Scholarlink

<p>Accurate photovoltaic (PV) power forecasting serves as a critical foundation for economic dispatch and reliable grid operation. To address the inherent uncertainty in PV power generation, this study proposes a short-term PV power interval prediction method based on Bayesian-optimized CNN-BiLSTM-attention (BO-CNN-BiLSTM-attention) that accounts for conditional dependencies in prediction errors. The methodology comprises three main stages: first, PV output data undergoes preprocessing and feature selection. Second, a Bayesian-optimized CNN-BiLSTM-attention model achieves high-precision point forecasting for target time periods. Finally, the K-shape time series clustering algorithm matches point predictions with temporally similar historical data, while adaptive bandwidth kernel density estimation models the probability distribution of prediction errors from similar patterns, thereby enabling interval prediction. Experimental validation on a photovoltaic plant in Xinjiang, China demonstrates that the proposed method achieves superior prediction accuracy compared to various single and ensemble forecasting models, while outperforming multiple interval construction approaches in terms of prediction effectiveness.</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

Photovoltaic power interval prediction with conditional error dependency using Bayesian optimized deep learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月13日

查看全文
  • Yixuan Chen,
  • Xinggang Wang,
  • Run Huang,
  • Guangzeng You

摘要

Accurate photovoltaic (PV) power forecasting serves as a critical foundation for economic dispatch and reliable grid operation. To address the inherent uncertainty in PV power generation, this study proposes a short-term PV power interval prediction method based on Bayesian-optimized CNN-BiLSTM-attention (BO-CNN-BiLSTM-attention) that accounts for conditional dependencies in prediction errors. The methodology comprises three main stages: first, PV output data undergoes preprocessing and feature selection. Second, a Bayesian-optimized CNN-BiLSTM-attention model achieves high-precision point forecasting for target time periods. Finally, the K-shape time series clustering algorithm matches point predictions with temporally similar historical data, while adaptive bandwidth kernel density estimation models the probability distribution of prediction errors from similar patterns, thereby enabling interval prediction. Experimental validation on a photovoltaic plant in Xinjiang, China demonstrates that the proposed method achieves superior prediction accuracy compared to various single and ensemble forecasting models, while outperforming multiple interval construction approaches in terms of prediction effectiveness.