Scholarlink

To address the challenges of strong randomness and limited capture of seasonal patterns in wind power prediction, this paper proposes a predictive model that integrates seasonal feature decomposition with hybrid deep learning (STL-CNN-GRU-Attention). The STL method is applied to split wind speed, wind direction, and power data into trend, seasonal, and residual components, successfully isolating periodic patterns from random fluctuations. For each component, a multi-dimensional feature matrix is constructed, from which local spatiotemporal features are extracted using a CNN. A GRU network is then utilized to capture temporal dependencies, and an attention mechanism is introduced to dynamically weight key features. Tests using actual operational data from wind farms indicate that the proposed model substantially surpasses the benchmark approaches, achieving superior performance in MSE (1.7884e-3), MAE (2.7351e-2), and R2 (0.9870). Furthermore, ablation studies confirm that STL decomposition and the attention mechanism play critical roles in enhancing prediction accuracy, establishing a dependable technical basis for incorporating wind power into integrated energy systems.

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

Power Forecasting Method for Wind Power Systems Based on Seasonal Wind Energy Characteristics and Hybrid Deep Learning Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • JunTuo Zhang,
  • Xin Guo

摘要

To address the challenges of strong randomness and limited capture of seasonal patterns in wind power prediction, this paper proposes a predictive model that integrates seasonal feature decomposition with hybrid deep learning (STL-CNN-GRU-Attention). The STL method is applied to split wind speed, wind direction, and power data into trend, seasonal, and residual components, successfully isolating periodic patterns from random fluctuations. For each component, a multi-dimensional feature matrix is constructed, from which local spatiotemporal features are extracted using a CNN. A GRU network is then utilized to capture temporal dependencies, and an attention mechanism is introduced to dynamically weight key features. Tests using actual operational data from wind farms indicate that the proposed model substantially surpasses the benchmark approaches, achieving superior performance in MSE (1.7884e-3), MAE (2.7351e-2), and R2 (0.9870). Furthermore, ablation studies confirm that STL decomposition and the attention mechanism play critical roles in enhancing prediction accuracy, establishing a dependable technical basis for incorporating wind power into integrated energy systems.