Events in an underground distribution system can be

Transcription

1 Classfcaton of Load Change Transents and Incpent Abnormaltes n Underground Cable Usng Pattern Analyss Technques Mrrasoul J. Mousav, IEEE Student Member, Karen L. Butler-Purry, IEEE Senor Member Rcardo Guterrez-Osuna, IEEE Member, Masseh Najaf Abstract Ths paper presents a feasblty study on the applcaton of pattern analyss technques to classfy load change transents and ncpent abnormaltes n an underground dstrbuton cable lateral. The data were collected usng an onlne montorng system nstalled n a resdental area n Dallas. A set of features obtaned from wavelet packet analyss was evaluated. Methods of dmensonalty reducton were employed to overcome the curse of dmensonalty whle preservng a good classfcaton rate. The classfcaton results usng k-nearestneghbor (KNN) classfers show that the proposed methodology can be used to classfy load change transents and ncpent abnormaltes. Index Terms Underground dstrbuton cable, ncpent abnormaltes, wavelet packet analyss, pattern analyss. I. INTRODUCTION Events n an underground dstrbuton system can be broadly dvded nto two man categores, normal and abnormal. When normal events take place, no acton needs to be taken. However, when an abnormalty occurs, a seres of correctve actons need to be used to ascertan the safe and relable operaton of the system. Abnormal events and transents can be categorzed n terms of ther severty level. For nstance, transents due to load changes may occur frequently yet they are consdered low level transents. In fact, t s the customer demand that dctates the amount of power to be delvered at any partcular tme. On the contrary, ncpent fault-based events ntroduce hgh-level abnormaltes that need to be dentfed so that the necessary correctve acton can be taken. Hence, t s crucal to be able to dstngush among these events and ssue an approprate alarm sgnal upon completon of the detecton of such abnormaltes. The Ths work was supported n part by Oncor Servce Delvery and the Texas Advanced Technology Program through Grant No M. J. Mousav and K.L Butler-Purry are wth Texas A&M Unversty, Department of Electrcal Engneerng, Power System Automaton Laboratory; College Staton, TX (emal: klbutler@ee.tamu.edu, mousav@ee.tamu.edu). R. Guterrez-Osuna s wth Texas A&M Unversty, Department of Computer Scence, College Staton, TX (emal: rguter@cs.tamu.edu). M. Najaf s wth Texas A&M Unversty, Department of Mechancal Engneerng, College Staton, TX , (emal: mnajaf@tamu.edu) unque and dssmlar characterstcs of these events suggest utlzng pattern analyss technques for automated dentfcaton. Successful classfcaton of abnormaltes and transent events would be a great beneft to the utltes, enablng them to detect severe faults at an early stage of ther development, and consequently preventng unscheduled outages due to falures n underground cable. Ongong research at the Power System Automaton Laboratory (PSAL) at Texas A&M Unversty ams to develop an ncpent fault detecton method havng the capablty of predctng the remanng lfe of underground cables. Ths paper presents prelmnary work on the evaluaton of a set of features obtaned by a tme-frequency mult resoluton technque to classfy load change transents and ncpent abnormaltes n an underground dstrbuton cable lateral. Ths work was conducted to determne the most nformatve features that could dstngush between ncpent abnormaltes and load change transents. In ths paper, a load change transent s defned as an apprecable ncrease or decrease n the current sgnals, and an ncpent abnormalty s any actvty n the low frequency or hgh frequency sgnal pertanng to an ncpent behavor. Secton II provdes a concse descrpton of the data collecton system and explans the formulaton of the classfcaton problem n terms of pattern analyss termnologes. In secton III a thorough dscusson of the classfcaton results s gven. Secton IV provdes conclusons. II. PROBLEM FORMULATION As n any pattern analyss problem, there are four dstnct steps to translate a classfcaton problem nto a pattern analyss formulaton, data collecton, preprocessng of data, feature extracton, and model selecton / classfer desgn [1]- [2]-[3]. A. Data Collecton An underground dstrbuton cable lateral nstalled n a resdental area n Dallas was chosen to collect on-lne data. Ths ste was selected as the most approprate locaton to capture possble ncpent abnormaltes. Fg. 1 shows the experment ste ncludng the dstrbuton transformer and the underground cable. The underground dstrbuton cable lateral s fed from a standard 7200 v dstrbuton feeder and supples

