Wind Power Prediction with Machine Learning Ensembles

For a sustainable integration of wind power into the electricity grid, precise and robust predictions are required. With increasing installed capacity and changing energy markets, there is a growing demand for short-term predictions. Machine learning methods can be used as a purely data-driven, spatio-temporal prediction model that yields better results than traditional physical models based on weather simulations. However, there are two big challenges when applying machine learning techniques to the domain of wind power predictions. First, when applying state-of-the-art algorithms to big training data sets, the required computation times may increase to an unacceptable level. Second, the prediction performance and reliability have to be improved to cope with the requirements of the energy markets. This thesis proposes a robust and practical prediction framework based on heterogeneous machine learning ensembles. Ensemble models combine the predictions of numerous and preferably diverse models to reduce the prediction error. First, homogeneous ensemble regressors that employ a single base algorithm are analyzed. Further, the construction of heterogeneous ensembles is proposed. These models employ multiple base algorithms and benefit from a gain of diversity among the combined predictors. A comprehensive experimental evaluation shows that the combination of different techniques to an ensemble outperforms state-ofthe-art prediction models while requiring a shorter runtime. Finally, a framework for model selection based on evolutionary multi-objective optimization is presented. The method offers an efficient and comfortable balancing of a preferably low prediction error and a moderate computational cost.