Prediction of Flight Departure and Arrival Routes with Gradient Boosted Decision Trees

Accurate trajectory predictions are of paramount importance to obtain representative traffic demand figures, and therefore to apply effective and efficient air traffic flow and capacity management measures. The trajectory of an aircraft can be divided in (1) the departure phase from the origin, (2) the en-route phase, and (3) the arrival phase to the destination. For the departure and arrival phases, air traffic controllers must provide the clearance to execute one of the routes pre-defined by the local authorities. The route that is assigned to an aircraft depends on several factors, such as the city-pair, the runway configuration, and environmental restrictions (e.g., those applied to mitigate noise in the surrounding communities). This paper proposes a machine learning model that, trained with historical data, is able to capture the influence of these factors and accurately predict which departure and arrival routes will be executed by a particular fight well before take-off and landing, respectively. Promising results of a case study for Amsterdam-Schipol airport using traffic and weather data from 2019 show that the model is able predict departure routes with an accuracy of 0.93 four hours before take-off, and arrival routes with an accuracy of 0.87 five hours before landing. Furthermore, a comprehensive feature importance analysis reveals which are the most important factors that determine the departure and arrival routes of an aircraft, therefore allowing to interpret the predictions of the model.