Electricity theft detection algorithm based on contrast learning and cluster combination discrimination

Abstract

Power system is susceptible to kinds of abnormal behavior of electricity consumption. Detection of electricity theft can improve the stability of power system. Aiming at the problems of dependence on label data and low accuracy of existing electricity theft detection methods, this paper proposes a theft detection algorithm based on contrastive learning and clustering combination discrimination. The algorithm is composed of representation extraction module and clustering combination discrimination module. The representation extraction module firstly uses BiGRU to extract the context information of energy consumption data, and then obtains the representation of power data based on dilated convolutional network, and uses hierarchical contrastive learning to improve the learning effect of the network. The cluster combination discrimination module is based on K-means and adaptive DBSCAN algorithm, and uses a combination discrimination mechanism to classify the representations and determine theft. In this paper, the effectiveness of the proposed method is verified by the public dataset SGCC, and the experimental results show that the proposed method is superior to other unsupervised anomaly detection methods.