There is a problem that the existing electricity theft detection of traditional power grids only uses a single amount of electricity and the detection accuracy is low under the actual dataset. Thus a method of electricity theft behavior detection based on a channel attention network improved convolutional neural network model is proposed. First, a multi-source data fusion electricity theft evaluation index system is established. This includes the trend of electricity consumption, the growth rate of line loss, and the terminal alarm, so as to construct the user's electricity consumption feature set. Then, the convolutional neural network (CNN) model is optimized based on the channel attention squeeze and excitation networks (SENet). Thus an electricity theft detection method based on an improved CNN model of the adaptive channel attention network is constructed. Finally, the dataset of the China Southern Power Grid is used to verify the validity and accuracy of the proposed method. Actual measurement results show that the proposed method can effectively realize the accurate detection of various electricity theft behaviors in the actual power grid situation, and the established evaluation index system can more clearly characterize the electricity theft behavior rules. The constructed electricity theft detection model can adaptively adjust the importance of feature channels to improve channel utilization. Compared with existing detection methods, it has higher accuracy and better generalization ability. This work is supported by the National Natural Science Foundation of China (No. 51777061).