Generation of Optimized Trajectories for Congestion Mitigation in Fukuoka Approach Control Area Using Deep Reinforcement Learning

Fukuoka airport is the busiest single runway airport in Japan and the congestion has been increasing year by year. The COVID-19 pandemic has temporarily eased the congestion, but it is expected to increase again after the pandemic. Excessive radar vector by air traffic controllers to maintain aircraft separation during congestion causes economic and environmental losses due to increased flight distances and times. We attempted to generate optimal trajectories in the approach control area of Fukuoka airport using a deep reinforcement learning method based on a centralized Deep Q Network (DQN). Wind information was considered in the trajectory optimization using Mesoscale Model (MSM) data from the Japan Meteorological Agency. As a result, the optimized trajectory was found to be useful because a radar vector reduction of 23.7% in distance and 33.6% in flight time were attained compared with the recorded radar data provided as CARATS open data. The trajectories have the characteristics that all of the Ground speed (GS) of aircraft are made slower while directed straight to the intermediate fix (IF) after entering the approach control area.