SeminaireIREENA 20180408

Home electrical appliances classification
for smart management

and cost accounting purposes
Sarra Houidi
PhD supervisors M. Francois Auger1 and Mme Houda Ben Attia2

Co-supervisors: Mme Laurence Miègeville and M. Dominique Fourer
1: IREENA, Université de Nantes

2: LSE-ENIT, Université de Tunis El Manar
Academic background
2013-2016: Electrical engineer (speciality: Electrical systems) ENIT

Tunis
End of study internship: Classification des charges électriques

résidentielles, IREENA Saint-Nazaire
Positioning
Co-supervision of IREENA, University of Nantes and LSE-ENIT,

University of Tunis El Manar
Introduction Research work progress Conclusion and prospectives
Outline
1 Introduction
Objective and scientific issues
Context and problematic
State of the art
General framework
2 Research work progress

Data acquisition system
Home Electrical Appliances database establishment
Abrupt change detection
3 Conclusion and prospectives

Conclusion
Prospectives
Scientific contributions
Modules de l’école doctorale validés
S. Houidi Séminaire IREENA 2 / 31

Objective and scientific issues Context and problematic State of the art General framework
1 Introduction
Objective and scientific issues
Context and problematic
State of the art
General framework

Research work goal and issues
Home electrical appliances recognition to perform detail energy

sensing from a load curve using a single current and voltage meter
Load scheduling and cost accounting

Context
Contribution of the residential sector to almost 40% of global
electricity consumption in 2017.
Population growth
HEA multiplication
⇒ Impacts on energy bills and environment

Solution
Monitoring of the energy consumption and relaying it back to the

consumers
Why does it matter?
Con
sumer
s Ut
i
li
ti
es
Bi
l
ls
’awar
enes
s Cu
stomer
s’beh
avi
orpr
ogn
osi
s
Mon
thl
ybu
dgetcon
trol
andu
nder
st
andi
ng Capaci
t
ypl
ann
ingi
mpr
ovemen
t
I
den
ti
fi
cat
i
onofen
ergy«
hogs
» I
den
ti
fi
cat
i
onofHEAsf
orDRpr
ogr
ams

Non Intrusive Load Monitoring
NILM: Process to estimate the energy consumed by individual

Home Electrical Appliances (HEAs) with a single meter in a
house electrical panel connected at the PCC.
⇒ Partition of the load curve into its main components
⇒ Assignment of energy expenses per HEA

NILM challenges
Different electrical behaviours: purely resistive / resistive inductive

/ harmonic polluting
Similar power consumption characteristics of HEAs
Multiple HEAs operation types:
Simultaneous operation of HEAs

NILM approaches: taxonomy

General framework of supervised NILM methods

Data acquisition system HEA database Abrupt change detection
1 Introduction

Data acquisition system
Home Electrical Appliances database establishment
Abrupt change detection


Acquisition system architecture


Experimental Environnement
Di
git
alos
cil
l
oscope Acquisition of i(t) and v(t):
Differ
ential ⇒ steady-state operating conditions
vol
tagepr
obe
⇒ 59 types of HEA
⇒ sampling frequencies: fs = 250 kHz
Instrumentation:
Mai
ns uppl
y
(
~230V,50Hz) ⇒ 10 mV/A sensitivity current probe
⇒ 1/100 attenuation differential
voltage probe
Homeappl
i
ance
⇒ 8-bit resolution digital
Cur
rentpr
obe
oscilloscope (RIGOL DS1104Z)

