Performance Evaluation of the Electrical Appliances Identification

System Using the PLAID Database in Independent Mode of House

Fateh Ghazali

, Abdenour Hacine-Gharbi

, Khaled Rouabah

and Philippe Ravier

LMSE Laboratory, University of Bordj Bou Arreridj, Bordj Bou Arréridj, Algeria

Electronics Department, University of Mohamed Boudiaf M'sila, Algeria

PRISME Laboratory, University of Orleans, Orléans, France

Keywords: Non-Intrusive Load Monitoring (NILM), Electrical Appliances Identification, Statistical Feature Extraction,

Discrete Wavelets Analysis, Wavelet Cepstral Coefficient (WCC), K-Nearest Neighbors (KNN),

Voting Rules Method, Independent Mode of House.

Abstract: In Electrical Appliances Identification (EAI) system, Plug Load Appliance Identification Dataset (PLAID) is

largely used to develop and benchmark new methods proposed for demand management in electricity

networks, more particularly, automated control, non-intrusive load planning and monitoring. Particularly,

this database contains electrical signals of 11 appliance electrical appliances, recorded in several houses. In

state-of-the-art, the EAI systems have used this latest PLAID designed, in two parts (one for training and the

other for testing). These parts can be organized on house-dependent mode or house-independent mode. In the

first mode, the signals of each appliance class and house in the testing part have examples in the training part.

In opposition, in the second mode, the houses in testing part have not any example in training part. In this

paper, we propose a comparative study between the performance of house-dependent EAI system and those

of house independent mode system. In addition, in order to more validate the results of the comparison study,

we propose the use of other classifiers like Gaussian Mixture Model (GMM), Linear Discriminant Analysis

(LDA) and Artificial Neural Network (ANN). The obtained results, based on the use of PLAID, have

demonstrated that the performances of this system, in independent mode, are relatively low compared to those

obtained in dependent mode. This shows that the house's electrical installation has a good footprint in the

input current signal.

1 INTRODUCTION

Electrical appliance identification (EAI) systems,

integrated into smart meters, are an important

function in ensuring proper management of

household electrical energy consumption and

distribution. An EAI system is considered as a pattern

recognition system containing two phases: (1) the

training phase (used to learn the different class

models) and (2) the testing phase (used to evaluate

system performances). These last phases are used to

match the data received, via a pattern recognition

system, with the information stored in a specific data

set. The Plug Load Appliance Identification Dataset

(PLAID) dataset (Gao, et al., 2014), a public,

https://orcid.org/0000-0003-2839-3259

https://orcid.org/0000-0002-7045-4759

https://orcid.org/0000-0002-0925-6905

collaborative dataset intended for load identification

research, is widely used in EAI systems in house-

dependent mode (Nait-Meziane, et al., 2016)- (Nait

Meziane, et al., 2017)- (Hacine-Gharbi, et al., 2018)-

(Ghazali, et al., 2019) (Ghazali, et al., 2020) (Ghazali,

et al., 2021). In this mode, the EAI systems are

designed in such a way that all houses have examples

of current signals in the training and test phases.

This present work aims to study and validate the

EAI systems proposed in (Ghazali, et al., 2019)

(Ghazali, et al., 2020) (Ghazali, et al., 2021), based on

the strategy of the voting rule, and realized in house-

dependent mode. Here, both validation and study of

the aforementioned works are carried out in house-

independent

mode using other classifiers, namely

880

Ghazali, F., Hacine-Gharbi, A., Rouabah, K. and Ravier, P.

Performance Evaluation of the Electrical Appliances Identiﬁcation System Using the PLAID Database in Independent Mode of House.

DOI: 10.5220/0012468900003654

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 13th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2024), pages 880-885

ISBN: 978-989-758-684-2; ISSN: 2184-4313

Figure 1: EAI system proposed in (Ghazali, et al., 2019).

Figure 2: EAI system proposed in (Ghazali, et al., 2021).

Gaussian Mixture Model (GMM), Linear

Discriminant Analysis (LDA) and Artificial Neural

Network (ANN). In addition, performances of the

latter EAI systems have been evaluated using the

PLAID database.

2 RELATED WORK

In (Ghazali, et al., 2019), the authors proposed an

electrical appliance identification system based on the

KNN classifier combined with the voting rule method,

using the statistical harmonics features from harmonic

analysis. These statistical features are estimated from

the sequences of the STFS coefficients vectors in order

to extract a single vector representing the complete

signal. Nevertheless, in practice, the measurements of

the current signals are supplied continuously. This

requires converting each signal into a sequence of

statistical harmonic features, considering time

segments of fixed duration. Figure 1 presents the

system proposed in (Ghazali, et al., 2019).

