Time in Advance Go-Around Predictions for Decision Support in Air Traffic Management

Maintaining the safety and cost-effectiveness of air transport operations, while increasing capacity, will push the next generation of ATM systems towards digitization. In the medium term, a digitized system in the human-managed ATM environment will be able to provide reliable predictive analytics based on automated information processing, thus providing decision support to human operators. This paper details the development of a machine learning based solution for go-around predictions, exemplified at two major European airports. Go-arounds are standard, high workload, procedures by which an aircraft in the final approach phase can safely discontinue the approach. The proposed machine learning solution is aimed at increasing safety levels in airport operations by enhancing air traffic controllers situational awareness and helping them better plan and adapt to go-around scenarios. This work leverages on more than two years worth of ADS-B and successfully uses oversampling technique to combat the high imbalanced in the data. In addition, we performed a benchmark study with a selection of the most common classification models. The final type of model selected was LightGBM for which a feasibility study for predictions at 2NM, 4NM, 6NM, and 8NM distance from the runway threshold was performed. The results for both airports showed that although the models’ recall decreases with the distance from the threshold, we were able to maintain a high prediction precision between 90% at 2NM to 80% at 8NM. Finally, a study of the explainability of the probabilistic predictions was carried out by evaluating the most important features of the models.