Passenger flow forecasting on transportation network: sensitivity analysis of the spatiotemporal features

Predicting the crowding level in train stations or the passenger load in trains can be useful to enrich the information available to passengers and improve train regulation processes or service quality levels. The main issue to handle when forecasting passenger flows is the structural variability of the related time series induced by the irregularity of train schedule and the influence of several contextual factors, such as calendar information and the characteristics of the served station. Forecasts depend on different contextual variables that generally have a spatial component, a temporal component, or both. We study the sensitivity of the spatiotemporal features of machine learning forecast models. Our main goal is to understand how the spatiotemporal features affect the performance of the models. First, we propose to study the impact of spatial and temporal inputs such as the served station, the train route or direction, and the type of day on the forecasting results to set up the best way to build a set of machine learning models to predict the passenger load of trains. Second, we address the effect of the temporal aggregation level on model performances for the forecasting task. The proposed models are based on ensemble machine learning approaches and have been deployed on a line of the Paris greater area railway network. A fine-grained evaluation is conducted as a support of the model’s sensitivity analysis.