Mathematical Modeling of Transit Network Spatiotemporal Dynamics: A Data-Efficient Graph-BiLSTM Architecture for Segment-Level Travel Time Forecasting

This study addresses the challenge of accurate short-term bus travel time prediction in data-constrained urban environments. While many existing models rely on extensive external data, we propose a data-efficient spatiotemporal framework that integrates sequential deep learning with multi-relational graph modeling using only standard operational transit data. The approach is implemented through a multi-relational graph spatio-temporal bidirectional LSTM (MRG-ST-BiLSTM) model, which captures temporal dependencies while preserving the physical topology of the road network. A real-world case study was conducted using GTFS-based data from bus operations in the Greater Sydney Area, covering five routes and more than four million stop-level records. In the real road network test, numerical experiments demonstrated that the proposed model achieved improved prediction accuracy compared to baseline methods. Specifically, it achieved an MAE of 15.82 s and RMSE of 31.38 s, outperforming the conventional LSTM (MAE 26.70, RMSE 45.61), Hybrid-BiLSTM (MAE 17.24, RMSE 32.73) and GCN-LSTM (MAE 17.27, RMSE 32.61). The proposed framework provides a scalable and interpretable practical solution for transit agencies.