Scholarlink

<p>With renewable energy’s growing share in the global energy matrix, precise photovoltaic (PV) power generation forecasting has become critical for grid stability and economic efficiency. Therefore, an accurate forecasting model can minimize energy waste and support informed operational decisions. This paper proposes a Dual Attention Gated Recurrent Unit Kolmogorov–Arnold Network (DAGKAN) model specifically designed for mid-term (7–30 day) PV forecasting. To address the dual challenges of long-term trend capture and short-term volatility in medium-term data, our model employs multi-criteria statistical screening. This screening uses Pearson, Spearman, and Kendall coefficients to identify dominant predictors: irradiance, panel temperature, conversion efficiency, and wind speed. We propose a dual-stream architecture that processes both statistically filtered data (ensuring robustness) and raw data (capturing deep feature interactions) through coordinated attention-GRU-KAN pathways. Evaluated on the 18th State Power Investment PV Station dataset from the AI-Operated Maintenance Big Data Competition, DAGKAN achieves state-of-the-art performance with 98% <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation> score, demonstrating superior generalization over benchmark models.</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

DAGKAN: Medium-Term PV Power Prediction via Dual-Attention GRU and Kolmogorov–Arnold Network

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月26日

查看全文
  • Jun Yuan,
  • Chengsi Yao,
  • Jiayu Zhou,
  • Rongjun Chen,
  • Xianxian Zeng,
  • Genhua Huang,
  • Guidong Zhang

摘要

With renewable energy’s growing share in the global energy matrix, precise photovoltaic (PV) power generation forecasting has become critical for grid stability and economic efficiency. Therefore, an accurate forecasting model can minimize energy waste and support informed operational decisions. This paper proposes a Dual Attention Gated Recurrent Unit Kolmogorov–Arnold Network (DAGKAN) model specifically designed for mid-term (7–30 day) PV forecasting. To address the dual challenges of long-term trend capture and short-term volatility in medium-term data, our model employs multi-criteria statistical screening. This screening uses Pearson, Spearman, and Kendall coefficients to identify dominant predictors: irradiance, panel temperature, conversion efficiency, and wind speed. We propose a dual-stream architecture that processes both statistically filtered data (ensuring robustness) and raw data (capturing deep feature interactions) through coordinated attention-GRU-KAN pathways. Evaluated on the 18th State Power Investment PV Station dataset from the AI-Operated Maintenance Big Data Competition, DAGKAN achieves state-of-the-art performance with 98% \(R^2\) score, demonstrating superior generalization over benchmark models.