An Improved Machine Learning Method to Speed up the Trajectory Prediction: Taking Melbourne Airport as a Study Case

The safety and efficiency of airspace operations largely depend on the accurate prediction of 4D trajectories in dense air traffic. Traditional methods are progressively giving way to more accurate machine learning (ML) techniques, among which the Long Short-Term Memory (LSTM) neural network emerges as an exceptionally promising tool and has been successfully applied, especially for time-series prediction tasks. In this study, we introduce an LSTM-based adjustable interpolation algorithm designed to significantly reduce computational time while maintaining accuracy at an acceptable level to meet operational constraints. The algorithm applies adjustable time intervals to input data based on ascent and descent rates, providing different data densities for different flight phases. A case study focusing on flight trajectories from Melbourne to Sydney is conducted, and the findings reveal that our proposed method can reduce computation time by half without significantly sacrificing prediction accuracy compared to the traditional linear interpolation method. Furthermore, it achieves accuracy improvements of at least 50% compared to raw data processing, with no substantial increase in computational time. Proven to be effective, our proposed algorithm can be an ideal solution for training dense air traffic data when regular training and high accuracy is required. This includes applications in Urban Air Mobility (UAM) and unmanned aircraft operations, as well as airport management and airspace sector handovers.