GREEND: An Energy Consumption Dataset of Households in Italy
arXiv:1405.3100v2 [cs.OH] 22 May 2014
GREEND: An Energy Consumption Dataset of
Households in Italy and Austria
Andrea Monacchi, ∗Dominik Egarter, Wilfried Elmenreich
Institute of Networked and Embedded Systems / Lakeside Labs,
Alpen-Adria-Universität Klagenfurt, Austria
[email protected]
Salvatore D’Alessandro, Andrea M. Tonello
WiTiKee s.r.l.
via Duchi d’Aosta 2, Udine, Italy
[email protected]
Abstract
Home energy management systems can be used to monitor and optimize
consumption and local production from renewable energy. To assess
solutions before their deployment, researchers and designers of those
systems demand for energy consumption datasets. In this paper, we
present the GREEND dataset, containing detailed power usage information
obtained through a measurement campaign in households in Austria and
Italy. We provide a description of consumption scenarios and discuss
design choices for the sensing infrastructure. Finally, we benchmark the
dataset with state-of-the-art techniques in load disaggregation, occupancy
detection and appliance usage mining.
Keywords: Energy demand modeling, demand forecasting, home energy
management, smart home, energy consumption dataset, smart appliance
1
Introduction
Stability issues arise from the progressive installation of renewable energy generation and the diffusion of electric vehicles. Demand response exploits a price
signal to reflect fluctuations in the availability of energy in the grid. This
incentives the coordination of electrical devices in order to optimize running
costs, fostering awareness to increase conservation and efficiency [1, 2] or to
∗ We want to thank Benjamin Steinwender, Micha Rappaport and Manfred Pöechacker for
the support provided while carrying out the measurement campaign. The work of Monacchi,
Egarter and Elmenreich is supported by Lakeside Labs, Klagenfurt, Austria and funded by
the European Regional Development Fund (ERDF) and the Carinthian Economic Promotion
Fund (KWF) under grant KWF 20214 — 23743 — 35469 u. 35470. The work of D’Alessandro
and Tonello is cofunded by the Interreg IV Italy-Austria ID-6462 MONERGY project.
1
automatically schedule the operation of selected devices to off-peak periods [3].
Nevertheless, control strategies might somehow disrupt user’s daily life routines.
Collecting information of reoccurring activities can increase the effectiveness of
control strategies, as they can consider necessities of inhabitants to minimize the
discomfort produced. Problems such as load detection and occupancy modeling
for HVAC optimization can be solved using well-established data mining and
machine learning techniques, which demand suitable datasets.
In order to enable research on energy and sustainability problems, it is
necessary to build upon publicly available data in order to elaborate solutions
that work in the real world. In this paper we discuss available energy consumption
datasets and present a new publicly available dataset that we name GREEND1 ,
containing power usage information at device level being obtained through a
measurement campaign in households in Austria and Italy. We describe the
hardware and software setup of the measurement campaign and elaborate the
modeling of loads for energy management applications. The paper focuses on the
dataset and it is meant to provide metadata describing the involved consumption
scenarios. Along with the dataset we are also releasing our code base for the
measurement campaign as a Sourceforge project. Moreover, we benchmark the
dataset by presenting experiences with load disaggregation, occupancy detection
and appliance usage mining using state-of-the-art techniques.
The remainder of the paper is as follows. To identify the main requirements
of an energy consumption dataset, we survey existing datasets in Section 2. In
Section 3, we report the design of the measurement campaign. We describe
our measurement infrastructure and provide a characterization of the deployed
scenarios, in terms of both residents and monitored electrical devices. In Section
4.1, we apply a state-of-the art load disaggregation technique based on particle
filtering to show a possible application of the dataset. Similarly, in section 4.2
and 4.3 we apply the dataset respectively to observe occupancy patterns and
mine appliance usage patterns. Section 5 concludes the paper and anticipates
future directions.
2
Other consumption datasets
To overcome the previously introduced challenges, researchers and designers
of Home Energy Management System (HEMS) require consumption datasets
where solutions can be assessed beforehand. A widely used dataset in the load
disaggregation community is the reference dataset (REDD), publically released
by MIT in 2011 [4]. Moved by a similar aim, various other datasets have been
shared. As shown in Table 1, the main classification attributes are sampling
frequency and characteristics of the signal being measured, active power (P),
reactive power (Q), apparent power (S), energy (E), frequency (f), phase angle
(Φ), voltage (V) and current (I). Certain datasets, such as REDD and BLUED [5]
monitor a small number of households at a high sampling frequency. This depends
on the requirements of load disaggregation, where a higher frequency allows for
extraction of more representative features capturing the transient behavior [1].
