Flight data clustering for offline evaluation of real-time trajectory optimization framework

Abstract

Real-time Trajectory Optimization (RTOP) is a machine learning-based flight path optimization framework that automatically performs in-flight re-planning continuously with the latest weather information. The framework leverages three core algorithms: (1) a supervised learning algorithm to predict winds, (2) an unsupervised learning algorithm to forecast areas of convective weather, and (3) a graph-based pathfinding algorithm to generate optimized flight trajectories. This research first proposes a methodology that uses an unsupervised learning-based clustering algorithm with the concept of Hausdorff distance to appropriately select representative flight cases to build up confidence in statistical analysis. Second, this research conducts an operational evaluation of the RTOP framework with the selected representative flight cases to highlight the statistical significance of simulation results and indicate the expected savings. The results show that optimized flight trajectories generated by the RTOP framework reduce the duration of the en-route phase up to 2% compared to the real-world flight trajectories in most cases.