological forecasting
models are being continuously developed, which reduces the number of
extreme events to verify the forecasts.
Meteorological ensemble prediction
Meteorological pre-processor
Deterministic hydrological model
Hydrological post-processor
Hydro-meteorological ensemble prediction
Decision support pre-processor
Figure 2.2 Elements of a hydro-meteorological ensemble prediction system
expanded with a Decision support pre-processor for end-use of the predictions in
Anticipatory Water Management
2.5.2 Risk
It has been discussed that hydro-meteorological forecasts are uncertain. This
would not be a problem if there would not be any risk of taking anticipatory
actions based on these forecasts. In water management problems, however,
taking unnecessary actions or not taking action when it is necessary usually
has negative implications. Meteorologists speak of false alarms and misses.
In the case of a false alarm economic damage will often occur. Evacuations,
releasing water from hydropower reservoirs and operating pumping stations
cost money. But also safety could be endangered. Imagine that a false alarm
of an intensive precipitation event is followed by anticipatory lowering of
reservoir levels below threshold values safeguarding the stability of
48
Anticipatory Water Management
embankments. The forecasted precipitation is expected to set up the water
levels soon after. If the precipitation event does not come, the water levels
will stay at this dangerous low level.
Misses of extreme events could be even more dangerous. If the flood
defence strategy of a water system relies on the forecast of intensive
precipitation events, missing such an event poses serious threats to the
community. If policy guidelines on safety have to be met by Anticipatory
Water Management, the frequency of forecasts missing extreme events
should be very low.
These risks in terms of safety and damage (economic, social, nature, etc.)
have to be assessed and taken into account when deciding on which levels of
uncertainty are acceptable.
2.5.3 Threshold-based decision rules for Ensemble
Prediction Systems
Threshold-based decision rules prescribe actions when a forecast exceeds a
predefined value (event threshold). For Anticipatory Water Management,
threshold based decision rules are very appropriate, because the first step in
the decision chain is to decide whether to switch from routine operational
management to anticipatory management.
The threshold-based decision rules for anticipatory actions may consider the
use of precipitation forecasts directly. For example, a decision rule based on
precipitation threshold can be defined by:
If forecasted precipitation > X mm day-1, then start control action A.
This decision rule is suited for a deterministic, single precipitation forecast
for a fixed 1-day forecast horizon. In the case of ensemble prediction
systems, there are a number of possible forecasts to consider for a range of
forecast horizons. The more of the ensemble members that exceed the
precipitation threshold, the higher the forecasted probability that the
precipitation threshold will be exceeded. A decision rule based on EPS has
to define how high this forecasted probability (P) should be, before an
anticipatory action is taken (see the probability threshold in Figure 2.3):
If forecasted probability P(precipitation Y days from now > X mm d-1) > N,
then start anticipatory action A.
Water Boards often have no information on how to set these decision rules.
The effects of the choice of event threshold, forecast horizon and probability
threshold on the performance of the management strategies are not known.
Anticipatory Water Management
49
There is no knowledge about the performance of the particular weather
forecasting system (e.g. ECMWF EPS), for the given water system, let alone
information about the performance of the decision rules that depend on these
forecasts.
25
Probability threshold
Precipitation (mm)
20
Precipitation threshold
15
10
Th 24
We 23
Tu 22
Mo 21
Su 20
Sa 19
Fr 18
Th 17
We 16
0
Tu 15
©2007, HydroNET
5
Forecast horizon
Figure 2.3 ECMWF EPS precipitation time series for location De Bilt (NL) (data
source: KNMI). When applying threshold-based decision rules for EPS, the event
threshold (Precipitation threshold), the forecast horizon and the probability threshold
have to be set. The probability threshold is the required forecasted probability that
the precipitation threshold will be exceeded. This is determined by the ensemble
members exceeding the precipitation threshold.
2.5.4 Cost-benefit analysis
The choice of a particular decision rule will not only depend on the accuracy
of the decisions. A cost-loss analysis is needed to optimize further the
decision rules and control strategies. The cost of an anticipatory controlaction and the loss if no action is taken when a critical event occurs are
different for every application, every particular water system, and for every
event. Furthermore, the anticipatory action does not have to be fixed, but
may be optimized for every particular forecast.
The cost-benefit analysis needs to take into account different types of
damage. The direct damage is the damage that occurs immediately after a
flood event due to the physical contact of the water with humans and
damageable properties (Smith and Ward, 1998). Indirect costs accumulate
50
Anticipatory Water Management
over time after the event, for example because the damaged infrastructure is
reducing economic output in the affected area. Next to the distinction
between direct an indirect costs, the differences between tangible and
intangible damage have to be taken into account. Tangible damage is
referred to in this dissertation as damage that can be expressed (and
estimated) in monetary units, while intangible damage can not. The most
important intangible damage is the loss of human life.
2.5.5 Decision Support
Water Management
Systems
for
Anticipatory
Incorporating probabilistic forecasts of several days ahead into a decision
support system (DSS) brings along high requirements of the system. The
DSS-system needs to be able to handle many data streams. For example, if
we get a 10-day forecast every 12 hours, we could save each forecast and
compare 14 forecasts for three days ahead. We also want to have a
probability distribution of each forecast, which could be represented by 50
possible forecasts of the same time frame (ECMWF EPS). Add to this the
measured data in the field for update or assimilation purposes with the
system response modelling output, and the complexity of the data
management becomes clear. The decision support system has to channel this
data through all its components, such as database, system response models,
optimisation models and presentation modules.
The optimisation problem also becomes very complex. Because anticipatory
management involves risks, it is preferred to stick to normal management.
Therefore, first the system has to be optimised using the constraints of
normal management. If normal management is expected to fail after, for
instance, four days (with enough probability), anticipatory actions will be
applied and then has to be optimised as well. This again is constrained
optimisation, e.g. to the system safety limits and capacity. At the next
decision moment both optimisations have to be applied again. Finally,
uncertainty must be handled as well, either in the objective functions or in
the presented results. The water manager can add to these results his expert
knowledge to complete the risk analyses and make the decision.
Complexity does not have to be a problem if there is ample time to solve the
optimisation problem. But water systems are dynamic and optimisation of
operational management therefore requires frequent decision moments.
Furthermore, anticipatory actions take time before becoming effective and
the lead-time of forecasts is limited. Thus the computational speed of
decision support systems could become a limiting factor.
The last issue to be raised on decision support systems is their reliability.
This research seeks to apply anticipatory management to reduce flooding
Anticipatory Water Management
51
problems. Thus the reliability of the entire system has to be high.
Anticipatory management will depend strongly on communication lines for
collecting the data and forwarding control decisions to operators. These
communication lines have to be secure. Another well-known problem is the
reliability of the computational models. Run-time errors are annoying for
analysis purposes, but very harmful for real-time control applications.
Note that in this section on DSS it is implicitly assumed that the decision
problem is mostly too complex to be solved by managers without the aid of
state-of-the-art support techniques. This assumption has to be treated
carefully. Many water managers have years of experience in operating the
control structure(s) of their water system. They also have considerable
knowledge on the system response to weather events. Thus anticipatory
management could be done by, for example, just providing weather
forecasts. The system response model will then be the experience based
models of the water managers and decisions will be based on their risk
analysis and priorities.
In practise many of the developed and operational decision support systems
are only partially used in the actual management of water systems. The role
of decision support systems is not formalised in management regulations.
Sometimes systems are not reliable enough or not practical to use (too slow
or too complicated). Most importantly they are not (and cannot be) 100%
accurate, while their deterministic outcomes suggest the opposite. Many
water managers, therefore, put little trust in weather forecasts and decision
support systems.
2.6 Knowledge gaps and hypotheses
To summarise, the main challenges of expanding the decision horizon by the
use of weather forecasts in Anticipatory Water Management are the high
uncertainty of the future system state and the risks associated with inappropriate anticipatory management actions.
From the literature review it appears that there are opportunities to enhance
AWM through the availability of ensemble weather prediction systems.
However there is a first knowledge gap on how the ensemble weather
forecasts perform over a long period of time for a particular catchment.
There are no universal guidelines on how to assess this. The hypotheses are
that:
The comparison of measured precipitation local to a
given water system, with ensemble precipitation
forecasts leads to an improvement in the use of those
forecasts. (Hypothesis 1)
52
Anticipatory Water Management
and that:
The comparison of measured water levels in a system,
with those predicted in response to ensemble
precipitation forecasts under the current management
strategy, leads to an improvement in the use of that
system. (Hypothesis 2)
The second knowledge gap is on how to define the best decision rules with
these probabilistic forecasts for whether to anticipate or not. It is
hypothesised that:
Effective decision rules can be found by hindcast
analysis. (Hypothesis 3)
Once the warning that anticipation is necessary has been given, the third
major knowledge gap is on what exactly the Anticipatory Water
Management action should be. It is hypothesised that:
Long-term simulation of the complete Antcipatory
Water Management strategy for historic time series
enables an optimisation of AWM. (Hypothesis 4)
Then, with the AWM strategy developed it should be decided by the water
authorities whether the strategy should be applied. It is hypothesised that:
A cost-benefit analysis, based on the continuous
simulation of water levels, generated in a water system,
with a prescribed management strategy, by forecasted
precipitation from a specific product for a historic time
series with critical events, is needed to assess whether
the forecasting product should be applied to the
particular system. (Hypothesis 5)
This implies that the costs and benefits of inappropriate and appropriate
anticipatory management actions have to be assessed dynamically, and an
adequate simulation model of the controlled water system has to be
available. These two requirements are often not readily available for a
particular water system, or with a particular water management body, e.g. a
Water Board.
A framework for developing Anticipatory Water Management for a
particular water system is needed to fill these interrelated knowledge gaps
and requirements. This framework is prepared in the next chapter. It is
Anticipatory Water Management
hypothesised that when implementing this framework:
The use of ensemble precipitation forecasts to decide on
anticipatory control actions, in preference to re-active
control, can reduce the damage over a long period of
time. (Hypothesis 6)
and that:
The expected benefits when applying AWM, despite
their uncertainty (due to limited availability of data,
changes in the cost-benefit relationships, model
uncertainty, etc.), more than compensate for the losses
that occur when AWM is not applied. (Hypothesis 7)
53
55
3
Framework for developing Anticipatory
Water Management (AWM)
3.1 Introduction
This chapter discusses the different steps, and the associated challenges, for
developing and evaluating an Anticipatory Water Management strategy for a
particular water system. The resulting theoretical framework is put into
practice in two cases studies described in Chapters 4 and 5.
3.2 Establishing the need and potential for AWM
3.2.1 For which events is AWM needed
Before exploring the possibilities of AWM it should be clear why and for
which events AWM is needed. Although this seems an obvious first step,
there are a number of reasons why this is often not taken and why it is not
straightforward to complete this task. Let us first consider why event
selection needs to be done and which criteria the event selection should
adopt. As in any scientific study, engineering project and management
review, there should be an incentive to change the present situation. For all
AWM applications the incentive to change is universal in the sense that there
is a desire to act on upcoming events, before they actually happen. This
means that from the past history or future scenario outlooks, it should be
known which events led (or will lead) to undesirable situations (damage).
For flood control this refers to past flood events. While recognising that not
all extreme events can be handled by improved control, AWM can
potentially reduce the frequency and magnitude of the damage due to these
events. In order to develope decision rules and to evaluate the effectiveness
of AWM as many events as possible should be identified. The higher the
number of selected events the better the confidence in the analysis results.
For added confidence in the findings, it is therefore also preferable to have
historic events, instead of hypothetical events.
In scientific studies, such as hydrological modelling, the event selection is
often not done. The focus is on reproducing (characteristics of) hydrographs.
Often continuous simulation is applied and the performance is measured in
terms of standard values, such as the squared error, in order to be able to
compare with other models and earlier publications. In the case that the
intended end-application is flood forecasting, emphasis is placed on the
peaks of these hydrographs; but still it is often not discussed exactly which
56
Anticipatory Water Management
peaks at what moment led to flooding historically. Then in the modelling
stage the model is calibrated and validated, and subsequently presented to
the end-user. The end-user usually does have some idea of which events are
critical for his water system, and based on forecasts of these he could make
warning decisions, but these end-user decision rules are rarely evaluated for
a series of past events.
Methods for event selection
The interesting thing is that if past events are selected, this appears not to be
as easy as expected. In many countries secondary data on flood events is
limited, or there is limited access. The latter arises either because of political
sensitivity or because of lack of (digital) infrastructure to search for this data.
The data can be limited in the sense that date of flooding may be known
from old newspapers, but the exact location of affected areas, the cause and
the duration of the flooding are not. This is particular truth for poor
countries, as is discussed in the Ethiopia case study (Chapter 5).
Secondly, in controlled water systems it is often difficult to determine
critical events, because of the human based regulation of the structure. For
instance when looking at water level records of a small reservoir, high peaks
do not necessarily mean that an extreme natural event occurred. The high
water levels may also have been caused by "wrong" operation of a discharge
structure (Figure 3.1a). Vice versa, an extreme load to the reservoir may not
be observed by looking for peaks in the water level data, because the
reservoir level was unusually low at the beginning of the event (Figure 3.1b).
This can be because release was maintained too long after a previous event,
or because a cautious operator decided to apply an anticipatory release
himself. In the following, solutions for both lack of data and deceptive data
are given.
(a)
(b)
06-Dec-98
16-Dec-98
26-Dec-98
27-Oct-98
-0.45
-0.50
-0.55
-0.60
-0.65
-0.70
06-Nov-98
16-Nov-98
-0.40
Modelled
Measured
Water level (m+Ref)
Water level (m+Ref)
-0.40
-0.45
-0.50
Modelled
Measured
-0.55
-0.60
-0.65
-0.70
Figure 3.1 Unnecessary high measured water levels (a) and high measured water
levels prevented by early lowering of storage level (b)
Framework for developing Anticipatory Water Management
57
Finding system thresholds in data scarce catchments
Anticipatory Water Management is necessary when the capacity of the water
system is not sufficient to cope with the hydrological load. Every water
system has limits in the amount of water that it can discharge and store,
meaning that up to the point where this limit is reached the system functions
well, and when the limits are exceeded and too much water is entering the
system, flooding will occur. On the other hand, every water system has
limits to the amount of water it can let in, recharge or store, meaning that if
for too long a period the net inflow is negative, water levels will drop and
drought may occur. To know whether Anticipatory Water Management
should be applied, therefore, involves finding these system thresholds.
Interviews of inhabitants of flood prone areas, government and NGO staff
members, and water professionals help to determine the timing and affected
area of the events.
In data scarce countries and catchments the first step is to look into
secondary data like (electronic) newspapers and humanitarian organisation
reports. This results in at least the years when floods and droughts occurred.
Most of the time, it is also possible to find from these sources the start of the
event with several days accuracy. Some idea of the severity of the event and
the cause may also be given. In these information sources, affected areas are
usually only roughly mentioned, like a regional district level, or affected
cities and villages.
A better source of information on the spatial impact of an event can be found
from satellite images, like radar-sat, which are increasingly available. Next
to a precise delineation of the affected areas, these images also provide a
confirmation of the timing of the event, because the exact data and time the
image was taken is always given. Note that the timing cannot be taken from
satellite imaginary alone, because in the case of flooding the affected area
may be inundated for several days up to a month. It cannot be seen from the
image whether it was taken at the beginning or at the end of the event.
All these secondary data can be compared with the primary data, the time
series data. If start dates are known, system thresholds may be found directly
if the dates are consistent with the highest (or lowest) recorded peaks. Errors
in the data can also be an indication of extreme events. For example
flattened peak discharges may result from water levels rising above the
measurement scale. Missing data can indicate the moment that a measuring
device was flushed away. Note that all this information has to be used with
care, because its interpretation can vary; for instance, the stopping of a
recording can also just mean that it was damaged by a floating log, or by
vandalism. Also, from the time-series data it cannot always be inferred what
was the cause of the event. It may be that the capacity of the system was still
sufficient, but that system failure led to the damage. In the case of flooding a
58
Anticipatory Water Management
(intentional) breach of the embankments may be the cause; in the case of
drought, illegal extractions may be the cause.
Also, the variable in which the threshold is expressed can not often be
decided upon beforehand. While in reservoir systems water level is often
chosen, precipitation depth is some cases may be used as well (avoiding the
dependency of rainfall-runoff, hydrodynamic and reservoir modelling), or
inflow volumes etc. A joint use of different variables can be used to cover a
wider range of forecast horizons and to cross-verify forecasts.
It can be that the available information is not sufficient to determine the
system thresholds. In that case it must be asserted what information is trusted
and to what detail the occurrence of extreme events can be determined. For
instance, it may be possible to assess only the years in which flooding
results. Then a range of thresholds can be estimated, for instance by
comparing the hydrographs of the years with flooding to the years without
flooding. An example three-level threshold system consists of an alert
threshold (AWM may be necessary), alarm level (AWM must be applied),
and disaster level (AWM will not help, calamity plans must be executed).
Finding system thresholds in controlled water systems
In controlled water systems more data are often available. In the channelled
reservoir systems used in the Netherlands for draining the low-lying
reclaimed lands called "polders", precipitation data, water level recordings,
and regulating structures operation and discharges are usually available for
several years. The problem comes from the not fully consistent, humanbased, operational management of the storage basin water level. Here a
thorough system analysis, taking into account all relevant variables must be
applied.
The primary variable from the systems approach for expressing thresholds is
Volume. This volume, however, is usually not directly measured or
measurable. In the Netherlands for instance the discharge of the smaller
pumping stations from the low-lying areas to the higher channelled storage
basin is not always available. Also, especially for the longer time scales as
with drought, seepage, percolation and evaporation become relevant, and yet
are difficult to measure. Therefore, it is often preferred to use the resultant
variable, water level, which is easy to measure. Water level is often the
primary variable from the point of view of impact on society, because it is
the high or low (ground)water levels that cause flooding and drought related
damage. From the system design it is already known what the water level
thresholds are. The historic water level recordings that exceed these
thresholds should be identified as a first step to identify water system
thresholds. In step two the selected events should be filtered by looking at
the precipitation data, the regulating discharge data and other available
volume data to exclude the extreme water level events that were caused by
Framework for developing Anticipatory Water Management
59
the control strategy or regulating structure failure, instead of exceedance of
the systems capacity. In the third and final stage, events should be added that
did not exceed the water level thresholds because a temporarily change of
control strategy was applied. For instance, lowering reservoir levels before
an extreme event occurs may prevent the water levels from exceeding the
threshold, although the inflow volume was more than the system's capacity.
Note that in this case, Anticipatory Water Management is already
successfully applied, even before it is formalised in the operational policy
and control strategies. These cases are not uncommon, because operators
have many years of experience and have a mental model that includes
system thresholds and they are used to dealing with additional information
like weather forecasts from the television. These events can be identified by
checking water level records for unusually low water levels just before an
unusually long rise of the water level.
After the critical events have been selected, based mainly on the resultant
variable, water level, the analysis can be expanded to find the volume based
system capacity as well. If, as was described in the beginning of this section,
discharge volume data is not available (small pumping stations in polders),
then the volume analysis can be taken further upstream in the hydrological
cycle, e.g. up to the precipitation input volume. Analysis of the relationship
between measured precipitation and extreme events can very well serve as a
cross-validation of event selection in controlled water systems or as
additional guidance for anticipation (Figure 3.2).
180.00
160.00
Precipitation (mm/ x days)
140.00
120.00
100.00
80.00
60.00
40.00
Min
20.00
Max
0.00
0
2
4
6
8
10
12
14
16
x days accumulated
Figure 3.2 Upper and lower precipitation thresholds for accumulated precipitation in
Rijnland. After seven days (veritical line) the minimum threshold does not increase
anymore.
60
Anticipatory Water Management
3.2.2 Potential for anticipatory management action
Reservoir level
After the critical events have been clearly defined in terms of thresholds, the
Anticipatory Water Management actions that may reduce the (frequency of)
exceedance of these thresholds have to be defined. Examples of such actions
are maintaining reservoir levels in anticipation of a dry-spell, and lowering
reservoir levels in anticipation of a peak discharge (Figure 3.3).
Conceptualising the pro-active actions can best be done by or together with
the operational managers and policy developers of the water authority
responsible for the particular water system.
Without anticipation
With anticipation
Time
Figure 3.3. Example of anticipatory action. Reservoir level is lowered in
anticipation of a flood event. As a result of the anticipatory lowering, the resulting
peak reservoir level is reduced.
Conceptualising range of possible AWM actions
In natural uncontrolled rivers where flooding is a problem, water
management actions can be; harvesting crops early, moving assets to higher
ground or floors, evacuating people, strengthening levees, or intentional
breaching of levees. Control structures in rivers that can be used to reduce
damage by flooding can be reservoirs and floodgates or weirs. The reservoirs
can accommodate (part) of the river discharge peak and lateral flood weirs
can divert part of the peak discharge to emergency storage basins. In some
cases the latter has the same effect as intentional flooding, only now it is
called controlled flooding. In reservoirs, early releases can done by opening
sluice gates, lowering weirs, or activating pumping stations. The early
releases lower the water level in the storage basin, thus increasing the
storage capacity for the upcoming peak inflow, or hydrological load.
Anticipatory actions for events invoking a shortage of water, range from
crop cultivation strategies, via water supply rationing, to keeping reservoirs
Framework for developing Anticipatory Water Management
61
full at the maximum level. From these examples it can be seen directly that
there is a wide variation in the time these actions take to become effective.
Therefore, after having an anticipatory control action in mind, the entire
process to apply this action has to be conceptualised in order to determine
the required forecast horizon.
Conceptualising the AWM process and estimating required forecast
horizon
Whereas there is a great variety of AWM actions and applications,
general process is always the same (Figure 3.4): forecast, communicate
forecast to a decision maker, choose anticipatory action, communicate
decision to operators, implement anticipatory action, allow time for
action to become effective (action response time).
the
the
the
the
Every management action takes time to become effective. Evacuation takes
hours to days or maybe even weeks to be completed. Sluices have to be
opened before water can be released from a reservoir. The response time of
the water levels in the reservoir and the upstream water system depends on
the discharge capacity.
Next to this action response time, the decision-making process also takes
time. The information that is used by the decision makers and decision
support systems is not in real time. Measurements have to be communicated
from the measurement station to the central control unit (often to a database),
where it might have to be processed before it can be used. Decision makers
have to interpret the information that is given to them. Hydrological models,
control models and optimisation models need time to run. In many cases
decisions have to be deliberated amongst several actors. Also the
communication of the decision to operators takes time.
Weather forecasts and analyses are made at the meteorological institutes and
it takes a considerable time before the forecasts are available for water
managers. In the Netherlands the delivery delay of the 10-day ensemble
forecasts of the ECMWF is about 15 hours.
The management action response time, decision time and data delivery
delays need to be taken into account when determining the required forecast
lead-time. Note that although the timing of these processes can be estimated
from a knowledge of the water system and experience with the operation of
the regulating structures, in many cases at this stage a computational model
of the water system is needed, because control actions are being considered
that have not been (regularly) applied in the past. We will return to water
system control modelling in Section 3.4. and Chapter 4.
62
Anticipatory Water Management
Monitoring
Weather forecast
Water system
prediction
Action response
Evaluation
Decision support
Decision models
Implement
anticipatory
actions
Instructions to
operators
Make decision
about
anticipation
Figure 3.4 General process of anticipatory water management
Setting boundary conditions
Boundary conditions are determined by the system design and its user
requirements. Reservoirs levels cannot drop below the dead storage level and
because of user requirements, such as power generation, the water level is
not allowed to drop below the upper active storage level. In channelled
storage basins, upper lower limits are governed by water depth requirements
for navigation, bank stability problems and groundwater level requirements
for crops.
Determining potential effectiveness of AWM actions
With the boundary conditions for the anticipatory management actions
known, the potential effectiveness of the actions can be determined. This can
be simple volume calculations with the storage basin level-area function and
the planned lowering or rising of the level before a flood (or drought event),
or model runs where historic extreme events are simulated with the lower (or
higher) antecedent storage levels. Because this is to calculate the potential
effect of the AWM, these model runs assume perfect forecasts, e.g. only
measured inputs are used. In the case of flood and drought forecasting and
early warning the effectiveness is often more difficult to quantify. Here the
effectiveness depends on agricultural processes, evacuation processes and
emergency mitigation measures like putting up temporary levees.
Identifying risks of adverse effects of AWM actions
It is inherent to AWM that proposed control actions deviate from the normal
operational guidelines. AWM deals with temporally expanding the normal
control range, which increases the risk of adverse effects. These risks have
already been taken into account in conceptualising the anticipatory control
Framework for developing Anticipatory Water Management
63
action and setting its boundary conditions, however, this is done for the case
when the AWM action is applied correctly. Additional risk analysis should
be performed for the cases that the AWM is not performed correctly. If
AWM is considered as a measure to replace structural measures such as an
increase in storage capacity, than the first risk is that an extreme event is not
forecasted and the AWM actions are not implemented or are started too late.
This leads to increased risk of flooding (or drought). The second risk is that
of "false alarms" or in a continuous sense over-predicting the upcoming
event; this always needs to be considered. Forecasts are never 100% certain
and because of the increased horizons applied for AWM, higher levels of
uncertainty are to be expected. These result in unnecessary (or too
aggressive) AWM actions. In the case of flood control in a reservoir this
means that the reservoir level is lowered because high inflows have been
forecasted, while subsequently the expected inflow does not arrive. This
could lead to prolonged low storage levels, with risks of water shortage,
falling groundwater levels, failure to meet power generation requirements
etc. The risk of false alarms is part of the "potential'' of AWM and needs to
be taken into account in the development of AWM.
3.3 Verification analysis
3.3.1 Product selection: time scales, spatial scales
When the aim of AWM has been identified, including the range of events for
which it should apply and the anticipatory actions that are to be taken, a
forecasting system can be selected. The forecasting horizon should be
enough for the AWM to be effective. The time step and spatial resolution of
the meteorological forecasts should match the water system characteristics.
After the forecasting system has been selected, the uncertainty and
associated analyses will have to assess up to which lead-time useful
information for management can still be deduced from weather forecasts
(Figure 3.5).
3.3.2 Continuous simulation of the real-time AWM
forecasting system
The uncertainty of the forecasts is of such importance for the effectiveness
of AWM and for the development of decision rules that its assessment must
be part of a fast screening analysis. It can be fast because the AWM strategy
does not yet have to be worked out. The only focus now is to make forecasts
with the current, normal, control strategy and to verify these forecasts with
historic data. The verification analysis is a screening of the forecasting
product that is being considered for use in AWM.
64
Anticipatory Water Management
P
F1
F2
F3
0
Tmax ?
t
Figure 3.5 Fictitious example of a three-member ensemble precipitation forecast. At
a certain lead-time the uncertainty might be considered too high for decision making
(Tmax).
This can be a meteorological forecasting product, hydrological,
hydrodynamic, or any considered change or addition to the tools used for
operation management decisions.
For this screening to be effective the real-time forecasting process that will
be used in the AWM needs to be emulated. This approach is called
hindcasting. A modelling system is prepared that allows continuous hindcast
simulation of long periods (multiple years). The layout of such a modelling
system is given in Figure 3.6. The available length of the time series depends
on the water system data in the water authority's archive and the forecast or
re-forecast archive of the new meteorological forecast product, provided by
the meteorological organisation.
First, the water system model used, must be validated for the continuous run
using measured input data, instead of archived forecasts. For controlled
water systems this is to test whether the normal control strategy can be
modelled well. This is a pre-requisite for the hindcasting to make sense. The
hindcasts are to predict which event could not be handled with the normal
control strategy, and can be handled better with AWM. Once the water
system control model performs satisfactorally it can serve to generate the
reference time series, instead of the measured time series. This is because the
model shows what would have happened if the control strategy had been
executed consistently, while the measured data has all the unpredictable,
human control, decisions incorporated in it. This water system control model
also helps to identify the critical events (Section 3.2.1) that were not visible
in the recorded data because accidentally AWM was applied instead of
normal control.
Finally, a good water system control model is important to build confidence
with the water authorities that the AWM strategies that will be developed
Framework for developing Anticipatory Water Management
65
and evaluated using this modelling system are realistic. Water system control
modelling is discussed separately in Section 3.4.
t
t+1
t+2
Meteorological forecast input
Meteorological measurements input
Forecast time
Forecast time
Forecast time
Water system state forecast
Spinup time
Water system control model
Spinup time
Normal control strategy input
Spinup time
Water system measurements input
Figure 3.6 Creating hindcasts. The forecasting process is repeated for every time
step t in the past.
3.3.3 Event based verification of a range of decision
rules for AWM
The next step of the screening is to verify the produced hindcasts with the
selected critical events. If the forecasts identify the critical events, it means
that the forecasting system can be used to decide when anticipatory actions
are needed. Note that this is only the first step in an actual AWM strategy,
because after the decision is made to undertake pro-active action, it has to be
decided what exactly the action will be. Since this is a dichotomous
(anticipate 'yes' or 'no') decision problem the verification results can be
classified according to a contingency table, where every forecast is either a
hit (event occurs and is forecasted), a false alarm (event is forecasted but
does not occur) or a correct rejection (event is not forecasted and does not
occur) or a miss (event occurs but was not forecasted) (Table 3.1). Note that
this is different from the usual verification of hydrological and
66
Anticipatory Water Management
hydrodynamic models. The verification has to be done for a range for
decision rules that are considered for the AWM strategy.
Table 3.1 Contingency table for ensemble precipitation forecasts
From this verification analysis important information can be derived. In the
first place this information concerns the absolute effectiveness of the
forecasts to identify the critical events. How many of the past critical events
could have been forecasted? How many false alarms occurred? Up to which
forecast horizon can still more than half of the critical events be forecasted?
Which forecast threshold results in all events being forecasted? For which
probability thresholds none of the critical events is forecasted, etc. In short,
this screening in absolute terms should tell whether the forecasting system is
effective and identify the range of useful decision rules.
The performance relative to other forecasting methods should be tested at
this stage as well. For this relative testing, scoring methods from the
meteorological sciences are very convenient. The forecasting system to
compare with can be a forecast product from another producer, the presently
used forecasting system, or should be at least a climate based forecast, a
forecast as previous, or a random forecast. The climate forecast is mostly
used as a baseline forecasting system. Here the forecasted probability of an
event is the climate frequency of the event. This frequency is often assumed
to be equal to the sample frequency from the analysis record. The new
forecasting system should of course be at least better than the climate
forecasts. Relative characteristics of a forecasting system are expressed in
"skill scores". For this screening the relative operating curve is often used
(Kok, 2000, p. 59) , because it visualises skills and decision rules of different
forecast systems in one graph.
