Actual Consumption Estimation Algorithm for Occupancy Detection
using Low Resolution Smart Meter Data
Shunichi Hattori and Yasushi Shinohara
Central Research Institute of Electric Power Industry, 2-11-1 Iwadokita, Komae-shi, Tokyo, Japan
Keywords:
Occupancy Detection, Smart Meter, Electricity Consumption, Non-intrusive Load Monitoring.
Abstract:
This paper proposes an actual consumption estimation algorithm that achieves highly accurate occupancy
detection using electricity consumption data derived from smart meters. In Japan, electricity consumption
data on households will soon be available because smart meters, which enable electric power companies to
monitor how much electric power people are using in each household, have been installed in all households.
Occupancy detection is a major technique that leverages electricity consumption data and can be applied
to various services such as ambient assisted living, sales promotions, and peak load shifting. However, it
is difficult to conduct high-accuracy occupancy detection using electricity consumption data automatically
derived from smart meters because of their low resolution: 30-min intervals and 100 Wh increments. An
actual consumption estimation algorithm is therefore proposed to generate data that reflects the characteristics
of a household’s state from low-resolution smart meter data. Occupancy detection is implemented using the
estimated consumption data, which are generated by the proposed algorithm, and the results of experiments
show that its performance is improved compared to the result obtained using raw smart meter data.
1 INTRODUCTION
In Japan, the restructuring of the electric power mar-
ket began in the 1990s and full deregulation of the
electric power market started in April 2016. Whole
new services creating additional value for the sales
departments are desired because this deregulation
brings intensive competition from new entrants for the
existing electric power companies.
At the same time, network-connected and remote-
controlled electricity meters called smart meters have
been installed in all households for the purpose of au-
tomated meter reading. The standard communication
method with a smart meter is called the A-route and
it enables electric power companies to monitor how
much electric power people are using in each house-
hold. That is, electricity consumption data for all
households are becoming available without any ad-
ditional cost through the use of A-route. Analysis tar-
geting the consumption data derived from the A-route
is therefore a promising approach for developing use-
ful and competitive services for electric power com-
panies.
Occupancy detection (Chen et al., 2013;
Kleiminger et al., 2015; Molina-Markham et al.,
2010) leveraging electricity consumption data is
known as one of the major non-intrusive load mon-
itoring techniques (Hart, 1992). This technique
was mainly developed in Europe and the United
States for the purpose of efficient control of Heating,
Ventilation, and Air Conditioning (HVAC) systems.
Moreover, occupancy detection is also considered
to be applicable in various useful services such as
ambient assisted living (AAL), sales promotions,
peak load shifting, and delivery route optimization.
However, it is difficult to conduct high-accuracy
occupancy detection using electricity consumption
data derived from the A-route, because the A-route
readout is low resolution data: 30-min intervals and
100 Wh increments (Komatsu and Nishio, 2015;
Nomura et al., 2014). The data contain phantom
demand variations because of these characteristics
and they degrade the performance of occupancy
detection using machine learning algorithms.
An actual consumption estimation algorithm is
therefore proposed to leverage low-resolution A-route
readout for occupancy detection. Occupancy detec-
tion is implemented using the estimated consumption
data generated by the proposed algorithm and its ac-
curacy, precision, and recall are improved better than
those of results obtained using raw A-route readout.
Related work about non-intrusive load monitor-
Hattori S. and Shinohara Y.
Actual Consumption Estimation Algorithm for Occupancy Detection using Low Resolution Smart Meter Data.
DOI: 10.5220/0006129400390048
In Proceedings of the 6th International Conference on Sensor Networks (SENSORNETS 2017), pages 39-48
ISBN: 421065/17
Copyright
c
2017 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
39
ing, occupancy detection, and its applications are re-
viewed in Section 2. Section 3 proposes the al-
gorithm that estimates actual electricity consump-
tion data from low-resolution A-route readout data.
The feasibility of the proposed algorithm is discussed
based on the performance of the occupancy detection
in Section 4. Section 5 concludes the paper.
