24th International Conference & Exhibition on Electricity Distribution (CIRED)
12-15 June 2017
Session 3: Operation, control and protection
Risk prediction in distribution networks
based on the relation between weather
and (underground) component failure
ISSN 2515-0855
doi: 10.1049/oap-cired.2017.0610
www.ietdl.org
Tongyou Gu ✉, Jerom Janssen, Edwin Tazelaar, Geert Popma
Alliander N.V., Duiven, Gelderland, The Netherlands
✉ E-mail: tongyou.gu@alliander.com
Abstract: Weather affects the reliability of distribution networks. Extreme weather conditions can initiate outages, but
already normal weather conditions affect the failure of the components in the system, such as cables and cable joints.
This study analyses the historical weather and failure records of Alliander, a Dutch distribution system operator. The
study discusses the correlation between failure rates and different weather factors. It presents a predictive model using
basket analysis. This predictive model is verified using a data set from recent component failures.
1
implemented within Alliander to assist the scheduling of crew
members for power restoration.
Introduction
Weather is a dominant factor affecting the reliability of distribution
systems [1]. Especially, the failures of components, such as cables
and joints, are heavily influenced by weather conditions. Studies
[2, 3] demonstrate the influences of extreme weather conditions
such as hurricanes and ice storms on power failures. However,
even normal weather conditions already correlate with the
reliability of components such as cables and joints. For example,
historical failure data of Alliander, a distribution system operator
(DSO), shows that in summer the System Average Interruption
Duration Index of the Amsterdam area is significantly higher than
other time of the year.
By analysing the impact of weather factors on (underground)
component failures, the expected risk in different regions of the
distribution networks can be predicted. Based on a daily or hourly
prediction, the grid operator can make a more efficient planning of
the number of staffs available for fault clearing service; more crew
members can be assigned to the region with higher failure risk.
Thus, in case of a power outage due to component failure, there
are sufficient crew members who can react immediately for the
power restoration. This helps to decrease the failure handling time
and reduce the outage duration.
This paper investigates the impact of different weather factors on
the failure of (underground) components in medium voltage (MV)
network. It cross-examines failure records from Alliander and
weather records from the Royal Dutch Meteorological Institute
(KNMI). A number of key weather factors over a day are selected
for big-data analysis, including maximum/minimum/average
temperature, the difference between maximum and minimum
temperature of 1 day, precipitation, vaporisation, and solar radiation.
Fig. 1 shows the approach of the study. To obtain the local weather
information for each failure, the weather data from 35 weather
stations in the Netherlands is interpolated to the location of the
failure. Weather data of up to 5 days before the failure is included
to investigate any delay effects. The (cor)relation between weather
and failures is statistically examined. The weather factors with the
most dominant impact on failures are further used to create a
predictive model which predicts the risk in different regions on the
next day. The (cor)relation between weather and the failures of
different types of components is also examined.
The predictive model is tested by the weather and failure data of
the summer of 2016. The test results show that it is possible to
give a moderately accurate prediction of power failure risk based
on the forecast weather information. This model will be further
1442
2
Interpolation of weather data
The KNMI has 35 weather stations over the country. Fig. 2 shows
their locations. To obtain local weather information for component
failures, we use the KNMI data.
For each rayon, an operational region, we use the data from the
stations inside this rayon. This provides regional average weather
data. We investigated several methods for interpolation of weather
data. Other methods provide more granulated weather data, which
slightly improves the correlation between component failure and
weather data. However, the method with a regional average per
rayon has the fastest calculating time. In some instances, there is
missing weather data, if this is the case we use weather from
another weather station nearby. In the rayon of Amsterdam, there
is no weather station. Here we use the nearest station.
3
Normalisation of failure data
For each of the key weather factors selected for this study, the
(expected) failure frequency is normalised to avoid the effect that
certain weather conditions are more common in the country and
occur in more days, which results in more failures occurring under
these weather conditions. For example, in the Netherlands an
over-average amount of failures occur at temperatures around 10
or 20°C simply because there are a lot of days with these
temperatures.
This normalisation is achieved by the following equation:
expected failure frequency =
total number of failures under these weather conditions
number of days with these weather conditions
Here, the total number of failures and the number of days are counted
based on the interpolated weather information.
