DST-TransitNet: A Dynamic Spatio-Temporal Model for Robust Station-Level Transit Ridership Prediction

Accurate short-term prediction of station-level transit ridership is crucial for effective transit operations, especially in dynamic urban environments. Traditional methods (e.g., ARIMA, SARIMA) fail to capture intricate spatial-temporal correlations and handle dynamic ridership fluctuations. This study introduces DST-TransitNet, a hybrid deep learning model integrating Graph Attention Networks (GAT) and Gated Recurrent Units (GRU) to model spatial-temporal dependencies among transit stations dynamically. A time-series decomposition framework further improves prediction accuracy and interpretability. Evaluated on Bogotá’s BRT system under normal operations, social disruptions, and pandemic conditions, DST-TransitNet achieved up to 94% R² and reduced prediction errors (MAAPE) by approximately 10%-15% compared to state-of-the-art benchmarks. These results demonstrate DST-TransitNet’s superior accuracy, robustness, and scalability across diverse prediction scenarios and changing time intervals.