Abstract: With the deepening of the coupling between information domain and physical domain in new power system and the rapid development of cyber attack technology, power industrial control systems are facing the threat of customized cyber attack, among which high-stealth false remote control command injection（HFCI） attacks have become one of the most destructive cyber attack types. This paper presents a HFCI attack detection method for power industrial control system. First, the optimized convolutional neural network model is used to detect HFCI and filter abnormal packets at the shallow application layer for IEC 60870-5-104protocol business traffic. Then, HFCI attack commands at deep application layer are detected through the factory-level command threat assessment module and the system-level command risk judgment module. Finally, the IEEE 30-bus simulation system verifies the accuracy and generalization ability of the proposed HFCI attack detection method.