A combined multivariate model for wind power prediction

Highlights: • A combined multivariate model is proposed for improving prediction accuracy. • Two-stages approach is used to reflect variables' influence on wind process. • More information is expressed by reconstructed data with a small dimension. • SVR models based on one variable are built with different kernel functions. • Data mining algorithms are used to combine results of univariate models. - Abstract: The intermittent and fluctuation of wind power has a harmful effect on power grid. To direct system operators to mitigate the harm, a combined multivariate model is proposed to improve wind power prediction accuracy. This model is built through two stages. First, valid meteorological variables for prediction are selected by Granger causality testing approach, and reconstructed in homeomorphic phase spaces. Then each variable is taken to build a wind power prediction model independently, and their effect on prediction is illustrated through different kernel functions in support vector regression models. Second, prediction results of univariate models are taken as inputs of a combined model predicting wind power. The final model is multivariate and expressive to reflect the interactive effects of selected meteorological variables on wind power prediction. Four data mining algorithms are trained for selecting the model with high accuracy. The industrial data from wind farms is taken as the study case. Prediction of models at two stages are tested, and performance of the proposed model is validated better at four error metrics.