To solve the problems of insufficient accuracy and low detection efficiency in existing detection methods of electricity theft in smart grids, a CAEs-LSTM detection model combining convolutional auto-encoders (CAEs) with long short-term memory networks (LSTM) is proposed. The model conducts two-dimensional conversion to power data, designs the encoder structure by analyzing the characteristics of data set, and reconstructs the two-dimensional space characteristics of the electricity data using pooling layers, down and up sampling layers. It adds Gaussian noise to improve its robustness, and builds long short-term memory networks to learn the global characteristics. Finally, spatial-temporal characteristics are fused to detect energy thieves, and parameter tuning is performed. Based on the public available real data set of the State Grid, the CAEs-LSTM model is optimal in the value of mean average prediction and area under curve, by comparing the CAEs-LSTM model with support vector machines, the LSTM model, and wide and deep convolutional neural networks. Simulation experiments show that the theft detection method based on the CAEs-LSTM model has higher detection efficiency and accuracy. This work is supported by the National Natural Science Foundation of China (No. 61873277).