基于集合经验模态分解与支持向量机回归的位移预测方法:以三峡库区滑坡为例

Displacement prediction method based on ensemble empirical mode decomposition and support vector machine regression——a case of landslides in Three Gorges Reservoir area

三峡库区滑坡地表位移-时间曲线多呈台阶型特征。基于位移响应成分模型的滑坡位移预测方法是该类滑坡位移预测的主要方法之一。针对目前水库滑坡波动项位移预测工作中尚未考虑主要诱发因素的高频成分与低频成分的问题,提出了基于时间序列集合经验模态分解(EEMD)与重构的粒子群优化-支持向量机回归(PSO-SVR)位移预测方法。以白水河滑坡2003年7月至2013年3月117个地表位移数据为例,采用EEMD法将位移时间序列分解为趋势项位移和波动项位移,该趋势项位移用最小二乘法的二次多项式拟合预测;根据EEMD和t检验法,确定高频降雨量、低频降雨量、高频库水位和低频库水位,结合其他常用因素,采用灰色关联分析确定白水河滑坡影响波动项位移的优势因素为高频降雨量和月间库水位变化,基于优势因素建立PSO-SVR模型预测波动项位移。结果表明,总预测值的平均相对误差为0.009 8,方差比为0.023 9,小误差概率为1,预测效果较好。利用该方法对三峡库区其他5个台阶型滑坡进行了预测,预测位移与实测位移较吻合,进一步证明了该方法的有效性,对同类滑坡的预测预报具有一定的借鉴意义。

Many landslides in Three Gorges Reservoir area are featured by stepwise increasing surface displacement along the time. Responding composition model is one of the main methods for displacement prediction of landslides. Currently, high-frequency and low-frequency components of inducing factors are usually ignored. A method based on reconstruction of time series by ensemble empirical mode decomposition（EEMD） and particle swarm optimization based support vector machine regression（PSO-SVR） prediction for displacement is proposed. The typical Baishuihe landslide in Three Gorges Reservoir is taken as an example. Firstly, the monitored surface displacement time series from July 2003 to March 2013 is decomposed into trend and fluctuant components by EEMD. The trend component can be predicted using quadratic polynomial equation fitted by the least square method. With EEMD and t-test methods, rainfalls and reservoir levels time series are reconstructed into high-frequency rainfalls, low-frequency rainfalls, high-frequency reservoir levels and low-frequency reservoir levels, respectively. Combined with other common factors, high-frequency rainfalls and monthly variations of reservoir levels are selected as predominant factors for fluctuant displacement components by method of gray relational analysis（GRA）. Finally, PSO-SVR is utilized for prediction purpose. Application results show that average relative error is 0.009 8, variance ratio is 0.023 9 and small error probability is 1, which demonstrate preferable effect of the proposed method. For verification and testing, 5 other typical landslides in Three Gorges Reservoir area are presented to test the effectiveness of our method, which can provide references for similar landslides.