Mapping wind speed distribution across large regions using machine learning and asymmetric kernel estimators.

Wind resource assessment studies over large regions provide the basis for the preliminary identification of locations with promising wind energy prospects. In past studies, several authors have mapped the mean wind speed across large regions using spatial interpolation methods or machine learning models. In recent studies, more emphasis has been placed on mapping the entire wind speed distribution to evaluate the wind resource variability at unsampled locations. Most of these studies have assumed that the wind speed distribution across the entire region belongs to a single family of probability distribution functions and then processed to map the distribution parameters. A flexible non-parametric approach for wind speed distribution mapping is proposed in this study. The new approach is based on mapping various wind speed quantiles at some fixed percentile points in the region using a machine learning model. Then, at any unsampled location, these quantiles are used as input of an asymmetric kernel estimator of cumulative distribution function to recover the whole wind speed distribution. Asymmetric kernel estimators solve the probability leakage problem that appears when fitting symmetric kernels to bounded variables such as wind speed. The non-parametric approach for wind speed distribution mapping was more effective than a traditional approach based on mapping the parameters of a distribution function. In the best scenario, an improvement was observed between 6% (test samples) and 9% (cross-validation) of the Kolmogorov-Smirnov statistic between the observed and estimated wind speed distribution. The non-parametric approach is recommended for regions with highly variable wind regimes that cannot be captured by a single family of distribution functions.