A dynamic Bayesian network approach to forecast short-term urban rail passenger flows with incomplete data

We propose an approach to forecast the short-term passenger flows of the urban rail network of Paris. Based on dynamic Bayesian networks, this approach is designed to perform even in case of incomplete data. The structure of the model is built so that the flows are predicted from their spatiotemporal neighbourhood, while the local conditional distributions are described as linear Gaussians. A first application carried out on a single station highlights the need to incorporate information on the transport service. In the presence of missing data, we perform the structural expectation-maximization (EM) algorithm to learn both the structure and the parameters of the model. Short-term forecasting is conducted by inference via the bootstrap filter. Finally, we apply the model to an entire Paris metro line, using on-board counting, ticket validation and transport service data. The overall forecasting results outperform the historical average and last observation carried forward (LOCF) methods. They also evidence the key role of the transport service in the modeling.