Artificial Intelligence Techniques Applications for Power Disturbances Classification

Transcription

1 Internatonal Journal of Electrcal and Computer Engneerng 3:5 28 Artfcal Intellgence Technques Applcatons for Power Dsturbances Classfcaton K.Manmala, Dr.K.Selv and R.Ahla Abstract Artfcal Intellgence (AI) methods are ncreasngly beng used for problem solvng. Ths paper concerns usng AI-type learnng machnes for power qualty problem, whch s a problem of general nterest to power system to provde qualty power to all applances. Electrcal power of good qualty s essental for proper operaton of electronc equpments such as computers and PLCs. Malfuncton of such equpment may lead to loss of producton or dsrupton of crtcal servces resultng n huge fnancal and other losses. It s therefore necessary that crtcal loads be suppled wth electrcty of acceptable qualty. Recognton of the presence of any dsturbance and classfyng any exstng dsturbance nto a partcular type s the frst step n combatng the problem. In ths work two classes of AI methods for Power qualty data mnng are studed: Artfcal Neural Networks (ANNs) and Support Vector Machnes (SVMs). We show that SVMs are superor to ANNs n two crtcal respects: SVMs tran and run an order of magntude faster; and SVMs gve hgher classfcaton accuracy. Keywords back propagaton network, power qualty, probablstc neural network, radal bass functon support vector machne I. INTRODUCTION HE Qualty of Electrcal Power delvered by utlty Tcompanes has been a matter of growng concern n recent tmes. Modern equpment lke Computers, Electronc Sensors, Programmable Logc Controllers and other senstve ndustral equpment are at rsk of malfuncton, nstablty or reducton of ther lfe-spans n the presence of power qualty devatons. In order to mprove the qualty of power, electrc utltes contnuously montor power delvered at customer stes. Thus the raw transent power sgnal dsturbance data could be analyzed usng data mnng technques to provde knowledge about the captured waveforms[-3]. The data mnng technque proposed n ths paper for knowledge dscovery n power qualty data conssts of two stages. Feature extracton usng wavelet transform and classfcaton usng data mnng technques lke SVM and ANN. An artfcal neural network (ANN) s an abstract computatonal model of the human bran. Smlar to the bran ANN s composed of artfcal neurons, regarded as the processng unts, and the massve nterconnecton among K.Manmala s wth Dr.Svanth Adtanar College of Engneerng, Truchendur,Inda (phone: ; e-mal: s_monmala@yahoo.com). Dr.K.Selv, s wth Thagaraar College of Engneerng, Madura (e-mal: kseee@tce.edu). R.Ahla s wth Dr.Svanth Adtanar College of Engneerng, Truchendur,Inda (e-mal: r.ahla@yahoo.co.n). them. It has the unque ablty to learn from examples and to generalze,.e., to produce reasonable outputs for new nputs not encountered durng a learnng process. The dstnct features of ANN are as followng: learnng from examples, generalzaton ablty, non-lnearty of processng unts, adaptablty, massve parallel nterconnecton among processng unts and fault tolerance. ANNs had attracted a great deal of attenton because of ther nherent pattern recognton capabltes and ther ablty to handle nosy data. Support Vector machne (SVM) s a two layer neural network employng hdden layer of radal unts and one output neuron. The procedure of creatng ths network and learnng ts parameters s organzed n the way n whch we deal only wth kernel functons nstead of drect processng of hdden unt sgnals. Ths paper wll summarze and compare these two networks: ANN and SVM. The comparson wll be done wth respect to the complexty of the structure as well as the accuracy of results for the soluton of Power dsturbances classfcaton problem. Specal emphass wll be gven to the generalzaton ablty of the learned structures acqured n dfferent learnng processes. Test results taken from three dfferent NN's tranng methods and three SVM methods ndcate the hgher ablty of the SVM for fve types of dsturbances