Machine learning based trajectory prediction to support demand forecasting

Abstract

Air traffic sector demand and capacity balancing is an important process to enable safe and efficient flight execution. In current operations, demand and capacity are determined based on schedules and flight plans. In reality, disruptions to flights create a different situation that may not have been anticipated by the Air Navigation Service Provider (ANSP). Wrong demand forecasts may cause unnecessary network regulations or inefficient flight execution. This research aims to improve air traffic sector demand forecasting, by exploring machine learning based trajectory prediction. In light of the Trajectory Based Operations (TBO) concept that is being developed within Air Traffic Management (ATM) research, a trajectory-based approach is taken to improve demand forecasts. To achieve this, the transformer neural network was identified as a suitable generative model that can predict aircraft trajectories. Using available traffic messages from the Eurocontrol Business-to-Business (B2B) connection, and actual trajectories obtained from the OpenSky ADS-B repository, a successful transformer neural network was built. This trajectory predictor could accurately generate trajectories, outperforming the flight plan and other neural network approaches by a large margin. For demand prediction, the introduction of improved trajectories provided small gains that could potentially lead to more stable predictions.