2 RELATED WORK
2.1 Non-intrusive Load Monitoring
The approach used to analyze the change of
appliance-level electricity consumption from aggre-
gated data is called disaggregation (Froehlich et al.,
2011; Zoha et al., 2012). Disaggregation is catego-
rized into non-intrusive load monitoring, which ana-
lyzes loads without any sensor or measuring equip-
ment inside buildings except for the aggregated elec-
tricity consumption meter (Hart, 1992).
Disaggregation is a well-studied problem and
there are many studies that deal with it based on
various characteristics and methods such as har-
monics (Nakano and Murata, 2007), neural net-
works (Chang et al., 2010), and factorial hidden
Markov models (Batra et al., 2014; Kim et al., 2011),
which is an expansion of the hidden Markov model.
It can be said that disaggregation is an approach
that is similar to occupancy detection in terms of ac-
tivity annotation on households or buildings. If such
annotated data regarding appliances are obtained,
the performance of occupancy detection algorithms
should be improve. In addition, more detailed state in-
formation such as sleep, bathing, and cooking behav-
ior might be classifiable. However, these approaches
require electricity consumption data with high fre-
quency, generally from 1 kHz to 100 MHz, because
they try to extract the features of appliance-level con-
sumption based on the characteristic pattern or elec-
trical noise of each appliance (Zoha et al., 2012). This
paper does not focus on this approach because the
proposed method is designed for electricity consump-
tion data with low resolution such as A-route readout.
2.2 Occupancy Detection
Occupancy detection is a technique that estimates the
occupancy state of commercial or residential build-
ings (e.g., whether a household is occupied or unoc-
cupied). This technique was mainly developed for the
purpose of energy-efficiency optimizations for HVAC
control and can be categorized into two approaches:
intrusive and non-intrusive. In the former approach,
various methods using environmental sensors such
as motion (passive infrared), CO
2
, acoustic informa-
tion, and doors (magnetic contact) have been pro-
posed (Nguyen and Aiello, 2013). In contrast, elec-
tricity consumption data are mainly employed in the
latter approach (Chen et al., 2013; Kleiminger et al.,
2015; Molina-Markham et al., 2010).
Non-intrusive occupancy monitoring has become
a well-studied problem since Molina-Markham et al.
suggested that occupancy state can be classified based
on aggregated electricity consumption data (Molina-
Markham et al., 2010). The advantage of the non-
intrusive approach is that it detects occupancy without
the need to install any sensors in buildings. Nowa-
days, it is a major research field in some regions of
Europe and the United States, where smart meters
have been installed before the rest of the world.
Fig. 1 shows the occupancy detection procedure
using electricity consumption data based on classifi-
cation criteria. Occupancy is generally detected using
the following steps.
Step 1: Extract features of a training set T =
{t
1
,t
2
, . . . } for discriminating occupancy
states and generate a classification criteria
based on the features.
Step 2: Input test set T
0
= {t
0
1
,t
0
2
, . . . } to the classifi-
cation criteria generated in Step 1.
Step 3: Obtain the set of classified occupancy states
for T
0
.
Using the procedure shown in Fig. 1, Chen et al.
proposed a threshold-based method that classifies oc-
cupancy state using simple criteria such as maximum
value, standard deviation, and range (the change be-
tween maximum and minimum value) (Chen et al.,
2013). Beckel et al. employed machine learning
algorithms such as support vector machine (SVM)
and hidden Markov models for state classifica-
tion (Kleiminger et al., 2015). They also provide the
ECO data set,
1
which includes pairs of 1 Hz electric-
ity consumption data and the occupancy state ground
truth of five Swiss households (Beckel et al., 2014).
These studies target electricity consumption data
whose frequency ranges from 1-s to 1-min intervals.
A method targeting low-resolution data, such as smart
meter data in Japan, has not yet been studied. The de-
tailed characteristics of smart meter data communica-
tion in Japan, which is called A-route, are described
in the next section.
