Inter national J our nal of Electrical and Computer Engineering (IJECE) V ol. 9, No. 2, April 2019, pp. 1002 – 1011 ISSN: 2088-8708, DOI. 10.11591/ijece.v9i2.pp1002-1011, 1002 T o wards automatic setup of non intrusi v e appliance load monitoring – featur e extraction and clustering Khaled Chahine Electrical and Computer Engineering Department, Beirut Arab Uni v ersity , Lebanon Article Inf o Article history: Recei v ed Apr 19, 2018 Re vised Okt 31, 2018 Accepted No v 9, 2018 K eyw ords: Automatic setup Ener gy ef ficienc y ESPRIT Feature e xtraction Load monitoring Rule-based clustering ABSTRA CT Gi v en climate change concerns and incessantly increasing ener gy dem ands of the present time, impro ving ener gy ef ficienc y becomes of significant en vironmental and economic impact. Monitoring household electrical consumption through a non-intrusi v e appliance load monitoring (NIALM) system achie v es significant ef ficienc y impro v ement by pro vid- ing appliance-le v el ener gy consumption and relaying this information back to the user . This paper focuses on feature e xtraction and clustering, which constitute tw o of the four modules of the proposed automatic-setup NIALM system, the other tw o being labeling and classifi- cation. The feature e xtraction module applies the Estimation of Signal P arameters via Rota- tional In v ariance T echniques (ESPRIT), a well-kno wn parametric estimation t echnique, to the dra wn electric current. The result is a compact representation of the signal in terms of comple x numbers referred to as poles and residues. These comple x numbers are then used to determine a feature v ector consisting of the contrib ution of the fundamental, the third and the fifth harmonic currents to the maximum of the total load current. Once a signature is e xtracted, the clustering module applies distance-base d rules inferred of f-line from v arious databases and decides either to create a ne w class out of the ne w signatur e or to discard it and increase the count of an e xisting signature. As a result , the feature space is clustered without the a priori kno wledge of the number of appliances int o singleton clusters. Results obtained from a set of appliances indicate that these tw o modules succeed in creating an unlabeled database of signatures. Copyright c 2019 Institute of Advanced Engineering and Science . All rights r eserved. Corresponding A uthor: Khaled Chahine Electrical and Computer Engineering Department, Beirut Arab Uni v ersity Debbieh Campus, Lebanon +961 7 985 858 Ext. 3488 Email: k.shahine@bau.edu.lb 1. INTR ODUCTION Soaring ener gy prices and gro wing global w arming concerns ha v e triggered go v ernments and polic y mak ers to promote ef ficient ener gy systems that can sa v e considerable amounts of ener gy and, as a result, reduce carbon emissions (e.g., [1], [2]). In 2007, the European Council set the follo wing ambitious tar gets to be achie v ed by 2020: reducing emissions of greenhouse g ases by 20% with respect to the 1990 le v els, impro ving ener gy ef ficienc y with the aim of sa ving 20% of the European Union ener gy consumption, and raising the rene w able ener gy sources share to 20% [3]. Load monitoring as a part of ener gy management system can substantially reduce ener gy consumption of a certain household [4]. Indeed, a detailed re vie w of more than 60 feedback studies suggests that considerable ener gy sa ving can be achie v ed using real-time appliance le v el consumption information as opposed to monthly bills or weekly advice on ener gy consumption [5]. Load monitoring can also pro vide critical information for the demand response programs of fered by utility companies that impro v es the load forecasting accurac y and pro vides a con v enient platform for utility companies to realize a real-time pricing [6], [7]. One method of load monitoring is distrib uted direct sensing, which requires a sensor at each de vice or appli- ance in order to measure consumption. Although conceptually straightforw ard and potentially highly accurate, direct sensing is often e xpensi v e due to time consuming installation and the requirement for one sensor for each de vice or appliance. In response to limitations with the direct sensing approach, researchers ha v e e xplored methods to infer dis- J ournal Homepage: http://www .iaescor e .com/journals/inde x.php/IJECE Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE ISSN: 2088-8708 1003 aggre g