Artificial Neural Network Based Backup Differential Protection of Generator-Transformer Unit

Transcription

1 Interntionl Journl of Eletronis nd Eletril Engineering Vol. 3, No. 6, Deemer 05 rtifiil Neurl Network sed kup Differentil Protetion of Genertor-Trnsformer Unit H. lg nd D. N. Vishwkrm Deprtment of Eletril Engineering, Indin Institute of Tehnology (HU) Vrnsi, Vrnsi, Indi Emil: {hrish.lg.rs.eee, H. Nth Deprtment of Eletril nd Instrumenttion Engineering, Thpr University, Ptil, Indi Emil: The NN-sed lgorithms hve een suessfully implemented in mny pttern or signture reognition prolems, s they n detet helthy onditions of genertor nd trnsformer sed on reognizing their wve shpes, more preisely, y differentiting them from the fult urrent wve shpes [3]-[5]. In [6], Neurl Network Priniple omponent nlysis long with Rdil sis Funtion Neurl Networks is used s pttern lssifier. In other words, this tehnique mkes the deision sed on the urrent signture verifition whih is more urte thn trditionl hrmoni restrint sed tehniques used for the protetion of trnsformer. This tehnique ould produe the tripping signl in the event of internl fult within 5ms fter fult ourrene. Optiml Proilisti Neurl Network (PNN) used in [7] s the ore lssifier to disriminte etween inrush nd internl fult. Prtile Swrm Optimiztion is used to otin optiml smoothing ftor for PNN. PNN requires lrger storge for exemplr ptterns & it is more diffiult to trin owing to numeril diffiulties. new pproh sed on deision tree for disrimintion etween inrush nd internl fult with etter ury is presented in [8]. This method lims to tke proessing time of 0.0se ( yle) with lssifition ury of 97.77%. Similrly, NN sed tehniques hve een used for the protetion of genertor too. One suh sheme with simple NN is presented in [9] for sttor winding protetion. Three prllel NNs hve een used in this sheme for lssifying three different fult ses. nother suh sheme is presented in [0] where two seprte NNs re used for fult detetion nd fult lssifition. n dvned version of this method using fuzzy logi in omintion with NN is presented in []. In oth ses, fult wveforms re simulted using diret phse quntities method. prtil protetion sheme is implemented in [] with NN developed on digitl signl proessor (DSP). lthough the importne of omined/unit protetion systems hs een identified in lte nineties, very few hve rried out reserh on unit protetion systems sine then. hyrid protetion sheme is presented in [3] for the strt This pper presents the use of rtifiil Neurl Networks (NN) s pttern lssifier for the omined differentil protetion of genertor-trnsformer unit with n im to uild kup protetion system to improve the overll reliility of the system. The proposed neurl network model is trined nd tested with n effiient Resilient k propgtion (RPROP) lgorithm nd Geneti lgorithm. The results re then ompred. The neurl network model mkes the disrimintion etween operting onditions (like norml, mgnetizing inrush, overexittion onditions in trnsformer) nd internl fults in trnsformer nd genertor sed on the differentil urrent wveform ptterns. The proposed method is independent of mplitudes of the wveforms. Vrious norml nd internl fult onditions of the trnsformer nd genertor re simulted using tooloxes in MTL/SIMULINK in order to otin the differentil urrent dt used for the trining nd testing of the NN. Index Terms rtifiil neurl networks, differentil protetion, geneti lgorithm, pttern reognition, resilient k propgtion, unit protetion I. INTRODUTION Trnsformer nd genertor re the most essentil elements of the power system with their protetion importne. Sine lst three dedes, reserhers hve een working on this prtiulr topi nd rose to mny new methods ut mostly onentrted on individul protetion system. There re vrieties of protetive relys to provide relile nd seure trnsformer protetion, of whih the differentil relys re found to e more effetive [] in fult disrimintion thn the old hrmoni restrint tehniques. The differentil relys should e designed in mnner tht it does not ml-operte during mgnetizing inrush nd over exittion onditions of trnsformer. The inrush urrents generted fter fult lerne re lso to e onsidered, s in [], while designing the rely. Most of the methods follow deterministi pproh, relying on fixed threshold. Mnusript reeived Otoer 0, 04; revised Ferury, Interntionl Journl of Eletronis nd Eletril Engineering doi: 0.70/ijeee

