Data-Driven Metro Train Crowding Prediction Based on Real-Time Load Data

This paper formulates the car-specific metro train crowding prediction problem based on real-time load data and evaluates the performance of several data-driven prediction methods (lasso, stepwise regression, and boosted tree ensembles). Two variants of the prediction problem are considered: 1) train-centered prediction, focusing on in-vehicle crowding information provision and 2) station-centered prediction focusing on crowding information provision at stations. The methodology is applied to a metro line in Stockholm, Sweden. The prediction accuracy is evaluated with respect to absolute passenger loads and discrete crowding levels. When available, predictions with real-time load data significantly outperform historical averages, with accuracy improvements varying in magnitude across target stations and train cars depending on load variability. The results suggest that the real-time crowding information can be provided sufficiently early to influence travelers' route, train, and car choices, in order to reduce in-vehicle crowding.