电气传动2023年第53卷第3期ELECTRICDRIVE2023Vol.53No.3摘要：准确的电力异常用户识别方法能为供电企业锁定存在窃电行为或其他违规行为的电力用户提供参考。大多数基于机器学习的异常识别模型采用了无监督算法，但模型的准确度还较低。针对上述问题，提出一种结合无监督的局部离群因子（LOF）算法与有监督的支持向量机（SVM）算法的两阶段异常用电用户识别方法。基于分析异常电能表区别于正常电能表的电流电压表现，构建异常识别模型的输入特征；采用无监督的LOF算法进行采样，筛选出可疑样本交给人工进行标记，然后利用标记样本训练有监督的SVM模型；在之后的检测工作中，直接将LOF算法筛选出可疑样本交给SVM模型进行识别。实例结果表明，该方法对电力异常用户的识别准确度高，对供电企业的窃电稽查工作具有指导意义。关键词：电力异常用户识别；机器学习；局部离群因子（LOF）；支持向量机（SVM）中图分类号：TM28文献标识码：ADOI：10.19457/j.1001-2095.dqcd23988PhasedIdentificationMethodofAbnormalElectricityUsersBasedonLOF+SVMGUZhen1，ZHUANGGewei1，HEQing1，ZHOULei1，ANBailong1，DUANYan2（1.PowerScienceResearchInstitute，StateGridShanghaiElectricPowerCompany，Shanghai200051，China；2.DepartmentofVehicleEngineering，SchoolofAutomobile，TongjiUniversity，Shanghai201804，China）Abstract:Accurateidentificationmethodofabnormalelectricityuserscanprovidereferenceforpowersupplyenterprisestolockinelectricitytheftorotherviolationsofpowerusers.Mostabnormaluseridentificationmodelsbasedonmachinelearningadoptunsupervisedalgorithms，buttheaccuracyofthemodelsislow.Tosolvetheaboveproblems，atwo-stageabnormalpoweruseridentificationmethodcombiningunsupervisedlocaloutlierfactor（LOF）algorithmandsupervisedsupportvectormachine（SVM）algorithmwasproposed.Basedontheanalysisofthecurrentandvoltageperformanceoftheabnormalenergymeterdifferentfromthenormalenergymeter，theinputcharacteristicsoftheabnormalidentificationmodelwereconstructed.TheLOFalgorithmwasusedtosample，andthesuspicioussampleswereselectedandhandedovertomanuallabeling.ThenthesupervisedSVMmodelwastrainedbythelabeledsamples.Inthesubsequentdetectionwork，thesuspicioussamplesscreenedbyLOFalgorith...
