Modeling train route decisions during track works. (June 2022)

• Record Type: Journal Article
• Authors: Schmid, Basil; Becker, Felix; Molloy, Joseph; Axhausen, Kay W.; Lüdering, Jochen; Hagen, Julian; Blome, Annette
• Is Part Of: Journal of rail transport planning & management. Volume 22(2022)
• Publication Date: 2022-06
• Subjects: Train route decisions -- Construction schedule -- Track works -- Discrete choice model -- Machine learning classifier -- Forecasting -- Behavioral outcomes; Railroads -- Periodicals; Railroads -- Planning -- Periodicals; Railroads -- Management -- Periodicals; Railroads; Railroads -- Management; Railroads -- Planning; Periodicals; 385.068
• DOI: 10.1016/j.jrtpm.2022.100320
• Languages: English
• ISSNs: 2210-9706

Abstract

Abstract: To better understand the choice behavior of train route schedulers and to predict their choices for optimizing the annual construction schedule, prospective data for 2020 on train route decisions are analyzed using discrete choice models and machine learning classifiers. The choice alternatives include (i) partial cancellation of the train schedule at the start, (ii) in the middle or (iii) at the end of the itinerary of the train service, (iv) detour and (v) delay/ahead of time, and are modeled using 39 train-, construction site-, and infrastructure variables. The top nine attributes account for about 80% of variable importance, including the travel time from the departure station to the construction site, total or line closure, travel time from the construction site to the terminus, length of the train and effective line capacity. The models are tested for 2021 and 2022 to verify whether they can be used to forecast choices in the following years. While Random Forest performs best in terms of prediction accuracy (2021: 60.8%; 2022: 58.6%), the improvements of about 6%-points compared to the Mixed Logit model are modest. Results indicate that a substantial amount of unobserved construction site heterogeneity is present, which Random Forest cannot capture either. Highlights: Large panel datasets to investigate choices of train schedulers during track works. Machine learning and Logit models for optimal performance in different outcomes. Simplification strategies to overcome the computational costs of Mixed Logit models. Accounting for unobserved heterogeneity increases out-of-sample prediction accuracy.