consdered n ths paper for classfcaton. II. FEATURE EXTRACTION A. Wavelet Analyss Wavelet analyss s a technque for carvng up functon or data nto multple components correspondng to dfferent frequency bands. Ths allows one to study each component separately. Wavelet analyss s a form of tme-frequency technque as t evaluates sgnal smultaneously n the tme and frequency domans [4]. It uses wavelets, small waves, whch are functons wth lmted energy and zero average, t dt () The functons are typcally normalzed, = and centered n the neghborhood of t =. It plays the same role as the sne and cosne functons n the Fourer analyss. In wavelet transform, a specfc wavelet s frst selected as the bass functon commonly referred to as the mother wavelet. Dlated (stretched) and translated (shfted n tme) versons of the mother wavelet are then generated [2]. Dlaton s denoted 949

2 Internatonal Journal of Electrcal and Computer Engneerng 3:5 28 by the scale parameter a whle translaton s adusted through b t b a b t (2), a a where a s a postve real number and b s a real number. The wavelet transform of a sgnal f (t) at a scale a and tme translaton b s the dot product of the sgnal f (t) and the partcular verson of the mother wavelet, a,b (t). It s computed by crcular convoluton of the sgnal wth the wavelet functon w t b a f a b f, a f t,, b A contracted verson of the mother wavelet would correspond to hgh frequency and s typcally used n temporal analyss of sgnals, whle a dlated verson corresponds to low frequency and s used for frequency analyss. Wth wavelet functons, only nformaton of scale a < correspondng to hgh frequences s obtaned. In order to obtan the lowfrequency nformaton necessary for full representaton of the orgnal sgnal f (t), t s necessary to determne the wavelet coeffcents for scale a >. Ths s acheved by ntroducng a scalng functon (t) whch s an aggregaton of the mother wavelets (t) at scales greater than.the scalng functon can also be scaled and translated as the wavelet functon, t a a, b t (4) a b Wth scalng functon, the low-frequency approxmaton of f(t) at a scale a s the dot product of the sgnal and the partcular scalng functon [3], and can be computed by crcular convoluton gven by (5). t a Lf a, b f,, f ( t) dt (5) a b a b Implementaton of these two transforms (3) and (5) can be done smoothly n contnuous wavelet transform (CWT) or dscretely n dscrete wavelet transform (DWT). Successful applcaton of wavelet transform depends heavly on the mother wavelet. The most approprate one to use s generally the one that resembles the form of the sgnal. Among the several wavelet functons that were mentoned n the lterature, the Daubeches famles of wavelets are the most wdely used. Among the dfferent dbn (N-order) wavelets, db4 s the most wdely adopted wavelet n power qualty applcatons. It has suffcent number of vanshng moments to brng out the transents whle mantanng a relatvely short support to avod havng too many hgh-valued coeffcents. In our work db4 mother wavelet s chosen. a (3) plane. They vary wth scale a but n the opposte manner, wth the tme spread beng drectly proportonal to a whle frequency spread to /a. The resolutons of DWT vary across the planes. At low frequency when the varaton s slow, the tme resoluton s coarse whle the frequency resoluton s fne. Ths enables accurate trackng of the frequency whle allowng suffcent tme for the slow varaton to transpre before analyss. On the contrary, n the hgh-frequency range, t s mportant to pnpont when the fast changes occur. The tme resoluton s therefore small, but the frequency resoluton s compromsed. Ths adustment of the resolutons s nherent n wavelet transform as the wavelet bass s stretched or compressed durng the transform [4]. Ths ablty to expand functon or sgnal wth dfferent resolutons s termed as multresoluton analyss, whch forms the cornerstone of many wavelet applcatons. In ths sense, a recorder-dgtzed functon a o (n), whch s a sampled sgnal of f(t), s decomposed nto ts