Bus Journey Time Prediction: A Comparison of Whole Route and Segment Journey Time Predictions Using Machine Learning

Accurately predicted bus journey times are essential for bus network reliability and making bus transport attractive. The most common approach when predicting bus journey times with machine learning (ML) is to predict journey times for each stop pair segment. Segment data can be very noisy, leading to inaccuracies. To investigate this, this paper compares the classic stop pair segment approach to three other methods. Firstly, a naive method of calculated historical averages is introduced as a baseline. We then explore two methods based on predicting the whole bus route journey time from origin to terminus. To estimate a passenger’s journey, where the whole route is not travelled, we estimate the proportion of the whole journey time the passenger’s journey will take. The first of these methods calculates this proportion from similar historical journeys, and the second proposed method trains an ML model to predict this proportion for each segment of the passenger’s journey. The results show that this novel proposed approach results in less error across most metrics, when compared to the segment prediction method. An interesting insight from the analysis shows the proposed approach has enhanced benefits during peak travel time and during the working week. Gains in prediction accuracy at these times would benefit the most commuters. This research can be applied to make robust scheduling decisions that will increase bus network reliability, improve bus network satisfaction and uptake, and lead to more sustainable cities.