Deep reinforcement learning-based air traffic flow coordination in flow-centric airspace

Air traffic flow coordination to avoid congestion at major flow intersections is a key enabler for the flow-centric airspace concept. This paper addresses the problem of air traffic flow coordination at major flow intersections by presenting a comprehensive solution encompassing flow identification, prediction, and re-routing at the Nominal Flow Intersections (NFIs). To identify the NFIs, a graph-based flow-pattern consistency approach is proposed to model and analyze daily air traffic flow patterns. With the identified NFIs, a transformer encoder-based neural network is adopted to learn the relations among the flow of flights at the NFIs to predict future demand. Finally, to avoid the foreseen demand exceeding the flow limit and reduce the congestion at NFIs, a reinforcement learning-based flow re-routing agent is designed and trained to dynamically assign alternative routes to air traffic flows based on the evolving flow states. The agent’s performance is quantified by the congestion reduction in the flows, quantified by the flight travel time. The proposed model is trained and tested using ADS-B data for December 2019 for two major en-route flows in the French airspace. The average travel time in each major flow is 30 min. Results show that, compared with the originally planned flows which have exceeded the flow limit, the per-flight travel time in the two flows is reduced by 3.34 min (11.1%) and 1.96 min (6.5%) through flow re-routing. Moreover, the overall travel time for flights around the two major flows (due to re-routing) is reduced by 1.45 min and 1.04 min respectively.