ANN‐based grid voltage and frequency forecaster Alessandro Massi Pavan.pdf

该【ANN‐based grid voltage and frequency forecaster Alessandro Massi Pavan 】是由【周瑞】上传分享，文档一共【5】页，该文档可以免费在线阅读，需要了解更多关于【ANN‐based grid voltage and frequency forecaster Alessandro Massi Pavan 】的内容，可以使用淘豆网的站内搜索功能，选择自己适合的文档，以下文字是截取该文章内的部分文字，如需要获得完整电子版，请下载此文档到您的设备，方便您编辑和打印。TheJournalofEngineeringThe9thInternationalConferenceonPowerElectronics,MachinesandDrives(PEMD2018)eISSN2051-3305ANN-basedgridvoltageandfrequencyReceivedon25thJune2018Acceptedon30thJuly2018forecasterE-Firston14thMay2019doi:.8162AlessandroMassiPavan1*,NadjwaChettibi2,AdelMellit2,3,ThomasFeehally1,,aTodd11SchoolofElectricalandElectronicEngineering,TheUniversityofManchester,Manchester,UK2RenewableEnergyLaboratory,JijelUniversity,Jijel,Algeria3TheInternationalCentreforTheoreticalPhysics(ICTP),Trieste,Italy*CurrentAddress:TheDepartmentofEngineeringandArchitecture,UniversityofTrieste,Trieste,ItalyE-mail:******@:ThispaperpresentsamethodfortheforecastingofthevoltageandthefrequencyatthepointofconnectionbetweenabatteryenergystoragesysteminstalledatTheUniversityofManchesterandthelocallow--(ANN)-basedtechniqueandcanpredictthegridquantitieswithtwodifferenttimewindows:-basedreal-timecontrollerandallforecastersshowverygoodperformance,withcorrelationscoefficients>,andmeanabsolutepercentageerrorsof<%.1?Introductionbeappliedtoanydevicethatcanabsorborinjectreactiveand/(DGs)inthe(V)of400?Vandoncethemeasuredvoltageisoutsideoftheformofrenewablessources,suchaswindandsolar,isanticipatednomdeadband,,andin2035,49%oftheUK'selectricitygenerationmeasuredvoltageisnormallyclosetothenominalvoltage,thenthecapabilityisanticipatedtobefromrenewables[1].,however,largeloadvariationsworkwork,thenthelocalvoltagewillexhibitsustaineddemand,andthiscanleadtochallengesinmanagingthelocalorrapidchangesawayfromthenominalvalueandthedevicemaynetworkvoltage[2]%[3],whichcanoftenpresenthighloadsonasingle-?Hzisthenominalfrequencyofthegridvoltageinathree-phaselow--voltageThispaperpresentsbothlocalvoltageandfrequencyforecastnetworksbyproviding/absorbingreactivepower[2].Accordingtotechniques,whichhavebeendevisedforuseinlow-voltagetheNationalGridcodes[4,5],thesemethodsallhavelocalcontroldistributiongrids,,workwhichforexamplecouldbeaDGoranelectricalstoragesystem,(ANN)-,-,-reactiveparallel-worksandcanlearnnewpowerandfrequency-activepowerprofiles,,newfunctionaldependencies,andnewpatterns[6].In[7],anoff-lineANNoneminuteaheadgrid-voltageforecasterwaswork,andthemaximumerroroftheirforecasterwas3%.,thepossibilitytoenableboththenominalvoltage(and/ornominalfrequency)andthegradientoftheabsorb/()tobeupdatedinrealtimebasedontheeedofthedistributionsystemoperator(DSO)(SoC)optimisationofaBESSorofagroupofBESSoperatingasavirtualpowerplant(VPP).Theforecastedgridfrequencyactsasaninputofanenergymanagementsystem(EMS)workmanagementstrategy,whichisresponsibleforensuringthesecurityofsupplyfortheDSOslow-[8]wherethevariablesareforecastedwith‘long’-,?Typicalvoltage–reactivepowerprofilehorizonismuchshorter,.,2019,,-36913687monsAttributionLicense(/licenses/by//)?Typicalfrequency–activepowerprofilevariablebeingforecasted(soeithervoltageorfrequencyhere),,onlyfrequencyisbeingforecast,andthedatausedtotrainthefrequencyforecasterareloggedeverysecondinrealtimebythedSPACEsystem,,oneforthefrequencyandoneforthevoltage,?Frequencyandvoltageofthegridloggedevery500?msduringtheloggedevery500?msinthedSPACEsystembetween2:–,thedatahavebeendown-sampledobtainingwheretimedelayscaninfluencethestabilityofthesystemsandn?=?8,4741-[9].Inthiscase,statisticaletimedelaysdue,,municationlatencies[10].anisedasfollows:Section2describestheTheRNNusedfortheonlineforecastingofthegridfrequencyisdesignoftheANN-basedforecasters,work[12]anditsgeneralstructureisshowninobtainedinthestudy,:(IL)correspondstothefrequencyf(k)attimek,therearefiveneuronsintheContextLayer(CL),fiveneuronsintheHidden2?ImplementationoftheforecastersLayer(HL),andoneneuronintheOutputLayer(OL)whichcorrespondstotheforecastedgridfrequency(attimek?