Electrical Appliances Identiﬁcation and Clustering using Novel Turn-on

Transient Features

Mohamed Nait Meziane

, Abdenour Hacine-Gharbi

, Philippe Ravier

, Guy Lamarque

Jean-Charles Le Bunetel

and Yves Raingeaud

PRISME Laboratory, University of Orl

eans, 12 rue de Blois, 45067 Orl

eans, France

LMSE Laboratory, University of Bordj Bou Arr

eridj, Elanasser, 34030 Bordj Bou Arr

eridj, Algeria

GREMAN Laboratory, UMR 7347 CNRS - University of Tours, 20 avenue Monge, 37200 Tours, France

{mohamed.nait-meziane, philippe.ravier}@univ-orleans.fr, gharbi07@yahoo.fr, {lebunetel, yves.raingeaud}@univ-tours.fr

Keywords:

Electrical Appliances Identiﬁcation and Clustering, Energy Disaggregation, Non-Intrusive Load Monitor-

ing (NILM), Sequential Forward Search (SFS) Algorithm, Supervised and Unsupervised Classiﬁcation,

Turn-on Transient Features, Wrappers Feature Selection.

Abstract:

Due to the growing need for a detailed consumption information in the context of energy efﬁciency, differ-

ent energy disaggregation, also called Non-Intrusive Load Monitoring (NILM), methods have been proposed.

These methods may be subdivided into supervised and unsupervised approaches. Electrical appliance classi-

ﬁcation is one of the tasks a NILM system should perform. Depending on the chosen NILM approach, the

classiﬁcation task consists of either identifying the appliances or grouping them into clusters. In this paper,

we present the results of appliance identiﬁcation and clustering using the Controlled On/Off Loads Library

(COOLL) dataset. We use novel features extracted from a recently proposed turn-on transient current model

for both identiﬁcation and clustering. The results show that the amplitude-related features of this model are

the most suited for appliance identiﬁcation (giving a classiﬁcation rate (CR) of 98.57%) whereas the envelope-

related features are the most adapted for appliance clustering.

1 INTRODUCTION

In the context of energy efﬁciency, Non-Intrusive

Load Monitoring (NILM) approaches aim to pro-

vide, in a non-intrusive manner, detailed energy con-

sumption. This detailed information helps increase

the awareness about energy consumption behavior

of consumers along with other beneﬁts. The ben-

eﬁts of such consumption feedback were discussed

in several previous works (Fischer, 2008) (Darby,

2010) (Hancke et al., 2012). The interest in this ﬁeld

pioneered by Hart’s work during the mid-1980s (Hart,

1985) (Hart, 1989) (Hart, 1992) started to grow

rapidly these past few years starting around the year

2010 (Parson, 2016).

Along with the appliance working periods and the

consumed energy, appliance class is a very important

output of a NILM system. NILM approaches may be

classiﬁed using different criteria (Zeifman and Roth,

2011). One possible taxonomy is subdividing the

approaches into supervised and unsupervised (Zoha

et al., 2012) depending on the chosen strategy for

appliance-related information inference.

Supervised NILM approaches are the most com-

monly found in the literature. These approaches need

labeled data for the training of the appliance classiﬁer.

A major drawback of these approaches is the non ro-

bustness with respect to unseen appliances especially

when the training dataset size is small.

To alleviate this problem, an alternative is the use

of unsupervised NILM approaches (Bonﬁgli et al.,

2015). These approaches try to solve the NILM prob-

lem (i.e. to obtain detailed consumption informa-

tion) without a priori information (Zoha et al., 2012).

Several challenges face theses approaches (Goncalves

et al., 2011). Nevertheless, they are more adapted to

solve the NILM problem in real case scenarios where

unseen and different appliance types may be encoun-

tered.

A mid-way approach that is worth mentioning

is the semi-supervised approach (Barsim and Yang,

2015) (Gillis and Morsi, 2016). This approach is a

mix between both supervised and unsupervised ap-

proaches where a training step (supervised) helps in

the prediction of the appliance type using a unsuper-

vised approach.

Meziane, M., Hacine-Gharbi, A., Ravier, P., Lamarque, G., Bunetel, J-C. and Raingeaud, Y.

