Scholarlink

<p>Accurate solar radiation forecasting is essential for optimizing the design, operation, and management of solar power systems. It ensures reliable energy trading, efficient maintenance scheduling and stable power generation. Deep learning (DL) models, known for capturing complex nonlinear relationships, have demonstrated potential in solar irradiance prediction. However, limitations such as data variability and privacy concerns persist. To overcome these challenges, this study introduces a novel hybrid DL model that combines self-attention, bidirectional long short-term memory, and bidirectional gated recurrent unit architectures. The model is trained using features selected through recursive feature elimination and ridge regression, ensuring the most relevant data is utilized for prediction. Using ground-measured global horizontal irradiance dataset, the proposed SA-Bi-LSTM-Bi-GRU model achieves exceptional forecasting performance. It significantly reduces MAE by 64.40–90.34 W/m<sup>2</sup>, RMSE by 16.15–57.30% W/m<sup>2</sup>, MAPE by 41.38–69.78%, NMRSE by 8.29–21.65% W/m<sup>2</sup> compared to considered standalone and hybrid models, while achieving an R<sup>2</sup> of 0.98. These improvements underscore the model's ability to deliver precise and reliable forecasts, enhancing grid stability and energy efficiency. By integrating advanced architectures and achieving substantial performance gains, this model offers a groundbreaking solution for sustainable solar energy management.</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

Enhancing solar radiation forecasting accuracy with a hybrid SA-Bi-LSTM-Bi-GRU model

  • 期刊论文

  • 发表日期：

    2025年2月19日

查看全文
  • Girijapati Sharma,
  • Subhash Chandra,
  • Arvind Kumar Yadav,
  • Rahul Gupta

摘要

Accurate solar radiation forecasting is essential for optimizing the design, operation, and management of solar power systems. It ensures reliable energy trading, efficient maintenance scheduling and stable power generation. Deep learning (DL) models, known for capturing complex nonlinear relationships, have demonstrated potential in solar irradiance prediction. However, limitations such as data variability and privacy concerns persist. To overcome these challenges, this study introduces a novel hybrid DL model that combines self-attention, bidirectional long short-term memory, and bidirectional gated recurrent unit architectures. The model is trained using features selected through recursive feature elimination and ridge regression, ensuring the most relevant data is utilized for prediction. Using ground-measured global horizontal irradiance dataset, the proposed SA-Bi-LSTM-Bi-GRU model achieves exceptional forecasting performance. It significantly reduces MAE by 64.40–90.34 W/m2, RMSE by 16.15–57.30% W/m2, MAPE by 41.38–69.78%, NMRSE by 8.29–21.65% W/m2 compared to considered standalone and hybrid models, while achieving an R2 of 0.98. These improvements underscore the model's ability to deliver precise and reliable forecasts, enhancing grid stability and energy efficiency. By integrating advanced architectures and achieving substantial performance gains, this model offers a groundbreaking solution for sustainable solar energy management.