JJEE Volume, Numbe, 5 Pages - Jodan Jounal of Electical Engineeing ISSN (Pint): 49-96, ISSN (Online): 49-969 Discimination between Tansfome Inush Cuent and Intenal Fault using Combined DFT-ANN Appoach Eyad A. Feilat Depatment of Electical Engineeing, The Univesity of Jodan, Amman, Jodan e-mail: e.feilat@ju.edu.jo Received: Januay 4, 5 Accepted: Mach 7, 5 Abstact This pape pesents a digital potection technique based on combined Discete Fouie Tansfom (DFT) and atificial neual netwok (ANN) fo discimination between the magnetizing-inush and intenal-fault cuents in thee-phase powe tansfomes. A full-cycle DFT is fistly applied as a pepocessing module to extact distinctive featues, namely the magnitudes of the fundamental and second hamonic fequency components I and I, espectively, fom tansient diffeential phase cuents. The 3-phase cuent signals ae sampled at a sampling ate of samples pe cycle. The featues of phases a, b and c ae then used to calculate the second hamonic atio (SHR), I /I. Secondly, the thee SHRs ae fed into an ANN fo classifying the tansient phenomenon into eithe magnetizing inush o intenal-fault cuent. The task of the ANN unit is to develop a block signal when the SHR exceeds a theshold value. As a esult, a needless elay tipping when a tansfome has an inush cuent can be avoided. The ANN has the achitectue of an input laye, hidden laye, and output laye. The input laye has thee neuons epesenting the SHR of each diffeential phase cuent. The neuons of the hidden laye wee selected based on speed and accuacy. The output laye has one neuon with an output (no tip) fo inush cuent o (tip) fo intenal fault. The ANN has been tained using Levenbeg-Maquadt (LM) algoithm with log-sigmoid tansfe functions in the hidden and output layes, espectively. Taining and testing pattens of inush and fault cuents ove a wide ange of inception angles have been obtained by compute simulation of a 3-phase non-linea tansfome bank using MATLAB/Simulink. Simulation esults show that the poposed technique can be consideed as an effective digital potection appoach fo fast and accuate discimination between inush and intenal fault-cuents of powe tansfomes. Keywods ANN, DFT, inush cuent, simulation, tansfome potection. I. INTRODUCTION Powe tansfomes suffe fom the phenomenon of magnetizing inush cuent duing enegization when thei coes etain some esidual flux. The effect of this inush cuent could cause mall-opeation of the potective elays. Since the magnetizing inush cuent geneally contains a lage second hamonic component in compaison to a shot-cicuit cuent, conventional tansfome potective schemes ae designed to avoid false tipping due to inush cuent. They as well block o estain the elay opeation by sensing the lage second hamonic o by implementing delays in ovecuent o diffeential potection, o by using diffeent appoaches. Howeve, the second method is undesiable because of the potential dange of delaying time duing shot-cicuit conditions. Othe techniques equie lots of computing time []-[6]. Thee has been extensive eseach on applications of impoved digital potective methods to powe tansfome potection ove the last two decades. In paticula, atificial neutal netwoks have been developed fo accuate and effective discimination between inush and intenal-fault cuents [7]-[]. These ANNs employ featue extaction techniques based on eithe time o fequency domain signals, o both. As opposed to conventional techniques, ANN techniques have seveal advantages ove conventional computing methods. These advantages include the ability to handle situations of incomplete and coupted data, the ability to lean fom examples, and the ability to genealize []. Based on these popeties, Coesponding authos' e-mail: e.feilat@ju.edu.jo