Electrical Appliances Identiﬁcation and Clustering using Novel Turn-on Transient Features.

DOI: 10.5220/0006245706470654

In Proceedings of the 6th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2017), pages 647-654

ISBN: 978-989-758-222-6

647

According to the above-mentioned approaches,

electrical appliances classiﬁcation problem can be

subdivided into two sub-problems: identiﬁcation and

clustering. Identiﬁcation is a supervised problem.

Having a set of data (called training dataset), labeled

with appliance type, the task is to identify the class

(appliance type) to which a new appliance belongs.

Clustering, on the other hand, is an unsupervised

problem. Having an unlabeled set of data (the appli-

ance types are unknown), the task is to ﬁnd clusters,

or groups, that deﬁne classes corresponding to appli-

ances with common characteristics.

The aim of this paper is to present the results

of both appliance identiﬁcation and clustering us-

ing novel features extracted from a recently proposed

turn-on transient current model (Nait Meziane et al.,

2015). Conclusions on the usefulness of these fea-

tures for both tasks are drawn.

The paper is organized as follows: section 2

describes the features used for appliance classiﬁca-

tion and the corresponding turn-on transient current

model. This section also discusses the relevance of

these features for the classiﬁcation task and motivates

the use of some chosen features instead of all the esti-

mated ones. Section 3 presents the identiﬁcation and

clustering results. It also gives a brief description

of the dataset used. The paper is concluded in sec-

tion 4 where conclusions are drawn and some possi-

ble tracks for the improvement of the presented work

are given.

2 TURN-ON TRANSIENT

FEATURES

2.1 Turn-on Transient Model

The work presented in this paper is based on

modeling the turn-on transient current signal us-

ing a recently proposed parametric mathematical

model (Nait Meziane et al., 2015). The parameters

of this model are then estimated and used as fea-

tures for electrical appliances classiﬁcation. One of

the goals of this work is to assess the usefulness of

these model parameters for classiﬁcation. For sim-

plicity, we suppose herein stationary amplitudes and

phases in contrast to the more general model pre-

sented in (Nait Meziane et al., 2015).

According to this mathematical model, the turn-

on transient current is an amplitude modulated sum-

of-sinusoids that can be written as:

s(t) = e(t)s

(t) + w(t), (1)

where

(t) =

∑

i=1

cos(2π f

t + φ

), (2)

e(t) =

(

+ 1 , if t ≥ 0

0 , otherwise

(3)

and w(t) is an additive white Gaussian noise

(AWGN)

. s

(t) is a sum of undamped sinusoids such

that A

(≥ 0), φ

∈ [−π, π] and f

are their amplitudes,

phases and frequencies, respectively. The number

of sinusoids d is supposed ﬁxed and known a priori.

The sinusoids frequencies are also known and are odd

order-harmonics such that f

= (2i −1) f

, i = 1, . . . , d

( f

= 50 Hz is the fundamental frequency also called

mains frequency).

The amplitude modulation, or envelope, e(t) de-

scribes the current amplitude variation from the

turn-on until reaching the steady-state phase. b =

, . . . , b

]

is a vector of n polynomial coefﬁcients

and t = [t, . . . , t

]

is a time vector such that b

t is a

degree polynomial allowing to tune the model am-

plitude variation to the real signal amplitude variation.

is a parameter that speciﬁes the initial amplitude of

e(t) i.e. when t = 0, e(t = 0) = A

+ 1.

For the work presented in this paper, and after dif-

ferent tests on real signals, we chose d = 5 harmonics

and a polynomial order n = 3. These values provide a

good ﬁt between the model and real signals.

All the model parameters can be put in-

side one vector θ

θ = [A

, b

, A

, φ

]

. The used algorithm for the es-

timation of θ

θ is based on a nonlinear least-squares

optimization algorithm called trust-region reﬂec-

tive (Coleman and Li, 1994). A detailed description

of the model, the theoretical limits of the variance of

the estimated parameters (Cram

er-Rao Bounds) and

the estimation algorithm will be considered in an up-

coming work.

2.2 Relevance of the Turn-on Transient

Features for Classiﬁcation

In this sub-section we will analyze the features in

order to select the most relevant ones for appliance

classiﬁcation amongst the elements of the θ

θ vector.