1 GREEND:
GREEND Electrical ENergy Dataset.
2
As noticeable in the table, most of datasets were collected in the USA, under a
120V voltage while European countries commonly work under 230V . Moreover,
certain peculiarities such as the weather and climate depend on the location of
the measurement campaign. To the best of our knowledge, GREEND is the first
1Hz consumption dataset for Austria and Italy. The type and number of devices,
as well as the number of monitored households significantly constraints the
final application of the dataset. A statistical analysis of consumption behavior
of residents would require a high number of households, such as in HES and
OCTES. Besides, seasonal consumption behaviors cannot be captured by a shortterm measurement campaign, for instance lasting days or months. Also, some
datasets monitor different households over different time windows, which makes
a comparison of the dwellings impossible. Moreover, the context under which
appliances are used over the day is an essential factor determining the complexity
of the demand. Therefore, collection should take place in real environments
and not in a lab or on a device testbench. For example TraceBase [6] and
ACS-F1 [7] provide consumption tracks from appliances collected in households
and offices. While they constitutes a great source of device signatures, they
currently do not provide information of consumption scenarios, which results
in the impossibility to analyze device interdependence. GREEND is meant to
overcome these limitations as it will be explained in the following section.
3
Table 1: Existing datasets for energy consumption in households
4
Dataset
ACS-F1 [7]
Location
Switzerland
AMPds [8]
BLUED [5]
GREEND
Greater Vancouver
Pittsburg, PA
Austria, Italy
HES
UK
iAWE [9]
India
IHEPCDS1
OCTES2
France
Finland, Iceland, Scotland
Boston, MA
Features
I, V, Q, f, Φ
Resolution
10 secs
1
1
9
#Sensors (per house)
100 devices in total (10
types)
19
Aggregated
9
I, V, pf, F, P, Q, S
I, V, switch events
P
1 min
12 Khz
1 Hz
251
13-51
P
2 min
1
V, I, f, P, S, E, Φ
1 Hz
1
33
33 sensors (10 appliance level)
3 circuits
Aggregated
I, V, P, Q
P, Energy price
1 min
7 secs
3 - 19 days
6
9-24
Sample dataset3 Austin, TX
Smart* [10]
Western Massachussets
7 days
3 months
12
25 circuits, 29 appliance monitors
Tracebase [6]
Germany
N/A
10
1 Sub-metered +2
(Aggregated + Submetered)
15
Aggregate: V, P; Sub- 15 Khz (aggr.), 3 sec
metered: P
(sub)
S
1 min
P, S (circuits), P (sub- 1 Hz
metered)
UK-DALE [11]
UK
499 days
4
REDD [4]
1
2
3
Duration
1 hour session (2 sessions)
1 year
8 days
1 year (3-6 months completed)
1 month (255 houses) 1 year (26 houses)
73 days
#Houses
N/A
4 years
4-13 months
http://tinyurl.com/IHEPCDS
http://octes.oamk.fi/final/
http://www.pecanstreet.org/projects/consortium/
158 devices in total (43
types)
5 (house 3) - 53 (house
1)
P
1-10 sec
Aggregated P, Sub P, 16 Khz (aggr.), 6 sec
switch-status
(sub.)
3
Dataset for Italy and Austria
3.1
The consumption dataset
The measurement campaign is carried out within the MONERGY2 project,
in which we aim at proposing solutions to reduce energy consumption in the
Austrian (AT) region of Carinthia and the Italian (IT) region of Friuli-Venezia
Giulia. In particular, we identified the following requirements:
• Features: For the selection of the features of interest we considered the
requirements of load disaggregation applications, as they are generally
stricter than user and appliance modeling [12]. Accordingly, we decided to
collect active power measurements at 1Hz, as this allows the identification
of more than 8 devices through load disaggregation algorithms [1]. Each
entry is associated to a UTC Unix timestamp so that measurements can
be matched to contextual data such as weather.
• Device selection: The selection of devices followed the energy hogs identified in [2]. We favored diversity in the dataset to promote multiple
applications. However, we faced issues in accessing certain devices in Austria. For instance, electric boilers are commonly connected to a separated
meter to be charged with a reduced energy tariff.
