Forecasting train arrival delays on the Ankara – Eskişehir high-speed line in Turkey

Railway operations may experience delays due to technical issues or weather conditions. Accurate prediction of such delays can enhance the quality of rail transport services and the effectiveness of railway operations. The study has developed the arrival delay prediction model using random forest regression based on the train operation data from the Ankara - Eskişehir high-speed train line in Turkey. The model can simultaneously predict arrival delays at all downstream stations on this line and continuously update these predictions as new information about train movements becomes available. The accuracy rates of the model vary from 76% to 99% under a 1-min prediction error. The results show that incorporating variables related to weather conditions and technical problems related to train control systems into the model improves prediction performance. The contribution of these variables to the model performance increases as the prediction horizon widens. The model results suggest that the model predictions may assist network managers in making better decisions about train operations. In order to evaluate the model's performance from the passengers' point of view, the study has proposed two methods: the proportion of late predictions and the stability of forecasts. The findings indicate that most trains (between 96.7% and 99%) have stable arrival delay predictions at target stations. The proportion of 2-min (or greater) late predictions, which means that the predicted delay exceeds the actual delay by 2 min or more, fluctuates from 14% to 0.5%, depending on the prediction horizon. Although the ratio for the short horizons (one station ahead) becomes relatively low, it is necessary to be cautious when using the model predictions to inform passengers because a prediction of more than 1 min late for short horizons might have negative consequences (e.g., misleading passengers to leave stations).