Wind Resource Assessment with Microscale Models and a Machine Learning Method

Abstract

Wind energy has been gaining more prominence among renewable energy sources, as it is an affordable and increasingly reliable technology. The precision in the evaluation of the wind resource is, of course, a fundamental factor to guarantee the continuous development of these types of projects. As installed capacity increases, it is natural that the new wind farms increasingly have to be installed on more complex terrain. Therefore the methodologies that have traditionally been used to predict mean wind speed will be subject to greater uncertainty, given the limitations of the models under these challenging conditions. A more demanding energy industry requires further investigation of reliable and robust methodologies to assess available resources accurately. In this master thesis, two approaches to predicting average wind speed in complex terrain were evaluated. These approaches were wind flow models and statistical methods. Regarding the wind flow models, one year of on-site measurements was used to validate two well-known microscale models, the Wind Atlas Analysis and Application Program (WAsP) and the WindSim model. The performance of each model was evaluated by using a crossprediction methodology. The second approach corresponds to a machine learning method called k-Nearest neighbor (k-NN) regression. As its name implies, measurements from neighboring sites were used to predict the mean speed at a target site. Terrain and climatic features were used as predictors in the method mentioned above. By using the statistical method, the prediction errors were reduced to 1.29%. Further improvements in the accuracy were achieved by implementing a weight-based ensemble model between the WAsP model and the k-NN regression, with an overall percentage error of 1.06% compared with the 5.09% and 4.31% obtained with the WAsP model and the WindSim model, respectively.