smoothed verson a (n) (contanng lowfrequency components), and detaled verson (contanng hgher-frequency components) d (n), usng flters h(n) and g(n), respectvely. Ths s frst-scale decomposton. The next hgher scale decomposton s now based on sgnal a (n) and so on Fg.. Fg. Multresoluton sgnal decomposton C. Classfcaton Of Varous Powerqualty Events The Daubauche db4 wavelet functon was adopted to perform the DWT. The dfferent levels of wavelet coeffcent over the scales can be nterpreted as uneven dstrbuton of energy across the multple frequency bands [2,5]. If the selected wavelet and scalng functons form an orthonormal (ndependent and normalzed) set of bass, then the Parseval theorem relates the energy of the sgnal to the values of the coeffcents. Ths means that the norm or energy of the sgnal can be separated accordng to the followng multresoluton expanson 2 2 (6) f t dt A 2 k D k k k These squared wavelet coeffcents were shown to be useful features for dentfyng power qualty events. The energy dstrbuton pattern n the wavelet doman can be computed as sums of the squared coeffcents as n (6). B. Multresoluton Analyss One mportant trat of wavelet transform s that ts nonunform tme and frequency spreads across the frequency 95

3 Internatonal Journal of Electrcal and Computer Engneerng 3:5 28 (a) (b) (c) (a) (b) (c) orgnal sgnal sample ponts Fg. 2 Sag at dfferent nstants energy dstrbuton level of decomposton Fg. 3 Energy dstrbuton dagram for sag at dfferent nstants type. The low-level energy dstrbuton wll show obvous varatons when the dstorted sgnal contans hgh-frequency elements. On the contrary, the hgh- level energy dstrbuton wll show obvous varatons when the dstorted sgnal contans low-frequency elements. Fgures 2-5 shows that the energy dstrbuton pattern remans the same for the event sag despte occurrng at dfferent nstants and wth dfferent ampltudes. D. Duraton of Transents In general, when a transent dsturbance occurs, the stable power sgnal wll generate a dscontnuous state at the start and end ponts of the dsturbance duraton. Employng the DWT technque to analyze the dstorted sgnal through one level decomposton of the MRA wll cause the wavelet coeffcents at the start and end ponts of the dsturbance to generate severe varaton [2]. Therefore, we can easly obtan the start tme and end tme of the dsturbance from the varatons n absolute wavelet coeffcents and calculate the dsturbance duraton. Fg 6 and 7 shows the plot of level, level 2 and level 3 DWT coeffcents for the dsturbance swell and pure sne wave. The coeffcents show varaton for the dsturbance swell from whch the dsturbance duraton can be determned but there s no varaton for pure sne wave because the sgnal s smooth. (a) (b) (c) Orgnal Sgnal Sample Ponts Fg. 4 Sag wth dfferent ampltudes v(t) Level Level 2 Level x x x Fg. 6 Plot of DWT coeffcents for swell (a) (b) (c) energy dstrbuton level of decomposton Fg. 5 Energy dstrbuton dagram for sag wth dfferent ampltudes The propertes of energy dsturbance features are The energy dstrbuton remans unaffected by the tme of dsturbance occurrence. The outlne of energy dstrbuton remans the same despte varatons n the ampltude of the same dsturbance Level Level 2 Level 3 V(t) Fg. 7 Plot of DWT coeffcents for pure sne III. DISTURBANCE CLASSIFICATION USING ANN A. Back Propagaton Network A two layer feedforward neural network s used for learnng the feature vectors. From experence, 27 neurons n the hdden 95

