For a dispatcher, an incident’s duration is only part of the problem. Similar incidents can have different consequences depending on their location, service frequency and the surrounding infrastructure. Useful decision support needs to connect the incident with its wider effect on the timetable.
Methods
We analysed Danish railway incident records from September 2017 to February 2019 and joined them with weather, location, network-centrality, headway, track-type and incident-severity attributes. Linear regression, generalised linear models, XGBoost and neural networks were tested for the two prediction tasks. The framework first estimated incident duration and then passed that estimate into a model of total delay, allowing us to examine both stand-alone performance and the effect of linking the predictions. Variable-importance analysis helped identify which information contributed most to each estimate.
Findings
XGBoost gave the strongest duration estimates, while a neural network performed best for total-delay prediction. Weather, incident characteristics, headway and network features helped explain disruption outcomes, but incident duration remained the harder target. The linked framework made a key operational dependency visible: error in the duration estimate propagates into the subsequent total-delay estimate. Manually entered records and relatively few severe incidents limited performance, pointing to better incident reporting and larger samples as priorities for improving prediction.