2 power to the 7200v/120v/240v, 100 KVA, 60 Hz dstrbuton transformer va a normally open dstrbuton loop. 1 three of the output sgnals s to remove the domnant fundamental frequency (60Hz), thereby mprovng the magntude resoluton n the gven frequency range [4]. TABLE I CATEGORIES OF CURRENT SIGNALS Category Frequency Range of Output Sgnals 4 2 Low Frequency Sgnal Hz 3 Notch Low Frequency Sgnal Notch Hgh Frequency Sgnal Scale 1 (NHF x 1) Hz, Notch at 60 Hz KHz, Notch at 60 Hz Notch Hgh Frequency Sgnal Scale 10 (NHF x 10) KHz, Notch at 60 Hz 1 Cable Lateral Beng Montored 2 Phase Current CT 3 Neutral Current CT 4 Pad Mounted Dstrbuton Transformer Fg. 1. Montorng Ste n Dallas, Transformer and Cable Data collecton was performed usng an on-lne montorng system nstalled n the ste. The montorng system, whose block dagram s shown n Fg. 2, comprses three basc components: sgnal transducers, analog sgnal condtonng unt, and a computer-based data acquston system. The sgnal transducers transform the voltage sgnals to levels acceptable by the sgnal-condtonng unt. They also transform the current sgnals nto equvalent voltage sgnals of acceptable range. The transformed sgnals are then fed to the analog sgnal-condtonng unt whose functons are to act as an solaton unt, and to flter the sgnals nto varous frequency ranges. Sgnals from the analog sgnal-condtonng unt are fnally fed to the dgtal data-acquston system embedded n the computer. Substaton Feeder (7200 V) Feeders Dstrbuton Transformer 7200:120/240 V Loads Cable Lateral Measurement Unts Fg. 2. Block Dagram of the Montorng System Phase Current Neutral Current Phase Voltage Analog Sgnal Condtonng Unt Output Sgnals Dgtal Computer (Data Acquston System) Usng the montorng system, three basc electrcal sgnals, namely voltage, phase and neutral currents are observed. The system records the sgnals for one-second duraton every 15 mnutes. Moreover, varous statstcal and frequency parameters of these sgnals namely average, maxmum, mnmum, standard devaton and magntude of the harmoncs are calculated and recorded. In the sgnal-condtonng unt, the phase and neutral current sgnals create four outputs, as shown n Table I, unque n frequency range and scale, to ncrease the magntude resoluton durng data-acquston. The notch n B. Preprocessng Mechansm Recorded current sgnals may contan normal or abnormal actvtes, however, most of the recorded data represent normal operaton of the system. Furthermore, transent events can be ntated by load changes, ncpent faults, or other events. A preprocessng scheme was employed n order to flter out normal events and categorze the remanng sgnals n terms of ther correspondng predefned classes. A second motvaton for the use of preprocessng s to reduce redundancy. The samplng rate for the notch low frequency sgnals was set at samples/sec for recordng, whch gave a frequency range of Hz. However, these sgnals were lmted to Hz by a low pass ant-alasng flter n the analog sgnal-condtonng unt. Thus, to reduce redundancy, these sgnals were decmated by a factor of eght yeldng an effectve samplng rate of 1920 samples/sec. C. Feature Extracton / Selecton The goal of feature extracton or selecton s to obtan a few features that dscrmnate classes wth a hgh degree of accuracy. Ths mportant procedure s a key step for the success of any classfer. The feature extracton/selecton process nvolves two steps. Frst, a number of raw features are obtaned. Ths was accomplshed n the present work by means of Dscrete Wavelet Packet Analyss (DWPA) method [5]. Second, the raw features are projected onto a lower dmensonal space by means of multvarate statstcal technques n order to reduce the dmensonalty whle preservng a good classfcaton rate. These technques are brefly explaned n the followng sectons. 1) Wavelet Packet Analyss Desgn Analyzng data usng DWPA nvolves three steps, selecton of the type of mother wavelet, the order of mother wavelet, and the level of decomposton. A number of wavelet famles wth unque propertes have been proposed n the sgnal processng lterature, but the most approprate famly s generally applcaton-dependant. After lterature revew and from earler wavelet analyss results, t was found that the fourth order Daubeches wavelet yelds the best performance