2 Interntionl Journl of Eletronis nd Eletril Engineering Vol. 3, No. 6, Deemer 05 out these inputs nd outputs is disussed in lter setions. protetion of genertor-trnsformer unit onsidering most of the fult types. This sheme is developed using three miroproessors sed on onventionl hrmoni restrint iruit method. This gve se ide for unit protetion systems. Lter, two NN sed tehniques were presented in [4], [5] in omintion with onventionl methods with fult detetion time of 0ms pproximtely. In oth ses, NN hd een trined with k propgtion lgorithm. ground fult unit protetion system is presented in [6] onsidering only the ground fults ourring in genertor. Mny of the proposed lgorithms produed good results in terms of ury. etter lgorithm n lwys improve the reliility of the protetion sheme. However, use of kup protetion system improves the reliility nd funtionlity of protetion devies. This pper presents model of deision system sed on NN onsidering the genertor-trnsformer unit s the proteted ojet. ll the internl fult onditions of trnsformer nd genertor hve een simulted to generte the required dtse for the trining of NN. lso, few ses of fults re generted using the method given in [0]. These ses re used only during testing of the networks. The developed NN hs een trined nd tested with RPROP nd Geneti lgorithm nd the results re ompred. During this proess, vrious rhitetures of NN hve een tested y vrying the numer of hidden neurons nd keeping the numer of input nd output neurons fixed. Detiled desription II. POWER SYSTEM SIMULTION FOR PTTERN GENERTION three-phse power system inluding 00MV, 3.8kV Genertor nd 00MV 3.8/3kV Δ-Yg Trnsformer long with 50 km trnsmission line hs een used to produe the required test nd trining ptterns. Fig. shows the sheme of the unit protetion system nd Fig. shows the power system model reted y mens of MTL Simulink softwre. Different types of fults re reted t different lotions. ll the genertor fults re ssumed to our t 00% of the sttor winding. lso, inrush urrent nd over exittion onditions re simulted t different voltge ngles nd with different lods. The generted wveforms re then smpled to feed the neurl networks to e tested with two different smpling rtes. Figure. Differentil protetion sheme of genertor - trnsformer unit Phse.0 Sutrt4 Tx Prim <Sttor voltge vd (pu)> Speed (pu) Vf (pu) Fult 3-Phse V-I Mesurement 3-Phse V-I Mesurement Fult T -i powergui Three-Phse Trnsformer (Two Windings) 00MV,50 HZ 3.8/3KV Fult3 Tx Phse Tx Phse Sutrt Tx Phse -i Disrete, Ts = 5e-06 s. Idiff_smpled V I T3 I Three-Phse 3-Phse V-I PI Setion Line Mesurement3 T4 Fult V T 3-Phse V-I Mesurement Tx Se Gx V & I 3 Phse Series olor RL Lod_50MW Lod HTG vref Synhronous Mhine vd Vf 00 MV 3.8 kv vq vst Exittion System Tx Prim V I V I T V-I Mesure m Vf_ 0 sp T5 Pm olor Phse &W V I <Sttor urrent> <Output tive power Peo (pu)> sp Phse Idiff4 <Rotor speed wm (pu)> wref Pref Pm we Pe0 gte dw Sutrt6 I (pu) <Rotor speed devition dw (pu)> 5 MW <Sttor voltge vq (pu)> Gin Tx_Idiff 7 i - 8 Sutrt i - -i 6 Sutrt3 9 i - Figure. Simulted three-phse power system model III. finl onfigurtion with gol of mximum ury. Keeping the numer of outputs fixed t, the numer of input neurons nd the numer of hidden neurons re vried on tril nd error sis until it produed minimum error. Two onfigurtions re finlized for testing fter NEURL NETWORK DESIGN ND SIMULTION The first step to formulte the prolem is identifition of proper input nd output set. Vrious rhitetures nd omintion of input sets were ttempted to rrive t the 05 Interntionl Journl of Eletronis nd Eletril Engineering 483

