Enhancing Wind Downscaling with Foundation Models

Abstract

Downscaling wind data is crucial for various applications such as climate modeling, weather forecasting, and renewable energy planning. However, dynamical downscaling is expensive to run, limiting the resolution, coverage, lead-time and number of ensemble members that can be produced. For example, ECCC’s High Resolution Deterministic Prediction System (HRDPS) covers Canada at 2.5 km resolution with a lead-time of up to 48 hours, while ECCC’s Global Deterministic System (GDPS) has a lead-time of up to 10 days for the whole globe, but at reduced resolution of 15 km. Recently published statistical downscaling methods employ super-resolution machine learning models that were originally applied to image super-resolution for the downscaling problem. For instance, Annau et al. (AIES 2023) investigate the use of generative adversarial networks (GANs) for wind downscaling. Interestingly, they utilize non-idealized low and high-resolution data for model training to alleviate shared-scale mismatches due to internal variability. However, some aspects of the physical processes in wind forecast may be lost due to the data-driven nature of machine learning methods. In this work, we explore how AI Foundation Models can improve downscaling under the hypothesis that these very large models can better encode physical aspects of weather phenomena. Specifically, we take the weather foundation model Prithvi-WxC as the backbone for various fine-tuning experiments of GANs. We utilize GDPS and HRDPS outputs as the training data, exploring a set of dynamic low-resolution predictors from variables both near the surface and at various pressure levels, as well as static high-resolution covariates such as topography. We validate the downscaled forecasts at different lead times against a mesoscale analysis and wind observations at stations using standard verification metrics. We also provide an analysis of the power spectrum and the statistical distribution of wind forecasts and inter-compare with HRDPS outputs and Annau’s approach.