ated ener gy usage via a single sensor . Pioneering w ork in this area is non-intrusi v e appliance load monitoring, first introduced by Geor ge Hart in the late 1980s [8]. In contrast t o the direct sensing methods, NIALM relies solely on single-point measurements of v oltage and current on the po wer feed entering the household. NIALM consists of four steps: data acquisition, e v ent detection, feature e xt raction, and e v ent classification. The ra w current and v oltage w a v eforms are transformed into a feature v ector , i.e . a more compact and meaningful representat ion that may include real po wer , reacti v e po wer , current-v oltage phase dif ference, and harmonics (e.g., [9], [10]). These e xtracted features are moni tored for changes, identified as e v ents (e.g., an appliance turning “on” or “of f ”), and classified do wn to the appliance or de vice cate gory le v el using a classification algorithm, which usually compares the features to a pree xist- ing database of signatures. Se v eral re vie ws and comparati v e studies of feature e xtraction and classification methods for electric loads in residential and commercial b uildings can be found in the literature [11–14]. Depending on the de gree of non-intrusi v eness, the literature dif ferentiates between manual-setup NIALM (MS-NIALM) and automatic-setup NIALM (AS-NIALM) systems. The manual setup is a non-intrusi v e appliance beha vior track er that requires a one-time intrusi v e period for setup. During the intrusi v e setup period, indi vidual appliances are manually switched on and of f to learn their signatures. On the other hand, the automatic setup is a process that sets itself up using prior information about potential appliances. AS-NIALM hence e xtracts the signatures and labels them without an y sort of manual interv ention, which w ould greatly f acilitate mass installation of smart meters. T o the author’ s kno wledge, no AS-NIALM system has hitherto been implemented. It is hence the main goal of this w ork to pa v e the w ay for such a solution. In this paper , the first tw o modules of an AS-NIALM system are presented. The system consists of the follo wing four modules: feature e xtraction, clustering, labeling and classification. The feature e xtraction module applies the ESPRIT method, a well-kno wn subspace-based estimation technique, to the dra wn electric current. The result is a compact representation of the current in terms of comple x numbers referred to as poles and residues [15, 16]. These comple x numbers are sho wn to be characteristic of the considered load and t hus can serv e as features for the subsequent classification layer . In [17], [18], poles and resi d ue s of current signals were estimated using the matrix pencil me thod. F or both synthetic and real data, results indicate that poles and residues e xtracted by ESPRIT allo w an almost perfect reconstruction of dra wn electric currents. Once a signature is e xtracted, the clustering module applies distance-based rules inferred of fline from v arious databases and decides either to create a ne w class out of the ne w signature or to discard it and increase the count of an e xisting signature. Signatures can hence be analyzed instantly and are not stored in memory . As a result, singlet on clusters are formed without the a priori kno wledge of the number of appliances. Results obtained from a database of a household indicate that these tw o modules succeed in distinguishing signatures of dif ferent appliances. The objecti v es of this paper are summarized in the follo wing three points: 1. sho w that the reduced number of poles and residues estimated by ESPRIT enable an accurate reconstruction of synthetic and real signals 2. sho w that the fundamental and higher harmonic currents determined from poles and residues yield a feature space with reduced inter -cluster o v erlap 3. propose a rule-based technique that assigns a set of signatures into singleton clusters. The rest of the paper is or g anized as follo ws. Section 2 presents the signal model, the co v ariance matrix, the principle of ESPRIT and the v alidation on simulated and real data. In Section 3, the feature space is gi v en. Sections 4 e xplains the rule-based clustering approach and e v aluates its performance on se v eral load combinations. Finally , section 5 pro vides the summary and conclusion. 