A Tunable Electricity Theft Detection Method Based on Usage Habits of Customers

Abstract

With the wide application of the Advanced Measurement Infrastructure in power grid, electricity theft detection methods based on data analysis become a main stream for diagnosis of customers with electricity theft behavior. However, the existing anomaly submergence problem may affect the accuracy of electricity theft detection. In addition, the threshold selection of the existing unsupervised learning algorithm is relatively fixed and cannot adapt to changing detection scenarios. To solve this problem, this paper proposes an electricity theft detection method with a variable threshold. Based on the user's load shape dictionary obtained by weighted clustering, the cosine distance between the user's load and the load shape dictionary is used as the standard of the user's consumption anomaly degrees, and the effectiveness and applicability of the proposed method are verified by numerical experiments.