HEA Profiling
Signature of a HEA recording n over one period of T = 20 ms
xn = {fn,1 (v, i), . . . , fn,j (v, i) . . . , fn,p (v, i)} (1)

v, i: voltage and current measurements
j th function of v, i of the nth HEA recording, where j ∈ {1, . . . , p} and n ∈ {1, . . . , N}
N = 59 × 8 = 472 number of recordings
p = 90 number of electrical features
   
x1 f1,1 f1,2 ... f1,j ... f1,p
 x2   f2,1 f2,2 ... f2,j ... f2,p 
   
 ..   .. .. .. .. 
 .   . . . . 
XN×p = =
    (2)
 xn   fn,1 fn,2 . . . fn,j . . . fn,p 
 ..   .. .. .. .. 
 .   . . . . 
xN fN,1 fN,2 . . . fN,j . . . fN,p

HEAs’ features extraction
Additivity property:
N
X N
X
f (v, i) = f (v, in ) for i= in (3)
n=1 n=1
-Active power P, and its harmonics Pk and PH

-Reactive power Q, and its harmonics Qk and QH
Features selection
Determination of the d 0 ≤ p0 relevant features for a correct grouping of nk = 8 recordings

of the same HEA within classes k ∈ {1, . . . , 59}
⇒ Dataset stored in a normalized matrix X̄N×p0
⇒ Euclidean-based K-Nearest Neighbours (K-NN) classifier (K=7)
3 methods for selecting the most relevant subsets of features among the p0 = 34 extracted:
⇒ Heuristic Forward Greedy Search
⇒ Principal Component Analysis (PCA)
⇒ Inertia Ratio Maximization using Feature Space Projection (IRMFSP)

Performance of the K-NN classifier across the

HEAs dataset for different feature subsets.
FS methods subset Features subset K-NN classification
size success rate
Heuristic 7 Q5 , Q11 , P9 , 96.8 %
search P, Q7 , P5 , PH
PCA1 9 P7 , P9 , Q, Q1 , P3 , 92.4 %
Q7 , Q3 , Q9 , PH
IRMFSP2 8 P, P15 , Q1 , P3 , P14 , 89.5 %
P10 , Q14 , Q8
P, Q 2 P, Q 84.4 %
Complete 34 - 84.3 %
feature set
1: Principal Component Analysis

2: Inertia Ratio Maximization Features Space Projection


Abrupt change definition
Fast transition that occurs between stationary states in a signal

⇒ NILM: On/Off and multiple operation modes appliances
Need of tools to decide whether a change occurs or not in the

signal

Mathematical problem statement

Xn = {xm ∈ R, m = n − k + 1, . . . , n}: vector of the last k available
samples of a signal at the current time n.
xm follows a probability density function (PDF) pθ (xm ) depending

on a deterministic parameter θ
Abrupt change: modification of θ at a change time nc .
⇒ Hypothesis test approach:
I H0 : “no change” versus H1 : “with a change at time nc ”
Decision rule:
At each time n, comparison of a decision function gn to a
threshold value h adjusted according to decision probabilities
decide H1 if gn > h otherwise decide H0

Detection algorithms
⇒ Three detection algorithms to be studied:
I The Effective Residual algorithm

I The CUmulative SUM (CUSUM) algorithm
I The Bayesian Information Criterion (BIC) algorithm
⇒ Assessment made in strictly the same conditions using

a sliding window of k = 5 samples
Statistical assessment through Monte Carlo test

Use of Performance metrics

Application of the detectors

Realization of a controlled consumption scenario using our own current and voltage
measurement system (Fs =1.2 kHz)
Xn
Active power signal P[n] = M1 v [k]i[k], with M = Fs/F
k =n−M+1
For each detector, h is set to allow the detection of the lowest ∆P = (P[n] − P[n − 1])

Detection results
Threshold values h and number of TP, FP and PR of the three considered detection algorithms
applied to the active power signal.
Effective CUSUM BIC

Residual
h 31.0 850.0 12.0
Number of 32 8 33
detected events
TP 7 7 7
FP 25 1 26
PR 21.3% 87.5% 21.2%
True Positives (TP): detection of a change when there is really one
False Positives (FP): detection of a change when there is not
Precision PR = TP/(TP + FP): fraction of correctly detected changes over all the times
⇒ The CUSUM algorithm appears to be the most effective with a high PR and a small
number of false positive.

