Machine learning based prediction of train punctuality in Sweden : Dynamic, multi-station delay forecasting with STGNN and tree-based models

Sammanfattning

Train arrival delays significantly impact the passenger experience, freight efficiency, and the workflow of planning and operations personnel. Traffic controllers play a crucial role in managing the rail network and directly influence the punctuality of the system. While improving overall train punctuality is a complex, multi-layered challenge, providing traffic controllers with accurate train delay forecasts enable them to prioritize effectively and make optimal decisions. This thesis explores how train punctuality can be predicted using dynamical multi-station regression, comparing ElasticNet, tree-based models, and Spatial Temporal Graph Neural Networks (STGNNs). The models were trained, validated, and tested using 2024 LUPP data from the railway segment between Stockholm and Hallsberg (via Mälarbanan), a railway corridor characterized by diverse traffic and a mix of single and multiple tracks. Furthermore, this paper contributes to the research on machine learning decision-support tools by developing a web application that operationalizes these models. Through this user interface, aggregated punctuality predictions and network status are actively communicated to traffic controllers to enhance oversight. Ultimately, the STGNN achieved a mean absolute error of 1.637 minutes, achieving an 11.4% improvement over the last-observation-carried-forward baseline.