The Classification of Power Swing Based on PNN and MLPNN

Transcription

1 Joural of Patter Recogitio ad Itelliget Systems Aug. 3, Vol. Iss., PP The Classificatio of Power Swig Based o P ad MLP Ahmad Farid Bi Abidi *, Azah Mohamed, Hussei Shareef 3 Ceter for Power Egieerig Studies, Faculty of Electrical Egieerig Uiversiti Tekologi MARA, Shah Alam, 445, Selagor, Malaysia, 3 Departmet of Electrical, Electroic ad Systems Egieerig Uiversiti Kebagsaa Malaysia, Bagi, 436, Selagor, Malaysia * ahmad94@salam.uitm.edu.my; azah@eg.ukm.my; 3 shareef@eg.ukm.my Abstract-A itelliget approach is developed to discrimiate a fault, stable swig ad ustable swig for correct distace relay operatio by usig the S-trasform ad the probabilistic eural etwork (P). To illustrate the effectiveess of the proposed techiques, simulatios were carried out o the IEEE 39 ad a practical test system usig the PSS/E ad MATLAB software. Test results show that the P gives a overall classificatio accuracy of 97.33% i which it performs better tha MLP i detectig ad classifyig ustable swig, stable swig, fault, fault clearace ad post fault evets. Such fast ad accurate itelliget detectio schemes are useful for prevetig distace relay from trippig durig power swig. Keywords- Ustable Swig; Stable Swig; S-trasform; Probabilistic eural etwork (P); Multi Layer Perceptro eural etwork (MLP) I. IRODUCTIO Power systems are large itercoected oliear systems where system wide istabilities ca occur whe the system is subjected to sudde evets as the loss or applicatio of large blocks of load, lie switchig, geerator discoectio ad faults []. The ever icreasig load demad ad the deregulated power market recetly pushed the power system to operate close to its stability limits which makes the system vulerable to wide area istabilities or collapses, ad fially cascadig blackout that ca affect millios of people ad cause huge ecoomic loss [-3]. Cascadig blackouts ca be iitiated due to may reasos ad oe of the promiet causes of such evet is uiteded operatio of distace relays [4]. Curretly, distace protectio with respect to trasiet istability is oe of the critical issues i trasmissio systems. The covetioal zoe 3 distace protectio relays o trasmissio lies may ot be able to distiguish betwee voltage istability ad short circuit faults [5]. This situatio ca lead to udesired operatio of relays ad as a cosequece, the system ca be exposed to voltage collapse. Power swigs or oscillatios occur followig a system disturbace such as load chage or fault clearace. Power swig has bee idetified as oe of major causes that brig a power system to a major blackout as reported i [6-7]. Whe a power swig occurs, the measured impedace ca eter the relay strippig zoe to iitiate trippig sigals of associated breakers [8]. A chage appears i the relative phase agle betwee two groups of geerators ad as a cosequece, the measured voltage, curret, apparet impedace, active power, reactive power ad agle vary due to oscillatios durig power swig [9]. From the aforemetioed electrical quatities, may techiques have bee developed to prevet relay trippig durig a power swig [- 4]. Apostolov et al. [5] has itroduced a superimposed compoet which is prove to be a very formidable tool for power swig detectio. A major advatage of the superimposed compoet based o faulted phase selectio method is that it does ot require ay settigs ad is ot sigificatly affected by the magitude of the pre-fault load curret. It also works very well uder evolvig fault coditios. However, the techique seems to be computatioally iefficiet as the computatio process requires three phase siusoidal currets to differetiate a fault ad a power swig. Jiao et al. [] has used a combiatio of the swig ceter s voltage waveform ad sythetic egative sequece vector to differetiate betwee a fault ad a power swig. The techique seems to be rigorous i discrimiatig power swig ad high fault resistace for protectio purposes. However, it requires demadig computatio of derivative operatio