1
https://www.vs.inf.ethz.ch/res/show.html?what=eco-data
[Accessed 4 October 2016]
SENSORNETS 2017 - 6th International Conference on Sensor Networks
40
Classifier
′
′
′
′
′
′
Electricity consumption
Test set
Occupied
Unoccupied
Occupied
Occupancy state
′ ′ ′ ′ ′ ′
Classified occupancy state of ′
Step1: Generate classifier
with the features of
Step3: Output
classified state
Step2: Input test set ′
to generated classifier
Occupied
Unoccupied
Electricity consumptionOccupancy state
Training set
Occupancy state of (Ground truth)
Figure 1: Procedure of occupancy detection from electricity consumption data based on classification criteria.
Table 1: Specification and Characteristics of Smart Meter Data in Japan.
A-route
Route Specification Advantages
From 5-s to 5-min intervals
(depending on equipment)
1 W increments
Disadvantages
B-route
30-min intervals
100 Wh increments
High-resolution data
Needs no dedicated equipment
Data on all households are
available
Need dedicated equipment to
obtain consumption data
Low-resolution data
Contain phantom power
variations
2.3 Smart Meters
A smart meter is a network-connected and remote-
controlled digital meter that monitors electricity con-
sumption online for meter reading. These meters
have been installed in some regions of Europe and the
United States. In Japan, smart meters will be installed
in all households as of 2023. Although the main pur-
pose of the installation is automated meter reading for
billing, smart meters also enable electric power com-
panies to obtain time series data regarding electricity
consumption for each household. Occupancy detec-
tion targeting households employing smart meter data
is therefore a promising approach to developing vari-
ous competitive services.
There are two communication routes, called A-
route and B-route, that are used with smart meters in-
stalled in Japan. Table 1 shows the specification and
characteristics of these routes. In A-route, electricity
consumption data of all households can be obtained
automatically through the power distribution sectors
of the electric power companies, although they are
low frequency and coarse sampling data (Komatsu
and Nishio, 2015; Nomura et al., 2014). In con-
trast, the route that acquires electricity consumption
data through the home area network is called B-route.
Whereas data with higher resolution are derived from
B-route, dedicated equipment such as a home en-
ergy management system (HEMS) is required to uti-
lize this route. Hence, the utilization of B-route is
currently totally impractical because HEMS is gen-
erally expensive equipment (it costs approximately
USD 1,500 as of 2016).
Hence, data acquisition using A-route is easy and
economical because it needs no extra equipment such
as HEMS. However, utilization of A-route data must
take into account its specification: low frequency and
coarse sampling. As shown in Table 1, A-route is
30-min intervals and 100 Wh increments consump-
tion data. In A-route, the fractions of meter readouts
less than 100 Wh are rounded off and carried over to
the next readout in 30 min.
Actual Consumption Estimation Algorithm for Occupancy Detection using Low Resolution Smart Meter Data
41
0
500
1000
1500
2000
2500
9:00 9:30 10:00 10:30 11:00 11:30
0
100
200
300
(b) A-route readout in 30-min intervals and 100 Wh increments
9:30
10:00 10:30 11:00 11:30
Time
12:00
Electricity consumption (W)
Electricity consumption(Wh)
(a) B-route data in 1-min intervals and 1 W increments
Time
Figure 2: Example of smart meter data: (a) B-route and (b)
A-route data.
Figs. 2(a) and (b) show an example of electricity
consumption data over the same period and house-
hold via B- and A-routes, respectively. The household
can be assumed unoccupied after 10:00 because the
consumption and its variation are relatively low com-
pared to those before 10:00. However, the A-route
readout shown in Fig. 2(b) oscillates between values
of 0 and 100 Wh after 10:30 because of the round off
and carry over. These characteristics are fatal for oc-
cupancy detection, which classifies occupancy state
based on features such as standard deviation, because
this phantom fluctuation tends to be regarded as the
characteristics of an occupied state. A method for es-
timating actual consumption which reduces the phan-
tom fluctuation during the unoccupied state is there-
fore needed to detect occupancy with high accuracy
using smart meter data.
2.4 Applications of Occupancy
Detection
Although occupancy detection has mainly been ap-
plied to building automation systems such as HVAC
control in Europe and the United States (Nguyen and
Aiello, 2013), some additional applications are ex-
pected.
AAL.
