.png){kind=icon}
.png){kind=icon}
Return
Wind power prediction method based on multi-stage signal processing and dynamic weighted attention
.png){kind=icon}
PRE
AI
DOI：10.1080/15435075.2025.2579936.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Accurate wind power prediction is essential for efficient dispatch and enhancing grid security. However, traditional methods face challenges with the non-stationary characteristics of wind energy and noise interference. This paper proposes a composite model integrating Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Dual-Tree Complex Wavelet Transform (DTCWT), a Residual Convolutional Attention Module (RCAM), Bidirectional Long Short-Term Memory (BiLSTM), and a Dynamic Weighted Multi-Head Attention (DWMHA) module. First, CEEMDAN decomposes the raw signal into high- and low-frequency components, after which DTCWT denoises the high-frequency portion. Subsequently, RCAM extracts features and BiLSTM captures long-term dependencies. Finally, the DWMHA module enables adaptive weight adjustment and feature fusion, enhancing the model’s adaptability to dynamic time series. Experiments at a Chinese wind farm show the proposed model achieves Root Mean Square Error (RMSE) values of 1.6937 MW, 5.4286 MW, and 7.6632 MW for one, six, and twelve-step-ahead predictions. Compared to the sub-optimal Informer model, our model reduces the RMSE by 3.31%, 2.83%, and 1.12% over the same horizons. These results significantly outperform competing models, confirming its strong potential for practical applications.
Keywords:
Wind power prediction
CEEMDAN decomposition
DTCWT denoising
Dynamic Weighted Multi-Head Attention
Bidirectional Long Short-Term Memory
Journal
IF:
3.1
Papers:
2.4K
Citations:
3.9K