• Household selection: The selection of householders was driven by the
findings identified in [2]. In particular, we wanted to promote diversity of
scenarios, for instance involving different types of dwellings and consumers.
A more detailed description of the scenarios is reported in Section 3.2.
• Campaign duration: The campaign was designed to last one year, in
order to observe and be able to model seasonal consumption behavior of
inhabitants. The first household is being monitored (#0) since the end of
December 2013. Most of other households followed in January 2014, while
the last two platforms (house #7 and #8) were deployed in April.
3.2
Deployment
At the time of writing, we are collecting consumption data in the following
scenarios:
• House #0 a detached house with 2 floors in Spittal an der Drau (AT).
The residents are a retired couple, spending most of time at home.
• House #1 a detached house with 2 floors in Villach (AT). The residents
are 3 university students, having irregular working-like days.
• House #2 an apartment with 1 floor in Klagenfurt (AT). The residents are
a young couple, spending most of daylight time at work during weekdays,
mostly being at home in evenings and weekend.
2 http://www.monergy-project.eu
5
• House #3 a detached house with 2 floors in Spittal an der Drau (AT).
The residents are a mature couple (1 housewife and 1 employed) and an
employed adult son (28 years).
• House #4 a detached house with 2 floors in Klagenfurt (AT). The residents
are a mature couple (1 working part-time and 1 full time), living with two
young kids.
• House #5 an apartment with 2 floors in Udine (IT). The residents are a
young couple, spending most of daylight time at work during weekdays,
although being at home in evenings and weekend.
• House #6 a detached house with 2 floors in Colloredo di Prato (IT).
The residents are a mature couple (1 housewife and 1 employed) and an
employed adult son (30 years).
• House #7 a terraced house with 3 floors in Udine, (IT). The residents are
a mature couple (1 working part-time and 1 full time), living with two
young children.
• House #8 a detached house with 2 floors in Basiliano (IT). The residents
are a retired couple, spending most of time at home.
The device configurations for the selected households are shown in Table 2.
3.3
Measurement setting
Our deployment consists of an ARM-based platform (e.g., Raspberry Pi or
BeagleBone) connected to an Anker Astro E5 15000mAh external battery3 , as
well as a Plugwise Basic kit4 . The Plugwise kit consists of a Zigbee network of 9
sensing outlets, each collecting active power measurements from the connected
load. The collection takes place in epochs. At each epoch, the system collects
power measurements from each node and sleeps for the remaining time. To
provide periodicity, in presence of failures such as mispelled packets, nodes are
skipped and retried at the end of the epoch if there is time. In addition, a
backoff time is used to manage faults, such as the ones resulting from temporary
erroneous states or disconnection of nodes from the network. The developed
daemon uses the open source python-plugwise library and it is freely available
both as a sourcecode and as a ready-to-use SD image5 . The daemon is based on
a collector script and a manager script. The manager is started at boot up by
crontab. Its task is to make sure that a collection script is running and that the
running version is the latest available. To this end, the manager periodically
checks (default is 5 hours) the presence of new versions on the Monergy servers,
and if needed, it replaces the current version with the newest available. In this
way we can push updates to all households without needing to access each and
every unit (see Fig.1). As for the data storage, the current version of the daemon
3 http://www.ianker.com/support-c1-g228.html
4 http://www.plugwise.com
5 http://sourceforge.net/projects/monergy
6
Table 2: Device configurations in the monitored households
House
0
1
2
3
4
5
6
7
8
Devices
Coffee machine, washing machine, radio, water kettle, fridge w/
freezer, dishwasher, kitchen lamp, TV, vacuum cleaner
Radio, freezer, dishwasher, fridge, washing machine, water kettle,
blender, network router
Fridge, dishwasher, microwave, water kettle, washing machine,
radio w/ amplifier, dryier, kitchenware (mixer and fruit juicer),
bedside light
TV, NAS, washing machine, drier, dishwasher, notebook,
kitchenware, coffee machine, bread machine
Entrance outlet, Dishwasher, water kettle, fridge w/o freezer,
washing machine, hairdrier, computer, coffee machine, TV
Total outlets, total lights, kitchen TV, living room TV, fridge w/
freezer, electric oven, computer w/ scanner and printer, washing
machine, hood
Plasma TV, lamp, toaster, hob, iron, computer w/ scanner and
printer, LCD TV, washing machine, fridge w/ freezer
Hair dryer, washing machine, videogame console and radio,
dryer, TV w/ decoder and computer in living room, kitchen TV,
dishwasher, total outlets, total lights
Kitchen TV, dishwasher, living room TV, desktop computer w/
screen, washing machine, bedroom TV, total outlets, total lights
7
implements 4 different modalities: i) local storage as a daily comma separated
value (CSV) file, ii) remote storage on a mysql server with visualization and
quick download (See Fig. 2), iii) a combination of i and ii for double backup, iv)
daily csv file uploaded via sftp to the Monergy server (selected method).