3 for studyng power system transents and ncpent behavor [4]-[6]-[7]. Thus, the fourth order Daubeches wavelet was chosen as the mother wavelet for the analyss. Roughly speakng, selecton of the level of decomposton depends on the desred frequency resoluton. To obtan the best frequency resoluton, the 4 th level decomposton was chosen as llustrated n Fg. 3. Level 1 Level 2 Level 3 Level 4 S Fg. 3. 4th Level Wavelet Packet Decomposton Tree The wavelet packet analyss was conducted on sample length of one-second duraton. Thus, for one-second data, the low frequency sgnals contan 1920 samples and the hgh frequency sgnals contan samples. To obtan a better frequency range n the sxteen detals, the samples are zeropadded symmetrcally (at the begnnng and the end of the sgnal) to acheve dyadc sample length (2048 samples for low frequency sgnals and samples for hgh frequency sgnals). The approxmate frequency ranges for each of the detals at level 4 are 64 Hz and 0.5 KHz for a sgnal length equal to 2048 and 16384, respectvely. It should be noted that, for the sgnal wth 2048 samples, the frequency ranges at level 4 are such that the harmoncs of 60 Hz (fundamental frequency) are evenly dstrbuted among the detals. Therefore, zero-paddng the sgnal facltate the nterpretaton of analyss results n terms of sgnal harmoncs. 2) Formaton of Raw Feature Vector After 4th level wavelet packet decomposton, the resultng 16 detals were stored along wth the orgnal sgnal. Raw features are defned to be the maxmum magntude of spkes n each of these sgnals. The magntude of the spkes s a measure of contrbuton of that frequency range to the orgnal sgnal. To provde a better comparson among the detals, the magntude of the maxmum spke s normalzed by the magntude of the correspondng spke n the orgnal sgnal. The normalzed magntude of the spke was then stored n the raw feature vector. The vector thus conssts of 17 elements, 16 of whch represent the normalzed magntude of the spkes n the detals plus the magntude of the spke n the orgnal sgnal as the 17 th feature. Ths process was performed on every nput sgnal, resultng n a N 17 data matrx where N denotes the total number of examples n the data set (2110 examples). 3) Dmensonalty Reducton The objectve of dmensonalty reducton s to keep the dmensonalty of the pattern recognton problem (.e. the number of features) as small as possble whle preservng good classfcaton accuracy. Dmensonalty reducton can be accomplshed by means of feature selecton or feature extracton. The term feature selecton refers to technques that select the best subset of the nput features set. Methods that create new features based on transformatons and combnatons of the orgnal feature set are called feature extracton methods. The choce between feature selecton and extracton depends on the applcaton doman. Prncpal Component Analyss (PCA) s the best-known lnear unsupervsed feature extracton method [3]. The lnear transformaton s defned by the egenvectors of the covarance matrx, whch leads to vectors that are uncorrelated regardless of the form of the dstrbuton. If the dstrbuton happens to be Gaussan, then the transformed vectors wll be statstcally ndependent. The objectve of PCA s to perform dmensonalty reducton whle preservng as much as the randomness (varance) n the hgh-dmensonal space as possble. PCA performs a coordnate rotaton that algns the transformed axes wth the drectons of maxmum varance. The man lmtaton, however, s that as an unsupervsed method, t does not consder class separablty nformaton. There s no guarantee that the drecton of maxmum varance wll contan good features for dscrmnaton. Lnear Dscrmnante Analyss (LDA) s another wellknown lnear feature extracton method, but unlke PCA t s supervsed [2]-[3]. The objectve of LDA s to perform dmensonalty reducton whle preservng as much of the class dscrmnatory nformaton as possble. In LDA, nterclass separaton s measured by Fsher crteron, whch fnds the egenvalues of the between-class scatter matrx to the wthn-class scatter matrx. The wthn-class scatter matrx s defned by: c S w = S = 1 where, T S = ( x µ )( x µ ) (2) x w x denotes the data, c s the number of classes and µ s the mean vector of class w. The between-class scatter s defned by : c T S B = N ( µ µ )( µ µ ) (3) where = 1 (1) N s the number of patterns of class, and µ s the mean of the entre dstrbuton. The soluton proposed by Fsher s to maxmze the functon that represents the dfference between the means of the classes (between-class scatter) normalzed by a measure of the wthn-class scatter. The projectons wth maxmum classseparablty are the egenvectors correspondng to the largest egenvalues of S 1. Ths method produces as many w S B

