A partially observable decision-making algorithm for network topology restoration in post-disaster power systems

The increasing frequency of extreme weather events and cyber-physical disruptions has made rapid and reliable power grid restoration a crucial challenge for modern energy infrastructures. This paper presents a Partially Observable Decision-Making Algorithm for Network Topology Restoration (PODAR) that integrates probabilistic inference, graph-based decision modeling, and hierarchical reinforcement learning to address recovery under incomplete information. The proposed method models the post-disaster restoration process as a partially observable Markov decision problem (POMDP), where system states, sensor readings, and communication links may be degraded or unavailable. By introducing belief-state updates, uncertainty propagation, and attention-driven graph embeddings, PODAR learns adaptive control policies capable of restoring network connectivity and energy flow efficiently. Simulation experiments on multi-scale grid models demonstrate that the proposed framework achieves higher restoration efficiency, faster decision convergence, and stronger robustness compared with existing methods such as DGC, MADDPG, and MPC. Furthermore, PODAR exhibits strong generalization across regional and topological variations, confirming its potential as a foundation for intelligent, self-healing power systems in future resilient energy networks.