First, we do a pre-analysis for these features to get

an insight and conclude on their usefulness for ap-

pliance classiﬁcation. Then, we compare the conclu-

sions with the results of a feature selection algorithm.

This assumption is veriﬁed for the measurements of the

COOLL as shown in (Nait Meziane et al., 2016) where the

measurement system used to create COOLL is described.

ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods

648

0 5 10 15

!"#$%& '()*+ ,-./

-4

-2

0*1*-(#(1 2*)3(

4*5 6 %7 8

;

Figure 1: An example showing the instability of the esti-

mated phase features φ

from measurements of the same ap-

pliance: “an electric saw” from the COOLL dataset. The

x-axis represent different action delays (0 to 19 ms) i.e. dif-

ferent measurements of the same appliance. Between action

delays the mains frequency is likely to slightly vary which

affects the phase estimates.

Note that one of the main desired properties that de-

ﬁne the usefulness for us is the low variability of the

feature value for different measurement instances of

the same appliance. Hence, we will give a special fo-

cus for this property in the sequel.

2.2.1 Pre-analysis of Features for Selection

Phase-related Features. The phase features φ

, i =

1, . . . , d specify the position (in radians) of the sinu-

soids cos(2φ f

t + φ

) at t = 0 with respect to the 2π

time-cycle. These features are subject to an ambigu-

ity in their deﬁnition. For example, the solution of

cos(φ

) = 0 is φ

= 2πm with m being an integer num-

ber. This means that the solution is not unique and

a set of solutions exists. This problem can be alle-

viated by only keeping the solutions that are in the

range [−π, π]. Still, we will end up having two so-

lutions to choose from (for example, the solutions of

cos(φ

) =

are

and −

Moreover, when working on real signals, the

slightest nonstationarity encountered in real signals

(especially the mains frequency very small variations

that are usually less than 0.5 % of 50 Hz) seemed

to affect negatively the estimated phase values (Fig-

ure 1).

For all the above-mentioned reasons, we chose not

to use the phase features for the classiﬁcation.

Amplitude-related Features. Unlike the phase fea-

tures, the amplitude features A

, i = 1, . . . , d, repre-

senting the amplitudes of the sinusoids (Eq. (2)), are

much more stable (Figure 2). The estimated valued

also show a variability between the estimated ampli-

tudes of different appliances, even for appliances of

the same type. This suggests that these features are a

0 5 10 15

!"#$%& '()*+ ,-./

0*1*-(#(1 2*)3(

4*5 6 %7 8

;

Figure 2: An example showing the stability of the estimated

amplitude features A

from measurements of the same ap-

pliance: “an electric saw” from the COOLL dataset. The

x-axis represent different action delays (0 to 19 ms) i.e. dif-

ferent measurements of the same appliance.

0 0.2 0.4 0.6 0.8 1

Time (s)

-1.5

-1

-0.5

Amplit ude

(t)

Figure 3: Functions g

(t) and g

(t).

good candidate for appliance identiﬁcation instead of

clustering.

In the sequel, these features are kept for use in

appliance classiﬁcation.

Envelope-related Features. The envelope features

are A

and b

, j = 1,. . . , n. Whereas we, ideally,

seek time-independent features for appliance classi-

ﬁcation, all b

depend on the time reference t

except

. To illustrate this, we consider the following expo-

nent of a simulated envelope with n = 2:

(t) = b

t + b

, t ∈ [0, 1] s (4)

with b

= 0.5 and b

= −2 representing the polyno-

mial coefﬁcients. Our time reference is t

= 0 s. Sup-

pose now that we shift our time reference from t

= 0 s

to 0.3 s (i.e. we deﬁne a new time interval starting at

0.3 s). We then obtain the function g

(t) = g

(t +t

)

which is the portion of g

(t) on the newly deﬁned in-

terval (Figure 3). Estimating the parameters b

and b

of g

(t), we ﬁnd that b

= −0.7 and b

= −2. This

corresponds to:

(t) = = b

(t + t

) + b

(t + t

)

= (b

+ b

) + (b

+ 2b

)t + b

= b

+ b

t + b

(5)

Electrical Appliances Identiﬁcation and Clustering using Novel Turn-on Transient Features

649

The parameter that is not affected by the time refer-

ence shift is b

. We show, using a similar reasoning

for n > 2, that the only time reference-independent

parameter is the last coefﬁcient b

adapted for use in

appliance classiﬁcation.

