Wavelet recurrent neural network with semi-parametric input data preprocessing for micro-wind power forecasting in integrated generation Systems

Abstract

Wind power penetration is increasing more and more in the modern power system and an accurate wind power forecasting is now required to provide an help to the system operator to consider this renewable source in economic scheduling and other typical tasks of the electrical power system or in smart grid applications. The novelty of this Wavelet Recurrent Neural Network (WRNN) based approach consists on the model construction for micro wind generations. The WRNN does not provides only a prediction for the wavelet coefficients like in other previous studies of the authors in this research area but it is able to reconstruct directly the power signal from band-selected coefficients. The presented approach does not provides only an accurate forecasting model respect to the state of art in the field, but it is also useful for case studies which suffer of a major lack of wind data regarding the geographic site and of accurate and long historical study of the wind speed time series. In fact due to the proposed method of training based on a semiparametric input data preprocessing as Parzen windows then wind power output forecasting is improved.