3 Interntionl Journl of Eletronis nd Eletril Engineering Vol. 3, No. 6, Deemer 05 mny trils, NN, 30-- neurons nd NN, For NN, eh of the differentil urrents (of eh phse) is typilly represented in disrete form s set of 0 uniformly sped (in time) smples otined over dt window of one yle i.e. t the smpling frequeny of 000Hz. For NN, set of 6 smples re otined over dt window of one yle i.e. t the smpling frequeny of 800Hz. These smples re used for trining nd testing the developed neurl networks. oth the proposed NNs generte outputs to represent 4 lssifitions s shown in Tle I. The si rhiteture of the NN is shown in Fig. 3. TLE I. Figure 3. NN rhiteture TRGET OUTPUT SES OF THE NN O O Output se 0 0 Norml 0 Trnsformer Inrush 0 Trnsformer Over Exittion Internl Fult in Trnsformer/Genertor IV. NN TRINING LGORITHMS. Resilient kpropgtion (RPROP) lgorithm Resilient kpropgtion is modifition of the ordinry grdient desent k-propgtion. To overome the inherent disdvntges of pure grdient-desent, Resilient kpropgtion (RPROP). This lgorithm ws pioneered y Mrtin Riedmiller [7]. The si priniple of RPROP is to eliminte the hrmful influene of the size of the prtil derivtive on the weight step. s onsequene, only sign of the derivtive is onsidered to indite the diretion of the weight updte ut not the mgnitude. The updte vlue for eh weight nd is is inresed y ftor Δ whenever the derivtive of the performne funtion with respet to tht weight hs the sme sign for two suessive itertions. The updte vlue is deresed y ftor Δ whenever the derivtive with respet to tht weight hnges sign from the previous itertion. If the derivtive is zero, then the updte vlue remins the sme. Whenever the weights re osillting, the weight hnge will e redued. In suh se, the updte vlue Δ ij is deresed y ftor η. If the derivtive retins its sign, the updte vlue is slightly inresed in order to elerte onvergene in shllow regions. This is shown in mthemtil form y () nd () [7]. The size of the weight hnge is exlusively determined y Et () ij ( t), if 0 wij Et () wij ( t) ij ( t), if 0 wij 0, else It should e noted, tht y repling the Δ ij y onstnt updte-vlue Δ, () yields the so-lled Mnhttn -updte rule. The seond step of RPROP lerning is to determine the new updte-vlues Δ ij (t). E( t ) E( t) ij ( t ), if * 0 wij wij E( t ) E( t) ij ( t) ij ( t ), if * 0 wij wij ij ( t ), else where 0<η <<η RPROP is generlly muh fster thn the stndrd steepest desent lgorithm s it onverges quikly nd it is sid to e the est trining lgorithm for pttern reognition & lssifition prolems [8].. G sed Trining of NN The geneti lgorithm (G) is well known optimiztion tehnique sed on the priniples of genetis nd nturl seletion nd doesn t require derivtive informtion for optimiztion. Unlike k propgtion lgorithm, it provides glol minim of optimiztion funtion. In the proposed method, G hs een used for finding weights nd ises of rtifiil Neurl Network. Then the next prt is to define fitness funtion whih n e used s n evlution funtion to optimize the weight set. The fitness funtion used here is men squre error (MSE), whih hs een otined y pplying ll trining sets (Input nd Trget) for eh weight set in the popultion. The lgorithm of fitness funtion used with G is given elow. { Let (I i, T i ), i=,, N, where I i =(I i, I i I li ) nd T i =(T i, T i, T ni ) represents the input-output pirs of the prolem to e solved y NN with onfigurtion l-m-n. For eh hromosome i =,, p elonging to the urrent popultion P i whose size is p { Extrt weights nd ises from i Keeping theses weights nd ises setting trin the NN for N input instnes; () () 05 Interntionl Journl of Eletronis nd Eletril Engineering 484

4 Interntionl Journl of Eletronis nd Eletril Engineering Vol. 3, No. 6, Deemer 05 lulte error E i for eh input instne using E i =T ji - O ji ; where Oi is the output vetor lulted y NN; Find root men squre MSE of the errors E i, i=,, N } Output fitness vlue F=MSE; }. Trining nd Testing of NN oth the NNs re trined seprtely with oth ove lgorithms. During RPROP sed trining, 0% sets of totl smples re used for vlidtion nd nother 0% re used for testing purpose. During G sed trining, the NN is trined y optimizing the weights nd ises of the network to minimize MSE. The totl numer of vriles is lulted s given elow. No. of vriles = input weights input ises lyer weights lyer ises =(I*H) H (H*O) O =(I O )*H O where I = No. of inputs; H = No. of Hidden neurons; O= No. of outputs. One the trining proess is ompleted the network is redy for testing. The network is then fed with new smples tht re not used for trining. For this purpose few test