Topology-aware graph reinforcement learning for generalizable cascading failure mitigation in hybrid AC/DC power systems

Abstract

Cascading failures in hybrid AC/DC power systems can propagate rapidly through DC-side disturbances, AC branch overloads, topology changes, and islanding, making emergency mitigation a constrained sequential decision-making problem under evolving network structures. This paper proposes a topology-aware graph reinforcement learning framework for active cascading failure mitigation. The proposed graph attention network-proximal policy optimization (GAT-PPO) policy combines an edge-aware graph attention encoder, shared action-conditioned scoring heads for branch disconnection and generation redispatch, a state-dependent physics-constrained action mask, and severity-aware scenario sampling. Case studies on an IEEE 39-bus hybrid AC/DC system show that GAT-PPO and a controlled multilayer perceptron-proximal policy optimization (MLP-PPO) baseline achieve similar performance on the seen base topology, while GAT-PPO attains higher load survival rate (LSR) on unseen rewired topologies. Across five random seeds, the mean unseen-topology advantage is +0.0130 ( p ≈ 0.012). Cross-scale validation yields a mean advantage of +0.0183 ( p ≈ 0.002) on the IEEE 118-bus system, accompanied by stronger reductions in branch tripping and cascade depth, while the IEEE 300-bus system further shows a mean advantage of +0.0181 ( p ≈ 0.011). These results show that topology-aware state representations improve the robustness of emergency mitigation policies under unseen connectivity changes.