+?1).Theworkhavebeenadaptedonlineduringthelow-voltagegridconnection,usingthedatatodevelopthevoltagelearningprocessusingtheadaptiveinteraction(AI)ruleasandfrequencyforecasters,anddemonstratingthevoltageanddescribedin[12,13].(k)attimeinstantkisscaledbythedevelopthevoltageandfrequencyforecastmethodsarerecordednominalvalueofgridfrequency,fgrid_nom,toformtheinputnodefromthelow-voltage(400?V)(BESS)installedatTheUniversityofManchesterthattheoutputofthenodeis[12]:[11].ThedSPACEreal-timecontroller,whichprovidesaninterfacetotheBESS,canrecordthevoltageandthefrequencyofthegridat(1)f(k)z(k)=(1)afixedsampletimeupto1?:onetypeistrainedonlineduringthereal-timeoperationofTheoutputsofthehiddennodesare:work(2)(2)(3)c(1)(RNN),andthismethodisusedtoforecastthefrequencyofthezi(k)=SjΣrzr(k)+w1jz(k)(2)gridonesecondahead;thesecondoneistrainedusingofflinehistoricaldatasetsandisafeed-work(FFNN),whererandjvaryintherange[1–5].whichcanforecastboththefrequencyandthevoltageoneminute(2):canefficientlyapproximateanynon-linearfunctionandarewidelyusedinthefieldsofcontrolsystemsandforecasting.(2)(2)1Sj(nj)=(2)(3)1+e?(2)-basedforecastersareofthenjfollowingtype:Z?=?(z,z,…,z),,workweightsarecalculatedas[12]:.,2019,,-3691monsAttributionLicense(/licenses/by//)?Structureoftheproposedoff-?Structureoftheproposedon-rainedRNN-basedgridfrequencyforecasters(oneminuteahead)forecaster(onesecondahead)Table1?Featuresofthetwooff-rainedFFNNsThe‘tansig’‘weights_2’‘weights_2’blockperformsthedotproducttrainingfunction‘trainscg’‘trainscg’oftheoutputsoftheHLoutputswithaone-dimensionalarraynumberofhiddennodes1212containing12gainvalues,andthesingleoutputissummedwithaHLtransferfunctionhyperbolictangenthyperbolictangentsinglebiasvalue,,whichinthiscaseisthe‘tansig’orthe‘purelin’function,toformtheforecastedvalueZ(t?+?1).OLactivationfunctiontangentsigmoidlinearThetwoFFNN-basedforecastersaretrainedofflineusingadatadivisionfunction‘dividerand'‘dividerand'scaledconjugategradientalgorithm(named‘trainscg’)[14]runinnumberofiterationsduring50005000aMatlab/(2)(1)workToolbox?.ThesettingsinTable1w1j(k)=w1j(k?1)?(1?zj)zwj(k?1)Δwj(k?1)weredeterminedwiththegoalofachievingameansquareerrorwkwkμek(4)0()=0(?1)?f(?1)(MSE)of10?×?10?9.(2)wj(k)=wj(k?1)?μzjef(k?1)3?Analysisoftheresultswherewcistheweightbetweentheinputneuronandthej-thparisonoftherealquantitieshiddennode,Δw(k???1)?=?w(k???1)???w(k???2),eisthefrequencyjjjf(voltagesandfrequencies)andthoseforecastedwiththeuseoftheerror,,so:-secondahead,onlineANNforecaster;theANNwas(3)(2)z(k)=zj(k?1)(5),,which(3)?s,andthebottomplotshowsamagnifiedviewTheoutputoftheOLisalineartransferfunction(TF)describedofthetracesbetween100and200?:between08::(2)initialisedbyrunningthesamestructureoftheRNNinanofflinef(k+1)=Σiwjzj(k)?w0(6),respectively,fortheone-minuteahead,offlineANNoutputnode,-forwardneuralsystemandhavebeenpreviouslytrainedwiththedatapresentedinnetwork(FFNN)(correspondingto4040?min)(IL)(t)andthefrequencyf(t)atthebetween2::,respectively,andtheOLgiveasaresultsthevoltageV(t?+?1)andthefrequencyf(t?+?1)-minuteahead;,bothfortheonlineandfortheoffline-Inthisimplementation,Z(t),‘weights_1’blockTheperformanceofthethreeforecasterscanalsobedescribedperformsthedotproductof12weightgainvalueswiththeinputquantitatively,(t)toform12outputs,andtheseoutputsareeachsummedwithacorrelationcoefficientr,thecoefficientofdeterminationR2,theseparatebiasvalue,containedintheone-(RMSE),.,2019,,-36913689monsAttributionLicense(/licenses/by//)?Measuredandestimatedfrequencyfortheon-?Measuredandestimatedvoltagefortheoff-lineoneminuteaheadFFNNforecasterTable2?–8ForecasterrR2RMSEMAPE,%on-?-?×?10?4?-?