Dissertation
Motivation
The continuous rise of fossil fuel's prices, together with environmental concerns, makes countries all over the world increase the penetration of wind power in their electrical systems, since this is a clean and renewable source of energy. The challenge in the integration of this type of generation in the electrical grid comes from the high variability and uncertainty of the wind. Therefore, wind power forecasting becomes an important tool to deal with this problem.
Accordingly to REN's data, in 2012, 37% of the energy demand was supplied by renewable sources, with wind power producing 20% of the total generation. This constant growth makes it more important to have accurate wind power forecasts, so that system operators are able to balance the demand and the supply in an efficient manner.
Objectives
There are a number of different forecasting models, of different types and using diverse approaches. All over the world there have been countless efforts and studies in trying to understand which ones are more appropriate in each case and in trying to improve wind forecasting overall.
The subject of this dissertation is framed within a project of SiNGULAR and it's main goal it to develop new wind power forecasting methods for a time horizon of about 7 days (short term), applied to wind farms from the island of Crete, Greece.
Throughout this paper, it is intended to conduct a study on optimizing both point forecast of expected production values, and the uncertainty associated with these. This is where the models under study differ from current ones. In a simplified way, it is intended that this work meets the following objectives:
- Development of methods to filter errors in historical series, used for the models training;
- Creation of forecasting models for the aggregate of wind farms used;
- Obtaining forecasts of wind generation, coupled with the uncertainty associated with these;
- Implementation of new forecasting techniques based on KDE, comparing them with neural networks and the persistence model;
- Development of methods for the parameterization of probabilistic models.
For this purpose, the project uses ensemble Numerical Weather Predictions (NWP) for the case study of the island of Crete. As software, KDE, Matlab and Microsoft Office Excel are the main tools applied.