Multi-Step Bidirectional LSTM for Low Frequent Bus Travel Time Prediction

Abstract

Accurate Bus Arrival Time (BAT) prediction is a measure of the quality of the public transport system. Intercity buses usually run for longer distances (e.g. 100 km+), and their frequency is lower compared to short-distance buses. It is essential to predict BAT accurately in order to improve the customer satisfaction of the passengers in the intermediate stops when the static schedules highly deviate from the displayed ones. We are introducing unidirectional and bidirectional multi-step LSTM Networks for link-based travel time prediction. We have derived two feature sets from the GPS data, weather data, and other augmented data considering the low frequency of the buses to test the models. To the best of our knowledge, this is the first work done to solve the BAT problem in Sri Lankan traffic conditions.