Using Machine-Learning Methods to Improve Surface Wind Speed from the Outputs of a Numerical Weather Prediction Model

The relationship between the wind speed derived from the outputs of a numerical-weather-prediction model and from observations is explored using statistical and machine-learning models. Eight years of wind-speed measurements at a height of 10 m (from 2010 to 2017) from 171 stations spread over mainland France and Corsica are used for reference. Operational analyses from the European Center for Medium Range Weather Forecasts (ECMWF) provide the model information not only on the surface flow, but on other aspects of the atmospheric state at the location (or above) each station. In a first step, a large number of explanatory variables are used as input to several models (linear regressions, k-nearest neighbours, random forests, and gradient boosting). The modelled wind speed in the ECMWF analyses, by itself, has root-mean-square errors over all stations distributed widely around a median of 1.42 m s-1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{-1}$$\end{document}. Using statistical post-processing and making use of a historical dataset for training, the median of the root-mean-square errors at all stations can be reduced down to 1.07 m s-1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{-1}$$\end{document} when modelled with linear regressions, and down to 0.94 m s-1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{-1}$$\end{document} with the machine-learning models (random forests or gradient boosting). Yet more significant decreases are found for coastal stations where the errors are largest. The random-forest models are further explored to reduce the list of explanatory variables: a list of 25 explanatory variables, mainly consisting of flow variables (wind speed, velocity components, horizontal gradients of geopotential on different isobaric surfaces, wind shear between 10 and 100 m) and including marginally some temperature variables, appears as a good compromise between performance and simplicity. Finally, as a preliminary test for further work, the relation thus captured between the model outputs and the observed wind speed at a given time is applied to forecasts of the numerical-weather-prediction model, for lead times up to 24 h. The machine-learning model is found to be essentially as relevant on the forecasts as it was on the analyses, encouraging further use and development of these approaches for local wind-speed forecasts.