4 projectons as the number of classes mnus one. If the classfcaton error estmates establsh that more features are needed, some other methods must be employed to provde addtonal features. LDA wll fal when the dscrmnatory nformaton s not n the mean of the data but rather n the varance. Raw features mght be expensve to obtan and mght not be numerc. Also, n some applcatons t may be mportant to extract meanngful rules from the classfer results. In such stuatons, feature extracton methods wll not work. Hence, feature subset selecton (FSS) methods must be employed. Feature subset selecton requres a search strategy to select canddate subsets and an objectve functon to evaluate these canddates. There are a large number of search strateges among whch Sequental Forward Selecton (SFS) a smple greedy approach was used n ths work. More detals about these methods can be found n [2] and [3] where statstcal pattern recognton technques are well ntroduced or revewed. Objectve functons are dvded nto two groups, flters and wrappers. Flters evaluate feature subsets by ther nformaton content; typcally nterclass dstance, statstcal dependence or nformaton-theoretc measures. Wrappers are essentally pattern classfers, whch evaluate feature subsets by ther predctve accuracy by statstcal resamplng or crossvaldaton. Flters are fast to be executed and ther results exhbt more generalty. However, they tend to select the full feature set as the optmal soluton. On the other hand, wrappers generally acheve better classfcaton rates than flters and have mechansm to avod overfttng. The man dsadvantage s slow executon. D. Model Selecton and Classfer Desgn Once the extracted/selected features are obtaned, the data set s organzed nto classes as shown n Fg. 4. There are two broad classes, load change transents and ncpent abnormaltes. Each class n turn contans four subgroups, Phase Notch Hgh Frequency (PH_NHF), Neutral Notch Hgh Frequency (NE_NHF), Phase Notch Low Frequency (PH_NLF), and Neutral Notch Low Frequency (NE_NLF). Based on these categores, three pattern analyss problems are defned. Two 4-class problems for each of load change transents and ncpent abnormaltes, whether the abnormalty s manfested n PH_NHF, PH_NLF, NE_NHF or NE_NLF sgnals. The thrd problem s a 5-class problem for whch the four classes of the load change transents are consdered as one class and categores of ncpent abnormaltes form the remanng four classes. Once a feature selecton or extracton procedure fnds a proper representaton, a classfer can then be desgned usng a number of possble approaches. In practce, the choce of a classfer s based on whch classfers happen to be avalable, or best known to the user. In ths study, k-nearest neghbor (KNN) classfers were used. In these classfers, the K closest examples n the tranng data set are found and the majorty class s determned and assgned to the unlabeled example. LOAD CHANGE TRANSIENTS PH_NHF PH_NLF NE_NHF NE_NLF Fg. 4. Problem Formulaton as a Classfcaton Problem III. CLASSIFICATION RESULTS INCIPIENT ABNORMALITIES PH_NHF PH_NLF NE_NHF NE_NLF The classfcaton results for each of the problems are shown n the followng fgures. The term orgnal n the fgures mples that all seventeen features n the orgnal feature space are used n the classfcaton wthout dmensonalty reducton. The classfcaton rate (CR) s a measure of the performance of the classfer defned by: number of correct assgnment s C R = (4) number of total assgnment s A. Results of the 4-Class Classfcaton Problem for Load change transents Fg. 5 summarzes the classfcaton results for the 4-class problem wth load change transents. As seen n the fgure, the classfcaton rate for classfyng data n the orgnal space was 66%. Applyng PCA, the classfcaton rate rose to 68%. The two largest egenvalues were 89.22% and 5.51% responsble for the varance of the data. Hence, usng PCA, only two features were determned to represent the data n the feature space. Applyng LDA, the classfcaton rate was around 57%, whch means that the dscrmnatory nformaton s not only n the mean of the