Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

The AAAI-17 Workshop on

Artificial Intelligence for Smart Grids and Smart Buildings


WS-17-05

Energy Disaggregation Methods for Commercial Buildings


Using Smart Meter and Operational Data
Shubham Bansal Mischa Schmidt
NEC Laboratories Europe, Heidelberg, Germany NEC Laboratories Europe, Heidelberg, Germany
École polytechnique fédérale de Lausanne, Switzerland Luleå University of Technology, Sweden
Email: shubham.bansal@alumni.epfl.ch Email: mischa.schmidt@neclab.eu

Abstract Related Work


One of the key information pieces in improving energy effi- The problem of energy disaggregation was first studied by
ciency of buildings is the appliance level breakdown of en- Hart more than two decades back (Hart 1992). The field re-
ergy consumption. Energy disaggregation is the process of mained dormant for a large period since then until recently
obtaining this breakdown from a building level aggregate data when with the proliferation of smart meters, a large amount
using computational techniques. Most of the current research of fine grained electricity data was available for algorithm
focuses on residential buildings, obtaining this information development and testing.
from a single smart meter and often relying on high frequency The approaches that have been developed so far can be
data. This work is directed at commercial buildings equipped broadly classified into event based and non-event based.
with building management and automation systems provid-
ing low frequency operational and contextual data. This pa-
Event based approaches rely on identifying the appliances
per presents a machine learning method to disaggregate en- by capturing features at switching events (Zeifman and Roth
ergy consumption of the building using this operational data 2011). The problem with this approach is that it assumes that
as input features. Experimental results on two publicly avail- only one appliance changes state at a time. If multiple appli-
able datasets demonstrate the effectiveness of the approach, ances switch states at a time, then they will occlude each
which surpasses existing methods. For all but one appliance other’s signature. This problem can be solved by sampling
of House 2 of the publicly available REDD dataset, improve- at high frequencies (1 Hz or more) where the probability
ments in normalized error in assigned power range between of such overlaps is low. Moreover, higher sampling rates al-
20% (Lighting) and 220% (Stove). For another dataset from low extracting features other than the changes in steady state
an educational facility in Singapore, disaggregation accuracy power such as switching time duration, transients & higher
of 92% is reported for the facility’s cooling system.
order harmonics.
Non-event based approaches do not rely on event detec-
Introduction tion to classify appliances. Instead all the power samples are
taken into account for inference. Work in this area has fo-
On a global scale, buildings represent a major share of the cused on modeling appliances as Hidden Markov Models
primary energy use (Cheung 2014; Sbci 2009). Research where the state of the appliance is hidden (Kim et al. 2011;
suggests that significant energy savings in the buildings Parson et al. 2014). The parameters of the model are learned
can be realized by providing direct feedback to the users using expectation-maximization algorithm and Gibbs sam-
regarding their energy consumptions as opposed to indi- pling is used for state estimation (Kim et al. 2011).
rect feedback in the form of monthly bills and weekly rec- Besides these works in single channel source separation,
ommendations (Ehrhardt-Martinez et al. 2010; Wood and researchers have also worked on using other information
Newborough 2003). Such direct feedback entails an under- sources to aid the process of energy disaggregation. Time of
standing of individual appliance consumption. Energy dis- day, day of week and frequency of appliance use have been
aggregation or non-intrusive load monitoring is a compu- used as features in (Kim et al. 2011). Contextual information
tational technique that deals with the task of identifying such as outside temperature and time of day which corre-
the individual appliances’ energy consumption from aggre- lates with the consumption of individual appliances has also
gate building energy consumption. Availability of this in- been utilized by the authors in (Wytock and Kolter 2014).
formation can help design behavioral energy efficiency pro- They formulate an optimization problem to simultaneously
grams, deploy demand response technologies and identify estimate the correlation as well as the source separation. In
faulty equipment. This work is directed at commercial build- (Reyes 2015), authors use location of the user to improve
ings equipped with building management and automation in- performance of the load disaggregation algorithm. This is
strumentation providing operational and contextual data at not useful in our case since we are dealing with a com-
minute intervals. mercial building which has several occupants throughout the
Copyright  c 2017, Association for the Advancement of Artificial working times. In (Kim et al. 2009), authors have used inex-
Intelligence (www.aaai.org). All rights reserved. pensive sensors to capture signals that are correlated with the

