Scholarlink

<p>Efficient and accurate scenario generation task is critical for modelling the inherent variability and unpredictability in wind and solar energy generation, especially in power systems with high renewable penetration. Scenario creation helps system operators and planners in making informed decisions by providing a range of future power possibilities. This paper explores the uses of deep convolutional generative adversarial networks (DCGANs) for generating renewable energy scenarios. In Phase I, orthogonal regularization is implemented in the generator network, whereas spectral normalization is employed in the discriminator network to improve the stability of DCGAN training. The approach utilizes Wasserstein distances to measure the disparity between actual and created scenarios, hence enhancing training efficacy. Upon completion of training, the model can successfully generate both unconditional and conditional time series scenarios for single-site and multi-site applications. Phase II presents a customizable DCGAN framework based on a generalized linear modelling (GLM) methodology. This phase incorporates a feature extraction process that uses a latent programmable vector in multivariate space, enabling controlled renewable scenario generation by adjusting latent vectors linked to specific features. The proposed machine learning strategy can detect nonlinear and dynamic renewable patterns and generate accurate wind and solar power generation time series data without model assumptions. The K-means clustering methodology is used to examine scenario pattern distributions on real clustered data and K-fold cross-validation is utilized to create controllable scenarios. Selected performance indicators compared generated scenarios to actual data, whereas cumulative density functions and kernel density estimator metrics evaluated scenario quality and controllability.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Renewable Energy Scenario Generation with Controllable Features Using Configurable Deep Convolutional Generative Adversarial Network

  • 期刊论文

  • 发表日期：

    2026年2月14日

查看全文
  • Mahesh Pal Singh,
  • Nabangshu Sinha,
  • Nidul Sinha

摘要

Efficient and accurate scenario generation task is critical for modelling the inherent variability and unpredictability in wind and solar energy generation, especially in power systems with high renewable penetration. Scenario creation helps system operators and planners in making informed decisions by providing a range of future power possibilities. This paper explores the uses of deep convolutional generative adversarial networks (DCGANs) for generating renewable energy scenarios. In Phase I, orthogonal regularization is implemented in the generator network, whereas spectral normalization is employed in the discriminator network to improve the stability of DCGAN training. The approach utilizes Wasserstein distances to measure the disparity between actual and created scenarios, hence enhancing training efficacy. Upon completion of training, the model can successfully generate both unconditional and conditional time series scenarios for single-site and multi-site applications. Phase II presents a customizable DCGAN framework based on a generalized linear modelling (GLM) methodology. This phase incorporates a feature extraction process that uses a latent programmable vector in multivariate space, enabling controlled renewable scenario generation by adjusting latent vectors linked to specific features. The proposed machine learning strategy can detect nonlinear and dynamic renewable patterns and generate accurate wind and solar power generation time series data without model assumptions. The K-means clustering methodology is used to examine scenario pattern distributions on real clustered data and K-fold cross-validation is utilized to create controllable scenarios. Selected performance indicators compared generated scenarios to actual data, whereas cumulative density functions and kernel density estimator metrics evaluated scenario quality and controllability.