Improving arrival time prediction of Thailand's passenger trains using historical travel times

Abstract

The State Railway of Thailand provides passengers with train location information on their Web site, which includes the name of the last station that each train arrives at or departs from, along with the timestamps and the accumulative train delay (in minutes) from the train timetable. This information allows passengers to intuitively predict the arrival time at their station by adding the last known train delay to the scheduled arrival time. This paper aims at providing a more accurate prediction of passenger train's arrival times using the historical travel times between train stations. Two algorithms that use train location information and historical travel times are proposed and evaluated. The first algorithm uses the moving average of historical travel times. The second algorithm utilizes the travel times of the k-nearest neighbors (k-NN) of the last known arrival time. To evaluate the proposed algorithms, we collected six months of data for three different trains and calculated prediction errors using mean absolute error (MAE). The prediction errors of the proposed algorithms are compared to the prediction errors of the baseline algorithm that predicts the arrival time by adding the last known train delay to the scheduled train arrival time. Both algorithms outperform the baseline prediction. The algorithm based on moving average travel time improves the prediction error by 22.9 percent on average, and the algorithm based on k-NN improves the prediction error by 23.0 percent on average (k=16).