基于Stacking集成学习的窃电检测研究  

Research on electricity theft detection based on Stacking ensemble learning

作　　者：冯小峰 姚诚智 杨俊华[2] FENG Xiaofeng;YAO Chengzhi;YANG Junhua(Metrology Center,Guangdong Power Grid Corporation,Guangzhou 510062,China;School of Automation,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东电网有限责任公司计量中心,广州510062 [2]广东工业大学自动化学院,广州510006

出　　处：《电测与仪表》2024年第11期211-218,共8页Electrical Measurement & Instrumentation

基　　金：中国南方电网有限责任公司科技项目(GDKJX M20185800)。

摘　　要：基于单一模型的传统窃电检测精度有待提高,应用Stacking集成学习策略,提出一种新的窃电模型,并根据实际业务需要构造MAP(mean average precision)检测指标。为实现降维效果,按时间多维度拆解用户日用电量特征,并采用Embedding特征选择,选取其中的重要度高特征;采用贝叶斯优化调参,结合XGBoost、LightGBM和CatBoost集成模型对数据进行交叉验证和预测;分别拼接验证结果和预测结果,Stacking的基分类器采用逻辑斯蒂回归进行集成训练,输出最终预测结果。以2016 CCF竞赛数据为算例,验证了所提出的窃电模型的有效性和可行性。Aiming at the shortcomings of traditional electricity theft detection method that only uses single classifier,a novel electricity theft detection model based on Stacking ensemble learning strategy is proposed,and a MAP detection indicator is built based on actual project.The daily electricity consumption characteristics of customers are disassembled in multiple dimensions according to time.And then,Embedding is employed in selecting the most important features,which reduces the feature dimensions.XGBoost,LightGBM,and CatBoost models are used to cross-validate and predict the data results by Bayesian optimization parameter tuning.The validation results and prediction results are spliced respectively,the base classifier of Stacking,whose integrated training with logistic regression,outputs the final prediction results.Taking CCF competition data in 2016 as an example,the example analysis results verify the effectiveness and feasibility of the proposed electricity theft model.

关 键 词：窃电检测 Stacking集成学习 特征选择 MAP检测指标 贝叶斯优化 

分 类 号：TM933[电气工程—电力电子与电力传动]