Forecasting of wind power production in the Netherlands THESIS

Wind power has become an important source of power for some countries because wind is renewable, wind power is clean and no pollutants are produced compared to fossil fuels which are mainly used for the generation of energy today. Because of these reasons also in the Netherlands attention towards the use of wind power has grown. In the past decade, a lot of research has been performed on the forecasting of wind power production over a period of minutes, days, months and years. This paper focuses on day-ahead forecasting and starts with a theoretical and economical overview of the electrical grid and energy market. The main reasons to focus day-ahead forecasting is to ensure the balance between the demand and supply of electricity and because the energy needs to be sold against a day-ahead spot price. Based on a literature study in the field of forecasting wind power it has been found that factors such as geographical location, data sources and grid sizes show influence on the accuracy of the data and therefore influence the prediction of wind power. Furthermore, based on the literature input parameters such as wind speed, wind direction, weather stability, availability, relative humidity and seasonal data have been found useful as input data for forecasting methods to forecast wind power dayahead. From a large set of forecasting methods it has been found that the most used techniques to predict wind power day-ahead are physical methods, and statistical or hybrid methods such as neural networks. This research has obtained forecasting results from a Random forest, Feed forward neural network and a hybrid model consisting of a combination of unsupervised k-nearest neighbour clustering and a neural network. These results have been compared with the forecasting results obtained from an external organization. Based on the comparison of monthly and average monthly MAPD and RMSPD we have found that the Feed forward neural network and the hybrid model are able to obtain a performance equally or even better compared to the external forecasting for a single turbine. The input parameters that made the difference were the u-vector, v-vector, the use of SCADA data and the wind speed time lag 1. Furthermore, the three forecasting models did perform less good compared to the external forecasting on forecasting wind power generated by a wind farm. Main reasons are because we did not take shadowing effects from other turbines into account and also the lack of fuzzy rules overfitted the neural networks at higher wind speed values. The random forest however was more robust and performed best of the three models.