A hybrid Bayesian network model for predicting delays in train operations. (January 2019)

• Record Type: Journal Article
• Authors: Lessan, Javad; Fu, Liping; Wen, Chao
• Is Part Of: Computers & industrial engineering. Volume 127(2019)
• Page Start: 1214
• Page End: 1222
• Publication Date: 2019-01
• Subjects: High-speed rail -- Train operation -- Punctuality -- Bayesian networks -- Delay prediction -- Performance evaluation; Engineering -- Data processing -- Periodicals; Industrial engineering -- Periodicals; 620.00285
• DOI: 10.1016/j.cie.2018.03.017
• Languages: English
• ISSNs: 0360-8352

Abstract

Highlights: Bayesian network is introduced to tackle delays in train operations. Train operation data is used to learn the BN structures. The heuristic-based structure is elaborated as a prediction model. The hybrid BN model is tested using different performance measures. Abstract: We present a Bayesian network-(BN) based train delay prediction model to tackle the complexity and dependency nature of train operations. Three different BN schemes, namely, heuristic hill-climbing, primitive linear and hybrid structure, are investigated using real-world train operation data from a high-speed railway line. We first use historical data to rationalize the dependency graph of the developed structures. Each BN structure is then trained with the gold standard k -fold cross validation approach to avoid over-fitting and evaluate its performance against the others. Overall, the validation results indicate that a BN-based model can be an efficient tool for capturing superposition and interaction effects of train delays. However, a well-designed hybrid BN structure, developed based on domain knowledge and judgments of expertise and local authorities, can outperform the other models. We present a performance comparison of the predictions obtained from the hybrid BN structure against the real-world benchmark data. The results show that the proposed model on overage can achieve over 80% accuracy in predictions within a 60-min horizon, yielding low prediction errors regarding mean absolute error (MAE), mean error (ME) and root mean square error (RMSE) measures.