Multilevel graph neural networks for bikeshare demand modeling using station and route features

Bikesharing systems are a key component of sustainable urban mobility, offering low-carbon and space-efficient alternatives to motorized travel. However, accurately forecasting demand remains challenging in cities with complex topography and heterogeneous infrastructure, where both station characteristics and route conditions influence ridership. This study proposes a multilevel spatiotemporal graph neural network that captures both temporal demand patterns and route-level context to improve predictive accuracy and interpretability. The framework represents the system using two complementary graphs, including a station-level graph capturing temporal demand correlations and a path-level graph modeling candidate routes enriched with features such as slope, traffic stress, land use, crime exposure, and cycling infrastructure. Using data from Pittsburgh’s POGOH system, the model outperforms attention-based, gated, and state-of-the-art benchmarks across multiple metrics. Beyond prediction, the framework supports more efficient operations and informs infrastructure planning, offering a scalable and interpretable approach for resilient and low-carbon bikeshare systems.