for the swig ceter s voltage waveform. There is also a time delay of about 3-4 ms before a fault ad a swig ca be idetified ad hece the method is relatively slow as compared to the techique proposed by Josso & Dalder []. Su et al. [] has developed a techique based o Vcos which activated a power swig detector i about 3-5 ms. The techique, however, requires further testig i large power systems before it ca be deployed to a relay. Xiagig et al. [3] used the derivatives of real ad reactive powers to develop a ublockig scheme for distace protectio durig symmetrical faults i power systems. The ublockig scheme seds the trip sigals after 3 ms i the evet of a fault. This techique is cosidered complicated ad computatioally iefficiet because it requires istataeous product of voltage, curret ad agle to obtai the real ad reactive powers. A more advaced techique usig adaptive euro fuzzy system has bee developed to block the relay trip sigals durig power swigs [4]. However, the techique has ot bee validated o the zoe 3 relay operatio settig

2 Joural of Patter Recogitio ad Itelliget Systems Aug. 3, Vol. Iss., PP cosiderig that this zoe is the most vulerable zoe durig power swigs. I additio, the relay respose time is greater tha 4 ms which is very slow as compared to other techiques developed i [], [-3]. A very fast ad reliable swig detector has bee proposed by Afzali ad Esmaeilia [6] ad Mahamedi ad Zhu [7]. However, both techiques were ot able to distiguish the stable ad ustable swig. Pag et al. [8] has itroduced the use of wavelet trasform (WT) to detect the power swig occurrece i power system. The method extracts the travellig waves from trasiet sigals iduced by faults ad calculates the eergy of high frequecy compoets extracted usig wavelet trasform. A approach which is based o Fast Fourier Trasform (FFT) for power swig detectig scheme is preseted by Mahamedi [9]. This techique is based o the detectio of frequecy compoet of the threephase active power. It is demostrated that durig power swigs the frequecy of the three-phase active power is equal to slip frequecy, whereas after symmetrical fault iceptio time the frequecy of the three-phase active power equals to 5Hz. oetheless, both techiques have ever bee tested for ustable swig coditio. From the literature, the existig techiques are cosiderably slow ad could trigger false relay operatio durig fast power swig ad fault clearace operatio at the adjacet lie. Hece, it is importat to develop a fast ad rigorous approach for detectig a fault, fault clearace operatio of a circuit breaker at a adjacet lie, stable power swig ad ustable to prevet udesirable distace relay operatio. Due to the promisig performace of AI techiques i various power system applicatios, these techiques are employed for more accurate ad selective operatio of distace relays. To address the eed for fast detectio of ustable swigs so as to improve the reliability of distace relay operatio, a ew scheme for detectig a fault, stable swig ad ustable swig at trasmissio lies is proposed by usig the S-trasform ad P. Here, the P is used because it is cosidered as a effective tool for solvig may classificatio problems. II. IELLIGE TECHIQUES FOR PREVEIG DISTACE RELAY OPERATIO DURIG POWER SWIG The S-trasform which is a advaced sigal processig techique ad artificial eural etwork techiques such as the P ad the multi-layer perceptro eural etwork (MLP) are applied for detectig stable swig, ustable swig ad fault i distace relay operatio. The backgroud theories of the MLP, P ad SVM are first preseted ad the followed by its implemetatio i the distace relay operatio. A. Multilayer Perceptro eural etwork ad Probabilistic eural etwork The MLP is a feedforward eural etwork with oe or more hidde layers []. A typical cofiguratio of the etwork cosists of a iput layer, oe or more hidde layers ad a output layer. A MLP with oe hidde layer is show i Fig.. X Y X Y X i i w ij j w jk k Y 3 X m l Y 4 Fig. Topological structure of MLP X Iput uit Patter uit Summatio uit X Output uit Y i X 3 X 4 Fig. The topological structure of P

