Abstract. The increase of renewable based generation as alternative power source brings an added uncertainty to power systems. The intermittent nature of renewable resources, such as wind speed and solar intensity, requires the use of adequate forecast methodologies to support the management and integration of this type of energy resources. This paper proposes a clustering methodology to group historic data according to the data correlation and relevance for different contexts of use. Using the clustering process as a data filter only the most adequate data is used for the training process of forecasting methodologies. Artificial Neural Networks and Support Vector Machines are used to test and compare the quality of forecasts when using the proposed methodology to select the training data. Data from the Brazilian city of Florianópolis, Santa Catarina, has been used, including solar irradiance components and other meteorological variables, e.g. temperature, wind speed and humidity. Experimental findings show that using the proposed method to filter data used for training ANN and SVM achieved promising results, outperforming the approaches without clustering.