基于CNN-SVM模型的电网窃电风险自动化预警算法  

Automatic Early Warning Algorithm of Power Theft Risk Based on CNN-SVM Model

作　　者：张春梅 袁杰生 许兴雀 黄彬系 王曦 ZHANG Chun-mei;YUAN Jie-sheng;XU Xing-que;HUANG Bin-xi;WANG Xi(Zhongshan Power Supply Bureau of Guangdong Power Grid Co.,Ltd.,Zhongshan 528400 China;Guangdong Electric Power Information Technology Co.,Ltd.,Guangzhou 510030 China)

机构地区：[1]广东电网有限责任公司中山供电局,广东中山528400 [2]广东电力信息科技有限公司,广东广州510030

出　　处：《自动化技术与应用》2025年第4期38-42,88,共6页Techniques of Automation and Applications

基　　金：中山供电局2021年综合域个性化数据应用建设(全业务合规管理监控应用)项目(032000HK42210004)。

摘　　要：针对电网窃电风险预警时效性差、精度低的问题,设计基于CNN-SVM模型的电网窃电风险自动化预警算法。分别从日电压均值、功率因数、电流信息等方面采集用户用电数据,利用最小-最大标准化方法预处理数据;构建窃电风险预警参数集合,确定暂态因子权值,将电压扰动、负载波峰作为特征指标,提取窃电风险边界特征;建立SVM分类器模型,获得最优分类平面和决策函数;通过卷积神经网络改进SVM分类器,构建CNN-SVM模型,提高模型的学习能力;经过模型逐层学习,输出最终特征向量,即窃电风险等级,结合该等级实现自动化预警。实验结果表明,CNN-SVM模型的收敛效果好,训练耗时短,能够准确识别出窃电用户,应用于电网窃电风险自动化预警中的预警精度高,响应速度快。Aiming at the problems of poor timeliness and low accuracy of power theft risk early warning,an automatic power theft risk early warning algorithm based on CNN SVM model is designed.It collects user power consumption data from aspects such as daily voltage average,power factor,and current information,and preprocesses the data using the minimum maximum standardization method,establishes a set of early warning parameters for power theft risk,determines the weight of transient factors,and extracts the boundary characteristics of power theft risk using voltage disturbances and load peaks as characteristic indicators,establishes a SVM classifier model to obtain the optimal classification plane and decision function.Using Convolutional Neural Networks to improve SVM classifiers,construct CNN-SVM models,and improve the learning ability of the models.After learning the model layer by layer,the final feature vector,namely,the risk level of electricity theft,is output,and combined with this level to achieve automatic early warning.The experimental results show that the CNN-SVM model has good convergence effect,short training time,and can accurately identify power theft users.The early warning accuracy and response speed applies to the automatic early warning of power theft risk in power grids are high.

关 键 词：支持向量机 卷积神经网络 窃电风险 自动化预警 决策函数 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]