ONLINE POWER SYSTEM CONTINGENCY SCREENING AND RANKING METHODS USING RADIAL BASIS NEURAL NETWORKS

Transcription

1 Internatonal Journal of Electrcal and Electroncs Engneerng Research (IJEEER) ISSN(P): X; ISSN(E): X Vol. 6, Issue 3, Jun 2016, 9-16 TJPRC Pvt. Ltd ONLINE POWER SYSTEM CONTINGENCY SCREENING AND RANKING METHODS USING RADIAL BASIS NEURAL NETWORKS KULDEEP SAINI 1 & AKASH SAXENA 2 1 Research Scholar, Swam Keshwanand Insttute of Technology, Management & Gramothan, Japur, Inda 2 Assocate professor, Swam Keshwanand Insttute of Technology, Management & Gramothan, Japur, Inda ABSTRACT Ths paper presents a supervsng learnng approach usng Mult Feed Forward Neural Network(MFFN) and Radal Bass Fucton Neural Network(RBFN) to deal wth fast and accurate statc securty assessment (SSA) and contngency analyss of a large electrc power systems. The degree of severty of contngences s measured by two scalar performance ndces (PIs): Voltage-reactve power performance ndex, PIVQ and lne MVA performance ndex, PIMVA. For each (N-1) contngency, theperformance Index (PI) s computed usng the Newton Raphson (NR) method. A correlaton coeffcent feature selecton technque has been utlzed to dentfy the nputs for the MFFN and RBFN. The proposed method has been appled on an IEEE 39-bus New England test system at dfferent operatng condtons comparng to sngle lne outage and the results demonstrate the sutablty of the methodology for on-lne power system securty assessment at Energy Management Center. The performace of the proposed ANN models s compared wthnewton Raphson (NR) method and the results shows that the proposed model s effectve and relable n terms of statc securty assessment of power systems. KEYWORDS: PerformanceIndex, Statc Securty Assessment, Contngency Analyss, Supervsed Learnng, Mult Feed Forward Neural Network, Radal Bass functon Network Orgnal Artcle Receved: Mar 28, 2016; Accepted: Apr 15, 2016; Publshed: Apr 25, 2016; Paper Id.: IJEEERJUN INTRODUCTION Securty assessment of a power systemplays an mportant role for on-lne applcatonsat Energy Management Centerwhch employs to examne the steady state performance of a power system after contngency. Nowdays, electrc power system move towards a new envronment that s deregulaton whch has forced modern electrc utltes to operate ther systems under hghly stressed condtons closer to ther securty lmts. Therefore, the system operators needs to develop quck and more precse rankng methods for analyzng the power system securty volatons and severty level of contngences to keep the power system n safe operatng lmts. As ndcated by the fast development of Machne Learnng (ML) applcatons n power system area, ths paper presents a revew of Artfcal Neural Network (ANN) n Statc Securty Assessment (SSA). Many research papers have dscussed dfferent methods to smulate and rank the contngences for nstance, automatc contngency selecton based on a pattern analyss as proposed by Rodrgues [Rodrgues, 1999]. Ths method s capable to dentfy the potental harmful contngences.on the other hand, Dynamc securty assessment (DSA) enables to determne whch contngences may cause power system operatng lmt volatons or system nstablty.so far, the only Transent Energy Functon (TEF) analytcal methods are suggested for dynamc securty assessment [2] and the use of Artfcal Intellgence ncludng Neural Networks and Expert Systems [3]. In [4] Transent stablty assessment of power system usng Probablstc Neural Network (PNN) and Least Squares Support Vector Machne edtor@tjprc.org

