Probabilistic Flight ETA prediction

This proposal builds on a separate project: Physics-Based Flight Trajectory Optimization in Japanese Airspace, by addressing the challenge of predicting aircraft Estimated Time of Arrival (ETA) and flight trajectories under uncertainty by leveraging advanced probabilistic and machine learning methods. The motivation stems from the limitations of traditional point-estimate approaches, which do not adequately capture the inherent uncertainties in flight operations, especially due to variable weather conditions and pilot actions. The key innovation proposed is to provide real-time probabilistic predictions of ETAs at each waypoint, quantifying uncertainty without assuming Gaussian noise, and improving estimation performance by integrating aircraft dynamics, weather impacts, and historical trajectory data.

The methodology involves modeling aircraft dynamics and observations using a state-space approach, where the evolution of aircraft states is influenced by control inputs, wind, and noise. The training process leverages historical trajectory data to quantify model uncertainties, using generative diffusion models and deep neural networks to estimate the distribution of future aircraft positions. During inference, the trained model generates samples of possible future trajectories, which are then mapped to ETA distributions at each waypoint. This approach allows for dynamic updating of ETA uncertainty as new observations become available, providing more robust and informative predictions for air traffic management and operational decision-making.