Scholarlink

<p>Photovoltaic power has strong randomness and intermittenity. A prediction approach utilizing graph convolutional neural network is proposed to improve the accuracy of photovoltaic power forecasting. Firstly, based on the graph theory, adjacent photovoltaic sites are modeled in a graphical form. Instead of estimating the relationships between each site based on the temporal correlations of their time series, this approach captures the future dynamic relationships between neighboring sites from a predictive perspective. Subsequently, a new neural network is proposed. The network combines a residual multi-layer graph convolutional network with gated linear unit to separately extract spatial features between nodes and temporal features of each node. When extracting temporal features, discrete wavelet decomposition is utilized to enable gated linear unit to extract more temporal features. Finally, Two case studies validate that the proposed method demonstrates good information capture performance and adaptability, significantly improving the prediction accuracy of each site for 4&#xa0;h-ahead prediction.</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

Ultra-Short–Term Forecasting of Multi-Site PV Power Based on Graph Convolution Network

  • 期刊论文

  • 发表日期：

    2025年4月29日

查看全文
  • Lingling Xie,
  • Jianguo Yang,
  • Bin Liu

摘要

Photovoltaic power has strong randomness and intermittenity. A prediction approach utilizing graph convolutional neural network is proposed to improve the accuracy of photovoltaic power forecasting. Firstly, based on the graph theory, adjacent photovoltaic sites are modeled in a graphical form. Instead of estimating the relationships between each site based on the temporal correlations of their time series, this approach captures the future dynamic relationships between neighboring sites from a predictive perspective. Subsequently, a new neural network is proposed. The network combines a residual multi-layer graph convolutional network with gated linear unit to separately extract spatial features between nodes and temporal features of each node. When extracting temporal features, discrete wavelet decomposition is utilized to enable gated linear unit to extract more temporal features. Finally, Two case studies validate that the proposed method demonstrates good information capture performance and adaptability, significantly improving the prediction accuracy of each site for 4 h-ahead prediction.