基于重加权策略平衡损失与LSTM的窃电行为检测研究  

Electricity Stealing Detection Based on Re-weighted Strategy Balancing Loss and LSTM

作　　者：吕笃良 刘梦爽 桓露 孙羽森 刘通宇 袁培森[3] LYU Duliang;LIU Mengshuang;HUAN Lu;SUN Yusen;LIU Tongyu;YUAN Peisen(State Grid Xinjiang Electric Power Co.,Ltd.Marketing Service Center,Urumqi 830000,China;NARI Technology Co.,Ltd.,Nanjing 210000,China;College of Artificial Intelligence,Nanjing Agricultural University,Nanjing 210095,China)

机构地区：[1]国网新疆电力营销服务中心,新疆乌鲁木齐830000 [2]国电南瑞科技股份有限公司,江苏南京210000 [3]南京农业大学人工智能学院,江苏南京210095

出　　处：《智慧电力》2022年第4期15-20,58,共7页Smart Power

基　　金：国家自然科学基金资助项目(61806097;62072247);江苏省农业科技自主创新资金项目(SCX(21)3059);上海市大数据管理系统工程研究中心开放基金(HYSY21022)。

摘　　要：窃电行为是导致电能损失与电力企业经济效益降低的重要原因。针对窃电问题,提出了一种基于有效数量加权策略的损失函数,改善数据集分布不均衡导致训练模型泛化性能下降的问题;基于该策略,设计了基于长短期神经网络的时间序列分类模型,用于用户日用电量的窃电行为检测任务;采用用户日用电量真实数据进行实验测试,结果表明基于有效数量的加权策略可一定程度解决数据集不平衡导致的模型泛化性下降问题。与现有方法相比,所提方法在精确度上有所提高,对窃电行为检测具有有效性与可行性。Electricity theft is an important cause of power loss and reduced economic efficiency of power enterprises.In the light of electricity theft,a loss function based on the effective quantity weighting strategy is proposed to improve the unbalanced dataset distribution,which is the main cause leading to the degradation of generalization performance of the training model.The time series classification model based on long and short term neural networks is designed for the task of detecting electricity theft behavior of daily electricity consumption of users.Experimental tests are conducted using real data of daily electricity consumption,and the results show that the effective number-based weighting strategy can solve the problem of model generalization degradation caused by the imbalance of data sets to a certain extent.Compared with existing methods,the proposed method has improved accuracy and is effective and feasible for electricity theft detection.

关 键 词：窃电检测 重加权策略 类别平衡损失 LSTM 时间序列分类 

分 类 号：TM769[电气工程—电力系统及自动化]