AAL is a key technology for tackling the prob-
lems of an aging society. Rashidi et al. indicated
that it is imperative to develop AAL technologies
that help older adults in their own homes (Rashidi
and Mihailidis, 2013). Non-intrusive AAL is ex-
pected to be based on occupancy detection be-
cause occupancy detection enables us to moni-
tor the daily behavior of older adults without in-
stalling any equipment or medical devices.
Sales Promotion.
Advertisement delivery systems are another
promising way to exploit occupancy detection
technology. Understanding the lifestyle in each
household could enable just-in-time ad serving,
although it might elicit a negative response re-
garding privacy issues.
Peak Load Shifting.
Mitigating electrical peak demand has been a crit-
ical issue for electric power companies because
it improves the efficiency of electric power gen-
eration and reduces generation capacity require-
ments. For example, some Japanese electric
power companies conducted experiments that dis-
tributed coupons to people as an incentive to go
to commercial facilities when electricity demand
is close to the limit of the existing power supply.
2
Occupancy detection is expected to contribute ef-
ficient encouragement to people who are always
at home during peak time.
Delivery Route Optimization.
Redelivery due to absence accounts for a quarter
of the total delivery distance in Japanese courier
companies.
3
Delivery route optimization employ-
ing household occupancy states is a promising ap-
plication to reduce redelivery cost and CO
2
emis-
sions.
The actual consumption estimation algorithm in
the next section, which detects occupancy with high
accuracy using electricity consumption data with low
resolution, is proposed as a fundamental technique for
realizing these applications.
3 ACTUAL CONSUMPTION
ESTIMATION ALGORITHM
In order to estimate occupancy state from A-route
data more accurately, an actual consumption estima-
tion algorithm is proposed. This algorithm estimates
actual electricity consumption based on cumulative
consumption. Figs. 3 and 4 show examples of ac-
tual consumption and A-route readout in 30-min in-
tervals, respectively. Although these two figures show
the same period and household consumption, A-route
2
http://www.ft.com/cms/s/0/5670af9e-0b22-11e4-9e55-
00144feabdc0.html [Accessed 4 October 2016]
3
http://www.mlit.go.jp/report/press/tokatsu01 hh 00023
4.html [Accessed 4 October 2016, written in Japanese]
SENSORNETS 2017 - 6th International Conference on Sensor Networks
42
0
100
0 0.5 1 1.5 2 2.5 3
Electricity consumption (Wh)
Elapsed time (h)
Figure 3: Actual electricity consumption in 30-min and 1
Wh increments.
0
100
0 0.5 1 1.5 2 2.5 3
Electricity consumption (Wh)
Elapsed time (h)
Figure 4: A-route readout in 30-min and 100 Wh incre-
ments.
readout data differ from the actual consumption be-
cause of the characteristics described in Section 2.3.
In this algorithm, the actual consumption data are
estimated using the cumulative consumption of the A-
route readout. In Fig. 5, which is converted into cu-
mulative consumption from Fig. 4, two stair-like solid
lines are drawn with cumulative A-route readout and
the values with 100 Wh added. The lower line shows
the lower limit of the A-route readout range. In addi-
tion, the upper line shows its upper limit. That is, the
area that consists of these two lines can be regarded
as the range of true cumulative electricity consump-
tion, which monotonically increases. The dashed line
which is the minimal distance within the range area
is calculated, as shown in Fig. 5. The estimated ac-
tual consumption data are finally calculated as each
increase of the dashed line in 30-min intervals.
This problem can be formulated as shown below.
In these formulas, the minimum increment of con-
sumption (100 Wh) is denoted by l, A-route readout
is x, and estimated consumption is y, respectively.
Minimize
T
t=1
q
1 + y
2
t
(1)
0
100
200
300
400
500
600
700
0 0.5 1 1.5 2 2.5 3
Cumulative consumption (Wh)
Elapsed time (h)
ݕ
ݕ
ݔ
ݔ
Lower limit based on
A-route readout
Upper limit (add 100 Wh
to A-route readout)
Figure 5: True cumulative electricity consumption and esti-
mated cumulative consumption.
sub ject to y
t
0
t
s=1
x
s
t
s=1
y
s
t
s=1
y
s
+ l
T
t=1
y
t
=
T
t=1
x
t
Equation (1) shows that the distance of the estima-
tion line consists of elapsed time and estimated con-
sumption y
t
. The distance of the estimation line can
be calculated as the sum of the oblique lines, which
show the estimated consumption such as y
1
and y
2
in
Fig. 5. The dashed line minimizing this distance is
determined as the estimated cumulative consumption
in the proposed algorithm.
