Comparative Study of Electricity-Theft Detection Based on Gradient Boosting Machine

Abstract

The detection of electricity theft is a crucial concern for electric utilities. The combination of the electricity consumption data and machine learning models bring a new way for analyzing the customer energy consumption pattern. This paper compares three gradient boosting machines for electricity theft detection, i.e., extreme gradient boosting, light gradient boosting machine and cat boosting. By taking the advantages of gradient boosting machines in processing numerical data, this paper conducts experiments on a realistic dataset released by State Grid Corporation of China with true malicious samples. We investigate the impact of different parameters of gradient boosting machines on the electricity theft detection performance by simulations. We also compare the performances of gradient boosting machines with the wide and deep convolutional neural network. Experimental results show that gradient boosting machines outperform the wide and deep convolutional neural network for electricity theft detection using customers' electricity consumption data.