It is difficult to obtain an accurate characterization of the uncertainty of renewable energy output. Thus an approach for generating stochastic scenarios of renewable energy based on improved maximum mean discrepancy generative adversarial networks (MMD-GAN) is proposed. First, the fundamental principles of GAN and MMD-GAN are described, and an improved scheme of MMD-GAN is proposed, one which enhances the discriminator’s loss function on the basis of MMD-GAN and uses spectral normalization and the bounded Gaussian kernel to improve the training stability of the generator and discriminator. Then, the process of stochastic scenario generation for renewable energy based on the improved MMD-GAN is designed. Finally, the effects of the proposed methods are analyzed. The performance of improved MMD-GAN, MMD-GAN, and typical GAN are compared. The results indicate that the improved MMD-GAN method can reduce the Wasserstein distance between the generated distribution and the real distribution by more than 50% contrasted with the comparison method, and the generated scenario accuracy can be effectively improved.