Quantifying and Improving Bus Delay Prediction Accuracy in Sydney

Abstract

The reliability of transit operations is a crucial factor in a user’s mode choice, as it affects their ability to make an accurate travel plan and arrive at their destination on time. Transit reliability, however, is not just about schedule adherence. Users can anticipate and adjust to regular deviations from a schedule and thus, for a user, transit reliability is often about schedule variance rather than absolute delay. With the advent of real-time data being made available to the public, users can make last-minute changes to their schedule based off the real-time delay reported in the transit network. These last-minute changes are dependent on the accuracy of the delay predictions, however, and so this thesis seeks to understand the accuracy of these predictions by quantifying the difference between them and the true delay. After quantifying this, the thesis proposes a series of post-processing additions to our current delay prediction models in order to improve the prediction accuracy. It was found that substantial improvements in delay prediction performance could be achieved, particularly when the buses were over 20 minutes away from a stop. Additionally, it was noted that this improvement varied depending on the error quantification method that was used, with the MAPE (Mean absolute percentage error) showing the greatest improvement. The thesis recommends the adoption of some of the improved prediction models and also discusses the implications of more transparently conveying the uncertainty in the prediction to users.