ses of genertor hve een developed using the diret phse quntities method given in [0]. For trnsformer fult ses, dtse is reted in MTL only. of fult, i.e., out 0ms for NN nd out 3ms for NN. This time is lulted sed on the numer of the smple t whih the NN produe vlue ove 0.98 t the output for trget vlue of fter the first smple of the fult wve is fed to it. lthough the results re not very good when the method is pplied s primry protetion system, the results n e onsidered stisftory when this system is used s kup protetion unit, whih generlly opertes fter some dely from the primry protetion unit. NN rhiteture TLE II. PERFORMNE ERRORS OF OTH NNS RPROP Trined NN est Performne Error G trined NN NN rhiteture TLE III. TRINING TIMES Time tken for trining (minutes) RPROP Trined NN G trined NN V. NETWORK PERFORMNE ND NUMERIL RESULTS The designed NN hs een trined nd tested with Resilient k Propgtion (RPROP) lgorithm nd Geneti lgorithm (G). The grphil representtions of the trining errors for oth rhitetures re given in Fig. 4-Fig. 7. Tle II shows the performne errors for ll ses. s one n find from these results, the RPROP lgorithm produed etter results thn G with the present network rhiteture. Further, NN with 30 inputs (hlf yle dt) give less error thn the NN with 48 inputs (full yle dt). However, further deresing the inputs didn t produe good results s the dt less thn hlf yle is insuffiient to reprodue the required wve shpe to tke the deision. Tle III gives the time tken for trining in eh se. prt from etter ury, RPROP took very less time for trining when ompred to G s it onverges quikly. The trining time lso depends on the proessor used in the P. Present methods re implemented on the ltest Intel ore i7 proessor sed system. To further inrese the trining speed of the G lgorithm, prllel proessing tehnology hs een used with the help of prllel proessing toolox ville in MTL. This llows G to use est speed of multi-ore tehnology of the proessor. The Intel i7 proessor hs 8 ores whih n e used in lusters or workers mode. It is worth mentioning tht oth lgorithms (RPROP nd G) tke lmost sme time to detet the ourrene Figure 4. Performnes of 30-- NN trined with RPROP Figure 5. Performnes of 30-- NN trined with G 05 Interntionl Journl of Eletronis nd Eletril Engineering 485

5 Interntionl Journl of Eletronis nd Eletril Engineering Vol. 3, No. 6, Deemer 05 Figure 6. Performnes of 48-- NN trined with RPROP Figure 7. Performnes of 48-- NN trined with G VI. ONLUSION In this pper, n rtifiil neurl network sed pttern reognition method hs een presented for the kup protetion of Genertor-Trnsformer unit. fter mny trils, two topologies of the network re finlized, one with hlf yle dt input nd the other with full yle dt input fed in moving window formt. oth topologies re trined seprtely with Resilient kpropgtion lgorithm (RPROP) nd Geneti lgorithm (G) for ll the possile ses of simulted dt under different operting onditions of trnsformer nd genertor. fter ompring the results, it is found tht the NN with hlf yle dt input is found more suitle thn the remining 3 omintions in terms of ury, trining speed, preision nd speed in fult detetion. The RPROP sed pttern reognition method is effiient in solving lssifition prolems nd differentil rely n e onsidered s lssifier whih identifies wht kind of event ours on the power system network. REFERENES [] M. Hi nd M. Mrin, omprtive nlysis of digitl relying lgorithms for the differentil protetion of three phse trnsformers, IEEE Trns. Power Syst., vol. 3, no. 3, pp , 988. []. Wiszniewski, W. Reiznt, nd L. Shiel, New lgorithms for power trnsformer inter-turn fult protetion, Eletr. Power Syst. Res., vol. 79, no. 0, pp , Ot [3] M. R. Zmn nd M.. Rhmn, Experimentl testing of the rtifiil neurl network sed protetion of power trnsformers, IEEE Trns. Power Deliv., vol. 3, no., pp , pr [4] Z. Morvej nd D. Vishwkrm, NN-sed hrmoni restrint differentil protetion of power trnsformer, Journl-Institution Eng. Indi Prt EL Eletr. Eng. Div., vol. 84, pp. -6, 003. [5] E.. Segtto nd D. V. oury, differentil rely for power trnsformers using intelligent tools, IEEE Trns. Power Syst., vol., no. 3, pp. 54-6, ug [6] M. Tripthy, Power trnsformer