基于知识图谱增强大语言模型强化微调的配电网动态重构OA

Dynamic reconfiguration of distribution network enables more flexible operational regulation through topology adjustment.However,the existing decision-making technologies struggle to balance computational real-time performance and reliability.Large language models(LLMs),with their powerful semantic understanding,in-context learning,and generative decision-making capabilities,demonstrate significant potential for complex decision-making problems such as dynamic reconfiguration of distribution network.Nevertheless,their direct application to dynamic reconfiguration optimization faces challenges including the absence of structured domain knowledge,insufficient task comprehension and adaptation capabilities,and difficulties in embedding physical constraints,thereby compromising decision reliability.To address these issues,this paper proposes a dynamic reconfiguration method of distribution network based on knowledge graph enhancing reinforcement fine-tuning of LLM.First,a dynamic knowledge graph is constructed to explicitly represent the temporal evolution and causal logical relationship of distribution network operation.Second,high-quality fine-tuning datasets are generated through subgraph sampling of knowledge graph,and the supervision and fine-tuning of the pre-trained LLM is completed to enhance its task comprehension and instruction-following capabilities.Finally,a multi-dimensional reward function is designed based on the structured prior knowledge from the knowledge graph,and the group relative policy optimization is introduced for reinforcement fine-tuning of the model.Case analyses demonstrate that the proposed method significantly improves the generation efficiency and physical consistency of dynamic reconfiguration strategies of distribution network,effectively improving the operational level of distribution network.