In recent years, the share of solar power in total energy production has gained a rapidincrease. Therefore, prediction of solar power production has become increasingly important for energy regulations. In this study we proposed an ensemble method which givescompetitive prediction performance for solar power production. This method firstly decomposes the nonlinear power production data into components with a multi-scale decomposition technique such as Empirical Mode Decomposition (EMD). These componentsare then enriched with the explanatory exogenous feature set. Finally, each componentis separately modeled by nonlinear machine learning methods and their results are aggregated as final prediction. We use two different training approaches such as Hour-basedand Day-based, for predicting the power production at each hour in a day. Experimentalresults show that our ensemble method with Hour-based approach outperform the examined machine learning methods.