面向电力数据异常检测的散度评估增强的联邦学习模型

打开文本图片集

中图分类号：TM933文献标志码：A

Abstract：To addressthe problems of inconsistent model convergence and slowconvergence speed caused bydierences indatadistributionacross local terminalsofmultipleparticipantsin federatedlearning，afederatedlearningmodel enhanced by divergenceevaluation on anomaly detection for power data was proposed.Based on the local power data storedbydiferentparticipants，asymmetric Jensen-Shannon divergence matrix was introduced torepresentthenon-independentand identicalldistributednatureof thedata.Softmax function was employed tocalculate divergence similarity scores，evaluating thecontribution of each client’s gradient during the agregation process，aiming at promoting the convergence andaccuracyof the proposed modelwhen meeting non-independentand identicalydistributed data.Anomaly detection inpower data wasultimatelyachieved basedon the wel-trained model.Thesimulation experiment foranomaly detectionin power data resultsshow that the proposed model demonstrates superioritycompared to baseline models in simulatioexperiments，whichefectivelysolvesthenon-independentandidenticalldistributedproblemoffederated learning，and verifiesthe practicalityoffederated learning technology inthe fieldof powerdataanomaly detection.

Keywords：electric powerdata analysis；anomalydetection;federatedlearning;divergence estimation；gradientaggregation

由于电力系统具有结构复杂与数据多样的特点，因此电力数据常常存在各种异常情况[1-2]，例如电压异常、电流异常、负荷异常等。（剩余12444字）