A deep learning based short-term wind power prediction model with Fourier self-attention mechanism

Traditional wind power forecasting methods typically focus on predictions at a single scale. However, changes in wind power generation are influenced by both long-term trends and cyclical fluctuations. This study employs an improved wavelet decomposition method for the multi-scale decomposition of wind power generation series, independently revealing its trend and frequency components. For the trend component, we combine convolutional neural networks (CNN) and Fourier transforms for prediction. Convolutional neural networks excel at capturing local features of wind power data, while Fourier transforms provide insights into the global fluctuation patterns of the data. For the frequency component, we design a Fourier self-attention mechanism, which combines the powerful frequency-domain analysis capability of the Fourier transform with the context-capturing properties of the self-attention model to achieve accurate prediction of frequency fluctuations. Experimental results demonstrate that our proposed prediction method outperforms traditional methods.