Wind Field Nowcasting and Forecasting using Denoising Diffusion Probabilistic Models with Aircraft-Derived Data
M. Slobbe (TU Delft - Aerospace Engineering)
J. Sun – Mentor (TU Delft - Operations & Environment)
Evert Westerveld – Mentor (Air Traffic Control The Netherlands)
J.M. Hoekstra – Graduation committee member (TU Delft - Operations & Environment)
Alessandro Bombelli – Graduation committee member (TU Delft - Operations & Environment)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
As with many aspects of modern life, wind nowcasting and forecasting are integral parts of aviation and Air Traffic Management (ATM). This study investigates the use of a Denoising Diffusion Probabilistic Model (DDPM) for nowcasting (inpainting) and forecasting (image-to-video) of wind fields using aircraft-derived meteorological data. The DDPM, implemented with a U-Net backbone, demonstrated strong performance in nowcasting tasks, outperforming previous models such as the Lagrangian transportation-based Meteo-Particle (MP) model and
Physically Inspired Neural Network (PINN) approach with a 29% improvement in magnitude error and a 62% reduction in directional error. The nowcasting model achieved a magnitude
error of 2.03 m/s and a directional error of 4.2°, based on 190 test samples from late 2024. A key contribution lies in the DDPMs ability to produce more consistent and lower-variance predictions than prior methods. The RMSE improved on the PINN results by 29%, to 3.99 m/s. Despite these successes, forecasting proved significantly more challenging, with no meaningful results achieved. The study used ECMWF CERRA reanalysis data for training
and evaluated model performance with simulated aircraft tracks on known wind fields and with real aircraft-derived data from the The Royal Netherlands Meteorological Institute (KNMI)’s EMADDC dataset split into model input and validation subsets. High computational demands restricted testing capabilities, and uncertainty quantification and severe weather conditions remain challenging for the model.