Development of a daily gridded wind speed observation product using artificial intelligence in Spain

Historical near-surface wind speed (NSWS; ~10 m above the ground) measurements from terrestrial weather stations are crucial for assessing NSWS changes and variability and its implications for various socioeconomic and environmental issues, such as wind energy. However, currently there is no all-Spain gridded NSWS observation product with higher spatial coverage than station-based wind series. A new methodological approach based on image reconstruction using artificial intelligence could help to solve this limitation. We use a partial convolutional neural network (PCNN) and station-based NSWS series from the Spanish Meteorological Agency (AEMET) to create a 0.1º daily gridded wind speed observation product over Spain for 1961-2021. The deep neural network is trained with wind data from the ERA5-Land reanalysis (at 9-km grid-spacing, ECMWF), and a mask where grid points with historical wind observations are identified. Thus, the 0.1º resolution wind distribution grid is treated as the pixel values of an image with the masked grid points being pixels to be reconstructed. The training process allows the PCNN model to learn the physical laws, such as momentum conservation, present as internal relationships between pixels in the reanalysis data. The learned laws were then implemented to estimate the wind speed of the masked grid points. During the training process, the PCNN model predictions are iteratively compared to the reanalysis data and improved according to the error (i.e., MAE or RMSE) between prediction and the original reanalysis data. Once trained, the model is applied to NSWS measurements in the target domain to predict wind at locations with no observations. The gridded NSWS product provides a high-resolution wind speed data for whole Spain that respects the available observations and reliably predict wind speed in unsampled places, which is useful to many applications requiring wind information.