Convolutional neural network downscaling to improve sub-seasonal wind-speed predictions in Europe

Wind power systems' maintenance, deployment, and management, as well as the balance between energy supply and demand, are highly dependent on wind speeds and their temporal variability. Wind speed prediction at the sub-seasonal time scale presents a challenge since the skill of surface wind prediction sharply declines after two weeks. Nevertheless, the predictability of large-scale variables is higher than that of surface variables at this time scale. Goutham et al. (2022) improved the skill of surface wind speed forecasts by downscaling 500 hPa geopotential height (Z500)t forecasts using redundancy analysis based on a linear framework. Leveraging their work, we investigate whether Convolutional Neural Networks (CNNs) can be used to further improve the skill of subseasonal wind speed predictions over Europe.

 

To answer this question, this study proposes a non-linear statistical sub-seasonal ensemble forecasting method for the boreal winter over Europe based on applying a supervised learning model to dynamic forecasts. Specifically, the proposed statistical CNN model regresses weekly-mean surface wind speeds from dynamic forecasts of Z500. The dynamical forecasts of surface wind speed from the European Centre for Medium-Range Weather Forecasts (ECMWF) and the statistical forecasts from Redundancy Analysis (RA, Goutham et al. 2022) are selected as benchmarks for comparison, known for their superiority to climatology in different domains, lead times, and assessment indicators. Initially, we access the skill of a deterministic version of the network, utilizing ERA5 reanalysis data (trained for 15 years and tested for 5 years) to regress wind speed from geopotential height, find that the network is, on average, more skillful than RA based on the root mean squared error. Subsequently, the same network (without retraining) is applied to the subseasonal predictions from the ECMWF  covering the boreal winters from 2015 to 2022, with the same variables. Several indicators, including the anomaly correlation coefficient, continuous ranked probability score, and corresponding skill scores, compare the skill of the statistical forecasts (RA and CNN) and the dynamical forecasts (ECMWF). This comparison aims to determine if the statistical CNN has superior skill in the primary wind energy-producing countries in Europe and whether different models exhibit specific spatiotemporal patterns of skill in the sub-seasonal range. This study also investigates the performance of statistical and dynamical wind speed forecasts under normal and extreme conditions by analyzing the probability density distribution of ensemble members at given areas. Our preliminary findings reveal that statistical forecasts exhibit superior skill in normal