Bir rüzgar çiftliğinden yapay sinir ağlarıyla kısa süreli elektrik üretim tahmini
xmlui.mirage2.itemSummaryView.MetaDataShow full item record
In recent years, wind energy systems become an important source for energy production around the world with increasing installations. When Turkey’s wind energy potential is taken into consideration, wind energy systems gain importance also in Turkey in parallel with this development. Because wind is an intermittent source, wind power forecasts are needed for the integration of wind energy systems into the grid and utilizing efficiently the generated energy. As energy production from wind is depends on the wind characteristic of the site, orography, obstacles around the site and such regional conditions, different mathematical models are built for different regions in the literature. In this study, short term electricity production forecasting with ANN was carried out with data obtained from a wind farm which is located in Turkey. Firstly, data sets were examined, data analysis was performed, and suitable data sets for ANN was constructed. Results given from developed models was evaluated according to persistence method which is accepted as a benchmark for short term wind power forecasts. The error for the estimates carried out with two basic ANN models, which are a static and a dynamic model, were found to be in an acceptable range according to persistence method. Generally, the lowest mean absolute percentage errors were obtained from the static ANN model. However, when the estimated average production values was taken into consideration, it was seen that results obtained from dynamic models for chosen time horizons were very close to the actual average production values for these time horizons. When results obtained from models with different activation functions were compared, it was seen that there are no considerable differences between the results. Because training data has a direct effect on learning of network, different networks with different training data were constructed for test days. In this case, while dynamic models was trained with less data compared to static model, static model, trained with a data set which includes data from each season, gave forecast results which had lower errors. Dynamic networks; trained with data belonging to two days preceding the test day and trained with data belonging to one week preceding the test day, gave similar forecast results, and errors of these results were in an acceptable range according to persistence method. In the second part, temperature was added as an input to dynamic ANN models. Because wind farm data was only used for short term forecasts in this study, adding temperature input to networks did not affect forecast results significantly. In the last part, random errors between 5% and 20% were added to test input (wind speed) to investigate the effects of the errors in meteorological data on ANN predictions. Then these test data was used for forecast and it was seen that meteorological data include 20% random error caused an increase in forecast errors only 2.69%.