233
measure of the goodness fit of an estimated statistical model,
Table 1: Illustrative example of a states table. and the Bayesian Information Criterion (BIC), which not
Time Light 1* Light 2* Microwave Stove Mains only rewards goodness of fit but also penalizes model com-
t1 1 0 0 0 150
plexity. Once the number of clusters are identified, the oper-
t2 1 0 0 1 300
t3 1 0 1 0 370 ational data is clustered and a state is assigned to the appli-
t4 0 1 0 0 200 ance at each instant.
*
Light is presumed to operate in 3 states (with an off-state). Base Algorithm: Multi-State Appliance Energy
Disaggregation
Assume n appliances of which each appliance i can operate
power consumption of the appliances. Subsequently, they in ki states. Thus, the operating state of an appliance  i at
formulate a convex numerical optimization problem to learn an instantt can be given by acolumn
 vector x i (t) = xi1 (t)
T
this mapping. In (Jung and Savvides 2010) and (Beckel et al. ....xiki (t) , where xij (t) ∈ 0,1 is a binary indicator vari-
2012), authors assume that the information about the on/off able indicating the on/off status of the j th state of the appli-
binary state of the appliance at each instant along-with the ance. Note that each appliance can be active only in one state
aggregate power is available and use this information to es- at a time. Therefore, power consumed by the device i at an
timate the average power consumption of the appliance in instant t is given by
each state.
pi (t) = Pi xi (t) (1)
Contributions and Problem Definition  
where Pi = Pi1 ....Piki and Pij is the power consump-
The present work builds on the work of Jung and Savvides
(Jung and Savvides 2010) to infer the energy consumption tion of the ith device in the j th state.
of individual appliances by using the information about their Since sum of the individual power consumptions should
operating state. In their study, the authors assume the infor- be equal to the total power consumption of the building
mation about the on/off state of the appliances to be available ptotal (t), a numerical optimization problem can be formu-
directly from sensors placed near the appliances. However, lated as:
this might not always be possible. Moreover, appliances may  T  n 2
have multiple modes of operation which would be harder min ptotal (t) − pi (t)
to detect for such sensors. The major contributions of this t=1 i=1
(2)
work are: 1) Infer the active state of the appliance from its subject to : Pi ≥ 0∀i
operational and contextual data 2) Use the state information
to perform disaggregation through an improved algorithm where ptotal (t) is the aggregate power consumption
3) Further improve the algorithm to make it robust against recorded by the smart meter. This is a convex optimization
the presence of appliances for which state information is not problem and gives the average power consumption vector
available. Pi of an appliance. This estimate can be combined with the
The problem of energy disaggregation can be formu- state information to obtain the power consumption of an ap-
lated as follows: given a time-stamped vector of appliance pliance at each instance using equation 1. Note that here the
states and an aggregate power consumption data, estimate L2-norm is used, but other norms may also be studied.
the power consumption of the appliance in each state over This is an extension of the approach described in (Jung
a given time interval. An illustrative states table is shown in and Savvides 2010) where appliances were modeled as only
Table 1 for three appliances. binary devices. As described in (Jung and Savvides 2010),
instead of solving equation 2, a weighted least square opti-
Algorithms for Energy Disaggregation mization is solved by giving higher weights to those instants
t, where fewer appliances are active as shown in equation 3.
This section introduces a multi-state energy disaggregation
algorithm based on (Jung and Savvides 2010) and two algo-
rithmic extensions to improve performance. The algorithm’s
T
  n
 2
prerequisite is information about the evolution of appliance min w(t) ptotal (t) − pi (t)
t=1 i=1
(3)
states over time. This state information may stem from ex-
pert knowledge or can be learned from clustering algorithms subject to : Pi ≥ 0∀i
applied to operational or contextual data available. To verify
the proposed approach on publicly available datasets sev- Extension 1: Average Power Assisted Optimization
eral different clustering algorithms were tested, out of which The approach mentioned above estimates the average power
k-means and Gaussian mixture model (GMM) reported the consumption of the appliances in different states for the
best results. Both these clustering techniques take number given interval. But it essentially discards the information that
of cluster centres as a hyper-parameter. This number may the sum of individual appliance power consumption should
be learned from experts, but this work inferred the optimal be equal to the aggregate power at all instances. Addition-
number of clusters using two well known information crite- ally, often information about the maximum power of the ap-
ria from statistics: the Akaike Information Criterion (AIC), a pliances may be available which can assist in disaggregation.

