A-FIRANN-as-a-differential-relay-for-three-phase-p

IEEETRANSACTIONSONPOWERDELIVERY,VOL.16,NO.2,APRIL2001215AFIRANNasaDifferentialRelayforThreePhasePowerTransformerProtectionÁngelL.Orille-Fernández,Member,IEEE,NabilKhalilI.Ghonaim,andJaimeA.ValenciaAbstract—ThispaperpresentsanapplicationofaFiniteIm-pulseResponseArtificialNeuralNetwork(FIRANN)asadiffer-entialprotectionforarealthreephasepowertransformer.ThreeFIRANNsaredesigned,trainedandtested.Thefirstonehasanoutput,whichidentifyinternalfaultsfromanyothercaseslikein-rushcurrentandexternalfaults.ThetwoothersFIRANNs,eachhavetwooutputsthatclassifybetweeninternalandexternalfaults,sothat,abackupprotectionisincluded.TheseFIRANNshavesixinputs,oneforeachsampledcurrentsignalfrombothtransformersides.Thesamplerateselectedis2kHzfora50Hzpowerfre-quency.AllFIRANNsweretrainedtohavea3.5msfaultdetectiontime,whichisconsideredasaveryfastprotection.ThetestresultsshowverygoodbehavioroftheFIRANNasadifferentialprotec-tionanditisplannedtobuildaprototype.IndexTerms—Artificialneuralnetwork,digitalrelays,trans-formerprotection.I.INTRODUCTIONTHEFIRSTuseofartificialneuralnetworks(ANN)appliedonpowertransformerdifferentialprotectionwasin1994[1].Sincethattimetherearereallyfewpaperssuggestedtheap-plicationofANNonpowertransformerprotectionandthereisalotofworktodointhissubject.TheusesofANNontrans-missionlineprotectionhavebeengivenmoreconsideration.TherearedifferentwayshowANNcanbeappliedtoadiffer-entialtransformerprotection.ThefirstinvestigationssuggestedtheuseofANNasaninrushcurrentidentifiertobeincludedasapartofadifferentialprotection.In[1],itwereproposedtouseaTimeDelayArtificialNeuralNetwork(TDANN)toprocessthenormalizedsampledcurrentsignals.Anothergroup[2],suggestedtheuseoftheDFT(discreteFouriertransform)tofilterthefundamentalcomponentandtheharmonicsfromthesecondorderthroughfifthorderofthecurrentsignalsandapplythemtoaMultilayerFeedForwardArtificialNeuralNetwork(MFANN).ArecentpaperkeptthesameideaandsuggestedanadditionalMFANNtoreconstructthedistortedsecondarycurrentsignalofthecurrenttransformerduetosaturationinordertoimproveoperationofthepowertransformerprotection[3].LaterpapersproposedMFANNasatransformerdifferentialprotection,onesuggestedtouseharmonicratiosasinputs[4];whiletheotherusedthenegativesequencecomponentofcurrentsignalsandthevoltagesignalsasinputsoftheneuralnetwork[5].ManuscriptreceivedFebruary7,2000.Á.L.Orille-FernándezandN.K.I.GhonaimarewiththeElectricalEngi-neeringDepartment,PolytechnicUniversityofCatalonia,Barcelona,Spain.J.A.ValenciaiswiththeElectricalEngineeringDepartment,UniversityofAntioquia,Colombia.PublisherItemIdentifierS0885-8977(01)03412-4.Fig.1.Three-phasetransformercircuitscheme.ThebehaviorofFIRANN(FiniteImpulseResponseArtificialNeuralNetwork)appliedtoatransformerdifferentialprotectionispresentedinthispaper.Thisneuralnetworkiswellknownforitsabilitytomanagetimevariablesignals[6].ThreedifferentFIRANNsweretrainedasdifferentialprotectionforathree-phasepowertransformer.Thefirstonehasanoutputtoclassifytheinternalfaultsfromanyothercase.TheothertwoFIRANNhavetwooutputstoclassifybetweeninternalandexternalfaults.This,infact,isanovelapproachtodifferentialprotectionthatimprovesselectivityoftheproposedrelayandincludesabackupprotectionfortheotherelements,whichareconnecteddirectlytothepowertransformer.TheinputsoftheFIRANNsarethenormalizedsampledcur-rentsignalsfrombothsidesofthetransformer.ItmeansthattheFIRANNsactsasasignalprocessorbasedrelay.TheFIRANNsstructure,trainingmethod,trainingstrategyandtestresultswillbereported.II.SIMULATEDSYSTEMThesystemusedinthisworktogeneratethetrainingpatternsoftheFIRANNwasa15kVA,220V/1300VYysolidlygroundedpowertransformer.Ithassixtapsoneveryphasewindingofthehighvoltageside.Theshortestsegmentbetweentwoconsecutivetapsis6%andthelongestis26%ofthetotalcoilturns.Themethodusedtosimulateinternalfaultsisexplainedin[7].TheAlternativeTransientProgram(ATP)wasusedtosimulateallcasesneededinthiswork.Thesimulationmodelwasvalidatedusingreallaboratorymeasurements.TheschemeofthesystemisshowninFig.1.Thetransformerismodeledas–coupledbranchestosimulateitscoilswind-ings.Ironlossesaresimulatedasthreeresistivebranch,Threenonlinearinductionbranches,,wereaddedtocon-siderthesaturationeffectforinrushsimulationcases.Thesam-pledcurrentsignalsaretakenfrommeasurementswitchtype,sotherewasnotincludedanycurrenttransformermodel.The0885–8977/01$10.00©2001IEEE216IEEETRANSACTIONSONPOWERDELIVERY,VOL.16,NO.2,APRIL2001TABLEITHREE-PHASETRANSFORMERSIMULATIONScurrentsensorsusedonlaboratorymeasurementwereHallef-fectelements,whichhavelinealcharacteristic.Thesamplingfrequencyusedis2kHz.III.SIMULATEDFAULTCASESOneofthemostimportantstagesduringthedesigningofANNistheselectionofthesimulatedcasesthatmustbein-cludedinthetrainingandtestingsets.Thetransformermodelwassimulatedtocoverallpossibleoperatingandfaultcondi-tionssuchasthesourceshortcircuitlevel,faultinceptionin-stants,faulttypesandinrushcurrent.Thesimulatedcasesaredividedintothreegroups.ThefirstisthetraininggroupanditspatternsareselectedrandomlyandnormallydistributedinordertomaketheFIRANNtogeneralizeandto