Modeling and Optimization of Power Grid Intelligent Restoration Strategy Combining Diffusion Models and Reinforcement Learning

Abstract

With the increasing proportion of renewable energy integration in modern power systems, grid fault recovery faces greater complexity and uncertainty, posing significant challenges for intelligent energy transmission infrastructures and reliable operation in electromagnetic power networks. To address these issues, this paper proposes a collaborative decision-making model (CDM-RL) that combines diffusion models (DM) and reinforcement learning (RL) for intelligent power grid restoration. The proposed framework generates physically feasible fault scenarios through a denoising diffusion probabilistic model (DDPM) and integrates graph neural networks (GNN) with proximal policy optimization (PPO) to achieve efficient restoration control under complex operating conditions. An alternating training mechanism is further introduced to enhance the generalization capability of the model across diverse fault scenarios. Experimental evaluation on the IEEE 33-bus system demonstrates that CDM-RL significantly outperforms conventional approaches, achieving an initial power restoration time of 8.9 s, a load recovery rate of 93.1%, and an average of only 3.9 switching operations while maintaining superior cross-scenario stability of 85.3%. The proposed AI-driven framework improves grid self-healing capability and provides an effective technical pathway for enhancing the resilience and operational reliability of intelligent electromagnetic energy transmission and distribution systems.