3 Joural of Patter Recogitio ad Itelliget Systems Aug. 3, Vol. Iss., PP The iput layer accepts the iput sigals ad redistributes these sigals to all the euros i the hidde layer. I order to determie the output of the euros, a et weighted output is computed ad passed through a activatio fuctio. The probabilistic eural etwork (P) is a kid of radial basis etwork suitable for solvig classificatio problems []. It uses kerel-based approximatio to form a estimate of the probability desity fuctios of classes i a classificatio task []. The P architecture is composed of may itercoected processig uits or euros orgaized i successive layers. The topological structure of P which comprises of four layers with oe iput layer, two hidde layers ad oe output layer is show i Fig.. B. Classificatio of Three Phase Fault, Stable ad Ustable Swigs Usig To address the eed for fast detectio of ustable swigs so as to improve the reliability of distace relay operatio, a ew scheme for detectig a fault, stable swig ad ustable swig at trasmissio lies is proposed by usig the S-trasform ad artificial eural etworks. The S-trasform is used to extract features of sigals obtaied durig a fault, stable swig ad ustable swig whereas artificial eural etworks based o MLP ad P are used to classify either a fault, stable swig or ustable swig for correct distace relay operatio as show i Fig. 3. Durig a stable swig, it is compulsory to block the trippig sigals, while for ustable swig the trippig sigals have to be triggered to the associated breaker for isolatio purposes. Iput data Features extractio usig derivative of voltage ad S-trasform Classificatio process usig MLP or P Iput to distace relay Fig. 3 Automated detectio of three phase fault, stable ad ustable swigs usig the S-trasform, MLP ad P C. S-Trasform as Feature Extractio Approach As for the detectio of a ustable swig, a fast trippig actio eeds to be triggered at trasfer lie to the associated breaker for isolatio purposes. S-trasform is a time-frequecy represetatio kow for its local spectral phase properties. A key feature of the S-trasform is its accurate time-frequecy (amplitude ad phase) domai by employig a movig ad scalable localizig Gaussia widow [3]. The basis fuctio for the S-trasform is the Gaussia modulated cosie wave whose width varies iversely with frequecy. where, The S-trasform of a discrete time series is give by k, m, =,,..., - T: samplig iterval : total of samplig poit S kt, Equatio () ca be further simplified as m m - H e e, m () m where, S kt, m m m H, () m e ad e Proper selectio of iput features is a importat step before implemetig the P ad MLP. The iput features of the MLP ad P are selected by cosiderig the derivative of voltage, the voltage ad active power processed by the S-trasform. The mathematical formulatio of the iput features is described accordigly. The first feature, F, is give by, j V F, if (3) T

4 F4 Joural of Patter Recogitio ad Itelliget Systems Aug. 3, Vol. Iss., PP where The secod feature, F, is derived from the S-trasform ad is give by, m V F, if, (4) T m 5 m 5 m 5 V V (5) F F m The third feature, F 3, which is also derived from the S-trasform is give by, m 5,if, (6),if, (7) where, V : voltage The fourth feature, F 4, is give by, m m 5 m 5 F3 V (8) where, m P : active power at a selected m 5 m m m 5 m F P P 4 (9) The iput features are usually ormalized i the rage of to for A applicatio. The iput features (F, F ad F3) ad (F, F ad F4) are plotted i three dimesios as show i Figs. 4 ad 5, respectively. The plots show that the features ca characterize the various evets which are grouped ito differet feature coordiates. The output features of P ad MLP have bee determied o the basis of the evets which might affect the relay operatio durig fault, stable swig ad ustable swig..5. fault ustable swig ormal stable swig fault clearace F F.6.8. Fig. 4 Plots of iput features F, F ad F3 As for the output features, the developed P classifies fault, stable swig, ustable swig, fault clearace ad post fault by represetig the evets as for fault, for stable swig, 3 for ustable swig, 4 for fault clearace ad 5 for post fault. As for MPL, the output features cosists of [ ] which deotes a fault, [ ] as ustable swig, [ ] as stable swig, [ ] as fault clearace ad [ ] as post fault. The MLP outputs are ot i crisp value of or,

