Abstract:

The integration of a large number of wind farms with uncertain output into the power grid will bring potential hazards in operation and uncontrollable risks. The uncertainty of wind power output is described by uncertain scenario sets generated from the variational autoencoder-based scenario generation method. Aimed at the complex spatiotemporal correlation of multi-wind farm output and the possible "KL collapse" during the traditional variational autoencoder model training, this paper proposes a daily scenario generation method of spatiotemporal power based on the Gaussian mixture model and improved conditional variational autoencoding. The two-dimensional convolution technique is introduced to extract the spatiotemporal correlation for dimension reduction, and the maximizing min-angle regularization technique is used to strengthen the independence of latent features. Hyperspherical distribution, instead of Gaussian distribution, is used to avoid "KL collapse" and improve the stability and accuracy of scene generation training. In addition, considering the diversity and flexibility of daily power scenarios of multi-wind farm, Gaussian mixture clustering technology is introduced to generate uncertain scenario sets with differentiated and changing characteristics, enabling the generation of certain scenario sets with varied characteristics according to specific condition labels. The results of real examples show that compared with conventional methods, the proposed method reduces the accumulated error distribution of probability by 17% to 71%, and the average error of temporal and spatial correlation by 85% to 97% and 55% to 91%, respectively. Besides, the proposed method can accurately generate daily power scenario sets of multi-wind farm in different wind conditions, improving scene generation's diversity and flexibility.

Key words: wind power scenario generation, Gaussian mixture model, feature extraction, conditional variational autoencoder, hyperspherical distribution, regularization technique