These meteorological skill scores are calculated by evaluating each forecast.
In Anticipatory Water Management it is better to apply an event based
verification approach. The comparison between the measured and forecasted
events should be done for the time at which a critical event begins. The
forecast of the beginning of an event is important to allow for effective
anticipatory control actions. The forecasted time at which the event begins
should be within a predefined range (e.g. one day) of the actual beginning of
the event.
Framework for developing Anticipatory Water Management
67
Note that this is different from a "forecast by forecast" analysis. Events that
last more than one day are considered as only one event, so that correctly
forecasting one long event does not count as multiple hits. This avoids
masking missed events of short duration by correctly forecasting one of the
long events. In the same way the missing of one long event, is considered as
one missed event. This avoids the disbenefit of correctly identifying several
short events whereas only one long event has been missed. Such
considerations are particularly important in the analyses for anticipatory
water-system control because they deal with infrequent critical events. Also,
false alarms are analysed as separate events. The drawback of the event
based approach is that forecasted duration of events is not scored, which in
some cases may result in less stringent verification than the "forecast by
forecast" verification.
3.4 Modelling controlled water systems
It has been made clear that for a successful verification analysis a reliable
water system control model is needed. Therefore, in this section, modelling
of controlled water systems is discussed.
In Section 2.4 the main challenges in modelling controlled water systems
have been identified as the high degree of freedom and the unpredictable
human based control strategies. In many cases the choice is made to model
only the rainfall-runoff part (Roulin, 2007) or to take a set of control rules
and consider the water system control model as providing a potential / or
perfect result. The main reason for this is that human behaviour is difficult to
predict and hence difficult to model. Most controlled surface water systems
have considerable or exclusive human supervised control. Therefore the
control is often not fully consistent over time.
Here it is argued that still it should always be attempted to model the current
(business as usual) control strategies and to show the model results together
with the measured results. The need for modelling control strategies, as
compared to modelling only the rainfall runoff process, is clear when the aim
is to evaluate new control strategies. The reasons for presenting the
comparison with measured data are three-fold:
1. For the scientist it is necessary to find out whether his control
model is capable of realistically modelling regulating structures
and the response of the water system to control actions
2. The current control strategy will always be the base-line control
strategy against which new control strategies will be measured
for their effectiveness and efficiency.
68
Anticipatory Water Management
3. If considerable local (in time or place) differences are found
between the modelled states and measured states, this may help
to identify events and locations where or when operators
decided to deviate from the normal control rules, for example, to
anticipate extreme events
3.4.1 Input data based on end-use of model
However obvious, it is not common practice to use the same input variables
and data sources for calibration and validation as will be used in operational
tasks of the model, and indeed it is not trivial to realise in many cases.
A first limitation is often that multi-year time series data are available only
from the (old) ground stations, while the model in the end will be fed with
data from enhanced ground station networks or remote sensing data such as
radar and satellite data.
The second limitation is that in operational forecasting applications the
models are often forced with several sources of data for the same variable.
For example precipitation input for rainfall-runoff modelling for flood
forecasting could use ground station data up to t = 0, radar data up to t = +2
hrs, and quantitative precipitation forecast (QPF) data up to t = +15 days.
Still, the model is usually only calibrated with the ground station data, while
temporal and spatial scale differences may well influence the performance of
these models with other data sources. If possible, the same source of data
should be used for calibration, validation and application. If this is not
possible, then a comparative analysis of the data should be made (Van Andel
et al., 2009a), to assess whether different sources can be used for calibration,
validation and application directly, or whether scaling of the data, or
combined calibration and validation is necessary.
3.4.2 Framework for modelling controlled water
systems
The framework discussed in this section concerns the model construction,
calibration and validation phases of the modelling process (Abbott and
Refsgaard 1996, p. 24) and emphasizes possibilities of iteration within these
steps (Nash and Sutcliffe 1970). The modelling framework suggested for
modelling of controlled water systems focuses on the problem of increased
degree of freedom, because of the control structures. The solution depends
on having more, and more reliable, measured data available (as is often the
case for controlled systems). The measured data allows for two modelling
steps that are not generally feasible and necessary with hydrological or
hydrodynamic modelling of natural systems.
Framework for developing Anticipatory Water Management
69
The first additional step is that with the increased availability of data, the
model validation can be expanded from testing the model for the target
variable for a period or event that was not used in the calibration, to a
validation of the non-target variables for long term simulation periods. This
extended validation allows identification and visualisation of any processes
that might have been omitted or wrongly presented in the model. In other
words, it allows visualisation, discussion and modelling of those processes in
the water system that have been overlooked or are simply not known (Figure
3.7). In the proposed framework, it is suggested not to leave any of these
deviations in the extended validation unresolved and un-modelled.
The second additional step can be taken if the increased data availability for
a number of control structures clearly separates one sub-system from
another. It is then often possible to replace sub-system models by time series
data input, to enable model calibration of one sub-system at the time. This
reduces considerably the danger of correcting one wrongly modelled subsystem or control structure by adjusting wrong parameter values to
connected sub-systems or control structures as well.
Together with data acquisition, model set-up, calibration and validation, the
modelling approach (Figure 3.7) is to first estimate all (physically based)
parameters on the basis of the expert knowledge and available data, second,
to calibrate the model, third, to compare the modelling results to check for
trends that indicate that some processes have not yet been modelled (data
driven approaches can be used), fourth, to model these deviations with either
physically based (known and separable processes) or data driven (unknown
or un-separable processes) model components, and finally to calibrate again
the estimated parameters. When several sub-catchments have to be
modelled, this methodology has to be used starting with the target variable
(often at the downstream-end of the system), using available measured data
of sub-systems as input, and then step by step to replace the measured input
with models, because measured data or external predictions may not be
available in operational forecasting mode.
In Section 4.4 the framework is applied to improve a water system control
model of the Rijnland water system, in the Netherlands (Van Andel et al.,
2009a).
3.5 Strategies for anticipatory water management
The available strategies for AWM can be described in three groups. These
are discussed in the following sections.
70
Anticipatory Water Management
Information and data acquisition
Data validation
Expert based best knowledge
model
Calibration
Validation
Finish
Accurate?
Yes
No
Visualise unknown
(time scale analysis)
Explainable?
Yes
No
Update model and parameters
Model unknown data driven
This may require expansion of
modelling software if missed process is
not included
(use stepwise, sub-system
calibration if possible)
(use stepwise, sub-system calibration if
possible)
Figure 3.7 Framework for modelling controlled water systems. Visualisation,
discussion and modelling of the unknown processes are key. If the processes can be
identified (e.g. by error analysis for different time scales) and isolated after
visualisation and discussion, they can be represented by an internal or external
physically based model, if not, a data driven approach can be used.
3.5.1 Rule-based
Rule-based strategies consist of a combination of heuristic rules and predefined anticipatory actions, e.g. "if this than do that, else do something
else". Examples could be decision trees when deciding whether to switch
from normal management to anticipatory management (van Andel et al.,
2008a), in combination with classes of pre-defined sub-optimal management
strategies to minimise the damage of false alarms. In this respect, it is
important to update the decision frequently.
Framework for developing Anticipatory Water Management
71
Q
Anticipate
then
If p(Q≥qflood) ≥ w
else
p
Normal control
Q
F1
F2
0
t
F3
Figure 3.8 Fictitious example of a decision rule, based on three members of an
ensemble hydro-meteorological forecast.
The decision tree accounts for probability (p, w) following Krzysztofowicz
(2002) by choosing a threshold number of ensemble members (Fi) that
forecast an alarm generating hydrological load or level (Q, qflood) within a
certain lead-time (Figure 3.8). Instead of or in addition to this probability
threshold type of rules, statistical measures (first moment, second moment,
etc.) describing the estimated Pdf, can be used to relate the forecasts to predefined strategies.
3.5.2 Pre-processing of ensemble forecasts to
deterministic forecast
In this approach the ensemble rainfall forecast is taken, and interpreted to
determine the inflow volume that identifies the control action for the coming
control time step. The most intuitive method is to take the average forecasted
precipitation for every time step from the ensemble forecast. Deterministic
optimisation methods can then be used for the strategy.
3.5.3 Risk-based
Risk based strategies refer to the use of decision rules on the basis of the
estimated probability of occurrence times the estimated associated cost. The
most widely used risk based decision rule is to decide on the alternative that
has the minimum risk. The reasoning behind this decision rule is that if you
apply it consistently over time, the actual cumulative total cost, after a long
period, will be minimum as well. This will only be true if the probability of
occurrence, e.g. the PDF of a flood event, can be forecasted accurately
enough. For risk-based decision making with ensemble forecasts the
72
Anticipatory Water Management
assumption is that each ensemble member has equal probability of
occurrence. The risk based method is illustrated for a "go- no go" decision
problem for flood warning in Figure 3.9 (Maskey et al., 2008). The risk of
either decision is estimated by multiplying the probability of occurrence of
flood/no flooding with the expected damage cost, given the decision that
would be taken. The risk of flooding, added with the risk of no flooding,
given that a warning was not issued, provides the total risk of not issuing a
warning. The total risk of issuing a warning is determined in the same way.
If the decision problem is risk neutral, then the decision with the minimum
risk should be taken.
Note that instead of "cost", "disutility" is used. The term "disutility" is
preferred for operational management decisions where not only direct
tangible costs are at stake, but also other sources of damage. These are
discussed in Section 3.6.2. Uncertainties in estimated damage costs are
generally high, but there is generally not much data or information about this
uncertainty and methods for incorporating these uncertainties (as well as
which uncertainty (not) to take into account) are being developed, debated or
still need to be developed.
FORECAST
ALTERNATIVE
FUTURE
STATE
F=0
DISUTILITY
DU00
Expected
Risk (W = 0)
p(Z)
W=0
F=1
DU01
Z
Zflood
W=1
F=0
DU10
Expected
Risk (W = 1)
F=1
DU11
Figure 3.9 Risk based decision tree for flood warning, where the alternatives are W
= {0, 1} and the future states of the system are F = {0, 1}. W = 0 and W = 1 imply
“do not issue warning” and “issue warning”, respectively. Similarly, F = 0 and F = 1
imply “the area is flooded” and “the area is not flooded”, respectively. (Cited from
Maskey et al., 2008)
Framework for developing Anticipatory Water Management
73
3.6 Cost-benefit of selected AWM strategies
3.6.1 Dynamic cost-benefit analysis
The next stage of the screening is an evaluation of a number of suitable
AWM strategies. Although the forecast verification informs us about the
numbers of hits, misses and false alarms, it is not known how much is gained
from a hit and how much is lost from a false alarm. Therefore, the
verification analysis is suitable for comparing different forecasting systems
and decision rules, while it does not determine whether it is beneficial for a
water authority to implement AWM. To resolve this an evaluation of the
hits, missed events and false alarms is needed. In meteorology this is often
done with a "cost-loss" analysis where the results are shown for a range of
cost/loss ratios. In such an analysis the "cost" refers to the costs of
anticipatory management actions (such as an evacuation) and "loss" refers to
the damage costs of a critical events when no measures are taken. Although
working with cost-loss ratios is good for inter-comparison, it is not suitable
for deciding whether or not to adopt a new strategy, because the method
does not work with absolute evaluation.
Cost-loss ratios can be based on absolute damage estimates, but the
assumption of a fixed ratio is too much of a simplification for AWM. In
water resources management every event is different and so are the AWM
actions. Therefore there is a need to prepare a dynamic, absolute valuation of
the operational water management. This would estimate the total value (cost)
over a given analysis period. Then current and alternative operational
management strategies can be compared in a cost-benefit analysis (Figure
3.10).
Total cost
Current strategy
Real-time control
Anticipatory control
Time
Figure 3.10 Total cost estimation for alternative operational water management
strategies
74
Anticipatory Water Management
3.6.2 Sources of damage
There are wide varieties of costs and of ways for these to be included in a
cost model. Traditional damage functions of direct tangible costs (e.g.
damage to houses because of flooding) are mostly used. In addition more
and more indirect tangible (disruption of economic activities ) and intangible
(loss of life, social disruption, loss of credibility of a warning system when
false alarms are being issued) damage costs are also taken into account. The
tangible costs are usually expressed mathematically as functions of water
system state variables such as water level, whereas the intangible costs are
often not expressed or expressed in terms of fuzzy membership functions.
Aside from the events, there are also continuous costs, such as operational
costs. Operational costs can be related to power costs for operating
structures, maintenance costs (lifecycle), environmental costs and social
costs.
In this research we propose an extension of the traditional damage cost
functions to the time duration cost functions. Mostly the damage functions
are constant for every time step, while in reality damage can grow
exponentially when the water system stays in an undesired state. For
example, flooding damage to crops can be highly dependent on the duration
of the inundation.
First, for the screening of new forecasting products and AWM strategies, the
most prominent sources of damage that can be estimated directly with the
same continuous modelling system as is used for the warning verification.
Defining the cost functions, expressed in monetary units or fuzzy numbers
etc, must be done by the water authority. They should define these costs,
because the results should give them convincing information about what
strategy and what decision rule to prefer over others.
The reduction of costs due to AWM actions, such as anticipatory control and
emergency protective measures, depends on the measures themselves, but
also on the lead-time provided and the severity of the high water or flood
event. In the case of evacuation, or unclear regulations, the reduction of costs
also depends on the credibility of the warning and the decision support
system. This credibility will for instance decrease with false alarms. As a
consequence of false alarms, people may choose to ignore evacuation orders,
and operational managers may choose to wait for more data to come in.
3.6.3 Anticipatory Water Management modelling
To perform the cost-benefit analysis, the continuous simulations with the
water system model have to include the emulation of the control strategy as
Framework for developing Anticipatory Water Management
75
well. A continuous simulation of operational management of the water
system has to be made. In the case of an AWM strategy, this means that
meteorological forecasts have to be input to the hydrological/hydrodynamic
model and the water system control model with the normal control rules.
Decision rules when the resulting warnings produce a temporary shift to
AWM control rules should be incorporated. Then the AWM control actions
are modelled and the effect on the water system state variables will be
known. This results in a continuous time series of water system state
variables, on the basis of which the total cost over the analysis period can be
calculated.
In real-life application every time a new precipitation ensemble forecast
comes in, the ensemble water level forecasts will be updated (or even more
often, e.g. for every time that the control strategy needs to be updated (e.g.
one hour)). If a yes/no decision has to be made first, whether AWM or
normal control will be applied, these updated EPS water level forecasts will
be done assuming normal control to see whether the water levels remain
within the target range. Updating can be important to reduce the number and
duration of false alarms. Note that because normal control is applied for the
forecast, the modelled system state will immediately start returning to its
normal range, regardless of current anticipatory actions. As a result, for fast
responding systems, and far forecast horizons, the effect of updating
forecasts will be limited.
On the basis of the warning verification and with the help of global
optimisation methods a limited number of suitable decision rules should be
evaluated in this way. The resulting total cost estimates can be compared
with the costs of the current control strategy and other non-AWM strategies.
This overview of costs of a number of suitable AWM strategies gives
valuable information to the water authorities about it may cost to adopt
AWM and what decision rules are efficient and what are not.
This ends the screening of AWM. If the screening results are satisfactory, a
further optimisation of the control strategy and decision rules needs to be
done at the next stage.
3.7 Optimisation of Anticipatory Water
Management
The main problem of optimising the control strategy of water systems is that
the variation in potential control strategies is usually very large and
dependent on a large number of different decision variables, not only on the
wide range of an individual variable. This multi-year optimisation problem,
in which per day several ensemble predictions are available and the best
76
Anticipatory Water Management
management strategy for the entire period needs to be defined, cannot be
captured in an analytical optimisation model. Therefore, global optimisation
methods with smart search methods, like evolutionary approaches, are used.
3.7.1 Objectives
For each case study the objectives need to be defined and agreed upon before
optimisation can take place. Usually the optimisation problem will be a
multiple-objective problem. In the case of flood control of a reservoir,
example objectives could be to:
1. Minimise flood damage cost
2. Minimise total damage cost
For a multi-objective problem (with conflicting objectives) to result in one
optimal solution, weights have to be given to the objectives. Because this is
always a heavily debated (before and after) process, of which the
consequences for the end result are not known up-front, preference is often
given to the provision of multiple possible optimal solutions in a Pareto
Front (Coello, 2005; Barreto et al., 2006).
3.7.2 Parameterisation of AWM strategies
The parameters that have to be optimised are the variables that make up the
warning and operation rules for the AWM strategy. Therefore the strategies
discussed in Section 3.5 have to be parameterised. For example, for a
strategy following a simple ensemble based threshold decision rule for the
early lowering of a storage basin water level to a fixed level:
If forecasted probability P(water level Y days from now > H m+Ref) > N,
then at A days from now start lowering water level to Ha m+Ref.
the decision parameters become:
- Water level threshold (H)
- Forecast horizon(Y)
- Probability threshold (N)
- Anticipation time (A)
- Anticipation water level (Ha)
However, when elaborating this strategy, an update frequency (control
decision time step) and a rule on how to deal with inconsistency in the
forecasts are needed. It can be seen that for complete control strategies the
number of optimisation parameters grows fast. Therefore, it has to be
considered whether sub-sets of the parameters can be optimised separately.
Framework for developing Anticipatory Water Management
77
3.7.3 Optimisation using perfect forecasts
In some cases of sub-optimisation, some of the decision variables can be
selected on the basis of perfect forecasts. In this way, the interpretation of
the probabilistic precipitation or water level forecasts is separated from the
actual anticipation action, since the decision to take action has been taken.
The perfect control actions for a given system load do not depend on the
forecast, but are in fact system characteristics. Once the decision variables
that are system characteristics have been defined, these can be separately
optimised, using the water system control model with perfect (measured)
forecasts. Examples of these parameters are the optimal anticipation time
and the update frequency. In addition optimisation with perfect forecaststs
shows the maximum benefit that can be achieved by applying AWM.
For example, in the case of flood control, the maximum anticipation time
(control horizon) required is determined by the maximum flood event in the
verification analysis and its antecedent and post event conditions. These
together determine how much time is needed to pump out the excess volume
of water (before (and or) after the event). Further expanding the control
horizon beyond this time has no effect on the analysis. Therefore the desired
maximum control horizon should be determined with the design storms.
Design storms do not necessarily have to be part of the verification archive.
Storms can be defined for larger return periods such that they take into
account expected climate change. Note that reference is made here to the
control horizon assuming a perfect forecasting system. A larger control
horizon may be chosen in reality, when working with imperfect forecasts, to
account for the possibility that the event is forecasted too late.
3.7.4 Optimisation with actual forecasts
A wide range of strategies can and may have to be tried in order to come up
with a reliable optimal control strategy for real, imperfect, forecasts. The
different kinds of strategy as discussed in Section 3.5 have to be optimised
separately. Rule based AWM strategies can be optimised by evolutionary
search methods, such as Genetic Algorithms. Rule based AWM strategies
need optimisation of two main components. The first is the interpretation
(pre-processing) of the hydro-meteorological forecasts in general (long-term
strategy optimisation). The second concerns the short-term optimal control
actions with the regulating structures. Through the simulation of a particular
strategy for a long historic period for which measured data is available, the
objective functions can be estimated. A Pareto front of the multi-objectives
can be produced (Figure 3.11). In this way the short term, real-time,
management actions, and the long term operational strategy can be optimised
simultaneously (layered optimisation).
Anticipatory Water Management
Criterium 2
78
AWM strategies
Criterium 1
Figure 3.11 Pareto front for a 2-objective (criteria) optimisation problem with
AWM strategies
3.8 Decision making for policy adoption of AWM
3.8.1 What-if analysis
The analyses described will be informative and convincing because they
make use of measured data and archived forecasts. Therefore, they clearly
show what could have been done with AWM in the past to improve the
management of critical events. However, for policy decisions to adopt AWM
we need to consider what these results in the past tell us about the future.
Design storms could be used, but the associated forecasts are not available,
so there appears to be no guarantee that a developed AWM strategy will
perform just as well for more extreme events, in changed climate conditions.
However, the atmospheric models are physically based and make use of realtime data assimilation. This implies that if these types of events occur in
places they did not occur before due to climate change, they will be
forecasted by the models just as well. Therefore, it can be assumed that the
performance of AWM, achieved with archived data will not deteriorate in
the future. With the continuous further development of the hydrometeorological numerical modelling, it may even be hoped that performance
will only improve.
What is more important, however, is that the reliability of the expectations
we will get from these analysis increases with increasing simulation periods.
Therefore archiving of measured data and re-analysis and hindcasting are
becoming ever more important.
Framework for developing Anticipatory Water Management
79
3.8.2 Re-analysis era
Developments in water system modelling have resulted in reduced
computational demand, and at the same time developments in parallel and
grid computing, together with the ongoing increase of processor speeds have
reduced computational time. Together these developments greatly enhance
the use of simulation models, scenario analysis, and optimisation in both
operational and strategic water management. The analyses described in this
work show how advantage can be taken of these new opportunities in
practical additions to the current analyses of the water authorities.
3.9 Framework for developing Anticipatory Water
Management
The methods described in the previous sections in response to the knowledge
gaps described in Section 2.6 together form a framework to develop,
evaluate and adopt an Anticipatory Water Management strategy for a given
water system (Figure 3.12).
This framework supports water managers to evaluate a new forecasting
product for application in Anticipatory Water Management (AWM). The
main part of the framework consists of steps that perform a screening of new
forecast products and control strategies. The outcome of the screening
should be twofold: It should indicate the range of suitable decision rules for
AWM and it should benchmark the proposed operational management
strategy against current management and alternative strategies.
Then if the forecasting product is selected and suitable AWM strategies
seem to be available, as a second stage, optimisation of the AWM strategy
can be performed.
After this follow the stages of implementing the decision support system,
learning how to use it, and building the confidence of the operational water
managers. At some stage the new operational strategies need to be
incorporated in the legislative policies of the particular water authority.
In general technological advances bring shifts of responsibilities from one
group to another. This poses strains on an organisation that adopts
technological change, as with any other change. These strains have to be
dealt with by involvement of all affected groups from the beginning,
building consensus and trust that the changes are for the better, and that the
new tasks for the different groups are clearly defined and satisfactory to all.
In chapters 4 and 5 the AWM framework is tested in case studies.
80
Anticipatory Water Management
Determine whether AWM is needed
Event selection, threshold selection
Identify necessary control action,
including forecast horizon
Identify forecasting product
Verify reliability of forecasting product
for critical events
Propose and evaluate new AWM control strategies
Perform cost-benefit analysis
Optimise AWM strategy, e.g for minimum cost, using
design events and re-forecasting
Test preferred system with historical data
Implement, monitor, review
Figure 3.12 Framework for developing Anticipatory Water Management. The main
part of the framework consists of steps for screening of new forecast products and
control strategies. If new control strategies perform well, in the next step the
optimisation of the AWM strategy can be performed.
81
4
Case study 1 - Rijnland Water System
4.1 Introduction
Rijnland is a polder area in the western part of the Netherlands, bordering
the North Sea (Figure 4.1). The total area is about 1000 km2 of which 72% is
occupied by low-lying land-reclamation areas, 15% by free draining areas
and 8% by dunes. A storage basin consisting of inner connected canals and
lakes, occupies 45 km2. The storage basin serves to collect all the excess
water of the Rijnland area, before it is discharged to the main water system
of the Netherlands and finally to the North Sea. The low-lying areas would
be subject to flooding if they were not protected by dikes and the excess
water not pumped to the storage basin. The water level in the storage basin is
kept between predefined bounds, mainly by the daily operation of four large
pumping stations: Halfweg, Katwijk, Spaarndam and Gouda. The total
capacity of these four pumping stations is 154 m3/s (13.3 mm/day). In case
of extreme events also pumping station Leidschendam may be used to
discharge water to the Delfland Water Board (8 m3/s).
The area consists of urban and rural parts. The rural parts can be sub-divided
into areas committed to horticulture, agriculture, and grass lands. The
dominant soil types are sandy in the free-draining and dune areas, clay in
part of the land reclamation areas, and peat in the main part of the land
reclamation areas.
Excess water is discharged from the urban areas to the main storage basin
through waste water treatment plants. Combined sewer overflow discharges
end up in the drainage network of the rural areas, and through small
pumping stations are pumped to the main storage basin.
During and after rainfall events, excess water from the land reclamation
areas is pumped directly to the main storage basin by over 200 small
pumping stations. Further hydrological load to the storage basin comes from
the free draining areas and excess water from the neighbouring water board
"Woerden", which is discharged to the Rijnland storage basin through an
inlet (Bodegraven inlet).
During summer and dry spells, the channelled storage basin is flushed by
combined operation of smaller inlets and sluices, mainly Gouda inlet and
KvL sluice, and the four main pumping stations.
82
Anticipatory Water Management
The Netherlands
North Sea
Rijnland water board
Pumping station
Figure 4.1 Principal Water-board of Rijnland: controlling a low-lying regional water
system in the western part of the Netherlands. A channelled storage basin collects all
the excess water of the area. The water level in the storage basin is controlled by
four pumping stations.
The response time, taken as time from peak of precipitation to the end of the
first half of the associated pumping period (representing the peak of the runoff response), varies between 0.5 to 1.5 day. Time to empty, taken as the
time from the peak precipitation to the moment that the water level is back to
normal, is about 3 to 4 days for big events (60 mm/3days).
A simulation model of the water system is used in a decision support system
(DSS) for operation of the four pumping stations to keep the storage basin
water level within a 0.05 m target range (Figure 4.2).
The following sections describe the results of each step in the framework for
developing Anticipatory Water Management (Figure 3.12 as applied to the
Rijnland water system.
Case study 1 - Rijnland Water System
83
Water level (m+Ref)
Critical upper-boundary
-0.50
Extreme range
4.5 x 106 m3
-0.60
Regular range
-0.65
-0.70
2.2 x
106
m3
Forecasting range
Critical lower-boundary
Without forecasting
With forecasting
Figure 4.2 Water level control of the Rijnland storage basin, with and without
forecasting. When using forecasts and temporarily allowing lower water levels, extra
storage of 2.2 x 106 m3 can be created before the extreme event occurs.
4.2 Problem description
The Rijnland area has faced both extreme precipitation events (1998, 2000)
and droughts (2003). Research has concluded that the required safety level
against floods is no longer being met. The estimated probability of
exceeding the critical water level in a year is more than 0.01, and therefore
remedial structural measures are planned by the Principal Water-board of
Rijnland (Rijnland, 2000). Emergency storage basins are to be allocated and
the pumping capacity is to be increased by 40 m3/s (3.5 mm/day), which is
26% of the present pumping capacity (13.3 mm/day). In addition, the
Rijnland water board would like to optimize the operational flood control of
its water system. The proposed anticipatory measure is to create extra
storage in the basin when extreme hydrological loads are expected.
Anticipatory pumping can lower the storage basin water level below the
regular range (Figure 4.2) before the extreme event occurs. The pumping
would create extra storage, which is comparable to that of the planned
emergency basins. Swinkels (2004) used offline control simulations to show
that for the extreme event of 2000, during which the –0.50 m + Ref (Dutch
reference level ~ mean sea level) level was exceeded, a forecast horizon of at
least 3 days would be necessary with an allowance of 0.08 m extra storage,
to prevent the water level from exceeding the maximum permitted value.
The problem is that low water levels may have adverse effects, such as
hindrance of navigation and damage to houseboats, and for very low water
levels the problem is the risk of embankments becoming unstable and soil
subsidence, bringing damage to nearby houses (generating an economic
risk). The water board started to apply heuristic rules for anticipatory
pumping to increase safety against floods. A daily precipitation threshold of
15 mm was chosen as the alert threshold, using the precautionary principle
that the forecast may underestimate the precipitation. If the 1-day
precipitation forecast exceeds this threshold, early pumping is considered an
84
Anticipatory Water Management
option to create extra storage in the storage basin. The water board would
like to extend the forecast horizon to 3 days or more.
4.3 Data
The Rijnland water board operates a telemetry system and has archived over
20 years of time series data. In this study 10 min. precipitation data of the
water board, and daily precipitation data of the Dutch National
Meteorological Institute (KNMI) are used as input. The meteorological data
is processed with HydroNet DSS (HydroNet, 2009) The daily data is
validated by the KNMI and has no missing data. The 10 min. precipitation
data contains outliers and missing data. When these occur, only the daily
data of the KNMI are used. For evaporation data the daily Makking
reference evaporation from 3 KNMI ground stations is used. These data are
validated and contain no outliers. If a station contains missing data, it is not
considered. For the analysis period of this study there was always at least
one of the three stations functioning properly. For sub-system calibration
(Section 3.4.2) also daily flow data from the water board is used. This data is
validated and contains no missing data or outliers.
Rainfall
In the future operational DSS, rainfall radar may be used as input to the
water system model. However, for the calibration and validation of the
model up to t = 0 the available archive of radar data at the time of calibration
was limited to one year, while the available ground station data was (more
than) 7.5 years. With respect to rainfall input the model is fully lumped
meaning that the area average rainfall is used as input. This is because the
area average water level is the target variable for operation requirements,
and because in the storage basin of connected canals, local effects of rainfall
flatten out quite fast. To check whether the use of radar would not change
the model performance that is calibrated with ground station data, the area
average rainfall from both sources is compared (Figure 4.3).
In the forecast verification analysis, daily precipitation data for 16 ground
stations from the KNMI and 10 min precipitation data from 6 stations from
Rijnland is used to estimate hourly area-average precipitation in the Rijnland
area. The Thiessen average daily sums of the KNMI stations are preserved.
The hourly distributions of the daily sums are taken from the hourly
Thiessen average of the 10 min data.
The radar data is stored in 3 hourly sums, updated every hour, for a grid of
regular 2.5 km by 2.5 km cells. For both sources the data has been
aggregated to daily sums for December 2004 and presented in Figure 4.3.
The result shows little difference between the sources (3-hourly rainfall data
is provided by the KNMI after calibration with ground station, so this is not
Case study 1 - Rijnland Water System
85
very surprising). The maximum daily difference is 0.9 mm, the difference in
the month sums is 0.6 mm. Therefore in this case study the archive of
ground station data can be used for the calibration and validation, without
expecting too many problems when switching to using radar data in the
operational application. Hourly precipitation data is used.