2. MODEL-B ASED FEA TURE EXTRA CTION 2.1. Signal Model F or a sinusoidal dri ving v oltage of the form v ( t ) = V p 2 sin( ! t ) , the dra wn electric current can be modeled as a linear combination of d cisoids (comple x-v alued sinusoidal signals) weighted by comple x residues according to the follo wing damped e xponential signal model: i ( t ) d X m =1 r m exp f ( m + j 2 f m ) t g + b ( t ) (1) where r m is the residue of the m th cisoid, m is its attenuation f actor , f m is its frequenc y , and b ( t ) is additi v e white Gaussian nois e with zero mean and v ariance 2 2 . After sampling, the time v ariable, t , is replaced by t k = k t s , where T owar ds automatic setup of non intrusive appliance load monitoring ... (Khaled Chahine) Evaluation Warning : The document was created with Spire.PDF for Python.
1004 ISSN: 2088-8708 t s = 6 : 25 10 4 s is the chosen sampling period. The discrete current signal becomes: i ( k ) d X m =1 r m z k m + b ( k ) k = 1 ; 2 ; : : : ; N (2) where z m = exp f ( m + j 2 f m ) t s g m = 1 ; 2 ; : : : ; d (3) is the m th comple x pole. Under matrix form, the signal model is e xpressed by: i = Ar + b (4) with the follo wing notational definitions: i = i (1) i (2) ::: i ( N ) T (5) A = a 1 a 2 : : : a d (6) a m = z m z 2 m : : : z N m T (7) r = r 1 r 2 ::: r d T (8) b = b (1) b (2) ::: b ( N ) T : (9) The superscript T denotes the transpose operator . The feature e xtraction problem can no w be stated as follo ws. Given the electric curr ent data sequence f i ( k ) g N k =1 use a featur e e xtr action method to e xtr act the comple x poles f z m g d m =1 and r esidues f r m g d m =1 of the load. 2.2. The Co v ariance Matrix The N N co v ariance matrix of the data v ector i is defined by: R = E ii H (10) where E and H are the e xpectati on and the conjug ate-transpose operat ors, respecti v ely . Using equation (4) and de v eloping, R can be written as R = AR i A H + 2 2 I (11) where R i and 2 2 I denote the co v ariance matrices of the cisoids and noise, respecti v ely . In practice, R is estimated from M independent realizations or snapshots of the data v ector i and is termed the sample co v ariance matrix ^ R = 1 M M X r =1 i r i H r : (12) The co v ariance matrix is Hermitian and positi v e semi-definite. According to the spectral theorem [19], e v ery Hermitian matrix can be diagonalized by a unitary matrix (a matrix is said to be unitary if and only if its in v erse is equal to its conjug ate transpose) and that the result ing diagonal matrix has only real entries. This means that all eigen v alues of R are real and, due to its positi v e semi-definiteness, nonne g ati v e. M ore im portantly , eigen v ectors with distinct eigen v alues are orthogonal. It is this remarkable property of orthogonality that forms the basis of subspace-based estimation methods. Being Hermitian, R admits the follo wing eigen v alue decomposition (EVD): R = U diag f 1 ; : : : ; N g U H = [ U i U n ] diag f 1 ; : : : ; N g [ U i U n ] H (13) where U is the unitary matrix of eigen v ectors, 1 2 d > d +1 = = N = 2 2 are the eigen v alues, and U i = [ u 1 : : : u d ] and U n = [ u d +1 : : : u N ] contain the eigen v ectors associated with the signal (lar gest d ) and noise (smallest N d ) eigen v alues, respecti v ely . Indeed, the subspace spanned by the eigen v ectors of U i is termed the signal subspace, whereas the subspace spanned by the eigen v ectors of U n is termed the noise subspace. IJECE V ol. 9, No. 2, April 2019: 1002 – 1011 Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE ISSN: 2088-8708 1005 2.3. Estimation of Signal P arameters via Rotational In v ariance T echniques The basic idea behind ESPRIT is to e xploit the rotational in v ariance of the underlying signal subspace induced by the V andermonde structure of the mode matrix A without ha ving to kno w A [20], [21]. In particular , this rotational in v ariance can be established by partitioning A into tw o adjacent frequenc y sub-bands, preferably with maximum o v erlapping. The reason for maxi mum sub-band o v erlapping is to maintain the resolution capability of the algorithm. F or both the damped and undamped e xponential models, it is then easily v erified that the follo wing relation holds: A 2 = A 1 (14) where A 1 and A 2 are ( N 1) d dimensional matrices obtained from A by remo ving its last and first ro ws, respecti v ely: A = A 1 = A 2 (15) and is a d d full rank diagonal matrix that g athers the signal poles: = diag f z 1 ; z 2 ; : : : ; z d g : (16) Therefore, in contrast