Graph-augmented temporal forecasting framework for airport traffic prediction
Abstract
Accurate short-term forecasting of airport traffic remains challenging due to the highly non-stationary and network-dependent characteristics of air transportation systems. This study proposes a graph-augmented temporal forecasting framework that integrates engineered temporal features (lags, rolling means, and calendar attributes) with Node2Vec-based airport connectivity representations. The approach combines the interpretability and robustness of ensemble learners with the spatial awareness offered by graph embeddings to predict hourly aircraft movements at King Khalid International Airport (RUH). A comprehensive experimental evaluation across multiple model families, including ensemble learners (Random Forest, XGBoost), neural networks (MLP), and statistical baselines (ARIMA, Ridge Regression), demonstrates the superiority of ensemble-based architectures. The two best-performing models were XGBoost and Random Forest. XGBoost achieved MAE = 0.833 and RMSE = 1.182, while Random Forest delivered a slightly lower MAE of 0.775 with RMSE = 1.190. These results highlight the strong generalization capabilities of ensemble learners and the added value of graph-informed features in capturing both temporal dynamics and network-driven variations. Overall, the proposed hybrid framework provides an accurate, robust, and interpretable forecasting solution suitable for real-time deployment within intelligent airport management and decision support systems.