Train Delay Prediction based on a Multimodal Deep-learning Method

High-speed rail (HSR) trains may suffer delays under emergencies, which will disturb the normal operation of trains and reduce passengers' travel satisfaction. Accurately predicting the delay time can help the dispatchers make advanced rescheduling strategies and thus improve train operation efficiency. Therefore, to quickly and effectively predict the delay time of trains for the HSR network under emergencies, a delay prediction model based on a multi-modal deep learning framework is proposed. The model considers the temporal and spatial characteristics and time-series characteristics of train operation, and then the Transformer and the three-dimensional convolutional neural network (3D CNN) are adopted to independently extract and merge these features. Specifically, the Transformer with the attention mechanism solves the interference mechanism among trains and stations on the delay of the train of interest. At the same time, 3D CNN is used to extract the temporal and spatial data characteristics of trains. Finally, the real operation data of the Beijing Railway Administration HSR network is adopted to verify the effectiveness of the proposed model. Results show that compared with the results obtained by other benchmark models, the predicted delays by the proposed model can lower the average absolute error (MAE), and root mean square error (RMSE) by 10% and 12%, specifically.