5 Jodan Jounal of Electical Engineeing. All ights eseved - Volume, Numbe and seveal successful applications, ANN seems to be a faste and moe eliable disciminating technique between tansfome inush and intenal-fault cuents. In this pape, a digital potection appoach of powe tansfomes is pesented. The poposed technique is used to disciminate a magnetizing inush cuent fom an intenal-fault cuent by combining DFT and ANN. A sliding DFT is fistly applied as a pe-pocessing module to extact distinctive featues fo each phase of the 3-phase tansient diffeential cuents detected by the digital diffeential elays. These featues epesent the instantaneous magnitudes of I and I of phases a, b and c extacted by applying a full-cycle sliding DFT to diffeential cuents. Secondly, the SHR of each phase is fed into an ANN fo classifying the tansient phenomenon into eithe magnetizing inush o intenal-fault cuent. The ANN is tained using LM algoithm with log-sigmoid tansfe functions in hidden and output layes. Taining and testing pattens of inush and fault cuents can be obtained by compute simulation of the tansfome equivalent cicuit duing inush and fault conditions and ove a wide ange of inception angles o fom on-line measuements. The output laye has one neuon with an output fo inush cuent o fo intenal-fault cuent, espectively. II. ARTIFICIAL NEURAL NETWORK Multilaye feed-fowad ANN is achitectue of highly inteconnected simple nonlinea pocessing elements (neuons) connected in paallel to pefom useful computational tasks such as patten ecognition o classification as an altenative to conventional computing appoaches. ANN computing chaacteistics ae distinguished fom conventional patten ecognition by thei capability to map complex and highly nonlinea input-output pattens. ANN can be used to classify pattens by selecting the output, which best epesents unknown input pattens in cases, whee an exact input-output elationship is not easily defined. It has been poven that a netwok with one hidden laye can pefom any nonlinea mapping and that no moe than two hidden layes ae needed fo most applications []. ANN has attacted attention in the last decades to solve poblems elated to electic powe system engineeing such as load foecasting, secuity assessment, economic dispatch, and fault detection and classification [3]. In its basic fom, an ANN consists of an input laye, one o moe hidden layes and an output laye. Each laye consists of a set of neuons o nodes that ae fully connected to the neuons in the next laye. The connections have multiplying weights associated with them. The node eceives its input fom eithe othe nodes o fom the outside wold. The sum of all weighted inputs epesents the node activation function. The output of the node is detemined by an output function which esponds to this activation. The numbe of neuons and hidden layes is poblem-based. The pocess of detemining the weights is called taining pocess. In the taining pocess, sets of input-output pattens ae associated by popely adjusting weights in the netwok, so that a sum of squaed eo function can be minimized: E = ΣE N p = N (p) (p) ΣΣ( t - O ) k k p k () whee E p is the patten eo, t k is the taget (desied) output, and O k is the actual output of the neual netwok. Vaious taining algoithms have been developed to adapt the weights in ANNs to educe the eo E p. The most popula leaning method fo taining ANNs is the back-popagation (BP) algoithm. This algoithm employs an iteative gadient descent ule to
5 Jodan Jounal of Electical Engineeing. All ights eseved - Volume, Numbe 3 adapt weights; and the eo is calculated and popagated backwads fom the output to the hidden and input layes []. Although the BP algoithm is effective, it suffes fom vey poo convegence ate. Theefoe, seveal appoaches including Newton s method, conjugate gadients, o the LM optimization technique have been developed. Among the above mentioned methods, the LM algoithm is widely accepted as the most efficient taining algoithm in the sense of ealization accuacy. It gives a good compomise between the speed of Newton s algoithm and the stability of the steepest descent method. Consequently, it constitutes a good tansition between these methods [4]. In this study, the poposed ANN has been tained using the LM leaning algoithm. III. CALCULATION OF THE SECOND HARMONIC RATIO Discimination between an inush and an intenal-fault condition is often based on inush detection algoithm to avoid the needless tip by magnetizing inush cuent. The discimination algoithm is based on the fact that a second-hamonic