Retrospective de-trending of wind site turbulence using machine learning

Abstract

This paper considers the removal of low-frequency trend contributions from turbulence intensity values at sites for which only 10-min statistics in wind speed are available. It is proposed the problem be reformulated as a direct regression task, solvable using machine learning techniques in conjunction with training data formed from measurements at sites for which underlying (non-averaged) wind data are available. Once trained, the machine learning models can de-trend sites for which only 10-min statistics have been retained. A range of machine learning techniques are tested, for cases of linear and filtered approaches to de-trending, using data from 14 sites. Results indicate this approach allows for excellent approximation of de-trended turbulence intensity distributions at unobserved sites, providing significant improvements over the existing recommended method. The best results were obtained using Neural Network, Random Forest and Boosted Tree models.