to polynomial-rooting-based methods such as root-MUSIC, ESPRIT can be adapted to parameter estimation of the damped e xponential model representing the dra wn electric current since the shift in v ariance structure is maintained. It is w orth noting that in the case of the undamped e xponential model (unit-modulus cisoids) z = 1 =z and becomes a unitary matrix since  H = H = I . Ho we v er , for the damped e xponential case, the comple x numbers on the diagonal of ha v e moduli less than one and becomes a contracti v e operator . The objecti v e of ESPRIT is hence to estimate the diagonal elements of from data samples. Since the subspace spanned by the mode v ectors of A is equal to the subspace spanned by the eigen v ectors of U i (the signal subspace) defined in section (2.2.), there must e xist a unique, nonsingular d d dimensional matrix G such that U i = A G : (17) P artitioning U i in the same w ay as done for A gi v es U i = U i 1 = U i 2 : (18) Using equation (17) allo ws to e xpress U i 1 and U i 2 as U i 1 = A 1 G (19) U i 2 = A 2 G = A 1 G : (20) Equation (19) implies A 1 = U i 1 G 1 : (21) Replacing the abo v e equation in equation (20) yields U i 2 = U i 1 G 1 G : (22) which can also be e xpressed as U i 2 = U i 1 : (23) where = G 1 G : (24) Therefore, and are similar matrices and the sought diagonal elements of are the eigen v alues of . The estimation problem then amounts to estimating from U i and computing its eigen v alues. Equation (24) is the k e y relationship in the de v elopment of ESPRIT and its properties. In practice, the co v ariance matrix is estimated from a finite number of noisy snapshots as described in section (2.2.). The result is that the subspace spanned by the eigen v ectors of ^ U i is only an estimate of the signal subspace and dif ferent from the subspace spanned by the mode v ectors of A . Moreo v er , t he spans of the eigen v ectors of ^ U i 1 and ^ U i 2 are dif ferent. Thus, the objecti v e of finding a matrix that satisfies equation (23) is no longer achie v able. T owar ds automatic setup of non intrusive appliance load monitoring ... (Khaled Chahine) Evaluation Warning : The document was created with Spire.PDF for Python.
1006 ISSN: 2088-8708 A commonly emplo yed criterion for such type of problems is the least-squares criterion that yields the follo wing approximate solution for : ^ = ^ U H i 2 ^ U i 2 1 ^ U H i 2 ^ U i 1 : (25) Kno wing that both ^ U i 1 and ^ U i 2 contain errors, estimating using the total least-squares (TLS) criterion is more appropriate. The TLS solution is gi v en by: ^ = W 12 W 1 22 (26) where the d d dimensional matr ices W 12 and W 22 are determined from a matrix W that results from the singular v alue decomposition of an ( N 1) 2 d dimensional matrix V = [ ^ U i 2 ^ U i 1 ] as follo ws: V = EW H (27) W = W 11 W 12 W 21 W 22 : (28) 2.4. Linear Loads T o v alidate ESPRIT as a feature e xtraction method, its poles and residues were first compared with those obtained from the theoretical e xpressions of the follo wing linear elementary loads: series RC, series RL, parallel RL and series RLC. The RC and RL circuits le ad to first order dif ferential equations in time, whereas the RLC circuit leads to a second order dif ferential equation. Using Euler’ s formula and rearranging allo w to re write the current e xpression obtained from the solution of the dif ferential equation characterizing the load in the form of equation (1). The poles and residues of each elementary load can then be readily identified. T abl e 1 gi v es the residues, a ttenuation f actors, frequencies, and dependent parameters of the four studied elementary loads. As can be seen from this table, first order circuits (RL and RC) are characterized by tw o purely imaginary conjug ate poles representing their forced response and one real pol e representing their natural response, whereas the second order circuit (RLC) has, bes ide the tw o purely imaginary conjug ate poles of its forced response, tw o conjug ate comple x poles related to its natural response. T able 1. The residues, attenuation f actors, frequencies, and dependent parameters of elementary loads. r m m f m Dependent P arameters Series RC V p 2 2 j R cos( ) e j 0 50 = R C 1 R v C 0 V p 2 sin( ) cos( ! t 0 ) e t 0 1 0 = arctan 1 R C ! Series RL V p 2 2 j R cos( ) e j 0 50 = L=R i L 0 V p 2 R cos( ) sin( ! t 0 ) e t 0 1 0 = arctan L! R P arallel RL V p 2 2 j R cos( ) e j 0 50 = arctan R L! i L 0 V p 2 R cos( ) sin( ! t 0 ) 0 0 Series RLC V p 2 2 j R cos( ) e j 0 50 = arctan[ 1 R ( L! 1 C ! ] ; ! 