GAN-based Extreme Conditional Distribution Estimation for Renewable Energy Systems

Given the intermittent and stochastic nature of renewable energy sources, along with the potential high consequences of extreme scenarios, there is a need to examine scenario distribution under such challenging conditions. This paper addresses the complex and high-dimensional nature of the Extreme Conditional Distribution in renewable power systems. We propose utilizing Generative Adversarial Networks (GANs) to replicate the distribution with a large number of generated extreme scenarios. While existing GANs tailored for extreme conditions often focus on generating individual extreme scenarios, our proposed approach, GMM-GAN, efficiently generates a substantial number of extreme scenarios by incorporating Gaussian Mixture Model (GMM)-based importance sampling, thereby enhancing the efficiency of extreme scenario generation. Experimental results demonstrate that our method excels in generating accurate empirical distributions with a high valid rate for scenarios meeting extreme conditions. GMM-GAN can be used to generate diverse and realistic scenarios that will enable more effective planning by grid operations in response to rare but high-impact events.