An intelligent optimisation method of train rescheduling for delay scenarios

Train delays induced by emergencies inconvenience passenger travel on high-speed railway network. Both manual rescheduling and solving mixed integer programming (MIP) models struggle with real-time demands. We propose an intelligent optimisation method based on reinforcement learning, leveraging its offline-to-online deployment capability to promptly and precisely reschedule trains. A micro-topological network integrating stations and sections is constructed. An 8-dimensional state variable is defined to capture spatiotemporal resource occupancy, track assignment, time information and train priority, thereby ensuring environment perception under time-varying constraints (safety protocols and infrastructure capacity). The reward function is mathematically aligned with MIP. Tests on nine regional delay scenarios show the proposed method reduces delays by up to 29.5% through flexible adjustments compared with the first-come-first-served (FCFS) approach and exhibits higher real-time responsiveness than MIP and FCFS, with a response time