基于堆叠去相关自编码器和支持向量机的窃电检测

摘要

已有窃电检测模型的准确率尚无法满足应用需求, 是因其均将建模重点放在了分类算法的选择或改进上, 而相对地忽视了特征提取过程.因此, 提出一种基于深度学习的特征提取方法, 即堆叠去相关自编码器.得益于深层结构和高度非线性, 其能够从用户用电数据中提取到高度抽象和简明的特征.随后支持向量机将这些特征映射到指示是否窃电的标签.基于真实数据的算例测试, 验证了所提窃电检测模型具有较高的检出率和较低的虚警率, 同时也验证了堆叠去相关自编码器能够提取到有效的特征.%The performance of existing consumption-pattern based models is not sufficiently satisfactory for application.This is partly because most of them have focused on the selection of classification algorithms rather than the design of features, and existing feature design methods with respect to electricity nontechnical loss (NTL) detection remain inefficient and unsatisfactory.A deep-learning based feature extraction model is proposed for NTL detection, namely the stacked uncorrelating autoencoder (SUAE).Due to the deep architecture and powerful uncorrelating ability of SUAE, features are extracted from load profiles concisely and effectively, which has thus enabled a great improvement in final NTL detection performance.Support vector machines (SVM) are applied as classifiers, which use the features extracted by SUAE to output a judgment result.Case studies on real datasets have demonstrated that the proposed NTL model has high relevance ratio and low false alarm rate, as well as the powerful feature extraction ability of the SUAE.