基于EMD分解法的大坝变形预测模型及应用

针对传统模型对脉动时间序列的预测效果较差的情况,结合经验模态分解(EMD)、相关向量机(RVM)理论以及改进粒子群算法(IPSO)的优点,提出一种基于EMD分解法的大坝变形预测模型。首先利用EMD分解法对大坝变形时间序列进行分解和重构,使非平稳的大坝变形时间序列平稳化,再以RVM理论为基础进行预测,核函数选用高斯核函数,并采用改进粒子群算法(IPSO)进行寻优,最终建立EMD-RVM(IPSO)大坝变形预测模型。通过实例计算得到,SVM、RVM和EMDRVM(IPSO)三种模型的平均残差分别为5.29 mm、3.13 mm、0.97 mm,并且EMD-RVM(IPSO)模型的预测值误差均控制在5%以内。这证明EMD分解法对非平稳时间序列的预处理可有效提高预测精度,相比于标准SVM模型和RVM模型,EMD-RVM(IPSO)模型的预测精度更高,且结构稀疏度更好,在实际工程中具有一定的可行性。

EMD decomposition method-based dam deformation prediction model and its application

Aiming at the situation of that the prediction effect from the conventional model on the fluctuation time series is poor，a EMD decomposition method-based dam deformation prediction model is put forward herein in combination with the merits of theempirical mode decomposition ( EMD) and the theory of the relevance vector machine ( ＲVM) as well as the improved particleswarm optimization ( IPSO) algorithm. At first，the dam deformation time series is decomposed and reconstructed with the EMDdecomposition method to make the non-stationary dam deformation time series stationarized，and then the prediction is carried outbased on the ＲVM theory，for which the gauss function is used as the kernel function and the improved particle swarm optimiza-tion ( IPSO) algorithm is adopted for the optimization，thus the EMD-ＲVM ( IPSO) prediction model for dam deformation is es-tablished finally. It is obtained through the relevant actual case that the mean residuals of the SVM，ＲVM and EMD-ＲVM ( IPSO) models are 5. 29 mm，3. 13 mm and 0. 97 mm respectively，while all the errors of the predicted values from the EMD-ＲVM( IP-SO) model are controlled within the range of 5% . It is demonstrated that the pre-processing of the non-stationary time series madeby the EMD decomposition method can effectively enhance the prediction accuracy，thus if compared with the standardized SVMmodel and ＲVM model，the prediction accuracy of the EMD-ＲVM( IPSO) is higher with better structure sparsity，and then has acertain feasibility in the engineering practice concerned.