4 Internatonal Journal of Electrcal and Computer Engneerng 3:5 28 layer gave the best results. The tan-sgmod functon s used as the transfer functon n the hdden-layer neurons. The output layer s comprsed of one neuron to dentfy the dsturbance class. The transfer functon used n the output layer s pureln because the output should ndcate the classes from to 5. Levnberg-Marquardt tranng functon (TranLM) s used to speed up the tranng process. The network s traned wth a backpropagaton algorthm. The error measure s gven as Q M 2 d k x k E (7) k where Q s the number of tranng samples, M the number of output neurons, and d k s the target value. The weghts are updated as wk we w k (8) where we the gradent of E wth respect to w, and s s the momentum constant and s the learnng rate. The gradent-descent algorthm was mplemented n batch mode. The performance of a gradent-descent algorthm s very dependent on the learnng rate. If the learnng rate were too large, the tranng would oscllate back and forth. If the learnng rate were too small, t would take a long tme to reach convergence []. To overcome ths problem an adaptve learnng rate that attempts to keep the step sze as large as possble wthout causng oscllaton s used. The learnng rate s made responsve to the complexty of the local error surface. The learnng rate used n ths work s.. B. Probablstc Neural Network A PNN s a varant of a radal bass functon neural network. It conssts of a radal bass layer and a compettve layer[]. The transfer functon n the hdden layer can be wrtten as h( x) ( x xc ) (9) where s the radal bass functon, x s the nput of the neuron, x c s the center of the neuron, and x- x c s usually taken to be the Eucldean dstance. In ths partcular case, a Gaussan functon was used as the radal bass functon. It s wrtten as 2 2 ( x x 2 ) h ( x ) e c () where s the spread of the Gaussan. If the spread of the Gaussan functon s too small, The Gaussan s sharply peaked and the neurons are not able to cover the nput space well. Ths wll result n poor generalzaton. However, f the spread s too large, there would be large overlaps n the nput space and all neurons wll gve large values for all nputs. The network was traned for spread value.. The nput layer has m unts to whch the m dmensonal nput vector s appled. The frst hdden layer has one pattern unt for each exemplar. The second hdden layer contans one summaton unt for each class. The output layer s the decson layer whch contans one neuron to specfy the class. The PNN [,6,7] can functon as a classfer and has the advantage of beng a fast learnng process as t requres only a snglepass network tranng stage wthout any teraton for adustng weghts. Further, t can tself adapt to archtectural changes. As the structure of PNN s smple and learnng effcency s very fast t s sutable for sgnal classfcaton problems. Hence PNN s consdered as the best neural network for power dsturbances classfcaton. The nput layer contans 4 neurons (3 energy features from DWT, tme of duraton), the frst hdden layer contans 3 neurons (3 tranng exemplars) and the second hdden layer contans 5 neurons (5 classes) and the fnal output layer contans one neuron. Fg. 8 PNN Classfer C. RBF Networks RBF networks can be used as unversal functon approxmatons. Ths conssts of a network wth a sngle hdden layer and a structure smlar to back propagaton networks [8,9]. Each hdden layer unt has a centrod c and smoothng factor.. These neurons compute the dstance between the nput x and the centrod c rather than the vector product of the weghts and nputs. The outputs are nonlnear, radal symmetrc functons of the dstance. Thus the output s the strongest when x s the closest to the value c. RBF networks apply real mappng functons f m whch have the general form M f m ( x) w K[( x c ) / ] () The functon K s a radal symmetrc kernel functon computed by M kernel unts. The Gaussan exponental functon s commonly used n RBF networks 2 f ( x) exp [( x c ) ] (2) The centrod c and constants and. have to be chosen accordngly to the tranng data set. General Gaussan actvaton functons are superor to sgmod functons n estmatng a broad class of functons. The network has only one hdden layer and the fact that the hdden nodes receve nput drectly from the nput layer wthout havng to calculate the weghted sums, makes t much faster to tran than a back propagaton network of comparable sze. 952