Gateway
(1) JSON samples
(2) Daily CSV file
Dev0
Collector
MONERGY
servers
1Hz
Campaign
Manager
Updated collector
Dev1
..
.
Devn
Figure 1: The sensing infrastructure
Figure 2: The website of the testbed
4
Case studies
The introduction of smart metering produces fine-grained energy consumption
data allowing for the extraction of more valuable information of energy production
and demand compared to having only an energy balance over an extended period
of time. Energy management systems (EMS) can be used to better manage
local consumption and production from renewable sources. Beside providing
consumers with feedback regarding used appliances and their operational costs,
this might also support appliance scheduling in order to minimize costs. In this
section, we report 3 case studies dealing with consumption information: i) load
disaggregation, ii) occupancy detection, and iii) appliance usage modeling.
8
4.1
Non-Intrusive appliance load monitoring
Non-intrusive load monitoring (NILM), also known as load disaggregation, is the
problem to detect and to classify appliances from the total household power draw,
based on specific characteristics of electrical devices. Specifically, NILM tries to
solve the disaggregation of aggregated appliance power loads under the influence
of noise in measurements and models. NILM was initially presented by Hart
[13], although various approaches were later presented. It is generally possible
to distinguish between supervised and unsupervised techniques, depending on
the necessity of labeled data to train the classifier. Nevertheless, there exists no
NILM algorithm able to solve the problem in all its aspects.
Our evaluation is based on the load disaggregator presented in [14], which
uses particle filtering (PF) to estimate the appliance state of all used appliances.
In detail, appliances are model as a hidden Markov model (HMM), whose
states define device states and are associated to the respective power demand.
Furthermore, the appliance models are combined into a fractional hidden Markov
model (FHMM) modeling the total household power demand. The PF is used to
estimate the household power demand according to the given appliance models.
Thereafter, a simple decision marker determines the appliance condition and
operational state, based on thresholding and the given appliance models. For the
evaluation we used the data of house with ID #0 and #2 for 7 consecutive days.
As the PF is mainly dependent the number of used particles, we empirically
identified 1000 as the appropriate number. The aggregated power draw is
composed by 6 different appliances which are listed in Table 3. Beside the
used appliances also the reached accuracy of the proposed load disaggregator is
presented. In this context, the accuracy (ACC) is defined as:
ACC =
TP + TN
,
N
(1)
where T P (true positives) is the number of times an appliance is correctly
detected as ON, T N (true negatives) is the number of times an appliance is
correctly detected as OFF, while N is the number of samples in the observation
window. The results in Table 3 show that the presented dataset can be applied
as a reference dataset to load disaggregation problems. The complexity of the
problem can be adjusted according to the used appliances and corresponding
appliance model (on/off device such as the water kettle or multi-state devices
such as the dishwasher) and the number of aggregated power loads.
4.2
Occupancy detection
Occupancy detection is the problem of inferring presence of people in environments. Many different approaches have been considered in the literature: motion
detection, doors opening, use of acoustic sensors and/or cameras, use of GPS and
localization systems through smart phones. Nevertheless, only a few works propose to use energy consumption data to develop occupancy detection techniques
[15, 16]. To this end, the assessment of occupancy detection requires suitable
datasets, offering either an aggregated or a disaggregated power draw, along with
9
Table 3: Accuracy of load disaggregator for the aggregated power draws of
houses #0 and #2
House ID 0
Type
ACC
TV
0.91
coffee machine
0.99
dishwasher
0.96
fridge
0.93
vacuum cleaner
0.99
water kettle
0.99
House ID 2
Type
ACC
hair dryer
0.99
light
0.97
dishwasher
0.46
fridge
0.90
water kettle
0.99
washing machine
0.80
a description of consumption scenarios. In particular, the presence of user-driven
devices is relevant to assume presence in the environment. It is important to
specify whether residents can postpone the operation of user-driven devices, as
this will condition consumption-based approaches. A simple approach based on
the use of energy consumption data to detect occupancy was presented in [15].