5 F3 Joural of Patter Recogitio ad Itelliget Systems Aug. 3, Vol. Iss., PP but rather i the rage of to. Hece, for classificatio purpose, a decisio rule is used so that if the MLP output is less tha.9 or greater tha., it is cosidered as a misclassified output.. fault ustable swig ormal stable swig fault clearace F F.6.8. Fig. 5 Plots of iput features F, F ad F4 III. RESULTS OF A FOR CLASSIFYIG THREE PHASE FAULT, STABLE AD USTABLE SWIGS Power swig simulatios were carried out to geerate traiig data. A fault with duratio 35 ms is triggered at differet locatios of the test systems i order to create differet cases of stable ad ustable swigs. Figs. 6 ad 7 show the examples of the time domai simulatios illustratig cases of stable ad ustable swigs Fig. 6 Time domai simulatio illustratig a stable swig Fig. 7 Time domai simulatio illustratig a ustable swig

6 Joural of Patter Recogitio ad Itelliget Systems Aug. 3, Vol. Iss., PP For the P ad MLP implemetatio, 6 traiig data sets cosistig of five particular evets have bee geerated, while 75 differet data sets have bee used for testig purpose. The results obtaied from the P ad MLP for detectig ad classifyig stable swig, ustable swig, fault, fault clearace ad post fault for distace relay operatio are preseted. The fault clearace ad post fault have to be cosidered as the P ad MLP outputs because the features i both situatios ca be misclassified as a fault or ustable swig if o specific outputs are assiged. The performace of both MLP ad P i predictig the detectio time of ustable swig has bee evaluated by comparig it with the detectio times obtaied from simulatios. A. MLP Results The architecture of the MLP is such that it has 4 iput euros represetig the 4 iput features, two hidde layers with euros ad 8 euros, respectively ad 5 output euros represetig stable swig, ustable swig, stable swig, fault clearace ad post fault coditios. It uses the hyperbolic taget trasfer fuctio ad the back propagatio algorithm i the eural etwork traiig. The mea squared error which is used as a goal for traiig the eural etwork is set at.. The performace goal was met at 56 epochs with a traiig time of secods. The MLP was tested with the 75 sets of testig data ad it was foud to give a overall accuracy of 93.33% with five misclassificatios, amog which, three misclassificatios are recorded as fault clearace evet while the remaiig misclassificatios are recorded as post fault evet. The MLP sample testig results are show i Table i which the bold umber i the table deotes the misclassificatio of evets. As show i Table, the MLP outputs are ot i crisp value of or, but rather i the rage of to. Hece, for classificatio purpose, a decisio rule is used such that if the MLP output is less tha.9 or greater tha., it is cosidered as a misclassified output. TABLE I THE MLP TESTIG RESULTS B. P Results Data Actual MLP SS SS SS SS PF PF PF PF PF PF PF FC FC FC FC FC FC FC F F F F Accuracy 93.33% The developed P classifies fault, stable swig, ustable swig, fault clearace ad post fault by represetig the evets as for fault, for stable swig, 3 for ustable swig, 4 for fault clearace ad 5 for post fault. Table shows the sample of P testig results i which bold umber i the table deotes the misclassificatio of evet. It was foud that P gives a overall accuracy of 97.33% with two misclassificatios recorded as stable ad ustable swigs. Comparig the testig results of P ad MLP, it ca be cocluded that the performace of P is better tha MLPP i classifyig differet types of power evets durig power oscillatio pheomea