12
Precipitation (mm)
Radar
10
Ground stations
8
6
4
2
0
29/11/2004 04/12/2004 09/12/2004 14/12/2004 19/12/2004 24/12/2004 29/12/2004 03/01/2005
Figure 4.3 Comparison of radar and ground station precipitation estimates for the
Rijnland area. The graphs show close resemblance for both dry and wet periods.
Evaporation
The input evaporation data is provided as daily Makkink Reference
Evaporation from three KNMI meteorological stations in the area. Forecasts
of evaporation are not used, because of the limited effect on peak discharge
events within a 10-day horizon.
Precipitation forecasts: ECMWF EPS
Precipitation forecasts of the ECMWF EPS are used. The ECMWF began
producing EPS forecasts operationally in December 1992 with 33 members
(different runs) of their global circulation model (Molteni et al. 1996). The
EPS is under continuous development. Since 1996 the model has been run
52 times for each forecast: one run with a high spatial resolution (operational
run), one run with the EPS spatial resolution and un-perturbed initial
conditions (control run), and 50 ensemble members with perturbed initial
conditions. In 2000, the spatial resolution of the operational run increased
from roughly 60 km to 40 km and the EPS from 120 to 80 km. In 2006, the
spatial resolution of the EPS was further increased to 50 by 50 km. Since
1998, also a scheme for model error has been included (stochastic forcing).
86
Anticipatory Water Management
The development of the operational EPS is described on the ECMWF
website (ECMWF, 2007).
The assumption is that the perturbed initial conditions are determined in such
a way that the 50 ensemble members are equally likely to occur (Persson and
Grazzini, 2007). If, for instance, five ensemble members predict a certain
precipitation, then the forecasted probability that this precipitation will occur
should be about 0.1. Research has shown that the ensemble members with
higher and lower precipitations have a higher probability of occurrence than
do the ensemble members with average precipitation (Bokhorst and
Lobbrecht, 2005). The EPS is run twice a day, and the output consists of
atmospheric states, expressed in grid-averaged values of a number of
variables, for every 6-hour time step. The forecast is made for 10 days
ahead.
ECMWF supplies the national weather institutes with the time series for
selected variables, such as precipitation and evaporation, interpolated to
requested locations (ECMWF, 2006). The Royal Netherlands Meteorological
Institute (KNMI) provides the forecast time series for a number of locations
in the Netherlands to Water Boards through the Internet. In this case study,
the ECMWF EPS precipitation forecasts of the 50 perturbed ensemble
members for forecast station De Bilt are used. This is the nearest available
location, located about 40 km to the east of the Rijnland water system. The
forecasts are compared with area-average measured precipitation to avoid
large spatial scale differences (local extremes versus large-scale
precipitation).
Downscaling techniques and bias analysis have not been applied in this
research; the aim is to establish what can be done by adjusting the decision
rules with the ECMWF EPS forecasts as they are.
4.4 Water system control model
4.4.1 Model structure
The Rijnland water system is modelled with the Aquarius modelling
software (van Andel, 2009a). Aquarius is an object oriented non-linear
reservoir model. Surface water, layered soil columns, ur-ban areas, and
green houses make up the elements of the reservoirs. Most important
additions to the rainfall-runoff process are the control structures. Weirs,
sluices, inlets and pumping stations can be modelled, including the control
methods used, such as PID, local switch on/off levels, and global control
(Lobbrecht, 1997).
Case study 1 - Rijnland Water System
87
The Rijnland water system is modelled with three distinct sub-areas,
representing the storage basin and dunes (Rijnland storage basin), landreclamation areas with clay soils (Rijnland Polder 1), and land-reclamation
areas with peat polders (Rijnland Polder 2). During events with excessive
rainfall, Rijnland receives excess water from the neighbouring water board
(Woerden). The Woerden area is modelled similarly with a storage basin
area, and a land reclamation area (Figure 4.4).
The characteristics of the sub-systems and the control structures have been
provided and estimated by the Rijnland Water Board.
Figure 4.4 Aquarius water system control model of Rijnland (Yufeng, 2003)
88
Anticipatory Water Management
4.4.2 Control strategy
The envisaged application of the Aquarius model is to operational use in
decision support for the operation of the four main pumping stations, and for
research to compare control strategies. The latter is done on the basis of long
term simulation of the Rijnland water system and its operation for a multiyear period. Therefore the simulation period applied for this building this
model is 1997 - 2002 for calibration and 2003-2004 for validation, for which
a comprehensive archive of water system data and reports was available.
To evaluate the model's capability to be used as a decision support tool for
operational management and for the development of control strategies, the
control strategy that was applied during the simulation period (1997-2004) is
modelled. This strategy can be described as a global control strategy, with
qualitative inclusion of rainfall-runoff forecasts up to 1-day a head.
When analysing the water system data, however, it could be seen that the
strategy could be well simulated with local automatic control. Therefore,
local automatic control will be used to calibrate and validate the water
system control model with the 1997-2004 data.
The application of forecasting for one day or more, requires that the inflow
into the storage basin is not taken from real-time measured data, but
modelled as well. Therefore, structures that have a discharge function during
excess rainfall events have to be fully modelled (to provide predictions in
operational mode). The inlet structures and sluices that have a regulatory
function, like flushing or acting as a water inlet during dry spells, do not
have to be modelled because they are not responsively operated, but are
scheduled tasks. When, for example, the scheduled task of flushing appears
not to be necessary it can be decided not to do it in real-time, on the basis of
measurements. Therefore, modelling of these structures control strategies for
the prediction of critical events with excessive water is not necessary. Even
more so, to predict whether Anticipatory Water Management is necessary,
all water inlets in the model have to be set to be closed (zero inlet). If then,
the modelled discharge structures still do not manage to maintain the water
levels below the upper target level, then anticipatory control actions are
needed.
Reduced pumping capacity during high tide, because of high sea level, is not
modelled because the reduction of pumping capacity is limited in volume
and time (tide) and the reduction is not incorporated in the current
operational decision support system. In addition the pumping system will be
adjusted in the near future to further reduce the capacity reduction during
high tides. Through shiplocks small amounts of water come into the storage
basin regularly. Because there are no data of this inflow, it is modelled as
part of an constant external inflow to the Rijnland storage basin, estimated
Case study 1 - Rijnland Water System
89
by the water board on the basis of information on hydrological loads and
yearly water balance studies.
4.4.3 Model calibration
In the Rijnland case study the target variable is water level in the storage
basin. For calibration of the water level data a critical rainfall event with
high storage basin levels, in November 2000 was selected.
Local automatic control is modelled, therefore, the control parameters to be
adjusted are the on- and -off set points of pumps and inlets, and open-close
set points for sluices. Whereas the set points of control structures can be
known or inferred from the water board's information on the applied control
strategy, the soil in- and outflow resistances are highly uncertain. Soil type
gives only ranges of possible resistances, and on top of this, soils in most
catchments are highly heterogeneous, often further complicated by a varying
drain network, which makes the area averaged soil in- and outflow
resistances highly uncertain. Therefore, the latter are the main calibration
parameters. For the three Rijnland sub-systems, two land reclamation areas
and the storage basin, three soil layers have been defined for each of which
separate in- and outflow resistances are defined. This makes a total of 18
calibration parameters. The Woerden sub-system has one reclamation area
with three soil layers, which makes six parameters to be calibrated. In
addition, the switch -on and -off levels of the Rijnland storage basin
pumping station that were received from the water board were further
refined through calibration (12 parameters).
Water level peaks are the most important events to be calibrated, because the
main operational purpose of the Aquarius Rijnland model will be decision
support for flood control. The modelled and measured slope of the rising
water level have to match well, the peak water level has to be modelled
accurately, and the slope of the receding water level is also a good indication
whether the runoff process are modelled well. The normal flow periods are
important indicators for the quality of the model as well. The slope of the
rising water level after discharge control actions have stopped, and the slope
of the receding water level when discharge is taking place are determined by
the rainfall runoff process, soil moisture content and groundwater level, and
groundwater in- and outflow resistances. Therefore the criteria for
calibration of the Rijnland model are visual similarity of measured and
modelled water level peaks, and normal flow periods.
Figure 4.5 shows the calibration result for the event of November 2000. The
rise of the water level around 10 November has been modelled especially
well. Note that the deviation in the beginning of the peak can be the result of
manual operation in reality, which defers from automatic local operation in
90
Anticipatory Water Management
the model. After the peak the modelled water level recedes not as fast as the
measured water level. This can be the result of errors in the rainfall input,
but also it could be the result of sub-optimal ground water flow resistance
parameters, initial conditions, or even surface area errors. As is often the
case in both hydrological and water system control modelling, if there are
still errors remaining, after the best of knowledge has been put into the
model, it is unknown what is causing the remaining error.
The simulation period for calibration has to be long enough to prevent
additional errors due to wrong initial state before the event. Note that when
changing the simulation start time, it needs to be checked that the initial
conditions (water level, groundwater level, soil moisture content) match the
new start date.
Low flow periods, such as in February 2000 (Figure 4.6), were simulated
accurately as well. Slopes of rising and receding water levels match well.
Again, small differences remain, but these most likely come from the
difference between manual and full automatic control.
As a next step validation was performed.
4.4.4 Model validation
The model has been validation for the year 2002 on the basis of monthly
discharge volume through the four main pumping stations. For comparison
the annual report monthly values have been used (Figure 4.7). April to
August, summer months, are the relatively dry months in the Netherlands.
Although the water level modelling results were reasonable the monthly
pumped discharge volumes are clearly under-estimated by the model. In
October and November 2002 the pumped volume is over-estimated by the
model. For the other months the validation is satisfactory.
In some cases, for the water boards, this modelling result could be
considered satisfactory, because in general flood events occur less frequent
in summer months than in winter months, and because despite the poor
validation in October and November the modelled water levels compared
quite well. Note that this is evidence of the point made in the introduction
that water-system control modelling has a high degree of freedom.
On the other hand, Figure 4.7 could raise the question why the pumped
volume is strongly underestimated in April and May and then gradually
improves to result in an over-estimation in October and November, and
whether this is a systematic error or not.
Case study 1 - Rijnland Water System
* 0.0
-0.30
-0.38
-2.5
-0.46
-5.0
-0.54
-7.5
-0.62
Precipitation (mm/h)
Water level (m+ref)
91
-0.70
-10.0
27-10-2000
4-11-2000
12-11-2000
20-11-2000
28-11-2000
Rijnland modelled: Water level (m+ref)
* Precipitation (mm/h) (-)
Rijnland modelled: Rural groundwater level (m+ref)
Rijnland measured: Water level (m+ref)
Figure 4.5 Calibration of Aquarius water system control model of Rijnland, the
Netherlands, for a peak water level event in November 2000.
* 0.0
-0.50
-2.5
-0.58
-5.0
-0.62
-7.5
-0.66
Precipitation (mm/h)
Water level (m+ref)
-0.54
-0.70
-10.0
8-2-2000
16-2-2000
24-2-2000
3-3-2000
11-3-2000
Rijnland modelled: Water level (m+ref)
* Precipitation (mm/h) (-)
Rijnland modelled: Rural groundwater level (m+ref)
Rijnland measured: Water level (m+ref)
Figure 4.6 Calibration of Aquarius water system control model of Rijnland, the
Netherlands, for a normal flow period in February and March 2000.
1.20E+08
Pumped volume (m3)
1.00E+08
modelled
Error total modelled
volume 2002:
reported
-11.40%
8.00E+07
6.00E+07
4.00E+07
2.00E+07
0.00E+00
Jan-02
Feb-02 Mar-02
Apr-02 May-02 Jun-02
Jul-02
Aug-02 Sep-02 Oct-02 Nov-02 Dec-02
Figure 4.7 Validation of the Aquarius water system control model of Rijnland, on
the basis of monthly pumped discharge volumes.
92
Anticipatory Water Management
Therefore, in this study the validation is expanded with cumulative pump
volume checks over a longer period, from 1997 to 2002 (Figure 4.8), and
with cumulative pumped discharge volumes during excessive rainfall events
in 2000 and 2001 (Figure 4.9). The horizontal sections of the modelled
cumulative discharge in Figure 4.8 confirm that the model is systematically
underestimating discharge volume in summer (dry season). The modelled
discharge of both winter events in Figure 4.9 confirm that the model is
systematically overestimating discharge in winter (wet season).
When discussed with the water board, it appeared that this is a known
problem and that from the expert knowledge of the system several possible
responsible processes have been identified, but that the issue in the
modelling had so far not been resolved. In the following sections, following
the approach proposed in Figure 3.7, the model results are further analysed
to visualise the unknown prosesses and try to explain them.
4.4.5 Visualise what is not known and explain
In Section 4.4.4 it was concluded that the pump discharge was systematically
underestimated by the model during dry season and over-estimated during
low season. In the flowchart (Figure 3.7) the question of the accuracy is
therefore answered with "no". The model error, "the unknown", is visualised
in Figure 4.8 and Figure 4.9, and the next step is to discuss with the water
system experts whether these differences between model and measurements
can be explained. This step brings into practice what is often described as
"learning from models".
The first question is what makes the modelled pumped out volume in the
summer too low. The answer requires the uncertain input variables or
calibration parameters that can cause errors in the long term (monthly,
seasonal) processes to be identified. The variables and parameters are as
follows:
- Actual evaporation
- Human water use (Dry weather flow)
- Variation of infiltration and seepage flows
- Rainfall-runoff process (groundwater flow resistance parameters)
- External inflows
- Groundwater storage change
- Contributing surface area in the dunes
First, when looking at Figure 4.7 for the reported volumes in the months
April to July it seems likely that part of the inflows and inlet discharges
(flushing, dry weather flow) during the summer is unaccounted for.
Therefore fixed inflows are increased in the model to improve the summer
pumping. However, because these fixed flows are assumed constant for the
Case study 1 - Rijnland Water System
93
4.50E+09
4.00E+09
Pumped volume (m3)
3.50E+09
3.00E+09
2.50E+09
2.00E+09
1.50E+09
Pumping modelled
Pumping Measured
1.00E+09
5.00E+08
0.00E+00
Jan 03
Jul 02
Jan 02
Jul 01
Jan 01
Jul 00
Jan 00
Jul 99
Jan 99
Jul 98
Jan 98
Jul 97
Jan 97
1.40E+08
1.40E+08
1.20E+08
1.20E+08
1.00E+08
8.00E+07
6.00E+07
4.00E+07
Pumping modelled
Pumping measured
2.00E+07
0.00E+00
1.00E+08
8.00E+07
6.00E+07
4.00E+07
Pumping modelled
Pumping measured
2.00E+07
0.00E+00
28/09/2001
26/09/2001
24/09/2001
22/09/2001
20/09/2001
18/09/2001
16/09/2001
20/11/2000
18/11/2000
16/11/2000
14/11/2000
12/11/2000
10/11/2000
08/11/2000
06/11/2000
04/11/2000
(a)
Pumped volume (m3)
Pumped volume (m3)
Figure 4.8 Cumulative pump discharge volume from the Rijnland storage basin.
Modelled volume is too low, because of underestimation during the dry summer
seasons.
(b)
Figure 4.9 Cumulative pump discharge volume for events in the wet winter season
in 2000(a) and 2001(b). For both events modelled volume is higher than measured
volume, indicating over-estimation of the model during excess water events in the
wet season.
whole year, this exacerbates the problem of too much pumping in the wet
season.
Secondly, increasing the soil inflow resistance helps to increase pumped
volumes in the dry season, because it reduces flow from the surface water to
the ground water. Groundwater is allowed to drop a little during dry season,
and surface water levels are maintained more easily. Increasing the soil
inflow resistance in the early winter months (September, October) also helps
94
Anticipatory Water Management
to reduce the modelled pumping, because the lowered groundwater levels
need longer time and therefore more volume to rise during the wet season,
before contributing to high flows during rain events.
The storage basin and storage basin land is a fast reacting system (due to
sandy soils), including the groundwater, so the soil inflow and outflow
resistances cannot be increased. In part of the land reclamation areas
however, high soil inflow resistances seem to be present and groundwater
levels have indeed been reported to drop during the summer by up to a
meter. It could be that soil inflow resistances become very high because of
siltation on the canal beds. Therefore in the land reclamation areas with clay
soils, high soil inflow resistances are modelled, which explains part of the
problem of too low pump discharge in summer and too high discharge in
winter.
The other possible sources of error could not be clearly explained in the
discussion with the water board experts. Therefore, first the model is run
again with the adjustments mentioned, following the left track of the
modelling framework (Figure 3.7). To allow the modelling of the Rijnland
area separately (Figure 3.7, sub-system calibration), the available measured
in- and outflows from the Woerden area and other boundaries (Gouda inlet,
KvL sluice) are used as input to the model.
Then, after this first iteration, the remaining errors have to be presented
again (Figure 3.7: Visualise the unknown). The daily difference between
measured and modelled discharge from the Rijnland storage basin is plotted
(Figure 4.10). The difference between measured and modelled daily pump
discharge shows short (1 day) deviations, both positive and negative, up to a
maximum of about 4*106 m3, which is less than a third of the total daily
pumping capacity (1.3*107 m3). These differences can be caused by small
timing differences between the local automatic (modelled) and the manual
(measured) control strategies.
The effort is to find out the longer term processes (e.g. seepage, water
consumption). These processes are expected not to vary too much from day
to day, but do have a monthly change and variability over the year, gradually
changing along patterns following the dry and wet seasons. One assumption
often made is to discretisise between summer and winter season, because
indeed in the Netherlands water management is changed from one day to the
other when it is decided to switch from winter to summer target levels and
vice versa. Still the natural process, cropping seasons, and domestic and
industrial water uses follow more graduate trends that are often not known
and not modelled in the water system control models for the water board.
Therefore a time scale analysis of the model errors has to be performed to
filter out the fast processes that are to be calibrated later, and to capture the
Case study 1 - Rijnland Water System
95
unknown slow varying (averaged) processes. When the time scale is scaled
up to 10, 20 up to 90 days a clear sine pattern emerges (Figure 4.10). This
sine pattern has a 1-year period, which was checked to hold for 6 years of
data.
5.00E+06
4.00E+06
3.00E+06
Pump discharge volume (m3)
2.00E+06
1.00E+06
0.00E+00
-1.00E+06
-2.00E+06
-3.00E+06
Daily
90 day moving average
-4.00E+06
10 day moving average
-5.00E+06
01/01/2000
10/04/2000
19/07/2000
27/10/2000
04/02/2001
15/05/2001
Figure 4.10 Time scale analysis of difference between measured and modelled
pump discharge. At 90-days moving average a clear sine function with a yearly
period becomes visible.
4.4.6 Modelling the unknown phenomena
Experts from the water board had two main notes on the remaining sine
function error. The first is that the errors probably come from not fully
capturing the soil and groundwater processes with the model. Small errors
are inevitable there, and because of the large volumes involved these small
errors cause big water level and volume balance differences between
modelled and measured values. Secondly the sine could result from
gradually changing groundwater levels in the dunes (including semicontrolled drinking water production) causing dynamics in the seepage flow
to the polders. Unknowns in the actual evaporation throughout the year
(dynamic land use, harvest times etc) and domestic water use (industrial
water use is limited) may be additional causes of the sine error shape,
although part of the changes in actual evaporation are covered by the
monthly Makking evaporation crop factors in the model.
96
Anticipatory Water Management
Since the model was already updated on the basis of the validation, and since
the discussion of the sine function with the Rijnland water board
representative did not result in a uniquely defined process that is causing the
sine, it was decided to apply a data driven model for the sine function and
use it to define the external inflows to the model.
4.00E+05
Sine function:
Qin = -3.51*10^5 + 3.26*10^5 sin(2PI/365*(t-T0)
T0 = 28 Feb
Pump discharge volume (m3)
2.00E+05
0.00E+00
-2.00E+05
-4.00E+05
-6.00E+05
90 day moving average
-8.00E+05
Sine function
-1.00E+06
Jan-2000
Mar-2000
May-2000
Jul-2000
Sep-2000
Nov-2000
Jan-2001
Mar-2001
May-2001
Figure 4.11 Sine function to model the slow processes error (90-days moving
average) of the Rijnland Aquarius model.
The model is deduced for the Rijnland land reclamation areas and storage
basin. The 3-parameter sine-function was optimised (Figure 4.11). Note that
only part of the 6-year period is presented. The slow cyclic behaviour can be
modelled well with the sine function. On top of this, there will still be other,
short-term errors that will partly be compensated by calibration and are
partly inevitable because of inconsistencies in the measured time series due
to manual operation of the control structures. The sine function is added in
the model as an external ground water inflow to the land reclamation areas.
In the same way, the sine inflow function to the neighbouring area,
Woerden, was prepared. A special feature there, is that there seems to have
been a system or policy change in January 2001, after which much more
water was discharged to Rijnland. This means that in the model two sets of
control rules for the control structures are used (before and after 2001) and
that two separate sine functions had to be prepared.
Case study 1 - Rijnland Water System
97
4.4.7 Final model results
After the physically based, and data driven adjustments were implemented
the final steps are to prepare the model in the form it will be used
operationally and to re-calibrate and validate this final model (Figure 3.7,
second loop).
For the sub-system modelling, control structure models were replaced by
measurements, but in the operational system not all these measurements can
be used. The model is planned to be used for decision support in operational
management by providing warnings for high inflow, and high water level
events. Therefore, structures that have a discharge function during high
flows have to be modelled. Structures that have a regulatory function during
normal or low-flow periods are operated according to schedules, so these do
not have to be modelled.
The Kocksluice or KvL sluice has no discharge function during excess
hydrological loads to the Rijnland storage basin, so it is modelled using the
measurement as input for times up to time 0. In predictive mode, the
discharge through the sluice will be put to zero.
The Leidschendam pumping station is used regularly in summer with small
volumes, for flushing of Delfland. In winter only it is used incidentally with
peak events like the one in 2000. Therefore, this pumping station is modelled
without its flushing function in summer (because it is not possible to model
the demand from Delfland, and because of the relatively small volumes), and
with pumping at peak events in summer and winter.
Gouda inlet is for flushing the system in the summer, and is therefore not
modelled in predictive mode: the measured data is taken as input.
The shiplocks are regarded as being closed, because the regular small
volume that comes in to the basin has already been taken into account in the
fixed inflows.
So in short, summer flushing is not modelled, because also in an operational
setting this is done according to schedule and will not be done during
excessive rainfall events. All control structures that do have a discharge
function during excessive rainfall events are modelled to show that the
model can accurately reproduce past critical events and is suitable to be used
for predictions. In operational mode, the model could be updated every
decision time step with the measured flows. All relevant discharge and inlet
structures are monitored in real-time with a telemetric system. The Aquarius
software can be controlled automatically by external modules, which allows
full incooperation in an online decision support system. This would ensure
that the model keeps an accurate initial state, during summer and winter.
98
Anticipatory Water Management
Calibration of the final model
With these settings the final model is calibrated by adjusting the soil outflow
resistances. Note that varying the soil outflow resistances does not have a
uni-directional effect, meaning that lower resistances do not always result in
steeper peaks, and that higher values do not always result in flatter peaks.
This behaviour of the soil outflow resistance is because different soil layers,
with different outflow resistances are used. Therefore, increasing soil
outflow resistance in one layer may force the groundwater level into the
upper layer, where lower soil outflow resistances may be used. So, with
calibration, a wide range of soil outflow resistances should be checked with
intervals that are not too big (20 days), for a wide range of events, to find the
best interval; then the optimum can be found with even smaller increments.
The final calibration results are presented in Figure 4.12 to Figure 4.15.
Figure 4.12 shows that the monthly discharge volumes of the year 2002 have
been improved considerably with respect to the first model (Figure 4.7).
Note also the accurately modelled total yearly volume (0.7% error) for 2002,
while only the total volume over 6 year simulation (1997-2002) was
calibrated with the sine functions and fixed flow adjustments. The modelled
volumes of the 2000 and 2001 events now match the measured volumes very
well (Figure 4.13).
Figure 4.14 shows that the improved representation of slow processes and
monthly and seasonal volumes, has also resulted in an improved water level
modelling of the Rijnland storage basin. The 2000 event is now calibrated
more accurately (Figure 4.14) as compared to the first model (Figure 4.5).
The results for normal flow periods remain the same (Figure 4.15).
1.20E+08
modelled
Pumped volume (m3)
1.00E+08
reported
Error total modelled
volume 2002:
0.67%
8.00E+07
6.00E+07
4.00E+07
2.00E+07
0.00E+00
Jan-02
Feb-02 Mar-02
Apr-02 May-02 Jun-02
Jul-02
Aug-02 Sep-02 Oct-02 Nov-02 Dec-02
Figure 4.12 Monthly pumped discharge volume from Rijnland storage basin in 2002
of the final model. Note the improvements in summer months and October and
November compared to Figure 4.7. Note also the accurately modelled total yearly
volume (0.7% error), while only the total volume over 6-year simulation (19972002) was calibrated.
1.40E+08
1.40E+08
1.20E+08
1.20E+08
1.00E+08
8.00E+07
6.00E+07
4.00E+07
Pumping modelled
Pumping measured
2.00E+07
0.00E+00
Pumped volume (m3)
Pumped volume (m3)
Case study 1 - Rijnland Water System
99
1.00E+08
8.00E+07
6.00E+07
4.00E+07
Pumping modelled
Pumping measured
2.00E+07
0.00E+00
28/09/2001
26/09/2001
24/09/2001
22/09/2001
20/09/2001
18/09/2001
16/09/2001
20/11/2000
18/11/2000
16/11/2000
14/11/2000
12/11/2000
10/11/2000
08/11/2000
06/11/2000
04/11/2000
Figure 4.13 Cumulative pump discharge volume for events in the wet winter season
in 2000(a) and 2001(b) after external modelling of unknown processes and
calibration.
-0.30
* 0.0
-0.35
-2.5
-0.45
-0.50
-5.0
-0.55
-0.60
-7.5
Precipitation (mm/h)
Water level (m+ref)
-0.40
-0.65
-0.70
-10.0
27-10-2000
4-11-2000
12-11-2000
20-11-2000
28-11-2000
Rijnland modelled: Water level (m+ref)
* Precipitation (mm/h) (-)
Rijnland modelled: Rural groundwater level (m+ref)
Rijnland measured: Water level (m+ref)
-0.50
* 0.0
-0.55
-2.5
-0.60
-5.0
-0.65
-7.5
Precipitation (mm/h)
Water level (m+ref)
Figure 4.14 Calibration of the event of November 2000, after the unknown
processes had been included as external data driven models. The modelling of the
peak has improved considerably with respect to the first model (Figure 4.5)
-0.70
-10.0
8-2-2000
16-2-2000
24-2-2000
3-3-2000
11-3-2000
Rijnland modelled: Water level (m+ref)
* Precipitation (mm/h) (-)
Rijnland modelled: Rural groundwater level (m+ref)
Rijnland measured: Water level (m+ref)
Figure 4.15 Model results for a normal flow period in February 2000, after the
unknown processes had been included as external data driven models. There are not
many differences with the first model (Figure 4.6)
100
Anticipatory Water Management
While the improvements with respect to the first model, after inclusion of the
sine function and re-calibration, are large and clear, the question remains
what the benefits are of modelling the unknown processes (in this case with
a sine function) compared to only correcting the volumes by applying a
fixed, constant flow correction. Comparative analysis shows that although
the use of a fixed flow correction does not result in consistantly large errors,
the overall accuracy is considerably less than the model with the sine
function. This is illustrated by the Nash-Sutcliffe (1970) coefficients of the
two models, derived for the top 20% of the inflow data, for different time
intervals (Figure 4.16).
1
Nash-Sutcliffe
0.95
0.9
Model with sine-function
for measured and
modelled inflow peaks
0.85
Model without sinefunction for measured and
modelled inflow peaks
0.8
0.75
0
1
2
3
4
5
6
Time step (days)
7
8
9
10
Figure 4.16 Nash-Sutcliffe coefficients for the Rijnland model, with the sine
function and with a constant flow correction, for different time intervals.
The gain in accuracy is larger for smaller time intervals, showing that the
model with the sine function has more accurate timing. Not using the sine
function may result in an under-estimation of some of the peaks. The model
with a fixed correction results in lowering the ground water table too far
during the dry season, and therefore may forecast the first peak too late and
underestimate it. It must be noted that the sensitivity of damage due to too
high and too low water levels in this case study area is very high. For critical
events 5 centimeters difference in water level or a couple of hours delay can
be the difference between little to no problems and serious damage. The
more extreme the event the damage costs increase exponentially (millions of
euros).
Validation of the final model
The final model has been validated for the years 2003 and 2004. Figure 4.17
shows that the cumulative pumped discharge volume from Rijnland is now
modelled very well for the calibration period (compared with the first model
Case study 1 - Rijnland Water System
101
result of Figure 4.8) and remains accurate during the validation in 2003 and
2004. The modelled volumes of two validation events match the measured
volumes well (Figure 4.18). Note that small differences here can be the
cause of expert based deviation from the operational routine by water
managers, which shows up in the measured discharge volume.
Validation results are particularly successful, when considering that 2003
was an extremely dry year in which exceptional control measures have been
taken. While part of these measures are included in the input data, it is
6.00E+09
calibration
validation
Pumped volume (m3)
5.00E+09
4.00E+09
3.00E+09
2.00E+09
Pumping modelled
1.00E+09
Pumping measured
0.00E+00
Jan 05
Jan 04
Jan 03
Jan 02
Jan 01
Jan 00
Jan 99
Jan 98
Jan 97
Figure 4.17. Cumulative pump discharge volume from the Rijnland storage basin
during calibration and validation. Note that the modelled cumulative volume now
matches very well compared to the first model (Figure 4.8) and that the
development of cumulative volume remains accurate during the validation period of
2003 and 2004.
6.00E+07
Pumping modelled
5.00E+07
Pumped volume (m3)
Pumped volume (m3)
6.00E+07
Pumping measured
4.00E+07
3.00E+07
2.00E+07
1.00E+07
5.00E+07
Pumping modelled
4.00E+07
Pumping measured
3.00E+07
2.00E+07
1.00E+07
0.00E+00
0.00E+00
30/12/2003
29/12/2003
28/12/2003
27/12/2003
26/12/2003
25/12/2003
24/12/2003
23/12/2003
05/01/2003
04/01/2003
03/01/2003
02/01/2003
01/01/2003
31/12/2002
30/12/2002
29/12/2002
28/12/2002
Figure 4.18. Cumulative discharge volume for validation events in January and
December 2003.