Therein, the authors developed an algorithm called non-intrusive occupancy
monitoring (NIOM). In particular, NIOM guesses the presence of people at home
by comparing the average, the variance and the maximum value of the current
power consumption with threshold values. The threshold values are computed
during inactivity periods, namely when the activity of residents does not add
power consumption to the baseline consumption, e.g., the consumption of devices
such as fridge, HVAC systems etc. Accordingly, the thresholds are computed
every night when people are supposed to sleep. As both power values and
threshold values are computed over time slots of fixed duration, the duration of
the slot is an important parameter affecting the estimation. Although occupancy
detection through energy consumption observation is a cheap and non-intrusive
solution, it is shown being rather inaccurate in determining the occupancy during
periods of inactivity.
In order to give an example on the use of the GREEND dataset to study
occupancy, we consider the house #5. We chose this scenario because we have
measured the total power consumption, which allows for better estimating the
baseline during inactivity periods. We consider two months of measurements
(February and March 2014) and we apply the NIOM algorithm as described in
[15]. The time slot duration is chosen equal to 15 minutes. We consider two cases,
the weekdays (WDs) and the weekend days (WEs). This is because the analyzed
power consumption greatly differs in these two circumstances as it can be seen
in Fig. 3c and Fig. 4b, where the 99th percentile of the energy consumption
during WDs and WEs are shown. The baseline, within which the baseline
threshold values are computed, was chosen at night between midnight and 6.30
AM. Fig. 3 reports the occupancy obtained applying the NIOM algorithm over
three consecutive days. Fig. 3b and Fig. 4a respectively show the probability of
occupancy obtained for WDs and WEs. We first notice that, as expected, the
results obtained between midnight and 6.30 AM do not indicate any activity.We
notice a good agreement between the derived probability of occupancy and the
10
Energy [KW*15mins]
habits of residents. Residents stated that during WDs, they usually wake up at
6.45 AM, leave around 8.00 AM to go to work and are back home around 6.00
PM. During WEs residents tend to spend more time at home during the day
and they go out at night coming back home around midnight.
Occupancy
Energy
2
1
0
0
6 12 18 24 6 12 18 24 6 12 18 24
Hours
Probability
(a) Occupancy over 3 days
1
0.8
0.6
0.4
0.2
0
0
2
4
6
8 10 12 14 16 18 20 22 24
Hours
Energy [KW*15mins]
(b) Weekday occupancy probability
3
2
1
0
0
2
4
6
8 10 12 14 16 18 20 22 24
Hours
(c) Weekday energy
Figure 3: Occupancy detection for weekdays in house #5
11
Probability
1
0.8
0.6
0.4
0.2
0
0
2
4
6
8 10 12 14 16 18 20 22 24
Hours
Energy [KW*15mins]
(a) Weekend occupancy probability
3
2
1
0
0
2
4
6
8 10 12 14 16 18 20 22 24
Hours
(b) Weekend energy
Figure 4: Occupancy detection for weekends in house #5
4.3
Appliance usage modeling
Mining appliance usage patterns concerns the extraction of models describing
how appliances are used by residents, given a sequence describing changes on
their operational state. Given log data describing status changes of electrical
devices, the extraction of models describing the usage of devices can be achieved
using different approaches, such as association rule mining [17], artificial neural
networks (ANNs) [18], episode-generating Hidden Markov models (EGH) [19]
and Bayesian networks [20]. To show a possible application of the dataset, we
propose to use a Bayesian network (BN) to predict the usage of user-driven
devices. A BN is a probabilistic graphical model encoding the joint probability
distribution of a set of random variables. A BN G = (V, E) is a directed acyclic
graph (DAG) whose nodes V = {X1 , . . . , Xn } are random variables, while
dependency between variable is represented by edges such as E = Xi → Xj , and
quantified by the conditional probability P (Xj |Xi ). In particular, each node is
associated to a conditional probability distribution (CPD) quantifying the effect
n
Q
the parents have on the node, that is P (X1 , . . . , Xn ) =
P (Xi |pa(Xi )) [21].
