Dynamic Bayesian networks with Gaussian mixture models for short-term passenger flow forecasting

A dynamic Bayesian network approach is proposed for short-term passenger flow forecasting. The graphical structure is based on the causal relationships between the flows and their spatiotemporal neighbourhood, and takes into account the transport service. In previous work, we described the local conditional distributions as linear Gaussians. In this paper, we extend the approach to Gaussian mixture models in order to better catch the nonlinear relationships between the variables. In the presence of incomplete data, the structure and the parameters are learned by the structural expectation-maximization (EM) algorithm, to which we add a new step for determining the optimal number of mixing components. The model is applied to the on-board passenger flows of Paris metro line 2 and outperforms the other testing methods.