102
Anticipatory Water Management
expected that the validation results would have been even better if all data
from exceptional inlets would have been available.
These results show that the model can be used for operational decision
support for warnings of peak storage basin water level events, and for
providing predictions of discharge volume.
4.4.8 Discussion
Parameter sensitivity or stability analysis is considered here as part of the
calibration process (Section 3.4.2, Figure 3.7), which is sufficient for
applications of forecasting and operational decision support. In applications
in which exploring hydrological processes or transferability of model and
parameters to other (ungauged) water systems, are the main objectives,
parameter stability should be explicitly mentioned as a requirement (next to
accuracy, Figure 3.7) in the framework (Nash and Sutcliffe, 1970).
The model developed for the Rijnland case study is not aimed at
representing the best physically based model. It is a conceptual model with
many calibration parameters (>25). Data driven model components have
been added to compensate for missing information about physical processes
in the Rijnland water system. The model is aimed at providing an operational
decision support tool, which reliably simulates the water system and the
control of the regulating structures. Confidence in the reliability of the model
is based on the sound system behaviour, analysed from multiple measurable
system variables (e.g. water level and discharge time series) for a range of
time scales (e.g. single events, months, and years).
In operational support of water system control there is often the possibility to
include data assimilation schemes in the modelling process. In the Rijnland
case, in operational use, the only measured data that will be used is
precipitation, which results in model updating, not in data assimilation. Data
assimilation with water levels is not done, because the Rijnland system is a
fast responding system, which can be controlled very well within the target
range during normal flow periods. Therefore, during these periods, for the
predictions of inflow and decision support for the coming control actions,
assimilating water levels will not have much added value. In the case of
extreme events, the modelled water level and measured water level may
deviate significantly, but then the control strategy is already clear and
effective (full discharge through all structures), so assimilation of water
levels has little added value. These considerations about the limited scope of
water level data assimilation can be valid for many fast responding
controlled water systems. Ground- water level measurements would be
valuable to improve the model's initial state by data assimilation, however
these measurements are currently not operationally available to Rijnland.
Case study 1 - Rijnland Water System
103
The challenges described, and the proposed modelling framework are
relevant for many operationally controlled water systems in the Netherlands,
because of the historic development of the use of models. Until recently the
focus was on water level simulations for design and scenario analysis, in
which case the results of the first model could well be satisfactory. Now
models are more and more used for prediction and decision support for
control structure operation, and water boards have to consider whether the
existing model performance is still sufficient and whether it is worth the
effort of solving unresolved issues, such as long term volume balances. The
presented framework for water system control modelling (Figure 3.7) in
many cases can be implemented within a reasonable amount of time (e.g.
only one additional iteration). Using this approach, solving or at least
modelling the unresolved long-term issues, does not only increase the
understanding of the water system, but has also shown how to improve the
short term water level and discharge predictions that are so important for
operational decision support.
4.5 Ensemble forecasts verification
To asses the reliability of the forecasting product for use in AWM (step 4 in
the AWM framework, Figure 3.12), a verification analysis of 7.5 years has
been performed with an ensemble precipitation archive and water level
hindcasts.
The archive and hindcasts were analysed for the period from 25 April 1997
to 31 August 2004.
4.5.1 Precipitation ensemble forecasts archive
The precipitation forecasts are the ECMWF EPS precipitation forecasts for
location De Bilt in the Netherlands. The 6-hour precipitation amounts were
accumulated according to the different precipitation thresholds used in the
experiments. For example, the first experiment compares the forecasts and
the measured precipitation for a 15 mm day-1 threshold. Therefore, for each
6-hour time step, the forecasted precipitation for the past 24 hours was
accumulated to give forecasted precipitation per day.
4.5.2 Water level hindcasts
For the water level hindcasts an Aquarius water-system control model was
used. The water level hindcasts were determined by feeding the ECMWF
EPS precipitation forecasts into the water-system control model. The model
was applied deterministically, without accounting for additional sources of
104
Anticipatory Water Management
uncertainty. Real-time forecasting was simulated using a spin-up period of
30 days with area averaged measured precipitation. Following this spin-up
period, the 50 ECMWF EPS precipitation members were fed to the
deterministic Aquarius model to generate the ensemble forecast of water
levels up to 10 days ahead.
Because the model contains the routine operational strategy, these forecasts
show when the current control strategy is not sufficient to prevent high water
levels, and therefore, when anticipatory control is needed.
4.5.3 Event based verification for water managers
An analysis method was chosen that best fits the needs of the operational
water managers. An analysis of the precipitation and water level hindcasts
for a period of 7.5 years was done to enable the water managers to verify the
analysis steps and results according to their own experience and measured
data. A verification tool was developed in which the user can define the
variables of a threshold based decision rule for EPS, being event threshold
(precipitation or water level threshold), forecast horizon and probability
threshold (minimum forecasted probability that the event threshold will be
exceeded).
The main concern of the Water Board is to limit damage due to floods. The
water managers want to know how many critical events, which can be
handled only by anticipatory control, will be forecasted by the system (hits).
Since unnecessary flood-control actions, such as water release and pumping,
may cause damage, the Water Board is also interested in the number of false
alarms.
The decision rule defines the event threshold (precipitation or water level),
the forecast horizon and the probability threshold. For every evaluated
combination of forecast horizon and probability threshold, the corresponding
forecasted value is determined using the percentile function (probability
threshold 0.1 corresponds to the 90th percentile of the ensemble members).
The forecasted value is compared with the event threshold. When the
forecasted value exceeds the event threshold, the date and time are marked
as a forecasted event.
The comparison between measured and forecasted events was done for the
time at which a critical event begins. The forecast of the beginning of an
event is important to allow for effective anticipatory control actions. The
forecasted time at which the event begins should be within a predefined
range (e.g. one day) of the actual beginning of the event. For every measured
critical event, it was determined whether it was forecasted (hit) or missed
Case study 1 - Rijnland Water System
105
(missed event). For every forecasted event, it was determined whether the
event actually occurred (hit) or not (false alarm).
4.5.4 Precipitation and water level thresholds
The numbers of hits, missed events and false alarms over a period of 7.5
years were determined for three different precipitation thresholds and one
water level threshold. For each precipitation or water level threshold,
forecast horizons up to nine days and the full range of probability thresholds
were analyzed.
4.5.5 Presently used precipitation threshold for
anticipatory pumping
The Rijnland Water Board currently applies a precautionary threshold-based
decision rule, with a precipitation threshold of 15 mm day-1 and a 1-day
forecast horizon. In order to test the possibility of extending the forecast
horizon from one day to three days or more, the 7.5 years of ECMWF EPS
forecasts and measured precipitation were compared for the 15 mm day-1
precipitation threshold. Eighty-five events exceeding this threshold were
identified (the sample climatology is 0.03). Of these 85 measured events, 78
could have been anticipated using a forecast horizon of three days and taking
the highest of the 50 forecasted precipitation values as probability threshold.
This anticipation of events would have been done at the expense of 352 false
alarms. For a period of 7.5 years, this means that on average once a week
there would be an alarm and approximately five out of six of these alarms
would be false. Fewer false alarms can be achieved by applying decision
rules with higher probability thresholds, but this also reduces the number of
hits.
Figure 4.19 shows this relationship between the decision rule and the
number of hits and false alarms, for the fixed precipitation threshold of 15
mm day-1 and a varying forecast horizon and probability threshold. It shows
that the number of hits decreases with increasing probability threshold. For
probability thresholds greater than 0.04, the number of hits also decreases
with increasing forecast horizon. The number of false alarms decreases with
increasing probability threshold but increases with increasing forecast
horizon for probability thresholds up to 0.07. For probability thresholds
greater than 0.10 the number of false alarms decreases with increasing
forecast horizon, indicating that for long forecast horizons the system is not
forecasting the event with high probability. Figure 4.19 also shows that the
decrease in the number of hits and false alarms with increasing probability
threshold increases for longer forecast horizons.
106
Anticipatory Water Management
Figure 4.19 Contours of number of hits (a) and number of false alarms (b) of the
ECMWF EPS precipitation forecasts for 85 precipitation events in the Rijnland
water system of 15 mm day-1 or more. (ECMWF EPS precipitation for location De
Bilt, from 25 April 1997 to 31 August 2004)
The benefit of Figure 4.19 is that the Water Board can see the performance
of a range of possible decision rules at a single glance. The data behind
Figure 4.19 is also directly accessible with the verification tool, enabling
further analyses of specific combinations of forecast horizon and probability
threshold. Figure 4.20a shows the number of false alarms and missed events,
together with the forecasts that were too early, for the 3-day forecast horizon
and a range of probability thresholds. It can be seen that for the lowest
probability thresholds some events were forecasted too early. This indicates
that when looking only at the maximum ensemble members the forecast may
overestimate the amount of precipitation. Figure 4.20b shows the actual
forecasts for a 3-day forecast horizon and a 0.04 probability threshold (96th
percentile). This combination results in good forecasts of dry weather
periods. The precipitation events also show good agreement, but most of the
forecasts overestimate the precipitation. This can result in false alarms, as is
the case for 10 September 2003 (Figure 4.20b).
Case study 1 - Rijnland Water System
107
(a)
500
450
Number of forecasts
400
Precipitation threshold: 15 mm day -1
Forecast horizon: 3 days
False alarm
Missed
Hit, but too early
350
Hit
300
250
200
150
100
50
0
0.00 0.01 0.02 0.04 0.06 0.08 0.10 0.15 0.20 0.30 0.40 0.50 0.65 0.80 1.00
Probability threshold (-)
(b)
35
Forecast horizon: 3 days
Precipitation (mm day -1)
30 Probability threshold: 0.04
Measured precipitation
Forecast precipitation
Precipitation threshold
25
20
15
10
5
0
03 Sep 2003
13 Sep 2003
23 Sep 2003
03 Oct 2003
13 Oct 2003
Figure 4.20 Detailed analyses of performance of threshold-based decision rules with
ECMWF EPS precipitation forecasts. (a) Number of hits, events that have been
forecasted too early, missed events and false alarms for a 15 mm day-1 precipitation
threshold and a 3-day forecast horizon. (b) Measured daily precipitation and
forecasted daily precipitation for a 3-day forecast horizon and a probability
threshold of 0.04 (96th percentile).
The number of hits of 15 mm day-1 events may be good, but the number of
false alarms is high. Furthermore, the Water Board knows that in the
analysis period only a few critical events occurred, not 85 as indicated, with
the precipitation threshold of 15 mm day-1. The Water Board uses this low
event threshold as a precautionary measure to account for the uncertainty of
the weather forecast. In reality an operational water manager takes into
account the measured precipitation of the previous days and the present
water levels, because the initial conditions of the water system determine
whether an additional 15 mm will cause flooding problems or not. There are
two ways to account for initial conditions. One is to determine precipitation
thresholds over a number of days (e.g. 30 mm per 3 days). The other is to
108
Anticipatory Water Management
look at the water level hindcasts as prepared with the water-system control
model and to use a threshold for forecasted water level.
4.5.6 3-Day accumulated precipitation threshold for
selected events
Measurements of precipitation, water levels and pumping discharge were
analyzed to identify critical events (Figure 3.12, step 1). Looking only at
precipitation is insufficient because initial conditions and pumping strategy
determine whether a certain amount of precipitation results in a critical
situation. Looking only at water levels would be the most logical thing to do,
since the Water Board knows that flooding problems start occurring at
certain water levels. However, the water level is strongly influenced by
variable pumping strategies. For instance, water levels may rise during the
night when manually operated pumping stations are preferably not used.
This could lead to relatively high water levels in the morning that are easily
reduced once the pumping stations are put into action. To avoid this type of
event being characterized as critical, events were selected that resulted in
high water levels for at least 12 hours despite continuous pumping at
maximum capacity. In this way, nine critical events were identified. All
further verification analysis was performed with these nine events.
To analyze the effect of accumulating precipitation forecasts over time, 3day accumulated precipitation thresholds were determined. An analysis of
the measured precipitation for the nine selected critical events showed that
the minimum 3-day accumulated precipitation for these events is 40 mm in
winter and 45 mm in summer. For higher precipitation thresholds at least
two of the nine selected events would not be recognized according to the
measured precipitation.
To compare the results of the precipitation forecasts with the water level
forecasts, a water level threshold was determined for the same nine critical
events. Water levels that were modelled using the measured precipitation as
input, were compared with the selected events. A water level threshold of
-0.57 m+Ref during winter and -0.55 m+Ref during summer would help to
identify the nine selected events.
Using these precipitation thresholds and water level thresholds, the
performance of the full range of threshold-based decision rules for ensemble
forecasts was determined as in Section 4.5.5. Of the nine selected events,
seven could have been forecasted using the precipitation thresholds, and six
could have been forecasted using the water level thresholds. Figure 4.21a
and Figure 4.21b show that fewer events are successfully forecasted for the
short forecast horizons (three and four days ahead). This could show that the
ECMWF EPS forecasting system has been optimized to provide good
Case study 1 - Rijnland Water System
109
probabilistic forecasts for three to five days. During the first two days, the
disturbances in the initial conditions have not grown enough to present the
full spread of possible events. 3-Day accumulated values of the 3- and 4-day
forecast horizon include the first two days and may therefore underestimate
the amount of precipitation. Figure 4.21c shows that with the precipitation
threshold of 40 mm per 3 days in winter and 45 mm per 3 days in summer,
the number of false alarms is decreased substantially compared to the
threshold of 15 mm day-1 (Figure 4.19b). Figure 4.21d shows that when
using the ensemble water level forecasts instead of the precipitation
forecasts, even fewer false alarms result. For example, for a 6-day forecast
horizon and a probability threshold of 0.1, Figure 4.21c reads close to 20
false alarms, while Figure 4.21d reads less than 5 false alarms.
Figure 4.21 Comparison of performance of decision rules based on precipitation
forecasts and water level forecasts. [Left] Contours of number of hits (a) and false
alarms (c) with ensemble precipitation forecasts for nine selected events. Winter
precipitation threshold: 40 mm per 3 days. Summer precipitation threshold: 45 mm
per 3 days. [Right] Contours of number of hits (b) and false alarms (d) with
ensemble water level forecasts for nine selected events. Winter water level
threshold: -0.57 m+Ref for 12 hours. Summer water level threshold: -0.55 m+Ref
for 12 hours.
4.5.7 5-Day accumulated precipitation threshold for
selected events
To analyze further the effect of accumulating precipitation forecasts over
time, a precipitation threshold for 5-day accumulated precipitation was used.
110
Anticipatory Water Management
When compared with the measured precipitation, a threshold of 65 mm per 5
days best identifies the nine selected critical events.
When using a precipitation threshold of 65 mm per 5 days, all of the nine
events could have been forecasted (8-day forecast horizon, highest forecast
value as probability threshold), which is more than the six events that were
forecasted using the water level forecasts. The decay of number of hits with
increasing probability threshold is stronger for water level forecasts than
precipitation forecasts (Figure 4.22a,b). Figure 4.22c shows that the number
of false alarms is further reduced compared to the 40 mm per 3 days and 45
mm per 3 days thresholds (Figure 4.21c). The number of false alarms with
precipitation forecasts is now comparable to the number of false alarms with
water level forecasts. For example, both Figure 4.22c and Figure 4.22d show
a maximum of approximately 150 false alarms. Further analyses of the
forecasts showed that when using the precipitation threshold, more events
were forecasted too early compared with the water level thresholds. This
indicates that the timing of the forecasts is more accurate when using water
level forecasts instead of precipitation forecasts.
Figure 4.22 Comparison of performance of decision rules based on precipitation
forecasts and water level forecasts. [Left] Contours of number of hits (a) and false
alarms (c) with ensemble precipitation forecasts for nine selected events. Winter
precipitation threshold: 65 mm per 5 days. Summer precipitation threshold: 65 mm
per 5 days. [Right] Contours of number of hits (b) and false alarms (d) with
ensemble water level forecasts for nine selected events. Winter water level
threshold: -0.57 m+Ref for 12 hours. Summer water level threshold: -0.55 m+Ref
for 12 hours.
Case study 1 - Rijnland Water System
111
Forecast horizons between 5 and 7 days seem to perform best for forecasting
the nine measured critical events. Relative Operating Characteristic (ROC) diagrams can be used to compare further the performance of these forecast
horizons and the precipitation and water level forecasts. For each probability
threshold the false alarm rate (probability of false detection, e.g. number of
false alarms divided by the number of non-events) and hit rate (number of
hits divided by the number of events) are plotted against each other. The
points for each probability threshold can be connected to form a curve. The
larger the area under the curve (ROC-area) the better the forecast skill
(Atger, 2001).
The difficulty in applying the ROC-diagram for a few critical events is that
the curves tend to be aligned with the y-axis of the graph, because of the
high number of correctly forecasted non-events (an event that is not
observed and not forecasted), making the lowest probability threshold
decisive for the ROC-area. This problem is reduced in the event-based
approach adopted here, by setting the duration of non-events to the duration
of false alarms (instead of one day or one forecast time step, as is often
applied for the non-event). Figure 4.23a shows that for the precipitation
forecasts the 6-day forecast horizon performs slightly better than the 5- and
7-day forecast horizon. The ROC-curves of the water level forecasts for 5, 6,
and 7-day forecast horizons are almost the same. Figure 4.23b compares the
6-day precipitation forecasts with the 6-day water level forecasts and
confirms that the precipitation forecasts perform slightly better.
When applying day-by-day verification the areas under the ROC-curves
(ROC-area) are slightly smaller than in Figure 4.23. The 6-day precipitation
forecast ROC-area would be 0.74 instead of 0.86 and the 6-day water level
forecast ROC-area would be 0.75 instead of 0.80. These differences are
caused by one long precipitation event in 2000 that lasted 4 days and was not
forecasted. The contingency table of the 6-day precipitation forecasts (Table
4.1) shows that even the lowest probability thresholds are still higher than
the sample climatology of 0.006 (9/1464) and that the decision rule based on
one ensemble member exceeding the threshold, leads to many false alarms.
Yet, these low probability thresholds with the highest number of hits, are of
interest to the Water Board, as its primary concern is to identify critical
events.
4.5.8 Discussion
The results show that for the case study of Rijnland Water Board in the
Netherlands, the ECMWF EPS precipitation forecasts can be used in flood
control to forecast critical events for which anticipatory control actions are
needed. The analysis of the different decision rules shows that low
probability thresholds (<0.05) should be used to identify critical events.
112
Anticipatory Water Management
Forecast horizons between five and seven days seem to perform best for
forecasting the nine measured critical events. The most important finding
from the present case study is that it is better to use high event thresholds
that identify only the truly critical events than to choose low thresholds
based on precautionary principles. A thorough event analysis is needed to
define these thresholds. The results show that a factor of two to three in the
reduction of false alarms can be reached while maintaining the same number
of hits (e.g. from 300 false alarms in Figure 4.21c to 150 false alarms in
Figure 4.22c).
(a)
(b)
1.0
1
Event based ROC-diagram
6-day forecast horizon
0.9
0.8
0.8
0.7
0.7
0.6
0.6
Hit rate
Hit rate
0.9
Event based ROC-diagram
65mm/5d
0.5
0.4
0.5
0.4
0.3
5-day
0.3
0.2
6-day
0.2
0.1
7-day
0.1
0.0
Precipitation
forecasts
Water level
forecasts
0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
False alarm rate
False alarm rate
Figure 4.23 Event based ROC-diagrams of ensemble precipitation (a) for 5, 6, and 7
days forecast horizons. For comparison the ROC curves of the 6-day forecast
horizon water level forecasts and precipitation forecasts for 65 mm/5 days have been
plotted (b). The curves show the relationship between hit rate and false alarm rate
for different probability thresholds. The lowest probability threshold is the upper
right end of the curves, for the highest probability thresholds the curves reach the
origin (no hits and no false alarms).
Table 4.1 Event based contingency table for precipitation forecasts with a 6-day
forecast horizon.
Case study 1 - Rijnland Water System
113
While at first it seems that the use of a water-system control model to
translate the precipitation forecasts into water level forecasts results in a
reduction of false alarms, this case study demonstrates that a good
precipitation threshold can be defined, which results in a similar reduction of
false alarms. The precipitation forecasts even forecasted three more critical
events than the water level forecasts. However, the timing of the critical
events is forecasted better using the water level forecasts.
For practical use, a water-system control model can be used to reduce the
risk of false alarms compared to using precipitation forecasts directly. If it is
decided to start anticipatory control actions, these can be modelled with the
water-system control model. The model shows when sufficient measures
have been taken (e.g. extra storage created) and anticipatory control actions
can be stopped. When using only the precipitation forecasts, anticipatory
control actions will be continued as long as the forecasted precipitation is
exceeding the threshold, despite the extra storage that has been created.
The analysis was performed to show how many of the critical events that
occurred during the full analysis period could have been identified by the
ensemble forecasting systems. Therefore, the effect of seasonal differences
and changes due to development of the ECMWF EPS system are not shown.
Ideally, for any up-grade of operational forecasting models, a new archive of
hindcasts would be created to allow for end-users to adjust their decision
rules accordingly.
The presented research applied decision rules where a forecast horizon is
fixed and this is the only horizon to look at from one decision point to
another. When looking at consecutive ECMWF EPS precipitation forecasts
they are not always consistent in time. This means that the forecast for day i
may show an extreme event coming up, while the forecast for day i+1 shows
no event at all. Therefore, dynamic decision rules that look at a range of
forecast horizons, e.g. three to eight days, may increase the number of hits.
In our particular case study, using this range of forecast horizons enables the
water level forecasts to identify eight of the nine selected events. Even the
last event is recognized by the forecasting system but the beginning of the
event is forecasted just over 24 hours too late. Since this can be dangerous
for flood-control actions, it is not considered a good forecast. An event that
is forecasted too early does not have to be a problem if the adverse effects of
prolonged anticipatory control actions are limited. In this case study,
forecasts that are too early do not limit the effectiveness of the control
actions to reduce flooding problems. Therefore, such forecasts are
considered good forecasts in the long-term verification analysis.
The analysis on the basis of only nine critical events has already resulted in
clear directions of what are the most effective decision rules. However, a
longer analysis period with more critical events would allow for a more
114
Anticipatory Water Management
detailed definition of the optimal decision rule and with more confidence.
Therefore it is important that Water Boards and Meteorological offices
generate multi-year (decades) hindcast archives.
The focus of this forecast verification analysis has been on numbers of hits
and false alarms, because this enables the Water Boards to verify their
absolute requirements in flood protection. Next to these absolute
requirements, a cost-loss analysis (Richardson, 2000) is an important factor
in the decision of the Water Board whether or not, and how, to apply AWM.
Therefore, after new control strategies are developed (Section 4.6), the costbenefit analysis of these strategies is performed (Section 4.7).
4.6 Anticipatory water management strategy
development
The verification results showed a decline in the hit rate with increasing
forecast horizon and probability threshold. The water board is interested in
the highest possible safety against flooding that can be achieved using the
ensemble forecasts. Therefore, an example AWM strategy is evaluated
(Figure 3.12) in which anticipatory control is applied if 1 of the 50 ensemble
members exceeds the water level threshold. The water level threshold is put
to -0.57 m+Ref, because this is a water level that normally occurs only if the
inflow into the system exceeds the total pumping capacity of the system
(outflow). To provide enough lead-time to lower the storage basin water
level before the precipitation event occurs, a forecast horizon of 3 days is
applied. To maximise the probability of identifying the critical events, not
only the present day forecast (t = 0) is considered for a 3-day horizon, but
also the forecasts that were received in the days before (t = -1, -2, -3, -4
days) are considered. The forecast of yesterday (t = -1 day) is evaluated for
its 4-day horizon to match with the 3-day horizon of the present day forecast.
In the same way the 5-, 6-, and 7-day horizons of respectively the t= -2, -3,
and -4 day forecasts are used. Note that because of the risk avers approach
adopted here, no relative weights have been given to the older forecasts to
express a decline of forecast skill with increasing forecast horizon. If one of
the forecasts exceeds the water level threshold, the water level in the storage
basin is drawn down to a level between -0.65 and -0.70 m+Ref to create
extra storage at the beginning of the event.
The effect of this control strategy was simulated using the combined
ECMWF EPS precipitation forecasts and the AQUARIUS water system
control model for a period of heavy precipitation in September 1998. One of
the ensemble forecasts is shown in Figure 4.24 (an animation of all forecasts
for that period can be found in the supplementary pages of the electronic
Case study 1 - Rijnland Water System
115
journal paper Van Andel et al., 2008b). In addition to the forecasts, the red
line presents the measured water level, and the blue line presents the water
level as modelled with the measured precipitation as input. The September
1998 peak exceeded the threshold water level of -0.57 m+Ref for about two
days. Several members of the ensemble forecast exceed the -0.57 m+Ref
level for several days as well. Given the proposed strategy for anticipatory
control, therefore, the model will activate the pumping stations and draw
down the water level. The results are presented in Figure 4.25.
The red line presents the historically measured water level in the storage
basin. The blue line presents the simulated water level with anticipatory
lowering of water levels to -0.70 m+Ref on the basis of the EPS water level
forecasts. It can be seen that the extra storage prevents the water level from
exceeding -0.57 m+Ref. In the periods before and after the event, the
modelled and measured water levels are about the same. This shows that for
this period no false alarms occurred and no unnecessary lowering of the
water level in the storage basin was performed.
EPS water level forecast Rijnland (6-hours) of 11 September 1998 12:00 UT
Water level (m+Ref)
19 Sep 1998
17 Sep 1998
15 Sep 1998
13 Sep 1998
-0.45
11 Sep 1998
-0.4
-0.5
-0.55
-0.6
-0.65
Measured
Modelled
Ensemble members
-0.7
Figure 4.24 Ensemble water level prediction of 11 September 1998 for the Rijnland
water system on the basis of ECMWF EPS precipitation forecasts (Van Andel et al.,
2008b).
Anticipatory Water Management
* 0.0
-0.30
Water level (m+Ref)
-0.40
2.5
-0.50
5.0
-0.60
7.5
-0.70
-0.80
6-9-1998
10-9-1998
14-9-1998
18-9-1998
Modelled Rijnland storage basin: Water level (m+ref)
* Precipitation (mm/h)
Measured Rijnland storage basin: Water level (m+ref)
Precipitation (mm)
116
10.0
22-9-1998
Figure 4.25 Effect of modelled Anticipatory Water Management on water level
control in the Rijnland water system. The modelled peak is lower than the measured
peak, because in the model the water level is lowered before the precipitation event
occurs. Exceedance of the -0.57 m+Ref level is prevented (horizontal line).
4.7 Cost-benefit of selected AWM strategies
With an effective AWM strategy in place, the next step is to evaluate costs
and benefits of this strategy and compare these with costs and benefits of the
normal control strategy, which the water board is currently adopting.
4.7.1 Water level - damage function
The damage costs as functions of water levels have been estimated by the
water board. The (winter) target water level is -0.62 m+Ref (≈ 0.62 m below
sea level). If the level in the channelled storage basin starts rising, moisture
sensitive horticulture is affected first. Starting from -0.55 m+Ref complaints
from farmers and public start to come in to the water board and need to be
handled. At -0.50 m+Ref the water board starts up its flood emergency
preparedness plan. As water levels rise further the foundations of nearby
houses are affected and shipping is hindered. From -0.40 to -0.35 m+Ref the
water board issues "milling stops" to halt the pumping of excess water from
the low-lying land-reclamation areas to the storage basin. As a consequence
damage to crops in these areas increases. Because of the presence of both
horticulture and staple food crops, the seasonality of damage to crops is
expected to be limited and not explicitly taken into account. From -0.30
m+Ref onward, damage increases rapidly because of damage to crops,
houses and infrastructure. With levels of -0.10 m+Ref and higher the
stability of the reservoir embankments is affected and flood damage then
depends on the location of the first breaches.
The damage costs for too high water levels have been quantified based on
maximum cost estimates (in case of inundation) per hectare per land use type
(HKV, 2006). The maximum flood damage for Rijnland was then calculated
Case study 1 - Rijnland Water System
117
by multiplying the cost estimates with the land use areas in the catchment.
The costs related to the different storage basis levels, and duration of their
occurrence, have been estimated as percentages of this maximum flood
damage by the water board.
When water levels become too low and drop below -0.65 m+Ref,
embankments start to be affected, shipping is hindered, and houseboats are
damaged. Also complaints from the public come in. Below -0.75 m+Ref the
foundations of houses start to be affected and the damage cost increases with
further lowering of water levels and their longer duration (Figure 4.26).
The costs of damaged embankments have been quantified based on costs per
meter embankment for repairs (1000 euro/m). The length of affected
embankments and foundations for different (low) storage basin levels has
been estimated by the water board.
For both too high and too low storage basin levels costs of hindrance of
shipping have been estimated as a multiplication of the costs per ship per day
that is unable to navigate (500 euro/ship/day). Costs of handling of
complaints have been estimated as multiplication of labour costs of 500
euro/day).
These estimates are all present worth estimates.
Operational costs can be related to fuel costs for operating pumps (this can
be related to the number of pump operating hours), timing of the operation,
e.g. pumping at night for manual operated stations requires operators to work
at night at extra high tariffs and social costs, also pumping at night causes
noise disturbance to nearby households and nature, and number of switching
on and -off, which is related to the maintenance of the structures and
shortens the life cycle. The operation at night becomes less of a problem
since in the Rijnland case study all main pumping stations are going to be
operated automatically, also with the pumps increasingly becoming
electrically powered and as the housing of pumps is improving the noise
disturbance becomes less. There remains only the effect on house boats
which, if situated nearby the pumping station, will go up and down with the
water level, and costs of more frequent switching on and -off. Since,
according to the water board representatives, the present yearly operational
and maintenance costs are so low compared to other costs that they are not
part of yearly budget planning or strategic consideration, operational costs
are left out of the initial cost estimates and of the optimisation analysis.
Intangible costs like human casualties are not taken into account, because
these are not likely to occur. Also indirect costs like the reduction of
confidence in the warning system are not considered.
