Multi-Route Aircraft Trajectory Prediction Using Temporal Fusion Transformers

Trajectory prediction plays a key role in modern air traffic management. The ability to predict the future position of aircraft in flight allows for greater predictability, safety and efficiency. In recent years, recurrent neural networks, and particularly LSTM (Long-Short Term Memory), have been successfully applied (alone or in combination with other kinds of network) to this problem. However, the criticality of the supervision of these operations and the difficulty of predicting trajectories in high density traffic zones, such the Terminal Area around the airports, require high accuracy methods that takes into account all factors involved in these operations. In this paper, we propose an architecture based on Temporal Fusion Transformer (TFT) for multi-route, long-term trajectory prediction using surveillance data (Automatic Dependent Surveillance - Broadcast, ADS-B). We conduct our experiments on the case study of the Madrid Barajas-Adolfo Suárez airport (Spain), using nine months worth of data. In particular, we focus on predicting the next 150 seconds at any point in the trajectory for flights arriving at this airport. Compared with other LSTM networks developed in this work, TFT provides an increased accuracy for 2D positioning, with mean absolute errors of 0.0091 and 0.0104 degrees for latitude and longitude, respectively, in the Terminal Area of the destination airport. These results have been shown to be competitive with, or even superior to, more consolidated approaches based on LSTM networks that have been proposed for single route, short-term predictions.