Short-term prediction of travel time is one of the central topics in contemporary ITS research and practice. Given the vast amount of prediction models it is a far from trivial task to select the most reliable and accurate prediction model for one particular scientific or commercial application. One possible way to deal with this is to develop a generic framework that can automatically combine multiple models running in parallel. Existing combination frameworks use the error in the previous time steps. In online applications however this method is not feasible, since travel times are available only after they are realized. This implies that errors on previous predictions are unknown. In this paper a Bayesian combination framework is proposed instead. The method assesses whether a model is likely to produce good results based on the current inputs in the light of the data with which it was calibrated. A powerful feature of this method is that it automatically balances between a good model fit and model complexity. Using two simple linear regression models as a showcase, we show this Bayesian combination improves prediction accuracy for real-time applications, but that the method is sensitive in case all models are biased in a similar way. It is therefore recommended to increase the number and the diversity of the prediction models to be combined.
Abstract