In machine learning, computer algorithms attempt to automatically distill knowledge from data, so as to construct a model capable of making predictions from novel data in the future. However there is no universal best learning method – e.g. good performance of a particular machine learning method used for classification does not guarantee good performance on regression problems. With this in mind, we combine a variety of machine learning methods to predict wind power production. We use a combination of random forest (RF), support vector regression with evolutionary strategy (SVR-ES) and supervised kernel principal component analysis (SKPCA) to build a robust model for forecasting wind power.

RF involves a combination of many decision trees. It is an attractive approach as it can handle a large number of predictor variables without the need for deleting irrelevant/redundant variables; it is not prone to overfitting and it is relatively fast compared with other machine learning methods. Similarly, support vector machine (SVM) for regression (SVR) has been successful used in nonlinear regression problems. Typically it has three hyper-parameters to be tuned and for this task we use a simple evolutionary strategy called "uncorrelated mutation with p step size". The SKPCA method is similar to conventional principal components analysis (PCA) except that the predictors are transformed in advance by a kernel transformation and added to the PCA which uses a subset of the predictors selected based on their association with the outcome. Wind power is then regressed onto the leading principal components.