Abstract: As an important part of the Energy Internet, microgrid is of great significance for the local consumption of distributed wind and solar energy. However, the randomness of new energy output and the temporal and spatial volatility of load have brought huge challenges to the optimal dispatch of microgrids. In response to this problem, this paper uses a deep deterministic strategy gradient model based on the classification experience replay mechanism, adapts to the uncertainty of wind and load through data-driven methods to realize the optimal dispatch of the microgrid. Based on the consideration of the time-of-use electricity price and the penalty of abandoning wind turbine and photovoltaic output, the design of reward mechanism is to minimize operating costs and accommodate new energy to the greatest extent, and the classification of the experience pool based on the instant reward value improves the training speed and convergence performance of the model. Finally, the IEEE14-node case is used for simulation verification. The results show that the DDPG model in this paper can generate an optimal dispatch strategy in real time, it does not require accurate modeling of wind turbine and photovoltaic output and load. Meanwhile, compared with the DQN, the cost of dispatch is reduced by 4.73%.