With the development of Internet of Things, the network security problem of smart substation has been widely concerned. In the face of information physical equipment failure, virus intrusion and other threats to network security, this paper proposes an anomaly detection method for smart substation network based on deep reinforcement learning. Firstly, combined with the characteristics of smart substation, the automation network architecture of smart substation is constructed, and the application scope of manufacturing message specification (MMS), general object-oriented substation event (GOOSE) protocol, sample value (SV) message, message queue telemetry transmission (MQTT) protocol and limited application protocol (CoAP) are defined. Then, the communication packets in the smart substation network are preprocessed to obtain the traffic characteristics which can reflect the abnormal network. Finally, deep reinforcement learning is used to judge whether the real-time data packets are abnormal. The detection results of each MMS/GOOSE/SV/MQTT/CoAP packet are saved as log records, and the records are added to the learning model to improve its detection performance. The experimental results based on the typical 110 kV smart substation automation system show that the proposed method can accurately detect the abnormal network traffic, and compared with other methods, its false alarm rate is the smallest and the delay is the shortest, so it has a good application prospect.