Two-stage scenario generation of hydro-wind-solar complementary system based on improved variational autoencoder and generative adversarial networks model

With the rapid development of new energy technologies, quantifying the spatial-temporal correlation between runoff and wind power and solar power generation capacity and generating representative high-dimensional coupled scenario sets have become urgent issues. As existing methods struggle to generate high-dimensional, long-term reliable scenarios under small sample conditions, this paper proposes a two-stage long-term scenario generation method for runoff-wind-solar integration. In the first stage, it couples generative adversarial network with variational autoencoder, and improves the model by introducing Wasserstein distance and loss weight coefficients to ensure training stability, scenario quality and diversity, and enable initial learning of distribution characteristics from small sample annual data. In the second stage, it incorporates Cholesky decomposition and quantile mapping to enhance the spatial-temporal correlation of scenario sets, thus obtaining highly reliable scenario sets that retain the spatial-temporal features of original data. Validation against scenarios from single-stage VAE-GAN, GAN, VAE and Copula proves the proposed method's effectiveness: in the three-element scenarios, the average absolute error of Kendall correlation coefficient reaches 0.03, 0.02 and 0.04 respectively. This method can provide key technical support for the long-term planning and dispatching of runoff-wind-solar complementary systems.