Note also that this time reference shift generates

a new term b

that we can pull out of the exponent

and multiply by A

to get A

= A

. Practically, we

always take the time reference as the time where the

current is at its extremum (max or min). Therefore,

the estimated A

(or A

) speciﬁes the highest ampli-

tude the current can reach and is important to keep.

Finally, we select for the appliance classiﬁcation

and b

along with A

, i = 1, . . . , 5.

2.2.2 Feature Selection using a Wrapper

Approach for Identiﬁcation

Following the pre-analysis sub-section and in order

to prove the soundness of the selected features, we

propose hereafter to use a wrapper-based algorithm

to perform the feature selection task.

The goal is to select the most relevant set of fea-

tures from the set of all available features (in our

case 14) for appliance identiﬁcation. We propose

to use the wrapper-based sequential forward search

(SFS) algorithm (Kohavi and John, 1997). This lat-

ter, sequentially, adds at each selection step the “rel-

evant” feature that gives the highest possible classiﬁ-

cation rate (CR). This selection procedure was used

in (Hacine-Gharbi et al., 2015) and is summarized in

Algorithm 1.

Algorithm 1: Wrapper-based sequential forward

search (SFS) algorithm.

1. F = {A

, b

, A

, . . . , A

, φ

, . . . , φ

S = {}, I = 14 (initial number of parameters),

j = 1 (iteration index).

2. • Evaluate the classiﬁcation rate CR for each

feature f

∈ F .

• Select the ﬁrst feature f

π1

such that:

π1

= argmax

∈F

(CR ( f

)).

• F = F − { f

π1

}, S = { f

π1

3. • j = j + 1.

• For each f

∈ F , evaluate CR using S ∪ { f

• Select the feature f

π j

such that:

π j

= argmax

∈F

(CR (S ∪ { f

})).

• F = F − { f

π j

}, S = S ∪ { f

π j

4. Repeat step 3 until j = I.

5. Give the output S that yields the maximum CR.

The chosen classiﬁer is based on the k nearest

neighbors (k-NN) algorithm. In order to study the

effect of the value of k on the result, we propose to

apply the selection procedure for k = 1, 2, . . . , 10.

The feature selection procedure is conducted us-

ing the COOLL dataset (sub-section 3.1). COOLL

was divided by keeping the ﬁrst 10 measurement in-

stances from each appliance for the training and the

remaining 10 instances for the test (each appliance

having 20 instances). This yielded 420 measurement

instances for training and 420 measurement instances

for the test.

Table 1 gives the corresponding indexes of the ini-

tial set of features that will be used to show the feature

selection result.

Table 1: Initial set of features and corresponding indexes.

index 1 2 3 4 5 6 7

index 8 9 10 11 12 13 14

The S matrix (Eq. (6)) gives the selection result

where each row k corresponds to a speciﬁc k nearest

neighbors choice and each column j corresponds to

the set of selected features, by relevance, at each it-

eration j (Algorithm 1). The matrix elements are the

indexes of the initial set of features (Table 1).

S =

5 6 7 8 9 1 2 4 3 10 11 13 12 14

5 6 7 8 9 1 2 4 3 10 11 14 12 13

5 6 9 8 7 1 2 3 4 11 10 14 13 12

5 6 7 8 9 1 2 4 3 11 10 13 14 12

5 6 7 8 9 2 1 4 3 11 10 14 13 12

5 6 7 8 9 2 1 4 3 13 10 12 11 14

5 6 7 8 9 1 2 3 4 11 10 13 12 14

5 6 7 8 9 1 2 3 4 11 14 12 10 13

5 6 7 8 9 1 2 3 4 11 10 12 13 14

(6)

In Eq. (6), and for each row, the element that corre-

sponds to the maximum CR is highlighted. The se-

lected features are, then, the set that corresponds to

this element (following Algorithm 1, for a speciﬁc

row, each element represents the last added feature

to S). In the ﬁrst row, for example, the highlighted el-

ement is the fourth element indexed 8. Hence, the se-

lected features are A

, A

and A

with correspond-

ing indexes 5, 6, 7, and 8, respectively. To better il-

lustrate this, Figure 4 gives the CR corresponding to

rows 1, 6 and 10 of matrix S.

