Research on Microgrid Resilience in Highway Service Areas Based on Federated Multi-Agent Deep Reinforcement Learning

This paper proposes a Federated Multi-Agent Deep Reinforcement Learning (FMADRL) framework to enhance the resilience of highway service area microgrids against extreme weather events. The method integrates Generative Adversarial Networks with Monte Carlo simulations to generate high-fidelity weather scenarios, enabling privacy-preserving collaborative optimization across distributed microgrids. A multi-objective approach using the Ripple-Spreading Algorithm yields balanced solutions for economic efficiency, reliability, and response speed. Large-scale simulations demonstrate significant improvements: the proposed method achieves an 88.3 score on the comprehensive system resilience metric, reduces the average fault recovery time from 46.6 min to 8.4 min, lowers annual operating costs by 69.3%, equivalent to 536,945.1 USD, and achieves annual carbon emissions reductions of 285 Mg. This approach provides an innovative solution for enhancing the resilience of distributed microgrids during extreme weather events.