Prediction of the estimated times of arrival of freight train based on operational and geospatial features

Abstract

In many railway systems, freight train schedules are often adjusted based on passenger train timetables at the operational level. Predicting the estimated time of arrival (ETA) for freight trains is a challenging task due to the high variability in transit times. This study introduces ETA prediction models developed using two years of operational data combined with geospatial features for freight trains operating within a sub-network of the Iranian railway. Prediction models for all origin-destination pairs in each direction (north-to-south and south-to-north) were created, predicting ETAs at three distinct locations along the routes. Four machine learning algorithms were evaluated, and the most accurate model was determined through comparisons with a baseline statistical model. The random forest algorithm demonstrated superior performance among the models at most locations. The performance improvements of the best prediction models with and without geospatial features were also investigated. Models incorporating geospatial features showed notably higher accuracy than those relying solely on non-geospatial predictors. These improvements were particularly more evident in the south-to-north direction and at locations closer to the destination. The results of this research offer practical insights for logistics centers, enabling optimized loading, unloading, and resource allocation strategies, thereby enhancing the efficiency of freight railway operations.