Probabilistic Prediction of Traffic States Using Bayesian Network

Abstract

Traffic state prediction is an important issue in traffic operations. One of the main purposes of traffic operations is to prevent flow breakdown. Therefore, it is necessary to perform traffic state predictions that reflects the stochastic process of traffic flow. However, traffic state transition is affected complexly and simultaneously by many factors, which lead to a lack of understanding and accurate prediction. Meanwhile, the Bayesian network is a methodology that not only is suitable for a problem with uncertainty but also can improve the understanding of a problem. Also, it is possible to derive fair probability with incomplete information, which allows the analysis of various situations. In this study, we developed a traffic state prediction model using the Bayesian network to reflect dynamic and stochastic traffic flow characteristics. In order to improve the structure of the Bayesian network, which has been used simply in transportation problems, we proposed a modeling procedure using mixture of Gaussians (MOGs). Also, spatially extended variables were used to consider the spatiotemporal evolution of traffic flow pattern. In particular, traffic state identification was performed by estimating the link speed in order to consider the spatial propagation of congestion. In the performance evaluation, the Bayesian network has better performance than logistic regression and has the same level of performance as artificial neural network based on machine learning. Also, by performing sensitivity analyses, we provided the understanding of traffic state prediction and the guidelines for model improvement. Therefore, the Bayesian network developed in this study can be considered as a traffic state prediction model with good prediction accuracy and provides insights for traffic state prediction.