Key points are not available for this paper at this time.
The swift restoration of infrastructure systems damaged by hurricane hazards (e.g., strong winds, heavy rainfall, and high surges/waves) hinges on efficiently managing the limited resources for different repair tasks during the recovery process, while prioritizing repair tasks can be significantly influenced by the interdependencies among these infrastructure systems. The restoration for interdependent infrastructure systems following hurricane damage can be formulated as a stochastic sequential decision problem using Markov decision process (MDP). This study employs knowledge-enhanced deep reinforcement learning (RL) to tackle the dynamic optimization of the MDP. Specifically, a deep neural network (DNN) is used to represent the recovery policy, which maps observations of the system's status to the allocation decision of repair resources. The optimal DNN weights are determined through RL techniques. Domain knowledge is integrated into the learning process by the mechanisms of knowledge-guided exploration and knowledge-based reward to enhance the training efficiency of the deep RL approach. To illustrate the effectiveness of the proposed framework, a case study is conducted to optimize the post-hurricane recovery of an interconnected traffic-electric power network. The simulation results show the improved training efficiency achieved by incorporating domain knowledge into the deep RL algorithm and highlight the advantages of collaborative decision-making over independent scheduling.
Li et al. (Thu,) studied this question.