Scenario Generation for Wind Power Using Improved Generative Adversarial Networks

TLDR

Wind power scenarios critically influence stochastic optimization in power systems, and while GANs can generate realistic renewable power scenarios without explicit modeling, overfitting remains a challenge on small training sets. The study proposes an improved GAN that enhances wind power scenario generation by refining the Lipschitz constraint on the discriminator. The method employs a gradient‑penalty Lipschitz constraint and a consistency term during training, and is evaluated on NREL wind integration time‑series data. Enforcing Lipschitz continuity reduces overfitting, and experimental results show the improved GAN outperforms existing methods.

Abstract

Wind power scenarios have a significant impact on stochastic optimization problems for power systems in which wind power is a significant component. Generative adversarial networks (GANs) are a powerful class of generative models, and can generate realistic scenarios for renewable power sources without the need for any modeling assumptions. However, the performance of GANs in generating scenarios can further be improved by modifying the way in which a Lipschitz constraint on discriminator network is imposed. Another critical problem of applying deep neural networks is overfitting, a phenomenon especially prone to appear on small training sets. In this paper, we propose an improved GAN for the generation of wind power scenarios. To improve the training speed, we use a gradient penalty term to enforce the Lipschitz constraint based on the output and input of the discriminator network. To improve the scenario quality, we further use a consistency term in the training procedure. Besides, the overfitting problem can be effectively alleviated by the enforced Lipschitz continuity. The proposed method is applied to actual time series data from the NREL wind integration data set. The experimental results demonstrate that our method outperforms the existing methods.