Mitigating cascading failures through intentional line removal: A conditional reinforcement learning approach

Cascading failures in power networks pose critical challenges to system resiliency, as they can trigger widespread blackouts through the uncontrolled propagation of failures. This paper proposes a novel conditional reinforcement learning approach to mitigate cascading failures via strategic transmission line tripping. Our method combines: (1) an interaction learning module that quantifies failure dependencies between network lines, and (2) a deep Q-network (DQN) architecture that integrates real-time failure probabilities as conditional inputs for remedial action selection. This work advances the integration of physics-informed deep reinforcement learning by explicitly modeling inter-component dependencies to derive real-time mitigation strategies that improve power system resilience. We train the framework using a cascading failure simulator built on the IEEE 300-bus, 118-bus, and 57-bus systems, and evaluate line-tripping actions using DC optimal power-flow analysis. The proposed algorithm’s input is the state of the power network, and the output is the transmission line whose tripping mitigates the unfolding cascading failure. Experimental results demonstrate that our conditional reinforcement learning method reduces cascade sizes and failed power compared to no-action baselines, outperforming standard Q-learning in reducing cascade size. Furthermore, experimental evaluations under modified network topologies and AC power flow conditions verify the robustness of the proposed approach. Representative scenarios are also provided to illustrate the step-by-step mitigation process. Finally, an analysis of decision latency highlights the severe impact of delayed line tripping, underscoring the critical necessity for real-time control algorithms.