SPSTN: Sequential Precoding Spatial-Temporal Networks for Railway Delay Prediction

Abstract

Predicting the delay time of trains is an important task in intelligent transport systems, as an accurate prediction can provide a reliable reference for passengers and dispatchers of the railway system. However, due to the complexity of the railway system, interactions of various spatio-temporal variables make it difficult to find the rules of delay propagation. We introduce a Sequential Precoding Spatial-Temporal Network (SPSTN) model to predict the delay of trains. SPSTN consists of a Transformer encoder that captures long-term dependencies in time series, and spatio-temporal graph convolution blocks that model delay propagation at both temporal and spatial levels. Experiments on a subset of the British railway network show that SPSTN performs favorably against the state-of-the-art, which verifies that the combination of sequential precoding and spatio-temporal convolution can effectively model delay propagation on railway networks.