A Disturbance-Driven Textual Model for Train Running Time Prediction on High-Speed Railways

Abstract

Accurately predicting train running times (TRTs) during disturbances is crucial for effective timetable rescheduling. Previous studies on train running prediction have overlooked the detailed textual description of disturbance events. To improve the prediction accuracy, this paper proposes a hybrid neural network model consisting of a transformer encoder and a fully connected neural network (FCNN) to predict the TRTs under disturbances, which is called transformer-FCNN here. In the proposed transformer-FCNN architecture, a transformer encoder is used to process textual disturbance events, while FCNN is used to deal with static features. The performance of transformer-FCNN is validated using data from the Wuhan-Guangzhou and Xiamen-Shenzhen high-speed railways. The results show that the models considering the disturbance category can substantially improve the predictive performance compared with those that do not. Further, based on the same algorithm, models which use textual disturbance event records are shown to reduce the mean absolute error and the root mean squared error, respectively, by over 10.8% and 8.3% on average compared with those that only consider the disturbance category. The proposed model can support potential strategies of train operation control, by providing accurate prediction results.