Scholarlink

<p>This paper presents a hybrid deep learning framework for wind power forecasting that integrates Long Short-Term Memory (LSTM) networks with a Multi-Layer Perceptron (MLP) through an attention-based feature fusion mechanism. The proposed methodology addresses the challenge of simultaneously capturing temporal dependencies in dynamic meteorological sequences and nonlinear mappings of static turbine characteristics. Dynamic features (temperature, pressure, humidity) are processed through LSTM layers, while static features (wind speed/direction at multiple heights) are encoded via MLP with adaptive fusion weights learned through attention. Empirical evaluation on a one-year dataset of 35,040 15-min records from a Xinjiang wind-solar hybrid farm demonstrates that the proposed model achieves a Root Mean Square Error of 26.83&#xa0;MW and R<sup>2</sup> of 0.71, outperforming direct concatenation (RMSE 28.50&#xa0;MW, R<sup>2</sup> 0.69) and single LSTM (RMSE 52.17&#xa0;MW, R<sup>2</sup> −0.05). An ablation study further validates that inclusion of photovoltaic parameters as auxiliary atmospheric descriptors reduces RMSE by 2.0% compared to the solar-excluded variant. While Random Forest achieves marginally lower Mean Absolute Error (16.97 vs. 18.18&#xa0;MW) through effective low-generation regime approximation, the proposed framework offers superior peak-tracking capability (lowest RMSE), balanced error distribution across operational regimes, and scalable modular architecture for evolving sensor infrastructure. The framework provides a reproducible solution for wind power prediction, with particular advantages in handling the intermittent and stochastic nature of wind resources through physics-aligned feature processing and adaptive multi-information fusion.</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 hybrid deep learning framework for wind power load forecasting with attention-based feature fusion

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月16日

查看全文
  • Jie Wu,
  • Hanyuan Li

摘要

This paper presents a hybrid deep learning framework for wind power forecasting that integrates Long Short-Term Memory (LSTM) networks with a Multi-Layer Perceptron (MLP) through an attention-based feature fusion mechanism. The proposed methodology addresses the challenge of simultaneously capturing temporal dependencies in dynamic meteorological sequences and nonlinear mappings of static turbine characteristics. Dynamic features (temperature, pressure, humidity) are processed through LSTM layers, while static features (wind speed/direction at multiple heights) are encoded via MLP with adaptive fusion weights learned through attention. Empirical evaluation on a one-year dataset of 35,040 15-min records from a Xinjiang wind-solar hybrid farm demonstrates that the proposed model achieves a Root Mean Square Error of 26.83 MW and R2 of 0.71, outperforming direct concatenation (RMSE 28.50 MW, R2 0.69) and single LSTM (RMSE 52.17 MW, R2 −0.05). An ablation study further validates that inclusion of photovoltaic parameters as auxiliary atmospheric descriptors reduces RMSE by 2.0% compared to the solar-excluded variant. While Random Forest achieves marginally lower Mean Absolute Error (16.97 vs. 18.18 MW) through effective low-generation regime approximation, the proposed framework offers superior peak-tracking capability (lowest RMSE), balanced error distribution across operational regimes, and scalable modular architecture for evolving sensor infrastructure. The framework provides a reproducible solution for wind power prediction, with particular advantages in handling the intermittent and stochastic nature of wind resources through physics-aligned feature processing and adaptive multi-information fusion.