data. Classfcaton Rate (%) Class Problem (Load) Orgnal PCA LDA FSS Feature Extracton/Selecton Method Fg. 5. Classfcaton Results for Load 4-Class Classfcaton Problem Sequental forward feature selecton wth wrapper objectve functon selected features 17, 3, 11, and 12. From the frequency range mentoned earler, t s nferred that features 3, 11 and 12 represent the contrbuton of the thrd, eleventh and twelfth harmoncs, respectvely. Feature 17 represents the contrbuton of all harmoncs n the sgnal. The classfcaton rate rose up to 82% when only those four features were used. Not only does FSS provde better performance than PCA or LDA, but also selected features can be techncally nterpreted n terms of harmonc contents of sgnals.

5 B. Results of the 4-Class Classfcaton Problem for ncpent Abnormaltes Fg. 6 summarzes the classfcaton results for the 4-class problem on ncpent abnormaltes. Classfcaton rate n the orgnal space was 68%. After applyng PCA, the classfcaton rate rose to 73%. The four largest egenvalues were 82.75%, 6.87%, 2.44%, and 1.41% responsble for the varance of the data. Applyng LDA, the classfcaton rate was not mproved, whch means that agan the dscrmnatory nformaton for ths class s not only n the mean. Sequental forward feature selecton wth a wrapper objectve functon selected features 17, 13, 11, and 12 whch correspond to the normalzed magntude of the spkes n the orgnal sgnal and ts 11 th, 12 th, and 13 th harmoncs. The classfcaton rate ncreased up to 82% usng these four features. number of ncpent abnormaltes, whch are domnant. Second reason les n the fact that the best feature (the magntude of the man frequency component) to dstngush load change transents from ncpent abnormaltes s not ncluded n the set of features. Recall that all the sgnals were fltered by a notch at 60 HZ. In general, the magntude of spkes n the ncpent abnormaltes s smaller than that of load change transents. As the order of harmoncs ncrease, the spke magntude becomes smaller and smaller. Therefore, t s qute reasonable to assume that the man frequency component wll carry dscrmnatory nformaton. It was observed that the classfcaton rate s rased to 83% after addng ths feature to the KNN classfer. Fg. 8 shows scatter plot for two features selected by FSS method. As shown, the boundary that separates the two classes s nonlnear. 4-Class Problem (Incpent) Classfcaton Rate (%) Orgnal PCA LDA FSS Feature Extracton/Selecton Method Fg. 6. Classfcaton Results for Incpent 4-Class Classfcaton Problem C. Results of the 5-Class Problem for Load Change Transents and Incpent Abnormaltes Fg. 7 depcts the classfcaton results for the 5-class problem defned earler. In ths case, the classfcaton rate n the orgnal space was 57%. After applyng PCA or LDA, the classfcaton rate was not mproved. Sequental forward feature selecton wth a wrapper objectve functon selected features 17, 13, 2, and 1 resultng n 72% classfcaton rate. Classfcaton Rate (%) Class Problem Orgnal PCA LDA FSS Feature Extracton/Selecton Method Fg. 7. Classfcaton Results for 5-Class Classfcaton Problem As t s seen, the classfcaton rate s lower than the other 4-class problems. By further nvestgaton, t was found that the msclassfed examples mostly belonged to the load change transent class whereas ncpent ones were correctly classfed. The frst reason has to do wth the number of exstng examples from load change transents relatve to the Fg.8. Scatter Plot of Two Selected Features IV. CONCLUSIONS Three classfcaton problems to categorze load change transents and ncpent abnormaltes n the underground dstrbuton cable were defned and solved. The classfcaton was performed usng seventeen features obtaned from wavelet packet analyss. In each classfcaton problem, methods of dmensonalty reducton were employed. It was observed that the feature subset selecton method had a better performance as compared to PCA or LDA feature extracton methods. The fnal classfcaton results usng KNN classfers were encouragng n all three cases. Future work ncludes explorng addtonal features and utlzng powerful classfers such as Support Vector Machnes (SVM) [8] to further mprove the classfcaton rate. ACKNOWLEDGMENT The authors acknowledge Rtesh Chaturbed for hs contrbuton to ths research. REFERENCES [1] R. O. Duda, P. E. Hart, D. G. Stork, Pattern Classfcaton, 2 nd Edton, New York, Wley, [2] A. Webb, Statstcal Pattern Recognton, London, Arnold, 1999.

