Air Traffic Flow Representation and Prediction using Transformer in Flow-centric Airspace

The air traffic control paradigm is shifting from sector-based operations to cross-border flow-centric approaches to overcome sectors’ geographical limits. Under the flow-centric paradigm, prediction of the traffic flow at major flow intersections, defined as flow coordination points in this paper, may assist controllers in coordinating intersecting traffic flows which is the main challenge for implementing flow-centric concepts. This paper proposes to predict the flow at coordination points through a transformer neural network model. Firstly, the flow coordination points, i.e., the major flow intersections, are identified by hierarchical clustering of flight trajectory intersections whose location and connectivity characterize daily traffic flow patterns as a graph. The number of coordination points is optimized through graph analysis of the daily flow pattern evolution. Secondly, air traffic flow features in the airspace during a period are described as a “paragraph” whose “sentences” consist of the time and callsign sequences of flights transiting through the identified coordination points. Finally, a transformer neural network model is adopted to learn the sequential flow features and predict the future number of flights passing the coordination points. The proposed method is applied to French airspace based on one-month ADS-B data (from Dec 1, 2019, to Dec 31, 2019), including 158,856 flights. Results show that the proposed prediction model can approximate the actual flow values with a coefficient of determination (R2) between 0.909 to 0.99 and a mean absolute percentage error (MAPE) varying from 27.4% to 11.7% with respect to a 15-minute to 2-hour prediction window. The sustainability of the prediction accuracy under an increasing prediction window demonstrates the potential of the proposed model for longer-term flow prediction.