Abstract Accurate flight trajectory prediction and reliable arrival time estimation are critical for enhancing the safety and efficiency of modern air traffic management, where minor deviations can lead to significant operational disruptions and safety hazards. Existing data-driven prediction methods typically omit explicit operational constraints, causing systematic deviations from prescribed airways and progressive error growth over extended forecast horizons.This paper presents FlightDiff, a novel diffusion-based framework that integrates dual-constraint guidance to generate physically coherent and operationally compliant flight trajectories. By incorporating flight plan waypoints as local constraints, FlightDiff ensures precise adherence to regulated airspace, while a global destination constraint maintains alignment with the overall flight objective. The framework employs a two-phase diffusion process: the first phase enforces both local and global constraints for physical plausibility, and the second phase refines predictions using a single constraint to balance accuracy and flexibility. A shared Variational Autoencoder (VAE) effectively encodes both dense flight trajectories and sparse waypoint data into a unified latent representation, significantly enhancing spatiotemporal coherence while reducing computational overhead required by prior work. Extensive evaluations on large-scale, real-world flight datasets from Qingdao, China (May–July 2024) demonstrate that FlightDiff outperforms state-of-the-art methods in trajectory forecasting and arrival time estimation across various prediction horizons, demonstrating FlightDiff’s transformative potential for next-generation data-driven air traffic management systems. In addition, the ablation experiments demonstrate that the shared VAE, the dual-constraint mechanism, and the two-phase denoising schedule each deliver complementary, incremental performance gains.
He et al. (Fri,) studied this question.