The selection results (see highlighted elements

of matrix S) indicate that the algorithm selected the

amplitude-related features as the most relevant for the

identiﬁcation. This is in agreement with the result

of sub-section 2.2.1 where the amplitude-related fea-

ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods

650

2 4 6 8 10 12 14

!"#$%"&'( &()#* !

100

+,%--&!.%"&'( $%"# "# /01

2344

5344

26344

(a) CR variation corresponding to rows k = 1, 6 and 10 (rep-

resenting also the number k of chosen nearest neighbors) of

matrix S (Eq. (6)))

2 3 4 5 6

!"#$%"&'( &()#* !

100

+,%--&!.%"&'( $%"# "# /01

2344

5344

26344

(b) Zoom of Figure 4a. The maximum values of CR for 1-

NN, 6-NN and 10-NN are found, respectively, for j = 4, 5

and 3.

Figure 4: Classiﬁcation rate CR variation function of itera-

tion index j (see Algorithm 1. j is also the column index of

matrix S (Eq. (6))).

tures were suspected to be good candidates for appli-

ance identiﬁcation.

Note also that the phases are the least adapted for

appliance identiﬁcation which, again, is in agreement

with the observations made in sub-section 2.2.1.

3 ELECTRICAL APPLIANCES

IDENTIFICATION AND

CLUSTERING

In this section we give the identiﬁcation and cluster-

ing results right after giving a brief description of the

used COOLL dataset.

In order to assess the usefulness of the model pa-

rameters for the classiﬁcation, we do three tests for

both the identiﬁcation and the clustering. In the ﬁrst

test we only use the envelope-related features A

and

. In the second one, we use solely the amplitude-

related features A

and in the last one we use all seven

features.

3.1 COOLL Dataset Description

Controlled On/Off Loads Library (COOLL) is a

dataset of high-sampled electrical current and voltage

measurements (840 current measurements and 840

voltage measurements) representing individual appli-

ances consumption. The measurements were taken in

June 2016 in the PRISME laboratory of the Univer-

sity of Orl

eans, France. The appliances are mainly

controllable appliances (i.e. we can precisely con-

trol their turn-on/off time instants). 42 appliances of

12 types were measured at a 100 kHz sampling fre-

quency (Table 2). A more detailed description of this

dataset and its speciﬁcities can be found in (Picon

et al., 2016).

Table 2: COOLL dataset summary. Source: (Picon et al.,

2016).

◦

Appliance type

# of appli-

ances

# of current

signals (20

per appliance)

1 Drill

6 120

2 Fan

2 40

3 Grinder

2 40

4 Hair dryer

4 80

5 Hedge trimmer

3 60

6 Lamp

4 80

7 Paint stripper

1 20

8 Planer

1 20

9 Router

1 20

10 Sander

3 60

11 Saw

8 160

12 Vacuum cleaner

7 140

Total

42 840

Note that we will use the current measurements

for the classiﬁcation. The voltage measurements

(sampled at 100 kHz) presenting low variability from

an appliance to another are, then, less adapted for

classiﬁcation and are discarded in this study.

The measurements of COOLL are 6 seconds long

with a pre-turn-on duration of 0.5 second duration and

post-turn-off duration of 1 second. Each appliance

has 20 measurement instances. Each instance corre-

sponds to speciﬁc action delay (a turn-on delay wrt

the mains voltage time-cycle) ranging from 0 to 19 ms

with a step of 1 ms (Picon et al., 2016).

