基于变量相关注意力机制的短期风速预测  

SHORT-TERM WIND SPEED FORECASTING BASED ON VARIABLE CORRELATION ATTENTION MECHANISM

作　　者：王旭光[1] 张可 李潇 白康[1] Wang Xuguang;Zhang Ke;Li Xiao;Bai Kang(Department of Automation,North China Electric Power University,Baoding 071003,China)

机构地区：[1]华北电力大学自动化系,保定071003

出　　处：《太阳能学报》2023年第8期467-476,共10页Acta Energiae Solaris Sinica

基　　金：国家自然科学基金(62076093);河北省机器学习与计算智能重点实验室开放基金(2020-2022-A)。

摘　　要：针对多因素风速预测问题,提出一种基于变量相关注意力机制的短期风速预测模型。该模型通过序列对齐与变量选择模块筛选出与风速相关性强的因素序列并生成其他因素矩阵,采用变分模态分解算法将历史风速序列分解并进一步生成模态分量矩阵。将其他因素矩阵与模态分量矩阵拼接,得到因素矩阵。利用变量相关注意力机制表达因素矩阵与风速序列的相关关系,在此基础上实现风速预测。采用中国南方某风电场一年的实测数据验证了序列对齐与变量选择模块的作用、变分模态分解算法的分解效果以及变量相关注意力机制的特征提取能力。该模型在4个季节的平均绝对误差分别为0.17、0.17、0.13和0.14 m/s,明显优于其他对比模型,充分说明该模型在短期风速预测方面的优越性能。A variable correlation attention mechanism-based short-term wind speed forecasting model is proposed for the multi-factor wind speed forecasting issue.The sequences strongly correlated with the wind speed sequence are screened out based on the sequence alignment and variable selection module,then the matrix made up of the samples of those sequences is generated.Meanwhile,a mode matrix is constructed from the VMD-decomposed historical wind speed sequence.The two matrixes are concatenated to generate the factor matrix.After that,the correlativity between the factor matrix and wind speed sequence is expressed by the variable correlation attention mechanism,and used to forecast the future wind speeds.Effectiveness of the sequence alignment and variable selection module,VMD module and variable correlation attention mechanism is validated on a one-year wind speed data measured from a wind power plant in southern China.The proposed model achieves the MAE values of 0.17,0.17,0.13 and 0.14 m/s in spring,summer,autumn and winter,respectively.Its performance is significantly better than that of the other comparative models,demonstrating the superiority of the proposed model.

关 键 词：风力发电 风速 预测 变分模态分解 最大信息系数 Transformer模型 变量相关注意力机制 

分 类 号：TM614[电气工程—电力系统及自动化]