With the construction of smart city and smart grid, energy Internet is booming, and the deployment of smart power terminals has increased, with a wide range of deployment locations, and the level of intelligence of the terminals has increased. However, most of the intelligent terminals are deployed in the non-controlled industrial control network environment, which brings new challenges to the network security. Therefore, the traffic anomaly detection of electric power industrial control becomes the first step to ensure the safety of the peripheral network. Currently, the traffic anomaly detection system of electric power industrial control based on machine learning has the problems of low computing efficiency and weak representation ability. To solve the problem of low efficiency in traditional deep learning, the deep belief network (DBN) is used to extract the non-linear features of data, which can improve the time efficiency to a certain extent. To solve the problem that the representation ability of traditional machine learning is weak, long short term memory network (LSTM) is used to improve the representation ability of traditional shallow machine learning. The experimental results suggest that the method is better than the traditional machine learning method in accuracy, and the recall rate reaches up to 97.3%, AUC runs up to 0.927.