Learning to Rank Flight Routes for Improved Air Traffic Demand Predictions

Effective air traffic flow and capacity management (ATFCM) relies on accurate predictions of both traffic demand and capacity. This paper focuses on the former because it presents the greatest uncertainty. Traffic demand predictions are typically based on flight plans submitted by airspace users to the Network Manager. However, to optimise their flights with the most up-to-date information, many users delay submitting their flight plans until just a few hours before departure. This delay leads to the implementation of ATFCM measures with incomplete traffic information. Currently, missing flight plans are estimated using the PREDICT system, which performs its task effectively. However, it operates based on straightforward rules and does not account for factors such as air traffic flow management regulations, convective weather activity, or the business strategies of airspace users. Prior efforts to improve PREDICT have utilised complex, city-pair-specific data-driven models that encountered practical constraints due to insufficient training data. Moreover, these models were only capable of predicting the most likely flight plans from those observed in the past, without ensuring their validity in the current environment. The main goal of this paper is to presents an alternative methodology, which consists of modelling the decision-making processes of flight dispatchers when submitting flight plans, leveraging historical data and learning-to-rank techniques. Preliminary results are presented, along with a discussion of key challenges encountered and lessons learned, offering insights for future research directions.