234
Equation 4 includes this information in a convex optimiza-
tion problem.
n 
T 
 2
min p̂i (t) − pi (t)
t=1 i=1
subject to : 0 ≤ p̂i (t) ≤ pmax,i (4)
 n
p̂i (t) = ptotal (t)
i=1

where pi (t) is obtained from equation 3 using equation 1,


p̂i (t) is the estimated power of the appliance i at instant t
and pmax,i is its maximum power.

Extension 2: Average Power Assisted Optimization


with Ghost Appliance Figure 1: Breakdown of the total power consumption by ap-
The limitation of the above formulation is that often in a pliances for House 2 of the REDD dataset for 5 weeks.
typical building operational state of all the appliances is not
available. These devices can introduce error in the estima-
tions by spreading the consumption of these devices over REDD Dataset The Reference Energy Disaggregation
the rest of the appliances. To limit these errors, an artificial Dataset (REDD) (Kolter and Johnson 2011) contains power
appliance (also called the ghost appliance) is introduced to data from 6 homes in Boston area collected in the summer of
represent all the un-instrumented appliances in the problem. 2011. It provides the mains as well as appliance power data
at high frequency (sampled at 15 kHZ) and low frequency
n 
T  2   (sampled at 1Hz for mains and 0.3 Hz for appliances) in the
 form of .csv files. We performed our study on House 2 so as
min p̂i (t) − pi (t) + p̂ghost (t)
to be able to compare our results with those reported in the
t=1 i=1
literature (Batra, Dutta, and Singh 2013). This house con-
subject to : 0 ≤ p̂i (t) ≤ Pmax,i tains data from 11 channels or meters including 2 mains and
0 ≤ p̂ghost (t) ≤ Pmax,ghost 9 appliances for 5 weeks.
 n The pie chart in figure 1 shows the contribution of the dif-
p̂i (t) + pghost (t) = ptotal (t) ferent appliances to the aggregate load. As noted in (Batra,
i=1 Dutta, and Singh 2013), washer dryer and disposal were re-
(5) moved from the analysis because of their low peak power
In equation 5, a L1-norm penalty has been introduced on consumption of 8 W and less than 0.1 % contribution to
the power allocated to the ghost term. In its absence, the op- the overall consumption respectively. Figure 1 indicates the
timization problem would have allocated all the power to the presence of non-metered appliances.(i.e, ghost appliances).
ghost giving out a trivial solution. The penalty term estab- Before the algorithm could be run on the data, certain
lishes a trade-off between the ghost power being completely preparation steps had to be carried out. In the dataset, the
ignored and all the power being allocated to it. appliance data starts 6 hours after the mains data and so
time series synchronization was performed to align them.
Metrics and Evaluation Second, the data was down-sampled to 1 min using a mean
filter to enable a fair comparison with the results of (Ba-
Datasets
tra, Dutta, and Singh 2013) and any gaps in the data were
To demonstrate and evaluate the proposed approaches, two forward filled. To demonstrate the effectiveness of the pro-
different datasets are used: posed approach in disaggregating low frequency data, data
• The REDD dataset contains electricity data from house- has also been down-sampled to 10-min resolution. Note that
holds & has been extensively used for building and testing there were two channels corresponding to the kitchen outlets
energy disaggregation algorithms (Kolter and Jaakkola which have been merged into one for this analysis. Also,
2012; Batra, Dutta, and Singh 2013; Parson et al. 2014). there were two mains channels in the dataset correspond-
Thus, it serves as a good benchmark to compare against. ing to the split phase power supply. Since the assignment of
appliances to each mains is not known both the values are
• The dataset from United World College (UWC) contains added to create one aggregate mains for the entire house.
electricity and building management system data from an
international school in Singapore. It was publicly released United World College Dataset The focus of this study is
by Miller et.al (Miller, Nagy, and Schlueter 2014) in 2014 to develop non-intrusive load monitoring techniques which
and can be exploited to build energy disaggregation tools can be used for commercial buildings. To validate that op-
for commercial buildings. erational data can be used to infer the appliance states and