3.2 Identiﬁcation

For the appliance identiﬁcation task presented here-

after, we use the supervised algorithm k-NN. This al-

gorithm allows the prediction of the class of a new

Electrical Appliances Identiﬁcation and Clustering using Novel Turn-on Transient Features

651

117

134

Drill

Fan

Grinder

Hair_drayer

Hedge_trimmer

Lamp

Paint_stripper

Planer

Router

Sander

Saw

Vacuum_cleaner

!"#$%&'#$ &()**#*

Drill

Fan

Grinder

Hair_drayer

Hedge_trimmer

Lamp

Paint_stripper

Planer

Router

Sander

Saw

Vacuum_cleaner

+",# &()**#*

100

120

Figure 5: Confusion matrix for identiﬁcation using A

and

. Classiﬁcation rate CR = 82.98%.

point (in the feature space representing an appliance)

by computing relative distances between this latter

and points in its neighborhood. The appliance class is,

then, decided with a majority vote between the classes

of the k nearest neighbors.

The parameters to be ﬁxed are the number k of

nearest neighbors to consider and the distance metric.

For our tests, we chose the Euclidean distance and k =

10. A low k value (< 5) may degrade the robustness

of the classiﬁer by increasing the risk of classifying

using isolated points (the extreme case being k =1).

Since we already know, for our dataset, that a well-

formed group should contain 20 points (the 20 mea-

surement instances of the same appliance), we chose

to compare each new point with half of the number

of points we are supposed to ﬁnd in its neighborhood.

Hence the k = 10.

Since our dataset is not big enough to have a lot of

measurement instances of all appliances in both the

training and the test datasets, and in order to get more

reliable test results, we chose to evaluate the iden-

tiﬁcation performance using K-fold cross-validation

with K = 10. The K-fold cross-validation consists

of dividing the dataset, randomly, into K sets, doing

K tests and averaging the results. For each test, we

choose one different set from all the K sets for the test

keeping the others for the training. We repeat this K

times and we average the obtained results.

Figure 5 gives the confusion matrix for the identi-

ﬁcation using the parameters A

and b

. With a classi-

ﬁcation rate (CR) of 82.98%, the identiﬁcation gives

several bad results which indicates that these features

are not the most adapted for appliance identiﬁcation.

Figure 6 gives the confusion matrix for the identi-

ﬁcation only using the features A

. We note the (very)

good CR value of 98.57%. This suggests that these

features are more adapted to appliance identiﬁcation

than the envelope-related features.

Figure 7, on the other hand, shows the obtained

120

160

140

Drill

Fan

Grinder

Hair_drayer

Hedge_trimmer

Lamp

Paint_stripper

Planer

Router

Sander

Saw

Vacuum_cleaner

!"#$%&'#$ &()**#*

Drill

Fan

Grinder

Hair_drayer

Hedge_trimmer

Lamp

Paint_stripper

Planer

Router

Sander

Saw

Vacuum_cleaner

+",# &()**#*

Figure 6: Confusion matrix for identiﬁcation using A

Classiﬁcation rate CR = 98.57%.

103

150

140

Drill

Fan

Grinder

Hair_drayer

Hedge_trimmer

Lamp

Paint_stripper

Planer

Router

Sander

Saw

Vacuum_cleaner

!"#$%&'#$ &()**#*

Drill

Fan

Grinder

Hair_drayer

Hedge_trimmer

Lamp

Paint_stripper

Planer

Router

Sander

Saw

Vacuum_cleaner

+",# &()**#*

100

150

Figure 7: Confusion matrix for identiﬁcation using A

, b

and A

. Classiﬁcation rate CR = 90.95%.

identiﬁcation confusion matrix obtained after using

all the model parameters. With a CR of 90.95%,

the use of all the features enhances the result of the

envelope-related features but also deteriorates the re-

sult of the amplitude-related features.

As a conclusion, we can say that the identiﬁcation

results conﬁrm the adaptability of the features A

for

the identiﬁcation task and the nonadaptability of A

and b

for this task.

3.3 Clustering

For the clustering, we use one of the most known un-

supervised algorithms i.e. the k-means. It requires the

user to specify a priori the number of clusters needed

to be formed. We chose k = 12 clusters based on

the number of device types we have in the COOLL

dataset.

