Generation of typical wind power scenarios based on spectral normalization generative adversarial networks and spectral clustering

To address the high-dimensional complexity of wind farm scenario data， a generation method based on spectral normalization generative adversarial networks and spectral clustering is proposed. First， adversarial training is performed on two deep neural networks of the generator and discriminator， with spectral normalization applied to the convolutional layers of the discriminator to enhance the Lipschitz continuity constraint， thereby improving the stability of data training and the quality of generated wind farm scenarios. Next， an improved Gaussian kernel-based spectral clustering method is used to extract wind power features and reduce the dimensionality of the data， transforming the generated scenarios into a set of typical wind farm scenarios. Finally， simulations are conducted using the publicly available WIND dataset. The simulation results indicate that the proposed method significantly reduces the mean squared errors of generated samples， accurately capturing the spatiotemporal correlations of wind power generation； the spectral clustering based on Gaussian kernel function effectively clusters the sample space.