Gaussian mixture variational autoencoder for runway configuration prediction in imbalanced data

Abstract

Runway configuration selection is a complex decision-making process that involves balancing the need to re-optimize traffic flow when configurations change against the risk of delays and operational inefficiencies when configurations are maintained. Additionally, the requirement for historical data, the evolving nature of decision-making policies, and the occurrence of infrequent configurations further complicate the prediction process. This paper proposes a Gaussian Mixture Supervised Variational Autoencoder, a deep probabilistic model designed to improve runway configuration prediction by addressing challenges such as data scarcity, temporal shifts, and imbalanced patterns. Experimental results using data from Changi Airport demonstrate that the proposed model significantly outperforms baseline models, including XGBoost and Random Forest, across different evaluation setups. Notably, the model achieves an F1-score of 68% in random-day split setup and 71% in weekly walk-forward setup. With a 30-min prediction horizon and updates every 15 min, the model offers a robust and efficient solution for accurate runway configuration management in real-world airport settings.