The obtained results show that the combination of

the mean and the standard deviation with 500

statistical features (250 for the mean and 250 for the

standard deviation) give a CR classification rate of

92.63%. Also, applying the voting rule strategy

improves the result to 94.97%.

In (Hacine-Gharbi, et al., 2018), the authors

proposed an electrical appliance identification system

based on the HMM classifier and the use of WCC

coefficients as features for compact data

representation to reduce dimensionality. This

descriptor estimated from other descriptor called

LWE (Log Wavelet decomposition-based Energy) by

the application of the Discrete Cosine Transform

(DCT). The last descriptor is also estimated from

another descriptor called DWE (Discrete Wavelet

Energy) by the application of the log10 of energy at

each decomposition level from the wavelet analysis.

The obtained results show that the WCC descriptor

give the best performances results.

In (Ghazali, et al., 2021) the authors proposed an

EAI system based on the KNN classifier combined

with the voting rule strategy based on the feature

extraction from wavelet analysis (DWE, LWE,

WCC). As well as, the concatenation of this

descriptors with the LOG_E descriptor can improve

the perform of the Figure 2 presents the EAI system

proposed by (Ghazali, et al., 2021). The results

obtained show good performance in terms of the

classification rate CR up to 98.51%.

It should be noted that this research used the

PLAID database. The Database is divided into two

subsets, one for learning and the other for testing, so

that each house will have examples in both bases

(dependent mode of the house). In this work, we are

Testing dataset of

signals and labels

STFS

features

extraction

Extraction of

selected statistical

STFS features

Training dataset of

signals and labels

STFS

features

extraction

Statistical STFS

features

Features selection

using JMI strategy

KNN modeling

of 11 Appliances

KNN

classifier

Voting rule

(decision)

Identified

class

Statistical DWE,

LWE, WCC,

LOG E features

Testing dataset

of signals and

labels

Discrete wavelet

analysis (DWE,

LWE, WCC)

features extraction

Extraction of

selected of statistical

DWE, LWE, WCC

and LOG_E features

Training

dataset of

signals and

Discrete wavelet

analysis (DWE,

LWE, WCC)

features and

LOG_E feature

KNN modeling

of 11 appliances

KNN

classifier

Voting rule

(

decision

)

Identified

class

Features selection

using wrapper

approach

Performance Evaluation of the Electrical Appliances Identiﬁcation System Using the PLAID Database in Independent Mode of House

881

interested in the independent mode of house, ie each

house will have examples in a single database (either

for testing or training).

3 THE EAI IN THE

INDEPENDENT MODE OF

HOUSE

3.1 Presentation of the PLAID Dataset

Our system is based on the Plaid dataset (The Plug

Load Appliance Identification Dataset) (Gao, et al.,

2014). PLAID is a public dataset of electric signatures

composed of 1074 instances recordings of currents

and voltages of 11 types of electrical appliances from

a variety of 55 households. These signals are sampled

at a 30 kHz rate. Figure 3 summarizes the appliances

found in the dataset with the different appliance types

and the number of instances for each type.

Figure 3: Distribution of appliances types and the number

of theirs instances in PLAID dataset.

3.2 PLAID Dataset Subdivision for

Training and Testing

It is clear that the design of an EAI system requires a

database. This latter is divided into two datasets, one

for training and the other for testing. The previous

works used the PLAID database in dependent mode

of house. In this work, we propose to study the case

of a distribution of the database in independent mode

of house. Initially the PLAID database is split into

two parts almost balanced in instance number. Table

1 presents the distribution of the database in the two

training and testing subsets for the different

appliances.

This distribution will be applied on the same

system proposed in (Ghazali, et al., 2021). the results

will be compared with those obtained in the previous

work. This situation is the most favorable for the real

case of the electrical appliance identification.

Table 1: PLAID database distribution in independent mode

of house for 50% for the training and 50% for the test.

Appliance type

Total

number

instances

Training

subset

Testing

subset

Compact ﬂuorescent

lamp

175 69 106

2 Vacuum cleane

38 22 16

3Hai

-drye

156 59 97

4 Microwave 139 64 75

5 Air conditione

66 51 15

6La

172 80 92

7 Fridge 38 5 33

Incandescent light

ulb

114 59 55

9 Fan 115 85 30

10 Washing Machine 26 11 15

11 Heate

35 25 10

Overall 1074 530 544

4 EXPERIENCES AND RESULTS

In this section, we present a number of experiments

in which we take an EAI system applied in (Ghazali,

et al., 2021). Different experiments are carried out to

evaluate the performance of our EAI system. This

system is based on WCC features extraction. The

WCC coefficients are extracted from the current

signals of the PLAID dataset (Gao, et al., 2014).This

latter is divided into two groups in independent mode

of house. For this study, we tested the following:

• A comparative study between different features

types, like DWE, LWE and WCC in dependent

and independent mode of house, taking the best

configurations obtained in the previous works.