5 Internatonal Journal of Electrcal and Computer Engneerng 3:5 28 RBF networks can be used as unversal functon approxmatons. Ths conssts of a network wth a sngle hdden layer and a structure smlar to back propagaton networks. Each hdden layer unt has a centrod c and smoothng factor. These neurons compute the dstance between the nput x and the centrod c rather than the vector product of the weghts and nputs. The outputs are nonlnear, radal symmetrc functons of the dstance. Thus the output s the strongest when x s the closest to the value c. The network has only one hdden layer and the fact that the hdden nodes receve nput drectly from the nput layer wthout havng to calculate the weghted sums, makes t much faster to tran than a back propagaton network of comparable sze. Smoothng factor s chosen as.3.radal bass functon wth exact ft s used as the tranng network. IV. DISTURBANCES CLASSIFICATION USING SVM In recent years, a new approach to construct and tran neural networks (NNs) was developed, whch s free of many dsadvantages. The new networks are called SVMs. Support vector machnes (SVMs) [8] were orgnally desgned for bnary classfcaton. How to effectvely extend t for multclass classfcaton s stll an ongong research ssue. Currently there are two types of approaches for multclass SVM. One s by constructng and combnng several bnary classfers whle the other s by drectly consderng all data n one optmzaton formulaton. In general, t s computatonally more expensve to solve a mult-class problem than a bnary problem wth the same number of data. Ths work s devoted to the second approach,.e. t solves a mult-class problem by decomposng t to several bnary problems n a herarchcal way. The three methods consdered n ths paper are oneaganst-all and one-aganst-one and dendogram based SVM. A. One aganst all method The earlest used mplementaton for SVM multclass classfcaton s probably the one-aganst-all method (for example, [8],[9]). It constructs k SVM models where k s the number of classes. The th SVM s traned wth all of the examples n the th class wth postve labels, and all other examples wth negatve labels. Thus gven l tranng data, (x,y ),.(x l,y l ), where x R n, =,.l and y k {,,k} s the class of x, the th SVM solves the followng problem: l T T mn ( w ) w C ( w ), w, b 2 T ( w ) ( x ) b, f y T ( w ) ( x ) b, f y,,..., l (3) where the tranng data x are mapped to a hgher dmensonal space by the functon and C s the penalty parameter. T Class of x arg max(( w ) ( x) b ) (4),.. k B. One aganst one method Ths method constructs k(k-)/2 classfers where each one s traned on data from the th and th classes, we solve the followng bnary classfcaton problem[]: T T mn ( w ) w C t ( w ) w b 2 T ( w ) ( xt ) b t, f y t T ( w ) ( xt ) b t, f y t (5) There are dfferent methods for dong the future testng after all k(k-)/2 classfers are constructed. Snce we have consdered 5 classes of dsturbances, the total number of SVMs s. C. Dendogram based SVM (DSVM) The proposed DSVM takes advantage of both the effcent computaton of the ascendant herarchcal clusterng of classes and the hgh classfcaton accuracy of SVM for bnary classfcaton. Although DSVM needs (N-) SVMs for N class problem n the tranng phase, for the testng phase DSVM requres an optmal set of SVMs selected n a descendant way from the root of the taxonomy through the selected class among the leaf nodes. The DSVM method conssts of two maor steps:() computng a clusterng of the known classes and (2) assocatng a SVM at each node of the taxonomy obtaned by (). The frst step of DSVM method conssts of calculatng N gravty centers for the N known classes. Then AHC clusterng s appled over these N centers. Dendogram s constructed through the AHC method to classfy PQ dsturbances. The basc thought s as follows: frstly the PQ dsturbance set needng to be classfed s dvded nto two subsets accordng to the smlarty of the chosen feature vectors, and then the two subsets are dvded nto two subsets separately agan accordng to the same prncple. The dvson wll contnue untl the classfcaton task s fnshed. The mult-class SVM classfcaton tree of PQ dsturbances s shown n Fgure 9. It can be seen that there are 4 SVMs n the mult-class SVM applcaton tree, and each SVM chooses dfferent feature vector to mplement bnary classfcaton t Fg. 9 Classfcaton way by DSVM 953

