基于重采样和混合集成学习的不平衡窃电检测  

Class Imbalanced Electricity Theft Detection Based on Resampling and Hybrid Ensemble Learning

作　　者：游文霞[1] 梁皓 杨楠[1] 李清清 吴永华 李文武[1] YOU Wenxia;LIANG Hao;YANG Nan;LI Qingqing;WU Yonghua;LI Wenwu(College of Electrical Engineering&New Energy(China Three Gorges University),Yichang 443002,Hubei Province,China;Xianning Power Supply Company,Hubei Electric Power Co.,Ltd.,Xianning 437000,Hubei Province,China;Xiaogan Power Supply Company,Hubei Electric Power Co.,Ltd.,Xiaogan 432000,Hubei Province,China)

机构地区：[1]三峡大学电气与新能源学院,湖北省宜昌市443002 [2]国网湖北省电力有限公司咸宁供电公司,湖北省咸宁市437000 [3]国网湖北省电力有限公司孝感供电公司,湖北省孝感市432000

出　　处：《电网技术》2024年第2期730-739,共10页Power System Technology

基　　金：国家自然科学基金项目(51607104)。

摘　　要：针对电力用户类别不平衡导致窃电检测具有偏向性问题,该文提出一种基于重采样和混合集成学习的不平衡窃电检测模型。首先以Easy-ensemble混合集成学习框架为基础确定最佳采样子集数;然后通过重采样自适应策略,即根据用户用电数据集的不平衡度以及最佳采样子集数确定检测模型的重采样方式,使用电数据达到平衡;最后按照先串行集成减小偏差、后并行集成降低方差的混合集成方式,对重采样后的均衡样本进行窃电检测。算例对比分析表明所提检测模型通过重采样和混合集成有效解决了传统集成算法在不平衡窃电检测中的偏向问题,降低了由于用电数据的不平衡性对集成结果的影响,提高了用户类别不平衡的窃电检测效果,在多种不平衡度下模型的准确率、F1值和G均值均表现优异。Aiming at the bias problem of electricity theft detection caused by the imbalance of power user classes,a class imbalanced theft detection model based on resampling and hybrid ensemble learning is proposed.Firstly,the optimal number of sampling subsets is determined based on the Easy-ensemble hybrid ensemble learning framework.Then,through the resampling adaptive strategy,that is,according to the imbalance of the user's electricity data set and the optimal number of sampling subsets,the resampling method of the detection model is determined,achieving the balance of the electrical data.Finally,according to the hybrid ensemble mode,i.e.,first serially ensembling the data to reduce deviation and then parallelly ensembling them to reduce variance,the resampled balanced sample is detected for power theft.The comparative analysis of the study example shows that the proposed detection model effectively solves the bias problem of the traditional ensemble algorithm in the detection of unbalanced electricity theft through resampling and hybrid ensembling.The influence of the imbalance of the electricity consumption data on the ensemble results is reduced,and the imbalanced electricity theft detection effect of the user category is improved,which shows that the proposed model performs very well in accuracy,F1 value and G mean under various imbalances.

关 键 词：窃电检测 不平衡数据 重采样 集成学习 Easy-Ensemble集成框架 

分 类 号：TM721[电气工程—电力系统及自动化]