Graph Reinforcement Learning-Based Fault Reconfiguration for Shipboard Integrated Power Systems

Objectives To address the problems of drastic local topological changes and the difficulty of rapidly restoring critical de-energized loads when modern shipboard integrated power systems encounter sudden faults, this study explores the optimal self-healing reconfiguration strategy to maximize load restoration and minimize the number of switch operations under constraints.. Methods The shipboard distribution network is abstracted into a graph, and an optimization model including node power balance and generator output constraints is established. Graph Convolutional Networks (GCN) are introduced to extract dynamic topological features, and combined with the Advantage Actor-Critic (A2C) algorithm to construct a Graph-A2C decision framework. A composite reward function integrating the restoration objective and limit-violation penalties is designed to guide the agent to optimize within a discrete action space. Results Simulations show that. Conclusions It provides a new solution for the intelligent self-healing of shipboard integrated power systems.