Scholarlink

<p>In recent years, photovoltaic (PV) power generation has received increasing attention. However, uncertainties in PV power output-especially the random variations caused by cloud cover-make accurate short-term forecasting essential for efficient power dispatch. This study proposes a novel ultra-short-term PV power forecasting framework that differs from previous methods by effectively integrating deep image features from ground-based sky images with multi-source meteorological and power generation data. Central to this approach is PatchDLinear, a precise and lightweight time series forecasting algorithm that improves upon traditional models by incorporating normalization and patch-based processing, which better captures local temporal patterns and trends. Initially, the Cloud Y-Net model is used to calculate cloud coverage and classify cloud types, extracting deep-level information from sky images. These features are then combined with meteorological and power generation data and fed into PatchDLinear for collaborative forecasting. Experiments using real-world images and data demonstrate that the improved module achieves RMSE enhancements of 5.10% and 7.84% over the original model structure, while the overall framework shows RMSE improvements ranging from 17.61 to 43.24% compared to baseline methods. These results confirm that PatchDLinear significantly enhances prediction accuracy, offering a robust solution for ultra-short-term PV power forecasting.</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

Distributed photovoltaic ultra-short-term power prediction using whole-sky images and multi-source data

  • 期刊论文

  • 发表日期：

    2025年4月16日

查看全文
  • Qinlong Zhang,
  • Beiping Hou,
  • Wen Zhu,
  • Jianwei Dong,
  • Aihua Yu,
  • Yuzhen Zhu,
  • Minyu Huang,
  • Feiyang Hu

摘要

In recent years, photovoltaic (PV) power generation has received increasing attention. However, uncertainties in PV power output-especially the random variations caused by cloud cover-make accurate short-term forecasting essential for efficient power dispatch. This study proposes a novel ultra-short-term PV power forecasting framework that differs from previous methods by effectively integrating deep image features from ground-based sky images with multi-source meteorological and power generation data. Central to this approach is PatchDLinear, a precise and lightweight time series forecasting algorithm that improves upon traditional models by incorporating normalization and patch-based processing, which better captures local temporal patterns and trends. Initially, the Cloud Y-Net model is used to calculate cloud coverage and classify cloud types, extracting deep-level information from sky images. These features are then combined with meteorological and power generation data and fed into PatchDLinear for collaborative forecasting. Experiments using real-world images and data demonstrate that the improved module achieves RMSE enhancements of 5.10% and 7.84% over the original model structure, while the overall framework shows RMSE improvements ranging from 17.61 to 43.24% compared to baseline methods. These results confirm that PatchDLinear significantly enhances prediction accuracy, offering a robust solution for ultra-short-term PV power forecasting.