Scholarlink

<p>Accurate wind speed forecasting is crucial for optimizing wind energy utilization. Existing models often overlook dynamic spatial correlations and prior knowledge, resulting in suboptimal performance. To address this, the study introduces SDG-WindNet, a novel spatiotemporal model combining static and dynamic graph learning for wind speed forecasting. Long-term static patterns are derived using weighted geographic and dynamic time warping (DTW) distances, while short-term dynamic relationships are captured via iterative node embeddings and multi-head attention mechanisms. The model integrates dynamic temporal features, dual-graph convolutional networks (GCN), and temporal convolutional networks (TCN), enabling deep spatiotemporal feature extraction from evolving wind speed data. Extensive simulations on three real-world datasets confirm the superiority of SDG-WindNet. For short-horizon forecasting, it achieves average improvements of 4.68% in Mean Absolute Error (MAE) and 10.84% in Mean Squared Error (MSE) on the Dutch dataset, 2.50% and 4.15% on the US dataset. For long-horizon forecasting, it delivers gains of 2.31% in MAE and 3.22% in MSE on the Danish dataset. Furthermore, the non-recurrent architecture of SDG-WindNet enables full parallelization on GPU-accelerated HPC platforms, achieving inference latencies suitable for real-time operational deployment in large-scale wind farm management systems.</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 spatiotemporal model with static-dynamic graph learning for multi-node wind speed forecasting

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Dabin Zhang,
  • Ni Hong,
  • Boting Zhang,
  • Huanling Hu,
  • Liling Zeng

摘要

Accurate wind speed forecasting is crucial for optimizing wind energy utilization. Existing models often overlook dynamic spatial correlations and prior knowledge, resulting in suboptimal performance. To address this, the study introduces SDG-WindNet, a novel spatiotemporal model combining static and dynamic graph learning for wind speed forecasting. Long-term static patterns are derived using weighted geographic and dynamic time warping (DTW) distances, while short-term dynamic relationships are captured via iterative node embeddings and multi-head attention mechanisms. The model integrates dynamic temporal features, dual-graph convolutional networks (GCN), and temporal convolutional networks (TCN), enabling deep spatiotemporal feature extraction from evolving wind speed data. Extensive simulations on three real-world datasets confirm the superiority of SDG-WindNet. For short-horizon forecasting, it achieves average improvements of 4.68% in Mean Absolute Error (MAE) and 10.84% in Mean Squared Error (MSE) on the Dutch dataset, 2.50% and 4.15% on the US dataset. For long-horizon forecasting, it delivers gains of 2.31% in MAE and 3.22% in MSE on the Danish dataset. Furthermore, the non-recurrent architecture of SDG-WindNet enables full parallelization on GPU-accelerated HPC platforms, achieving inference latencies suitable for real-time operational deployment in large-scale wind farm management systems.