Diffusion-Based, Data-Assimilation-Enabled Super-Resolution of Hub-height Winds

TL;DR

WindSR addresses the challenge of generating high-resolution hub-height wind fields from sparse observations by coupling denoising diffusion probabilistic models with data assimilation. The method conditions the diffusion process on terrain and a dynamically blended field that fuses sparse HRRR observations with coarse WTK simulations, using a dynamic impact radius $d$ to modulate observation influence. Key contributions include a DA-conditioned diffusion framework with terrain-aware conditioning, a dynamic radius assimilation mechanism, and thorough comparisons against CNN/GAN baselines showing improved accuracy and roughly 20% bias reduction relative to independent observations. The approach provides a scalable, physics-aware pathway to improve wind-resource assessments and gust-risk evaluations for wind farms and critical infrastructure. The framework points toward future extensions incorporating temporal conditioning and additional auxiliary data (e.g., land use, temperature) to further enhance fidelity.

Abstract

High-quality observations of hub-height winds are valuable but sparse in space and time. Simulations are widely available on regular grids but are generally biased and too coarse to inform wind-farm siting or to assess extreme-weather-related risks (e.g., gusts) at infrastructure scales. To fully utilize both data types for generating high-quality, high-resolution hub-height wind speeds (tens to ~100m above ground), this study introduces WindSR, a diffusion model with data assimilation for super-resolution downscaling of hub-height winds. WindSR integrates sparse observational data with simulation fields during downscaling using state-of-the-art diffusion models. A dynamic-radius blending method is introduced to merge observations with simulations, providing conditioning for the diffusion process. Terrain information is incorporated during both training and inference to account for its role as a key driver of winds. Evaluated against convolutional-neural-network and generative-adversarial-network baselines, WindSR outperforms them in both downscaling efficiency and accuracy. Our data assimilation reduces WindSR's model bias by approximately 20% relative to independent observations.

Figures (7)

• Figure 1: The figure demonstrates the DDPM’s forward process of adding noise and the reverse denoising process.
• Figure 2: Correlation coefficient (left) and root mean square error (right) between WTK and HRRR hourly data averaged over 5 days over each 128x128 tile. Despite advances in NWP, significant model biases remain, especially at coarser resolutions and in areas where complex terrain makes parametrization difficult. Data driven approaches can rectify these differences without the need for computationally expensive physics.
• Figure 3: Image-generation workflow combining DA and SR to downscale windspeed. Sparse observations are interpolated to the inference grid and softly blended via a mask, then inpainted into the simulation filed to form a composite that conditions the diffusion model during generation; terrain information is included as an additional conditioning input during the reverse process.
• Figure 4: Cumulative distribution functions of wind speeds across different terrain conditions of the US, comparing values recorded by HRRR, WTK, and our customized SR with and without DA. In both flat and mountainous regions, DA brings the distribution of wind speeds closer to HRRR data which is used to emulate observations.
• Figure 5: Visual comparison of different SR models