The failure data used in this analysis is the historical failure
records of Alliander from January 2007 to June 2016. The failure
records contain the type of component, date and time, and the
location.
CIRED, Open Access Proc. J., 2017, Vol. 2017, Iss. 1, pp. 1442–1445
This is an open access article published by the IET under the Creative Commons
Attribution License (http://creativecommons.org/licenses/by/3.0/)
Fig. 1 Approach of investigating influence of weather conditions on failure rates
4.1
Correlation
The correlation between weather and the normalised (expected)
failure frequency is analysed. The impact of weather conditions on
component failure may be cumulative. To study the time delay
effect, the correlation between the failure rate and the weather data
over five preceding days is calculated. For example, Fig. 3 shows
the correlation coefficient between the failure rate and different
weather factors over 5 days in Amsterdam area. Temperature
difference (3 days average) and global radiation (2 days average)
give the highest correlation. This result verifies that weather
conditions have an impact on component failures. Moreover, the
weather impact could be cumulative over the preceding days.
The time delay effect for each weather factor is also obtained. For
each weather factor, there is a highest correlation with the x-day
average, e.g. for the temperature difference the 3 days’ average
value is the highest, thus the representative for this weather factor.
The representative values (x-day average) of all the weather factors
are further used to generate predictive rules.
4.2
Fig. 2 Weather stations in the Netherlands
4
Statistical analysis
The interpolated weather data and the normalised failure frequency
are further used in the statistical analysis.
Predictive model
A predictive model is made to predict the risk in different regions
based on the interpolated weather forecast information.
Basket analysis, an a priori method [4], is applied to generate
prediction rules. Basket analysis is a modelling technique based
upon the theory that if you buy a certain group of items, you are
more (or less) likely to buy another group of items. In our case,
we would like to generate rules that in certain weather conditions,
Fig. 3 Correlation between failure rate and weather conditions in Amsterdam area
CIRED, Open Access Proc. J., 2017, Vol. 2017, Iss. 1, pp. 1442–1445
This is an open access article published by the IET under the Creative Commons
Attribution License (http://creativecommons.org/licenses/by/3.0/)
1443
Table 1 Definition of high, medium, and low for basket analysis
high
medium
low
TG
TN
TX
TD
Q
EV24
RH
Number of failures
(25, +∞)
(−10, 25]
(−∞, −10]
(18, +∞)
(−10, 18]
(−∞, −10]
(30, +∞)
(−5, 30]
(−∞, −5]
(15, +∞)
—
(−∞, 15]
(2500, +∞)
—
(−∞, 2500]
(5, +∞)
—
(−∞, 5]
(20, +∞)
—
(−∞, 20]
(0, +∞)
—
0
Table 2 Predictive rules for expected failure rate in Amsterdam area
Table 3 Predictive rules for expected failure rate in Amsterdam area
If
{TX_2 = high,
Q_2 = high}
{TX_2 = high}
{EV24_4 = high}
Then
Support
(>0.001)
Confidence
(>0.65)
If
num_fault = high
0.001142531
0.8000000
num_fault = high
num_fault = high
0.002285061
0.001142531
0.7272727
0.6666667
{EV24_1 = high}
{TX_2 = high,
Q_2 = high}
it is more likely to have higher failure frequency. The basket analysis
targets high failure risk and seeks relevant conditions with a certain
confidence level.
The input items of the basket analysis need to be discrete
quantities. Therefore, the historical data of the weather factors are
categorised as ‘high’, ‘medium’ or ‘low’. The boundaries are
defined based on the statistical distribution of the data. Table 1
shows the category boundaries. Since the number of failures in the
MV network of a certain region is relatively low from a statistical
perspective, it is defined as ‘high’ when there is one (or more) failure.
Table 2 illustrates the predictive rules for Amsterdam area. The rules
state that when the maximum temperature (2 days average) or
evapotranspiration (4 days average) is high, more failures are expected
in Amsterdam area with a confidence level >0.65. The support value
is larger than 0.001, which indicates that >4 days (0.001×3501 days
in January 2007–June 2016) in the dataset have the weather condition
as in each rule. Grid operators can use the predictive rules as an
indication about which regions are with higher risk.
