Scholarlink

<p>Accurate photovoltaic (PV) power forecasting is essential for enhancing grid stability and optimizing renewable energy integration. Conventional long short-term memory (LSTM) neural networks are widely applied in this field due to their strength in modeling sequential dependencies, but they face limitations in capturing global features and maintaining computational efficiency. To overcome these drawbacks, this article proposes a novel feature-cascade enhanced LSTM model that integrates the Transformer encoder (TE) for global representation learning and the interactive convolution block (ICB) for local feature extraction and cross-scale feature fusion, followed by an LSTM for temporal sequence modeling. Experiments conduct on a real-world dataset collected from a PV power station in Ningxia, China, demonstrate that the proposed model significantly improves forecasting accuracy compared with other baseline models. Compared with the traditional LSTM model, the proposed method achieves substantial improvements in prediction accuracy in different weather patterns. Specifically, the improvements in MAE, RMSE, WAPE, and MGF reach 36.1139%, 15.1343%, 36.1118%, and 3.0241% in the rainy-day pattern, 38.6747%, 26.3807%, 38.6753%, and 5.9870% in the cloudy-day pattern, and 50.4565%, 42.1669%, 50.4577%, and 2.5000% in the sunny-day pattern, respectively. These results demonstrate the effectiveness and robustness of the proposed approach for practical photovoltaic power 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

Feature Cascade Enhanced LSTM for Short-Term PV Power Prediction

  • 期刊论文

  • 发表日期：

    2026年2月25日

查看全文
  • Yiwei Ma,
  • Xu Mai,
  • Dong Yan

摘要

Accurate photovoltaic (PV) power forecasting is essential for enhancing grid stability and optimizing renewable energy integration. Conventional long short-term memory (LSTM) neural networks are widely applied in this field due to their strength in modeling sequential dependencies, but they face limitations in capturing global features and maintaining computational efficiency. To overcome these drawbacks, this article proposes a novel feature-cascade enhanced LSTM model that integrates the Transformer encoder (TE) for global representation learning and the interactive convolution block (ICB) for local feature extraction and cross-scale feature fusion, followed by an LSTM for temporal sequence modeling. Experiments conduct on a real-world dataset collected from a PV power station in Ningxia, China, demonstrate that the proposed model significantly improves forecasting accuracy compared with other baseline models. Compared with the traditional LSTM model, the proposed method achieves substantial improvements in prediction accuracy in different weather patterns. Specifically, the improvements in MAE, RMSE, WAPE, and MGF reach 36.1139%, 15.1343%, 36.1118%, and 3.0241% in the rainy-day pattern, 38.6747%, 26.3807%, 38.6753%, and 5.9870% in the cloudy-day pattern, and 50.4565%, 42.1669%, 50.4577%, and 2.5000% in the sunny-day pattern, respectively. These results demonstrate the effectiveness and robustness of the proposed approach for practical photovoltaic power prediction.