基于不同Bootstrap方法和KELM-BPNN模型的滑坡位移区间预测

Landslide displacement interval prediction based on different Bootstrap methods and KELM-BPNN model

为解决传统滑坡位移点预测模型无法对自身预测结果的可靠程度进行有效描述这一问题,引入区间预测方法,提出一种基于不同Bootstrap方法和KELM-BPNN模型的滑坡位移区间预测模型。该模型以4种常用的Bootstrap方法为基础,首先对由各种外界触发因素与滑坡地表位移的监测信息组成的原始数据集,分别进行B次有放回的等概率随机抽样;然后基于不同Bootstrap方法得到的B个伪数据集,分别训练B个KELM模型对系统误差的方差进行估计,并根据估计结果,训练一个BPNN模型对随机误差的方差进行回归逼近;最终将采用相同Bootstrap方法得到系统误差方差和随机误差方差相结合,构造出在不同置信水平下的滑坡位移预测区间,并通过综合对比分析,提出与实际滑坡变形特征相适宜的位移区间预测模型。以三峡库区内具有阶跃式变形特征的典型堆积层滑坡——白水河滑坡为例,选取ZG93和ZG118两个监测点在2004年7月～2013年12月期间的数据进行研究。结果表明,与传统点预测模型相比,该模型不仅能够提供具有一定精度的点预测结果,还能构造出较为清晰、可靠的位移预测区间将真实的位移曲线完全包裹在内。此外,通过预测区间宽度的实时变化,该模型能够较好地量化与解释滑坡演化过程中外界触发因素的动态变化对滑坡变形造成的不确定性影响,为滑坡灾害的预报预警提供了一种新的思路和方法。

To solve the problem that the traditional landslide displacement point prediction model cannot effectively describe the reliability of the prediction result,a landslide displacement interval prediction model based on different Bootstrap methods and KELM-BPNN model was proposed by introducing the method of interval prediction.In this model,firstly,the original dataset consisting of the monitoring information from various external trigger factors and landslide surface displacement was randomly sampled with an equal probability for B times and then,B dummy datasets were obtained based on different Bootstrap processes.B KELM models were trained to estimate the variance of the system error respectively and consequently,a BPNN model was trained to regress the variance of the random error.Finally,the variances of the system error and the random error,both obtained from the same Bootstrap process,were combined to construct the landslide displacement prediction intervals with different confidence levels.Through comparisons,an optimal displacement interval prediction model fitting in the deformation characteristics of actual landslides was proposed.Baishuihe Landslide,a typical colluvial landslide with step-like behaviour in the area of Three Gorges Reservoir,was taken as an example.The monitoring data of ZG93 and ZG118 from July 2004 to December 2013 were analysed.The results show that,compared with the traditional point prediction model,the developed model not only provides a relatively accurate point prediction result but also constructs a clear and reliable displacement prediction interval to cover the landslide displacement curve completely.In addition,the dynamic variation of the prediction interval width can be used to better quantify and explain the uncertain impact of the dynamic change of external triggering factors on the landslide evolution,which offers a new idea or option for the forecasting and the early warning of landslides.