摘要
为进一步挖掘风速时序规律与多气象变量的关联信息,提出一种融合“特征筛选-二次分解-时空特征学习”的深度学习模型来提升短期风速预测精度。首先,从原始数据集提取风速序列与气象变量,通过相关性筛选与风速高相关的气象变量,剔除冗余特征;其次,提出了称之为TVF-EMD-VMD的二次分解方法,对风速序列进行分解,为后续神经网络提供了更高质量的输入数据;最后,将二次分解后的所有子序列与筛选后的气象变量整合,通过CNN提取多变量间的空间关联特征,输入BiLSTM学习时序演化规律,输出得到最终风速预测值。通过对比模型得出本文方法相较一次分解和已有的二次分解在预测精度上有所提升;将其应用于内蒙古风电场风速序列中,表明该方法具有较好的预测效果。
To further explore the temporal patterns of wind speed and the correlation information among multiple meteorological variables,a deep learning model integrating'feature selection-secondary decomposition-spatiotemporal feature learning'is proposed to improve short-term wind speed prediction accuracy.Firstly,wind speed sequences and meteorological variables are extracted from the original dataset,and meteorological variables highly correlated with wind speed are selected through correlation screening,eliminating redundant features.Secondly,a secondary decomposition method called TVF-EMDVMD is proposed to decompose the wind speed sequences,providing higher-quality input data for subsequent neural networks.Finally,all sub-sequences after secondary decomposition and the selected meteorological variables are integrated,and CNN is used to extract spatial correlation features among multiple variables,which are then input into BiLSTM to learn temporal evolution patterns,outputting the final wind speed predictions.Comparative models show that the proposed method improves prediction accuracy compared to single decomposition and existing secondary decomposition methods.Its application to wind speed sequences in Inner Mongolia wind farms demonstrates that the method has good predictive performance.
作者
马腾
施三支
MA Teng;SHI Sanzhi(School of Mathematics and Statistics,Changchun University of Science and Technology,Changchun 130022)
出处
《长春理工大学学报(自然科学版)》
2025年第5期121-131,共11页
Journal of Changchun University of Science and Technology(Natural Science Edition)
基金
国家自然科学基金(11601039)
吉林省教育厅项目(JJKH20210809KJ,JJKH20230791KJ)。