0 = 1 p LC Ae k 1 ! 0 t 0 k 1 ! 0 0 k 1 = p 2 1 , k 2 = + p 2 1 B e k 2 ! 0 t 0 k 2 ! 0 0 A = k 1 k 1 k 2 ( A 2 k 2 A 1 ) , B = k 2 k 2 k 1 ( A 2 k 1 A 1 ) The dependent parameters, A and B , of the RLC circuit are e xpressed in terms of A 1 and A 2 gi v en by: A 1 = v c 0 C L + V p 2 R cos( ) cos( ! t 0 ) ! 0 ! (29) A 2 = i L 0 V p 2 R cos( ) sin( ! t 0 ) (30) 2.5. Nonlinear Loads A nonlinear load is one for which the relationship between the current through the load and the v oltage across the load is a nonlinear function. A simple vie w of the nature of nonlinear loads can be presented using Ohm’ s La w , IJECE V ol. 9, No. 2, April 2019: 1002 – 1011 Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE ISSN: 2088-8708 1007 which states that the v oltage is the product of the load resistance and the current ( V = R I ). F or a linear load, the resistance ( R ) is a const ant; for a nonlinear load, the resistance v aries. When A C po wer is supplied to a nonlinear load, the result is the creation of curr ents that do not oscil late at the supply fr equenc y . These currents are called harmonics. Harmonics occur at multiples of the supply (fundamental) frequenc y . F or instance, if the fundamental frequenc y is 50 Hz, the so-called second harmonic is 100 Hz, the third harmonic is 150 Hz, and so on. An y number of harmonics can be created by a particular piece of equipment depending on that equipment’ s electrical c haracteristics. Therefore, the current dra wn by nonlinear loads can be still represented by equation 1 where harmonics appear in the form of pole-residue couples at frequencies multiples of 50 Hz. 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 −40 −30 −20 −10 0 10 20 30 40 T i m e ( s ) C u r r e n t ( A)     An al y t i c c u r r e n t R e c on s t r u c t e d c u r r e n t F or c e d r e s p on s e Nat u r al r e s p on s e (a) Series RC 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 −100 −50 0 50 100 150 T i m e ( s ) C u r r e n t ( A)     An al y t i c c u r r e n t R e c on s t r u c t e d c u r r e n t F or c e d r e s p on s e Nat u r al r e s p on s e (b) Series RL 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 −400 −300 −200 −100 0 100 200 300 400 500 T i m e ( s ) C u r r e n t ( A)     An al y t i c c u r r e n t R e c on s t r u c t e d c u r r e n t F or c e d r e s p on s e Nat u r al r e s p on s e (c) P arallel RL 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 −60 −40 −20 0 20 40 60 80 100 T i m e ( s ) C u r r e n t ( A)     An al y t i c c u r r e n t R e c on s t r u c t e d c u r r e n t F or c e d r e s p on s e Nat u r al r e s p on s e (d) Series RLC Figure 1. The analytic and reconstructed currents of the series RC, series RL, parallel RL, and series RLC circuits along with their forced and natural responses. 2.6. Results Assuming zero initial conditions ( i L 0 = 0 and/or v C 0 = 0 ), the follo wing numerical v alues were used to determine the electric current data sequence from which ESPRIT e xtracted poles and residues: f R = 100 , C = 0 : 1 mF g for the series RC circuit, f R = 10 , L = 100 mH g for both the series and parallel RL circuits, and f R = 1 , L = 20 mH, C = 60 mF g for the series RLC circuit. A duration of ten periods or 0 : 2 s w as chosen for the current which, at t s = 6 : 25 10 4 s, is equi v alent to 320 samples, and ESPRIT w as applied at each period. Figures 1a-1d sho w the current obtained from the analytic e xpression of poles and residues in table 1 and its reconstruction obtained from the poles and residues e xtracted by ESPRIT . An almost perfect agreement can be seen between the tw o curv es indicating the accurac y of the characteristic comple x numbers e xtracted by ESPRIT . In addition, the figures sho w the forced and natural responses of each of the four elementary circuits. T o e v aluate the performance of ESPRIT on nonlinear loads, a current consisting of a fundamental ( I 1 = 100% ) and four harmonics ( I 5 = 18 : 9% , I 7 = 11% , I 11 = 5 : 9% , and I 13 = 4 : 8% ) w as considered. This current can be represented by ten pairwise comple x conjug ate pole-residue couples. ESPRIT w as then used to e xtract these ten couples, which serv ed to rec on s truct the current as sho wn in Figure 2. As can be seen, ESPRIT is successful in estimating the pole-residue couples of the load. T owar ds automatic setup of non intrusive appliance load monitoring ... (Khaled Chahine) Evaluation Warning : The document was created with Spire.PDF for Python.