118
Anticipatory Water Management
The above described damage cost estimates lead to the water level - damage
- duration function shown in Figure 4.26. Note that non-linear increase of the
costs with prolonged duration of damaging water levels has been taken into
account.
45
Damage due to too high and too low water levels
40
35
30
25
Damage Million €
20
15
10
10
5
5
2
0
-0.45
1
-0.5
-0.55
-0.6
Storage basin water level
Duration (days)
0.5
-0.65
-0.7
0.25
-0.75
Figure 4.26 Water level-cost function Rijnland storage basin (Van Andel et al.,
2009b).
4.7.2 Inter-comparison of costs for selected strategies
The Aquarius water system control model of the Rijnland water system is
used together with the damage cost function to estimate the total damage
cost and flood damage cost of a selection of control strategies for a given
evaluation period (7.5 years).
Reference scenario
The reference scenario is the current control strategy. There are two options:
to use the measured water levels to estimate the flood and total costs over the
analysis period, or to use the modelled water levels with measured
precipitation as input. The approach is first to get a reliable and accurate
water system control model (Van Andel, 2009a), and based on the calibration
and validation results to agree on the suitability of the model. Then in the
subsequent cost-benefit analysis it is better to use the model results, because
human inconsistencies in applying the control strategy and measurement
errors in the water levels are filtered out. The simulated water levels are fed
Case study 1 - Rijnland Water System
119
to the water level-cost function to calculated the cumulative costs for the
simulation period.
Flood risk averse strategy
The flood risk averse strategy developed in Section 4.6 is evaluated as an
example Anticipatory Water Management strategy. The flood costs and total
costs over the analysis period are compared for the flood risk averse strategy
and the reference strategy (Figure 4.27 and Figure 4.28). The results show
that the flood damage is reduced considerably by applying the rule based,
risk averse AWM strategy, but that the total damage cost (of both too high
and too low water levels) becomes higher in this case. Note that presenting
the cumulative costs also highlights the important passed events. For
example, the strong increase in costs in the year 2000 Figure 4.27 points to
the November event that is well known with the water board as a critical
event.
Because already with this simple strategy flood damage reduction is
considerable, and because the increase of total costs may be seen by the
Water Board as investment in reducing flood risk, it was decided to go on to
the next phase in the AWM framework; namely, the optimisation of the
AWM strategy.
Flood damage
1800000
1600000
1400000
Cost (Euro)
1200000
1000000
800000
600000
Modelled with normal
control
400000
AWM flood risk averse
200000
May-05
Sep-04
Jan-04
May-03
Sep-02
Dec-01
Apr-01
Aug-00
Dec-99
Mar-99
Jul-98
Nov-97
Mar-97
0
Figure 4.27 Comparison of the flood damage cost estimate of the normal control
strategy with a flood risk averse AWM strategy
120
Anticipatory Water Management
Total damage
1800000
1600000
1400000
Cost (Euro)
1200000
1000000
800000
600000
Modelled with normal
control
400000
AWM flood risk averse
200000
May-05
Sep-04
Jan-04
May-03
Sep-02
Dec-01
Apr-01
Aug-00
Dec-99
Mar-99
Jul-98
Nov-97
Mar-97
0
Figure 4.28 Comparison of the total damage estimate of the normal control strategy
with a flood risk averse AWM strategy
4.8 Optimisation of Anticipatory Water
Management strategy
First, optimisation using deterministic perfect (synthetic) forecasts is applied
to find the maximum cost reduction by AWM. This also helps to identify
ranges of decision parameters for the optimisation with real forecasts in the
second stage.
4.8.1 Optimisation with perfect forecasts
The Rijnland case study generates a two-objective optimisation problem:
1. Minimise costs of too high water levels (flood damage cost)
2. Minimise costs of too low water levels (drought damage cost)
Both flood and drought damage costs have been expressed as a function of
the storage basin water level and its duration (Figure 4.26). When too high
water levels incur damage costs, this is attributed to 'Flood damage'. When
too low water levels incur damage costs, this is attributed to 'Drought
damage'.
Perfect water forecasts are prepared by taking the measured precipitation as
input to the simulation model of the Rijnland catchment. The lowering of the
storage basin level, in anticipation of the simulated inflows, is optimised
Case study 1 - Rijnland Water System
121
1.00E+09
1.00E+08
Damage costs (Euro)
1.00E+07
1.00E+06
1.00E+05
1.00E+04
1.00E+03
1.00E+02
Reactive control
1.00E+01
Optimal AWM with perfect forecast
1.00E+00
0
10
20
30
40
50
60
70
80
90
100 110
Return period (years)
Figure 4.29 Theoretical potential of total cost reduction by applying AWM with
perfect (synthetic) forecasts to the Rijnland water system for extreme events with
return periods between 10 and 100 years. Note the logarithmic scale of the cost-axis.
7.00E+05
Damage costs (Euro)
6.00E+05
Flood costs
Total costs
5.00E+05
4.00E+05
3.00E+05
2.00E+05
1.00E+05
0.00E+00
0
20
40
60
80
100
120
140
Control horizon (hrs)
Figure 4.30 Optimisation of control horizon by minimising estimated total damage
costs. Both the flood costs and the total costs become stable after 70 hrs. Expansion
of the control horizon beyond 70 hrs has no use. The analysis has been performed on
the basis of perfect (synthetic) forecasts for a 1/100 year event.
using the genetic algorithm NSGAII (Deb et al., 2002; Barreto et al., 2006).
The start-time and end-time of anticipation by pumping with full capacity
are optimised for extreme events with estimated return periods between 10
and 100 years (Hoes, 2007). The total costs of these extreme events, both in
case AWM is applied and when AWM is not applied (re-active control),
122
Anticipatory Water Management
have been presented in Figure 4.29. The potential of AWM increases with
the severity of flood events (note the logarithmic cost-axis). The cost
reduction from applying AWM with a perfect forecast for a 1/100 year
event, is estimated around 140 million Euro.
If for the same event the optimisation is repeated for different forecast
horizons, it is found that increasing the control horizon beyond three days is
not useful, because the damage costs will remain the same (Figure 4.30).
4.8.2 Optimisation with actual forecasts
The verification analysis and the cost-benefit analysis of the risk averse
AWM strategy showed that the high number of false alarms cause high total
damage costs (Figure 4.28). Therefore, a two-step approach might improve
the AWM strategy further. An early warning may be used to temporarily
switch to enhanced, more operationally demanding, real-time control, while
still staying within a low-cost water level range. Another warning rule
should be defined to decide on an AWM action, which temporarily lowers
the water level further to a level that induces damage costs, but is considered
necessary at the time to limit flood risk.
First a continuous improvement of real-time control of the Rijnland water
system is simulated by adjusting the switch-on and switch-off levels of the
pumping stations of the model, such that the storage basin levels are kept
more strictly between -0.65 and -0.60 m+Ref. This improvement leads to a
reduction of the estimated damage costs of control without AWM from
1.2*106 to 0.8*106 Euro.
Then ensemble water level predictions are used to decide on anticipatory
lowering of the storage basin water levels, on the basis of water level
thresholds, forecast horizons and probability thresholds. Also the time at
which the anticipatory lowering is started (anticipation time) and the level to
which the storage basin is lowered (anticipation level) are defined. The
resulting basin levels are simulated with Aquarius, and the water level-cost
function is used to estimate the flood and drougth costs between 1997 and
2004. The result of this analysis for one particular rule-based AWM strategy
corresponds to one data point in Figure 4.31. The genetic algorithm NSGAII
is used to optimise the AWM parameters. Probability threshold, forecast
horizons, warning levels, draw-down levels (anticipation levels) and
anticipation time are optimised for the wet seasons in the years 1997 to
2004, resulting in strategies with minimum flood or drought damage costs
(damage costs of too high and too low water levels). In the summer no
anticipation is applied, because the improved RTC would already prevent
summer flood damage). The flood and drought costs of all the sampled
AWM strategies are plotted in Figure 4.31. The strategies with the least
flood damage costs (closest to the y-axis) and with the least drought damage
Case study 1 - Rijnland Water System
123
1.00E+06
Flood damage cost (Euro)
Flood damage cost (Euro)
costs (closest to the x-axis) make up the Pareto front of optimal AWM
strategies.
AWM strategy
9.00E+05
8.00E+05
7.00E+05
5.50E+05
AWM strategy
5.40E+05
5.30E+05
6.00E+05
Strategy with least total costs
5.00E+05
0.00E+00
5.00E+05
1.00E+06
5.20E+05
0.00E+00
Drought damage costs (Euro)
5.00E+04
1.00E+05
Drought damage costs (Euro)
Figure 4.31 Estimatied drought (too low water levels) and flood (too high water
levels) damage costs for AWM strategies generated with NSGAII optimisation.
Costs are evaluated for the period between 1-9-1997 and 24-4-2004. The lower-left
corner of the Pareto front shows strategies with total cost reductions of around
2.4*105 Euro compared to strategies without anticipation (Van Andel et al., 2009b).
850000
Flood damage costs (Euro)
800000
AWM strategy
750000
700000
650000
600000
550000
500000
0
0.2
0.4
0.6
0.8
1
Probability threshold (-)
Figure 4.32 Estimated Flood damage costs versus Probability threshold for the 150
least total cost AWM strategies determined by NSGAII optimisation. Costs are
evaluated for the period between 1-9-1997 and 24-4-2004. The graph clearly shows
how flood damage reduces when low probability thresholds are applied. Meaning
that when only 1 or 2 ensemble forecast members are required to exceed the warning
level, then most critical events will be identified and the forthcoming damage
reduced by AWM strategy.
124
Anticipatory Water Management
It can be seen in Figure 4.31 that the maximum estimated flood damage is
8.1*105 Euro. This corresponds to the situation where no anticipation is
applied, hence no drought damage costs occur (upper, left points). When
AWM is applied the flood damage can be reduced with 2.8*105 Euro to a
minimum of 5.3*105 Euro (Figure 4.31). The strategy with the minimum
total costs of 5.7*105 Euro, shows that also the total damage costs can be
reduced by AWM with around 2.4*105 Euro (30%).
From the same analysis overviews can be generated to show the influence of
decision parameters on the objectives. In Figure 4.32 an example is given of
the influence on the flood damage of the probability threshold for deciding
on an anticipatory lowering of the storage basin water level. It shows that
only probability thresholds of 0.1 and lower can be used to reduce flood
damage costs on the basis of the ensemble water level forecasts applied in
this analysis.
Figure 4.31 and Figure 4.32 show how optimisation can help in selecting
decision variables in the rule based AWM strategies. For this case study, for
this analysis period, the least-cost strategy found consists of the following
decision parameter values:
Table 4.2 Optimal decision parameter values and minimum damage estimation
Decision parameters
Water level threshold A
Anticipation level A
anticipation time A
Water level threshold B
Anticipation level B
Anticipation time B
First forecast horizon
Last forecast horizon
Probability threshold
flood damage
drought damage
total damage
-0.57
-0.66
25
-0.47
-0.74
5
76
139
0.05
€532,000.00
€38,000.00
€570,000.00
The denotations "A" and "B" refer to the two-stage lowering of the water
level. First, on the basis of a low warning threshold (Water level threshold
A; -0.57 m+Ref) the level is lowered, pre-cautiously, to Anticipation level A
of -0.66 m+Ref. This level causes only limited damage costs. If water level
forecasts also exceed the warning threshold, Water level threshold B, then
the water level is lowered further to Anticipation level B, -0.74 m+Ref. The
forecast horizons included, range from 3 to 6 days. This is consistant with
the results of the verification analysis. The anticipation time A, however, is
only 1 day, while optimisation with perfect forecasts indicated that 3-days
would be needed for extreme events. The reason that this did not come back
Case study 1 - Rijnland Water System
125
in the optimisation with real forecasts is that none of the real forecasts
managed to predict the extreme event of November 2000. Hence, the AWM
strategy is optimised to the smaller events that could be well predicted. For
flood risk averse strategies it would therefore be needed to apply apply a
longer anticipation time. This shows that for the adoption of the AWM
strategy the results of the optimisation analysis should be handled with care.
4.9 Adoption of AWM in operational management
policy
AWM can be applied in Rijnland both for the reduction in the costs of too
high water levels, and of the total damage costs. The analysis period needs to
be expanded in order to generate more reliable optimal control strategies.
The rule-based AWM strategy is satisfactory for this case study.
Because the cost-water level function is for a large part estimated from
country average unit flood damage estimates and expert judgement from the
water board, the absolute cost estimates are to be considered as indicative.
Verification of the estimated costs is difficult, because most of the
components are hidden costs in the sense that they are usually not actually
determined and declared. Only for the extreme range of the water levels,
with dike breach or inundation of the polders, the direct damage costs are
analysed, e.g. for other Water Boards in the Netherlands in 1998 and 2000.
When monitored also the cost estimates for the less extreme water levels
could be validated and further detailed and improved.
As alternative to AWM, further improving the real-time control can be
investigated as well. The timing of the switching-on of pumping stations at
the beginning of the event is crucial. The water board could already reduce
the risk of flood damage by starting to pump earlier with all pumping
stations at full capacity, instead of a stepwise approach of starting up the
pumping stations. Because this strategy would lead to a more frequent
switching on and off of the pumps with short intervals, it may not be a
preferred strategy in terms of operational costs (electricity, maintenance
costs). It can, however, be used as a strategy in combination with flood risk
averse warnings. Even if a large part of the cost reduction can also be
achieved by further improving the RTC, AWM is still preferable because it
reduces the risk of flooding.
An additional incentive to somehow include medium-range ensemble
precipitation forecasts in the operational water management, is that water
authorities will increasingly be held accountable for having used all the
information available. In todays information society individual citizens,
when confronted with water damage, will look up the weather forecasts of
126
Anticipatory Water Management
the days before on the internet and ask the water authorities why they did not
anticipate the forecasted rain event.
If the water board decides to implement AWM, it is expected that the rule
based AWM will easily fit within the current legislative structure of
operational water management, because the Principal Water-board of
Rijnland has already started applying AWM on the basis of rainfall forecast
thresholds. A decision support system, including deterministic weather
forecasts, a rainfall runoff-model, and a decision model for control of the
pumping stations is already in place and operational. This would only have
to be expanded to include the real-time streaming of the ensemble forecasts
and the AWM strategy proposed. Then, while running the forecasts in
parallel (off-line), operational water managers can familiarise themselves
with AWM, the verification and cost-benefit analysis periods can be
continued, and the operational reliability can be assessed.
127
5
Case study 2 - Upper Blue Nile
5.1 Introduction
While the countries of the Nile may be mostly known for their droughts and
subsequent famines, the basin also faces frequent flooding problems. Mainly
the Blue Nile river in Ethiopia and Sudan, up to Khartoum, overflows its
banks (almost) every year. Despite this recurrent problem, few flood
forecasting and early warning systems are in place. This has partly been
attributed to the limited data exchange among the riparian countries before
the Nile Basin Initiative (NBI) was established. Now it is possible for
researchers within the relevant disciplines to function effectively in Nile
flood management projects. This case study focuses on flood forecasting and
early warning for two sub-catchments, Ribb and Gumara, of Lake Tana in
the Upper-Blue Nile in Ethiopia (Amare, 2008). These areas have been
identified as two of the target areas for the Flood Preparedness and Early
Warning project of the Eastern Nile Technical Regional Office (ENTRO).
Lake Tana can be considered as the source of the Blue Nile. The lake is
surrounded by sub-basins with a total area of 12000 km2. The surrounding
sub-basins are drained by several small streams and 11 major rivers that flow
into the lake. The eastern portion of the basin is drained by the Ribb and
Gumara rivers that account for 28% of the basin area (Kebede, 2006). These
two rivers flow to the lake passing through the flat fields of the Fogera flood
plain (Figure 5.1).
5.2 Problem description
Flooding is not new to Ethiopia. Floods have been occurring at different
places and times, with varying, but often with manageable or ‘tolerable’
severity. In recent years, however, the country has been threatened by more
extreme flooding and severe damage. Most of these flood disasters are
attributed to rivers that overflow or burst their banks and inundate
downstream flood plains, following torrential rains in the upstream
highlands, with duration of several days.
The flooding problems of Ribb and Gumara rivers are of similar nature. The
river flow increases from continuous rainfall on the upstream part of the
catchments and local rainfall on the flood plain. Areas in the Fogera flood
plain that are most at risk from flooding are located between these two
128
Anticipatory Water Management
rivers. During high floods, people have to live in chest-high water levels,
roads become impassable and communication between affected people gets
limited to swimming. Fogera is in an administrative district (a Woreda). The
Fogera Woreda comprises a land area of 1095 km2, and has a total
population of 243000 people (SMEC, 2006). Within Fogera, 6 or 7 subdistricts are particularly flood-prone. This amounts to approximately a
quarter of the Woreda land area.
In spite of the recurrent flood problem, the existing disaster management
mechanism is primarily aimed at strengthening rescue and relief
arrangements during and after major flood disasters. No decision support
systems and anticipatory management strategies to mitigate the flood
damage are present. Regional and national flood management authorities
(ENTRO) want to research the potential of flood forecasting and early
warning for mitigation measures.
Because of limited financial resources, as an additional requirement for the
forecasting system, it was to be composed of free and open source weather
forecasting and hydrological modelling products.
5.3 Data
5.3.1 Geographical data
The digital elevation model (DEM) data of the Shuttle Radar Topographic
Mission (SRTM) was used (SRTM, 2008). The DEM has a spatial resolution
of 90 by 90m at the equator.
A soil data set, following the FAO classification, and land use and land
cover datasets were obtained from the Ministry of Water Resources of
Ethiopia.
5.3.2 Meteorological data
The meteorological data was provided by the National Meteorological
Agency (NMA) of Ethiopia. Data from some of the stations, which were not
available at the NMA, were collected from the Bahirdar metrological office.
The locations of the meteorological stations is shown in Figure 5.1. A period
of seven year (2000-2006) was used for analysis.
The rainfall data is daily total rainfall. Most of the stations exhibit significant
gaps. Hamusit station is excluded from analysis because 24% of the data is
missing. The gaps for most of the stations lie in the rainy season of the year,
when it affects the results of hydrological simulations most. For example one
of the rainfall time series is shown below (Figure 5.2).
Case study 2 - Upper Blue Nile
µ
129
Legend
ribb Q station
$
1 gumara Q station
Metreological gauging
* stations
#
2
%
Lake Tana
Fogera Flood
plain
river
7
#
Gondar
Ribb watersheds
Gumara watersheds
Lake Tana
7
#
7
#
Addis Zemen
Yifag
Lake Tana
2
%
7
#
Wereta
1
$
7
#
Debre Tabor
7#
#
7wanzaye
Hamusit
7
#
Bahir Dar
Decimal Degrees
00.025
0.05 0.1 0.15 0.2
Figure 5.1 Ribb and Gumara catchments with the locations of hydro-metrological
gauging stations.
130
Anticipatory Water Management
80
70
60
50
40
30
20
10
0
06-05-01
06-09-01
05-09-01
06-01-01
05-01-01
05-05-01
04-09-01
04-01-01
04-05-01
03-09-01
03-05-01
02-09-01
03-01-01
02-05-01
02-01-01
01-09-01
01-01-01
01-05-01
00-09-01
00-01-01
00-05-01
99-09-01
99-01-01
99-05-01
98-05-01
miss
98-09-01
98-01-01
Rainfall(mm)
Rainfall
(mm)
AddisZemen_station
Addis
Zemen station
Figure 5.2 Data gaps of Addis Zemen gauging station, indicated by circles for long
periods of missing data and an arrow for a short period of missing data. In the
analysis only the data between 2000 and 2006 was used.
In addition, no rainfall station exists in the Gumara catchment, while
neighbouring stations exhibit significant data gaps in similar periods.
Therefore, satellite based rainfall estimates are used to fill the data gaps.
Tropical Rainfall Measurement Mission (TRMM) datasets are freely
available through the National Aeronautics and Space Agency (NASA,
2008). The datasets provide the opportunity to have rainfall estimates in
regions where conventional rainfall data are scarce (Kummerow, 2000). This
study makes use of the daily 0.25o x 0.25o TRMM and other rainfall data set
(3B42 V6) from 2000 to 2006 for four pixels covering the study area
(NASA, 2008).
Piche evaporation data (PET) from Bahirdar station is used. The monthly
data from this gauge has been correlated with long time series of PET data at
Gondar airport station to convert the PET evaporation to Potential
evaporation.
5.3.3 Streamflow data
Daily flow records of the Gumara river at the station near Bahirdar, and of
the Ribb river at the station near Addis Zemen (Figure 5.1), were obtained
from the Hydrology Department of the Ministry of Water Resources,
together with stage-discharge relationships and river cross-sections at the
gauging sites. The location of the gauging stations are shown in Figure 5.1.
Data was available from 1998 to 2006, which covers the analysis period
(2000-2006).
The data of the Ribb River was almost complete with only seven days
missing, except for the year 1998, where peak discharges are missing (Figure
5.3). The data for the Gumara river, however, has gaps, which are all
Case study 2 - Upper Blue Nile
131
observed to occur during the dry period of the year. Therefore, the recession
curve method is used to fill the gaps (Maidment, 1994). The complete time
series of both rivers are plotted in Figure 5.3. It can be observed from this
plot that the Gumara River, while having a smaller catchment area, has more
discharge in all seasons than the Ribb river.
Ribb
06-09-01
06-05-01
06-01-01
05-09-01
05-05-01
05-01-01
04-09-01
04-05-01
04-01-01
03-09-01
03-01-01
03-05-01
02-09-01
02-05-01
02-01-01
01-05-01
01-09-01
Gumara (m3/s)
350
01-01-01
300
0
00-09-01
250
50
00-05-01
200
100
00-01-01
150
150
99-09-01
200
99-05-01
100
99-01-01
50
250
98-09-01
0
300
98-05-01
350
98-01-01
Ribb (m3/s)
Ribb and Gumara river discharge
Gumara
Figure 5.3 Gumara and Ribb daily river discharge from 1998 to 2006. The discharge
data of the Ribb river in the wet season of 1998 seems too flat (circled), pointing to
measurement errors. The data between 2000 and 2006 was used for analysis.
5.4 Hydrological model
The Hydrologic Engineering Center-Hydrologic Modelling System (HECHMS) is a physically-based semi-distributed model (USACE, 2003). It is
designed to simulate the rainfall-runoff processes of dendrite watershed
systems. The software is freely available. It has been selected for use in this
study partly because it has been tested in Upper Blue Nile, and has resulted
in good performance (Bashar and Zaki, 2006).
5.4.1 Model set-up
The SRTM 90m DEM was used for catchment delineation. The HEC-HMS
Soil Moisture Accounting model was used to allow for continuous
simulation. For the direct run-off computation the Clark unit hydrograph
method was used. The linear reservoir was adopted for base flow calculation
methods, because this module is suitable with the soil moisture accounting
model (USACE, 2003). The Muskingum method is used for flood routing in
this study for the reason of data limitation to employ the conceptual
kinematic wave model.
132
Anticipatory Water Management
The Ribb and Gumara catchments have been modelled separately, each with
three sub-catchments (Figure 5.1). The area-average rainfall was estimated
by making use of the gauge weighting method. The first estimate of the
gauge weights was made by making use of the Thiessen polygon method.
The Thiessen polygons ware not used alone because of the scarce
distribution of the gauges, especially for Gumara catchment, where more
than 50% of the catchment lies outside of the Thiessen polygon. Therefore,
gauge weights were adjusted based on expert judgement.
5.4.2 Calibration and validation
In this study the daily streamflow data from 1 January 2000 to 31 December
2003 has been used for calibration and the period from 1-Jan-2003 to 31Dec-2005 for validation. First, manual calibration with visual inspection of
the measured and monitored streamflow data was performed to provide a
good estimate of the parameters. Then, automatic calibration was applied
using the Peak-Weighted Root Mean Square Error (PWRMSE) and Volume
Percent Error (VPE) objective functions. PWRMSE was selected as it gives
greater overall weight to errors near the peak discharge without significantly
affecting the VPE calibrated parameters. Both the default hard-constraints,
which limit the range of parameter values within reasonable physical
intervals, and soft-constrains on the basis of physical implications of the
parameters, were used to limit the range of possible values. The Univariate
Gradient search method was used.
Figure 5.4 and Figure 5.5 show a reasonable fit for the calibration of both the
Ribb and the Gumara model. The validation of the Gumara river (Figure 5.6
and Figure 5.8b) shows that the model over-predicts, but the trend is
reasonable and over-prediction for flood warning applications can be
considered positive from a flood risk averse approach. The validation of the
Ribb river (Figure 5.7 and Figure 5.8a), however, shows strong underestimation of streamflow for long periods of the wet seasons. The highest
peaks of the wet season are captured well.
Seasonality (wet and dry seasons), next to, for example, spatial distribution
of rainfall, and soil and land cover heterogeneity, may be an important
sources of error in the hydrological modelling. Developing a seasonal
parameterisation approach where each simulated year is divided into two
simulation periods (wet and dry seasons) and accordingly one parameter set
is obtained for each period could be a good step for improvement of the
model.
Case study 2 - Upper Blue Nile
133
400
350
Discharge m3/s
300
Measured
Modelled
250
200
150
100
50
0
19-04-2001 08-06-2001 28-07-2001 16-09-2001 05-11-2001 25-12-2001 13-02-2002
Figure 5.4 Gumara calibration result for 2001
120
100
Measured
Discharge m3/s
Modelled
80
60
40
20
0
19-04-2001 08-06-2001 28-07-2001 16-09-2001 05-11-2001 25-12-2001 13-02-2002
Figure 5.5 Ribb calibration for 2001
134
Anticipatory Water Management
400
350
Measured
Modelled
Discharge m3/s
300
250
200
150
100
50
0
08-04-2005 28-05-2005 17-07-2005 05-09-2005 25-10-2005 14-12-2005 02-02-2006
Figure 5.6 Gumara validation for 2005
120
Measured
100
Modelled
Discharge m3/s
80
60
40
20
0
08-04-2005 28-05-2005 17-07-2005 05-09-2005 25-10-2005 14-12-2005 02-02-2006
Measured (m3/s)
Figure 5.7 Ribb validation for 2005
Ribb validation
(a)
150
Gumara validation
(b)
400
y = 1.10x + 2.43
R2 = 0.72
100
y = 0.62x + 8.11
R2 = 0.72
300
200
50
100
0
0
0
50
100
Simulated (m3/s)
150
0
100
200
300
Simulated (m3/s)
Figure 5.8 Validation results: Ribb correlation (a) , Gumara correlation (b)
400
Case study 2 - Upper Blue Nile
135
5.5 Ensemble forecasts verification
5.5.1 Event selection
Threshold based decision rules can be used for issuing a flood warning. A
warning is issued whenever the forecasted flow or water level exceeds a
threshold. Thresholds for streamflow need to be related to actual flood
events. Flood thresholds for Gumara and Ribb rivers were computed by
making use of three different criteria (see also 3.2.1):
- Recorded flood damages (data from the study area)
- Flood damages and dates of occurrence from Dartmouth Flood
Observatory
- Bank-full discharge
Each of the three criteria are discussed in the following sections.
Recorded flood damages
The Ribb and Gumara critically high discharges were retrieved by referring
to flood damages recorded during the analysis period. Table 5.1 indicates
different levels of flood damage in the area. The peak floods in those years
are taken as first estimates of the different warning threshold levels.
Table 5.1 Recorded flood damages (DPCC, 2007)
Fiscal
year
Affected
land area
(ha)
Production
(kg)
Cost
estimate
(Eth Birr)
No. of
affected
Kebeles
2000/01
2001/02
2003/04
2005/06
1566
3697
1155
39
14,562
21,617
22,937
590
2,184,300
2,594,040
3,440,550
118,000
5
7
3
2
Peak floods of
the recorded
flood years
(m3/s)
Gumara Ribb
278
297
269
223
102
87
84
91
Satellite information
Dartmouth Flood Observatory is an international clearinghouse for GIS data
concerning flood inundation, mapped using satellite data. The observatory
uses remote sensing to detect, measure, and map major river floods.
Information on flood incidences, the dates of occurrence and the damages
from the flood can be retrieved free of charge (DFO, 2008). The information
for the case study area is presented in map (Figure 5.9) and tabular format
(Table 5.2).
It can be observed from the flood map (Figure 5.9) that, on the mentioned
dates, the flood extent in Fogera flood plain was the greatest of all the other
flood prone areas around Lake Tana. This is further checked with gauge flow
136
Anticipatory Water Management
records in the corresponding dates and it is found that those dates are
recorded with highest flood peaks in that year. A peak of 274 m3/s is
recorded on 15 August 2006 at Gumara river. The Ribb hydrograph also
indicates continuous peak flows from 90 m3/s to 99 m3/s within the
Dartmouth ranges of dates (13 to 18 August 2006).
Fogera
flood plain
Figure 5.9 Flood areas around Lake Tana sub basin (Aug 13-27, 2006) (DFO, 2008)
Table 5.2 Flood damage in lake Tana sub basin (DFO, 2008)
Areas
flooded
Ethiopia
in Blue
Nile,
Bereka,
Ribb,
Gumara
Cause of
flood
Heavy
rain
Date of flood
Flood damage
13-Aug to
27-Aug 2006
Ethiopia - 10,000 displaced around
Lake Tana, the source of the Blue
Nile River.
38,000 displaced by flooding in
Amhara region.
River bank-full discharges
One conservative measure of a “flooding flow” is the bank-full discharge.
This definition of “flooding” is physically based, but is considered
conservative as more than bank full flow is generally needed to cause
damage (Carpenter et al., 1999). The bank full discharges and cross-sections
at the gauging sites were taken as a reference for the lower warning
Case study 2 - Upper Blue Nile
137
threshold (Figure 5.10). Based on three criteria discussed, a low, medium
and high streamflow warning threshold were determined (Table 5.3).
