Empirical error correction and feature identification for long term wind resource assessment using support vector regression

Abstract

A technique that makes use of supervised machine learning (support vector regression) to correct systematic errors in long-term model wind time series is presented. The technique—Empirical Wind Output Correction (EWOC)—is computationally efficient and is particularly effective over complex terrain. In addition to error correction, EWOC acts as a non-linear feature identification tool by ranking the input site characteristics according to their ability to reduce the overall error. Independent regression functions are constructed for the mean and variance of the wind speed using an observational training data set comprising 109 onshore sites throughout Europe. Corrections are applied to Virtual Met Mast™ (VMM) corrected time series, hindcast Numerical Weather Prediction output (Euro4), and raw Modern-Era Retrospective analysis for Research and Applications and ECMWF Re-Analysis (ERA)-interim reanalysis data. Reductions of approximately 40% in the Mean Absolute Error of the mean and variance of the wind speed compared to the source data are obtained. The quality of fit of the EWOC corrected wind speed distributions is examined through their Kullback-Leibler Divergences (KLD). Reductions of approximately 40% and 60% in the KLD are achieved compared to VMM corrected and raw ERA data, respectively. Possible methods for improving and extending the technique are suggested, including ways of maximising the size and quality of the training data set using a limited set of wind speed observations.