1008 ISSN: 2088-8708 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 −1.5 −1 −0.5 0 0.5 1 1.5 T i m e ( s ) C u r r e n t ( A)     An al y t i c c u r r e n t R e c on s t r u c t e d c u r r e n t Figure 2. The analytic and reconstructed currents of a nonlinear load. The inset zooms in on a half period of the dra wn current in order to sho w the accurac y of reconstruction. 2.7. V alidation on Real Data In this section, the v alidation of ESPRIT is carried out on currents of three representati v e loads: a tele vision set, a v acuum cleaner , and an economy lamp. As for the case of synthetic data, ESPRIT w as applied at each period. Figures 3a-3c sho w the current dra wn by the appliances and its reconstruction based on the pole-residue estimates of ESPRIT . The close agreement sho wn in the figures indicates that the e xponential model of equation 1 and its parameters estimated by the ESPRIT accurately predict the response of the actual loads. It is w orth menti o ni ng that the number of pole-residue couples d increases with the nonlinearity of the load. F or instance, the current of the v acuum cleaner could be accurately reconstructed from four pole-residue couples, whereas that of the economy lamp needed up to twelv e couples. 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 −15 −10 −5 0 5 10 15 T i m e ( s ) C u r r e n t ( A)     M e as u r e d c u r r e n t R e c on s t r u c t e d c u r r e n t (a) TV 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 −40 −30 −20 −10 0 10 20 30 40 T i m e ( s ) C u r r e n t ( A)     M e as u r e d c u r r e n t R e c on s t r u c t e d c u r r e n t (b) V acuum cleaner 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 −10 −8 −6 −4 −2 0 2 4 T i m e ( s ) C u r r e n t ( A)     M e as u r e d c u r r e n t R e c on s t r u c t e d c u r r e n t (c) Economy lamp Figure 3. The measured and reconstructed currents of a TV , v acuum cleaner and economy lamp. IJECE V ol. 9, No. 2, April 2019: 1002 – 1011 Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE ISSN: 2088-8708 1009 3. FEA TURE SP A CE The feature space cont ains 900 signatures uniformly distrib uted among the follo wing nine appliances: incan- descent lamp, halogen lamp, ec on om y lamp, w ater heater , electric con v ector , o v en, tw o-b urner hot plate, tele vision set, and computer . As sho wn in Figure 4, each signature (represented by a point in the the three-dimensional feature space) is characterized by three pole-residue products corresponding to the maxima of the fundamental, third and fifth harmonic currents. The restriction to three frequencies has the sole aim of representing the feature space graphically . From the feature space, ten clusters representing the nine appliances can be clearly distinguished. The additional cluster is due to the tw o-b urner hot plate which is represented by tw o clusters, one for each b urner . It can hence be concluded that the studied appliances can be f airly distinguished using the fundamental and higher harmonics. 0 5 10 15 20 −0.2 0 0.2 0.4 0.6 −0.2 0 0.2 0.4 0.6   F u n d am e n t al ( A) T h i r d Har m on i c ( A)   F i f t h Har m on i c ( A) T e l e v i s i on E c on om y Lamp C omp u t e r I n c an d e s c e n t Lamp Hal oge n Lamp Hot p l at e 1 b u r n e r E l e c t r i c C on v e c t or O v e n Hot p l at e 2 b u r n e r s W at e r He at e r Figure 4. The feature space sho wing the disaggre g ated contrib ution of the fundamental and harmonic currents to the maximum of the total current for se v eral appliances. 4. R ULE-B ASED CLUSTERING Rule-based clustering is accomplished by applying specific kno wledge rather than specific technique. This is a k e y idea in e xpert systems technology . It reflects the belief that human e xperts do not process their kno wledge dif ferently from others, b ut the y do possess dif ferent kno wledge. W ith this philosoph y , when one finds that their e xpert system does not produce the desired result s, w ork be gins to e xpand the kno wledge base, not to re-program the procedures. The clustering module has the task of reducing the feature space into singleton clusters. Moreo v er , applying clustering prior to classification ha s been recently sho wn to impro v e the accurac y of classification methods [22]. F orm- ing singleton clusters