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 〇ptimized by PS〇 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-supportvectormachine,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supportvectormachine) 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(waveletanalysis-backpropagation,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 waterleve1 forlandslide stability analysis. For the problem that the evolution process of groundwaterleve1 in reservoirlandslide is a highly non-1inear and non-stationary time series affected by many factors, to predictlandslide groundwaterleve1 time series, a coupling mode1 based on phase space reconstruction and wavelet analysis-support vector machine(WA-PSVM) optimized by particle swarm optimization is proposed. Firstly, the groundwaterleve1 time series was decomposed into severa1 different frequency compo- nents to transform the non-stationary groundwaterleve1 time series into stationary time series. Secondly, the PSVM mode1 was established for each component prediction based on the phase-space reconstruction. Atlast, the fina1 prediction result was ob- tained by adding the predicted values of a11 frequency components. Taking daily average groundwaterleve1 time series of STK- 1 hydro1ogy hole on Sanzhouxi Landslide in the Three Gorges Reservoir Area for example, the influencing factors of landslide groundwaterleve1 lluctuation were analyzed and WA-PSVM mode1 was used to predict the STK-1 groundwaterleve1 values. Meanwhile, the single PSVM mode1 and wavelet analysis-back propagation neura1 network(WA-BP) mode1 were also used for groundwaterleve1 prediction. The results show that reservoir waterleve1 fluctuation and rainfa11 are the main factors of ground- waterleve1 fluctuation in the reservoirlandslideleading edge. We also find that the root-mean-square error(RMSE) of the pro- posed mode1 for groundwaterleve1 time series prediction in STK-1 hydro1ogy holes is 0.073 m, the goodness of fit is 0.966, re- spectively. The prediction accuracy of WA-PSVM mode1 is higher than the single PSVM mode1 and WA-BP mode1. What is more, WA-PSVM modelsolves the non-1inear and non-stationary problem. WA-PSVM mode1 also has a high operating effi-