Two approaches for real-time robust rolling stock rescheduling given disruption duration uncertainty
针对中断持续时间不确定导致调度效率低下的问题，提出两种鲁棒实时机车车辆重新调度方法，利用荷兰铁路数据验证其能显著降低中断延长时的重调度成本。
In railway planning, effective disruption management tools can limit the impact of a disruption. A key challenge is that disruption durations are typically uncertain, leading to inefficiencies such as unnecessary shunting, seat-shortages, end-of-day inventory mismatches, and trip cancellations. To address this challenge, we propose two different approaches for incorporating robustness in real-time rolling stock rescheduling under uncertain disruption duration, an aspect that is typically assumed to be known in existing rescheduling models. The first approach, Light trip robustness (LTR) requires that a pre-specified percentage of critical trips are robust with respect to their incoming composition across different disruption durations. The second approach, Strict composition robustness (STR), ensures feasibility across all possible disruption durations without further changes to compositions or shunting operations. Both methods extend a well-known rolling stock rescheduling model by explicitly accounting for multiple possible disruption durations. We develop an online evaluation framework and conduct a computational study using real-world data from the Dutch railway network. Our results demonstrate that explicitly accounting for disruption duration uncertainty can significantly reduce rescheduling costs when disruptions last longer than initially expected. For instance, the STR model achieves up to 21% lower total costs compared to models that do not consider robustness, while the LTR model offers flexibility and performs well under moderate uncertainty. However, these benefits can come at a cost when the initially expected disruption duration is accurate. Overall, our results show that robust rolling stock rescheduling can significantly improve operational resilience and reduces the need for reactive updates. The appropriate level of robustness should be chosen based on the likelihood of disruption duration extensions.