Deep Learning-Based Spatiotemporal Prediction of Public Transport Ridership: Integrating Network-Constrained Isochrone Features

Abstract

Accurate station-level public transport ridership prediction is crucial for urban planning, yet it is often hindered by unrealistic spatial representations and noisy data. This study aims to predict Bus and Rail ridership volumes in Singapore using a spatiotemporal deep learning approach. The novelty of this research lies in the integration of Network Service Area (NSA) or isochrone-based Point-ofInterest (POI) features to capture realistic pedestrian accessibility, replacing conventional circular buffer methods. A TabNet model was developed using a twostage hyperparameter optimization strategy via Optuna (TPE) and Ensemble Learning within a 5-Fold Expanding Window Cross-Validation framework. Evaluation results demonstrate that the Tuned Extreme model significantly outperforms baseline models, achieving R2 scores exceeding 0.98 for both modes. Specifically, the Rail model exhibited highly stable performance (sMAPE ~11- 13%), whereas the Bus model faced higher volatility challenges (sMAPE ~19-22%) due to stochastic travel characteristics in industrial and peripheral zones. Interpretability analysis confirmed that while temporal lag features are the dominant predictors, the integration of isochrone features particularly residential density for Rail In and commercial amenities for Rail Out plays a vital role in capturing spatial demand variations. This research offers a robust framework for transport capacity management in complex urban networks.