A Seasonal Modelling Approach Capturing Spatio-Temporal Correlations for Dynamic Bus Travel Time Prediction

The accuracy of arrival time information provided to passengers plays a vital role in the success of developing mobility solutions. For better prediction accuracy, the prediction method should be able to capture both temporal and spatial variations. Approaches capturing spatio-temporal dependencies in bus travel time are limited in literature. The present study proposes a novel approach, where a seasonal time series model that can capture spatio-temporal correlations in the travel-time data. In particular, the study rearrange the data intelligently in such a way that the one-dimensional seasonal auto-regressive model captures both the spatial and temporal correlations in the data. We further augment the predictive model by using additional observations optimally chosen from the historic data. To make dynamic real-time prediction feasible using the above seasonal model, we recast the predictive model as a Linear Dynamical System (LDS) in state-space form. Particle Filter is used for dynamic prediction of the bus travel times in subsequent sections (relative to current bus position). Results obtained clearly indicate superior performance of the proposed approach in comparison to a diverse set of current approaches.