Gumara river cross section at Gauging site
Bank full
Discharge = 96 m3/s
8
6
Elevation (m)
Elevation (m)
Ribb River cross section at the gauging site
4
2
0
0
10
20
30
40
Width (m)
50
Bank full
Discharge = 218 m3/s
8
6
4
2
0
60
0
10
20
30
40
50
Width (m)
Figure 5.10 River cross-sections of Ribb and Gumara
Table 5.3 Suggested warning threshold
Thresholds (m3/s)
/Rivers
Low (threshold 1)
(m3/s)
Medium (threshold 2)
(m3/s)
High (threshold 3)
(m3/s)
210
60
250
85
300
110
Gumara
Ribb
Threshold based flow comparison between Ribb and Gumara rivers
When comparing the peak flows of the Ribb and Gumara rivers it seems that
the hydrographs generally follow the same pattern (Figure 5.11). Therefore,
getting peak flow warnings for one of the rivers would contribute to flood
warnings for the Fogera flood plain. Because of the better performance of
the Gumara model and the greater discharges of the Gumara river, it is
assumed that the best results for flood warning will be based on the Gumara
predictions. Therefore, in the remainder of this case study focus is on the
Gumara river only.
measured Ribb
Figure 5.11 Gumara and Ribb river flows comparison.
06-09-01
06-05-01
06-01-01
05-05-01
measured Gumara
05-09-01
04-09-01
05-01-01
04-05-01
04-01-01
03-09-01
03-05-01
03-01-01
02-05-01
02-09-01
02-01-01
01-09-01
01-05-01
01-01-01
00-09-01
00-05-01
0
50
100
150
200
250
300
350
400
450
500
Gumara flow(m3/s)
180
160
140
120
100
80
60
40
20
0
00-01-01
Ribb flow(m3/s)
Gumara and Ribb River flows comparison
138
Anticipatory Water Management
5.5.2 Ensemble precipitation hindcasts
Five different rainfall forecast archives have been prepared for input to the
HEC-HMS model for the period of the years 2000-2006. The re-forecasting
(hindcasting) is done with daily rainfall forecasts with a 1 to 10 day forecast
horizon on the basis of:
1. No rainfall
2. Monthly mean rainfall
3. Minimum of the ensemble rainfall forecast
4. Mean of the ensemble rainfall forecast
5. Maximum of the ensemble rainfall forecast
As a reference forecast "no-precipitation" is taken. A first improved forecast
is prepared by taking the monthly average daily precipitation as input.
Monthly mean values are computed from the rainy seasons of the 7 years
analysis period (2000-2006).The third, fourth and fifth forecast methods are
the Min, Mean, and Max from the ensemble precipitation hindcast archive
from the National Centers for Environmental Prediction (NCEP) Global
Forecasting System (NCEP, 2008). The ensemble forecasts are freely
available on the internet for the entire globe in a grid size of 2.5o x 2.5o. Note
that this 'frozen version' of the GFS for hindcasting contains a different
model with lower spatial resolution than the currently operational GFS. The
12-hourly ensemble forecast from NCEP consists of fifteen members to a
forecast horizon of fifteen days.
5.5.3 Ensemble streamflow hindcasts
The five different rainfall forecast archives are used as input to the HECHMS model to produce streamflow hindcasts. For automatically performing
the re-forecasting the Hydrologic Engineering Centre Data Storage System
Utility Program (DSSUTL) is used. Resulting hindcasts for the wet season of
2001 are presented in Figure 5.12.
5.5.4 Verification analysis
The resulting stream flow forecasts are compared with flows simulated with
measured rainfall as input, by making use of three different verification
methods.
First, statistical analysis is used as an aid in screening the better ones from
the five different forecasts (zero, monthly mean, minimum EPS, mean EPS
and maximum EPS). Normalized Root Mean Square Error (NRMSE) and
correlation (R2) are used for comparison.
Case study 2 - Upper Blue Nile
139
Modeled with measured precip
28-Aug-01
21-Aug-01
14-Aug-01
7-Aug-01
31-Jul-01
24-Jul-01
17-Jul-01
10-Jul-01
3-Jul-01
26-Jun-01
19-Jun-01
5-Jun-01
12-Jun-01
Discharge (m3/s)
Forecast horizon 4 days
300
250
200
150
100
50
0
with Zero precip
With Monthly mean
Figure 5.12 Example streamflow hindcasts with a 4-day forecast horizon. After 4
days the underestimation by assuming no rainfall becomes clear. Assuming Monthly
mean rainfall shows a better comparison with the reference streamflow (simulated
by using measured rainfall as input), but streamflow peaks, particularly in the
beginning of the wet season, are underestimated.
Secondly flood warning verification is applied to compare forecasts of
discharges above thresholds in terms of number of hits, missed events and
false alarms (Van Andel et al., 2008a). Each of the measured discharge peaks
that exceed the threshold is considered as one event. If the peak stays above
the threshold for more than one day, this is still considered as only one
event. If the forecasted discharge also exceeds the threshold, then the
forecast is considered as hit. If the measured event is not forecasted, then it
is called as missed event. The allowable time lag between the forecasted and
measured events is taken as 2 days. The number of false alarms is the other
important criterion used in warning verification. If the forecasted discharge
exceeds the threshold when the measured discharge peak is below the
threshold, then the forecast is considered as false alarm. If the forecast
discharge stays above the threshold for more than one day, this is still
considered only one false alarm. Such incidences can be identified with
visual comparison of the measured and forecasted flows.
Visual inspection is the third important method of analysing forecast results.
The interpretation of the statistical and flood warning verification results
requires visual inspection.
5.5.5 Statistical verification
First the flow forecasts from each of the above discussed forecast methods
are analysed by NRMSE. The simulated flow hydrograph from measured
precipitation values is taken as reference when comparing with the different
140
Anticipatory Water Management
flow forecasts. The statistical analysis is conducted on a yearly basis. Results
for 2000 are shown in Figure 5.13. The NRMSE values show that the 'zero
precipitation forecast' and the 'max EPS' forecast perform worse, while the
other options do not show large differences, with the 'monthly mean'
precipitation forecast performing best. Correlation analysis for the 3-day
forecast horizons showed the same pattern: Monthly mean performs best (R2
= 0.79), while Min EPS (R2 = 0.66) and Mean EPS (R2 = 0.66) do not show
much difference.
0.3
NRMSE
1-day forecast horizon
0.25
2-day forecast horizon
3-day forecast horizon
0.2
0.15
0.1
0.05
0
zero precip
monthly mean
min EPS
mean EPS
max EPS
Figure 5.13 NRSME for 1, 2, and 3-day forecast horizons for the year 2000
It could be concluded from Figure 5.13 that the monthly mean, min EPS, and
mean EPS forecasts result in better estimates than the other forecasts.
However, this result alone cannot lead to a conclusion that these three
forecasts are good estimates for flood forecasting, because statistical
measures describe only general performance of models without special
consideration of peak errors. Therefore, the monthly mean, min EPS, and
mean EPS forecasts were further analysed by visual inspection.
5.5.6 Comparison by visual inspection
Forecasts with monthly mean precipitation and Min EPS forecasts as input
underestimating peaks (Figure 5.14, Figure 5.15). The forecasts with Mean
EPS precipitation as input are better catching the peaks (Figure 5.16). This
shows that while Monthly mean forecasts showed the best performance with
correlation analysis, the Mean EPS forecasts perform better with visual
inspection. Capturing the peaks is of course crucial in flood forecasting
applications. Therefore, the Mean EPS were analysed further for their
applicability in flood forecasting and warning.
Case study 2 - Upper Blue Nile
141
Hydrographs of Gumara river, 3-day forecast horizon
300
Monthly mean
Streamflow (m3/s)
250
200
Simulated with
measured
150
100
50
0
03/06/2000 23/06/2000 13/07/2000 02/08/2000 22/08/2000 11/09/2000 01/10/2000
Figure 5.14. Gumara streamflow forecasts with Monthly mean precipitation
forecasts as input to HEC-HMS, 3-day forecast horizon, wet season 2000
Hydrographs of Gumara river, 3-day forecast horizon
300
Min EPS
Streamflow (m3/s)
250
Simulated with
measured
200
150
100
50
0
03/06/2000 23/06/2000 13/07/2000 02/08/2000 22/08/2000 11/09/2000 01/10/2000
Figure 5.15. Gumara streamflow forecasts with Min EPS precipitation forecasts as
input to HEC-HMS, 3-day forecast horizon, wet season 2000
142
Anticipatory Water Management
Hydrographs of Gumara river, 3-day forecast horizon
350
Mean EPS
Streamflow (m3/s)
300
250
Simulated with
measured
200
150
100
50
0
03/06/2000 23/06/2000 13/07/2000 02/08/2000 22/08/2000 11/09/2000 01/10/2000
Figure 5.16. Gumara streamflow forecasts with Mean EPS precipitation forecasts as
input to HEC-HMS, 3-day forecast horizon, wet season 2000
5.5.7 Flood early warning verification
The warning verification analysis is conducted on seasonal basis. The typical
rainy season when flood threat is common in Ethiopia (June to September) is
considered for a streamflow threshold indicated average flood events. The
verification result for each of the forecast type is described in terms of
number of hits, missed events and false alarms.
Figure 5.17 shows the number of hits and false alarms for the mean EPS as
input to the streamflow model. The number of hits alone can not give
sufficient information to decide on forecast performance. The other
important aspect to be considered in issuing early flood warning is the
number of false alarms. False alarms need to be minimized, which otherwise
would cause the warning users to loose trust and confidence in the
forecasting centre that issues the warning.
It can be seen from Figure 5.17 that while for a 1-day forecast horizon still 8
events are forecasts, only 5, and 4 hits out of 9 events are recorded in the
forecast horizons of 2, and 3 days respectively. The number of false alarms
is 11 for 3-day forecast horizon. The number of false alarms drops down for
forecast horizons greater than 5-days. This shows that these forecasts do not
anymore predict peak flows for more than the 5-days forecast horizon.
Figure 5.17 shows clear decrease of forecast skill with increasing lead-time.
The number of hits of 5 out of a total of 9 events for the 2-day forecast
horizon is not very high. Again, visual inspection of the reference and
forecasted hydrographs may clarify the patterns in the verification results.
Case study 2 - Upper Blue Nile
143
Number of forecasts
Middle threshold (2000-2006)
Total events = 9
12
10
8
6
4
2
0
1
2
3
4
5
6
7
8
9
10
Forecast Horizon(days)
Mean EPS Hits
Mean EPS False alarms
Figure 5.17 Number of hits and false alarms with mean EPS
Figure 5.18 shows that in 2002 the 2-day forecasts with mean EPS
precipitation as input show good resemblance with the reference streamflow
time series. The chosen warning threshold of 250 m3/s leads to warnings for
the peak events, however the timing should be further improved (there seems
to be a delay in the forecasts). Application of the lower warning threshold of
210 m3/s would increase the number of identified flood events, and the
number of hits by the forecasts. Another way of increasing the number of
hits is to look to ensemble forecasts between the Mean EPS and the Max
EPS.
Hydrographs of Gumara river, 2-day forecast horizon
350
300
Streamflow (m3/s)
250
Mean EPS
Simulated with
measured
Threshold
200
150
100
50
0
24/05/2002
13/06/2002
03/07/2002
23/07/2002
12/08/2002
01/09/2002
Figure 5.18 Mean EPS based flood forecast (2002, 2-days forecast horizon)
144
Anticipatory Water Management
5.6 Anticipatory management strategy
development
Although the EPS forecasts predicted only part of the modelled flood peaks
in the analysis period it is valuable to discuss the methods for applying the
EPS in operational flood warning, because further developments may
improve the forecasts in the future. A local quantitative rainfall forecast may
be nested in the global NCEP forecasts when the National Meteorological
Agency of Ethiopia (NMA) succeeds to fully build up an MM5 model. The
agency is now in semi-operational level with some constraints in initializing
the model with local data.
Two different warning thresholds (medium and low) are suggested based on
the results of the EPS forecasts verification analyses. The lower threshold
could be used for an early warning for alert of operational services and
decision makers, while the medium threshold could be used for issuing flood
warnings. The verification analyses results suggest that the mean-EPS as
input to the rainfall-runoff model provides the best predicitions.
The following two questions need to be addressed in order to issue an
effective warning.
- Who should first receive the warning and
- When should the public receive it?
Following the countrywide severe flooding in 2006, a Flood Task Force
(FTF) was set up in Ethiopia under the coordination of the Disaster
Prevention and Preparedness Agency (DPPA).
The institutions involved are:
- National metrological agency
- Ministry of water resources
- Non governmental organizations (USAID,WHO,FAO)
- And DPPA itself
The early alarm (3 days before) could be forwarded to this Flood Task
Force. The task force could meet together and discuss what kind of measures
to take and how to evacuate the public. The next day, the streamflow
prediction would be updated with the new precipitation forecasts, now with a
2-days horizon. Based on this forecast result and real-time data then, the
outcome of the FTF discussion may be communicated to the public to help
them pack their belongings, harvest crops if they are nearly matured stage
and be psychologically prepared. The early warning on the basis of the 3-day
forecast horizon takes into account the 48 hrs allowed too early time in the
verification analyses. If a forecasted event would occur 2 days earlier as
predicted, still mitigation measures would have been started.
Case study 2 - Upper Blue Nile
145
5.7 Adoption of AWM in operational management
policy
The results provide direction for further EPS research. The result guide to
further research on EPS by considering ensemble forecasts between mean
and max EPS using percentiles for probability thresholds as in the Rijnland
case study (Chapter 4). Secondly, a short analysis of the forecasts showed
that many of the flood forecasts are too late. This tendency of too late
forecast has to be improved in either the rainfall forecast or the HEC-HMS
model.
Both the HEC-HMS model and the ensemble precipitation forecasts need to
be further improved before continuing with development of AWM
strategies. The performance of the forecasting system is not good enough to
consider application in the present form. Calibration and downscaling of the
NCEP-EPS precipitation hindcasts, or replacement with the available higher
resolution NCEP-GFS ensembles, are the preferred first steps to try and
improve the forecasts. Calibration and statistical downscaling and analogues
methods are unlikely to be effective, because of the limited number of
monitoring stations in the area. Expanding the analysis area for the weather
forecasts and complementing ground station data with remote sensing and
re-analysis data can be used to overcome this problem. Dynamic
downscaling will be possible by making use of the limited area MM5 model.
147
6
Conclusions and recommendations
6.1 Contributions to Anticipatory Water
Management
Anticipatory Water Management (AWM) is defined as daily operational
water management that pro-actively takes into account expected future
conditions and events on the basis of weather forecasts. Anticipatory Water
Management is an efficient way to optimise further the operational use of
our water systems.
An approach to the development of Anticipatory Water Management
strategies has been presented. This approach makes use of recent
developments in weather forecasting, ensemble forecasting (providing
forecasts of the dynamic probability distribution of the target variables) and
water system control modelling. Flexible water system control models allow
a wide range of control strategies to be applied in multi-year hindcast
analysis. As archives of weather forecasts and water system state variables
increase, hindcast verification analysis will become the basis for the
development and optimisation of new control strategies.
Threshold based decision rules for early warning of critical events, on the
basis of ECMWF EPS rainfall forecasts and hydrological simulation, have,
for the first time, been verified and optimised for a hindcast archive of
multiple years for a regional water system in the Netherlands. This is a
valuable contribution, because increasingly water boards in the Netherlands
are using ECMWF EPS rainfall forecasts for operational decision support.
Freely available NCEP EPS rainfall forecasts (both real-time and archived)
and HEC-HMS hydrological simulation software were used to generate
ensemble streamflow forecasts for a sub-catchment of Lake Tana in the Blue
Nile basin, Ethiopia. This contribution shows that today much information
for water management is freely available through the Internet and that the
previously prohibitive costs and the lack of infrastructure for application of
hydroinformatics decision support tools in less privileged countries is
disappearing.
A method has been described to perform local, long-term verification
analyses that are customized in order to evaluate probabilistic weather
forecast products and to help in choosing the probability-threshold based
decision rules for application in water management. Verification analysis
methods from meteorology have been used and adapted. In meteorology,
verification is done in terms of right or wrong decisions, e.g. hits, false
148
Anticipatory Water Management
alarms and misses. A suggested modification to the verification analyses
applied in meteorology, is to verify on an event basis, instead of a fixed time
step (e.g. daily).
Another step in the development of AWM that has been emphasised, is the
simulation of controlled water systems. The modelling of a controlled water
system is tested particularly by the high degree of freedom in both the
system and the model, because of the control structures. The right output for
the wrong reasons is a risk in using water system control models. On the
other hand, many systems are (partly) manually operated or at least
supervised by an operational water manager. The decisions of these
managers are not as rigid as a computer simulated control strategy.
Therefore getting a very close fit with a water system control model is
mostly not possible.
A modelling approach has been formulated that takes advantage of the
availability of a large amount of measurement data in controlled water
systems. Water level and flow data at control structures allow for intensive
validation and sub-system calibration to reduce the degree of modelling
freedom, and to model separately the natural rainfall-runoff and hydrodynamic processes. The remaining, unexplained, phenomena, which could
not be captured by physically based modelling, are to be simulated with data
driven modelling. The modelling approach has been applied to the Rijnland
water system model, and has resulted in a clear improvement of the model as
compared to a straightforward calibration and validation. It has been shown
that by improving the long term volume balances of the model, also the short
term water level simulations could be further improved. The resulting water
system control model is more reliable for both design studies and operational
decision support.
For integrated evaluation of AWM strategies, end-users (water boards) need
to define their own criteria upfront. It has been shown for the Rijnland case
study how these criteria can be expressed in a cost function. Then this cost
function can be coupled to continuous simulation runs with the water system
control model to analyse the dynamic cost-benefit analysis over a long
period. This allows a search for least-cost alternatives and further
optimisation of Anticipatory Water Management. It has been shown for the
Rijnland case study that rule-based AWM strategies can be optimised using
the water system simulation model with a Genetic search Algorithm
(NSGAII). The presentation of a range of strategies along a Pareto Front,
allows water managers to relate values of decision variables to requirements
for different objectives, e.g. reducing flood or drought damage costs.
The process of developing AWM has been analysed and the identified steps
have been cast in a framework. The application of this framework to the case
Conclusions and recommendations
149
studies prompts a review of the hypotheses posed at the beginning of this
dissertation (Section 2.6).
6.2 Discussion of the hypotheses
Performance of hydro-meteorological ensemble forecasts over a long period
of time for a particular catchment has been assessed by verification analysis
with continuous simulation. The verification analysis of the ensemble
precipitation and water level forecasts for the Rijnland case study confirms
that:
The comparison of measured precipitation and water
level local to a given water system, with hydrometeorological ensemble forecasts leads to an
improvement in the use of those forecasts (hypotheses 1
and 2).
The verification, on the basis of decision rules for early warning for the need
of anticipation, for different event thresholds showed that the number of
false alarms decreased when higher thresholds were applied. The need for
applying low probability thresholds and the high hit rate for forecast
horizons from 5-8 days could also not have been known without the
verification analysis. This is confirmed by the deviating warning rules as
currently applied by the water board for the heuristic anticipation rules for
their water system. Also, the case study of the Blue Nile shows that without
verification analysis of the ensemble streamflow forecasts and warning
thresholds, flood early warnings run the risk of missing all the flood events.
With all of the critical events of the Rijnland case study forecasted it can be
concluded that:
With hindcast analysis effective decision rules for early
warning of the need for anticipation could be found
(hypothesis 3).
With a maximum hit rate of 60% with respect to simulated reference
streamflow, for the Blue Nile case study it is yet pre-mature to conclude the
effectiveness of the forecasting system applied. Downscaling and calibration
of the NCEP-EPS precipitation hindcasts, or replacement with the available
higher resolution NCEP-GFS ensembles, and improvements of the rainfallrunoff models can further improve the effectiveness of the Blue Nile
ensemble forecasts.
150
Anticipatory Water Management
The long-term, continuous, simulation of the complete
AWM strategy for historic time series has enabled an
optimisation of AWM (hypothesis 4) for the Rijnland
case study.
A Pareto front of least flood damage or least drought damage cost (damage
costs of too high and too low water levels) using AWM strategies showed
clear convergence to optimal decision variables defining the dynamic switch
to AWM and the start and extent of the Anticipatory Water Management
action.
Unless the water authorities are forced to reduce the
flood damage costs, regardless of the costs of adverse
effects, a dynamic cost analysis, as applied in the
Rijnland case study, is needed to support the water
authorities in the decision whether or not to adopt
AWM (hypothesis 5).
Simpler cost-benefit analysis, e.g. with cost-loss ratio's, are not sufficient
because AWM does not concern a yes/no decision with constant cost-loss
ratio. Every event will be different from its predecessors. Note that even the
dynamic, continuous, cost analysis serves as a decision support analysis, not
as a decision model. It may be a strong argument in the decision process, but
incompleteness of the cost-model, uncertainties of the occurrence of critical
events and the performance of the forecasting system in the future, and
institutional, social and political arguments will be taken into account by
decision makers.
The main purpose of the cost-benefit analysis is to benchmark operational
water management strategies to assess whether the current strategy can be
improved and which alternative strategy is most efficient in doing so. As
such, the cost-benefit analysis with the optimisation approach can be used to
assess the current potential of AWM. For the Rijnland case study it can be
concluded that:
The use of ensemble precipitation forecasts to decide on
anticipatory management actions, in preference to reactive management, can reduce the damage over a long
period of time (hypothesis 6).
A rule based, two-stage lowering of reservoir levels for flood control, on the
basis of warnings from ensemble water level forecasts, was shown to be
effective in reducing the estimated total costs of too high and too low water
levels, over an analysis period of 8 years, with 30% (Section 4.8.2).
Conclusions and recommendations
151
Based on the AWM strategies found for the Rijnland
case study, which reduce damage costs of too low and
too high water levels, it is likely that the benefits when
applying AWM, more than compensate for the losses
when AWM is not applied (hypothesis 7). However,
more research into the uncertainties of the expected
benefits and losses is needed to confirm this hypothesis.
One source of uncertainty is how the analysed performance of the
forecasting system for previous years corresponds to the performance of the
coming years. Remaining research questions in this respect are discussed in
the section on recommendations for further research (Section 6.5).
6.3 Conclusions
Anticipatory Water Management outperforms re-active operational water
management or management on the basis of hydrological predictions alone.
ECMWF ensemble precipitation forecasts contain valuable information for
anticipatory water management of regional water systems in the
Netherlands. Hindcast verification analyses for the Rijnland water systems
show that these forecasts are effective in reducing flood damage. The skill of
the forecasts is such that the estimated reduction of flood damage is more
than the increase in damage due to false alarms. Therefore, the use of
ECMWF EPS in Anticipatory Water Management strongly reduces the total
damage costs.
Freely available forecasting products, such as NCEP GFS, and hydrological
simulation modelling systems, such as HEC-HMS, can be used to develop
Anticipatory Water Management strategies world-wide at a low cost-level.
Given the variation in the ratio between the costs of false alarms and missed
events on the one hand, and the benefits of hits and correct rejections on the
other, a continuous cost model including the multi-objectives is the preferred
evaluation criterion. Because of the complex, non-linear relationships
between forecast, interpretation, management action, water system state, and
long-term cost, evolutionary search algorithms are the preferred tool to
expose the pay-offs between different AWM strategies and to enable the
water authority to choose their optimal strategy.
Forecast archives are limited and contain only a few relevant events. Reforecasting with new or updated meteorological products has to be
performed for many previous years and including many relevant events. This
is crucial to enable water authorities to develop anticipatory water
152
Anticipatory Water Management
management strategies and evaluate with confidence whether the water
authority would benefit from applying AWM.
Anticipatory Water Management Strategies should be developed and
optimised using continuous hindcast simulation to verify the accuracy of the
forecast local to the water system of interest, and any decision rule as
defined by the water authority.
The level of today's hydroinformatics tools to simulate off-line the complete
process of real-time Anticipatory Water Management is such that continuous
hindcast simulations for multi-year periods can be executed in a limited
amount of time, with a level of realism that builds confidence, and with a
degree of flexibility in defining decision rules that permits experimentation
with different strategies.
6.4 Recommendations for management practice
The framework for developing Anticipatory Water Management (Section
3.9) is recommended for use as a process guideline to evaluate current
operational water management strategies, and to improve these strategies
with enhanced application of weather forecasts. Through experience with the
case studies it is shown that the Anticipatory Water Management framework
helps water managers and water management policy makers answer
questions on how to develop, evaluate and decide on the adoption of AWM.
In this research the current operational management strategy of the Principal
Water-board of Rijnland has been benchmarked with rule-based AWM
strategies. The rule-based type of strategy matches with current operational
practice. Not only the water board of Rijnland, but an increasing number of
other water boards in the Netherlands, are working with rule-based early
warnings using ensemble forecasts from the ECMWF. These rules are being
set up on the basis of expert judgement and adjusted on a trial-and-error
principle. There is an urgent need to verify these rules using hindcast
analysis. For Rijnland, for example, the optimal decision rules found are
different from the currently applied heuristic rules.
The ensemble water level forecasts for the Rijnland case study are currently
running in real-time, in parallel to the operational DSS of Rijnland, to enable
evaluation of the forecasting system by the operational water managers.
Re-analysis, hindcasting, and verification as facilitated by modelling systems
are the vehicles for bridging the gap between theory and practice. In order to
increase the application of weather forecast products meteorological
organisations should provide more hindcast data sets to allow the end-users
Conclusions and recommendations
153
to assess the performance of the product for their intended use. For water
authorities, in turn, it is essential for them to build up their water-system data
and forecast archives.
Meteorological organisations are mostly limited in staff and computational
power resources to perform hindcasting, while not risking interference with
their operational tasks. Therefore, the responsibility of running and
providing the hindcasts should be separated from the meteorological
organisations that have operational forecasting responsibilities. Independend
hindcasting institutes should be established. A funding model needs to be
found with support from national governments and a wide variety of enduser groups.
A cost model of inappropriate and appropriate anticipatory management
actions, and an adequate simulation model of the controlled water system are
key in applying the analyses as described in this dissertation. These two
requirements are often not readily available for a particular water system, or
with a particular water authority. Development of these cost-models and
simulation models of controlled water systems is recommended.
Space and time variability in predictions of the atmosphere and the water
systems, is such that water authorities cannot rely on general performance
indicators of the weather forcasts as provided by the meteorological
institutes. Water authorities themselves should apply hindcast analyses local
to the water system they are responsible for. The increasing availability of
data and forecast archives, and simple-to-use re-forecasting and simulation
technologies, have taken away the former practical and economic restrictions
in doing this.
6.5 Recommendations for further research
Expanded set of decision rules
This research focussed on development, verification and optimisation of
AWM strategies based on heuristic decision rules with probability threshold
to capitalise on the ensemble forecasts. The choice was made to study the
possibilities of training the decision rules on readily available forecasting
products in the case study areas. The range of heuristic decision rules applied
is not exhaustive. Further research on strategies with additional heuristic
rules to find other important decision variables is recommended. One
example of a potential additional decision variable is 'consistency'. This
variable would support a decision for AWM actions if a certain number of
subsequent forecasts indicate an upcoming critical event, whereas AWM
would not be advised if only one of the subsequent forecasts indicates a
critical event.
154
Anticipatory Water Management
Anticipatory flood control with drainage canals
The anticipatory actions covered in the Rijnland case study concerned a
temporary storage increase in the main discharge canals that make up the
storage basin. In addition, the water level could be lowered in all the small
drainage canals or ditches in the low-lying areas (polders) in anticipation of
a critically excessive rainfall event to create extra storage. The potential of
the anticipatory control for the drainage canals is larger than for the storage
basin (about 60% more storage volume). Following the lower drainage level,
also the groundwater level will drop, which again creates substantial extra
storage in the soil (Schultz, 1992, p. 136-137).
Research into the anticipatory control of polder drainage networks will
become relevant due to the current steps being taken to link the monitoring
networks in the polders and the storage basins. The many small pumping
stations that control the drainage network (200 in the Rijnland area),
compared with 4 pumping stations for the storage basin in the Rijnland area,
have to be controlled centrally and automatically, while presently most small
pumping stations are still operated in local automatic or manual control
modes. Research will have to be performed to check the controllability of the
polder drains, because with their limited dimensions their discharge capacity
might be the limiting factor instead of the discharge capacity of the pumping
stations. The sensitivity of the land use (agricultural) and soil subsidence to
the temporary groundwater level changes due to the anticipatory lowering of
water levels in drainage canals has to be analysed. This again will result in
particular requirements on the accuracy of the forecasting system in terms of
hits and false alarms.
Anticipatory Water Management compared with structural measures
While this research has shown that the application of AWM will both reduce
the frequency of flooding and damage costs of deviations from target water
levels in the Rijnland water system, it will be even more beneficial to the
water authorities if AWM could replace structural measures such as
expanding the storage or discharge capacity.
In the Netherlands the required storage and discharge capacity of the
channelled storage basins in regional water systems is determined on the
basis of estimated return periods of exceedance of a critical water level. The
required return periods depend on the potential flood damage in the
catchment (risk-based approach; IPO, 2006). For example, for most parts of
the channelled storage basin of Rijnland the return period of exceedance of
the critical water level for flood damage should be 100 years or longer. The
return period is estimated using historic extreme rainfall events applied to a
water system simulation model. The design of structural measures to
increase the water system's capacity to meet the return period requirement is
done assuming no failure of these measures, e.g. pumping stations or the use
of emergency storage basins. Indeed, the operational reliability of pumping
Conclusions and recommendations
155
stations (with adequate back-up systems in place) and, more disputable, of
emergency storage basins, allows this assumption to be made.
The reliability of AWM in reducing flood frequency, compared with the
reliability of a structural measure is less, simply because some events are
missed by the forecasting system. The extra storage capacity made available
at the beginning of an event by anticipatory pumping cannot be guaranteed.
Therefore, in the first place, AWM should be seen as an optimisation of the
use of an existing system, not as a (structural) change to the system.
Secondly, AWM will increasingly need to be applied due to societal
demands. This is because weather forecasts are becoming available to
professional and public stakeholders, who increasingly take note of the
information. Then, for example, if flood damage occurs, and forecasts of
extreme rainfall have been given, water authorities will be asked to justify
their decisions if anticipatory actions were not taken. There will be growing
demands to the water authorities to use all the information available.
Assessing the effectiveness of AWM
Although AWM should not be viewed as a structural measure to increase the
capacity of a water system, when applied it will reduce the frequency of
system failure (e.g. exceedance of critically high water levels) and thus
reduce the need for a structural increase in the system capacity. The main
research question then becomes: How much will AWM reduce the frequency
of failure? Therefore, the use of AWM should be taken into account when
estimating the return periods of failure. How to do this is an important and
challenging topic for further research.
