基于相空间重构和小波分析-粒子群向量机的滑坡地下水位预测

Landslide Groundwater Level Time Series Prediction Based on Phase Space Reconstruction and Wavelet Analysis-Support Vector Machine Optimized by PSO Algorithm

预测滑坡地下水位的动态演变过程对滑坡稳定性分析具有重要意义,三峡库区库岸滑坡地下水位时间序列受多种因素影响,呈现出高度非线性非平稳的特征.为对其进行预测,提出一种基于相空间重构的小波分析-粒子群优化支持向量机(wavelet analysis-support vector machine,简称WA-PSVM)模型.该模型引入小波变换法对地下水位序列进行时频分解,将非平稳的地下水位序列转变为多个不同分辨率尺度下的较平稳的地下水位子序列;然后重构各子序列的相空间,再利用PSVM(全称support vector machine)模型对地下水位各子序列进行预测,最后将各子序列预测值相加得到最终预测结果.以三峡库区三舟溪滑坡前缘STK-1水文孔日平均地下水位序列为例,首先分析滑坡前缘地下水位变化的影响因素,再将WA-PSVM模型应用于地下水位预测,并与单独PSVM模型和小波分析-BP网络模型(wavelet analysis-back propagation,简称WA-BP)作对比.结果表明:滑坡前缘地下水位受降雨和库水位影响较大,利用WA-PSVM模型对STK-1水文孔地下水位进行预测的均方根误差为0.073m、拟合优度为0.966,WA-PSVM模型预测精度高于单独PSVM模型和WA-BP模型.WA-PSVM模型解决了地下水位序列非线性非平稳的问题,在不考虑影响因素的情况下能获得满意的预测效果,具有较高的建模效率和较强的实用性.

It is of great significance to predict the dynamic evolution process of landslide underground water level for landslide stability analysis. For the problem that the evolution process of groundwater level in reservoir landslide is a highly non-linear and non-stationary time series affected by many factors, to predict landslide groundwater level time series, a coupling model based on phase space reconstruction and wavelet analysis-support vector machine （WA-PSVM） optimized by particle swarm optimization is proposed. Firstly, the groundwater level time series was decomposed into several different frequency compo- nents to transform the non-stationary groundwater level time series into stationary time series. Secondly, the PSVM model was established for each component prediction based on the phase-space reconstruction. At last, the final prediction result was ob- tained by adding the predicted values of all frequency components. Taking daily average groundwater level time series of STK- 1 hydrology hole on Sanzhouxi Landslide in the Three Gorges Reservoir Area for example, the influencing factors of landslide groundwater level fluctuation were analyzed and WA-PSVM model was used to predict the STK-1 groundwater level values. Meanwhile, the single PSVM model and wavelet analysis-back propagation neural network （WA-BP） model were also used for groundwater level prediction. The results show that reservoir water level fluctuation and rainfall are the main factors of ground- water level fluctuation in the reservoir landslide leading edge. We also find that the root-mean-square error （RMSE） of the proposed model for groundwater level time series prediction in STK-1 hydrology holes is 0.073 m, the goodness of fit is 0.966, respectively. The prediction accuracy of WA-PSVM model is higher than the single PSVM model and WA-BP model. What is more, WA-PSVM model solves the non-linear and non-stationary problem. WA-PSVM model also has a high operating efficiency and strong applicability without considering the impacts of reservoir water level fluctuation and seasonal rainfall.