This study introduces a novel approach that integrates dynamic Bayesian network with attention based spatio-temporal graph convolutional network to forecast railway train delays, capturing the intricate operation interactions between train events and the dynamic evolution of train delays. Initially, train delay patterns are identified using the K-means clustering algorithm and incorporated as additional variables into the prediction model. To capture dynamic causality in delay propagation, we utilize a dynamic Bayesian network-based dynamic causality graph, incorporating train delay data and domain knowledge to effectively model the train delay propagation. Leveraging these insights on delay dynamic propagation, we propose an attention based spatio-temporal graph convolutional network that effectively models the dynamic spatio-temporal dependency among train events and enhances the accuracy of delay predictions. The proposed method is assessed using operational data from the Wuhan-Guangzhou high-speed railway. Results show that the proposed model outperforms the baseline models, particularly with the expansion of the prediction horizon. The learned dynamic causality of the train delay propagation enhances interpretability and results in a 6.97% reduction in mean absolute error. Furthermore, train delay patterns and weather variables contribute to a respective 12.85% and 4.37% reduction in mean absolute error. The statistical tests further validate the efficacy of the proposed model.
Bayesian Spatio-Temporal Graph Convolutional Network for Railway Train Delay Prediction
IEEE Transactions on Intelligent Transportation Systems ; 25 , 7 ; 8193-8208
2024-07-01
4585427 byte
Article (Journal)
Electronic Resource
English