Deep Reinforcement Learning-Based Rescue Resource Distribution Scheduling of Storm Surge Inundation Emergency Logistics

Abstract

A high-intensity storm surge hazard exposes residents and facilities to inundation danger. A continuously operating robust inundation emergency logistics system is vital to guarantee lives and support the control of hazardous materials. A massive inundation area, large number of affected facilities, and instantly changing flooding situation bring tremendous challenges to rescue resource allocation. In this article, a rescue resource distribution scheduling of storm surge inundation logistics is proposed to quantitatively formulate the rescue time minimization problem in emergency logistics. The mixed-integer linear programming (MILP) method is proposed for the emergency logistics scheduling model validation and optimality comparison. To enhance the efficiency of creating a good quality allocation strategy when facing large-scale problems, a deep reinforcement learning algorithm—deep deterministic policy gradient (DDPG)—is utilized to search the solutions. Based on a rescue resource scheduling model of storm surge inundation logistics targeting storm surge Mangkhut in September, 2018, a case study of the Futian District, Shenzhen, China, was conducted to verify the correctness and efficiency of the MILP and the DDPG. The optimal schedule solution designed by the MILP had a duration of 3.5138 h, while the solution calculated by DDPG had a duration of 5.065 h for the rescue. The execution time of DDPG was stable and under a second, while the execution time of the MILP was over two hours.