7 Joural of Patter Recogitio ad Itelliget Systems Aug. 3, Vol. Iss., PP C. Detectio Time of Ustable Swig TABLE II THE P TESTIG RESULTS Data Actual P Data Actual P Data Actual P SS PF 5 5 F6 SS PF 5 5 F7 SS3 PF3 5 5 F8 SS4 PF4 5 5 F9 SS5 PF5 5 5 F SS6 FC 4 4 F SS7 FC 4 4 F SS8 FC3 4 4 F3 SS9 FC4 4 4 F4 SS FC5 4 4 F5 SS FC6 4 4 US 3 3 SS FC7 4 4 US 3 3 SS3 FC8 4 4 US3 3 3 SS4 FC9 4 4 US4 3 3 SS5 FC 4 4 US5 3 3 PF 5 5 FC 4 4 US6 3 3 PF 5 5 FC 4 4 US7 3 3 PF3 5 5 FC3 4 4 US8 3 3 PF4 5 5 FC4 4 4 US9 3 3 PF5 5 5 FC5 4 4 US 3 3 PF6 5 5 F US 3 3 PF7 5 5 F US 3 3 PF8 5 5 F3 US3 3 3 PF9 5 5 F4 US4 3 5 Accuracy 97.33% The compariso with the frequecy deviatio techique which has proposed by So et al., (7) is coducted i order to ascertai the validity ad advatages of the propose approaches. I this part, the P has bee chose istead of MLP due to the high accuracy as metioed i earlier part of this paper. The frequecy deviatio is oe of the reliable approaches to idetify the ustable swig. The results of the techiques have bee preseted at Table 3. From the result it ca be observed that P approach is faster as compared to the existig techique. Moreover, the techiques are applicable at trasmissio etwork protectio with less complexity as compared to frequecy deviatio which acquires the data from geerator to be set to the trasmissio. TABLE III THE P TESTIG RESULTS RESULT DETECTIG TIME OF USTABLE SWIG Data P (secod) /d Secod) Data.5.6 Data.8.45 Data3.5.6 Data Data Data Data7.6.6 Data8.6.6 Data Data.3.5 Data Data.45.5 Data Data Data

8 Joural of Patter Recogitio ad Itelliget Systems Aug. 3, Vol. Iss., PP IV. COCLUSIOS To automate the evet detectio ad classificatio process, results of itelliget classificatio of three phase fault ad voltage collapse usig SVM for correct distace relay operatio are preseted. The test results showed that the SVM performs better tha the P i detectig ad classifyig three phase fault ad voltage collapse whe compared with the actual simulatio outputs. Such detectio scheme is useful for prevetig distace relay from trippig durig voltage collapse. Fially, the results of usig the proposed combied S-trasform ad P to detect ad classify fault, stable swig, ustable swig, fault clearace ad post fault are preseted. The results showed that the P gives a better performace compared to the MLP i terms of accuracy. Apart from that, the P also proves to be faster o detectig ustable swig evet as compared to the frequecy detectio techique which is proposed by So et. al. V. ACKOWLEDGME The authors gratefully ackowledge the Uiversiti Tekologi MARA for fiacial support i terms of scholarship ad Uiversiti Kebagsaa Malaysia for fiacial support o the project. REFERECES [] Sea, C., Fraco, R. & Giusto, A. 8. Assessmet of power swig blockig fuctios of lie protective relays for a ear sceario of the Uruguaya system. IEEE/PES Trasmissio ad Distributio Coferece ad Expositio: Lati America 8, pp. -6. [] Zi, A. A. & Karim, S. P. 7. The Applicatio of Fault Sigature Aalysis i Teaga asioal Berhad Malaysia. IEEE Trasactio o Power Delivery (4): [3] Istardi, D., Abba-Aliyu, S., Bergqvist, A., Rouch,., Abdalrahma, A. M. & Tua, L. A. 9. Uderstadig Power System Voltage Collapses Usig ARISTO: Effects of Protectio Itercoected Power Systems. IEEE PowerTech, 9, pp. -7. [4] Krebs, R., Ziegler, G., Sieme, J. J., Balasiu, F. & Lazar, F. 9. Avoidig Cascadig Protectio Trippig i Trasmissio Systems ad Assessmet of Quality ad Security of the Protectio System Power Systems. IEEE PowerTech 9, pp. -7. [5] M. Ji ad T. S. Sidhu, A adptive Load Ecroachmet Scheme for Distace Protectio. Electr. Power Syst. Res. J., vol. 78, pp , 8. [6] Martuscello, L., Krizauskas, E., Holbach, J., & Lu, Y. 9. Tests of distace relay performace o stable ad ustable power swigs reported usig simulated data of the august 4th 3 system disturbace. 9 Power Systems Coferece: Advace Meterig, Protectio, Cotrol, Commuicatio, ad Distributed Resources, PSC 9, pp [7] Zhag,., & Kezuovic, M. 5. A study of sychroized samplig based fault locatio algorithm performace uder power swig ad out-of-step coditios. 