Scholarlink

<p>The incorporation of both spatial and temporal characteristics is vital for improving the predictive accuracy of photovoltaic (PV) power generation forecasting. However, in multivariate time series forecasting, an excessive number of features and traditional decomposition methods often lead to information redundancy and data leakage, resulting in prediction bias. To address these limitations, this study introduces the Time-Embedding Temporal Convolutional Network (ETCN), providing an innovative solution. The ETCN model follows three key steps: First, a heatmap-based feature selection method identifies and prioritizes the most influential features, enhancing performance and reducing computational complexity. Second, a novel embedding architecture using multi-layer perceptrons captures both periodic and non-periodic information. Finally, a fusion strategy integrates spatial and temporal features through a DCNN with residual connections and BiLSTM, enabling effective modeling of complex data relationships. The ETCN model achieved outstanding performance, with RMSE, MAE, and <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="41598_2025_4630_Article_IEq1.gif" Format="GIF" Height="17" Rendition="HTML" Resolution="72" Type="Linedraw" Width="19" /> </InlineMediaObject> <EquationSource Format="TEX">\(\hbox {R}^{2}\)</EquationSource> </InlineEquation> values of 0.5683, 0.3388 and 0.9439, respectively, on the Trina dataset. On the Sungrid dataset, the model achieved top results of 0.1351, 0.0964 and 0.9894, further demonstrating its superior accuracy across both datasets.</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

Short term prediction of photovoltaic power with time embedding temporal convolutional networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年7月1日

查看全文
  • Jingxin Wang,
  • Guohan Li,
  • Jin Gu,
  • Zhengyi Xu,
  • Xinrong Chen,
  • Jianming Wei

摘要

The incorporation of both spatial and temporal characteristics is vital for improving the predictive accuracy of photovoltaic (PV) power generation forecasting. However, in multivariate time series forecasting, an excessive number of features and traditional decomposition methods often lead to information redundancy and data leakage, resulting in prediction bias. To address these limitations, this study introduces the Time-Embedding Temporal Convolutional Network (ETCN), providing an innovative solution. The ETCN model follows three key steps: First, a heatmap-based feature selection method identifies and prioritizes the most influential features, enhancing performance and reducing computational complexity. Second, a novel embedding architecture using multi-layer perceptrons captures both periodic and non-periodic information. Finally, a fusion strategy integrates spatial and temporal features through a DCNN with residual connections and BiLSTM, enabling effective modeling of complex data relationships. The ETCN model achieved outstanding performance, with RMSE, MAE, and \(\hbox {R}^{2}\) values of 0.5683, 0.3388 and 0.9439, respectively, on the Trina dataset. On the Sungrid dataset, the model achieved top results of 0.1351, 0.0964 and 0.9894, further demonstrating its superior accuracy across both datasets.