6 Internatonal Journal of Electrcal and Computer Engneerng 3:5 28 C-sag, C2-swell, C3-Interrupton, C4-Harmoncs, C5- Flcker Fgure 9 shows an example of a taxonomy done by AHC [6] algorthm over the N classes. In the second step, each SVM s assocated to a node and traned wth the elements of the two subsets of ths node. For example, n Fgure whch llustrates clusterng of 5 classes SVM s traned by consderng elements of {C2} as postves and elements of {C,C5,C3,C4} as negatves; SVM2 s traned by consderng elements of {C4} as postves and elements of {C,C3,C5} as negatves. SVM4 s traned by consderng elements of {C} as postves and elements of {C5} as negatves. In fnal, we wll tran (N - ) SVM for N-class problem. The advantage of tranng n DSVM s to a pror separate the classes n a herarchcal way. That facltates the class separaton for the SVM. In fact, SVM found easly boundary separaton between {C2} and {C,C3,C4,C5}. The level of dffculty for boundary separaton ncreases from the root through the leaves. The dea of DSVM s that t s preferable to solve many small problems n a herarchcal way than to solve a complex great problem. For classfyng a pattern query, DSVM presents t to the root SVM whch provdes an output for rght or left on the taxonomy. The procedure s repeated for each selected node n the way of the classfcaton (Fgure 3) untl arrvng to a leaf whch fnally represents the assocated class for our pattern query. V. APPLICATION AND RESULTS A. Sgnal Modelng Sgnal modelng by parametrc equatons for classfers tests was advantageous n some aspects. It was possble to change testng and tranng sgnal parameters n a wde range and n a controlled manner. Sgnals smulated that way were very close to realty. On the other hand, dfferent sgnals belongng to the same class gave the opportunty to estmate the generalzaton ablty of classfers based on Neural Networks. Sgnals belongng to sx man groups of dsturbances [9], were smulated. The classes and respectve parametrc equatons for smulaton of sgnals are summarzed n Table. As a specal group, the nondsturbed sgnals were chosen (pure snusod).the ranges of sgnals parameter varaton are shown n Table 2.The varaton range corresponds to values measured n real power systems. The parameters ss and sw correspond to the depth of sag and swell respectvely. The step functon (t) s used to determne sag and swell duraton. Flcker s characterzed by ts frequency fw and f ampltude. In order to obtan representatve sgnals for the most common power qualty dsturbances to serve the purpose of tranng, as well as the testng of the PNN classfer, power qualty dsturbance sgnals are smulated usng Matlab. Sx categores of dsturbances are smulated, namely, undsturbed snusod, sudden swell, sudden sag, nterrupton, harmoncs, and voltage flcker. The dsturbances are based on ten cycles of voltage waveform. These waveforms are generated at a samplng rate of 256 samples / cycle for a total of 3 ponts. TABLE IPARAMETRIC EQUATIONS FOR SIMULATION OF DISTURBED SIGNALS Event Pure Snusod TABLE II PARAMETERS VARIATION IN SIMULATED SIGNALS Event Pure Snusod Sag Swell Interrupton Parameters Varaton Ampltude : Frequency : 5 Hz Duraton : t 2 t 9 Ampltude : ss.3.8 T t t T Duraton : 2 8 Ampltude : ss.3.7 Order : 3,5,7 Harmoncs Ampltude :. 