Abstract: In order to effectively deal with the inherent high nonlinearity and non-convexity of the distributed generation composite load system and improve the identification accuracy and efficiency of the model, a parameter identification method based on deep reinforcement learning and intelligent exploration mechanism is proposed. The parameter sensitivity analysis was carried out by using the data-driven feature kernel Lasso method, and the sensitivity weights reflecting the contribution of parameters to the model dynamics were obtained. The improved deep reinforcement learning with intelligent detection function was used for parameter identification. The numerical experiment results show that the method has high identification accuracy, can effectively avoid falling into local optimum, and has fast learning speed.