Downscaling wind fields using deep learning

Description

Deep learning methods such as Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs) can potentially greatly accelerate the production of near- surface wind fields while limiting the amount of computational and time intensive dynam- ical downscaling. Such methods have recently achieved great success in the field of image super-resolution (SR), a problem closely related to downscaling. This thesis aims to quantify the applicability such methods for wind field downscaling in the purpose of pre- liminary wind resource assessment, typically associated with extensive regional weather model simulations. Advanced SRGAN and SRCNN models are trained to estimate plau- sible HR wind velocity values conditioned on given low-resolution (LR) wind fields and additional high-resolution (HR) topographic data. Mesoscale wind fields from the New European Wind Atlas (NEWA) project are used as ground truth HR targets and LR input data is obtained by averaging the true HR wind fields to mimic the resolution of the forcing ERA5 reanalysis data. The trained super-resolution is thus 10 x, operating on scales from roughly ∼30 km to 3 km. This idealized approach aims to quantify model errors associated with unresolved atmospheric scales and terrain effects, circumventing issues related to the projection mismatch between global domain and regional numerical weather simula- tions. An extensive analyses of the reconstruction errors shows that systematic recon- struction errors are associated with prominent topographic features. When provided, the models are able to effectively use HR topographic information to reduce these systematic errors leading to a near zero wind speed magnitude bias while angular deviations persist in complex terrain.

Period	Jan 2022 → Jun 2022
Examinee	Gísli Björn Helgason
Examination held at	Department of Wind and Energy Systems
Degree of Recognition	International