9 Flcker Frequency : 5 Hz..2 f The descrbed experments were mplemented usng the MATLAB7 envronment, on a Pentum IV, 2.88 GHz, wth 256 Mb of memory. The parameters used to vary the classes of events are the depth, the angle, the startng tme and the duraton of the events, whch are defned below. The depth of the event s defned as the change n the ampltude of a sgnal. The angle represents the phase shft at whch the sgnal s captured. The startng tme s the tme at whch the event starts. The duraton s the tme perod of the event. In ths study the above descrbed parameters were vared accordng to the IEEE recommended practce n [2]. The angle of the sgnal s vared from % to % of the entre perod (whch s a realstc assumpton snce the captured waveforms n a practcal montorng system could have a phase shft that may vary from to 2). In addton, the startng tme of the sag s vared from % to 8% of the total length of the sgnal. Moreover, the duraton of the sag s vared from 5% to % of the total length of the waveform. Equaton v t snwt Sag v t ss t t t t2 snwt Swell v t sw t t t t2snwt Interrupton v t t t t t2snwt Harmoncs k snwtk3 sn3wt v t wt wt k5 sn 5 k7 sn 7 Flcker v t sn wtsnwt f f 954

7 Internatonal Journal of Electrcal and Computer Engneerng 3:5 28 For the nterrupton and the swell events, the four parameters (depth, angle, startng tme and duraton) were also vared as descrbed above. The depth of the swell s ncreased from % to 9% of the magntude of the pure sne waveform. For the harmoncs events, 2nd, 3rd, 5th, 7th, 9th, and th harmoncs are used to randomly contamnate the deal waveforms. Durng the generaton of such events, the total harmonc dstorton (THD) of the waveform was kept greater than 5%, as suggested n [2]. In the case of the flcker events, the ampltude of the smulated sgnals was changed perodcally to ntroduce the effect of a flcker. To acheve ths, the magntude of the target waveform was vared as a functon of another sne wave. Ths results an oscllaton n the ampltude of the target waveform, whch vared randomly from 5% to 7% of the fundamental frequency. In order to obtan representatve sgnals for the most common power qualty dsturbances to serve the purpose of tranng, as well as the testng of the PNN classfer, power qualty dsturbance sgnals are smulated usng Matlab. Fve categores of dsturbances are smulated namely, sudden swell, sudden sag, nterrupton, harmoncs, and voltage flcker. Undsturbed pure sne wave s consdered as a specal case. The dsturbances are based on twelve cycles of voltage waveform. These waveforms are generated at a samplng rate of 256 samples / cycle for a total of 3 ponts. B. Result of the Proposed Classfcaton Method Based on the feature extracton by the Wavelet-Transform method, we wll perform a 3-level decomposton of each dscrete dstorted sgnal to obtan the detaled verson coeffcents. The features extracted from wavelet transform were appled to the ANN Classfers and SVMs for recognzng and classfyng the dstorted sgnals. The -Fold Cross Valdaton Evaluaton Results of the ANN and SVM based Classfers for the fve data sets s shown n Table 3&5 respectvely. Cross-valdaton, s the practce of parttonng a sample of data nto subsets such that the analyss s ntally performed on a sngle subset, whle the other subset(s) are retaned for subsequent use n confrmng and valdatng the ntal analyss.the test result shows that the SVM classfer attans better recognton rates when compared wth the ANN classfer. TABLE III -FOLD CROSS VALIDATION EVALUATION RESULT OF THE ANN CLASSIFIER Class BPN RBF PNN Sag Swell 7 8 Interrupton 8 8 Harmoncs Flcker Overall Classfer TABLE IV PERFORMANCE OF ANN Tranng Testng Tme (sec) Tme (sec) Accuracy (%) BPN RBF 5 88 PNN TABLE V -FOLD CROSS VALIDATION EVALUATION RESULT OF THE SVM CLASSIFIER Class One One DSVM Aganst All Aganst One Sag Swell Interrupton Harmoncs Flcker Overall Classfer One Aganst One One Aganst All Dendogram SVM TABLE VI PERFORMANCE OF SVM Tranng Tme(sec) Testng Tme( sec) Accuracy (%) VI. CONCLUSION AND FUTURE WORK Artfcal Neural Networks (ANN) and Support Vector Machnes (SVM) are the two most popular machne learnng technques now used n classfcaton. The numercal experments performed for both: ANN and SVM networks have confrmed that both solutons are very well suted for classfcaton. ANN reles mostly on heurstcs whereas SVM s mathematcally well founded. Out of the three ANN technques consdered n ths paper, the PNN performs well n terms of accuracy and tme of computaton. But the maor drawback of PNN s that all