GPS滑坡位移监测时序分析与组合建模预测

Time Series Analysis and Combined Modeling Prediction of GPS Landslide Displacement Monitoring

GPS位移以时间序列的形式存在,客观真实地记录了滑坡位移的发展演化特征。针对三峡库区滑坡灾害预报预警需求,以白水河滑坡为例,根据GPS时序专业监测分析滑坡的变形特征和诱发因素,并结合机器学习建立组合预测模型,对滑坡位移进行预测。首先,采用小波去噪去除滑坡位移时序的随机噪声,并使用HP滤波器将位移时序分解为周期项和趋势项2个分量。在此基础上,采用核主成分分析法提取滑坡位移主要影响因子特征,建立一种联合粒子群寻优支持向量回归机(PSO-SVR)和二次指数平滑(DES)的混合预测优化模型,分别构造周期项与趋势项训练集实现滑坡位移预测。定量评价结果显示,混合预测模型的预测结果与GPS实际监测值具有很好的一致性,对复杂的滑坡位移曲线拆分并针对性预测,能够有效地提高滑坡位移预测精度。

GPS displacement exists in the form of time series,which objectively and truly records the development and evolution characteristics of landslide displacement.In order to meet the needs of landslide disaster prediction and early warning in the Three Gorges Reservoir area,we take the Baishuihe landslide as an example to analyze the deformation characteristics and induced factors of landslide according to GPS time series professional monitoring,and establish a combined prediction model to achieve landslide displacement prediction by combining with machine learning.Firstly,we removed the random noise of landslide displacement time series by using the wavelet denoising,and decomposed the displacement time series into two components:periodic term and trend term by HP filter.On this basis,we extracted the characteristics of the main influencing factors of landslide displacement through the kernel principal component analysis(KPCA),and established a hybrid prediction optimization model combined with particle swarm optimization support vector regression(PSO-SVR)and quadratic exponential smoothing(DES).Training sets of periodic term and trend term are constructed respectively to realize landslide displacement prediction.The quantitative evaluation results show that the prediction results of the hybrid prediction model proposed in this paper are in good agreement with the actual monitoring values of GPS,and the segmentation of complex landslide displacement curve and targeted prediction can effectively improve the accuracy of landslide displacement prediction.