Enhancing Air Traffic Management: A Latent Representation Learning Approach

Efficient management of the National Airspace System (NAS) relies heavily on strategic Traffic Management Initiatives (TMIs), such as Ground Delay Programs (GDPs) and Ground Stops (GSs), to balance fluctuating demand with constrained airport capacity during adverse conditions. This paper introduces a representation learning approach designed to assist air traffic managers in identifying and retrieving similar historical TMI events. Using autoencoder-based latent feature extraction, our method jointly optimizes predictive accuracy of TMI triggers while providing meaningful, interpretable representations clustered in latent embeddings. Evaluations conducted using data from three major airports in the New York Metroplex—JFK, EWR, and LGA—demonstrate improved scenario clustering compared to traditional clustering and dimensionality reduction techniques. Our methodology facilitates scenario-driven recommendations, enhancing decision-making precision and interpretability in TMI implementation.