ts tranng vectors must be stored requrng large amount of memory. In the case of SVM, the dendogram based SVM acheves 96% accuracy and outperform the other two methods n terms of computaton tme. The general percepton for ths power dsturbances classfcaton problem s that SVM may outperform ANN n terms of classfcaton accuracy. The observed dfferences n performance are n most cases neglgble. However the man dfference s n the complexty of the neural networks. Because the dstorted sgnals n ths study were generated by smulaton, employng real dstorted sgnals measured by the dgtal recorder to mprove the proposed method for more number of dsturbances s one of our future works. 955

8 Internatonal Journal of Electrcal and Computer Engneerng 3:5 28 REFERENCES [] Mshra S, C. Bhende N, Pangrah B. K., Detecton and Classfcaton of Power Qualty Dsturbances Usng S-Transform and Probablstc Neural Network, IEEE Transactons On Power Delvery, January 28, Vol. 23,pp [2] Gang Zwe-Lee, Wavelet-Based Neural Network for Power Dsturbance Recognton and Classfcaton, IEEE transactons on Power delvery,october 24,vol. 9, pp [3] Jank Przemyslaw and Lobos Tadeusz, Automated Classfcaton of Power-Qualty Dsturbances Usng SVM and RBF Networks, IEEE Transactons On Power Delvery, July 26,Vol. 2, Pp [4] S. Chen and H. Y. Zhu, Wavelet Transform for Processng Power Qualty Dsturbances, EURASIP Journal on Advances n Sgnal Processng Volume 27, Artcle ID 47695, 2 pages. [5] Borras Dolores, Castlla M, Moreno Narcso and Montano J.C., Wavelet and Neural Structure: A New Tool for Dagnostc of Power System Dsturbances, IEEE Transactons on Power Delvery, January 2,Volume 37, pp [6] Specht, D. F., "Probablstc neural networks", Internatonal Neural Network Socety, Neural Networks, Vol3, 99, pp [7] Z.Chen,Senor Member, IEEE, and P.Urwn Power Qualty Detecton and Classfcaton Usng Dgtal Flters IEEE Porto Power Tech Conference 2. [8] Peter G. V. Axelberg, Irene Yu-Hua Gu, Senor Member,, Support Vector Machne for Classfcaton of Voltage Dsturbances, IEEE, and Math H. J. Bollen, Fellow, IEEE, IEEE Transactons On Power Delvery, Vol. 22, No. 3, July 27 [9] Chh-We Hsu and Chh-Jen Ln, A Comparson of Methods for Multclass Support Vector Machnes, IEEE Transactons On Neural Networks, Vol. 3, NO. 2, March 22 pp [] Khald Benabdeslem and Youn`es Bennan, Dendogram- based SVM for Mult-Class Classfcaton, Journal of Computng and Informaton Technology - CIT 4, 26, [] I.W.C.Lee, Member,IEEE, and P.K. Dash, Senor Member, IEEE S- Transform-Based Intellgent System for Classfcaton of Power Qualty sturbance Sgnals, IEEE Trans on Industral Electroncs,Vol 5,No.4, Aug 23. [2] IEEE Recommended Practce for Montorng Electrc Power Qualty, IEEE/Std Manmala.K earned a bachelor of engneerng n Electrcal and Electroncs from Government College of Engneerng, Trunelvel, Inda. She earned a Master s degree n Computer scence and Engneerng from Manonmanam sundaranar Unversty, Inda n 24. She has 2 years teachng experence. She s currently an Assstant Professor at the computer Scence and Engneerng Department of Dr.Svanth Adtanar College of Engneerng, Truchendur, Inda. She s pursung PhD n the area of Data Mnng. Dr.K.Selv obtaned B.E(EEE) wth Honours,M.E (Power System) wth Dstn, from Madura Kamara Unversty n the year 989 and 995 respectvely.she obtaned Ph.D n Electrcty Deregulaton n June 25 from Madura Kamara Unversty. She s currently workng as Assstant Professor n Department of Electrcal Engneerng, n Thagaraar college of Engneerng, Madura.,Tamlnadu, Inda. She has obtaned Young Scentst Fellowshp from Dept. of Scence and Technology. Her research nterests are Electrcty deregulaton and AI technques. Ahla.R receved the B.E. degree n Electrcal and Electroncs engneerng n 995 and the M.E. degree n Computer Scence and Engneerng n 24 from Manonmanum Sundaranar Unversty,Trunelvel, Inda.She s currently an assstant professor at the computer Scence and Engneerng Department of Dr.SvanthAdtanar College of Engneerng, Truchendur,,Inda. 956