Traffic flow data are now routinely collected for many roads. These data can be used as part of a traffic management system to assess highways facilities and performance over time, or for real-time traffic control to prevent and manage congestion. The data can also be used as part of a traveller information system. Good short-term traffic flow forecasting models of (multivariate) traffic flows are vital for the success of both traffic management and traveller information systems and Bayesian graphical dynamic models are proving to be very promising for this purpose.
Bayesian graphical dynamic models use the direction of traffic flow and the possible routes through a traffic network to represent the multivariate time series of traffic flows by a graph. This graph is then used to break the multivariate forecasting problem into much simpler univariate forecasting problems. In particular, the multivariate time series of traffic flows is modelled by a set of (conditional) Bayesian dynamic linear models. As such, the models are computationally straightforward to use in practice and established univariate dynamic linear modelling techniques can be used directly in the multivariate setting.
Catriona Queen leads the research on Bayesian graphical dynamic models for short-term traffic flow forecasting in the school.