Distributed Robust Dispatch for Networked Microgrids With Coalition Game-Guided Multiagent Adversarial Safe Reinforcement Learning

The integration of distributed energy resources introduces demand-side flexibility and uncertainty, challenging dispatch of networked microgrids (MGs). This article proposes a coalition game-guided multiagent adversarial safe reinforcement learning (SRL) method to achieve distributed robust dispatch under privacy protection. Leveraging SRL for rapid decision-making with safety guarantees, we further overcome the reliance on predefined source-load distributions inherent in existing SRL approaches. Specifically, we employ the Wasserstein generative adversarial network to assume additional distributional information on source and load transition probabilities. This unique design for MG dispatch enhances decision robustness under source-load fluctuations compared to existing adversarial training, which gives adversary the same action space as the protagonist. To ensure privacy-preserving multiagent training, we construct a coalition game-guided training framework. This framework adjusts protagonists’ rewards based on game theory to ensure global convergence only with MGs’ boundary power, and uses Hash-based message authentication code to guarantee data integrity. Unlike existing multiagent reinforcement learning methods relying on global critic, our method avoids sharing privacy-sensitive local states and actions. Simulations on a modified IEEE-123 bus system and an actual power grid-derived test system validate the effectiveness.