Bayesian inference for short term traffic forecasting

Abstract

In intelligent transport systems, short term traffic forecasting is one of the most important problems, reflecting the network state in the near future and feeding information to other application modules. Even though there have been quite a lot of works in this area, most of them are univariate models which may not be able to exploit the spatial relationship of traffic variables. So, this thesis explores the domain of two renowned modelling classes, the auto regressive moving average model and the dynamic model, taking into account the spatial dependency. The sparse-form vector autoregressive moving average model is applied to the short term traffic forecasting problem with different preprocessing methods. Network information is used to constrain the matrix parameters of the model, reducing the number of parameters. For the estimation problem, an improved MCMC method is proposed to tackle the variable correlation problem, using the marginalisation and the correlation direction information. Multi-step-ahead prediction results of different models are compared with two Dublin traffic datasets. A second model, consisting of four sub-models, targets the multi-step-ahead flow prediction for traffic data with incidents, where the data pattern may shift unexpectedly. The model is designed to satisfy the scalability property so that the inference of each component can be done conditionally independently. Furthermore, each submodel supports sequential inference, which is essential for real-time applications. The first two sub-models are analysed with a VISSIM dataset and the discussion of the last two is given at the end. A sequential approximation method is developed for both the state vector and the parameters of the dynamic model that is part of the second model. To avoid the degeneracy problem of the sampling-based particle filter, the method uses a continuous functional approximation which is a modified implementation of the iterated Laplace approximation (Bornkamp, 2011a). Both the modified iterated Laplace approximation and the sequential approximation method are illustrated and analysed with several examples.