Accurate predictions of aircraft Estimated Time of Arrival (ETA) are key enablers for air traffic controller Decision Support Tools (DSTs) aimed at decreasing delays, fuel burn and emissions. In addition, reliable aircraft arrival time predictions, when displayed on sequence timelines, provide valuable situational awareness to controllers. In this paper, we present a machine learning approach for estimation of aircraft landing times using tree-based ensemble methods, namely Random Forests (RF) and Quantile Regression Forests (QRF). These tree-based ensemble methods have high performance and are suitable for real-time applications, provide robust and accurate solutions in the presence of uncertainties in the input parameters, and allow for interpretability of results through confidence intervals tied to each individual prediction. Predictions made by RF and QRF are based on a selection of variables describing each individual aircraft and its spatiotemporal track, such as weight class and speed, as well as along-track weather and airport congestion. Unlike existing models, the RF and QRF models make no a priori assumptions about the functional relationship between the predictor variables and the response ETA. Rather, these models “discover” the appropriate (nonlinear) response to the given input based on historical data. A by-product of these ensemble approaches is a measure of the relative importance of the individual predictor variables. These properties of RF and QRF are quite attractive for the ETA prediction problem, given the large number of predictor variables and the complex (unknown) inter-relationships between them. Our approach has been tested for arrivals at Dallas/Fort Worth International Airport. High prediction accuracies are obtained across a range of days encompassing a variety of operational conditions. For example, in our preliminary experiments, we are able to predict ETA with absolute errors of less than 1 minute for 54% of flights immediately outside the terminal airspace (60-80NM); for distances of 20-35NM, 68% of flights are predicted to within 1 minute, and for 10-25NM, 78% of predictions are accurate to 1 minute. The widths of 90% confidence intervals for individual predictions decrease considerably as distance to the runway decreases, demonstrating increased certainty in ETA predictions as flights approach the wheels-on event.

This work was sponsored by the National Aeronautics and Space Administration (NASA) under Air Force Contract FA8721-05-C-0002. Opinions, interpretations, conclusions, and recommendations are those of the authors and are not necessarily endorsed by the United States Government.