Feature Selection for Prediction of Railway Disruption Length

Abstract

Disruptions such as accidents, facility failures and power outage occur frequently during daily railway operation, which can disrupt the railway timetable and cause financial cost in terms of lost patronage. Recent study indicates that skilled dispatchers can estimate the disruption length based on the analysis of related features. To select the significant features of railway disruption length (RDL), in this study, bidirectional gated recurrent unit with focal loss (FL-BiGRU) and maximal information coefficient (MIC) were applied to analyze the Chinese railway accident reports. FL-BiGRU was proposed to classify the disruption causes into seven categories and MIC algorithm was applied to select the related RDL features. Experiment results showed that the FL-BiGRU model outperforms rest models and can provide good accuracy especially on categories with fewer samples. Moreover, the set of predictors built by the maximum information coefficient can significantly improve the accuracy of the prediction model.