is rather challenging gi v en the cluster dispersion of some appliances and the unkno wn number of clusters. T o this end, t h e clustering module calculates the Euclidean distance between the first three elements of the present signature and those of each of the signatures in the unlabeled database. Based on the follo wing rules inferred of fline by a human e xpert from se v eral household appliances, the clustering module then decides either to create a ne w unlabeled signature in the database or to increase the count of an e xisting signature: 1: if the unlabeled database is empty then 2: add present signature 3: else 4: Determine the minimum distance from the present signature to signatures in the database 5: if MAX CURRENT > 3 and MIN DIST > 0 : 5 then 6: add present signature 7: else if 0 : 3 < MAX CURRENT < 3 and MIN DIST > 0 : 2 then 8: add present signature 9: else if 0 < MAX CURRENT < 0 : 3 and MIN DIST > 0 : 05 then 10: add present signature 11: else 12: Increase the count of the nearest signature 13: end if 14: end if T owar ds automatic setup of non intrusive appliance load monitoring ... (Khaled Chahine) Evaluation Warning : The document was created with Spire.PDF for Python.
1010 ISSN: 2088-8708 MAX CURRENT and MIN DIST represent the maximum of the current and the minimum distance in Amperes, respecti v ely . The outcome of this phase is a compact database of unlabeled signatures to which additional features such as the acti v e po wer and the current-v oltage phase dif ference can be added. In other w ords, the unlabeled database is an M N array where M is the number of identified signatures and N is the number of features in each signature. Applying the clustering module to the 900 signatures of the feature space sho wn in Figure 4 yields nine singleton clusters depicted as blue squares in Figure 5. It can be noticed that the clustering module represented the tw o clusters corresponding to the electric con v ector and the one-b urner hotplate by one cluster due to their hea vy o v erlap, thus missing the tenth cluster . 0 5 10 15 20 −0.2 0 0.2 0.4 0.6 −0.2 0 0.2 0.4 0.6   F u n d am e n t al ( A) T h i r d Har m on i c ( A)   F i f t h Har m on i c ( A) T e l e v i s i on E c on om y Lamp C omp u t e r I n c an d e s c e n t Lamp Hal oge n Lamp Hot p l at e 1 b u r n e r E l e c t r i c C on v e c t or O v e n Hot p l at e 2 b u r n e r s W at e r He at e r Figure 5. The feature space sho wing the singleton clusters determined by the clustering module. 5. CONCLUSION This paper discussed the first tw o modules of an automatic-setup NIALM system. First, the feature e xtraction module emplo yed ESPRIT to e xtract comple x poles and residues from the electric current. These comple x numbers were then used to determine a feature space with reduced inter -cluster o v erlap. Second, a cl ustering approach based on rules deri v ed of fline by a human e xpert w as proposed to assign a set of signatures into singleton cluste rs. Results indicated that these tw o modules succeed in creating an unlabeled database of signatures. Further research w ould be to update the clustering rules through testing the de v eloped modules on other datasets. REFERENCES [1] M. S. H. Lipu and T . F . Karim, “Ener gy Ef ficienc y Opportunities and Sa vings Potential for Electric Motor and its Impact on GHG Emissions Reduction, ” International Journal of Electrical and Computer Engineering (IJECE) , v ol. 3, no. 4, pp. 533-542, Aug 2013. [2] A. Rasheed and G. K. Rao, “Impro v ement of Po wer Quality for Microgrid Using Fuzzy Based UPQC Controller , ” International Journal of Electrical and Computer Engineering (IJECE) , v ol. 4, no. 2, pp. 37-44, Jun 2015. [3] Brussels European Coucil, Presidenc y Conclusions, 8/9 Mar , 2007. [4] S. Darby , “The Ef fecti v eness of Feedback on Ener gy Consumption: A Re vie w for DEFRA of the Literature on Meter ing, Billing and Direct Displays, ” T echnical Report for En vironmental Change Institute , Uni v ersity of Oxford: Oxford, UK, 2006. [5] K. Ehrhardt-Martinez, et al., “ Adv anced Metering Initiati v es and Residential Feedback Programs: A Meta-Re vie w for Household Electricity-Sa ving Opportunities, ” T echnical Report E105 for American Council for an Ener gy- Ef ficient Economy (A CEE) , W ashington, DC, USA, 2010. IJECE V ol. 9, No. 2, April 2019: 1002 – 1011 Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE ISSN: 2088-8708 1011 [6] Y .-H. Lin and M.