Conclusion Prospectives Scientific contributions Modules de l’école doctorale validés
1 Introduction

Conclusion
Prospectives
Scientific contributions
Modules de l’école doctorale validés

Conclusion

Prospectives
Student t-test: to differentiate On from Off
Adaptation of the detection threshold h:

⇒ to reduce false alarms FP&
⇒ to maximize detection performance TP%
Adaptation of the detectors to multi-component signals
Taking an interest to the design and the performance evaluation of

a real-time scheduling algorithm for household expenses in order
to minimize the energy bill

Scientific papers
S. Houidi, F. Auger. H. Ben Attia Sethom, D. Fourer, L. Miegeville,
”Relevant feature selection for home appliances recognition”,
Electrimacs, Toulouse, July 2017.
S. Houidi, F. Auger. H. Ben Attia Sethom, L. Miegeville, D. Fourer
”Statistical assessment of abrupt change detectors for
Non-Intrusive Load Monitoring”, IEEE International Conference on
Industrial Technology (ICIT), Lyon, February 2018.

Formations scientifiques
Intitulé Durée Total
Intelligence artificielle : images, apprentissage 20h 20h/40h
et reconnaissance de formes
Formations professionelles
Intitulé Durée Total
Initiation à la création de site web 13h
La Recherche et l’Entreprise 12h 40h/60h
Comprendre les enjeux d’un état de l’art 15h
et de la propriété intellectuelle et industrielle
Activité d’enseignement
Travaux pratiques: Circuits électriques avec PSIM (15h)

Thank you for your attention.

How much energy can be really saved?
Better control of electricity consumption through direct feedback

12.
0%
9.
2%
8.
4%
6.
8%
gs
Real-Ti
me
in
Dail
y/ Pl
u s
av
Weekl
y Real
-
Time Feedback
tS
Es
ti
mat
ed
F
eedback Real
cen
3.
8% eedback F
F eedback -T
imei
nf
o
Hou s
ehol
d- downtothe
er
Web-based Real
-Ti
me
E
nhanced speci
fi
cin
fo, pr
emis
elevel appl
i
ance
P
energyaudi
ts
al
advis
eon l
evel
Bi
ll
i
ng wit
hinfoon in
fo
u
ongoingbasi
s dail
yor
n
Househol
d-
An
s
pecifi
cin
fo weekl
ybasis
advi
ce
«I
ndi
rect
»Feedback «
Dir
ect»F
eedback
P
rov
idedaft
erConsumpt
ionOccu
rs P
rov
idedRealT
ime
ACEEE energy savings resulting from 36 different studies between 2000-2010

*ACEEE: American Council for an Energy Efficient Economy
Effective Residual decision function

Absolute variation δm between 2 consecutive signal samples:
δm = |xm − xm−1 | for n−k +2≤m ≤n
Residual rm : difference between 2 consecutive variations
rm = |δm − δm−1 | for n−k +3≤m ≤n
Effective Residual decision function: sum of the last 3 residuals

H1
gn >
≤ h with gn = rn + rn−1 + rn−2
H0
⇒ Detection of a mean change at nc = n from the last k = 5 samples

of the signal
CUSUM algorithm principle

PDFs of Xn under hypotheses H0 and H1 :
n
Y c −1
nY n
Y
p(Xn |H0 ) = pθ0 (xm ) p(Xn |H1 ) = pθ1a (xm ) pθ1b (xm )
m=n−k +1 m=n−k+1 m=nc
If θ0 =θ1a , log-likelihood ratio L(Xn , nc ):
n
!
pθ1b (xm )
X
p(Xn |H1 )
L(Xn , nc )=ln = sm with sm= ln
p(Xn |H0 ) m=n pθ1a (xm )
c
CUSUM decision rule: maximization of the log-likelihood ratio over nc
H1 n
X
gn >
≤ h, with gn = max sm
n−k +1≤nc ≤n
H0 m=nc
CUSUM decision function