5 IEEE Russia Power Tech, PowerTech, pp. -7. [8] Moore, P., & Johs, A ew method of power swig blockig for digital distace protectio. IEE Proceedigs: Geeratio, Trasmissio ad Distributio, 43 (), 9-6. [9] Blumschei, J., Yelgi, Y., & Kereit, M. 8. Proper detectio ad treatmet of power swig to reduce the risk of Blackouts. 3rd Iteratioal Coferece o Deregulatio ad Restructurig ad Power Techologies (DRPT 8), pp [] M. Josso ad J. Daalder, Distace protectio ad voltage stability. Iteratioal Cof. o Power Syst. Techology, pp ,. [] Jiao, S., Bo, Z., Liu, W., & Yag, Q.. ew Priciples to Detect Faults Durig Power Swig. Seveth IEE Iteratioal Coferece o Developmets i Power System Protectio, hlm [] Su, B., Wag, J., & Xu, Y. 7. Directioal elemet of distace relay for symmetrical fault durig power swig. 8th Iteratioal Power Egieerig Coferece, (IPEC 7), hlm [3] Xiagig, L., Qig, Z., Weju, L., Kecheg, W., & Hali, W. 6. Fast Ublockig Scheme for Distace Protectio to Idetify Symmetrical Fault Occurrig Durig Power Swigs. IEEE Power Egieerig Society Geeral Meetig, hlm. -8. [4] Zadeh, H., & Li, Z. 8. Artificial eural etwork based load blider for distace protectio. IEEE Power ad Eergy Society 8 Geeral Meetig: Coversio ad Delivery of Electrical Eergy i the st Cetury, PES, p. art. o [5] A. Apostolov, D. Tholomier, ad S. Richards, Superimposed compoets based sub-cycle protectio of trasmissio lies. 4 IEEE PES Power Syst. Cof. ad Expositio, vol., pp , 4 [6] M. Afzali, Esmaeilia, A ovel algorithm to idetify power swig based o superimposed measuremets, th Iteratioal Coferece o A. Eviromet ad Electrical Egieerig (EEEIC), pp. 9-3,. [7] B. Mahamedi, Jiaguo Zhu, A ovel Approach to Detect Symmetrical Faults Occurrig Durig Power Swigs by Usig Frequecy Compoets of Istataeous Three-Phase Active Power, IEEE Trasactios Power Delivery, vol. 7, iss. 3, pp ,. [8] C. Pag ad M. Kezuovic, Fast Distace Relay Scheme for Detectig Symmetrical Fault durig Power Swig, IEEE Trasactios o Power Delivery, vol. 5, iss. 4,. [9] B. Mahamedi, A Very Fast Ublockig Scheme for Distace Protectio to Detect Symmetrical Faults durig Power Swig, Coferece Proceedigs IPEC,. [] egevitsky, M. 5. Artificial Itelligece A Guide to Itelliget Systems. Essex: Addiso Wesley. [] Mao, K. Z., Ta, K.-C., & Ser, W.. Probabilistic eural-etwork Structure Determiatio for Patter Classificatio. IEEE Trasactios o eural etworks, (4), 9-6. [] Zakich, A. 3. eural etworks for itelliget sigal processig. Sigapore: World Scietific Publishig Co. Pte Ltd. [3] R. G. Stockwell, L. Masiha, ad R. P. Lowe, Localizatio of the complex spectrum: the S trasform. IEEE Tras. o Sigal Processig, vol. 44, pp. 998-,

9 Joural of Patter Recogitio ad Itelliget Systems Aug. 3, Vol. Iss., PP Ahmad Farid Bi Abidi received the B.E.E.S degree from Uiversity Kebagsaa Malaysia (UKM), Malaysia, i, M.S.E.E. from Uiversiti Tekologi MARA (UiTM), Malaysia, i 5 ad Ph.D. from Uiversity Kebagsaa Malaysia (UKM), Malaysia, i. He is curretly a lecturer at the Faculty of Electrical Egieerig, UiTM Shah Alam, Malaysia. His research iterests are i power system protectio ad power quality. Azah Mohamed received her B.Sc. from Uiversity of Lodo i 978 ad M.Sc. ad Ph.D. from Uiversiti Malaya i 988 ad 995, respectively. She is a professor at the Departmet of Electrical, Electroic ad Systems Egieerig, Uiversiti Kebagsaa Malaysia. Her mai research iterests are i power system security, power quality, artificial itelligece ad distributed geeratio. She is a seior member of IEEE Hussai Shareef received his B.Sc. with hoours from IIT, Bagladesh, M.S. degree from METU, Turkey, ad Ph.D. degree from UTM, Malaysia, i 999, ad 7, respectively. He is curretly a associate professor i the Departmet of Electrical, Electroics, ad Systems Egieerig, Uiversiti Kebagsaa Malaysia. His curret research iterests are power system deregulatio, power quality ad power system distributio automatio