2 10 Kuldeep San & Akash Saxena (LS-SVM) was presented.for a gven severty of contngences any on-lne TSA tool must provde a fast and accurate stablty evaluaton and system securty analyss under random load perturbatons.system securty can be assessed usng contngency analyss. Contngency analyss dentfed those cases whch are harmful to the system and ranked them accordng to ther severty s referred to as contngency selecton.rankng methods employs two scalar performance ndex (PI) to measure the severty of each sngle lne outage [5-7].In [8] a cascade neural network-based approach s proposed for fast lne flow contngency selecton and rankng. The developed cascade neural network s a combnaton of a rankng module and a flter module. Generally PI-based methods are dffcult for onlne applcatons because of hgh computatonal tme needed. Recently, artfcal neural networks have shown a great promse for contngency screenng and rankng due to ther ablty to accurate nput-output mappngs. To estmate the severty of contngences a pattern recognton technque s proposed n [9-11].Over the past few years, several approaches have been proposed for statc securty assessmentusng artfcal neural networks (ANN) [12] [14]. In [15] author employs ANN for onlne contngency screenng and rankng and found that ANN s effectve n terms of accuracy and speed. In ths paper,two four-ed Mult Feed Forward Neural Network(MFFN) and Radal Bass Functon Neural Network(RBFN) are proposedwhch predct the degree of severty of contngences.the study nvolves RBFN s more accurate and faster than MFFN as RBF networks reduce the computaton tme requred for tranng. Further due to mnmum tranng error, the optmal learnng can be acheved by RBF networks. The proposed ANN models are traned wth Reslent back propagaton algorthm for nput output mappng [16]. SYSTEM PERFORMANCE INDEX FORMULATION In ths paper the Performance Index (PI) s defned n terms of both bus voltage volatons and lne overloads s used for securty level classfcaton and to rank contngences n order of ther severtyfollowng a gven lst of contngences. The severty of a lne contngency s measured by two scalar performance ndces (PI), namely the lne MVA performance ndex (PIMWA) and voltage reactve power performance ndex (PIVQ). Lne MVA Performance Index (PIMVA) Most of the lterature on contngency rankng show that the PI rankng methods s greatly nfluenced by lne flow performance of the system. Lne MVA Performance Index (PIMVA) [11] s expressed by the followng scalar PI: PIMVA = post 2n N l W L S l max l 1 M = l S (1) Where S post l s the post-contngency MVA flow of lne l, S lmax the MVA ratng of lne l, N l the number of lnes n the system, W L the real non-negatve weghtng factor (=1). n s the nteger exponent. In ths paper the value of n s fxed as 4 for the IEEE 39-bus test system. Voltage-Reactve Performance Index (PIVQ) The severty of a contngency to out of lne voltage lmts and generator reactve powervolatons s gven by: Impact Factor (JCC): NAAS Ratng: 2.40

3 Onlne Power System Contngency Screenng and Rankng 11 Methods usng Radal Bass Neural Networks PIVQ = N B = 1 W M V V V V Lm sp M + N G = 1 W M G Q Q max 2n (2) Where V Lm = voltage devaton lmt; V = post contngent voltage at the th bus; V sp = rated voltage magntude correspondng to bus ; WV s the weghng coeffcent(=1); Q = reactve power at bus ; Q max= upper lmt for reactve power generaton at bus ; NG the number of generatng unts, NB number of buses, n s the nteger exponent where n=4; W G = real non-negatve weghtng factor(=1). Accordng to the PIs value, three dfferent securty states have been consdered for the contngences whch are mentoned as; Secure state f PI < 0.2, Crtcal state f 0.2 < PI < 0.8, Insecure state f PI > 0.8. SUPERVISED LEARNING APPROACH FOR ONLINE STATIC SECURITY ASSESSMENT In ths paper, Supervsed learnng approach s used whch automatcally learn to recognze complex patterns and make ntellgent decsons based on nput/output parameters. The proposed model s based on Supervsed Machne learnngapproach employsmult feed forward neural network (MFFN) and Radal Bass Functon Neural Network (RBFN) for onlne statc securty assessment andfast contngency screenng and rankng of power systems. Ths paper uses Reslent Back Propagaton Algorthm for tranng the neural network. It s an adaptve weght learnng algorthm, whch adapt the weght step based on the local gradent nformaton.due to ts weght update propertes the reslent back propagaton algorthm converges much faster than the conventonal back propagaton algorthm.two types of ANNs are used for onlne statc securty assessment and contngency analyss; MFFN and RBFN. PROPOSED METHODOLOGY Proposed ANN Methodology for Performance Indces The proposed ANN model selected for on-lne securty assessment s a mult feed forward neural network (MFFN) and Radal Bass Functon Neural Network (RBFN) traned wth Reslent back propagaton algorthm.to avod msclassfcaton, separate rankng s obtaned for PIVQ and PIMVA usng two MFFNs and RBFN as shown n the block dagram of the proposed model n Fgure 1(a) and Fgure 1(b). For each contngency, the performance ndexes are calculated by off-lne Newton Raphson method. YI=[PG10,QG1,QG2, QG3,QG4,QG5,QD14, QG7,QG8,QG9,QG10] MFFN-I YO=PIVQ Input Hdden Output MFFN-II YO=PIMVA Input Hdden Output Fgure 1(a): Proposed MFFNModel edtor@tjprc.org