235
assist in disaggregation, another publicly available dataset
from United World College of South-east Asia, an interna- Table 2: Power states of appliances estimated for House 2
Appliances Power for state 1 (W) Power for state 2 (W)
tional school located in Singapore is used. This building is
Refrigerator 156 394
certified at the Platinum level of the Singapore Green Mark Lighting 25 148
Building rating scheme. The dataset contains almost 3 years Dishwasher 201 1183
of time series data from 105 electrical sub-meters, 633 air- Microwave 54 1375
side systems, 385 light points and 414 water-side systems. Stove 67 374
The campus is fitted with a 5630 kW cooling system com- Kitchen 29 924
prised of 3 electrical chillers (Miller, Nagy, and Schlueter
2014). Cooling takes up a significant chunk of the energy
consumption of the campus and the building management Table 3: Performance for different algorithms for 1-min data.
systems tracks a number of different operational variables Metric Appliances
Simple Literature* Multi- PAO**
Mean Results State W/o With
related to cooling system. Its chiller plant comprises of three ghost ghost
chillers. In this study, the power consumption of the individ- NEP (%) Refrigerator 65.1 25 42.9 45.5 38.7
Lighting 102 63 44.9 43.1 42.4
ual chillers is inferred from the aggregate meter data for the Dishwasher 188.4 73 10.5 124.4 12.5
chiller system. The operational variables available for the Microwave 117.4 113 77.7 89.6 65.0
Stove 163.0 271 59.8 739.0 51.8
chillers are flow rate, condenser supply & return tempera- Kitchen outlets 131.9 165/100 85.3 118.9 75.5
tures and chilled water supply & return temperatures. The Accuracy (%) - 69.9 NA 83.9 77.6 85.5
*
available data is sampled at 3 minute intervals. Lots of gaps Reported in (Batra, Dutta, and Singh 2013). ** PAO stands
and drops in different channels were noticed and hence it for Power Assisted Optimization
was decided to use the data from Nov.11,2011 - Dec.31,2011
period for this analysis as it was of higher quality. Any gaps
in the data were forward filled. tify power levels of the appliance. Subsequently, the closest
power level at each instance is assigned to the appliance.
Metrics Table 2 shows the states obtained via simple k-means clus-
Different performance measures can be used depending on tering with two clusters. Also note that appliances are con-
the application area of the disaggregation algorithm. For us sidered to be off when their power consumption is below 10
it was important to understand how well the algorithm is W. A similar strategy has been used in the NILMTK tool kit
able to disaggregate the energy consumption at each instance (Batra et al. 2014). Based on these power levels, appliance
for every appliance and also the overall performance of the state vectors similar to that in table 1 are constructed.
algorithm. Batra et al. defines Net Error in assigned Power Table 3 summarizes the results achieved with the pro-
(NEP) as an appliance level error measure as follows: posed algorithms. It compares these with the results reported
T in (Batra, Dutta, and Singh 2013) and a simple mean predic-
|p̂i (t) − pi (t)| tion heuristic. The simple mean algorithm predicts at each
NEP = t=1 T (6) time instant a breakdown proportional to the overall proba-
t pi (t)
bility of the signal. This serves as the base case performance
where pi (t) is the measured and p̂i (t) is the estimated measure obtained from naive intuition. Note also that here
power of the appliance i at instant t. It measures the devia- in our implementation we have combined both the kitchen
tion of estimated power from the actual power in each sam- outlets into one appliance.
pling interval. Lower NEP indicates a higher performance in
Table 3 shows that the multi-state algorithm performs bet-
estimating the appliance’s consumption.
ter than the benchmark study for all the appliances except for
Apart from this appliance level error measure, we also
the refrigerator. For the power level assisted optimization,
want a single metric that can capture the overall perfor-
the performance degrades for the variant which does not
mance of the algorithm. For this, the Total energy correctly
have a ghost component in the optimization. This is because
assigned as defined in (Kolter and Johnson 2011) is used:
in this case the algorithm forces the constraints that the sum
T n of the appliance powers is equal to the mains power whereas
t=1 i=1 |p̂i (t) − pi (t)| as the chart in figure 1 shows, about 35 % of the power is un-
Accuracy = 1 − T (7)
2 t=1 ptotal (t) accounted. Hence, when the ghost appliance is included per
equation 5, the performance increases beyond that of multi-
where ptotal (t) is the aggregate consumption for the state approach. Compared to the benchmark, power level as-
building at instant t. sisted optimization achieves the following results: while the
The higher the total energy correctly assigned is (hence- refrigerator worsens by 13.7%, all other appliances improve
forth referred to as accuracy), the better is the overall per- between 20.6% (Lighting) - 219.2% (Stove).
formance of the algorithm.
The overall performance of the algorithms is compared
using the accuracy measure defined earlier. As is clear from
Results and Discussion table 3, power level assisted optimization performs bet-
Experiments on REDD Household Dataset ter with the inclusion of the ghost appliance and outper-
In the REDD dataset, operational variables are not avail- forms the multi-state algorithm. Moreover the performance
able. Hence, appliance level sub-meter data is used to iden- is higher by about 15% over the simple mean algorithm.

