Learning-based scope reduction for real-time train rescheduling optimization in station areas

Authors:

Abstract
Real-time train rescheduling requires the rapid generation of revised train timetables in response to unforeseen disruptions. Solving such problems using exact methods is computationally challenging, especially in case of infrastructure blockages at complex station areas with dense traffic. This paper introduces a Learning-based Scope Reduction (LSR) framework for local train rescheduling that effectively prunes the search space by fixing a subset of decision variables prior to optimization. The approach fixes nominal routing decisions of trains that are predicted to be unaffected by infrastructure disruptions. Logistic Regression, Random Forest, and Neural Network classifiers are trained on generated disruption scenarios and their corresponding high-quality rescheduling solutions obtained by solving a Mixed Integer Linear Programming model. To balance the trade-off in computation time savings and potential solution quality degradation, we propose an effective tuning method that strengthens prediction robustness to improve computational efficiency. We evaluate the proposed LSR framework on a real-world medium-sized station area in Germany. Results show that LSR achieves computation time reductions of up to 97%, while largely preserving solution quality in comparison to the full scope solutions. In the best configuration, the approach solves 95% of the most complex instances in real time, with only minor solution quality degradation of 2.3%. This work demonstrates the potential of a hybrid train rescheduling approach that combines supervised learning with exact optimization and offers a promising tool for intelligent real-time dispatching support.