Scholarlink

<p>To address the strong non-stationarity of wind power sequences and the complex mechanisms of disturbance propagation, existing studies on wind power forecasting often struggle to simultaneously capture time-varying spatial disturbance propagation and multi-scale temporal dependencies, leading to limited performance under non-stationary fluctuations and long-horizon forecasting scenarios. To this end, this paper proposes a multi-turbine wind power forecasting approach that integrates dynamic disturbance awareness with spatiotemporal feature modeling. Specifically, we first construct a disturbance-mechanism-driven dynamic adjacency matrix that jointly accounts for inter-turbine geographical proximity, wind-direction relationships, wake effects, and turbulence intensity, thereby dynamically characterizing the structural evolution of disturbances across the wind farm. We then employ a graph neural network composed of GraphSAGE and GAT to extract structural dependencies among turbines for spatial modeling, while introducing TimesNet for multi-scale temporal representation learning. Finally, spatiotemporal features are fused to perform wind power prediction. Experimental results demonstrate that the proposed method consistently outperforms multiple competitive baselines in both single-step and multi-step forecasting tasks and maintains higher stability and generalization capability under longer forecasting horizons.</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

Disturbance-aware spatiotemporal feature fusion for wind power forecasting

  • 期刊论文

  • 发表日期：

    2026年5月14日

查看全文
  • Xunwen Su,
  • Hangyuan Wei,
  • Jiahao Geng,
  • Yu Han,
  • Zijing Wu,
  • Siqi Zhang

摘要

To address the strong non-stationarity of wind power sequences and the complex mechanisms of disturbance propagation, existing studies on wind power forecasting often struggle to simultaneously capture time-varying spatial disturbance propagation and multi-scale temporal dependencies, leading to limited performance under non-stationary fluctuations and long-horizon forecasting scenarios. To this end, this paper proposes a multi-turbine wind power forecasting approach that integrates dynamic disturbance awareness with spatiotemporal feature modeling. Specifically, we first construct a disturbance-mechanism-driven dynamic adjacency matrix that jointly accounts for inter-turbine geographical proximity, wind-direction relationships, wake effects, and turbulence intensity, thereby dynamically characterizing the structural evolution of disturbances across the wind farm. We then employ a graph neural network composed of GraphSAGE and GAT to extract structural dependencies among turbines for spatial modeling, while introducing TimesNet for multi-scale temporal representation learning. Finally, spatiotemporal features are fused to perform wind power prediction. Experimental results demonstrate that the proposed method consistently outperforms multiple competitive baselines in both single-step and multi-step forecasting tasks and maintains higher stability and generalization capability under longer forecasting horizons.