An online approach for integrated railway schedule replanning and passenger reassignment with unknown passenger reactions under hub disruptions

Abstract

Major railway disruptions, such as city-wide disasters, in hub areas often persist for a long time and affect multiple routes simultaneously. Original railway schedules, including line plans, timetables, and rolling stock dispatching, become invalid and require strategic replanning to maintain serviceability. Under such conditions, original trains may be delayed or canceled, resulting in passengers stranded in overcrowded stations or abandoning trips and leaving independently. An effective replanned railway schedule should minimize not only passengers’ delay but also the number of passengers abandoning their trips and the risk of overcrowding during the disruption period. However, passengers’ independent reactions make the effective passenger demand that practically complies with the replanned railway schedule unknown beforehand. Therefore, this paper proposes an online approach that dynamically and integrally optimizes railway schedule replanning and passenger reassignment by referring to the real-time passenger demand remaining at stations. This study designs a hierarchical deep reinforcement learning framework that operates without requiring deterministic forecasts of future passenger demand. A real-world disaster case in Chinese railway network demonstrates the advantage of our approach in offline computational efficiency by benchmarking against single-level deep reinforcement learning, polynomial surrogate models, rule-based strategies and rolling horizon approaches. The online application is validated across diverse disruption scenarios. This result shows that our approaches have strong potential to mirror the dispatcher’s role for real-time disruption management.