4 12 Kuldeep San & Akash Saxena Input X1 Weghts Radal bas functon f1 W1 W0 Lnear weghts Input X2 f2 W2 Output Y WK Output Layer fm Input Xk Hdden Layer Input Layer Fgure 1(b): Proposed RBFN model Mult-Layer Feed Forward Network (MFFN) In ths paper, MFFN consstng of three hdden s wth nonlnear actvaton functons s proposed for power system statc securty assessment. Real and reactve power generaton at varous generator buses and reactve power demand at load buses are chosen as the nputs to the MFFN. The actvaton functon used n the hdden unts s the Tansg and the output unts, the lnear functon s used. The network s traned wth Reslent back propagaton algorthm [16] due to ts good convergence propertes. For two MFFNsand RBFN the parameters are shown n Table 1. Radal Bass Functon Neural Networks (RBFN) The RBFN used n ths paper s shown n Fgure 1(b). The RBF network conssts of sngle hdden feed forward structure. The nput nodes pass the nput varables drectly to the hdden wthout any connecton weghts. The output of the th hdden unt α (X) s gven by 2 X µ ( X ) = exp (3) 2σ α 2 WhereX s the nput vector elements, µ s the vector whch determnes the center of the bass functon α, σ j s ther wdths. The kthoutput node value Yks gven as Y k H = = 1 w ( X ) ( X ) + k φ (4) w ko Where, wk = connecton weght between the output and hdden node wko = bas term and p s the number of the bass functon. In ths paper, RBFNgves faster convergence than the conventonal MFFN. Impact Factor (JCC): NAAS Ratng: 2.40

5 Onlne Power System Contngency Screenng and Rankng 13 Methods usng Radal Bass Neural Networks Feature Selecton One of the key ssues n ANN-based approach s the proper selecton of nputs to the neural networks tranng. In ths paper correlaton coeffcent technque has been used to select the approprate tranng features for the MFNN and RBFN. The correlaton coeffcent between the jth and kth varable s calculated usng (3) as C jk { } { } E { } X j X k X σ j σ k E E j X K = j, k = 1,2,3,4,,n (5) The varables that are less correlated are selected as features for the MFNN and RBFN. In ths paper precontngent real and reactve power generaton at the generator buses (PG, QG), reactve power loads (QD) are selected as the nput features for the proposed ANN tranng. Thus, 11 features (out of 62 features) are selected for the tranng of MFNN and RBFN. Data Generaton, Tranng, and Testng In ths paper, sngle lne contngency s consdered. The load patterns are generated by changng the loads randomly at each bus n the range of 80 and 143 % of ther base case values. For each load pattern (N-1) contngency s smulated by Newton Raphson method. For each contngency, the Performance Index (PI) s calculated usng (1) and (2). Nearly tranng patterns are generated for the proposed ANN model. The nput features selected by correlaton coeffcent method s normalzed n the range of 0.1 and 0.9 for each load pattern. Once the tranng of ANNs s accomplshed, the traned network s evaluated by test data. The contngences are ranked n terms of severty based on the performance ndex PI. TEST SYSTEMS AND SIMULATION RESULTS In ths paper the smulaton results of IEEE 39-bus New England system s dscussed. The effectveness of the proposed method showsthe sutablty of the methodology for on-lne power system securty assessment at Energy Management Center. The test system conssts of 10 generators, 12 transformers and 46 transmsson lnes [20]. The dagram of IEEE 39-bus New England system s shown n Fgure 2. Fgure 2: Sngle Lne Dagram of IEEE 39-Bus New EnglandTest System For contngency rankng bus no. 39 s taken as slack bus.the number of neurons n the hdden unts s 30 and edtor@tjprc.org

6 14 Kuldeep San & Akash Saxena % Base Case Lne No. Out Age No. Table 1: Tranng Parameters Parameter MLFFN RBFN Maxmum epochs Performance goal 0 - Mnmum gradent 1x e-10 - Learnng rate Momentum coeffcent Gaussan functon spread Table 2: Sample Results of PI Calculatons by MFNN and RBF PIVQ Class (PIVQ) PIMVA Class (PIMVA) NR MFN MFN MFNN- NR RBF NR NR RBF N-1 N III III II II II II I I II II II II II II I I I I I I I I I I II II II II Class I and Class II (crtcal); Class III (non-crtcal). 11,000 load patterns were generated by varyng the loads randomly at all the buses and generaton n the range of % of ther base case values. For 240 dfferent loadng condtons, 46 sngle lne outages are smulated for each loadng, to obtan dfferent operatng condtons. Contngency analyss has been performed by utlzng the pre-contngency data and the lne performance ndces for each load scenaro and each outage at a tme. A total of 8840 patterns have been taken to analyze the performance of the proposed model and for remanng cases Newton Raphson (NR) faled to converge. The test results of the proposed MFNN and RBF model for contngency screenng and rankng s shown n Table 2. It s observed that PI values obtaned by the proposed MFNN and RBF model are close to desred values of PI obtaned from Newton Raphson method. It can be concluded from the table that for performance ndces, PIVQ and PIMVA a separate rankng must be done. For sample result for % of base case, t s found that operatng constrants are volated for outage of lne 9 39, resultng n system s nsecure as both the PIs are n nsecure classes. Such result s expected snce bus No. 39 s a generator bus and lne outage 9 39 causes bus voltage lmt volaton, the reactve power generaton lmt volaton and the overloadng of the transmsson lnes connected to bus No. 9. The outage of lne for both 142.5% of base case makes the system crtcal as agan all the operatonal constrants are volated. Smlar analyss can be drawn for other lne outages. The number of cases that belong to the Class (I to III) obtaned from the proposed MFNN and RBF model and ther comparson wth Newton Raphson method s shown n Table 2. Table 1 shows the parameters for the proposed MFNN and RBF model. The PI values obtaned from MFNN and RBFneural network are very close wth PI values obtaned from Newton Raphson method. CONCLUSIONS In ths paper power system statc securty assessment has been nvestgated. The test results presented on IEEE 39-bus system provdes the followng observatons. MFN N-2 Impact Factor (JCC): NAAS Ratng: 2.40