Each constrained condition corresponds to non-
negativity, the range of actual consumption, and the
consistency of the cumulative consumption. There-
fore, the dashed line that fulfills these conditions cor-
responds to non-negativity cumulative consumption,
which is in the area between the two solid lines in
Fig. 5.
This problem equals the problem shown in for-
mula (2) and its dual problem in formula (3). Formula
(3) is a convex optimization problem called the total
variation optimization problem.
Minimize
T
t=1
y
2
t
(2)
sub ject to y
t
0
t
s=1
x
s
t
s=1
y
s
t
s=1
y
s
+ l
T
t=1
y
t
=
T
t=1
x
t
Actual Consumption Estimation Algorithm for Occupancy Detection using Low Resolution Smart Meter Data
43
Minimize
1
2
T
t=1
(y
t
x
0
t
)
2
+
1
2
T
t=1
|y
t
y
t1
|
2
(3)
x
0
0
= x
0
+
1
2
, x
0
t
= x
t
(t > 0)
Optimized cumulative consumption (the minimal-
length dashed line) can be calculated with the algo-
rithm shown below. The algorithm sequentially up-
dates the upper and lower lines, which indicate the
upper and lower limits of acceptable gradients. The
dashed line is fixed when both lines depart from the
range of true cumulative consumption. Finally, the
starting point is updated based on the end-point of the
fixed dashed line.
(1) The starting point and upper and lower lines are
determined (see Fig. 6(a)).
(1-1) Starting point: the terminal of the determined
dashed line.
(1-2) Upper line: A line connecting the starting point
and upper limit of the range of true cumulative
consumption.
(1-3) Lower line: A line connecting the starting
point and lower limit.
(2) Increment the time by 30 min and update the up-
per and lower lines (see Fig. 6(b)).
(2-1) When the upper line is larger than the upper
limit, update the upper line.
(2-2) When the lower line is smaller than the lower
limit, update the lower line.
(3) If both lines depart from the range of true cumu-
lative consumption, the dashed line and starting
point are updated (see Figs. 6(c), (d), and (e)).
(3-1) When the lower line is smaller than the up-
per cumulative consumption, the line connecting
the starting point and contact point with the lower
limit of the area are determined as the dashed line.
(3-2) When the upper line is larger than the lower
cumulative consumption, the line connecting the
starting point and contact point with the upper
limit of the area are determined as the dashed line.
(3-3) The contact point is determined as the new
starting point. The upper and lower lines are up-
dated.
(4) Return to Step 2 unless the time is terminated.
Fig. 7 shows an example of the estimated con-
sumption data, which are generated by the proposed
algorithm from the A-route readout in Fig. 4. It can
be seen that the estimated data in Fig. 7 are closer to
the actual data in Fig. 3 than the A-route readout in
Fig. 4.
4 EXPERIMENTS
4.1 Outline of Experiments
Occupancy detection using actual electricity con-
sumption data was conducted to consider the fea-
sibility of the proposed algorithm. The ECO data
set (Beckel et al., 2014) referred to in Section 2.2
was used for evaluating performance. This data set
consists of the 1 Hz electricity consumption data and
occupancy state on five Swiss households from June
2012 to January 2013.
In the experiment, the following four types of
electricity consumption data were prepared from the
ECO data set for the performance comparison.
B-route.
Actual instantaneous data in 1-min intervals and 1
W increments. The B-route data is the consump-
tion data with the highest resolution in this exper-
iment, and it corresponds to data that could be de-
rived from dedicated equipment such as HEMS.
30-min/1 Wh.
Actual watt-hour data in 30-min intervals and 1
Wh increments. These data correspond to the an-
swer data that the proposed algorithm described
in Section 3 aims to estimate.
A-route.
