Wind turbines has played an important part in the energy grid and in the growth of renewable energy sources. When it comes to producing electrical energy from wind turbines, the challenging part is the accurate prediction of power generation due to high volatility and intermittency of the wind. The knowledge about the accuracy of different prediction models is important for better production planning, load balancing, optimal trading on electricity market, and better utilization of the wind resource.
In order to model wind power prediction, an exploratory data analysis is performed on wind speed data and wind power data to characterize the power curve behavior. The analysis is performed from statistical, quantitative, temporal and probabilistic aspects.
Moreover, this work proposes a method for data preprocessing and outlier removal to improve the performance of prediction model. The preprocessing step is an important step prior to building a prediction model and it can affect the prediction accuracy. The proposed method is based on detecting extremely large changes within power series and wind speed series. The method yields good results and its performance is evaluated.
Furthermore, this work addresses the problem of assessing the prediction models for different time horizons and different types of data resolutions that are used to make the prediction. In this work several types of prediction models are built and tested with different types on input data. The accuracy of models is evaluated in terms of the time horizon, as well as the resolution of the input data, and the prediction strategies for very-short term and short-term prediction horizons are established. The prediction models include an ARIMA model, Support Vector regression, Artificial Neural Network Multi Layer Perceptron, as well as a new model: Probabilistic State Transition model.