differentil protetion using neurl network prinipl omponent nlysis nd rdil sis funtion neurl network, Simul. Model. Prt. Theory, vol. 8, no. 5, pp , My 00. [7] M. Tripthy, R. P. Mheshwri, nd H. K. Verm, Power trnsformer differentil protetion sed on optiml proilisti neurl network, IEEE Trns. Power Deliv., vol. 5, no., pp. 0-, Jn. 00. [8] S. R. Smntry nd P. K. Dsh, Deision tree sed disrimintion etween inrush urrents nd internl fults in power trnsformer, Int. J. Eletr. Power Energy Syst., vol. 33, no. 4, pp , My 0. [9]. Tl, H.. Drwish, nd T.. Kwdy, NN-sed novel fult detetor for genertor windings protetion, IEEE Trns. Power Deliv., vol. 4, no. 3, pp , Jul [0]. I. Meghed nd O. P. Mlik, n rtifiil neurl network sed digitl differentil protetion sheme for synhronous genertor sttor winding protetion, IEEE Trns. Power Deliv., vol. 4, no., pp , 999. []. hlj, R. P. Mheshwri, S. Nem, nd H. K. Verm, Neuro- Fuzzy-sed sheme for sttor winding protetion of synhronous genertor, Eletr. Power omponents Syst., vol. 37, no. 5, pp , pr [] H.. H.. Drwish,.-M. I.. I. Tl, nd T.. T.. Kwdy, Development nd implementtion of n NN-sed fult dignosis sheme for genertor winding protetion, IEEE Trns. Power Deliv., vol. 6, no., pp. 08-4, pr. 00. [3] I. Korsiewiz, miroproessor sed protetive system for genertor-trnsformer units, in Pro. The Fourth Interntionl onferene on Developments in Power Protetion., 989, pp [4]. Hlink nd M. Szewzyk, NN sed detetion of eletril fults in genertor-trnsformer units, in Pro. The Eighth IEE Interntionl onferene on Developments in Power System Protetion, 004, vol. 004, pp [5] Y. Lu, L. Li, nd G. Tng, Neurl network sed genertortrnsformer protetion, in Pro. 004 Interntionl onferene on Mhine Lerning nd yernetis, 004, vol. 7, pp [6] M. Zielihowski nd T. Szlezk, new digitl ground-fult protetion system for genertor-trnsformer unit, Eletr. Power Syst. Res., vol. 77, no. 0, pp , ug [7] M. Riedmiller nd H. run, diret dptive method for fster kpropgtion lerning: The RPROP lgorithm, in Pro. Interntionl onferene on Neurl Networks, Sn Frniso, 993, pp [8] M. Shilee,. hndr, nd P. K. Klr, Lerning of geometri men neuron model using resilient propgtion lgorithm, Expert Syst. ppl., vol. 37, no., pp , De. 00. Hrish lg is urrently ssistnt Professor t GMRIT, Rjm. He hs sumitted his Ph.D. thesis t the Indin Institute of Tehnology (HU) Vrnsi, Indi. He otined his.e. (Eletronis nd Instrumenttion Engineering) from ndhr University nd M.Teh (ontrol Systems) from the Ntionl Institute of Tehnology Kurukshetr, Indi. His re of reserh interest is pplitions of Miroomputers nd rtifiil Intelligene to Power Systems. He hs ontriuted reserh ppers in IEEE nd other Interntionl onferenes. He is student memers of IET nd IEEE. 05 Interntionl Journl of Eletronis nd Eletril Engineering 486

6 Interntionl Journl of Eletronis nd Eletril Engineering Vol. 3, No. 6, Deemer 05 Hrshit Nth is presently finl yer student of E in Eletril Engineering t Thpr University, Ptil, Indi. He hs undergone prtil internship trining t HVD k to k Susttion of Power Grid orportion of Indi. His re of reserh interest is pplitions of Miroomputers nd rtifiil Intelligene to Power Systems. He hs worked on pplitions of rtifiil Intelligene to Power System Protetion t IIT (HU), Vrnsi during summer vtions. He hs ontriuted reserh ppers in IEEE nd other Interntionl onferenes. He is student memers of IET nd IEEE. Devendr Nth Vishwkrm is urrently Professor of Eletril Engineering t the Indin Institute of Tehnology (HU) Vrnsi, Indi. He otined his.s. (Engineering), M.S. (Engineering) nd Ph.D. from Ptn University, Ptn. He hd erlier served s ssoite professor of eletril engineering t the ihr ollege of Engineering, Ptn (Presently Ntionl Institute of Tehnology Ptn). He hs tehing nd reserh experiene of over 35 yers nd hs ontriuted out 65 ppers to vrious ntionl nd interntionl journls nd onferenes. He is outhor of the ook, Power System Protetion nd Swithger pulished y MGrw Hill edution (Indi) privte limited. He is senior memer of IEEE (US), fellow of the Institution of Engineers (Indi) nd fellow of the Institution of Eletronis nd Teleommunition Engineers. 05 Interntionl Journl of Eletronis nd Eletril Engineering 487