Abstract: This paper studies the topology control strategy of distribution network fault recovery based on deep reinforcement learning. First, design the distribution network topology state representation and decision-making action rules to support the combined optimization solution. Secondly, use the improved pointer network structure and deep reinforcement learning algorithm to achieve model self-learning and end-to-end calculation suitable for multiple types of failure recovery strategies. Finally, by improving the mask mechanism to reduce the complexity of exploration and solving, and effectively improve the efficiency of training and learning. By randomly setting fault combinations on the preset lines, the effectiveness of the improved mechanism and model proposed in this paper is verified on the single and mixed initial state sample sets, which provides an effective reference for the application of deep learning technology in the optimization of distribution network operation mode.