6 [3] Anl K. Jan, Robert P.W. Dun and Janchange Mao, Statstcal Pattern Recognton: A Revew, IEEE Transactons on Pattern Analyss Intellgence Vol. 22, No.1, Jan 2000, pp [4] R. Chaturbed, Characterzaton and Detecton of Incpent Underground Cable Falures, MS Thess, Dept. of Electrcal Engneerng, Texas A&M Unversty, College Staton, August [5] Matlab reference book for wavelet toolbox, The Math Works, Verson 2, [6] Chul H. Km and Raj Aggarwal, Wavelet transforms n Power Systems, Power Engneerng Journal, August 2001, pp [7] A. W. Gall and O. M. Nlsen, Wavelet Analyss for Power System Transents, IEEE Computer Applcatons n Power, vol.12, no.1, January 1999, pp [8] N. Crstann, J. Shawe-Taylor, An ntroducton to Support Vector Machnes and other Kernel-Based Learnng Methods, London, Cambrdge Press, BIOGRAPHIES Mrrasoul Jaafar Mousav receved the B.Sc. degree from the Isfahan Unversty of Technology n 1996 and the M.Sc. degree from Sharf Unversty of Technology n 1999 all n Electrcal Power Engneerng. He worked for Nroo Research Insttute (NRI) from 1999 to Hs research nterests are related to fault detecton, antcpaton and dagnoss, transent studes, protecton and dstrbuton automaton. He s a student member of IEEE and IEEE Power Engneerng Socety (PES). He s currently pursung hs PhD degree n the Electrcal Department at Texas A&M Unversty. Karen L. Butler-Purry s an assocate professor n the department of electrcal engneerng at Texas A&M Unversty. In , Dr. Butler- Purry was a Member of Techncal Staff at Hughes Arcraft Co. n Culver Cty, Calforna. Her research focuses on the areas of computer and ntellgent systems applcatons n power, power dstrbuton automaton, and modelng and smulaton of power systems and vehcles. Dr. Butler-Purry s a senor member of IEEE and IEEE Power Engneerng Socety (PES), and a member of the Lousana Engneerng Socety. She s a regstered professonal engneer n the states of Lousana, Texas, and Msssspp. Rcardo Guterrez-Osuna (M 00) receved the B.S. degree n Industral/Electroncs Engneerng from the Polytechnc Unversty of Madrd n 1992, and the M.S. and Ph.D. degrees n Computer Engneerng from North Carolna State Unversty n 1995 and 1998, respectvely. From 1998 to 2002 he served on the faculty at Wrght State Unversty. He s currently an assstant professor n the Department of Computer Scence at Texas A&M Unversty. Hs research nterests nclude pattern recognton, machne learnng, bologcal cybernetcs, machne olfacton, speechdrven facal anmaton, computer vson and moble robotcs. Masseh Najaf receved hs B.Sc. degree from Unversty of Tehran n Iran n He s currently pursung hs graduate studes n the Mechancal Engneerng Department at Texas A&M Unversty. He s a research assstant n Energy System Laboratory (ESL) at Texas A&M Unversty. He s a student member of ASME. Hs research nterests are fault detecton, sensor dagnoss, ntellgent systems, and control.