A data-driven mixed-integer linear programming approach for real-time rescheduling of urban rail transit under rolling stock faults

Urban rail transit operations are susceptible to unexpected disturbances or disruptions, with rolling stock faults being a particularly common cause. Therefore, this paper focuses on the integrated rescheduling of the train timetable and rolling stock circulation in an urban rail transit line under rolling stock faults. Three typical scenarios arising from such faults are studied simultaneously, i.e., delay, out-of-service, and rescue. Taking general key practical constraints and scenario-specific constraints into account, multi-objective mathematical models are formulated for each scenario to optimize various dispatching measures, such as retiming, cancellation, short-turning, and backup rolling stock utilization. For computational tractability, the proposed models are transformed into equivalent mixed-integer linear programming (MILP) reformulations using some linearization techniques. In order to satisfy the real-time requirements of train rescheduling, a data-driven approach is developed to accelerate the solving process by fixing some decision variables in advance. Specifically, the prediction of binary variable values is treated as a classification task. After creating a dataset including different rolling stock faults and their respective optimal solutions generated by GUROBI, the correlations between optimal solutions and instance features are extracted through supervised learning based on the multilayer perceptron. By generalizing the extracted correlations to unseen instances, high-quality solutions can be found in a short time. Finally, numerical experiments are carried out based on the Beijing Yizhuang Metro Line. Compared to directly solving the original model using GUROBI, the proposed solution approach can reduce the average computation time by up to 91.49% with an average optimality gap of only 0.77%.