Abstract
Precise wind power output predictions is of great importance for a successful integration of wind power into the power grid and market operations. A popular approach to short-term wind power prediction is to use machine learning methods. Machine learning methods are data-drove, and the data it utilize as input is significant to prediction performance. There exits many proposed models in literature for short-term wind power prediction, but a variation in choice of input is apparent. Particularly a difference in the use of forecast weather data from numerical weather predictions models is observed. In this thesis a wide variety of input types is examined by measuring their effect on prediction accuracy using a state of the art method for short-term wind power prediction. The inputs examined stems from an in-depth review of previous work on the subject. In addition, some input types not commonly utilized is considered. In order to give a general account on the importance of the specific feature types, experiments have been done for two wind turbines in separate wind parks located in Scandinavia using real world data measurements and correlating weather forecast data from a publicly available numerical weather prediction model. Further, in addressing the challenge of battling computation costs, this work compares two strategies for predicting the wind power output for the whole wind park. In the first, and expensive strategy, separate models are developed for each turbine, and the final wind park power prediction is obtained by summarizing the individual predictions from each turbine. The second strategy treats the wind park as a single entity by aggregating the input data from all turbines before model development.