Benchmarking Deep Learning based Downscaling of Wind Speed

The efficient placement of wind turbines relies on strategic assessment of local wind speed. Recent
studies highlight the crucial role of spatial resolution in accurately forecasting wind speed and
estimating the associated wind energy potential [1].

However, climate models typically fail to provide the spatial data resolution necessary for precise
energy resource assessment. To address this challenge, various downscaling methods have been
proposed to infer high-resolution data from coarser resolution data. Notably, image super-resolution
methods, a class of image processing techniques originally developed in computer vision to enhance
the resolution of natural images, have emerged as a promising approach for statistical downscaling.
By interpreting gridded data as images, these techniques are amenable to increasing the spatial resolution
of climate [3] and weather data [2].

We provide a comprehensive benchmark to compare the performance of various state-of-the-art image
superresolution models on weather data, such as ERA5 reanalysis data. The benchmark ranges from
interpolation baselines to all prominent deep learning based models, including a CNN-based model,
an attention-based model and a spatio-temporal model.

 

[1] Jung, C. and Schindler, D. [2022], ‘On the influence of wind speed model resolution on the global technical
wind energy potential’, Renewable and Sustainable Energy Reviews 156, 112001.
[2] Kurinchi-Vendhan, R., Lütjens, B., Gupta, R., Werner, L. and Newman, D. [2021], ‘Wisosuper: Bench-
marking super-resolution methods on wind and solar data’, arXiv preprint arXiv:2109.08770 .
[3] Stengel, K., Glaws, A., Hettinger, D. and King, R. N. [2020], ‘Adversarial super-resolution of climatological
wind and solar data’, Proceedings of the National Academy of Sciences 117(29), 16805–16815.