基于高维时空张量CP分解的风速监测缺失数据恢复  

Wind speed data recovery based on CP decomposition of a higher-dimensional spatial-temporal tensor

作　　者：许学方 胡诗婷 时培明[1] 李瑞雄 李志 XU Xuefang;HU Shiting;SHI Peiming;LI Ruixiong;LI Zhi(School of Electrical Engineering,Yanshan University,Qinhuangdao 066004,China;School of Energy and Power Engineering,Xi’an Jiaotong University,Xi’an 710049,China;Jiangsu Goldwind Science&Technology Co.,Ltd.,Yancheng 224100,China)

机构地区：[1]燕山大学电气工程学院,河北秦皇岛066004 [2]西安交通大学能源与动力工程学院,西安710049 [3]江苏金风科技股份有限公司,江苏盐城224100

出　　处：《振动与冲击》2023年第11期163-169,共7页Journal of Vibration and Shock

基　　金：秦皇岛市科学技术研究与发展计划(202101A345);河北省自然科学基金青年项目(E2022203093);国家自然科学基金项目(61973262);燕山大学基础创新科研培育项目(2021LGQN022)。

摘　　要：“大数据”时代的到来为风机健康监测带来了新机遇,风机往往运行在极端恶劣环境下,监测数据中夹杂了大量缺失值,数据质量无法保障,进而会制定有误的运维指导策略。为保证风速监测数据质量,提出了基于高维时空张量CP分解的风速监测数据缺失值恢复方法。构建包含时空信息的四阶张量,利用CP分解将张量分解为多个因子矩阵,通过加权张量将恢复缺失数据转化为求解目标函数最小值,根据因子矩阵重构张量,从而获得缺失处原始信息值。利用提出方法与GPR、GRU、LSTM、SWLSTM等传统方法对某风电场的缺失数据进行恢复,结果表明,相比传统方法,提出方法的R^(2)最接近1,MAE等误差指标均为最小,具有最高拟合度,从而验证了该方法的有效性。The advent of“big data”era brings new opportunities for wind turbine health condition monitoring.However,wind turbines often operate in harsh environment,and thus monitoring data is mixed with a large number of missing values which reduce data quality.As a result,wrong operation and maintenance strategies will be developed based on these low-quality data.A method based on CP decomposition of spatial-temporal tensor was proposed to recover missing data to improve the quality of monitoring data.Firstly,a four-dimension tensor containing spatial-temporal information was constructed.Then CP decomposition was applied to decompose the estimating tensor into factor matrices.Afterwards,a weighted tensor was used to translate the recovery issue into the solving of a minimization function.Finally,the tensor was reconstructed according to the factor matrices,and the original value of the missing data was obtained.The actual monitoring data of a wind farm was used to recover the missing values by different methods including GPR,GRU,LSTM,SWLSTM.The results show that compared with that of traditional methods,the R^(2) of proposed method is closest to 1 and the other recovery error such as MAE are minimum,which has higher fitting degree with the real data.Therefore,the case verifies the effectiveness of the proposed method.

关 键 词：风机健康监测 数据质量 缺失值数据 张量分解 数据恢复 

分 类 号：TH165.3[机械工程—机械制造及自动化]