236
Figure 2: Effect of unaccounted power (ghost appliances) on
the disaggregation performance for House 2 of REDD

Figure 4: Predicted versus measured power signal of all the


appliances obtained from the power assisted optimization
with ghost approach for REDD dataset

Table 4: Comparison of performance of the algorithms on


REDD dataset down-sampled to 10 min resolution
Simple Multi- PAO*
Metric Appliances
Mean State Without With
ghost ghost
NEP (%) Refrigerator 59.8 46.4 48.6 42.9
Figure 3: Histograms of appliance power consumption Lighting 96.6 48.5 41.4 44.4
Dishwasher 188.0 27.6 136.8 25.7
Microwave 111.2 87.5 103.2 74.7
Stove 153.6 64.1 757.9 52.7
Next, to investigate the effect of ghost power on the per- Kitchen outlets 121.3 77.9 112.6 68.6
formance of the different algorithms, we select the refrig- Accuracy (%) - 74.3 82.7 76.6 84.4
erator as a single base appliance and progressively remove *
PAO stands for Power Assisted Optimization
the state information corresponding to the other appliances.
The choice of the appliance whose state information is re-
moved is based on its percentage in the overall household microwave and kitchen outlets are very high because most
consumption. Figure 2 shows the overall accuracy and NEP of the time they operate in the low power states as is clear
for refrigerator for the different algorithms as the state infor- from their histogram in figure 3 with occasional very high
mation of the appliances is progressively removed causing peaks. Because of this skewed distribution, the clustering al-
increasing amount of ghost power. It is clear from the figure gorithm does not learn good cluster centres for their high
that performance for the power level assisted optimization power states. Finally, figure 4 shows the measured and esti-
remains stable when ghost appliance is modeled in the op- mated power profiles for all the appliances over two days.
timization until un-instrumented appliances account for as The above analysis was performed on data sampled at 1
much as 53% of the total consumption. This robustness is min intervals. Since, the objective of this project is to de-
very important as in a typical building state information may velop approaches for low frequency data which is usually
not be available for all the appliances. the case with the commonly installed smart meters, the per-
After comparing the disaggregation performance across formance results of all the algorithms for data sampled at
algorithms, we now discuss disaggregation results across 10-min intervals are presented in table 4. As is clear from
different appliances for the best performing algorithm power the data, there is an expected drop in the performance from
level assisted optimization with ghost. It can be seen that the 85.5% to 84.4% but still the accuracy is much better than
performance is particularly good for refrigerator and dish- the baseline of 74% for the power assisted optimization with
washer with about 39% and 13% error respectively. This ghost appliance approach. This indicates that the approaches
can be ascribed to both these appliances having a repeatable presented in this work can be applied to disaggregate low
pattern in power consumption. The error for lighting is on frequency smart meter data.
the higher end because the 3-state model (including the off
state) might not be appropriate for it. The power histogram Experiments on UWC: Commercial Building Data
for lighting in figure 3 confirms this fact which shows that The United World College dataset provides data for several
5 state model might be more suitable. Its a similar case for operational variables of 3 chillers along-with the individual
the stove which appears to be a 4 state device. The errors for and aggregate chiller power data. As no literature bench-

