| dc.description.abstract | Wind power is currently one of the fastest growing power generation sectors in the world. The intermittent and random nature of wind energy, however, poses difficulties when integrating it into the electricity grid. Furthermore, it makes it harder to perform necessary regulation actions. Accurate forecasting methods are thus essential for efficient operation of the wind turbines. This study has an emphasis on very-short term wind speed forecasting, i.e. from a few seconds to 30 minutes ahead. Two different methods were implemented when forecasting, namely linear prediction and neural networks (using a NARX network). The analyses were done by using wind data acquired from the Fr{o}ya Wind Measuring Station, and the persistence method was used as a benchmark for evaluating the performance of the two methods. When forecasting with the linear prediction method, only data from a single measurement station was used in the analyses. For the neural network approach, data acquired from one more measurement station was included so as to study the effects of spatial correlation. The linear prediction method was found to perform marginally better than the persistence method when predicting 5 seconds ahead. Better predictions were obtained when first lowpass filtering the original signal, but the relative improvement over the persistence method did not increase. Also, there were troubles with finding optimal model parameters and the prediction horizon was very limited. The NARX neural network, on the other hand, managed to predict further into the future. The most promising results were found when increasing the size of the training interval to 150000 samples. Here, the mean average prediction error was found to be significantly lower than for the persistence method. | nb_NO |
