A Bayesian Spatiotemporal Approach for Bus Speed Modeling

Abstract

Bus speed modeling is essential for effective operation and management of public transit systems. Space-time interaction patterns are being ignored when modeling bus speed, and this would lead to biased statistical inferences. This paper proposed a spatiotemporal Bayesian model to characterize space-time interaction patterns among road segments using large-scale bus GPS data and to further develop the bus speed prediction model based on that. Results showed that a type II interaction pattern existed in the data, and the mean absolute percentage errors (MAPEs) of the test sets were 11.3% for the AM peak and 22.5% for the PM peak. Results were further compared with existing work. It was found that the proposed model presented a superior predictive performance while keeping the interpretability of contributing factors and space-time interaction patterns.