Importantly, the extensive archive of recorded extreme rainfall events used
for the frequency analysis is generally not accompanied by weather forecasts
for the same events, which are needed to simulate AWM. Whether and how
these large archives of coupled extreme events and probabilistic weather
forecasts can be created still needs to be determined. How many critical
events and forecasts are needed to determine the effectiveness of AWM with
sufficient statistical reliability is another question to be answered. If, in
addition to the effectiveness, also the (economic) efficiency is to be
determined, then also the statistics of normal hydro-meteorological
conditions and their forecasts (false alarms or correct rejections) need to be
assessed.
Risk-based AWM versus Rule-based AWM
In a follow-up of this research, the risk-based AWM strategy will also be
verified and benchmarked for the same data set. It is hypothesised that the
rule-based AWM strategy might be more successful than the risk-based
approach. This is because the risk-based approach assumes perfect
probabilistic predictions by minimising the expected damage for every timestep. The ensemble hydro-meteorological forecasts are not perfect
156
Anticipatory Water Management
probabilistic predictions. However, probabilistic hydro-meteorological
forecasts are continuously improving and pre- and post-processing
techniques (e.g. downscaling and bias correction) can be used to fine-tune
the forecasts local to the case study area.
Establishing the level of quality of the probabilistic forecasts for which riskbased AWM becomes more cost efficient than rule-based AWM would be
interesting additional research. A second limitation of the (minimum) riskbased approach is that it is expected to be highly governed by the ensemble
average, which may result in too little anticipation to reduce damage of
extreme events considerably. Weights can be used in the objective functions
to take preferences for risk-averse decisions into account, but as soon as
weight factors are introduced, the strategy moves towards strategies of
heuristic rules, which need similar optimisation approaches as described in
this dissertation.
Next to the arguments described above, there is another reason why the rulebased AWM approach may be preferred from an operational water
management point of view. This is because risk-based approaches, with risk
defined as probability of occurrence times damage, inherently provide only
an expected cost over a long period of time. In day-to-day decisions, from
event-to-event, it is not 'Risk' that matters. The water board will not be
confronted by the average, expected damage, but always with either the
maximum, or the minimum damage as a consequence of the momentary
decision. Operational managers want to be aware of the maximum damage
that may occur. Research into the comparison and combination of minimum
risk and rule-based AWM strategies is recommended. The most favourable
strategy will be somewhere in between, differently for every case study,
depending on the quality of the probabilistic forecast, and the requirements
of the water authority.
AWM and climate change
AWM permits a more flexible use of water systems to optimise the
management of critical and extreme events. Because it is based on real-time
meteorological forecasts it increases the preparedness and adaptivity to
climate change. At the same time climate change adds to the uncertainty of
the performance of weather forecasts in the future. This cannot be an excuse
not to use all the currently available information as effectively as possible in
managing our water systems.
However, we should research and monitor the potential challenges for AWM
in a changing climate. For example, statistics of critical events and forecast
accuracy may change. This means that optimal decision rules may also
change and that statistical downscaling and bias correction methods may fail.
Because atmospheric simulation models are physically based, and because of
the use of monitoring data for the initial state, it may be assumed that for a
Conclusions and recommendations
157
large part forecast accuracy is independent of climate change. Frequent recalibration of the models would further contribute to maintaining and
improving forecast skill in a changing climate. Research to enhance methods
of calibration and optimisation methods that accomodate sudden changes in
trends remain of the upmost value in this respect.
While the development of numerical weather prediction so far has shown
improvement or at least the maintenance of weather forecast performance, it
is not inconceivable that for some catchments, in case of sudden climate
change, the frequency of events that are difficult to predict increases faster
than the numerical weather prediction can keep up with (for example
occurrence of convective rainfall). Such (temporal) decrease of forecast
accuracy should be signalled quickly and an adjustment of the AWM
strategy should be considered. Research that verifies assumptions about the
behaviour of the peformance of weather forecasts in a changing climate is
needed.
Wider applicability of AWM
This research focussed on AWM for flood forecasting, early warning,
control and evacuation applications through the two case studies. The
potential for wider application was illustrated in Chapter 1, Figure 1.1.
Applications identified there were, amongst others, hydropower, water
supply, irrigation and urban drainage. These applications are expressed in
terms of the end-use of the water system, but often the requirements for
these different end-uses need to be met simultaneously for the same water
system. Therefore, the applicability of AWM can also be described in more
general terms. The need and effectiveness of AWM depends for each case
study on the spatial scale of the water system, the requirement for the
forecast range, the type of hydrological problem, and the controlability of
the water system. These characteristics will determine the accuracy of the
hydro-meteorological forecasts available and the potential effect of the
anticipatory actions. The (economic) efficiency in addition depends on the
benefit from effectively anticipated events and the adverse effects of false
alarms and missed events. The benefits and adverse effects should be
compared to current practice in which damage occurs as well, due to taking
actions late or taking no control actions at all.
In general the larger the spatial scale of the water system, the lower the
resolution requirements for the meteorological forecasts. With regard to the
forecast range, monthly and seasonal forecasts of precipitation and potential
evaporation, for example, are applied to reservoir control to assure water
supply throughout the year. It becomes clear that there is a potential to use
weather forecasts for all hydro-meteorological variables for all forecast
ranges, from nowcasting to short range (up to 2 days) and from medium
range (2-10 days) to long range (monthly and seasonal).
158
Anticipatory Water Management
The hydrological problems for which AWM is needed can be grouped in
traditional problem descriptions for water management as 'too much water',
'too little water', and 'poor water quality'. While the flood control
applications concern problems with too much water, AWM can also be
applied for problems with too little water (e.g. decisions for water inlets
during dry spells can be taken on the basis of rainfall forecasts for the
coming days) and for problems with water quality (e.g. control can be
optimised to minimise CSO's from urban systems on the basis of now-casts
and short term rainfall forecasts).
The controlability of the water system determines the type and effectiviness
of anticipatory actions. For example, the controlability for the Rijnland case
study is limited because of the strong adverse effects of too low water levels.
Other regional water systems in the Netherlands can lower the water levels
further with less adverse effects, and vice versa with less controlability and
more adverse effects.
We will continue to research the applicability of AWM to case studies
covering the full range of characteristics described above. Scientists,
engineers and practitioners are called on to join in an effort to maximise the
use of hydro-meteorological forecasts in operational water management.
159
References
Abbott, M.B., 1991: Hydroinformatics: Information Technology and the
Aquatic Environment. Ashgate, Aldershot, UK, and Brookfield, USA.
Abbott M.B., Refsgaard J.C., 1996: Distributed Hydrological Modelling,
Kluwer Academic Publishers, Dordrecht, The Netherlands, pp.321
Abbott, M.B., 1999: Introducing
Hydroinformatics, Vol. 1, No. 1, 3-20
Hydroinformatics,
Journal
of
Abbott, M.B, 2005: The waterknowledge initiative, Knowledge Engineering
and European Institute for Industrial Leadership, Brussels
Andel, S.J. van, Price, R.K., Lobbrecht, A.H., Kruiningen, F. van, Mureau,
R., 2008a: Ensemble Precipitation and Water-Level Forecasts for
Anticipatory Water-System Control, J. Hydrometeor., 9, 776–788
Andel, S.J. van, Lobbrecht, A.H., Price, R.K., 2008b: Rijnland case study:
hindcast experiment for anticipatory water-system control, Atmospheric
Science Letters, Vol. 9, No 2, 57-60
Andel, S.J. van, Price, R.K., Lobbrecht, A.H., Kruiningen, F. van, 2009a:
Modelling controlled water systems, J. of Irrigation and Drainage. in press
Andel, S.J. van, Price, R.K., Lobbrecht, A.H., Kruiningen, F. van, Mureau,
R., 2009b: Framework for Anticipatory Water Management: application in
flood control for Rijnland reservoir system, submitted
Akhtar, M. K., Corzo, G. A., van Andel, S. J., Jonoski, A., 2009: River flow
forecasting with artificial neural networks using satellite observed
precipitation pre-processed with flow length and travel time information:
case study of the Ganges river basin, Hydrol. Earth Syst. Sci., 13, 1607-1618
Assefa, K.A., 2008: Flood forecasting and early warning in Lake Tana subBasin, Upper Blue Nile, Ethiopia, UNESCO-IHE MSc Thesis WSE-HI.0808
Atger, F., 2001: Verification of intense precipitation forecasts from single
models and ensemble prediction systems, Nonlinear Processes in
Geophysics, Vol. 8, 401-417
160
Bálint, G., Csík, A., Bartha, P., Gauzer, B., 2005: Application of
Meteorological Ensembles for Danube Flood Forecasting and Warning,
NATO Advanced research Workshop Proceedings, Trans–boundary Floods:
Reducing Risks and Enhancing Security through Improved Flood
Management Planning, Oradea B.F. (ed.), 282-293, Oradea, Romania, Treira
S.R.L.
Barreto, W., Price, R.K., Solomatine, D.P., Vojinovic, Z., 2006: Approaches
to multi-objective multi-tier optimization in urban drainage planning, 7th
International Conference on Hydroinformatics, 2006, Nice, FRANCE
Bashar, K.E., Zaki, A.F., 2006: SMA Based Continuous Hydrologic
Simulation of The Blue Nile, Proceeding of the International Conference for
the FRIEND/Nile Project Towards a Better Cooperation, Sharm El Shiekh,
Egypt, 12-14 November 2005
Belaineh, G., Peralta, R.C., Hughes, T.C., 1999: Simulation/Optimization
Modeling for Water Resources Management, Journal of Water Resources
Planning and Management, 125(3), 154-161
Beven, K., Freer, J., 2001: Equifinality, data assimilation, and uncertainty
estimation in mechanistic modelling of complex environmental systems
using the GLUE methodology, Journal of Hydrology, Vol. 249, 11-29
Bhattacharya, B., Lobbrecht, A. H., Solomatine, D. P., 2003: Neural
Networks and Reinforcement Learning in Control of Water Systems, Journal
of Water Resources Planning and Management, Vol. 129, No. 6, 458-465
Boetzelaer, M. van, Schultz, B., 2005: Historical Development of
Approaches and Standards for Flood Protection along the Netherlands Part
of the Rhine River, ICID 21st European Regional Conference, 2005,
Frankfurt (Oder) and Slubice, Germany and Poland
Bokhorst, J., Lobbrecht, A.H., 2005: Neerslag-kansverwachting voor het
waterbeheer. (Probability precipitation forecasts for water management).
H2O, 13, 26–29 (in Dutch)
Buizza, R., Houtekamer, P.L., Toth, Z., Pellerin, G., Wei, M., Zhu, Y., 2005:
Comparison of the ECMWF, MSC, and NCEP Global Ensemble Prediction
Systems. Mon. Wea. Rev., 133, 1076-1097
Carpenter, T.M., Sperfslage, J.A., Georgakakos, K.P., Sweeney, T., Fread,
D.L., 1999: National threshold runoff estimation utilizing GIS in support of
operational flash flood warning systems, J. Hydrol. 224(1), 21-44
161
Chu, X., Steinman, A., 2009: Event and Continuous Hydrologic Modeling
with HEC-HMS, Journal of Irrigation and Drainage Engineering, 135(1),
119-124
Clark, M.P., Gangopadhyay, S., Hay, L., Rajagopalan, B., Wilby, R., 2004:
The Schaake shuffle: A method for reconstructing space–time variability in
forecasted precipitation and temperature fields, J. Hydrometeor., 5, 243-262
Clemmens, A.J., Bautista, E., Wahlin, B.T., Strand, R.J., 2005: Simulation of
Automatic Canal Control Systems, Journal of Irrigation and Drainage
Engineering, 131(4), 324-335
Coello, C.C.A., 2005: Twenty Years of Evolutionary Multi-Objective
Optimization: A Historical View of the Field
Deb, K., Pratap, A., Agarwal, S., Meyarivan, T., 2002: A Fast and Elitist
Multiobjective Genetic Algorithm: NSGA–II., IEEE Transactions on
Evolutionary Computation, 6(2):182-197
DFO., 2008; Dartmouth Flood observatory
www.dartmouth.edu/~floods/index, Accessed; March, 2008
DPPC, 2007: Woreda Agriculture and Rural Development Office
ECMWF, 2006: The Catalogue of ECMWF Real-Time Products,
www.ecmwf.int/products/catalogue
ECMWF, 2007: The evolution of the ECMWF analysis and forecasting
system. (www.ecmwf.int/products/data/operational_system/evolution)
Falkovich, A., Kalnay, E., Lord, S., Mathur, M. B., 2000: A New Method of
Observed Rainfall Assimilation in Forecast Models, Journal of Applied
Meteorology: Vol. 39, No. 8, 1282-1298
Ferraris, L., Gabellani, S., Rebora. N., 2003: A comparison of stochastic
models for patial rainfall downscaling, Water Resources Research, VOL. 39,
NO. 12, 1368, SWC 12
Franz, K., N. Ajami, J. Schaake, R. Buizza, 2005: Hydrologic Ensemble
Prediction Experiment Focuses on Reliable Forecasts, Eos, Trans. Amer.
Geophys. Union, 86(25), 239
Hamill, T., 2009: An overview of the use of reforecasts for precipitation
forecast calibration, HEPEX workshop, 15-18 June, 2009, Toulouse
HKV LIJN IN WATER, 2006: Toekomstig Waterbezwaar Rijnland,
162
Hoofdrapport fase 2 en 3, Project report, PR981.10, 59 pp. (in Dutch)
Hlavcova, K., Szolgay, J., Kubes, R., Kohnova, S., Zvolenský, M., 2005:
Routing of Numerical Weather Prediction System Outputs through a
Rainfall–Runoff Model. NATO Advanced research Workshop Proceedings,
Trans–boundary Floods: Reducing Risks and Enhancing Security through
Improved Flood Management Planning, Oradea B.F. (ed.), 105–116, Oradea,
Romania, Treira S.R.L..
Hoes, O.A.C., 2007: Aanpak wateroverlast in polders op basis van
risicobeheer, Optima Grafische Communicatie, Rotterdam, ISBN 978-909021597-6, 188 pp. (in Dutch)
Holly, F.M. Jr., Merkley, G.P., 1993: Unique Problems in Modeling
Irrigation Canals, Journal of Irrigation and Drainage Engineering, 119(4):
656-662
Huang, W.-C., Yuan, L.-C., 2004: A drought early warning system on realtime multireservoir operations, Water Resources Research, Vol. 40, No. 6.,
W06401
HydroNet, 2009: Software, www.hydronet.nl
IPO, 2006: IPO-richtlijn ter bepaling van het veiligheidsniveau van
boezemkaden (Guideline for determining safety standards of embankments
of regional storage basin channels), (in Dutch)
Kebede, S., 2006: Water balance of Lake Tana and its sensitivity to
fluctuations in rainfall, Blue Nile basin, Ethiopia, Journal of Hydrology 316,
Issues 1-4, 233-247
Kok, C.J., 2000: On the behaviour of a few popular verification scores in
yes/no forecasting, KNMI Scientific Report, WR-2000-04, De Bilt, The
Netherlands, 73 pp.
Krzysztofowicz, R., 2001: The case for probabilistic forecasting in
hydrology. J. Hydrol., 249, 2-9
Krzysztofowicz, R., 2002: Bayesian system for probabilistic river stage
forecasting, Journal of Hydrology, Vol. 268, 16-40
Kummerow, C., 2000: The status of the Tropical Rainfall Measuring
Mission (TRMM) after two years in orbit, J. Appl. Meteor., 39, 1965-1982
163
Legg, T., Mylne, K., 2003: Numerical Weather Prediction Severe weather
early warnings from ensemble forecast information, Forecasting Research
Technical Report No. 415, Met Office, Bracknell, UK
Lobbrecht, A.H., 1997: Dynamic Water-System Control, Design and
Operation of Regional Water-Resources Systems, Dissertation, A.A.
Balkema Publishers, ISBN 9054104325
Lobbrecht, A.H., Dibike, Y.B., Solomatine, D.P., 2002: Applications of
Neural Networks and Fuzzy Logic to Integrated Water Management,
STOWA Project Report, Delft, the Netherlands
Lobbrecht, A.H., Hiemstra, G., Talsma, M., Vonk, Z.,
Neerslaginformatie voor het waterbeheer, H2O, Vol. 23 (in Dutch)
2003:
Lobbrecht, A.H., Loos, S., 2004: From Precipitation Data to Flood
Forecasting, 6th International Conference on Hydroinformatics - Liong,
Phoon & Babovic (eds), World Scientific Publishing Company
Lobbrecht, A.H., Solomatine, D.P., 1999: Control of water levels in polder
areas using neural networks and fuzzy adaptive systems, Water Industry
Systems: Modelling and Optimization Applications, Vol.1, D. Savic, G.
Walters (eds.). Research Studies Press, Ltd. Baldock, UK,. 509-518
Lutgens, F. K., Tarbuck, E. J., 2001: The Atmosphere, An Introduction to
Meteorology, 8th edition, Prentice Hall, ISBN 0130879576, New Jersey, US
Maidment, D.R., 1994: Hand book of Hydrology
Maskey, S., 2004: Modelling Uncertainty in Flood Forecasting Systems,
Dissertation, ISBN: 978-90-5-809694-4
Maskey, S., Andel, S.J. van, Venneker, R., Uhlenbrook, S., 2008: A fuzzyprobabilistic approach to support flood warning decision making, submitted
Moll, J.R., Parmet, B.W., Sprong, T.A., 1996: Country paper of the
Netherlands, Proceedings of Management and Prevention of Crisis
Situations: Floods, Droughts and Institutional Aspects. No. 2, EurAqua,
European Network of Fresh Water Research Organizations: 123-142
Molteni, F., Buizza, R., Palmer, T.N., Petroliagis, T., 1996: The ECMWF
Ensemble Prediction System: Methodology and validation. Quart. J. Roy.
Meteor. Soc., 122, 73-119
164
Mureau, R., Molteni, F., Palmer, T.N., 1993: Ensemble Prediction using
dynamically conditioned perturbations, Q.J.R. Meteorol. Soc., 119, pp. 299323
NASA, 2008: TRMM data interface.
http://disc2.nascom.nasa.gov/Giovanni/tovas/. Last visited: 7-10-2008
Nash, J. E., Sutcliffe, J. V. , 1970: River flow forecasting through conceptual
models. Part 1: A discussion of principles, Journal of Hydrology, 10(3): 282290
NCEP, 2008: Global Forecast System,
ftp://ftp.cdc.noaa.gov/pub/Datasets.other/refcst/ensdata/
Nestler, J.M., Goodwin, R.A., Loucks, D.P., 2005: Coupling of engineering
and biological models for ecosystem analysis, Journal of Water Resources
Planning and Management, 131(2): 101-109
NHS, 2002: Information for Action, A good practice guide on anticipatory
management in healthcare, Department of Health, Healthcare Operational
Intelligence Project,
http://www.dh.gov.uk/assetRoot/04/06/30/16/04063016.pdf
Olsson, J., Lindstrom, G., 2008: Evaluation and calibration of operational
hydrological ensemble forecasts in Sweden, Journal of Hydrology, Volume
350, Issues 1-2, 14-24
Overloop, P.J. van, 2006: Model Predictive Control on Open Water Systems,
Ph.D. dissertation, 2006
Park, J., Obeysekera, J., Zee, R. van, 2007: Multilayer Control Hierarchy for
Water Management Decisions in Integrated Hydrologic Simulation Model,
Journal of Water Resources Planning and Management, 133 (2): 117-125
Persson, A., Grazzini, F., 2007: User Guide to ECMWF forecast products,
Meteorological Bulletin M3.2, ECMWF, UK
Price, R.K., 2006: Engineering and technical aspects of managing floods:
hazard reduction, operational management and post-event recovery.
Frontiers in Flood Research, Kovacs Colloquium, IAHS publ. 305, 73-91
Price, R.K., 2008: Hydroinformatics in retrospect, UNESCO-IHE, Delft, The
Netherlands, ISBN 978-90-73445-20-8
165
Rayner, S., Lach, D., Ingram, H., 2005: Weather Forecasts are for Wimps:
Why Water Resource Managers Do Not Use Climate Forecasts, Climatic
Change, 69 (April 2005), 2–3, 197-227
Richardson, D.S., 2000: Skill and relative economic value of the ECMWF
ensemble prediction system. Quart. J. Roy. Meteor. Soc., 126, 649-667
Rijnland, 2000: Studie Toekomstig Waterbezwaar. (Study of future
hydrological loads). Project report, Rijnland Water Board, Leiden, the
Netherlands (in Dutch)
Roo, A.P.J. de, and Co-authors, 2003: Development of a European flood
forecasting system. Intl. J. River Basin Management, 1 (No. 1), 49-59
Roulin, E., 2007: Skill and relative economic value of medium–range
hydrological ensemble predictions. Hydrol. Earth Syst. Sci., 11, 725-737
Roulin, E., Vannitsem, S., 2005: Skill and Medium–Range Hydrological
Ensemble Predictions. J. of Hydrometeor., 6, 729-744
Sattler, K., Feddersen, H., 2003: Final report for the special project
SPDKEFFS: Heavy rain in Europe, Danish Meteorological Institute,
Copenhagen, Denmark
Schaake, J.C., Franz, K., Bradley, A., Buizza, R., 2006: The Hydrologic
Ensemble Prediction EXperiment (HEPEX), Hydrol. Earth Syst. Sci.
Discuss., 3, 3321-3332
Schaake, J.C., Hamill, T.M., Buizza, R., Clark, M., 2007: HEPEX: The
Hydrological Ensemble Prediction Experiment, BAMS, 88(10), 1541-1547
Schilling, W., 1990: Foreword, in Applications of Operations Research to
Real Time Control of Water Resources Systems, Proceedings of the First
European Junior Scientist Workshop, Kastanienbaum/Luzern, Switzerland,
15-18 March, edited by Einfalt, T., Grottker, M., Schilling, W.,
Schriftenreihe der EAWAG, CH-8600, Dübendorf, Switzerland
Schultz, E., 1992: Waterbeheersing van de Nederlandse droogmakerijen,
(Water management of the drained lakes in the Netherlands), Dissertation
T.U. Delft, ISBN 90-369-1087-0
SMEC, 2006: Project preparation, Flood prepardness and early warning,
Technical back ground paper, volume1: Main report. Addis Ababa, Ethiopia,
ENTRO
Smith, K., Ward, R., 1998: Floods: Physical processes and human impacts,
John Wiley & Sons, Chichester, England
166
SRTM, 2008: Digital Elevation Model data,
www.ambiotek.com/srtm. Last visited: February 2008
global
coverage,
Swinkels, M.A.J., 2004: Control rules for operational management of polder
storage basin systems, M.Sc. Thesis, Technical University Delft (in Dutch)
Thielen, J., Bartholmes, J., Ramos, M.-H., Roo, A. de, 2009: The European
Flood Alert System – Part 1: Concept and development, Hydrol. Earth Syst.
Sci., 13, 125-140
Todini, E., 1999: Using phase-state modelling for inferring forecasting
uncertainty in nonlinear stochastic decision schemes, Journal of
Hydroinformatics, 01.2, 75-82
USACE, 2003: Hydrological Modelling System HEC-HMS, Technical
Manual, Version 3.1., US Army Corps of Engineers, Hydrological
Engineering Center
Vriend, H.J. de, 2002: Model-based morphological prediction: from art to
science, In: River Flow 2002, Eds. D. Bousmar, D.; Zech, Y.; Proc. Int.
Conf. Fluvial Hydraulics, Louvain-la-Neuve, Belgium, 4-6 September 2002,
3-12
Wandee, P., 2005: Optimization of water management in polder areas : some
examples for the temperate humid and the humid tropical zone, Dissertation,
Taylor & Francis, London
WB21, 2000: Waterbeleid voor de 21e eeuw, Geef water de ruimte en de
aandacht die het verdient, Advies van de Commissie Waterbeheer 21e eeuw,
Water policy for the 21th century (in Dutch), pp 97
Weijs, S.V., Leeuwen, P.E.R.M. van, Overloop, P.J. van, Giesen, N. van de,
2007: Effect of uncertainties on the real time operation of a lowland water
system in the Netherlands, IAHS Publ. 313, 463-470
Werner, M., Reggiani, P., Roo, A. de, Bates, P., Sprokkereef, E., 2005:
Flood Forecasting and Warning at the River Basin and at the European
Scale, Natural Hazards, 36, 25–42
WMO, 2004: Report of the CHy Working Group on Hydrological
Forecasting and Prediction, Commision for Hydrology, CHy-XII/Doc. 6,
7.VII.2004
Yufeng, G., 2003: Flood Modelling and Forecasting for Rijnland Area Using
Physically Based and Data-Driven Methods, M.Sc. Thesis, HH 470,
UNESCO-IHE, Delft, the Netherlands
167
List of Figures
Figure 1.1 Significance of meteorological forecasts for operational water
management applications ........................................................................... 22
Figure 1.2 Basic representation of the process of Anticipatory Water Management 25
Figure 2.1 Elements of a hydro-meteorological ensemble prediction system .......... 44
Figure 2.2 Elements of a hydro-meteorological ensemble prediction system
expanded with a Decision support pre-processor for end-use of the
predictions in Anticipatory Water Management ........................................ 47
Figure 2.3 ECMWF EPS precipitation time series for location De Bilt (NL) (data
source: KNMI). When applying threshold-based decision rules for EPS, the
event threshold (Precipitation threshold), the forecast horizon and the
probability threshold have to be set. The probability threshold is the
required forecasted probability that the precipitation threshold will be
exceeded. This is determined by the ensemble members exceeding the
precipitation threshold................................................................................ 49
Figure 3.1 Unnecessary high measured water levels (a) and high measured water
levels prevented by early lowering of storage level (b).............................. 56
Figure 3.2 Upper and lower precipitation thresholds for accumulated precipitation in
Rijnland. After seven days (veritical line) the minimum threshold does not
increase anymore........................................................................................ 59
Figure 3.3. Example of anticipatory action. Reservoir level is lowered in anticipation
of a flood event. As a result of the anticipatory lowering, the resulting peak
reservoir level is reduced............................................................................ 60
Figure 3.4 General process of anticipatory water management................................ 62
Figure 3.5 Fictitious example of a three-member ensemble precipitation forecast. At
a certain lead-time the uncertainty might be considered too high for
decision making (Tmax)............................................................................. 64
Figure 3.6 Creating hindcasts. The forecasting process is repeated for every time
step t in the past.......................................................................................... 65
Figure 3.7 Framework for modelling controlled water systems. Visualisation,
discussion and modelling of the unknown processes are key. If the
processes can be identified (e.g. by error analysis for different time scales)
and isolated after visualisation and discussion, they can be represented by
an internal or external physically based model, if not, a data driven
approach can be used.................................................................................. 70
Figure 3.8 Fictitious example of a decision rule, based on three members of an
ensemble hydro-meteorological forecast.................................................... 71
Figure 3.9 Risk based decision tree for flood warning, where the alternatives are W
= {0, 1} and the future states of the system are F = {0, 1}. W = 0 and W = 1
imply “do not issue warning” and “issue warning”, respectively. Similarly,
F = 0 and F = 1 imply “the area is flooded” and “the area is not flooded”,
respectively. (Cited from Maskey et al., 2008) .......................................... 72
Figure 3.10 Total cost estimation for alternative operational water management
strategies..................................................................................................... 73
Figure 3.11 Pareto front for a 2-objective (criteria) optimisation problem with AWM
strategies..................................................................................................... 78
168
Figure 3.12 Framework for developing Anticipatory Water Management. The main
part of the framework consists of steps for screening of new forecast
products and control strategies. If new control strategies perform well, in
the next step the optimisation of the AWM strategy can be performed. .... 80
Figure 4.1 Principal Water-board of Rijnland: controlling a low-lying regional water
system in the western part of the Netherlands. A channelled storage basin
collects all the excess water of the area. The water level in the storage basin
is controlled by four pumping stations. ...................................................... 82
Figure 4.2 Water level control of the Rijnland storage basin, with and without
forecasting. When using forecasts and temporarily allowing lower water
levels, extra storage of 2.2 x 106 m3 can be created before the extreme
event occurs................................................................................................ 83
Figure 4.3 Comparison of radar and ground station precipitation estimates for the
Rijnland area. The graphs show close resemblance for both dry and wet
periods........................................................................................................ 85
Figure 4.4 Aquarius water system control model of Rijnland (Yufeng, 2003)......... 87
Figure 4.5 Calibration of Aquarius water system control model of Rijnland, the
Netherlands, for a peak water level event in November 2000.................... 91
Figure 4.6 Calibration of Aquarius water system control model of Rijnland, the
Netherlands, for a normal flow period in February and March 2000. ........ 91
Figure 4.7 Validation of the Aquarius water system control model of Rijnland, on
the basis of monthly pumped discharge volumes....................................... 91
Figure 4.8 Cumulative pump discharge volume from the Rijnland storage basin.
Modelled volume is too low, because of underestimation during the dry
summer seasons.......................................................................................... 93
Figure 4.9 Cumulative pump discharge volume for events in the wet winter season
in 2000(a) and 2001(b). For both events modelled volume is higher than
measured volume, indicating over-estimation of the model during excess
water events in the wet season. .................................................................. 93
Figure 4.10 Time scale analysis of difference between measured and modelled pump
discharge. At 90-days moving average a clear sine function with a yearly
period becomes visible. .............................................................................. 95
Figure 4.11 Sine function to model the slow processes error (90-days moving
average) of the Rijnland Aquarius model................................................... 96
Figure 4.12 Monthly pumped discharge volume from Rijnland storage basin in 2002
of the final model. Note the improvements in summer months and October
and November compared to Figure 4.7. Note also the accurately modelled
total yearly volume (0.7% error), while only the total volume over 6-year
simulation (1997-2002) was calibrated. ..................................................... 98
Figure 4.13 Cumulative pump discharge volume for events in the wet winter season
in 2000(a) and 2001(b) after external modelling of unknown processes and
calibration................................................................................................... 99
Figure 4.14 Calibration of the event of November 2000, after the unknown processes
had been included as external data driven models. The modelling of the
peak has improved considerably with respect to the first model (Figure 4.5)
.................................................................................................................... 99
Figure 4.15 Model results for a normal flow period in February 2000, after the
unknown processes had been included as external data driven models.