237
Table 5: Comparison of disaggregation results using differ-
ent approaches across the chiller system
Simple Best state Multi
Metric Appliance PAO*
Mean information State
Chiller 1 60.9 4.4 11.3 9.3
NEP (%) Chiller 2 85.8 4.4 24.6 20.8
Chiller 3 117.1 5.1 19.6 18.3
Accuracy (%) - 56.1 97.6 90.9 92.1
*
Figure 5: Scatter plots for the operational data for chiller 1. PAO stands for Power Assisted Optimization
Different colors represent different power levels identified
from the sub-meter data.

marks are available to compare against, the best-case per-


formance of the proposed approach is estimated. For this,
a states table is constructed by binning each chiller’s sub-
meter power data using 4, 5 and 5 bins respectively. This
states table is then used for disaggregation using power level
assisted approach. This represents the case when informa-
tion on the true power state is available for each appliance
and hence serves as the upper performance limit for the pro-
posed algorithms. As second benchmark, the simple mean
heuristic is used again. Figure 6: Predicted versus measured power consumption
After establishing the benchmarks, the chillers’ states profiles of the chillers in the UWC campus over a typical
need to be inferred from the operational data. The opera- week.
tional data relevant to the chillers is clustered using GMM.
This clustering step groups together data points which often
occur together indicating that they might be associated to the data of the appliances to assist in energy disaggregation.
same power state. But before that the data is normalised (by
deducting the mean and dividing by the standard deviation).
Figure 5 shows the scatter plots for the operational data for Conclusions and Future Work
chiller 1. From the figure, it is clear that the points corre-
sponding to the off-state (0 W) are clearly separable from Inspired by related work (Jung and Savvides 2010), a novel
the rest. Moreover, clustering algorithm should also be eas- approach to obtain the breakdown of electricity consump-
ily able to identify the cluster corresponding to the power tion of commercial buildings was presented in this work,
level of 216 kW from the points with ≥ 250kW power. relying on low frequency operational information available
It is important to note that clustering henceforth has from building management systems. An extensive evalua-
been done in an unsupervised framework without using the tion of the proposed methods were performed on REDD, a
ground truth consumption data for the individual chillers. publicly available dataset and its results were compared to
Investigating the AIC and BIC measures yielded the opti- those reported in literature. For all but one appliance a clear
mal number of clusters for the three chillers: 6 (chiller 1), improvement in performance is reported. An overall disag-
5 (chiller 2) and 6 (chiller 3). After building the states table gregation accuracy of 85.5% is achieved in the best case.
with these clusters, the multi-state and power-level assisted Applied to the publicly available dataset of the United World
disaggregation algorithms are applied. Table 5 summarizes College in Singapore, a energy disaggregation accuracy of
the results obtained along-with with the benchmarks (perfect 92% is achieved for the chiller equipment. This is 5.5% be-
state information and mean prediction heuristic). low a best case benchmark and 36% better than a naive esti-
Compared to the naive mean prediction heuristic, power mation algorithm.
assisted optimization improved accuracy by around 36%. Future work will focus on alternative clustering methods
Also, the NEP significantly improved across all the chillers. so that the cluster centers of operational data better corre-
A slight improvement of 1% in the performance for power spond to the power states of the appliances. To obtain good
assisted optimization over a multi-state approach is also ob- clusters representative of appliance states, feature synthe-
served. This improvement can be attributed to the second sis is of paramount importance. Expert knowledge about
level of optimization. Compared to the benchmark (state in- the appliances in a typical commercial building can be used
formation deduced from chillers individual power meters), for constructing such features. Alternatively, we plan to ex-
power assisted optimization falls short only by 5.5%. Figure plore deep neural networks to simultaneously learn the fea-
6 shows the estimated profiles along-with the ground truth tures and cluster mapping as has been proposed in (Xie, Gir-
data for a typical week. The estimated power profile for all shick, and Farhadi 2016). This can improve the scalability of
the three chillers traces the measured profile very closely. our approach by eliminating the need for expert knowledge
This proves the concept of using low frequency operational about the appliances on the building premises.

