GooDFlight: Goal-Oriented Diffusion Model for Flight Trajectory Prediction

Flight trajectory prediction is an essential task in the air traffic control field. Previous approaches to this problem usually follow a single-stage or short-term intention-guided prediction paradigm, which suffers from problems such as insufficient trajectory prediction diversity, limited accuracy, and interpretability. Different from existing paradigms, in this article, we present GooDFlight—A goal-oriented diffusion model for flight trajectory prediction. GooDFlight is a long-term intention-guided, diversity-emphasizing framework that decouples the flight trajectory prediction process into two stages: goal estimation and trajectory prediction. In the first stage, we propose a one-then-all goal estimation method to sufficiently cover the macro-uncertainty in flight patterns and then tailor the interaction-aware joint goal distribution, which extends the flight intention from a single, deterministic ground truth to the empirical intention distribution from the similar experience. In the second stage, we employ a transformer-based diffusion model to generate stochastic flight trajectories conditioned on the intention estimations, modeling the micro-uncertainty in flight operations under each pattern estimated in stage one. In terms of evaluation metrics, existing metrics have difficulties in accurately reflecting the model's ability to handle the natural uncertainty of trajectories. We further propose a simple yet effective global-local endpoints variance (GLeV) metric for evaluating the diversity of predicted trajectories under social acceptance. Our proposed method is validated in-depth on TrajAir, a large-scale dataset collected from the real-world air traffic control environment at the Pittsburgh-Butler Regional Airport, a nontowered general aviation airport. The experimental results demonstrate that the proposed method significantly outperforms other methods in terms of both accuracy and diversity with superior interpretability.