A Two-Layer Simulation-Based Optimization Framework for Coordinated Metro Rescheduling: Integrating Machine Learning Surrogates with Metaheuristics

Abstract Metro systems are prone to unplanned disruptions that can trigger cascading delays across the network, particularly when transfer stations are involved. Existing rescheduling approaches often treat the disrupted line in isolation or rely on indirect operational metrics (e.g., train delay) that fail to accurately reflect passenger utility during major disruptions. To address these challenges, this paper proposes a novel passenger-oriented coordinated rescheduling framework for a critical scenario in which a transfer station is located within a disrupted section. This framework coordinates multifaceted recovery tactics on the disrupted line with subtle timetable fine-tuning on the intersecting line. We formulated a two-layer simulation-based optimization model aimed at directly minimizing passenger travel time: a macroscopic layer optimizes strategic variables (e.g., headways, route choices, and dwell times), and a microscopic multiagent simulator evaluates these strategies by generating conflict-free timetables. To overcome the computational burden of high-fidelity simulations, a machine learning (ML) surrogate-assisted metaheuristic, named the surrogate-assisted tornado optimizer with Coriolis force (SA-TOC), was developed. Specifically, a CatBoost-based surrogate model is integrated to approximate the complex mapping between operational decisions and total passenger travel time, thereby accelerating the search process. A case study based on the Ningbo Metro validated the proposed approach. The results demonstrate that the coordinated strategy reduces the total passenger travel time by 23.2% compared with basic operational responses. Furthermore, the SA-TOC algorithm reduces the computational time by over 98.4% compared with standard metaheuristics, identifying near-optimal solutions with sufficient rapidity for real-time application.