Bus Arrival Time Modeling Based on Auckland Data

Abstract

Inaccurate bus arrival time predictions are counterproductive to changing transport habits and promoting public transport use. This research sought to improve the bus passenger experience in terms of bus arrival time prediction by investigating various time series and regression-based techniques suitable for bus arrival time modeling. The models developed in the current study included: random walk with drift, multivariate linear regression, decision tree, artificial neural networks, and gene expression programming models. Historic automatic vehicle location and passenger flow data obtained for four bus routes spanning Auckland city, in both travel directions, were used as model inputs. Specifically, 10 independent variables were incorporated in the regression models, with distance between bus stops being the most significant predictor for bus travel time. Research results indicated that time series models outperformed regression techniques, with the time series artificial neural network being the most successful of the seven models developed. Moreover, the alternative models all performed significantly better than the prediction engine currently utilized by an Auckland bus company for arrival time prediction. However, these results require corroboration with manually collected field data, on account of the quality concerns afflicting the raw data reported by the transport company.