Figure 8 shows the results as a confusion matrix

using A

and b

. The algorithm formed, especially,

one big cluster (cluster 1) and different smaller clus-

ters. The smaller clusters contain only lamps whereas

the big cluster contains mostly motor-driven appli-

ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods

652

120

160

140

!"#$% &'#()*+(

Drill

Fan

Grinder

Hair_drayer

Hedge_trimmer

Lamp

Paint_stripper

Planer

Router

Sander

Saw

Vacuum_cleaner

,+#* &'-((*(

100

120

140

160

Figure 8: Confusion matrix for clustering using A

and b

ances. This result is very interesting. The appliances

of the COOLL dataset are mostly motor-driven ap-

pliances (i.e. appliances with a motor that is respon-

sible for the main task the appliance is supposed to

perform) that were gathered inside that big cluster.

This indicates that the envelope-related features are

suitable for distinguishing appliances having different

working principles.

Lamps have a different working principle than the

motor-driven loads. One of the reason that may ex-

plain the scatter of the rest of the appliances (not

motor-driven) in different clusters is that the lamps are

of different types that have different working princi-

ples. Actually, the four lamps of COOLL are, respec-

tively, a 1.6 W light emitting diode (LED), a 15 W

compact ﬂuorescent lamp (CFL), a 105 W halogen

lamp (HL) and 100 W halogen lamp (HL).

The tests we did show that the feature b

is related

to the envelope amplitude decrease rate. Higher (neg-

ative) values indicate faster amplitude decrease.

Note also (Figures 8) that 46 lamps were grouped

inside the big cluster. These lamps are seemingly

lamps with no transient (Figure 9) (the current goes

almost directly from zero to the steady-state and no

transition is observed; they are most likely LED and

CFL lamps) and, hence, are different from lamps with

high amplitude variation transients (halogen lamps).

Figure 10 shows the clustering result using the A

features. Clearly, these features are not adapted for

the clustering since no link between the found clus-

ters and the true classes is clear and apparently no de-

terministic pattern seems visible to distinguish well

deﬁned clusters.

The result of Figure 11 shows that the use of all

available features still allows us to retrieve the motor-

driven cluster even with the use of the A

features in

contrast to what happened with identiﬁcation.

0 1 2 3 4 5 6

−0.04

−0.03

−0.02

−0.01

0.01

0.02

0.03

0.04

Time (s)

Amplitude (A)

Lamp1:0ms

Figure 9: Turn-on transient current of a light emitting diode

(LED). The interval [0, 1] s represents the pre-turn-on pe-

riod whereas the interval [5, 6] s represents the post-turn-off

period.

100

!"#$% &'#()*+(

Drill

Fan

Grinder

Hair_drayer

Hedge_trimmer

Lamp

Paint_stripper

Planer

Router

Sander

Saw

Vacuum_cleaner

,+#* &'-((*(

100

Figure 10: Confusion matrix for clustering using A

120

160

140

!"#$% &'#()*+(

Drill

Fan

Grinder

Hair_drayer

Hedge_trimmer

Lamp

Paint_stripper

Planer

Router

Sander

Saw

Vacuum_cleaner

,+#* &'-((*(

100

120

140

160

Figure 11: Confusion matrix for clustering using A

, b

and

4 CONCLUSIONS

In this paper, novel features extracted from a recently

proposed mathematical model for modeling the turn-

on transient current were presented and used in order

to classify electrical appliances. These features were

analyzed for the sake of selecting a set of features

that is relevant for appliance classiﬁcation. From a

Electrical Appliances Identiﬁcation and Clustering using Novel Turn-on Transient Features

653

set of fourteen features seven were selected. A se-

quential forward search (SFS) wrapper-based selec-

tion algorithm was also used and its results validated

the soundness of the previously selected features.

The Controlled On/Off Loads Library (COOLL)

was used for the classiﬁcation. A comparison be-

tween the appliance identiﬁcation and clustering re-

sults using the turn-on transient features was con-

ducted. The results indicate that the amplitude-based

features A

, i = 1, . . . , 5 are the most relevant for appli-

ance identiﬁcation whereas the envelope-based fea-

tures A

and b

are the most relevant for appliance

clustering.

Future work may investigate further the robust-

ness of the obtained results by testing the classiﬁca-

tion on other datasets with bigger sizes than COOLL

and containing other families of appliance types (TV,

washing machines, refrigerator, etc.). Other problems

like model selection (parameters d and n) for the turn-

on transient current model may also be addressed.

