Scholarlink

<p>Accurate day-ahead wind power forecasting is crucial for power system dispatch and planning. This work proposes a novel Transformer-Gated Recurrent Unit (GRU) prediction model based on data feature reconstruction to enhance the accuracy of wind power forecasting. Firstly, the feature dataset is reconstructed using correlation analysis and Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to address the non-stationarity inherent in wind power sequences. Secondly, a Transformer-GRU prediction model is designed. This model integrates a multi-head attention mechanism to capture long-range temporal dependencies and a GRU decoder to extract partial features. Finally, the proposed model is evaluated against six benchmark models using real-world datasets from multiple wind farms in Gansu Province, China. Compared to a standard Transformer model, the proposed method achieves an improvement in coefficient of determination exceeding 5.56%, reduces the mean squared error by more than 1.043 MW relative to the baseline, and reduces computational time by over 50%. The results demonstrate that the proposed method can accurately forecast day-ahead wind power generation across different seasons and diverse wind farm locations. This model holds significant potential for improving the quality of future large-scale grid integration.</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

A Novel Transformer-GRU Day Ahead Wind Power Forecast Considering Data Feature Reconstruction

  • 期刊论文

  • 发表日期：

    2025年10月27日

查看全文
  • Tong Yang,
  • Minan Tang,
  • Hanting Li,
  • Jinping Li,
  • Hongjie Wang

摘要

Accurate day-ahead wind power forecasting is crucial for power system dispatch and planning. This work proposes a novel Transformer-Gated Recurrent Unit (GRU) prediction model based on data feature reconstruction to enhance the accuracy of wind power forecasting. Firstly, the feature dataset is reconstructed using correlation analysis and Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to address the non-stationarity inherent in wind power sequences. Secondly, a Transformer-GRU prediction model is designed. This model integrates a multi-head attention mechanism to capture long-range temporal dependencies and a GRU decoder to extract partial features. Finally, the proposed model is evaluated against six benchmark models using real-world datasets from multiple wind farms in Gansu Province, China. Compared to a standard Transformer model, the proposed method achieves an improvement in coefficient of determination exceeding 5.56%, reduces the mean squared error by more than 1.043 MW relative to the baseline, and reduces computational time by over 50%. The results demonstrate that the proposed method can accurately forecast day-ahead wind power generation across different seasons and diverse wind farm locations. This model holds significant potential for improving the quality of future large-scale grid integration.