Watt-hour values in 30-min intervals and 100 Wh
increments. This data corresponds to the readout
derived from a smart meter via A-route.
Estimated.
Watt-hour estimated data in 30-min intervals and
1 Wh increments. The data mean estimated con-
sumption data generated from the A-route readout
by the proposed algorithm.
Occupancy detection was conducted every hour.
Random forests and SVM, which are known as
high-accuracy machine learning algorithms, were em-
ployed for classification. The explanatory variables
shown in Table 2 are employed as the feature for oc-
cupancy detection. Note that the temporal resolution
is different from the others in the B-route data.
Although occupancy detection was conducted ev-
ery hour, its state (occupied or unoccupied) is in 1-s
intervals in the data set. The hourly state is defined as
follows.
1. When the entire hourly state is unoccupied, the
state is regarded as unoccupied.
2. When the hourly state is mixed occupied and un-
occupied states, the state is regarded as occupied.
3. When the entire hourly state is occupied, the state
is regarded as occupied.
SENSORNETS 2017 - 6th International Conference on Sensor Networks
44
ݐ ݐ ݐ
ݐ ݐ ݐ
ݐ ݐ ݐ
ݐ
ݐ
ݐ
(a) Determine a starting point and
upper and lower lines at
ݐ
(b) Update the upper and lower
lines at
ݐ
(c) Both lines depart from the
range at
ݐ
(d) Fix the dashed line and update
the starting point
ݐ
ݐ
ݐ
(e) Re-update the upper and
lower lines at ݐ
Starting point
Upper or lower line
Fixed dashed line
(estimated consumption)
Range of true
cumulative consumption
Figure 6: Procedure of the actual consumption estimation algorithm.
0
100
0 0.5 1 1.5 2 2.5 3
Electricity consumption (Wh)
Elapsed time (h)
Figure 7: Estimated actual consumption data in 30-min in-
tervals from A-route readout.
That is, the state becomes occupied when it in-
cludes an occupied state of at least 1 s. The state is
used as an objective variable in occupancy detection.
In this experiment, the target hours of occupancy
detection are from 6:00 to 21:59 because almost all
the states during night and early-morning are occu-
pied in the ECO data set.
4.2 Results of Actual Consumption
Estimation
The actual consumption estimation algorithm de-
scribed in Section 3 was applied to the A-route read-
out, which was prepared from the ECO data set. Fig. 8
Table 2: Explanatory Variables for Occupancy Detection.
Variable Description
id Unique ID of household
mean Average consumption in the target interval
max Maximum consumption in the target interval
min Minimum consumption in the target interval
range Difference between maximum and minimum values
std Standard deviation in the target interval
hour Hour when occupancy detection is conducted
temp Average temperature in the target interval
season Summer or winter (dummy variable)
shows the visualized results of the four types data in-
cluding B-route, 30-min/1 Wh, A-route and estimated
consumption. As described in Section 4.1, 30-min/1
Wh consumption data shown in Fig. 8(b) corresponds
to the answer that the proposed algorithm aims to es-
timate. Fig. 8(c) shows that the A-route readout in-
cludes sequentially iterated values consisting of 0 and
100 Wh when consumption is relatively low, as de-
scribed in Section 2.3. In contrast, Fig. 8(d) shows
the estimated result, which suppresses the fluctuations
and represents the actual consumption more precisely
than the raw A-route readout, although it is smoother
than the 30-min/1 Wh data.
As described in Section 3, the estimated consump-
tion data are generated based on the dashed line,
which is the minimal distance within the range of true
cumulative consumption. Given these conditions, the
Actual Consumption Estimation Algorithm for Occupancy Detection using Low Resolution Smart Meter Data
45
Figure 8: Visualized consumption data of the ECO data set using (a) B-route, (b) 30-min/1 Wh, (c) A-route, and (d) Estimated
data.
proposed algorithm presupposes the continuation of
nearly equal consumption. For that reason, the pro-
posed algorithm tends to smooth the estimated results
when the actual consumption is low or surges quickly.
4.3 Performance of Occupancy
Detection
Occupancy detection was conducted based on the
conditions described in Section 4.1. The variables
shown in Table 2 were employed as features and the
occupancy state was classified into a binary class: the
household is occupied or unoccupied.
