Forecasting traffic flows in road networks: A graphical dynamic model approach

Congestion on roads is a major problem worldwide. Many roads now have induction loops implanted into the road surface providing real-time traffic flow data. These data can be used in a traffic management system to monitor current traffic flows in a network so that traffic can be directed and managed efficiently. Reliable short-term forecasting and monitoring models of traffic flows are crucial for the success of any traffic management system. Traffic flow data are invariably multivariate so that the flows of traffic upstream and downstream of a particular data collection site S in the network are very informative about the flows at site S. Despite this, most of the shortterm forecasting models of traffic flows are univariate and consider the flow at site S in isolation. In this paper we use a Bayesian graphical dynamic model called the Linear Multiregression Dynamic Model (LMDM) for forecasting traffic flow. An LMDM is a multivariate model which uses a graph in which the nodes represent time series of flows at the various data collection sites, and the links between nodes represent the conditional independence and causal structure between flows at different sites. All computation in LMDMs is performed locally, so that model computation is always simple, even for arbitrarily complex road networks. This allows the model to work in real-time, as required by any traffic management system. LMDMs are also non-stationary and can readily accommodate changes in traffic flows. This is an essential property for any model for use with traffic management systems where series often exhibit temporary changes due to congestion or accidents, for example. Finally, LMDMs are often easily interpretable by non-statisticians, making them easy-to-use and understand. The paper will focus on the problem of forecasting traffic flows in two separate motorway networks in the UK.