238
References Sbci, U. 2009. Buildings and climate change: Summary
Batra, N.; Kelly, J.; Parson, O.; Dutta, H.; Knottenbelt, W.; for decision-makers. United Nations Environmental Pro-
Rogers, A.; Singh, A.; and Srivastava, M. 2014. NILMTK: gramme, Sustainable Buildings and Climate Initiative, Paris
an open source toolkit for non-intrusive load monitoring. In 1–62.
Proceedings of the 5th international conference on Future Wood, G., and Newborough, M. 2003. Dynamic
energy systems, 265–276. ACM. energy-consumption indicators for domestic appliances: en-
Batra, N.; Dutta, H.; and Singh, A. 2013. INDiC: Improved vironment, behaviour and design. Energy and buildings
Non-intrusive Load Monitoring Using Load Division and 35(8):821–841.
Calibration. 79–84. IEEE. Wytock, M., and Kolter, J. Z. 2014. Contextually supervised
Beckel, C.; Kleiminger, W.; Staake, T.; and Santini, S. 2012. source separation with application to energy disaggregation.
Improving device-level electricity consumption breakdowns CoRR abs/1312.5023.
in private households using on/off events. ACM SIGBED Xie, J.; Girshick, R. B.; and Farhadi, A. 2016. Unsu-
Review 9(3):32–38. pervised deep embedding for clustering analysis. CoRR
Cheung, H. 2014. Detailed energy data collection for mis- abs/1511.06335.
cellaneous and electronic loads in a commercial office build- Zeifman, M., and Roth, K. 2011. Nonintrusive appliance
ing. 2012 ACEEE Summer Study on Energy Efficiency in load monitoring: Review and outlook. IEEE Transactions
Buildings, Asilomar, CA. on Consumer Electronics 57(1):76–84.
Ehrhardt-Martinez, K.; Donnelly, K. A.; Laitner, S.; et al.
2010. Advanced metering initiatives and residential feed-
back programs: a meta-review for household electricity-
saving opportunities. American Council for an Energy-
Efficient Economy Washington, DC.
Hart, G. W. 1992. Nonintrusive appliance load monitoring.
Proceedings of the IEEE 80(12):1870–1891.
Jung, D., and Savvides, A. 2010. Estimating building con-
sumption breakdowns using on/off state sensing and incre-
mental sub-meter deployment. In Proceedings of the 8th
ACM Conference on Embedded Networked Sensor Systems,
225–238. ACM.
Kim, Y.; Schmid, T.; Charbiwala, Z. M.; and Srivastava,
M. B. 2009. ViridiScope: design and implementation of
a fine grained power monitoring system for homes. In Pro-
ceedings of the 11th international conference on Ubiquitous
computing, 245–254. ACM.
Kim, H.; Marwah, M.; Arlitt, M. F.; Lyon, G.; and Han,
J. 2011. Unsupervised Disaggregation of Low Frequency
Power Measurements. In SDM, volume 11, 747–758. SIAM.
Kolter, J. Z., and Jaakkola, T. 2012. Approximate infer-
ence in additive factorial hmms with application to energy
disaggregation. In International conference on artificial in-
telligence and statistics, 1472–1482.
Kolter, J. Z., and Johnson, M. J. 2011. REDD: A public
data set for energy disaggregation research. In Workshop on
Data Mining Applications in Sustainability (SIGKDD), San
Diego, CA, volume 25, 59–62. Citeseer.
Miller, C.; Nagy, Z.; and Schlueter, A. 2014. A seed dataset
for a public, temporal data repository for energy informatics
research on commercial building performance. In In pro-
ceeding of: 3rd Conf. on Future Energy Business & Energy
Informatics, Rotterdam, Netherlands, 1–2.
Parson, O.; Ghosh, S.; Weal, M.; and Rogers, A. 2014.
An unsupervised training method for non-intrusive appli-
ance load monitoring. Artificial Intelligence 217:1–19.
Reyes, A. R. L. 2015. Location-aware Energy Disaggre-
gation in Smart Homes. Ph.D. Dissertation, TU Delft, Delft
University of Technology.

239

You might also like