Deep Reinforcement Learning for Emergency Resource Allocation in Urban Rail Stations

As the lifeline of modern urban transportation, urban rail transit faces complex challenges caused by rapid urbanization and population growth. Especially for subway stations, due to the spatiotemporal complexity of their passenger flow, the operational problems have become increasingly complex, and the challenge lies in the effective scheduling of subway station resources. Currently, the resource scheduling mainly depends on the work experience of the station leaders. However, excessive reliance on experience can easily lead to resource waste and low decision-making efficiency, especially in emergency or unexpected situations that may affect the overall system’s operational quality and passenger safety. To address this issue, this study proposes a deep reinforcement learning scheduling decision model based on a single objective and multiple constraints. This model comprehensively considers factors such as employee skill matching, task urgency, and multi-site linkage. By collecting on-site survey data and resource scheduling plans from subway stations, this model aims to intelligently optimize station resource scheduling. We implemented the Proximal Policy Optimization (PPO) algorithm and constructed two feedforward neural networks as critics and actors. In a reward-based personnel scheduling environment, this model shows higher scores compared to manual experience allocation and traditional heuristic algorithms.