• In order to improve the performance of the system,

we seeking the optimal configuration of the KNN

classifier by choosing the parameters K (nearest

neighbor vectors) and the optimal distance.

• We applied the wrappers selection method for

studying the relevance of the concatenation of the

LOG_E descriptor with different descriptors

(DWE, LWE and WCC)

• In order to validate the previews works we applied

other classifiers like GMM, ADL ...

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

882

4.1 Comparative Study Between

Dependent and Independent Mode

of House

In this experiment, we present the CR obtained in

dependent and independent mode of house for a

PLAID database distribution. taking the best

configurations obtained in the previews work

(Ghazali, et al., 2021), which are the mean and the

standard deviation statistical features extracted from

the window durations of 8 cycle length using DB5

mother wavelet with a decomposition level equal to

5, and 15 voting vectors for WCC, LWE and DWE

statistical features.

Table 2: CR obtained in dependent and independent mode

of house (NBF is the number of features).

Dependent mode of

house

Independent mode of

house

Descriptors

NBF

DWE

LWE

WCC

DWE

LWE

WCC

(

)

94.04 97.57 98.13 66.54 75.55 75.55

From this table, we show that the classification

rate (CR) in independent mode of house is relatively

low compared to that in dependent mode of house for

all descriptors type. These results show that the

electrical installation of the house has an influence on

the performance of our EAI system, the objective then

to look for another configuration of our system to

improve the identification results in independent

mode of the installation. In this context, the next

experiment consists in finding the right configuration

of the KNN classifier in terms of the optimal number

of nearest neighbor vectors as well as the

corresponding optimal distance.

4.2 Optimal Configuration of the KNN

Classifier

The objective of this experiment is to seek the optimal

configuration of the KNN classifier by choosing the

number of nearest neighbor vectors K as well as the

optimal distance in terms of the classification rate of

the TCS signals. Table III.3 presents the TCS

classification rates for the different values of k

varying from 1 to 100 and for the different distances

(Euclidean, Cosine, Correlation, Cityblock) using the

LWE descriptor. Figure 4 shows the evolution of the

classification rates of the different distances for the

different values of the variable k.

From this figure we show that, the maximum

classification rate of 78.49% is obtained by choosing

'Euclidean' distances with k equal to 80. The

Figure 4: CR as a function of k for KNN classifier using 12

LWE features.

'Cityblock' distance gives an equally acceptable

classification rate of 78.30% with the value of k equal

69. On the other hand, the 'Cosine', 'Correlation'

distances, give classification rates of 69.30%, 69.48%

respectively with the value of k equal 1. This

configuration (Euclidean, k=80) is taken for the rest

of the study.

4.3 Features Selection Results

The feature selection-based wrapper method is

applied to select the most relevant features among the

12 statistical features extracted from the wavelet

analysis. The three descriptors DWE, LWE and WCC

are concatenated with the two statistical features of

LOG energy. The initial 14 features subset is given

as follows:

F_DWE = {

, M

, Std

Std

, Std

, M

, Std

F_LWE = {LM

, LM

LStd

, LStd

, M

, Std

F_WCC = {wcc

, wcc

wcc

, wcc

, M

, Std

Table 4 gives the CR and the selected features at

each iteration j following the same selection

procedure as in (Ghazali, et al., 2021).

From the table 3 we can observe the following

points:

• The concatenation of the LOG energy descriptor

adds an improvement in the classification rate for

all the descriptors.

with the WCC descriptor the classification rate CR

reaches the value obtained with all 14 features with

only 6 features, and exceeds this value by a maximum

of 79.77% with 8 features only.

Performance Evaluation of the Electrical Appliances Identiﬁcation System Using the PLAID Database in Independent Mode of House

883

Table 3: CR% as a function of the selected features, j is the iteration number, Sel is the selected features number and Feat is

the feature name.

DWE + LOG_E

(

14 features

)

LWE + LOG_E

(

14 features

)

WCC + LOG_E

(

14 features

)

