LLM‐Powered Multi‐Agent RL Framework for Real‐Time Power Grid Dispatching

Abstract

Real‐time power grid dispatching (PGD) is an essential task in power‐system operation, but it remains challenging because of high dimensionality, nonlinearity, renewable‐energy uncertainty, and strict response‐time requirements. Classical optimization‐based methods may become computationally burdensome or insufficiently adaptive under rapidly changing renewable generation. Existing multi‐agent reinforcement learning (MARL) strategies provide a promising data‐driven alternative, yet their practical use is limited by coordination blindness under partial observations, weak semantic understanding of grid topology and fault evolution, and limited generalization under rare events. To address these issues, this paper proposes a large‐language‐model‐powered agent‐orchestration framework, termed LLM‐AOM, for real‐time PGD. The framework contains a high‐level orchestration layer and a low‐level execution layer. The LLM is used only as a safety‐filtered supervisory orchestrator: it converts grid states into structured situational‐awareness descriptions and issues macro‐level coordination guidance. Physical dispatch actions are generated by MAPPO‐based low‐level agents within predefined operational limits. For reproducibility and feasibility, the implementation uses Grid2Op‐based simulation, an IEEE 39‐bus test system, structured LLM‐output parsing, bounded reward shaping, safety filtering, fallback execution, latency measurement, and matched‐seed comparison. Simulation results on the IEEE 39‐bus benchmark indicate that LLM‐AOM improves security‐related indicators and post‐contingency recovery relative to DCOPF, PPO, and MAPPO in the tested scenarios. The study therefore provides a proof‐of‐concept for using LLMs as interpretable, safety‐filtered high‐level orchestrators rather than direct physical controllers in critical infrastructure control.