Generally, it is known that classifying a minority
class is harder than the majority one when machine
learning-based classification is conducted on imbal-
anced data sets (Japkowicz et al., 2000). In the ECO
data set, classifying an unoccupied state is difficult
because it accounts for only about 20% of the data.
Precision and recall of the unoccupied state are there-
fore employed as the performance criteria in addition
to accuracy, which shows the total rate of classifica-
tion for occupied and unoccupied states.
Figs. 9 and 10 respectively show the performances
of occupancy detection using random forests and
SVM methods, which employ four types of electricity
consumption data: B-route, 30-min/1 Wh, A-route,
and estimated consumption. The performances of
B-route are basically highest, followed by estimated
consumption, 30-min/1 Wh, and A-route in descend-
ing order. The accuracy, precision, and recall of
the estimated consumption data outperformed the 30-
min/100 Wh and A-route data results for both ma-
0.925
0.839
0.771
0.903
0.782
0.715
0.861
0.662
0.623
0.917
0.816
0.752
0.0 0.2 0.4 0.6 0.8 1.0
Performance
B-route 30-min/1 Wh A-route Estimated
Accuracy
Precision
(Unoccupied)
Recall
(Unoccupied)
Figure 9: Performance of occupancy detection by random
forests.
chine learning algorithms. Moreover, these results
are comparable to the results obtained with B-route
data, which corresponds to high resolution consump-
tion data.
The contribution ratio of features by the random
forests classifier is shown in Table 3. The contribution
ratio is used to evaluate the efficiency of each explana-
tory variable. This table shows the variables that are
independent of electricity consumption, such as hour
and temp, have relatively high ratios in the results ob-
tained with A-route, whereas max is the highest ratio
in the results obtained with the others. As described in
Section 2.3, A-route readout contains phantom varia-
tion that consists of alternating 0 and 100 Wh val-
ues. The discrepancy between the actual consump-
SENSORNETS 2017 - 6th International Conference on Sensor Networks
46
0.902
0.752
0.762
0.877
0.690
0.695
0.874
0.707
0.633
0.898
0.750
0.734
0.0 0.2 0.4 0.6 0.8 1.0
Performance
B-route 30-min/1 Wh A-route Estimated
Accuracy
Precision
(Unoccupied)
Recall
(Unoccupied)
Figure 10: Performance of occupancy detection by SVM.
Table 3: Contribution Ratio of Each Explanatory Variable
by Random Forests.
Variable B-route 30-min/1 Wh A-route Estimated
id 0.103 0.131 0.049 0.121
mean 0.183 0.165 0.149 0.173
max 0.211 0.236 0.072 0.208
min 0.149 0.138 0.116 0.130
range 0.129 0.084 0.028 0.032
std 0.093 0.088 0.033 0.023
hour 0.056 0.074 0.232 0.147
temp 0.005 0.073 0.316 0.160
season 0.070 0.009 0.006 0.007
tion and A-route readout tends to be large when con-
sumption is small. As a result, its contribution ratio of
explanatory variables regarding electricity consump-
tion decreased because it does not reflect the charac-
teristics of occupied and unoccupied household well.
In contrast, the estimated consumption data are con-
sidered to reflect the characteristics appropriately be-
cause its ratio of explanatory variables regarding elec-
tricity consumption is higher than that of A-route.
5 CONCLUSION
This paper proposed an actual consumption estima-
tion algorithm that detects occupancy with high accu-
racy using electricity consumption data with low reso-
lution. The proposed algorithm estimates actual con-
sumption using the cumulative consumption of the A-
route readout and the segmented line that exists within
the range of true cumulative consumption. The exper-
imental results of occupancy detection using the con-
sumption data estimated by the proposed algorithm
show an improvement in performance compared to
the result obtained with raw A-route readout. The re-
sults also show that the estimated consumption data
reflects the characteristics of occupied and unoccu-
pied states appropriately.
The proposed algorithm is expected to be useful
for various tasks such as profile analysis of household
attributes based on A-route data. Future work also
includes occupancy detection targeting Japanese do-
mestic households and efficient feature selection for
improving performance.
