Comparison of several measure-correlate-predict models using support vector regression techniques to estimate wind power densities. A case study

Abstract

The long-term annual mean wind power density (WPD) is an important indicator of wind as a power source which is usually included in regional wind resource maps as useful prior information to identify potentially attractive sites for the installation of wind projects. In this paper, a comparison is made of eight proposed Measure-Correlate-Predict (MCP) models to estimate the WPDs at a target site. Seven of these models use the Support Vector Regression (SVR) and the eighth the Multiple Linear Regression (MLR) technique, which serves as a basis to compare the performance of the other models. In addition, a wrapper technique with 10-fold cross-validation has been used to select the optimal set of input features for the SVR and MLR models. Some of the eight models were trained to directly estimate the mean hourly WPDs at a target site. Others, however, were firstly trained to estimate the parameters on which the WPD depends (i.e. wind speed and air density) and then, using these parameters, the target site mean hourly WPDs. The explanatory features considered are different combinations of the mean hourly wind speeds, wind directions and air densities recorded in 2014 at ten weather stations in the Canary Archipelago (Spain).The conclusions that can be drawn from the study undertaken include the argument that the most accurate method for the long-term estimation of WPDs requires the execution of a specially trained model which considers the variability of the wind speeds of the reference stations, as well as of the wind directions and air densities, and in addition the functional manner in which these variables participate in the proposed MCP models. It is also concluded that it is important to consider the annual variation of air density even in regions at sea level. It is further concluded that, of the eight MCP models under comparison, the one that predicts the WPDs based on two sub-models (which estimate the wind speeds and air densities in an unlinked manner) always provides the best MAE (Mean Absolute Error), MARE (Mean Absolute Relative Error) and R2 (Coefficient of determination) metrics, with the differences being statistically significant (5% significance) for most of the cases assessed. Additionally, the regulatory capacity of the SVR technique was sufficient to manage most of the overfitting problems, and hence the contribution of the wrapper method was not relevant in our study.