Abstract: The deep integration of cyber-physical systems enables efficient operation of smart grid systems while also exposing them to security threats posed by cyber-physical attacks. By injecting false data, attackers can achieve no change in measurement output and thus deceive the traditional detection methods based on chi-square. By considering the impact of false data attack on the internal state change of the system, a detection method against false data attack based on neural network observer is proposed. Based on the established smart grid physical dynamics model, the stealthy characteristics of the false data attack are analyzed. Considering the impact of the false data attack on the internal state change of the system, the state residual detection method based on the neural network observer is proposed. In addition, considering the impact of perturbation on the threshold, adaptive thresholds are designed to replace the traditional empirical thresholds for cutting the false data attack detection time. Finally, the superiority of the proposed state residual attack detection method based on neural network observer is verified in IEEE 3-generator 6-bus grid system.