5
Analysis for different types of components
Predictive rules are also generated for different types of components
in the MV network: cables, cable joints, transformers, and secondary
installations.
Then
Support
(>0.001)
Confidence
(>0.2)
num_fault = high
num_fault = high
0.002158114
0.002189851
0.2054381
0.2005814
As an example, Fig. 4 shows the correlation between failure rate of
cable joints and different weather factors over 5 days. Precipitation
(on the day of failure), temperature difference (3 days average),
and global radiation (2 days average) give the highest correlation.
Again, the representatives for each weather factor are used to
generate the predictive rules. Table 3 illustrates the resulted rules,
in which the reference evapotranspiration, maximum temperature
and global radiation are the most predictive factors. The
confidence level in Table 3 is lower than in Table 2, because the
total number of failures of a certain component type is much
smaller than the number of all the failures, which leads to lower
confidence in statistical analysis. It is discovered that for different
components the dominant weather factor varies.
6
Test results
The predictive model is tested using the weather and failure data of
the summer of 2016. Within the 60 days of July and August, there
were 159 power failures in 10 Alliander operational regions. For
every region on each day (in total 620 cases), the model is applied
to predict the risk of power failures.
Fig. 5 illustrates the summary of the test results. When it is
predicted as ‘low risk’, in 72% of the cases there was indeed no
failure. When it is predicted as ‘high risk’, in 64% of the cases,
there was power failure(s). These results show that based on the
Fig. 4 Correlation between failure rate of cable joints and weather conditions
1444
CIRED, Open Access Proc. J., 2017, Vol. 2017, Iss. 1, pp. 1442–1445
This is an open access article published by the IET under the Creative Commons
Attribution License (http://creativecommons.org/licenses/by/3.0/)
Fig. 5 Test results of predictive model
forecast weather information, it is possible to give a moderately
accurate prediction.
7
discovered that the dominant weather factor varies for different
components.
Using the x-day average representative for each weather factor, a
basket analysis is performed to build the predictive model. The
basket analysis targets high failure risk and seeks relevant
conditions with sufficient confidence level. The predictive model
can be used to give a high-risk alarm when certain weather
conditions occur. The model is tested by the recent failures and
weather records in the summer of 2016. The test results show
moderate accuracy of the prediction.
Weather impact is one of the (indirect) factors that could cause a
component failure. The study suggests to take the root causes into
account to make a more complete predictive model of failure risk,
e.g. weather influence on the physical aging process of
components and/or digging damages of underground components
because of improper weather conditions for digging work.
Alliander is implementing the model in the optimisation of the
scheduling of outage service crew members. This helps to
decrease the failure handling time and reduce the outage duration.
Conclusion
This paper presents an analysis of historical weather and failure
records. It first discusses the correlation between the failures and
different weather factors, and then constructs a predictive model
using basket analysis.
The results show that some weather factors are clearly correlated
with the failure rates of certain component type. The dominant
weather factors (highest correlation coefficient) vary between
different regions. A possible explanation is that the type and age
of the cables in each region are different due to the diversification
of regional grid development over the past decades. It is also
8
1
2
3
4
References
Zhou, Y., Pahwa, A., Yang, S.S.: ‘Modeling weather-related failures of overhead
distribution lines’, IEEE Trans. Power Syst., 2006, 21, (4), pp. 1683–1690
Billinton, R., Acharya, J.: ‘Consideration of multi-state weather models in reliability
evaluation of transmission and distribution systems’. Canadian Conf. Electrical and
Computer Eng., 2005, Saskatoon, Sask., 2005, pp. 916–922
Brostrom, E., Ahlberg, J., Soder, L.: ‘Modelling of ICE storms and their impact
applied to a part of the Swedish transmission network’. 2007 IEEE Lausanne
Power Tech, Lausanne, 2007, pp. 1593–1598
Trnka, A.: ‘Market basket analysis with data mining methods’. 2010 Int. Conf.
Networking and Information Technology, Manila, 2010, pp. 446–450
CIRED, Open Access Proc. J., 2017, Vol. 2017, Iss. 1, pp. 1442–1445
This is an open access article published by the IET under the Creative Commons
Attribution License (http://creativecommons.org/licenses/by/3.0/)
1445
Download PDF
Similar pages