i=1
The approach undertaken follows the work presented in [22, 23] from which
we derived the network structure. Because of the limited length of the dataset at
time of writing we decided to omit seasonal information expressed through the
12
month (see Fig. 5). As a first step, we processed the daily CSV files collected in
the household ID#0 and extracted starting events for each monitored device. The
BN was implemented using the Netica6 tool. In particular, we used expectation
maximization (EM) to perform parameter learning for the given network and we
used the junction tree (JT) algorithm to perform exact inference. Netica provides
APIs for various programming languages, making the learned model applicable in
working solutions. The plot in Fig.6 reports the posterior probability of starting
the coffee machine of household ID#0, given the observations day of the week
(i.e. weekday, Saturday and Sunday) and hour of the day. As noticeable, the
patterns of coffee making tend to be regular over the day. The modeled couple
tends to wake up earlier during weekdays, which was confirmed by an interview
with the householders.
Figure 5: Bayesian network for the coffee machine of House #0
Probability to Start
1
SAT
SUN
WD
0.8
0.6
0.4
0.2
0
0
4
8
12
Hour
16
20
24
Figure 6: Usage forecasting for the coffee machine of House #0
6 https://www.norsys.com
13
5
Conclusions and future work
We have presented the GREEND consumption dataset, obtained through a
measurement campaign in households in Austria and Italy. The dataset is
available for open use and consists of consumption data of selected devices.
Along with the dataset we have also released our code base for the campaign as a
Sourceforge project. This includes the scripts to collect consumption data and a
ready-to-use SD image for the Raspberry Pi, as well as processing scripts that can
be used to extract edges and consumption events, based on configurable devicespecific thresholds. We expect our contribution to be of value for researchers
and engineers dealing with domestic energy management systems. To this
end, we showed the use of GREEND to develop and assess techniques for load
disaggregation, occupancy detection and appliance usage mining. We expect to
extend the monitoring campaign to collect aggregated consumption data, as this
allows for modeling the whole demand, which is a valuable information when
emulating microgrids and assessing control strategies at a wider scope. Similarly,
we plan to include data concerning production from renewable energy, as it can
be of value to forecast energy generation. The presence of temporal information,
along with the locality and therefore the weather, will allow for the extraction of
seasonal consumption patterns. The long term measurement campaign carried
out within MONERGY will provide the time span to allow this type of studies.
We expect to ultimately provide ready-to-use models of the regions, that can be
imported in current simulation tools for smart grid applications.
References
[1] K. C. Armel, A. Gupta, G. Shrimali, and A. Albert, ‘‘Is disaggregation
the holy grail of energy efficiency? The case of electricity,’’ Energy Policy,
vol. 52, no. 0, pp. 213 -- 234, 2013.
[2] A. Monacchi, W. Elmenreich, S. D’Alessandro, and A. M. Tonello, ‘‘Strategies for Domestic Energy Conservation in Carinthia and Friuli-Venezia
Giulia,’’ in Proc. of the 39th Annual Conference of the IEEE Industrial
Electronics Society, Vienna, Austria, 2013.
[3] P. Palensky and D. Dietrich, ‘‘Demand side management: Demand response,
intelligent energy systems, and smart loads,’’ IEEE Trans. on Ind. Informat.,
vol. 7, no. 3, pp. 381--388, Aug 2011.
[4] J. Z. Kolter and M. J. Johnson, ‘‘REDD: A Public Data Set for Energy
Disaggregation Research,’’ in Proceddings of the SustKDD Workshop on
Data Mining Applications in Sustainability, 2011.
[5] K. Anderson, A. Ocneanu, D. Benitez, D. Carlson, A. Rowe, and M. Berges,
‘‘BLUED: a fully labeled public dataset for Event-Based Non-Intrusive load
monitoring research,’’ in Proc. of the 2nd KDD Workshop on Data Mining
Applications in Sustainability (SustKDD), 2012.
14
[6] A. Reinhardt, P. Baumann, D. Burgstahler, M. Hollick, H. Chonov,
M. Werner, and R. Steinmetz, ‘‘On the Accuracy of Appliance Identification Based on Distributed Load Metering Data,’’ in Proc. of the 2nd IFIP
Conference on Sustainable Internet and ICT for Sustainability (SustainIT),
2012.
[7] C. Gisler, A. Ridi, D. Zujferey, O. Khaled, and J. Hennebert, ‘‘Appliance
consumption signature database and recognition test protocols,’’ in 8th International Workshop on Systems, Signal Processing and their Applications
(WoSSPA), May 2013.
