Enhancing operational wind downscaling capabilities over Canada: Application of a Conditional Wasserstein GAN methodology

TL;DR

This work tackles extending high-resolution wind forecasts over Canada from 48 hours to 10 days by downscaling GDPS outputs to HRDPS resolution. It advances the DownGAN framework by employing a Conditional WGAN-GP with a UNET generator conditioned on high-resolution static covariates such as topography, land-water mask, and roughness length, enabling large-domain wind downscaling. A Frequency Separation strategy further refines the high-frequency content, reducing artifacts and improving spectral fidelity. Across independent test data covering the Canadian domain, the proposed approach yields lower RMSE and LSD compared with baselines, demonstrating operationally scalable wind downscaling with potential to bridge the 10-day low-resolution global forecast window.

Abstract

Wind downscaling is essential for improving the spatial resolution of weather forecasts, particularly in operational Numerical Weather Prediction (NWP). This study advances wind downscaling by extending the DownGAN framework introduced by Annau et al.,to operational datasets from the Global Deterministic Prediction System (GDPS) and High-Resolution Deterministic Prediction System (HRDPS), covering the entire Canadian domain. We enhance the model by incorporating high-resolution static covariates, such as HRDPS-derived topography, into a Conditional Wasserstein Generative Adversarial Network with Gradient Penalty, implemented using a UNET-based generator. Following the DownGAN framework, our methodology integrates low-resolution GDPS forecasts (15 km, 10-day horizon) and high-resolution HRDPS forecasts (2.5 km, 48-hour horizon) with Frequency Separation techniques adapted from computer vision. Through robust training and inference over the Canadian region, we demonstrate the operational scalability of our approach, achieving significant improvements in wind downscaling accuracy. Statistical validation highlights reductions in root mean square error (RMSE) and log spectral distance (LSD) metrics compared to the original DownGAN. High-resolution conditioning covariates and Frequency Separation strategies prove instrumental in enhancing model performance. This work underscores the potential for extending high-resolution wind forecasts beyond the 48-hour horizon, bridging the gap to the 10-day low resolution global forecast window.

Figures (9)

• Figure 1: Example of zonal and meridional winds on the GDPS and HRDPS grids. Note that the GDPS grid was cropped over North America. The domain of HRDPS grid is indicated by a black rectangle.
• Figure 2: Schema of the Critic Network. The input consists of either a block of high-resolution NWP predictands or downscaled data generated by the Generator. The output is a real-valued score representing the Critic's confidence that the input originates from actual NWP data—a higher score indicates a more physically realistic appearance. The labels $k$, $n$ and $s$ below the blocks denote the kernel size, the number of filters, and the stride used in each layer, respectively. The numbers above the blocks represent the tensor dimensions as a function of the number of feature maps, width, and height.
• Figure 3: Schematic of the Generator Network. The inputs consist of static high-resolution covariates (top-left) and low-resolution predictors (buttom-left). The static covariates go through three successive coarsening passes and are then concatenated with the low-resolution predictors. The resulting combination then goes sixteen times through the core component called "residual in residual dense block". Three refinement passes with skip connections followed by a last group of convolution yield the final high-resolution output.
• Figure 4: Training and validation curves: DownGAN vs conditional WGAN-GP. Conditional WGAN-GP outperforms DownGAN in terms of the validation MSE curve and achieves faster convergence during training.
• Figure 5: Effect of Frequency separation in the validation curve. Frequency separation with filter sizes of (5,5), (9,9) or (13,13) has a positive effect on the MSE validation score (0.86, 0.87 and 0.9, respectively), but partial frequency separation does not improve the scores or only slightly (MSE validation score of 0.93) for filter size of (5,5).