Enhancing resilience of electric vehicle charging management in hydrogen–electric coupled distribution networks: A risk-characterization multi-agent reinforcement learning approach

With the rapid growth of electric vehicle (EV) ownership, the deep integration of power grids, renewable energy, and transportation systems has led to the emergence of highly coupled hydrogen–electric distributed networks. In this complex environment of multi-energy coordinated operation, EV charging management systems face not only supply–demand imbalances caused by renewable fluctuations but also multiple external risks such as extreme weather, natural disasters, and cyberattacks, which impose higher demands on system resilience. To address the limitations of existing studies—most of which focus on steady-state or small-disturbance scenarios and lack coordinated optimization strategies for extreme events and uncertainties—this paper proposes a centralized training and decentralized execution (CTDE) multi-agent reinforcement learning framework integrated with a risk characterization mechanism. The framework builds a dynamic simulation environment integrating EV charging facilities and hydrogen–electric hybrid energy storage systems, and introduces diffusion models to enrich the distribution of risk features in training data, thereby improving the perception and identification of rare and extreme risk events. An attention-based information filtering module and a low-frequency, high-efficiency communication strategy are designed to reduce communication costs and latency while enhancing coordination efficiency among agents in high-dimensional, long-horizon scenarios. Experimental results, evaluated on multi-dimensional resilience indicators including risk loss, response capability, and overall system resilience, demonstrate that the proposed method outperforms other reinforcement learning algorithms in enhancing resilience, reducing operational costs, and improving cross-scenario generalization. The diffusion model also shows strong adaptability to extreme risk disturbances. The proposed algorithm achieves an average reduction of approximately 27.8 % in operational cost compared to the best-performing baseline across all test scenarios and disturbance levels. This result is obtained from repeated trials and averaged outcomes, covering a wide range of risk types and intensities, and demonstrates high statistical reliability.