Sequential Restoration of Active Distribution Networks via LLM-Guided Optimistic Distributional Reinforcement Learning

Data-driven reinforcement learning (RL) often struggles to process unstructured human semantics during distribution network restoration. To address this challenge, this paper proposes a method integrating large language model (LLM)-based semantic space guidance with optimistic distributional RL (DistRL). A fine-tuned LLM parses complex human semantics into dynamic semantic observations and action mask constraints. Concurrently, an optimistic implicit quantile network (O-IQN) employs a high-quantile optimistic preference mechanism to guide the agent toward restoration paths with higher rewards and success rates. The method is validated under random N-1, N-2 and N-3 contingencies on modified 70-bus system and 136-bus systems. Compared with baselines, it accelerates convergence and enhances cumulative rewards. Comprehensive evaluations across 50,000 scenarios on the 70-bus system and 300,000 scenarios on the 136-bus system show that O-IQN consistently ensures high recovery rates. This approach effectively integrates unstructured human semantics with data-driven control, enhancing the robustness and safety of distribution network restoration.