cuent component is pesent within the diffeential cuent when tansfome coe becomes satuated. Conventional method of inush cuent detection fo tansfome potection is based on compaing the second-hamonic and fundamental components. The atio I /I is so-called the SHR. If the SHR is geate than a set value, an inush cuent condition is assumed and tipping is pevented. Othewise, an intenal fault condition is assumed and a tipping signal is issued to disconnect the tansfome. The magnitudes of I and I can be digitally extacted using the Fouie appoach. In this pape, a sampling ate of samples pe cycle ( Hz) with onecycle window length is chosen. The elay stoes a full-cycle of cuent samples and feeds them to a full-cycle DFT to extact the magnitudes of I and I cuent components. Assume that the cuent wavefom is sampled N times pe peiod of the fundamental, and let the samples be denoted by i k =i(k t), the eal and imaginay pats of the n th hamonic (a n and b n ) can be found. In tems of cuent samples stating at the th sample, a, b and the n th hamonic can thus be calculated: () n () n () In of a ( ) n = + N + N k= πk ik cosn, N n=, () b ( ) n + N = + i N k= k πk sinn, N n=, (3) ( ) ( ) ( a ) ( b ) I = + ( ) n n n (4) The esult can be updated iteatively as each new sample becomes available. This is done by dopping the ealiest sample and adding the new sample: a b ( + ) n ( + ) n ( ) = an + N+ N ( ) = bn + N+ N π N [ i i ] cosn π N [ i i ] sinn (5) (6)
4 5 Jodan Jounal of Electical Engineeing. All ights eseved - Volume, Numbe whee i and i N+ ae the oldest and newest samples, espectively. Having detemined I and I magnitudes, the tansfome discimination potection is then implemented [5]. When a new input sample aives, the oldest sample is discaded. The SHR of each phase is calculated and fed to the ANN to identify the tansient cuent as the inush o intenal fault based on the value of the SHR. IV. NEURAL NETWORK TRAINING The poposed ANN consists of thee layes as shown in Fig.. The input laye has 3 neuons. The numbe of neuons of the hidden laye was selected based on a compomise between the speed and accuacy of the ANN duing taining and testing phases. W h SHR a W O SHR b SHR c W h W hk h W Om O / Input Laye Hidden Laye Output Laye Fig.. Multilaye feed fowad ANN An attempt was made to minimize the numbe of hidden neuons. The mean squaed eo (MSE) was used as a pefomance citeion in this wok. The output laye has one neuon, which gives a binay output of o. To tain the poposed ANN, a lage numbe of input-output pattens has been geneated. Input pattens epesent the instantaneous 3-phase SHRs of phases a, b and c detected by the diffeential elays of the tansfome. Associated output pattens ae eithe coesponding to an inush cuent duing inush tansient o coesponding to an intenal fault cuent. In this pape, a total of 5 simulations of inush and intenal-fault cuents have been geneated by compute simulation of a 3-phase nonlinea tansfome bank ove a wide ange of inception angles α ( o -8 o ), as shown in Fig..
5 Jodan Jounal of Electical Engineeing. All ights eseved - Volume, Numbe 5 Fig.. MATLAB/Simulink schematic of a 3-phase nonlinea tansfome To simulate inush cuents, the cicuit beake is kept open-cicuited, wheeas to simulate intenal faults, the cicuit beake is kept closed. The name plate of the 3-phase tansfome bank is given in the Appendix. Examples of cycles of simulated inush and intenal-fault cuent signals, coesponding to α= o and α=9 o, along with thei hamonic spectum ae shown in Fig. 3. 4 3 5 Inush Cuent (A) Inush Cuent (A) 5 - -..4.6.8...4.6.8. Time (s) -5 3 4 5 6 7 8 9 3 Hamomic Ode a) Taining 3-phase inush cuents and thei hamonic spectums at α= o
6 5 Jodan Jounal of Electical Engineeing. All ights eseved - Volume, Numbe 3 5 IA IB IC Inush Cuent (A) - Inush Cuent (A) 5 - -5-3..4.6.8...4.6.8. Time (s) - 3 4 5 6 7 8 9 3 Hamomic Ode b) Taining 3-phase inush cuents and thei hamonic spectums at α=9 o 5 4 3 IscA IscB IscC 4 35 3 IA IB IC 5 SC Cuent (A) - - -3-4 SC Cuent (A) 5 5-5 -5..4.6.8...4.6.8. Time (s) - 3 4 5 6 7 8 9 3 Hamonic Ode c) Taining 3-phase-fault cuents and thei hamonic spectums at α= o 5 4 3 IA IB IC 5 4 IA IB IC SC Cuent (A) - SC Cuent (A) 3 - -3-4 -5..4.6.8...4.6.8. Time (s) - 3 4 5 6 7 8 9 3 Hamonic Ode d) Taining 3-phase-fault cuents and thei hamonic spectums at α=9 o Fig. 3. Samples of 3-phase inush and intenal-fault cuents
