Data-driven Aircraft Trajectory Predictions using Ensemble Meta-Estimators

Aircraft trajectory prediction is nowadays a core task in the development and implementation of new concepts and tools within the Air Traffic Management and Control environment. Traditionally, this problem has been tackled by the usage of sophisticated model-based approaches that do not consider tactical situations that greatly influence the actual evolution of the aircraft's trajectory. Currently, the focus has shifted towards the application of data-driven methods, which enable the adoption of these factors thanks to learning algorithms trained with recorded trajectory information. In this paper, the aircraft trajectory prediction problem is formulated as a regression problem to be solved by state-of-the-art ensemble machine learning techniques. The selected algorithms have been trained using reconstructed trajectory datasets derived from actual surveillance recorded data. Once trained, to compute a trajectory prediction, they are fed only by information available prior to the flight departure (i.e. filed Flight Plans and weather forecasts). Then, the predictions issued by the different families of ensemble meta-estimators are compared to weigh their suitability and accuracy as data-driven trajectory predictors. The main results show that the data-driven predictors presented in this paper are potentially able to estimate parameters at a designated trajectory points such as the Estimated Time of Arrival within an average error of ≈ 10 sec, which represents extraordinary results for ATM purposes, and the aircraft mass within an average range of ≈75 kg, thus possibly enabling very highly accurate future environmental impact assessments.