-S. Tsai, “ An Adv anced Home Ener gy Manage ment System F acilitated by Nonintrusi v e Load Monitoring with Automated Multiobjecti v e Po wer Scheduling, ” IEEE T ransactions on Smart Grid , v ol. 6, no. 4, pp. 1839-1851, Jul 2015. [7] D. He, et al., “Incorporating Non-Intrusi v e Load Monitoring into Building Le v el Demand Response, ” IEEE T rans- actions on Smart Grid , v ol. 4, no. 4, pp. 1870-1877, Dec 2013. [8] G. W . Hart, “Nonintrusi v e Appliance Load Monitoring, ” Proceedings of the IEEE , v ol. 80, no. 12, pp. 1870-1891, Dec 1992. [9] C. Laughman, et al., “Po wer Signature An a lysis, ” IEEE Po wer and Ener gy Mag azine , v ol.1, no. 2, pp. 56-63, Mar -Apr 2003. [10] S. Semw al and R. S. Prasad, “Identifying Appliances using NIALM with Minimum Features, ” International Journal of Electrical and Computer Engineering (IJECE) , v ol. 4, no. 6, pp. 909-922, Dec 2014. [11] Y . Du, et al., “ A Re vie w of Identification and Monitoring Methods for Electric Loads in Commercial and Resi- dential Buildings, ” Ener gy Con v ersion Congress and Exposition , Atlanta, USA, 2010, pp. 4527-4533. [12] M. Zeifman and K. Roth, “Nonintrusi v e Appliance Load Monitoring: Re vie w and Outlook, ” IEEE T ransactions on Consumer Electronics , v ol. 57, no. 1, pp. 76-84, Feb 2011. [13] A. Zoha, et al., “Non-Intrusi v e Load Monitoring Approaches for Disaggre g ated Ener gy Sensing: A Surv e y , ” Sensors , v ol. 12, pp. 16838-16866, 2012. [14] K. Khalid, et al., “Performance Comparison of Artificial Intelligence T e chniques for Non-Intrusi v e Electrical Load Monitoring, ” Bulletin of Electrical Engineering and Informatics , v ol. 7, no. 2, pp. 143-152, Jun 2018. [15] H. Najme d di ne, et al., “Smart Metering by using the Matrix Pencil, ” in Proceedings of the 9th International Conference on En vironment and Electrical Engineering , Prague-Czech Republic, May 2010, pp. 238-241. [16] H. Najmeddine, et al., “Method and De vice for the Non-Intrusi v e Deter mination of the Electrical Po wer Con- sumed by an Installation, by Analysing Load T ransients, ” French P atent, FR 0856717 , Oct 2008. [17] K. Chahine, et al., “Electric Load Disaggre g ation in Smart Metering using a No v el Feature Extraction Method and Supervised Classification, ” Ener gy Procedia , v ol. 6, pp. 627-632, 2011. [18] K. Chahine and K. El Khamlichi Drissi, “ A No v el Feature Extraction Method for Nonintrusi v e Appliance Load Monitoring, ” Applied Computational Int elligence and Soft Computing , v ol. 2013, Article ID 686345, 7 pages, 2013. [19] P . R. Halmos, “What does the Spectral Theorem Say?” The American Mathematical Monthly , v ol. 70, no. 3, pp. 241-247, Mar 1963. [20] R. Ro y and T . Kailath, “ESPRIT -A Subspace Rotation Approach to Estimation of P arameters of Cisoids in Noise, ” IEEE T ransactions on Signal Processing , v ol. 34, no. 5, pp. 1340-1342, Oct 1986. [21] R. Ro y and T . Kailath, “ESPRIT -Esti mation of Signal P arameters via Rotational In v ariance T echniques, ” IEEE T ransactions on Acoustics, Speech, and Signal Processing , v ol. 37, no. 7, pp. 984-995, Jul 1989. [22] N. Mathi v anan, et al., “Impro ving Classification Accurac y Using Clustering T echnique, ” Bulletin of Electrical Engineering and Informatics , v ol. 7, no. 3, pp. 465-470, Sep 2018. BIOGRAPHY OF A UTHOR Khaled Chahine recei v ed his B .E. in Electrical and Electronics Engineering from the Lebanese Uni v ersity , Beirut, Lebanon, in 2007, his M.S. in Electronic Systems and Electrical Engineering from Polytech’Nantes, Nantes, France, in 2007, and his Ph.D. in Electrical and Electronics Engi- neering from the Uni v ersity of Nantes, Nantes, France, in 2010. In 2010, Dr . Chahine w as a post- doctoral researcher with LASMEA laboratory , Blaise P ascal Uni v ersity , Clermont-Ferrand, France. From 2011 to 2016, he w as an assistant professor with the Department of Electrical and Electronics Engineering of the Lebanese International Uni v ersity , Beirut, Lebanon. Since September 2016, he has been an assistant professor with the Electrical and Computer Engineering Department of Beirut Arab Uni v ersity , Debbieh campus, Lebanon. His research interests include statistical signal pro- cessing, control systems, nondestructi v e testing and e v aluation, and non-int rusi v e appliance load monitoring. T owar ds automatic setup of non intrusive appliance load monitoring ... (Khaled Chahine) Evaluation Warning : The document was created with Spire.PDF for Python.