Graph-Based Multi-Horizon Forecasting of Airport Delay Propagation in the U.S. Air Transportation Network
Existing airport delay prediction models often focus on isolated flight-level or airport-level forecasts and do not explicitly represent the spatial propagation of disruptions across a national air transportation network. This study develops and evaluates a spatio-temporal graph learning framework for multi-horizon airport-level delay forecasting in the U.S. domestic network. Using 2,878,854 flight records from January 2019 to August 2023, flights are aggregated into hourly node-time tensors over 370 airports and 5334 directed connections. The proposed framework combines graph-based airport connectivity, temporal delay persistence, rare-disruption detection, and conformal uncertainty estimation over a 1–3 h forecasting horizon. The results show that LSTM attains the lowest aligned point-forecast errors, while ST-GCN provides competitive multi-horizon performance with substantially lower parameter complexity and explicit graph-aware outputs; the direct XGBoost baselines provide nonlinear tabular benchmarks but do not dominate the final multi-horizon evaluation. These findings indicate that graph-based models are most useful not as universal point-forecast winners but as interpretable network-aware tools for monitoring delay propagation, disruption risk, and operational uncertainty.