Optimisation of Airspace Sectorisation based on Open Data and Reinforcement Learning

Abstract

In response to the growing demand for air travel and limited airspace capacity, the aviation industry is seeking to optimise airspace sector design, targeting the reduction of Air Traffic Controller (ATCo) workload - a significant bottleneck in airspace capacity. However, re-designing airspace sectorisation is complicated due to the high-dimensional search space, sector interdependencies, dynamic air traffic, and data accessibility. As a machine learning approach, Reinforcement Learning (RL) algorithm is able to adaptively learn optimal sector configurations from the complex air traffic environment, and thus holds promise in addressing these challenges. Hence, this project proposed a method that utilises an RL algorithm in conjunction with Voronoi diagram, a widely used method to produce convex polygons, to generate newsector centroids and their corresponding boundaries. To this effect, we developed a learning-based algorithm that divides airspace into sectors. Proximal Policy Optimisation(PPO) was adopted to address continuous state and action space. We trained a PPO agent to optimise sector center locations and integrated a Voronoi diagram to fulfill the sectorisation solution. The proposed methodology was applied to the daily operations of ATM system within the UK airspace. OpenSky API is an important tool used in this project to retrieve historical real-life flight data. To better understand the effectiveness of our proposed method, we compared it with approaches in two other studies focusing on ATCo workload reduction and balance: both of them employed Voronoi partitioning and non-learning based algorithms. Results show that our proposed approach leads to a notable reduction in ATCo workload and a simplification in the configuration of airspace sectorisation simultaneously. Workload reduction of up to 40% is achievable with appropriate hyperparameter values, and the reduction can be even more significant with extended training time. The training process duration varies depending on learning time, but upon completion, it can consistently provide real-time sectorisation with reduced ATCo workload. This advantage holds promising implications for dynamic airspace sectorisation.