Integrated wind speed forecasting model based on Fast Fourier transform, factor decomposition and Gated Recurrent Unit network

The operational efficiency of wind farms and the reliability of the grid depend heavily on accurate wind speed forecasts. However, this critical task is complicated by the inherent complexity of meteorological data, coupled with dynamic timevarying nature and random fluctuations. To overcome the limitations of existing decomposition methods in periodicity identification and specific component prediction, this paper proposes a hybrid prediction framework by combining the Fast Fourier Transform (FFT), factor decomposition and Gated Recurrent Unit (GRU) network. Firstly, the FFT is employed to accurately extract the periodicity from wind speed series. Then, with the aid of the obtained periodicity, two factor decomposition strategies in terms of the multiplication decomposition and the mixing decomposition are utilized to decompose the original wind speed series into trend, seasonal and irregular components. Finally, the following three prediction experiments are designed and conducted: in Experiment 1, only the irregular component is forecasted via the GRU network; in Experiment 2, both the seasonal and irregular components are forecasted through the GRU network; While Experiment 3 is designed by predicting all of the three decomposed components according to the GRU network. Experimental results indicate that the mixing decomposition strategy outperforms the multiplication way for the wind speed data adopted in this paper. This study demonstrates the effectiveness of factor decomposition and highlights that irregular component prediction remains a key challenge for future improvements in wind speed forecasting.