Adaptive Bayesian Network for Traffic Flow Prediction

Traffic control is essential for the achievement of a sustainable and safe mobility. Monitoring systems for traffic control collect a great amount of data that must be efficiently processed by estimation/prevision models to support operations of traffic management. In this paper we investigate a statistical method for traffic flow forecasting based on graphical modeling of the spatial-temporal evolution of flows. We propose an adaptive Bayesian network in which the network topology changes following the non-stationary characteristics of traffic. Two major stationary areas are recognized as principal phases of traffic flows. Graph optimization is implemented for each phase using mutual information as learning metric. Experimental tests on data provided by the PeMS project in the area of Los Angeles showed that the proposed method can reliably predict traffic dynamics.