Integrating graph and neural relational inference for network-wide train delay prediction: An equilibrium between accuracy and interpretability

Abstract

Accurate train delay predictions contribute to real-time decision-making, and comprehending the intricate interactions among diverse elements of delay evolution holds paramount significance for tactical timetabling. However, existing research struggles to strike an equilibrium between the accuracy and interpretability of delay prediction models. This paper introduces NRI-GraphSAGE, a predictive model for railway network delay evolution, successfully harmonizing interpretability and accuracy by integrating neural relational inference (NRI) and Graph Neural Networks (GNNs). The proposed model follows a standard encoder-decoder structure. The model’s encoder module employs a variational autoencoder structure to learn train-train interactions. In the model’s decoder module, heterogeneous GNNs are used to process the acquired train-train interactions and other information guided by domain knowledge. Case studies on two local networks of the Chinese high-speed railway affirm the rationality of each module within NRI-GraphSAGE and showcase its outstanding predictive accuracy. Through experiments, we affirm the significance of interactions between elements (station-train, disturbance-train, station-station) in the railway network, alongside the sensitivity of influencing features. Furthermore, an analysis of the learned train-train interactions reveals that multiple adjacent trains can interact, and the strength of interactions increases with the decrease of headways or growth of train delays. Compared with existing approaches that rely on predefined relationships, our model automatically infers these interactions from historical data, more accurately capturing critical train interactions. Consequently, the high predictive accuracy of NRI-GraphSAGE furnishes dispatchers with a foundation for crafting rescheduling decisions, while explaining the interactions of different elements during the delay evolution lends support to the allocation of recovery time in timetable planning.