Deep reinforcement learning-based resilience optimization for infrastructure networks restoration with multiple crews

Restoration of infrastructure networks (INs) following large disruptions has received much attention lately due to examples of massive localized attacks. Within this challenge are two complex but critical problems: repair route identification and optimizing the sequence of the repair actions for resilience improvement. Existing approaches have not, however, given due consideration to globally optimal enhancement in resilience, especially with multiple repair crews that have uneven capacities. To address this gap, this paper focuses on a resilience optimization (RO) strategy for coordinating multiple crews. The objective is to determine the optimal routes for each crew and the best sequence of repairs for damaged nodes and links. Given the two-layered decision-making required—coordinating between multiple crews and optimizing each crew’s actions—this study develops a deep reinforcement learning (DRL) framework. The framework leverages an actor-critic neural network that processes IN damage data and guides Monte Carlo tree search (MCTS) to identify optimal repair routes and actions for each crew. A case study based on the 228-node power grid, simulated using Python, demonstrates that the proposed DRL approach effectively supports restoration decision-making.