5 Jodan Jounal of Electical Engineeing. All ights eseved - Volume, Numbe 7.9.8.7 3 6 5.8 x -6.6.4 5 8.6 8. SHR of Phase A.5.4.3 SHR of Phase A.8.6 3 9..4. 9. 6 3 4 5 6 7 8 9 a) Phase a 3 4 5 6 7.7 6.5 x -6 SHR of Phase B.6.5.4.3 3 9 5 8 SHR of Phase B.5 3 6 9 5 8..5. 3 4 5 6 7 8 9 b) Phase b 3 4 5 6 7.8.7.6 6.5 x -6 3 5 8 SHR of Phase C.5.4.3 3 9 5 8 SHR of Phase C.5 6 9..5. 3 4 5 6 7 8 9 c) Phase c 3 4 5 6 7 Fig. 4. Taining input pattens of SHRs of 3-phase inush and fault cuents These signals ae used to extact SHR of the thee phases (SHR a, SHR b and SHR c ) ove a - cycle tansient cuent at o, 3 o, 6 o, 9 o, o, 5 o and 8 o inception angles, as depicted in
8 5 Jodan Jounal of Electical Engineeing. All ights eseved - Volume, Numbe Fig. 4. As can be seen in Fig. 4, the SHR fo the inush tansient vaies with time and is much highe than that of the intenal fault. Theefoe, the two cases ae sepaable and can be easily classified by the ANN. Once the taining matices of the input-output pattens ae obtained, the LM algoithm is applied to tain the ANN. The tainlm of the MATLAB Neual Netwok Toolbox has been used with log-sigmoid functions fo the hidden and output layes [4]. V. SIMULATION RESULTS In this study, seveal tests have been pefomed to detemine the optimum numbe of neuons in the hidden laye based on accuacy and speed. It was found that an ANN of 3 neuons in the hidden laye is a good choice. It leads to an MSE<e -6 in epochs, as shown in Fig. 5. Theefoe, this achitectue (3 3 ) has been adopted in this pape. The taining pefomance of the poposed ANN is depicted in Fig. 6. As can be seen fom the plots, the ANN output pefectly fits the taget values ( fo inush cuent and fo intenalfault). It is evident that the ANN is able to classify the tansients of the thee phase cuents to eithe inush o intenal-fault. P e fo m a nc e is. 85 4 e - 7, G o a l is e - 6 T a in in g- B lu e G o al -B la c k - - -3-4 -5-6 -7 3 4 5 6 7 8 9 E p o c h s Fig. 5. MSE taining convegence of the poposed (3 3 ) ANN Intenal Fault.8 NN Output of Phase A.6.4. Inush 5 5 Patten Numbe a) Taining actual and taget output pattens of phase a
5 Jodan Jounal of Electical Engineeing. All ights eseved - Volume, Numbe 9 Intenal Fault Intenal Fault.8.8 NN Output of Phase B.6.4 NN Output of Phase C.6.4.. Inush Inush 5 5 Patten Numbe 5 5 Patten Numbe b) Taining actual and taget output pattens of phase b c) Taining actual and taget output pattens of phase c Fig. 6. Discimination taining pefomance of the poposed ANN The poposed ANN opeates in a static manne. The ANN is tained off-line. Once the desied pefomance is achieved, the weights of the ANN ae fozen. Upon completion of the taining phase, the genealization capability of the poposed ANN, when exposed to test pattens that ae diffeent fom the taining pattens, is tested. The testing pefomance of the ANN has been examined using input-output testing pattens epesenting inush and intenal-fault tansients at inception angles (5 o, 45 o, 75 o, 5 o, 35 o and 65 o ) as shown in Fig. 7. The esults of the testing pefomance ae shown in Fig. 8. The testing phase simulations illustate the efficiency of the poposed ANN. It can be seen that almost % of the tested inush and fault pattens have been successfully classified. This esulted in the coect no tip/tip output signal..9.8 x -6.85.8.75 75 5 35.6.4 SHR of Phase A.7.65.6 5 45 65 SHR of Phase A..8.55.5.6.45.4.4 3 4 5 6 7 8. 3 4 5 6 a) Phase a
5 Jodan Jounal of Electical Engineeing. All ights eseved - Volume, Numbe x -6.9.8.8.6 SHR of Phase B.7.6.5 5 45 75 5 35 65 SHR of Phase B.4..8.4.6.3.4. 3 4 5 6 7 8. 3 4 5 6 b) Phase b.8 5.5 x -6.7.6 45 75 35 SHR of Phase C.5.4.3 5 65 SHR of Phase C.5..5. 3 4 5 6 7 8 3 4 5 6 c) Phase c Fig. 7. Testing input pattens of SHRs of 3-phase inush and fault cuents Intenal Fault.8 NN Output of Phase A.6.4. Inush 4 6 8 Patten Numbe a) Testing actual and taget output pattens of phase a