There are not many differences with the first model (Figure 4.6).............. 99
169
Figure 4.16 Nash-Sutcliffe coefficients for the Rijnland model, with the sine
function and with a constant flow correction, for different time intervals.
.................................................................................................................. 100
Figure 4.17. Cumulative pump discharge volume from the Rijnland storage basin
during calibration and validation. Note that the modelled cumulative
volume now matches very well compared to the first model (Figure 4.8)
and that the development of cumulative volume remains accurate during
the validation period of 2003 and 2004.................................................... 101
Figure 4.18. Cumulative discharge volume for validation events in January and
December 2003. ....................................................................................... 101
Figure 4.19 Contours of number of hits (a) and number of false alarms (b) of the
ECMWF EPS precipitation forecasts for 85 precipitation events in the
Rijnland water system of 15 mm day-1 or more. (ECMWF EPS
precipitation for location De Bilt, from 25 April 1997 to 31 August 2004)
.................................................................................................................. 106
Figure 4.20 Detailed analyses of performance of threshold-based decision rules with
ECMWF EPS precipitation forecasts. (a) Number of hits, events that have
been forecasted too early, missed events and false alarms for a 15 mm day1 precipitation threshold and a 3-day forecast horizon. (b) Measured daily
precipitation and forecasted daily precipitation for a 3-day forecast horizon
and a probability threshold of 0.04 (96th percentile). .............................. 107
Figure 4.21 Comparison of performance of decision rules based on precipitation
forecasts and water level forecasts. [Left] Contours of number of hits (a)
and false alarms (c) with ensemble precipitation forecasts for nine selected
events. Winter precipitation threshold: 40 mm per 3 days. Summer
precipitation threshold: 45 mm per 3 days. [Right] Contours of number of
hits (b) and false alarms (d) with ensemble water level forecasts for nine
selected events. Winter water level threshold: -0.57 m+Ref for 12 hours.
Summer water level threshold: -0.55 m+Ref for 12 hours. ...................... 109
Figure 4.22 Comparison of performance of decision rules based on precipitation
forecasts and water level forecasts. [Left] Contours of number of hits (a)
and false alarms (c) with ensemble precipitation forecasts for nine selected
events. Winter precipitation threshold: 65 mm per 5 days. Summer
precipitation threshold: 65 mm per 5 days. [Right] Contours of number of
hits (b) and false alarms (d) with ensemble water level forecasts for nine
selected events. Winter water level threshold: -0.57 m+Ref for 12 hours.
Summer water level threshold: -0.55 m+Ref for 12 hours. ...................... 110
Figure 4.23 Event based ROC-diagrams of ensemble precipitation (a) for 5, 6, and 7
days forecast horizons. For comparison the ROC curves of the 6-day
forecast horizon water level forecasts and precipitation forecasts for 65
mm/5 days have been plotted (b). The curves show the relationship
between hit rate and false alarm rate for different probability thresholds.
The lowest probability threshold is the upper right end of the curves, for the
highest probability thresholds the curves reach the origin (no hits and no
false alarms). ............................................................................................ 112
Figure 4.24 Ensemble water level prediction of 11 September 1998 for the Rijnland
water system on the basis of ECMWF EPS precipitation forecasts (Van
Andel et al., 2008b)................................................................................... 115
Figure 4.25 Effect of modelled Anticipatory Water Management on water level
control in the Rijnland water system. The modelled peak is lower than the
170
measured peak, because in the model the water level is lowered before the
precipitation event occurs. Exceedance of the -0.57 m+Ref level is
prevented (horizontal line). ...................................................................... 116
Figure 4.26 Water level-cost function Rijnland storage basin (Van Andel et al.,
2009b). ...................................................................................................... 118
Figure 4.27 Comparison of the flood damage cost estimate of the normal control
strategy with a flood risk averse AWM strategy ...................................... 119
Figure 4.28 Comparison of the total damage estimate of the normal control strategy
with a flood risk averse AWM strategy.................................................... 120
Figure 4.29 Theoretical potential of total cost reduction by applying AWM with
perfect (synthetic) forecasts to the Rijnland water system for extreme
events with return periods between 10 and 100 years. Note the logarithmic
scale of the cost-axis. ............................................................................... 121
Figure 4.30 Optimisation of control horizon by minimising estimated total damage
costs. Both the flood costs and the total costs become stable after 70 hrs.
Expansion of the control horizon beyond 70 hrs has no use. The analysis
has been performed on the basis of perfect (synthetic) forecasts for a 1/100
year event. ................................................................................................ 121
Figure 4.31 Estimatied drought (too low water levels) and flood (too high water
levels) damage costs for AWM strategies generated with NSGAII
optimisation. Costs are evaluated for the period between 1-9-1997 and 244-2004. The lower-left corner of the Pareto front shows strategies with total
cost reductions of around 2.4*105 Euro compared to strategies without
anticipation (Van Andel et al., 2009b). ..................................................... 123
Figure 4.32 Estimated Flood damage costs versus Probability threshold for the 150
least total cost AWM strategies determined by NSGAII optimisation. Costs
are evaluated for the period between 1-9-1997 and 24-4-2004. The graph
clearly shows how flood damage reduces when low probability thresholds
are applied. Meaning that when only 1 or 2 ensemble forecast members are
required to exceed the warning level, then most critical events will be
identified and the forthcoming damage reduced by AWM strategy......... 123
Figure 5.1 Ribb and Gumara catchments with the locations of hydro-metrological
gauging stations........................................................................................ 129
Figure 5.2 Data gaps of Addis Zemen gauging station, indicated by circles for long
periods of missing data and an arrow for a short period of missing data. In
the analysis only the data between 2000 and 2006 was used. .................. 130
Figure 5.3 Gumara and Ribb daily river discharge from 1998 to 2006. The discharge
data of the Ribb river in the wet season of 1998 seems too flat (circled),
pointing to measurement errors. The data between 2000 and 2006 was used
for analysis. .............................................................................................. 131
Figure 5.4 Gumara calibration result for 2001 ....................................................... 133
Figure 5.5 Ribb calibration for 2001 ...................................................................... 133
Figure 5.6 Gumara validation for 2005 .................................................................. 134
Figure 5.7 Ribb validation for 2005 ....................................................................... 134
Figure 5.8 Validation results: Ribb correlation (a) , Gumara correlation (b).......... 134
Figure 5.9 Flood areas around Lake Tana sub basin (Aug 13-27, 2006) (DFO, 2008)
.................................................................................................................. 136
Figure 5.10 River cross-sections of Ribb and Gumara .......................................... 137
Figure 5.11 Gumara and Ribb river flows comparison........................................... 137
171
Figure 5.12 Example streamflow hindcasts with a 4-day forecast horizon. After 4
days the underestimation by assuming no rainfall becomes clear. Assuming
Monthly mean rainfall shows a better comparison with the reference
streamflow (simulated by using measured rainfall as input), but streamflow
peaks, particularly in the beginning of the wet season, are underestimated.
.................................................................................................................. 139
Figure 5.13 NRSME for 1, 2, and 3-day forecast horizons for the year 2000 ........ 140
Figure 5.14. Gumara streamflow forecasts with Monthly mean precipitation
forecasts as input to HEC-HMS, 3-day forecast horizon, wet season 2000
.................................................................................................................. 141
Figure 5.15. Gumara streamflow forecasts with Min EPS precipitation forecasts as
input to HEC-HMS, 3-day forecast horizon, wet season 2000................. 141
Figure 5.16. Gumara streamflow forecasts with Mean EPS precipitation forecasts as
input to HEC-HMS, 3-day forecast horizon, wet season 2000................. 142
Figure 5.17 Number of hits and false alarms with mean EPS ................................ 143
Figure 5.18 Mean EPS based flood forecast (2002, 2-days forecast horizon) ........ 143
173
About the author
Schalk Jan van Andel was born 10th of May, 1978, in Amsterdam, The
Netherlands. He lived in Kamoto, Zambia, for three years, and has worked
as a lecturer at the secondary school of Fowakabra, Ghana. In 2003 he
graduated (with distinction) for his MSc degree in Integrated and
quantitative water management from Wageningen University. This
programme included courses in hydrology, computational hydraulics and
water management. During his MSc study he has been involved in national
and international research projects, like the design of innovative flood
reduction measures along the Dutch branches of the Rhine at Delft
Hydraulics (now Deltares) and the development of Earth System Models at
the Potsdam-Institut für Klimafolgenforschung (PIK). He specialised in the
development and application of hydrological and hydrodynamic models.
After graduating he worked as a specialist water management with
HydroLogic, The Netherlands, and as a project officer with the Netherlands
Water Partnership (NWP). By the end of 2004 he joined UNESCO-IHE with
the Hydroinformatics and Knowledge Management department and the
Hydroinformatics core, to start the PhD research presented in this
dissertation. He has published several papers in meteorological and
hydrological international journals, and is a member of the international
HEPEX initiative on Hydrological Ensemble Prediction EXperiments.
At present Schalk Jan is a lecturer in Hydroinformatics at UNESCO-IHE,
Delft, The Netherlands. He is involved in a number of national and
international research projects on operational water management and realtime control of water systems. His research interest concerns the application
of meteorological data and forecasts in operational water management.
Publications in peer-reviewed journals
Andel, S.J. van, Price, R.K., Lobbrecht, A.H., Kruiningen, F. van, Mureau,
R., 2008a: Ensemble Precipitation and Water-Level Forecasts for
Anticipatory Water-System Control, J. Hydrometeor., 9, 776-788.
Andel, S.J. van, Lobbrecht, A.H., Price, R.K., 2008b: Rijnland case study:
hindcast experiment for anticipatory water-system control, Atmospheric
Science Letters, Vol. 9, No 2, 57-60
Andel, S.J. van, Price, R.K., Lobbrecht, A.H., Kruiningen, F. van, 2009a:
Modelling controlled water systems, J. of Irrigation and Drainage, in press
174
Andel, S.J. van, Price, R.K., Lobbrecht, A.H., Kruiningen, F. van, Mureau,
R., 2009b: Framework for Anticipatory Water Management: application in
flood control for Rijnland reservoir system, submitted
Akhtar, M.K., Corzo, G.A., Andel, S.J. van, Jonoski, A.: River flow
forecasting with artificial neural networks using satellite observed
precipitation pre-processed with flow length and travel time information:
case study of the Ganges river basin, Hydrol. Earth Syst. Sci., 13, 16071618, 2009
Conference papers
Lobbrecht, A.H., Andel, S.J van, 2005: Integrated urban and rural water
management using modern meteorological data, Proc. 10th International
Conference on Urban Drainage, 21-26 August 2005, Copenhagen, Denmark
Andel, S.J. van, Lobbrecht, A.H., 2005: Ensemble weather forecasts Applicability and use in flood prevention, Proc. Actif conference,
Innovation, advances and implementation of flood forecasting technology,
17 to 19 October 2005, Tromsø, Norway
Andel, S.J. van, Lobbrecht, A.H., 2006: Ensemble weather forecasts and
operational management of regional water systems, 7th International
Conference on Hydroinformatics (ed. by P. Gourbesville, J. Cunge, V.
Guinot & S.Y. Liong), Research Publishing Services, 1351-1358, Nice,
France
Lobbrecht, A.H., Andel, S.J. van, Kruiningen, F. van, 2006: Operational
management of hydrological extremes using global-scale atmospheric
models, in: Climate Variability and Change—Hydrological Impacts
(Proceedings of the Fifth FRIEND World Conference held at Havana, Cuba,
November 2006), IAHS Publ. 308, 2006., Havana, Cuba
Griensven, A. van, Akhtar, M.K., Haguma, D., Sintayehu, R., Schuol, J.,
Abbaspour K., Andel, S.J. van, Price, R.K., 2007: Catchment Modelling
with Internet based Global Data, 4th International SWAT conference, July 27, Delft, the Netherlands
Andel, S.J. van, Lobbrecht, A.H., Price, R.K., 2007: Rijnland case study:
anticipatory control of a low-lying regional water system, in Thielen., J., J.
Bartholmes J., and J. Schaake (Eds.) (2007) 3rd HEPEX workshop, Book of
Abstracts, European Commission EUR22861EN
175
Andel, S.J. van, Lobbrecht, A.H., Price, R.K., 2008: Anticipatory Water
Management; cost-benefit analysis, Geophysical Research Abstracts, Vol.
10, EGU2008-A-06996, 2008
Andel, S.J. van, 2008: Anticipatory water management for advanced flood
control, in Flood Risk Management: Research and Practice – Samuels et al.
(eds), Taylor & Francis Group, London
S.Loos, S.J.van Andel, A.H.Lobbrecht, R.K.Price, 2008: Anticipatory water
management, decision support for real-time operational and long term
strategic use of new meteorological forecast products in flood control,
Hydropredict, International Interdisciplinary Conference on Predictions for
Hydrology, Ecology, and Water Resources Management: Using Data and
Models to Benefit Society, Czech Republic
Assefa, K.A., Andel, S.J., Jonoski, A., Lobbrecht, A.H., 2009: Combining
Different Verification Methods for Analysis of Flood Early Warnings:
Fogera Plain, Lake Tana, Upper Blue Nile Case Study, 7th ISE & 8th HIC,
Chile, 2009
177
Samenvatting
Water is nauw verweven met onze leefomgeving. Ontwikkelingen in onze
maatschappij hebben via de ruimtelijke ordening invloed op het
watersysteem. De watergerelateerde omgeving waarin we leven legt op zijn
beurt beperkingen op aan het gebruik van de ruimte om ons heen. We richten
onze leefomgeving zodanig in dat we onder normale omstandigheden goed
gebruik kunnen maken van het water, zonder dat het water overlast
veroorzaakt. Extreme omstandigheden kunnen echter tot problemen leiden
met overstromingen en droogte als gevolg. Deze kritische gebeurtenissen
kunnen worden geclassificeerd in te veel water, te weinig water, of water
van een slechte kwaliteit. Met het waterbeheer trachten we voortdurend de
frequentie en omvang van de schade die het gevolg is van kritische
gebeurtenissen, te minimaliseren. We onderscheiden in het algemeen
strategisch waterbeheer en operationeel waterbeheer. Strategisch
waterbeheer is verweven met het landgebruik in een stroomgebied en de
ruimtelijke ordening en hierbij spelen aspecten van het ontwerp van het
watersysteem en effecten op lange termijn. Operationeel waterbeheer, het
onderwerp van dit onderzoek, richt zich op de dagelijkse beheersing van het
watersysteem, waarbij inzet van regelkunstwerken van belang is.
Een grote groep aan kritische gebeurtenissen in het waterbeheer zijn
meteorologisch van aard. Het komt regelmatig voor dat waterbeheerders te
laat zijn geïnformeerd over een op handen zijnde kritische gebeurtenis zoals extreme neerslag - om daarop nog effectief te kunnen reageren. De tijd
die beschikbaar is tussen het moment van een hydrologische of een
meteorologische meting en het moment dat de waterbeheerder deze meting
tot zijn beschikking heeft en kan ingrijpen, is te kort.
Weersverwachtingen en voorspellingen van het gedrag van het watersysteem
bieden uitkomst en kunnen worden gebruikt om de beschikbare reactietijd
voor de waterbeheerder te vergroten. De periode waarover vooruit kan
worden gekeken noemen we de voorspellingshorizon. Het beheersen van het
watersysteem, op basis van een hydrometeorologische voorspelling, wordt
‘anticiperend waterbeheer’ genoemd. Anticiperend waterbeheer stelt
waterbeheerders in staat om op tijd maatregelen te nemen om de schade van
kritische gebeurtenissen te beperken. Een voorbeeld van een anticiperende
maatregel is het verlagen van de waterstand in een boezemstelsel om een
overstroming te voorkomen, wat ook wel wordt aangeduid met ‘voormalen’.
Zoals schakers die hun kansen op het winnen van het spel vergroten door op
zetten van hun tegenstander te anticiperen, kunnen ook waterbeheerders de
prestaties van hun watersysteem verbeteren door zich voor te bereiden op
178
aankomende gebeurtenissen zoals extreme neerslag,
overstromingen of juist droogte en slechte waterkwaliteit.
hoogwater,
De hydrometeorologische voorspellingen, die voor anticiperend waterbeheer
worden gebruikt, zijn niet altijd correct en zijn omgeven met een mate van
onzekerheid. Die onzekerheid hangt samen met de weersverwachting en met
de berekening van het effect van het verwachte weer op het watersysteem.
Vooral de weersverwachtingen hebben een hoge mate van onzekerheid,
omdat de atmosfeer, waarin het weer zich afspeelt, een chaotisch systeem is,
waarin kleine verstoringen snel kunnen uitgroeien tot een niveau waarop ze
ook grootschalige invloed hebben. Anticiperend ingrijpen met
waterbeheerstechnische maatregelen is daardoor in de praktijk een complexe
taak. Als gevolg van de onzekerheid in de voorspellingen en de complexiteit
van waterbeheersing zullen maatregelen soms niet op tijd worden genomen,
of achteraf niet nodig blijken te zijn geweest. Omdat anticiperende
maatregelen mogelijk nadelige effecten met zich meebrengen, moeten de
onzekerheid van de voorspelling en de risico’s van een achteraf onjuist
ingrijpen, worden meegenomen bij anticiperend waterbeheer.
De onzekerheid van de weersverwachting en de veranderlijkheid daarvan in
de tijd, kan worden geschat met zogenaamde ‘ensemble’
voorspellingssystemen. Bij een ensemble voorspelling wordt een
kansverwachting samengesteld, waarmee voor een zekere tijd vooruit de
nauwkeurigheid van de voorspelling berekend is. Deze verwachting wordt
bepaald door het computermodel dat wordt gebruikt voor de
weersverwachting herhaaldelijk te draaien met variërende beginwaarden. Dit
wordt op zo’n manier gedaan dat de variatie van de modeluitkomsten een
maat is voor de onzekerheid van de verwachting. Op deze manier kunnen we
rekening houden met het feit dat we maar in beperkte mate in staat zijn de
actuele staat van de atmosfeer nauwkeurig te meten of te schatten.
Overheden en bedrijven in de watersector maken in toenemende mate
gebruik van deze ensemble voorspellingen. De kansverdeling die voor iedere
verwachting beschikbaar is, stelt een waterbeheerder in staat de risico's van
mogelijke waterbeheerstechnische maatregelen mee te nemen in zijn
beslissing om een maatregel al dan niet in te zetten. Veel onderzoek richt
zich op het leveren van zo goed mogelijke hydrometeorologische ensemble
voorspellingen. Het voorliggende onderzoek, richt zich juist op het zo
effectief mogelijk gebruiken van beschikbare ensemble voorspellingen voor
anticiperend waterbeheer.
In dit onderzoek is een raamwerk opgesteld voor het ontwikkelen van
beheersstrategieën voor anticiperend waterbeheer. In de eerste plaats wordt
in dit raamwerk nadruk gelegd op de beschikbaarheid van instrumenten uit
de hydroinformatica, die het mogelijk maken om op flexibele wijze
computersimulaties uit te voeren van gereguleerde watersystemen. Door
179
gebruik te maken van dergelijke simulatiemodellen, kan een gangbare
waterbeheerstrategie worden nagebootst en worden vergeleken met
alternatieve, anticiperende strategieën.
In de tweede plaats wordt in het raamwerk benadrukt dat waterbeheerders
zelf de prestaties van de hydrometeorologische voorspellingen voor hun
beheersgebied zouden moeten verifiëren, waarbij achteraf een vergelijking
wordt gemaakt met metingen. In het bijzonder geldt voor de meteorologische
verwachtingen dat een verificatie voor het eigen beheersgebied nodig is
omdat de prestatiescores, die worden geleverd door de meteorologische
instituten, vaak zijn bepaald voor een regionale of wereldwijde schaal. Voor
een lokaal stroomgebied kan de prestatie van de weersverwachtingen anders
zijn.
Het is nodig om een verificatie op maat uit te voeren om de effectiviteit van
anticiperend waterbeheer te kunnen vaststellen. Dit betekent bijvoorbeeld
voor hoogwaterbeheersing dat de verificatie zich moet richten op neerslag en
neerslag-afvoermodellering. Daarnaast moet de verificatie niet zijn
gebaseerd op een vast tijdsinterval (bijvoorbeeld van een dag) maar op
gebeurtenissen (bijvoorbeeld een extreme neerslaggebeurtenis die enkele
dagen aanhoudt).
De verificatie moet worden uitgevoerd met continue meerjarige tijdreeksen
en simulaties, en niet op basis van een aantal geïsoleerde kritische
gebeurtenissen, zoals tot voor kort gebruikelijk was. Alleen met continue
simulatie kunnen de volledige gevolgen van het toepassen van anticiperend
waterbeheer worden bepaald, waaronder ook de risico's van onnodige
alarmeringen. Dit laatste is een ‘false alarm’, of ook wel een waarschuwing
voor een kritische gebeurtenis terwijl deze in werkelijkheid niet blijkt op te
treden.
De beschreven aanpak duiden we ook wel aan met ‘verificatie-analyse’.
Hiervoor zijn historische meetreeksen van variabelen in het watersysteem,
historische meteorologische gegevens en historische weersverwachtingen
nodig. Als er geen historie van weersverwachtingen beschikbaar is, moeten
die historische weersverwachtingen alsnog worden gegenereerd. Dit kan
worden gedaan door het numerieke weersverwachtingsmodel opnieuw te
draaien voor de analyseperiode. Dit wordt ook wel aangeduid met ‘reforecasting’ of ‘hindcasting’.
Door de resultaten van de historische weersverwachtingen toe te passen op
een simulatiemodel van het watersysteem, is het mogelijk om voor situaties
uit het verleden alternatieve anticiperende waterbeheerstrategieën na te
bootsen. Deze simulaties laten waterbeheerders zien wat er gebeurd zou zijn
als ze in het verleden weersverwachtingen hadden gebruikt bij het
180
operationele waterbeheer. Hiermee kan de effectiviteit van anticiperend
waterbeheer voor kritische gebeurtenissen worden vastgesteld.
Voor veel overheden in de watersector zal zicht op een verbeterde
effectiviteit van het waterbeheer alleen, niet voldoende zijn om te besluiten
deze techniek toe te passen. In de meeste gevallen zal ook de efficiëntie
moeten worden aangetoond. Omdat waterbeheer zeer dynamisch is, kunnen
geen vaste kosten-baten verhoudingen worden gebruikt voor deze
efficiëntie-analyse. Elke gebeurtenis is anders dan een vorige en hierdoor is
ook de kosten-baten verhouding steeds weer anders.
Daarom wordt in het raamwerk voor anticiperend waterbeheer, in de derde
plaats, benadrukt dat de waterbeheerder een kostenmodel zou moeten
opstellen voor de relatie tussen toestandsvariabelen - zoals waterstanden - en
de opbrengsten of schade in het watersysteem. Hiermee kan de continue
simulatie van de waterbeheersing worden vertaald in een tijdreeks van
kosten. De totale kosten van kritische gebeurtenissen, en de ontwikkeling
van deze kosten door de jaren heen, kunnen worden bepaald en worden
vergeleken voor verschillende voorspellingsproducten en strategieën voor
anticiperend waterbeheer.
Als uit het bovenstaande blijkt dat anticiperend waterbeheer efficiënt is, dan
kan als extra analyse de beheerstrategie worden geoptimaliseerd. Een
belangrijk doel van deze optimalisatie is het minimaliseren van de kosten
van kritische gebeurtenissen, en tegelijkertijd het minimaliseren van de
totale kosten over een meerjarige periode, waarbij ook de reguliere situaties
en onjuist voorspelde situaties horen. Voor de optimalisatie van de
operationele maatregelen lijkt het voor de hand te liggen gebruik te maken
van een minimale risicobenadering, waarbij de kansverdeling uit de
ensemble verwachting wordt gebruikt. Een belangrijke reden waarom deze
techniek hier echter niet is gebruikt is dat de ensemble methode weliswaar
een schatting van de kansverdeling geeft, maar dat dit niet altijd
representatief hoeft te zijn voor de werkelijke kansverdeling. Dit probleem
wordt ondervangen door van een meerjarige periode de effecten van het
gebruik van ensembles te analyseren, waarbij alle onnauwkeurigheden in de
verwachtingen
impliciet
zijn
meegenomen.
Het
meerjarige
optimalisatieprobleem, waarbij per dag meerdere ensemble kansverdelingen
van toepassing zijn en de beste maatregelenstrategie voor de hele periode
moet worden bepaald, is niet in een mathematisch optimalisatiemodel te
vatten en daarom is gekozen voor een ‘random search’ methode, in dit geval
een genetisch algoritme. Belangrijk is dat deze methode ook problemen
aankan met meer doelfuncties tegelijk en daarmee een reeks alternatieve
beheerstrategieën genereert. Zo wordt aan de waterbeheerders de
mogelijkheid geboden om de door hen gewenste waterbeheerstrategie te
selecteren, afhankelijk van de afweging van het belang van de onderkende
doelfuncties. Een waterbeheerder kan er bijvoorbeeld voor kiezen een
praktische strategie toe te passen waarbij op optimale wijze de verwachte
181
kosten van hoogwatergebeurtenissen worden teruggebracht, terwijl de
verwachte totale kosten daardoor misschien niet op het minimum liggen.
Het raamwerk voor het ontwikkelen van strategieën voor anticiperend
waterbeheer is toegepast op twee praktijkonderzoeken die betrekking hebben
op hoogwatervoorspelling, -alarmering en -beheer. Eén van deze
onderzoeken betrof het polder-boezemsysteem van het Hoogheemraadschap
van Rijnland in Nederland. De ander betrof een deelstroomgebied van de
Blauwe Nijl, bovenstrooms van Lake Tana in Ethiopië. De ensemble
neerslagvoorspellingen van het ECMWF Ensemble Prediction System (EPS)
en het NCEP Global Forecasting System (GFS) zijn in de onderzoeken
gebruikt. Het EPS wordt al operationeel ontvangen door het
hoogheemraadschap. Het GFS is vrij beschikbaar via het internet, wat het
een zeer interessant onderzoeks- en operationeel instrument maakt voor
ontwikkelingslanden, met beperkte investeringscapaciteit in meteorologische
verwachtingsinformatie.
In
het
Nederlandse
praktijkonderzoek
werden
de
meeste
hoogwatergebeurtenissen in de meerjarige analyseperiode van 8 jaar goed
voorspeld. De optimalisatie van de anticiperende waterbeheerstrategie voor
Rijnland resulteerde in een reductie van 35% van de hoogwaterschade en een
reductie van 30% in de totale schade. Dit laat zien dat anticiperend
waterbeheer tot beduidend betere resultaten leidt dan het traditionele
reactieve operationele waterbeheer. In Nederland kunnen de ECMWF EPS
voorspellingen worden gebruikt om de voorspellingshorizon uit te breiden
tot drie dagen of meer. De aanmerkelijke verschillen tussen de gevonden
optimale beslissingsregels en de beslissingsregels die op het moment door
Rijnland worden toegepast, bevestigen dat het toepassen van hindcastinganalyses positieve resultaten oplevert voor het verbeteren van anticiperende
waterbeheerstrategieën.
De resultaten van het praktijkonderzoek in de Blauwe Nijl laten zien dat vrij
beschikbare weersverwachtingen en hydrologische modellerings-software
goed kunnen worden gebruikt bij onderzoek naar voorspellingssystemen en
strategieen voor anticiperend waterbeheer. In een gebied waar momenteel
geen vorm van waarschuwing voor hoogwater beschikbaar is, is dat van
grote waarde. In dit specifieke praktijkonderzoek kon maximaal 60% van de
gesimuleerde referentie hoogwatergebeurtenissen worden voorspeld. Het
voorspellingssysteem moet eerst verder worden verbeterd voordat
operationeel gebruik realistisch is. Bij het realiseren van deze verbeteringen
moeten bias-correctie en neerschalingsmethoden worden gebruikt. Een brede
internationale onderzoeksgemeenschap op het gebied van ensemble
voorspellingen richt zich op de ontwikkeling van deze methoden. Deze
methoden, die tot doel hebben om betrouwbare probabilistische
voorspellingen te genereren met een zo klein mogelijke onzekerheidsmarge,
182
worden thans ook gebruikt om de efficiëntie van anticiperend waterbeheer
voor waterschappen in Nederland nog verder te vergroten.
Het belangrijkste onderdeel bij het ontwikkelen van succesvolle
anticiperende waterbeheerstrategieën, is de verificatie-analyse met continue
simulaties voor perioden van verscheidene jaren. Consistente meerjarige
historische reeksen met weersverwachtingen zijn noodzakelijk vanwege de
lage frequentie van kritische gebeurtenissen. Dergelijke historische gegevens
zijn over het algemeen niet beschikbaar omdat weersverwachtingsystemen
continu worden vernieuwd en er tot voor kort binnen de meteorologie weinig
aandacht was voor systematische opslag van deze informatie. Er is daarom
een grote behoefte aan het creëren van hindcast reeksen ter ondersteuning
van de ontwikkeling van nieuwe toepassingen, zoals in het waterbeheer.
Omdat het generen van hindcasts conflicteert met de operationele taken van
meteorologische instituten, en wel in verband met de beschikbare mens- en
computercapaciteit, moet de taak van het hindcasten liefst aan aparte,
gespecialiseerde organisaties worden overgelaten die hier onafhankelijke
financiering en rekencapaciteit voor beschikbaar hebben. Dit zal een grote
bijdrage leveren aan het praktisch gebruik van weersverwachtingen voor
operationeel waterbeheer.
Het gebruik van weersverwachtingen is belangrijk voor het waterbeheer
wereldwijd. Deze verwachtingen zijn tegenwoordig van zodanige kwaliteit,
dat de vraag gesteld moet worden of waterbeheerders het zich nog kunnen
veroorloven geen gebruik te maken van deze beschikbare informatie. Dat
geldt niet alleen voor het hoogwaterpraktijkonderzoek dat in dit proefschrift
is gepresenteerd, maar voor veel meer toepassingen, zoals droogtebeheer,
irrigatie, en voor een breed scala aan typen watersystemen. Daarom wordt
hier een beroep gedaan op wetenschappers, ingenieurs en waterbeheerders
om gezamenlijk het volledige toepassingsbereik van anticiperend
waterbeheer te ontwikkelen en het gebruik van hydrometeorologische
voorspellingen in het operationele waterbeheer te stimuleren.
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement