Predict ATFCM weather regulations using a time-distributed recurrent neural network

Abstract

In recent years, prior to COVID-19, capacity shortfalls in airspace and airports inevitably caused an increase in aircraft delays. Therefore, when it returns to normal conditions, the airspace will exhibit the same capacity limits, even under normal weather conditions. To ensure that air traffic remains safe, reliable, and efficient in adverse weather conditions, planning and coordination activities through a Collaborative Decision Making process are required to deliver the most effective Air Traffic Flow and Capacity Management services to Air Traffic Control and Aircraft Operators. Nowadays, this task is based on air traffic controllers’ experience and historical data. That means that the Flow Manager Positions and the Network Manager operators have to process a huge amount of information, and the detection of future overloads is based on past experiences. Moreover, due to the inherent uncertainty of weather information, a reliable decision support framework is required to handle these situations as efficiently as possible. We propose a Deep Learning model able to extract the relationship between both the historical data and the implemented actions, accurately identifying the intervals of time that must be regulated. The proposed model achieves an accuracy between 80% and 90% across six traffic volumes belonging to both the MUAC and REIMS regions, a recall higher than 85%, and an F1-score higher than 0.8 in all the cases. Furthermore, the confidence-level analysis shows a really high activation when making a prediction. Finally, the SHapley Additive exPlanations method is applied to identify the most relevant input features.