摘要
针对抽水蓄能机组振动趋势预测中振动信号时间序列非线性、非平稳性极强导致常规预测方法难以进行精确预测的问题,提出变分模态分解(Variational Mode Decomposition,VMD)结合时序模式注意力(Temporal Pattern Attention,TPA)机制改进的门控循环单元(Gated Recurrent Units,GRU)神经网络的抽水蓄能机组振动预测方法。利用VMD算法首先将振动信号序列分解为若干个本征模态分量(IMF),降低时间序列的非平稳性,结合其他特征参数,构建预测输入矩阵。将输入矩阵放入TPA改进的GRU神经网络中训练,利用神经网络强大的非线性特征提取能力,达到精准的预测效果。最后将本方法与GRU-TPA、结合常规注意力机制(AM)的VMD-GRU预测方法进行对比发现,基于TPA改进的VMD-GRU预测方法效果更好,能够更加准确地预测振动信号的时间变化趋势。
Aiming at the difficulty of accurate forecasting of vibration trend of pumped storage units due to the non-linearity and nonsmoothness of the signal,the Variational Mode Decomposition(VMD)combined with the improved Gated Recurrent Units(GRU)neural network with Temporal Pattern Attention(TPA)mechanism is proposed as a vibration forecasting method for pumped storage units.The VMD algorithm is used to decompose the vibration signal sequence into several intrinsic mode components(IMF)firstly to reduce the non-smoothness of the time series,the combine with other feature parameters to construct the prediction input matrix.The input matrix is put into the TPA-improved GRU neural network for training,and the powerful nonlinear feature extraction ability of the neural network is utilized to achieve accurate forecasting results.Finally,comparing the new method with GRU-TPA and VMDGRU-AM(normal attention mechanism)forecasting method,it is found that the TPA-improved VMD-GRU forecasting method is more effective and accurate.
作者
杨雄
聂赛
章志平
刘泽
卢俊琦
汤络翔
冯陈
张玉全
郑源
YANG Xiong;NIE Sai;ZHANG Zhiping;LIU Ze;LU Junqi;TANG Luoxiang;FENG Chen;ZHANG Yuquan;ZHENG Yuan(Jiangxi Hongping Pumped Storage Co.Ltd.,Jingan,330600,China;College of Energy and Electrical Engineering,Hohai University,Nanjing 211000,China)
出处
《水电与抽水蓄能》
2023年第1期31-38,共8页
Hydropower and Pumped Storage
基金
国网新源控制有限公司科技项目(项目编号:SGXYKJ-2022-034)。
关键词
时序模式注意力机制
变分模态分解
门控循环单元
抽水蓄能机组
振动信号
预测
temporal pattern attention mechanism
variational mode decomposition
gated recurrent units
pumped storage units
vibration signal
forecating