基于注意力组合模型的风速预测
Wind Speed Prediction Based on Attention-Combined Models

鉴于风速具有非线性、非稳定性和高度随机性等特点，本文提出了CEEMDAN-CNN-Transformer混合模型，旨在提高风速预测的准确性。模型结合CEEMDAN算法对风速序列进行多尺度分解，降低噪声并提取多尺度特征，将复杂的非线性序列转化为更易于建模的子序列；再利用CNN网络提取特征分量的局部时空特征，捕捉风速序列的关键模式；最后通过Transformer模型的自注意力机制捕获序列间的长期依赖关系，得到最终预测风速。实验结果表明，该模型在预测精度上优于传统深度学习模型，有效提升了风速预测的准确性，展现了多模态深度学习架构在处理风速数据非平稳性和高波动性方面的优势。
Considering the characteristics of wind speed, such as nonlinearity, instability, and high randomness, this paper proposes a CEEMDAN-CNN-Transformer hybrid model aimed at enhancing the accuracy of wind speed forecasting. The model employs the CEEMDAN algorithm to perform multi-scale decomposition on wind speed sequences, reducing noise and extracting multi-scale features, thereby transforming complex nonlinear sequences into more easily modeled sub-sequences. Subsequently, a CNN network is utilized to extract the local spatiotemporal features of the feature components, capturing key patterns in wind speed sequences. Finally, the self-attention mechanism of the Transformer model is employed to capture long-term dependencies between sequences, yielding the final predicted wind speed. Experimental results demonstrate that the model outperforms traditional deep learning models in terms of forecasting accuracy, effectively improving the precision of wind speed prediction and showcasing the advantages of multi-modal deep learning architectures in dealing with the non-stationarity and high volatility of wind speed data.