5 Jodan Jounal of Electical Engineeing. All ights eseved - Volume, Numbe Intenal Fault Intenal Fault.8.8 NN Output of Phase B.6.4 NN Output of Phase C.6.4.. Inush Inush 4 6 8 Patten Numbe 4 6 8 Patten Numbe b) Testing actual and taget output pattens of phase b c) Testing actual and taget output pattens of phase c Fig. 8. Discimination testing pefomance of the poposed ANN VI. CONCLUSION In this pape, an ANN fo inush and intenal-fault cuents discimination based on the SHR has been developed. The ANN has been tained using LM algoithm to geneate a no tip () o tip () output signal accoding to the values of the SHR that ae extacted fom the tansient diffeential cuent detected by the diffeential elay. The SHR has been calculated using sliding full-cycle DFT. Compute simulations of inush and intenal-fault tansients of a 3-phase nonlinea tansfome bank, ove a wide ange of switching angles, show that the pefomance of the poposed ANN-based disciminato is eliable and vey encouaging. Simulation esults show that the poposed ANN is able to classify cuent tansients that have not been exposed to duing the taining phase. The poposed ANN-disciminato can be implemented by using dedicated digital diffeential elay; and it can be used to suppot o eplace the conventional tansfome diffeential elays. APPENDIX The paametes of the studied 3-phase 8/45-V, 5-Hz, 4.5-kVA nonlinea distibution tansfome bank ae as follows: =.5Ω, =.34Ω, x l = x l =.56Ω, x m =78.8Ω. All paametes given above ae efeed to the -V pimay winding. The nonlinea chaacteistic of the tansfome is shown in Fig. A..5 Voc (Vpu).5..4.6.8...4.6.8. I (Apu) Fig. A. Magnetization cuve of the nonlinea tansfome
5 Jodan Jounal of Electical Engineeing. All ights eseved - Volume, Numbe REFERENCES [] M. A. Rahman and B. Jeyasuya, A state-of-the-at eview of tansfome potection, IEEE Tans. Powe Delivey, vol. 3, no., pp. 534-544, 988. [] J. A. Skyes and I. F. Moison, A poposed method of hamonic estaint diffeential potection of tansfomes by digital compute, IEEE Tans. Powe Appaatus and systems, vol. 9, pp. 66-7, 997. [3] T. S. Sidhu and M. S. Sachdev, On-line identification of magnetizing inush and intenal faults in thee-phase tansfomes, IEEE Tans. Powe Delivey, vol. 7, no. 4, pp. 885-89, 99. [4] A. K. Al-Othman and K. M. El-Nagga, A New digital dynamic algoithm fo detection of magnetizing inush cuent in tansfomes, Electic Powe Components and Systems, vol. 37, no.4, pp. 355-37, Apil 9. [5] R. Sedighi and M. R. Haghifam, Detection of inush cuent in distibution tansfome using wavelet tansfom, Electic Powe Enegy Systems, vol. 7, pp. 36 37, 5. [6] K. N. Al-Tallaq, A. Dahab, and M. Abo Zahhad, Impedance-based algoithm fo the discimination between inush and shot-cicuit cuent in single-phase tansfomes, Electic Powe Components and Systems, vol. 3, no.6, pp. 593-64, June 3. [7] L. G. Peez, J. Fleching, J. L. Meado, and Z. Obadovic, Taining an atificial neual netwok to disciminate between magnetizing inush and intenal faults, IEEE Tans. Powe Delivey, vol. 9, no., pp. 434-44, 994. [8] P. Bastad, M. Meunie, and H. Regal, Neual-netwok-based algoithm fo powe tansfome diffeential elays, IEE Poc. Geneation, Tansmission and Distibution, vol. 4, no. 4, pp. 386-39, 995. [9] P. L. Mao and R. J. Aggawal, A novel appoach to the classification of the tansient phenomena in powe tansfomes using combined wavelet tansfom and neual netwok, IEEE Tans. Powe Delivey, vol. 6, no. 4, pp. 654-66,. [] J. Pihe, B. Gca, and D. Dolina, Impoved opeation of powe tansfome potection using atificial neual netwok, IEEE Tans. Powe Delivey, vol., no. 3, pp. 8-36, 997. [] Z. Moavej, D. N. Vishwakama, and S. P. Singh, ANN-based potection scheme fo powe tansfome, Electic Machines and Powe Systems, vol. 8, pp. 875 884,. [] S. Haykin, Neual Netwoks: a Compehensive Foundation, nd ed., Pentice Hall, NJ, 999. [3] M. El-Shakawi and D. Niebu, Atificial Neual Netwoks with Applications to Powe Systems, IEEE Sevice Cente, NJ, 988. [4] H. Demuth and M. Beale, Neual Netwok Toolbox Use s Guide fo Use with MATALB, The MathWoks.. [5] A.T. Johns and S. K. Salman, Digital Potection fo Powe Systems, Pete Peegines, 995.