基于多变量自优化动态神经网络的“阶跃型”滑坡变形预测  

Step-type Landslide Deformation Prediction Based on Multivariable Self-optimizing Dynamic Neural Network

作　　者：徐志华 杨旭 孙钱程 何钰铭 张国栋 叶义成 XU Zhihua;YANG Xu;SUN Qiancheng;HE Yuming;ZHANG Guodong;YE Yicheng(National Field Observation and Research Station of Landslides in Three Gorges Reservoir Area of Yangtze River,Yichang 443002,China;Hydrology and Engineering Geology Institute in Hubei Geological Bureau,Yichang 443002,China)

机构地区：[1]湖北长江三峡滑坡国家野外科学观测研究站,湖北宜昌443002 [2]湖北省地质局水文地质工程地质大队,湖北宜昌443002

出　　处：《金属矿山》2024年第3期74-82,共9页Metal Mine

基　　金：国家自然科学基金项目(编号:51909136)。

摘　　要：传统累计变形预测方法在曲线结构分解和表征模型选择上具有多样性,由此引起了工作量大、预测精度低以及预测方法适用对象较局限等问题,为此考虑降雨量、库水位、库水位变化对滑坡累计变形的影响,基于非线性自回归模型建立了多变量自优化动态神经网络,并将其应用在三峡库区典型的“阶跃型”滑坡——白家包滑坡累计位移预测中。通过对滑坡变形累计曲线时间序列的分析,采用神经网络方法对全曲线模型进行求解,形成了非线性自回归神经网络模型,利用多种群遗传算法对神经网络的参数和结构进行优化训练,并将适应度函数均方误差作为预测模型误差偏离标准。结果表明:所提出的自优化动态神经网络对滑坡多个测点的累计位移拟合精度高,误差可控制在1%左右,预测过程减少了主观因素引起的误差,考虑了滑坡发展过程的动态性,可为“阶跃型”滑坡累计位移的实时预测提供参考。The traditional cumulative deformation prediction methods are diverse in curve structure decomposition meth-ods and characterization model selection,which brings about the problems of large workload of prediction methods,low predic-tion accuracy,and restricted applicability objects.To address the above problems,a multivariate self-optimizing dynamic neural network is established based on a nonlinear autoregressive model considering the effects of rainfall,reservoir water level and reservoir water level changes on the cumulative deformation of landslides.The neural network is applied to the prediction of the cumulative displacement of the typical stepped landslide in the Baijiabao landslide of the Three Gorges Reservoir.By analyzing the time series of the cumulative curve of landslide deformation,a nonlinear autoregressive neural network is composed by using a neural network to solve the full curve model.The parameters and structure of the neural network are optimized and trained u-sing multiple swarm genetic algorithms,and the mean square error of the fitness function is used as the prediction model error deviation criterion.The results show that the self-optimized dynamic neural network proposed in this paper has high accuracy in fitting the cumulative displacement of multiple measurement points of landslides.Its error can be controlled within 1%,and the prediction process reduces the error caused by subjective factors.The neural network can provide a reference for the prediction of cumulative displacement of such step-type landslides.

关 键 词：累计位移预测 “阶跃型”滑坡 多因素影响 多变量自优化 动态神经网络 

分 类 号：TD325[矿业工程—矿井建设]