Real-Time Prediction of Go-Around Probability During Final Approach with Deep Learning

This paper presents a real-time system that esti- mates the probability of a go-around (GA) during final approach using deep learning. Historical approach data from Barcelona–El Prat and Madrid–Barajas — including alignment with the runway, range and descent geometry, aircraft energy state, and in-trail spacing — are fused with meteorological variables from Open-Meteo to train three neural architectures: a Long Short- Term Memory (LSTM) network, an attention-augmented LSTM, and a Transformer. Performance is assessed with distance-to- ideal, temporal accuracy, area under the accuracy curve, and confusion-matrix metrics. Model explainability is provided via SHAP. The Transformer consistently delivers the best results, achieving >80% accuracy up to two minutes before a GA and ∼87% at 30 seconds, and has been integrated into a real-time prototype (LandIA) designed to support air traffic controllers.