Robust Reinforcement Learning-Based Resilient Operation of Networked Microgrids

As critical hubs of regional economies, process-oriented industrial parks (PIPs) are typically highly dependent on the main grid for power supply due to their continuous production processes and significant energy consumption. However, unexpected extreme events or disasters may cause main grid outages, resulting in production line shutdowns in PIPs. To enhance the resilience of PIPs, research has focused on transforming PIPs into microgrids (PIPMGs) that integrate renewable energy and local generation, utilizing energy dispatching methods to facilitate island operation during main grid outages. However, relying solely on its internal power generation equipment may not meet the high energy demands of PIPMGs, and existing dispatching methods frequently overlook the discrepancy between simulation models and real world systems. This paper introduces a networked microgrid (NMG) model, which connects PIPMGs with community microgrids (CMGs). By prioritizing power allocation to critical loads during outages and enabling the transfer of excess energy between different microgrids, this approach achieves the dynamic transfer of redundant power supply capacity, thereby improving the resilience of PIPMGs. Additionally, a robust reinforcement learning algorithm is constructed to optimize dispatching policies. By optimizing the agent's reward lower bound within the environmental uncertainty set, the algorithm ensures that its performance in real-world environments is at least as good as in simulated environments, thereby improving the robustness and adaptability of the dispatching policy. Finally, through a comparative analysis under various scenarios and algorithms, combined with metrics such as cumulative reward and task completion rate, the effectiveness of the proposed method in enhancing the resilience of PIPMG and the adaptability of the dispatching policy has been validated.