Aiming at the problem of low detection accuracy in existing methods for detecting electricity theft users, a multi-time scale deep learning method for detecting electricity theft users is proposed based on the smart grid data acquisition system. The bidirectional long short-term memory network extracts daily electricity consumption features, the residual network ResNet extracts weekly electricity consumption features, the deep convolutional neural network AlexNet extracts monthly electricity consumption features, and the AdaBoost classifier classifies whether users steal electricity. The superiority of the proposed method is verified through case analysis. The results indicate that, the proposed method fully utilizes the powerful feature extraction capabilities of deep learning and the efficiency of AdaBoost in classification tasks, by analyzing electricity consumption data at different time scales, it improves the detection performance of electricity theft users. Compared with conventional electricity theft detection methods, the proposed method can more comprehensively reflect the characteristics of normal users and electricity theft users, and has the best performance in multiple indicators, with an accuracy rate of 91.28%, ROC-AUC value of 0.950 5, and PR-AUC value of 0.946 9. The proposed method not only helps reduce energy loss, but also provides some assistance for achieving the dual carbon target.