Diploma Thesis
Untersuchung zu einem Windleistungsprognosesystem mittels künstlicher neuronaler Netze auf der Grundlage der Daten des meteorologischen Meßnetzes Mecklenburg-Vorpommern
Stefan Müller (07/2000-12/2000)
Support: Thomas Foken
The intention of this work was to examine the suitability of artificial neural networks for predicting windspeed. This is the most important parameter of a windprognosis system. An advantage of artificial neural networks is the lack of necessity to identify processes, as the meteorological as well as the coherence of time and space can be learned by the system itself. Artificial neural networks based on the biological process of information management. The database were hourly means of measure results from seven meteorological stations in Mecklenburg-Vorpommern which were recorded within a meteorological measuring project between 1993 and 1997.
Artificial neural networks of 1st and 2nd order for predicting times of one, two, three and five hours were created and analysed. Different combinations of input parameters were applied. It turned out that a combination of the different model input parameters did not play an important role for the quality of prediction with the exception of the windspeed. In contrary there was an positive influence to the prognosis quality using artificial neural networks of 2nd order (with shortcut connections) especially for short prediction times. It was proved, that these empirical models predict those situations best which were presented them most often throughout the trainingphase i.e. above all-through windspeed around the modal-value or the winds from the main wind directions. For that reason the accuracy of the prediction of very low and very high windspeeds is poor. This can be a problem for the energy provider in predicting windspeeds at turn-on and cut-off windspeed for wind-power plants. It was found out the simple model of persistence which was taken as a reference could be exceeded starting with a prediction time of two hours. Compared to the numerical Deutschlandmodell of the German Weather Service the artificial neural networks have a better accuracy. Artificial neural networks are therefore able to close the gap between the simple persistence from two hours on and the numerical models up to six hours.