A Comparative Analysis of Train Delay Prediction Models for Markov Chains

Delay is a service quality measure in urban and intercity rail systems, reducing their operational efficiency and reliability. Accurate estimation of train delays is one of the main challenges in railway operation due to complex interrelations among the operating elements and inherent uncertainty associated with their behavior. Markov chains have recently been used, in this respect, to model variability in departure and arrival delays of trains in running and dwelling processes. In this study, we measured the performance of delay-based homogeneous and non-homogeneous Markov models and benchmarked them with artificial neural networks, support vector machine, and random forest models. The results show that Markov models can make comparable predictions. Moreover, due to their interpretability and transparency, the Markov chains allows us to gain insights into delay transitions in various processes and to make statistical inferences, such moment estimations and occurrence probabilities for delays. Model comparisons also imply that the process of train running and dwelling is memoryless.