Abstract: The traditional power theft detection method can only identify whether the user has power theft, but cannot perform rapid and accurate inspections for various types of power theft users. Aiming at the characteristics of medium-voltage users with large and regular power consumption, this paper proposed a method for detecting the type of power theft in medium-voltage distribution lines based on robust regression and convolutional neural networks. Firstly, considering abnormal data due to factors such as communication delay interruption, a robust regression algorithm is used to reduce its impact and improve the accuracy of regression analysis. Secondly, the correction coefficient and error term of each user obtained by regression are taken as the characteristics of the user stealing electricity and input into the convolutional neural network model for training to complete the identification of stealing electricity type. Finally, the method is verified by simulation and measured data. The results show that under different disturbance conditions, the proposed method can accurately identify different types of power stealing behaviors, which can better assist the on-site investigation, narrow the investigation scope, and improve the verification rate.