Focusing on the problems of low accuracy, poor real-time performance, and no feature extraction in the current grid single learner power-theft detection method, a power-theft behavior detection method based on the Convolutional Neural Network-Light Gradient Boosting Machine (CNN-LG) model is proposed. First, the power fea? tures of user electricity data are extracted through the Convolutional Neural Network (CNN), and the extracted fea? tures are input into the Light Gradient Boosting Machine (LightGBM, LG) classifier based on the decision tree in or? der to train the data. On this basis, a detection method of electricity theft based on the CNN-LG model is estab? lished. Finally, the State Grid Corporation of China and Irish Smart Energy Trail（ISET）datasets are used to conduct experiments to verify the accuracy and effectiveness of the method proposed in this paper. The experimental results show that the method proposed in this paper can quickly and accurately realize the detection of various power theft behaviors in the power grid. Compared with the existing detection methods, it has higher accuracy, better generaliza? tion performance, and real-time performance.