From Numerical Weather Prediction Outputs to Accurate Local Surface Wind Speed: Statistical Modeling and Forecasts

Abstract

Downscaling a meteorological quantity at a specific location from outputs of Numerical Weather Prediction models is a vast field of research with continuous improvement. The need to provide accurate forecasts of the surface wind speed at specific locations of wind farms has become critical for wind energy application. While classical statistical methods like multiple linear regression have been often used in order to reconstruct wind speed from Numerical Weather Prediction model outputs, machine learning methods, like Random Forests, are not as widespread in this field of research. In this paper, we compare the performances of two downscaling statistical methods for reconstructing and forecasting wind speed at a specific location from the European Center of Medium-range Weather Forecasts (ECMWF) model outputs. The assessment of ECMWF shows for 10 m wind speed displays a systematic bias, while at 100 m, the wind speed is better represented. Our study shows that both classical and machine learning methods lead to comparable results. However, the time needed to pre-process and to calibrate the models is very different in both cases. The multiple linear model associated with a wise pre-processing and variable selection shows performances that are slightly better, compared to Random Forest models. Finally, we highlight the added value of using past observed local information for forecasting the wind speed on the short term.