Discerning Primary and Secondary Delays in Railway Networks using Explainable AI
David Rößler, Julian Reisch, Florian Hauck, Natalia Kliewer
Abstract
In this paper, we study the problem of discerning different reasons for which train delays occur. Given the total amount of delay a specific train builds up at a specific station, we discern the primary delays that would have occurred if there was no other train in the network, such as vehicle problems, from secondary delays which are knock-on delays. Our approach is to train an ML model that predicts the additional delay of a train, given a set of primary features such as weather and secondary features such as the delays of nearby other trains. Methods from explainable AI help to classify to which amount the primary features and to which amount the secondary features contribute to a specific prediction of the model. In particular, we apply SHAP values for the use case of delay management. In addition, we propose a novel method for the use case of railway simulations. In both cases, we use the classification to discern the different reasons for the specific delay.