Changing parameter: mean value µ in a Gaussian process N (µ, σ)
under H0 , µ = µ0 for n−k+1≤m≤n

µ
1a for n−k+1≤m≤nc −1
under H1 , µ =  µ1b for nc ≤m≤n
Instantaneous log-likelihood ratio sm :

(xm −µ1b )2 (xm −µ1a )2 µ1b +µ1a
sm = + = ∆µ
2 ( xm − ) with ∆µ=µ1b −µ1a
2σ 2 2σ 2 σ 2
⇒ The CUSUM decision rule gn :

For an abrupt change occurring at nc = n in a sliding window of k = 5
samples
n−1
X n−1
X
with µ̂1b = xn , µ̂1a = 14 xm and σ̂ 2 = 14 (xm − µ̂1a )2 ,
m=n−4 m=n−4
H1
(xn − µ̂1a )2
gn >
≤ h, with gn = sn =
2 σ̂ 2
H0
BIC algorithm principle

The BIC of Xn under hypothesis Hi , i ∈ {0, 1}: likelihood criterion
penalized by the model complexity
λ
BIC(Hi ) = ln(p(Xn |Hi )) − M ln(k )
2
p(Xn |Hi ) Maximized data likelihood for the given model
λ Penalty factor (ideally equal to 1)
k Last available samples
M Number of parameters in the probabilistic model
Probabilistic model:
H0 : xn−k+1 , . . . , xn ∼ N (µ0 , σ0 )
H1 : xn−k+1 , . . . , xnc −1 ∼ N (µ1a , σ1a );
xnc , . . . , xn ∼ N (µ1b , σ1b )
Model parameters:
I Under H0 : µ0 and σ0 (M = 2)
I Under H1 : µ1a , σ1a , and µ1b , σ1b (M = 4)
⇒ Maximization of BIC(Hi ) when µ and σ 2 are replaced by their MLEs
µ̂ and σ̂ 2
BIC decision function

The BIC decision function gn is:
H1
gn >
≤ h with gn = max ∆BIC(nc )
n−k +1≤nc ≤n
H0
where ∆BIC(nc ) = BIC(H1 ) − BIC(H0 )
k (nc −n+k −1) (n−nc +1)
= ln(σ̂02 )− 2
ln(σ̂1a )− 2
ln(σ̂1b )−λ ln(k)
2 2 2
⇒ BIC decision rule:

For an abrupt change occurring at nc = n − 1 in a sliding window of k = 5 samples:
H1 !
1 σ̂ 10
gn0 >
≤ h
0
with gn0 = ln 6 0 4 , h0 = h + λ ln(5)
2 σ̂1a σ̂1b
H0
Xn n−2
X Xn
with σ̂02 = 1
5
(xm − µ̂0 )2 , σ̂1a
2 = 1
3
(xm − µ̂1a )2 and σ̂1b
2 = 1
2
(xm − µ̂1b )2
m=n−4 m=n−4 m=n−1
Test bench
Monte Carlo Test repeated 100 000 times
⇒ Xn is filled with 5 i.i.d samples xm ∼ N (0, σ) with σ = 1
⇒ Under H1 , addition of ∆µ = SNR × σ:
I to the last sample for Effective Residual and CUSUM
I to the last 2 samples for the BIC
⇒ Assessment made for:
I fixed SNR values
I varying SNR values ranging from 0.5 to 10
⇒ Use of 400 logarithmically spaced values of h
Performance metrics
Basic performance metrics:
TN
FP
True Positive TP → detection of a change H0
TP
when there is really one H1 FN
l
it
iy
→
b
True Negative TN no detection of a change
o
Pr ba
when there is not
False Positive FP → detection of a change
when there is not Xo
False Negative FN → no detection of a change gn<h gn>n
when there is really one