7 Onlne Power System Contngency Screenng and Rankng 15 Methods usng Radal Bass Neural Networks A new method has been developed for calculatng voltage performance and lne performance ndex for contngency rankng whch elmnates msrankng and maskng problems. The performace of the proposed models s compared wth Newton Raphson (NR) method and the results shows that the RBF model s more effectve and relable n terms of statc securty assessment of power systems. Tranng s very fast as the RBF network has the capablty of handlng large amount of data. Testng tme s less than 0.25sec. REFERENCES 1. Rodrgues, M S, Souza, J C S, Do CouttoFlho, M B and Th. Schllng, M, Automatc Contngency Selecton Based on a Pattern Analyss Approach, Proceedngs of IEEE Internatonal Conference on Electrc Power Engneerng, Power Tech Budapest 1999, pp G.D. Irsar, G.C. Ejebe, J.G. Waght, W.F. Tnney, Effcent Soluton for Equlbrum Pont n Transent Energy Functon Analyss, IEEE Trans. on Power Systems, May 1994, pp EPRI Report, lw : Dynamc Securty Analyss-Feasblty Evaluaton Report, Aprl Noor W, Azah M, An H. Transent stablty assessment of a power system usng PNN and LS-SVM methods. J ApplSc 2007; 7(21): Swarup KS, Sudhakar G. Neural network approach to contngency screenng and rankng n power systems. Neuro computng 2006; 70: Chen, Y.; Bose, A. Drect rankng for voltage contngency selecton. IEEE Transactons on 7. Power Systems 1989; 4(4): Ghosh S, Chowdhury BH. Desgn of an artfcal neural network for fast lne flow contngency rankng. Electr Power Energy Syst 1996; 18(5): Sngh R, Srvastava L. Lne flow contngency selecton and rankng usng cascade neural network. Neurocomputng 2007; 70: K. L. Lo, L. J. Peng, J. F. Macqueen, A. O. Ekwue, and D. T. Y. Cheng, Fast real power contngency rankng usng a counterpropagaton network, IEEE Trans. Power Syst., vol. 13, pp , Nov Swarup KS. Artfcal neural network usng pattern recognton for securty assessment and analyss. Neurocomputng 2008; 71: Souza J, Flho M, Schllng M. Fast contngency selecton through a pattern analyss approach. Electr Power Syst Res 2002; 62: V. S. Vankayala and N. D. Rao, Artfcal neural network and ther applcaton to power system A bblographcal survey, Elect. Power Syst. Res., vol. 28, pp , 1993.S. Saeh and A. Kharuddn, Statc securty assessment usng artfcal neural network, n Proc. IEEE Int. Conf. Power and Energy, Dec. 2008, pp T. S. Sdhu and C. Lan, Contngency screenng for steady-state securty analyss by usng FFT and artfcal neural networks, IEEE Trans. Power Syst., vol. 15, no. 1, pp , Feb R. Fschl, Applcaton of neural networks to power system securty: Technology and trends, n Proc. IEEE World Congr. Computatonal Intellgence, Jul. 1994, vol. 6, pp edtor@tjprc.org

8 16 Kuldeep San & Akash Saxena 16. Redmller M, Braun H. A drect adoptve method for faster backpropagaton learnng: the PROP algorthm. IEEE IntConf Neural Networks 1993; 1: Subraman, C.; Dash, S.S.; ArunBhaskar, M.; Jagadeeshkumar, M.;Sureshkumar, K.; Parthpan, R. Lne outage contngency screenng and rankng for voltage stablty assessment ICPS '09. Internatonal Conference on Power Systems, Jan T, Srvastava L, Sngh SN. Fast voltage contngency screenng usng radal bass functon neural network. IEEE Trans Power Syst 2003; 18(4): Smon Haykn. Neural networks-a comprehensve foundaton. 2nd ed. Prentce Hall; Power system test cases at < 21. The Math Works, Inc., MATLAB programmng, Matpower 4.0 User s Manual: < manual>. Impact Factor (JCC): NAAS Ratng: 2.40