ACKNOWLEDGEMENTS

This study was supported by the R

egion Centre-Val

de Loire (France) as part of the project MDE–MAC3

(Contract n

◦

2012 00073640).

REFERENCES

Barsim, K. S. and Yang, B. (2015). Toward a semi-

supervised non-intrusive load monitoring system for

event-based energy disaggregation. In 2015 IEEE

Global Conference on Signal and Information Pro-

cessing (GlobalSIP), pages 58–62.

Bonﬁgli, R., Squartini, S., Fagiani, M., and Piazza, F.

(2015). Unsupervised algorithms for non-intrusive

load monitoring: An up-to-date overview. In Environ-

ment and Electrical Engineering (EEEIC), 2015 IEEE

15th International Conference on, pages 1175–1180.

Coleman, T. F. and Li, Y. (1994). On the convergence of

interior-reﬂective newton methods for nonlinear min-

imization subject to bounds. Mathematical Program-

ming, 67(1):189–224.

Darby, S. (2010). Smart metering: what potential for house-

holder engagement? Building Research & Informa-

tion, 38(5):442–457.

Fischer, C. (2008). Feedback on household electricity con-

sumption: a tool for saving energy? Energy efﬁciency,

1(1):79–104.

Gillis, J. M. and Morsi, W. G. (2016). Non-intrusive load

monitoring using semi-supervised machine learning

and wavelet design. IEEE Transactions on Smart

Grid, PP(99):1–8.

Goncalves, H., Ocneanu, A., Berges, M., and Fan, R.

(2011). Unsupervised disaggregation of appliances

using aggregated consumption data. In The 1st KDD

Workshop on Data Mining Applications in Sustain-

ability (SustKDD).

Hacine-Gharbi, A., Petit, M., Ravier, P., and N

emo, F.

(2015). Prosody based automatic classiﬁcation of the

uses of french ouias convinced or unconvinced uses.

In International Conference on Pattern Recognition

Applications and Methods (ICPRAM), number ISBN

978-989-758-077-2, pages 349–354.

Hancke, G. P., Hancke Jr, G. P., et al. (2012). The role of

advanced sensing in smart cities. Sensors, 13(1):393–

425.

Hart, G. W. (1985). Prototype nonintrusive appliance load

monitor. In MIT Energy Laboratory Technical Report,

and Electric Power Research Institute Technical Re-

port.

Hart, G. W. (1989). Residential energy monitoring and

computerized surveillance via utility power ﬂows.

Technology and Society Magazine, IEEE, 8(2):12–16.

Hart, G. W. (1992). Nonintrusive appliance load monitor-

ing. Proceedings of the IEEE, 80(12):1870–1891.

Kohavi, R. and John, G. H. (1997). Wrappers for feature

subset selection. Artiﬁcial intelligence, 97(1):273–

324.

Nait Meziane, M., Picon, T., Ravier, P., Lamarque, G.,

Le Bunetel, J.-C., and Raingeaud, Y. (2016). A

measurement system for creating datasets of on/off-

controlled electrical loads. In Conference on Envi-

ronment and Electrical Engineering (EEEIC), Pro-

ceedings of the 16th IEEE International, pages 2579–

2583.

Nait Meziane, M., Ravier, P., Lamarque, G., Abed-Meraim,

K., Le Bunetel, J.-C., and Raingeaud, Y. (2015). Mod-

eling and estimation of transient current signals. In

Signal Processing Conference (EUSIPCO), 2015 Pro-

ceedings of the 23rd European, pages 2005–2009.

Parson, O. (2016). Overview of the nilm ﬁeld.

http://blog.oliverparson.co.uk/2015/03/overview-

of-nilm-ﬁeld.html.

Picon, T., Nait Meziane, M., Ravier, P., Lamarque, G., Nov-

ello, C., Le Bunetel, J.-C., and Raingeaud, Y. (2016).

COOLL: Controlled on/off loads library, a public

dataset of high-sampled electrical signals for appli-

ance identiﬁcation. arXiv preprint arXiv:1611.05803.

Zeifman, M. and Roth, K. (2011). Nonintrusive appli-

ance load monitoring: Review and outlook. Consumer

Electronics, IEEE Transactions on, 57(1):76–84.

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.

ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods

654