[8] S. Makonin, F. Popowich, L. Bartram, B. Gill, and I. V. Bajic, ‘‘AMPds:
A Public Dataset for Load Disaggregation and Eco-Feedback Research,’’
in Proceeding of the IEEE Conference on Electrical Power and Energy
(EPEC), 2013.
[9] N. Batra, M. Gulati, A. Singh, and M. B. Srivastava, ‘‘It’s Different:
Insights into Home Energy Consumption in India,’’ in Proc. of the 5th
ACM Workshop on Embedded Systems For Energy-Efficient Buildings, ser.
BuildSys’13. New York, NY, USA: ACM, 2013.
[10] S. Barker, A. Mishra, D. Irwin, E. Cecchet, P. Shenoy, and J. Albrecht,
‘‘Smart*: An open data set and tools for enabling research in sustainable
homes,’’ SustKDD, August, 2012.
[11] J. Kelly and W. Knottenbelt, ‘‘”UK-DALE”: A dataset recording UK
Domestic Appliance-Level Electricity demand and whole-house demand,’’
ArXiv e-prints, Apr. 2014.
[12] M. Zeifman, ‘‘Disaggregation of home energy display data using probabilistic
approach,’’ IEEE Trans. Consum. Electron., vol. 58, no. 1, pp. 23--31, 2012.
[13] G. Hart, ‘‘Nonintrusive appliance load monitoring,’’ Proc. of the IEEE,
vol. 80, no. 12, pp. 1870--1891, 1992.
[14] D. Egarter, V. P. Bhuvana, and W. Elmenreich, ‘‘Appliance State Estimation Based on Particle Filtering,’’ in Proc. of the 5th ACM Workshop on
Embedded Systems For Energy-Efficient Buildings, ser. BuildSys’13, 2013.
[15] D. Chen, S. Barker, A. Subbaswamy, D. Irwin, and P. Shenoy, ‘‘Nonintrusive occupancy monitoring using smart meters,’’ in Proc. of the 5th
ACM Workshop on Embedded Systems For Energy-Efficient Buildings, ser.
BuildSys’13, 2013.
[16] W. Kleiminger, C. Beckel, T. Staake, and S. Santini, ‘‘Occupancy detection
from electricity consumption data,’’ in Proc. of the 5th ACM Workshop on
Embedded Systems For Energy-Efficient Buildings, ser. BuildSys’13, 2013.
[17] K. J. Kang, B. Ka, and S. J. Kim, ‘‘A service scenario generation scheme
based on association rule mining for elderly surveillance system in a smart
home environment,’’ Engineering Applications of Artificial Intelligence,
15
vol. 25, no. 7, pp. 1355 -- 1364, 2012, advanced issues in Artificial Intelligence
and Pattern Recognition for Intelligent Surveillance System in Smart Home
Environment.
[18] M. Aydinalp, V. I. Ugursal, and A. S. Fung, ‘‘Modeling of the appliance,
lighting, and space-cooling energy consumptions in the residential sector
using neural networks,’’ Applied Energy, vol. 71, no. 2, pp. 87 -- 110, 2002.
[19] N. C. Truong, L. Tran-Thanh, E. Costanza, and D. S. Ramchurn, ‘‘Activity prediction for agent-based home energy management,’’ in Agent
Technologies for Energy Systems (ATES 2013), May 2013.
[20] P. Aguilera, A. Fernández, R. Fernández, R. Rumı́, and A. Salmerón,
‘‘Bayesian networks in environmental modelling,’’ Environmental Modelling
& Software, vol. 26, no. 12, pp. 1376 -- 1388, 2011.
[21] D. Koller and N. Friedman, Probabilistic Graphical Models: Principles and
Techniques - Adaptive Computation and Machine Learning. The MIT
Press, 2009.
[22] L. Hawarah, S. Ploix, and M. Jacomino, ‘‘User behavior prediction in energy
consumption in housing using bayesian networks,’’ in Artificial Intelligence
and Soft Computing, ser. Lecture Notes in Computer Science. Springer
Berlin Heidelberg, 2010, vol. 6113, pp. 372--379.
[23] C. Chang, P.-A. Verhaegen, J. R. Duflou, M. M. Drugan, and A. Nowe,
‘‘Finding Days-of-week Representation for Intelligent Machine Usage Profiling,’’ Journal of Industrial and Intelligent Information, vol. 1, no. 3, pp.
148--154, 2010.
16
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

Related manuals

Download PDF

advertisement