MAPPO-LLM: Multi-Agent Deep Reinforcement Learning with Large Language Model-Guided Priors for Coordinated Smart Grid Control

This paper proposes MAPPO-LLM, a multi-agent reinforcement learning framework enhanced by Large Language Model (LLM)-guided priors for smart grid control. A two-stage data augmentation mechanism is designed, combining MMD-based sample selection and LLM-based generation under both quantifiable constraints and unstructured expert knowledge. A dual-buffer architecture and hybrid loss function integrate prior-driven actions with real environment experience. Experimental results in a 24-hour multi-energy system scenario show that MAPPO-LLM achieves 27.6% lower average loss, 63.8% less volatility, and up to 0.1432 improvement during peak hours over baseline MAPPO, verifying the effectiveness of LLM in enhancing control robustness and training efficiency.