Renewable scenario generation using controllable generative adversarial networks with transparent latent space

With the growing penetration of renewable energy sources in power systems, it becomes increasingly important to characterize their inherent variability and uncertainty. Scenario generation is a key approach to provide a series of possible power scenarios in the future for the system planner and operator to make decisions. In this paper, a data-driven method is presented for renewable scenario generation using stable and controllable generative adversarial networks with transparent latent space (ctrl-GANs). The machine learning based algorithm can capture the nonlinear and dynamic renewable patterns without the need for modeling assumptions and complicated sampling techniques. The orthogonal regularization and spectral normalization are adopted to improve the training stabilization of the GAN model. To control the generation process, a relationship is built between features of the generated scenarios and latent vectors on the manifold. Moreover, several new metrics for GANs are used to evaluate the quality of the scenarios. The proposed approach is applied to generate realistic time series data of wind and photovoltaic power. The results demonstrate that our method has a better performance on numerical stabilization and is able to control the generation process with latent space.