基于多异学习器融合Stacking集成学习的窃电检测  

Electricity Theft Detection Based on Multiple Different Learner Fusion by Stacking Ensemble Learning

作　　者：游文霞[1] 李清清 杨楠[1] 申坤 李文武[1] 吴泽黎 YOU Wenxia;LI Qingqing;YANG Nan;SHEN Kun;LI Wenwu;WU Zeli(School of Electrical and New Energy,China Three Gorges University,Yichang 443002,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

出　　处：《电力系统自动化》2022年第24期178-186,共9页Automation of Electric Power Systems

基　　金：国家自然科学基金资助项目(51607104)。

摘　　要：针对窃电检测中用户用电数据类别不平衡、采用投票法作为结合策略的集成学习方法无法充分发挥多个不同学习器优势等问题,提出一种利用Stacking集成学习融合多异学习器的模型应用于窃电检测。首先,从影响电量计量的因素出发,根据常见的5种窃电方法模拟6种窃电行为模式;其次,采用合成少数类过采样技术(SMOTE)对不平衡的用电数据进行处理,并利用K折交叉验证法对平衡后的训练集进行划分以缓解因重复学习造成的过拟合;然后,使用评价指标和多样性度量优选模型的不同初级学习器和元学习器,构建融合不同学习器优势和差异的Stacking集成学习窃电检测模型;最后,算例对比分析结果表明所提窃电检测模型能有效解决用电数据类别不平衡,充分发挥不同学习器的优势,评价指标良好。Aiming at the problems that the consumer power consumption data categories are unbalanced for electricity theft detection, and the ensemble learning method using voting as a combination strategy cannot give full play to the advantages of multiple different learners, a model using Stacking ensemble learning to fuse multiple different learners is proposed and applied to electricity theft detection. First, starting from the factors affecting electricity metering, six electricity theft behavior modes are simulated according to five common electricity theft methods. Secondly, synthetic minority oversampling technique(SMOTE) is used to process the unbalanced power consumption data, and K-fold cross-validation method is used to divide the balanced training sets to alleviate the overfitting caused by repeated learning. Then, the evaluation indicators and diversity metrics are employed to optimize different primary learners and meta-learners of the model, and a Stacking ensemble learning electricity theft detection model integrating the advantages and differences of different learners is constructed. Finally, the comparative analysis results of examples show that the proposed electricity theft detection model can effectively solve the imbalance of power consumption data categories, give full play to the advantages of different learners, and the evaluation index is good.

关 键 词：Stacking结合策略 集成学习 窃电检测 合成少数类过采样技术 K折交叉验证 

分 类 号：TM73[电气工程—电力系统及自动化] TP181[自动化与计算机技术—控制理论与控制工程]