Sel Feat CR% Sel Feat CR% Sel Feat CR%

1 5 M

53.86 5 LM

51.65 1 wcc

55.69

2 6 M

73.52 6 LM

70.22 2 wcc

70.03

3 11 St

74.44 3 LM

75.55 3 wcc

75.73

4 12 St

74.44 11 LSt

76.65 14 St

77.20

5 10 St

74.44 2 LM

76.83 6wcc

77.94

6 9 St

72.42 13 LM

78.12 13 M

78.49

7 14 St

72.79 4 LM

79.41 5 wcc

79.77

8 8 St

73.71 1 LM

79.41 9 wcc

79.96

9 2 M

73.16 7 LSt

79.41 10 wcc

79.96

10 4 M

72.79 8 LSt

79.41 11 wcc

79.96

11 7 St

72.42 9 LSt

79.41 7 wcc

79.77

12 3 M

69.85 14 LSt

79.59 8 wcc

79.77

13 1 M

63.97 10 LSt

79.59 12 wcc

79.77

14 13 M

58.08 12 LSt

79.59 4 wcc

79.41

• The LWE descriptor gave the CR of 79.41 % at

only with 7 features. It reaches a maximum of

79.59% with 12 parameters.

with the DWE descriptor the CR classification rate

reaches its maximum value of 74.44% with only 3

features, this value exceeds that obtained with all 14

features of 58.08%.

4.4 Validation of the System with Other

Classifiers

Table 4: CR obtained in independent mode of house for the

different classifiers and the different descriptors.

Classifiers

DWE

12 features

LWE

12 features

WCC

12 features

KNN

(

Euclidean, k=80

)

63.97 78.49 78.49

GMM

(

GN=5

)

42.09 78.49 65.25

ANN (HLS = 100) 49.26 76.47 78.86

ADL (DFT= quadratic) 27.20 66.36 67.09

In order to validate the electrical appliances

identification system architecture, we applied several

classifiers such as GMM, LDA, ANN. Table 4

presents the classification rate results for the different

classifiers and for the different descriptors (with HLS:

hidden Layer Size; GN: Gaussian Number; DFT=

Discriminant Function Type).

From the results of Table 4, we can observe that

the performance of the system may vary between

27.20% and 78.49%, depending on the classifier and

the type of features. This shows the important role of

both factors.

5 CONCLUSIONS

In this work, we have investigated a comparative

study between the performances of EAI system in

house-dependent and house-independent modes. The

comparative study is carried out firstly using EAI

system based on the KNN applied on statistical

wavelet features vectors, and combined with the

voting rule strategy. The performances of the EAI

system are evaluated in two previous modes using the

PLAID database. The comparative study is extended

using other classifier such as GMM, LDA and ANN

classifiers.

The obtained results in independent mode which

represent the real case are relatively low with those

obtained in dependent mode. These results

demonstrate that the electrical installation in the

house will have an imprint in the input current signal,

and has an influence on the performance of our EAI

system.

REFERENCES

Gao J [et al.] Plaid: A public dataset of high resolution

electrical appliance measurements for load

identification research: demo abstract. [Conférence] //

In: Proceedings of the 1st ACM Conference on

Embedded Systems for Energy Efcient Buildings. -

New York, NY, USA : [s.n.], 2014. - pp. pp. 198–199..

Ghazali F. [et al.] Extraction and selection of statistical

harmonics features for electrical appliances

identification using kNN classifier, combined with

voting rules method [Revue] // Turkish Journal of

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

884

Electrical Engineering & Computer Sciences. - 2019. -

No. 4 : Vol. Vol. 27. - pp. pp. 2980-2997.

Ghazali Fateh, Hacine-Gharbi Abdenour et Ravier Philippe

Selection of statistical wavelet features using a wrapper

approach for electrical appliances identification based

on a KNN classifier combined with voting rules method

[Revue] // Int. J. Computational Systems Engineering.

- 2021. - No. 5 : Vol. Vol. 6. - pp. pp. 220–230.

Ghazali Fateh, Hacine-Gharbi Abdenour et Ravier Philippe

Statistical features extraction based on the discrete

wavelet transform for electrical appliances

identification [Conférence] // International conference

of intelligent systems and pattern recognition .. - 16-18

October 2020, Hammamet , Tunisia : [s.n.], 2020.

Hacine-Gharbi Abdenour et Ravier Philippe Wavelet

Cepstral Coefficients for Electrical Appliances

Identification using Hidden Markov Models

[Conférence] // 7th International Conference on Pattern

Recognition Applications and Methods. - Funchal,

Portugal : [s.n.], 2018.

Nait Meziane M. [et al.] Electrical Appliances

Identification and Clustering using Novel Turn-on

Transient Features [Conférence] // 6th International

Conference on Pattern Recognition Applications and

Methods (ICPRAM). - Porto, Portugal: [s.n.], 2017. -

pp. pp. 647-652.

Nait-Meziane M. [et al.] HMM-based transient and steady-

state current signals modeling for electrical appliances

identifcation [Conférence] // 5th International

Conference on Pattern Recognition Applications and

Methods. - Rome, Italy : [s.n.], 2016. - pp. 670-677.

Performance Evaluation of the Electrical Appliances Identiﬁcation System Using the PLAID Database in Independent Mode of House

885