基于VMD-PE-CNN的混凝土坝变形预测模型

VMD-PE-CNN-based deformation prediction model of concrete dam

为了进一步提高混凝土坝变形预测精度,基于“先分解再重构”的思想,将变分模态分解(VMD)、排列熵(PE)与卷积神经网络(CNN)相结合,提出了一种混凝土坝变形预测模型。通过VMD和计算模态分解余量的PE将原始实测变形时间序列数据自适应地分解为一系列具有不同频域尺度特征的模态分量,然后将每个模态分量作为单独的子序列,采用CNN直接对各子序列进行时域建模并预测,最后将各个子序列的预测值叠加重构得到最终的大坝变形预测值。实测数据计算结果表明:采用计算模态分解余量PE的方法可以得到最优的模态分量个数,实现实测数据的最优分解;较之于CNN和LSTM模型,VMD-PE-CNN模型在测试数据上的均方根误差分别降低了61.8%和65.5%,显示出更强的预测能力。

In order to further increase the prediction accuracy of the deformation of concrete dam,a concrete dam deformation prediction model is proposed herein by combining variational mode decomposition(VMD)and permutation entropy(PE)with convolution neural network(CNN)on the basis of the idea of decomposing at first and then reconstructing.Through VMD and calculating the permutation entropy(PE)of modal decomposition residual,the originally measured deformation time series data are adaptively decomposed into a series of modal components with the characteristics of different frequency scales,and then each modal component is taken as an individual subsequence to directly make the time domain modeling and prediction for all of them.Finally,the final predicted value of dam deformation is obtained by means of superimposing and reconstructing the predicted values of all the subsequences.The calculation results of the measured data show that the optimal number of the modal components can be obtained by calculating the PE of modal decomposition residual to realize the optimal decomposition of the measured data.Compared with the CNN model and the LSTM model,the root mean square errors of the VMD-PE-CNN model on the testing data are reduced by 61.8%and 65.5%respectively,thus the model exhibits a stronger prediction capacity.