摘要
针对变速变载条件下风电机组监测参量中包含大量缺失数据导致基于深度学习的状态监测模型预警精度低的难题,提出一种多头注意力双向长短时记忆网络缺失数据修复方法(Multi-headed attention bidirectional long and short term memory network,MA-BiLSTM)。所提方法利用多头注意力机制抑制复杂工况条件下变速变载对神经网络特征提取时的干扰,采用跨层连接残差单元增加模型的特征提取能力,充分学习已有监测数据的隐藏特征以及多源参量间的关联关系;采用双向长短时记忆网络同时对风电机组监测数据的复杂变化规律进行学习,实现监测参量中缺失数据的预测修复。实例应用结果表明所提多头注意力双向长短时记忆网络能够抑制复杂工况条件下的变速变载干扰,实现单变量或多变量中缺失数据的预测修复,有效提升风电机组状态监测精度。
To address the problem of low early warning accuracy of condition monitoring models due to a large number of missing data in wind turbine monitoring parameters under variable speed and variable load,multi-headed attention bidirectional long and short-term memory networks(MA-BiLSTM)is proposed for repair those data.The multi-headed attention machine is used to suppress the interference of variable loads on the neural network feature extraction under complex working conditions.In addition,the model feature extraction ability is increased by constructing a cross-layer of residual units,and the hidden features of existing monitoring data and the correlation between multi-source parameters are fully learned.The Bi-LSTM cells are used to simultaneously learn the law of the monitoring data of wind turbines to achieve the prediction and repair of incomplete data.The application results show that the proposed MA-BiLSTM networks can suppress the multivariate load disturbance under complex working conditions and realize the repair of incomplete data for improving fault detection accuracy of wind turbines.
作者
余晓霞
汤宝平
王伟影
吴宣勇
李彪
YU Xiaoxia;TANG Baoping;WANG Weiying;WU Xuanyong;LI Biao(The State Key Laboratory of Mechanical Transmission,Chongqing University,Chongqing 400044;CSSC Longjiang Guanghan Gas Turbine Co.,Ltd.,Harbin 150078)
出处
《机械工程学报》
EI
CAS
CSCD
北大核心
2023年第14期1-9,共9页
Journal of Mechanical Engineering
基金
国家重点研发计划(2020YFB1709800)
重庆市自然科学重点基金(cstc2019jcyj-zdxm X0026)资助项目
关键词
风电机组
缺失数据预测
多头注意力机制
双向长短时记忆网络
wind turbine
deteriorated data repair
multi-headed attention mechanism
bi-directional long and short term memory network
