P4. Machine learning for aircraft trajectory prediction: a solution for pre-tactical air traffic flow and capacity management

The goal of air traffic flow and capacity management (ATFCM) is to ensure that airport and airspace capacity meet traffic demand while optimising traffic flows to avoid exceeding the available capacity when it cannot be further increased. In Europe, ATFCM is handled by EUROCONTROL, in its role of Network Manager (NM), and comprises three phases: strategic, pre-tactical, and tactical. This thesis is focused on the pre-tactical phase, which covers the six days prior to the day of operations.

During the pre-tactical phase, few or no flight plans (FPLs) have been filed by airspace users (AUs) and the only flight information available to the NM are the so-called flight intentions (FIs), consisting mainly of flight schedules. Trajectory information becomes available only when the AUs send their FPLs. This information is required to ensure a correct allocation of resources in coordination with air navigation service providers (ANSPs). To forecast FPLs before they are filed by the AUs, the NM relies on the PREDICT tool, which generates traffic forecasts for the whole European Civil Aviation Conference (ECAC) area according to the trajectories chosen by the same or similar flights in the recent past, without taking advantage of the information on AU choices encoded in historical data.

The goal of the present PhD thesis is to develop a solution for pre-tactical traffic forecast that improves the predictive performance of the PREDICT tool while being able to cope with the entire set of flights in the ECAC network in a computationally efficient manner. To this end, trajectory forecasting approaches based on machine learning models trained on historical data have been explored, evaluating their predictive performance.

In the application of machine learning techniques to demand trajectory prediction, three fundamental methodological choices have to be made: (i) approach to trajectory clustering, which is used to group similar trajectories in order to simplify the trajectory prediction problem; (ii) model formulation; and (iii) model training approach. The contribution of this PhD thesis to the state of the-art lies in the first two areas. First, we have developed a novel route clustering technique based on the area comprised between two routes that reduces the required computational time and increases the scalability with respect to other clustering techniques described in the literature. Second, we have developed, tested and evaluated two new modelling approaches for route prediction. The first approach consists in building and training an independent machine learning model for each origin-destination (OD) pair in the network, taking as inputs different variables available from FIs plus other variables related to weather and to the number of regulations. This approach improves the performance of the PREDICT model, but it also has an important limitation: it does not consider changes in the route availability, thus being unable to predict routes not available in the training data and sometimes predicting routes that are not compatible with the airspace structure. The second approach is an airline-based approach, which consists in building and training a model for each airline. The limitations of the first model are overcome by considering as input variables not only the variables available from the FIs and the weather, but also route availability and route characteristics (e.g., route cost, length, etc.).

The airline-based approach yields a significant improvement with respect to PREDICT and to the OD pair-based model, achieving a route prediction accuracy of 0.896 (versus PREDICT’s accuracy of 0.828), while being able to deal with the full ECAC network within reasonable computational time. These promising results encourage us to be optimistic about the future implementation of the proposed system.