基于联邦学习的用电异常检测系统设计与实现.pdfVIP

关键词：联邦学习；数据安全；卷积神经网络；异常检测；

Abstract

Theabnormalpowerconsumptionbehaviorofpowerusershasseriouslydisrupted

thenormalorderofpowerconsumptionandendangeredthesafeoperationofthepower

grid.Nowadays,artificialintelligencehasemergedinthefieldofelectricityanomaly

detection,graduallyreplacingtraditionalmanualinspectionmethodswiththerapid

developmentofmachinelearning.However,duetotheslowpopularizationof

intelligentpowerequipment,somepowercompanieshavepoorhistoricaldatadueto

variousreasonssuchasbackwardequipmentconditions,falsealarmsandomissionsof

terminalalarms,andjudgementerrorsbyinspectors,resultinginpoorperformanceof

thetrainedmodels.Inordertocombinethedataofmultiplepowercompaniesunder

thepremiseofprotectingdatasecurity,andsolvetheproblemsofpoordataqualityand

insufficienteffectivedataofpowercompanies,thisprojecthasdesignedand

implementedasetofelectricityabnormaldetectionsystembasedonfederatedlearning

toassistpowercompanymanagesthestandardizedelectricityconsumptionofusers

andimprovestheefficiencyofelectricityinspection.

Differentfromthetraditionalmethodoftrainingmodelsincentralizeddata,thejoint

electricityanomalydetectionsystemdesignedandimplementedinthisprojectadopts

adistributedmodeltrainingmethod,usinglocaldatadistributedinvariouspower

companiestotraintheelectricityanomalydetectionmodel.Aggregatemodel

parametersuploadedbypowercompanyclientsparticipatinginfederatedlearningto

updateandsharetheglobalmodel.

Atthesametime,inordertosolvetheproblemthattheglobalmodeloffederated

learningiseasilyaffectedbytheclientmodelwithpoordataqual