REFERENCES
Batra, N., Kelly, J., Parson, O., Dutta, H., Knottenbelt,
W., Rogers, A., Singh, A., and Srivastava, M. (2014).
Nilmtk: An open source toolkit for non-intrusive load
monitoring. In Proceedings of the 5th International
Conference on Future Energy Systems, pages 265–
276. ACM.
Beckel, C., Kleiminger, W., Cicchetti, R., Staake, T., and
Santini, S. (2014). The ECO data set and the perfor-
mance of non-intrusive load monitoring algorithms. In
Proceedings of the 1st ACM Conference on Embedded
Systems for Energy-Efficient Buildings, pages 80–89.
ACM.
Chang, H.-H., Lin, C.-L., and Lee, J.-K. (2010). Load
identification in nonintrusive load monitoring using
steady-state and turn-on transient energy algorithms.
In Computer Supported Cooperative Work in Design
(CSCWD), 2010 14th International Conference on,
pages 27–32. IEEE.
Chen, D., Barker, S., Subbaswamy, A., Irwin, D., and
Shenoy, P. (2013). Non-intrusive occupancy monitor-
ing using smart meters. In Proceedings of the 5th ACM
Workshop on Embedded Systems For Energy-Efficient
Buildings, pages 1–8. ACM.
Froehlich, J., Larson, E., Gupta, S., Cohn, G., Reynolds,
M., and Patel, S. (2011). Disaggregated end-use en-
ergy sensing for the smart grid. IEEE Pervasive Com-
puting, 10(1):28–39.
Hart, G. W. (1992). Nonintrusive appliance load monitor-
ing. Proceedings of the IEEE, 80(12):1870–1891.
Japkowicz, N. et al. (2000). Learning from imbalanced data
sets: A comparison of various strategies. In AAAI
Workshop on Learning from Imbalanced Data Sets,
volume 68, pages 10–15.
Kim, H., Marwah, M., Arlitt, M. F., Lyon, G., and Han,
J. (2011). Unsupervised disaggregation of low fre-
quency power measurements. In SDM, volume 11,
pages 747–758. SIAM.
Kleiminger, W., Beckel, C., and Santini, S. (2015). House-
hold occupancy monitoring using electricity meters.
In Proceedings of the 2015 ACM International Joint
Conference on Pervasive and Ubiquitous Computing,
pages 975–986. ACM.
Komatsu, H. and Nishio, K. (2015). How can smart data
analytics improve the way of information provision?
–the trend of methods and an exploratory analysis–
. In Proceedings of 8th International Conference on
Actual Consumption Estimation Algorithm for Occupancy Detection using Low Resolution Smart Meter Data
47
Energy Efficiency in Domestic Appliances and Light-
ing.
Molina-Markham, A., Shenoy, P., Fu, K., Cecchet, E., and
Irwin, D. (2010). Private memoirs of a smart meter.
In Proceedings of the 2nd ACM Workshop on Embed-
ded Sensing Systems for Energy-efficiency in Build-
ings, pages 61–66. ACM.
Nakano, Y. and Murata, H. (2007). Non-intrusive elec-
tric appliances load monitoring system using har-
monic pattern recognition-trial application to com-
mercial building. In International Conference on
Electrical Engineering.
Nguyen, T. A. and Aiello, M. (2013). Energy intelligent
buildings based on user activity: A survey. Energy
and Buildings, 56:244–257.
Nomura, K., Kashiwagi, T., Yamashita, T., and Kawade, T.
(2014). Solution to utilization of smart meter data.
FUJITSU Sci. Tech. J, 50(2):58–66.
Rashidi, P. and Mihailidis, A. (2013). A survey on ambient-
assisted living tools for older adults. IEEE Journal of
Biomedical and Health Informatics, 17(3):579–590.
Zoha, A., Gluhak, A., Imran, M. A., and Rajasegarar, S.
(2012). Non-intrusive load monitoring approaches
for disaggregated energy sensing: A survey. Sensors,
12(12):16838–16866.
SENSORNETS 2017 - 6th International Conference on Sensor Networks
48