A graph attention network-deep deterministic policy gradient model for resilient urban rail transit network repair decisions

As urban rail transit networks (URTNs) continue to expand and evolve, integrating artificial intelligence technology into URTNs to make them more resilient against high-consequence disasters such as earthquakes is becoming increasingly paramount. With the aim of enhancing URTN recovery capability after earthquakes, this paper proposes a graph attention network-deep deterministic policy gradient (GAN-DDPG) model to support station repair decisions. Firstly, a framework for assessing the earthquake recovery capability of URTN is developed, which includes hazard intensity evaluation, station damage assessment, post-earthquake system recovery simulation, passenger flow relocation and resilience quantification. In the established GAN-DDPG model framework, the topological connectivity matrix of the URTN and the operational state of each station are encoded as inputs to the GAN; the output of GAN is the repair sequence corresponding to each damaged station, which is evaluated by the critic network and added to the DDPG process to generate a high-quality station repair sequence from a broad range of actions to maximize URTN resilience. By considering various earthquake damage scenarios of the Nanjing subway network, the superiority and robustness of the developed GAN-DDPG model were verified. The findings show that the station repair decisions determined by the GAN-DDPG model enable the URTN to restore system performance with the lowest resilience loss in the shortest time among those obtained by two greedy search-, one genetic algorithm-, one topology- and one function-based repairing prioritization methodologies. Meanwhile, the pre-trained GAN-DDPG model achieves rapid response in repair decision-making with high computational efficiency under new URTN stochastic earthquake damage scenarios through transfer learning.