Meta-parameter selection for Support Vector Machines in wind energy forecasting models
|Abstract||Support Vector Machines (SVMs) is a machine learning technique that allows the combination of the simplicity and uniqueness of linear models with the possibility of a highly nonlinear, kernel based, preprocessing into a possibly infinite dimensional extended feature space. This results in powerful models that can be applied to classification and regression problems. While quite simple, SVMs require the selection of two structural parameters, the penalty term that is applied to margin slack values and, in the case of Support Vector Regression (SVR), the tolerance threshold under which errors are not penalized. If, as usually done, Gaussian kernels are used, a third parameter, the kernel width has also to be selected.|
The choices made may greatly affect the performance of SVMs and their correct selection is an important task when applying them to concrete problems in machine learning. In this Master Thesis methods for this choice will be considered, coming either from general, model independent, approaches to parameter selection or from concrete procedures that rely on the SVM structure. After reviewing the basic facts on SVMs and the state of the art literature on SVM parameter selection, three of such methods will be selected for implementation and application, first to data sets available at machine learning repositories and then to the concrete problem of building SVR models to predict wind energy production at individual farms.
It is well known that parameter selection is a heavy computational task but suitable to execution time savings when properly parallelized. Because of this, and as a technological complement, the Open MP standard for parallel processing will also be considered and applied in the implementation and application of the selected parameter selection techniques.
|Type||Master's thesis [Academic thesis]|
|Publisher||Technical University of Denmark, DTU Informatics, E-mail: email@example.com|
|Address||Asmussens Alle, Building 305, DK-2800 Kgs. Lyngby, Denmark|
|Note||Supervised by Associate Professor Jan Larsen, firstname.lastname@example.org, DTU Informatics|
|BibTeX data|| [bibtex]|
|IMM Group(s)||Intelligent Signal Processing|
Back :: IMM Publications