Computation of performance rates:
True Positive Rate TPR
TPR = TP/(TP + FN)
False Positive Rate FPR
FPR = FP/(TN + FP)
Precision PR
PR = TP/(TP + FP)
Performance metrics
Receiver Operating Characteristics (ROC):
Plot of the TPR versus the FPR for varying values of h
lim TPR = 1, lim FPR = 1, lim TPR = 0, lim FPR = 0
h→−∞ h→−∞ h→+∞ h→+∞
SNR = 3 Uniformly distributed SNR between 0.5

and 10
⇒ CUSUM ROC curve is the closest to the “optimal” point (FPR=0;TPR=1)
Performance metrics
Area Under the Curve (AUC):
Numerical integration of the ROC curve
AUC values computed for the three considered
detection algorithms
Effective
Residual CUSUM BIC
SNR = 0.5 0.51 0.59 0.53
SNR = 3 0.75 0.89 0.87
SNR = 6 0.96 0.98 0.97
variable SNR 0.85 0.91 0.89
⇒ CUSUM: best detection ability SNR-AUC profile of the Effective Residual,

CUSUM and BIC algorithms.
IRMFSP method
Maximization of the features subset relevancy for the recognition
task by:
⇒ Selection of features that maximize the ratio r
K
X
nk (µj,k − µj )2
k=1
r (j) = (4)
XN
(x̄n,j − µj )2
n=1
µj : center of gravity of the feature j ∈ {1, . . . , p}

µj,k : center of gravity of the feature j ∈ {1, . . . , p} for data belonging to class k ∈ {1, . . . , K }
x̄n,j : normalized value of feature j affected to individual n ∈ {1, . . . , N}
Minimization of features subset redundancy by performing an iterative
Gram-Schmidt orthogonalization process
IRMFSP results
1
0.9
0.8
0.7
0.6
Ratio r
0.5
0.4
0.3
0.2
0.1
0
P15
P14
P10
P12
P13
P3
Q14
Q13
Q15
Q10
P6
P4
P7
Q11
P9
P11
P2
P8
Q12
P1
P5
P
Q1
Q8
Q9
Q5
Q6
Q4
Q7
PH
Q2
Q3
QH
Heuristic Forward Greedy Search method

Addition of features one at time until the maximal classification
success rate is reached
98
96
Classification success rate (%)
94
92
90
88
86
84
82
P11
P15
P14
P13
Q11
P9
P5
P2
P1
P8
P3
P7
P6
Q10
P4
Q
P10
P12
Q12
Q13
Q14
Q5
PH
Q15
P
Q7
QH
Q9
Q4
Q3
Q1
Q6
Q8
Q2
Principal component analysis

Statistical conversion of p0 correlated features into q 0 linearly
decorrelated synthetic features (principal components)
Determination of the d 0 most correlated original features to the q 0
principal components
⇒ The clothest ones to the circumference of the unit correlation circle
1
Distance to the unit correlations circle center
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
P15
P11
P14
P12
P10
P13
P7
P9
Q
P3
P6
P2
Q11
P5
P8
P1
Q15
P4
Q10
Q12
Q14
Q13
Q1
Q7
Q3
Q9
PH
Q5
QH
Q6
Q8
Q4
Q2
Classification success rate against number of

features for two classifiers
100
95
Classification success rate (%)
90
85
80
Euclidean−based K−NN classifier

LDA Classifier
10 20 30 40 50 60 70 80 90
Number of features
Principle of simultaneously connected HEAs

recognition method
Y = AX + E (5)
Mixture matrix Y = [y1 , . . . , yp ] ∈ R(1×p)
Dataset of HEAs X = [x1 , . . . , xN ] ∈ R(N×p)
Activation matrix A = [a1 , . . . , aN ] ∈ {0, 1}(N×p)
Error matrix E = [e1 , . . . , ep ] ∈ R(1×p)
⇒ Problem: find the activation matrix which let the Root Mean Square Error
(RMSE) to be the minimum.
v
u p n
u1 X X
RMSE = t (yj − a1k xkj)2 (6)
p
j=1 k =1
K
X N!
Activation matrix: generation of all the N-dimensional possible binary
(N − k)!k!
k =1
vectors with k bits set.
Adapation of the threshold h

