Extraction and Interpretability of Reinforcement Learning Control Strategies for Distribution Network Fault Recovery Based on Decision Tree

In recent years, Reinforcement Learning (RL) has demonstrated potential applications in the field of power systems due to its ability to adapt to dynamic environments, and has achieved preliminary results in multiple domains. However, due to the "black box" nature of Deep Reinforcement Learning (DRL) strategies, their decisions are difficult to interpret and interact with, significantly limiting their application in power systems where safety and reliability are highly demanding. This paper proposes a decision tree-based reinforcement learning strategy extraction framework for distribution network fault restoration. Firstly, an RL model for distribution network fault restoration is trained through interaction with a power system simulation environment, and a corresponding dataset is generated. Then, a decision tree algorithm is used to extract control strategies and evaluate their interpretability. Finally, a distribution network fault restoration case study is conducted to verify the effectiveness and practical value of the proposed method. The research results show that this method can effectively balance complexity and interpretability, providing a new approach for the reliable application of reinforcement learning in practical power system scenarios.