B0730
Title: Probabilistic modelling of passenger movements to predict onboard loads
Authors: Remi Coulaud - SNCF (France)
Christine Keribin - INRIA - Paris-Saclay University (France) [presenting]
Gilles Stoltz - CNRS (France)
Abstract: Transilien trains operate in the dense zone of Paris and its suburbs. Some of them are equipped with infrared sensors at their doors, allowing the measure of the number of boarding and alighting passengers at each door. These trains have communicating areas, i.e., once passengers have boarded, they can move along the coaches, so even with an error-free knowledge of boarding and alighting, it is not possible to deduce the load on board in each train area. However, this knowledge is crucial for smoothing passenger flows. A stationary model is first considered, where passengers move only depending on their boarding doors, independently of the station. Hence, a unique transition matrix modelling the displacement probability is estimated, either with least squares or maximum likelihood. This model is then refined by making the transition matrix depend on the station of boarding. This induces a complex situation due, on the one hand, to the number of latent variables representing intermediate onboard loads and, on the other hand, to the approximation of sums of multinomial distributions by simple distributions. Estimation is discussed using an EM algorithm or resorting to a neural network paradigm.