Z +∞
PFA = pgn|H0 (yj ) dyj (7)
h
⇒ Fix PFA to a constant value and determine

h = f (PFA )
Z +∞
PD = pgn|H1 (yj ) dyj (8)
h=f (PFA )
y = {yj ∈ R, j = 1, . . . , NR }: experiment outcomes of the decision function gn , under H0 and H1

NR : number of trials.
⇒ Threshold selection: optimization problem to balance the two conflicting
objectives (i.e., maximize PD while minimizing PFA ).

SeminaireIREENA 20180408

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

SeminaireIREENA 20180408

Uploaded by

Copyright:

Available Formats

Home electrical appliances classification

for smart management

PhD supervisors M. Francois Auger1 and Mme Houda Ben Attia2

1: IREENA, Université de Nantes

2013-2016: Electrical engineer (speciality: Electrical systems) ENIT

End of study internship: Classification des charges électriques

Co-supervision of IREENA, University of Nantes and LSE-ENIT,

2 Research work progress

3 Conclusion and prospectives

S. Houidi Séminaire IREENA 2 / 31

2 Research work progress

3 Conclusion and prospectives

S. Houidi Séminaire IREENA 3 / 31

Research work goal and issues

Home electrical appliances recognition to perform detail energy

S. Houidi Séminaire IREENA 4 / 31

⇒ Impacts on energy bills and environment

Monitoring of the energy consumption and relaying it back to the

Why does it matter?

S. Houidi Séminaire IREENA 6 / 31

Non Intrusive Load Monitoring

NILM: Process to estimate the energy consumed by individual

S. Houidi Séminaire IREENA 7 / 31

Different electrical behaviours: purely resistive / resistive inductive

Simultaneous operation of HEAs

NILM approaches: taxonomy

S. Houidi Séminaire IREENA 9 / 31

General framework of supervised NILM methods

S. Houidi Séminaire IREENA 10 / 31

2 Research work progress

3 Conclusion and prospectives

S. Houidi Séminaire IREENA 11 / 31

S. Houidi Séminaire IREENA 12 / 31

Acquisition system architecture

S. Houidi Séminaire IREENA 13 / 31

S. Houidi Séminaire IREENA 14 / 31

S. Houidi Séminaire IREENA 15 / 31

Signature of a HEA recording n over one period of T = 20 ms

xn = {fn,1 (v, i), . . . , fn,j (v, i) . . . , fn,p (v, i)} (1)

S. Houidi Séminaire IREENA 16 / 31

HEAs’ features extraction

-Active power P, and its harmonics Pk and PH

Determination of the d 0 ≤ p0 relevant features for a correct grouping of nk = 8 recordings

S. Houidi Séminaire IREENA 18 / 31

Performance of the K-NN classifier across the

1: Principal Component Analysis

S. Houidi Séminaire IREENA 19 / 31

S. Houidi Séminaire IREENA 20 / 31

Abrupt change definition

Fast transition that occurs between stationary states in a signal

Need of tools to decide whether a change occurs or not in the

S. Houidi Séminaire IREENA 21 / 31

Mathematical problem statement

xm follows a probability density function (PDF) pθ (xm ) depending

decide H1 if gn > h otherwise decide H0

⇒ Three detection algorithms to be studied:

I The Effective Residual algorithm

⇒ Assessment made in strictly the same conditions using

Statistical assessment through Monte Carlo test

S. Houidi Séminaire IREENA 23 / 31

Application of the detectors

S. Houidi Séminaire IREENA 24 / 31