PETA: Combining Machine Learning Models to Improve Estimated Time of Arrival Predictions

All aviation stakeholders require accurate estimated times of arrival in order to run flight operations as efficiently as possible. The time of arrival, however, is difficult to predict because it is affected by the uncertainties of the previous flight phases, with take-off time variability being the most significant contributor. At present, estimated time of arrival predictions are computed by the Enhanced Traffic Flow Management System, which collects data from a variety of sources to provide the best estimate throughout the entire duration of the flight. This paper introduces a novel approach that leverages existing machine learning models to enhance the accuracy of estimated time of arrival predictions, also during the pre-departure phase. More specifically, the first model (FADE) forecasts the evolution of air traffic flow management delays for regulated flights; the second model (KNOCK-ON) anticipates rotational reactionary delays arising from unrealistic available turn-around times; and the third model was trained to identify systematic discrepancies between reported and actual airborne times. Using a dataset comprised of historical traffic and meteorological data collected from March to June 2023, this paper presents a comprehensive evaluation of this ensemble of models, referred to as PETA, against the current predictions across various time horizons, ranging from 6 hours before departure to the moment of take-off. The results indicate that the proposed solution surpasses the existing system in approximately two-thirds of the predictions. When the proposed solution performs better, the average and median improvements are 14 minutes and 7 minutes respectively. However, when it underperforms, the average and median deteriorations are 7 minutes and 4 minutes respectively.