应用倒传递类神经网络模拟预测山棕寮地滑地位移之研究

Displacement Predicted for the Backpropagation Neural Network Applying in Shonzongliao Landslide Area

地滑为陡坡地最具破坏之灾害形态之一,藉由实施地滑地监测与评估,可以协助政府拟订适当的管理对策。研究以台东县池上乡山棕寮地滑地为例,利用倒传递类神经网络(BPNN)具有建构高复杂且非线性关系方式预测坡面位移变化。研究所发展之倒传递类神经网络分析系应用MATLAB程序之Levenberg-Marquardt算法求解。网络输入系以直接关系位移之7个物理因子为变数,建构最佳之4层网络。根据监测资料共取12次台风暴雨事件,作为网络训练及模拟测试,并分别以6批次、8批次、10批次等3种情况网络训练对应模拟测试。研究结果显示所需监测台风暴雨事件至少8批次,即可达到13.1%预测位移误差。其误差精度可作为中危害度以上地滑地执行后续监测管理值与利用管理之参考。

Landslides has become one disaster type of the most serious destroy on the steep lands. The way to monitoring and assessment for landslide area can help government agencies to select suitable management and plan mitigation in unstable landslide areas. This research presents a case study of landslide moni- toring and assessment at Shonzongliao Landslide area, Township, Taitung County, attempt to predict slope movements using backpropagation neural network （BPNN）, as well as use powerful tools to model and investigate various complex and non-linear phenomena. The BPNN can performed calculation to use MATLAB program with the Levenberg--Marquardt algorithm. The data from the case study are used to train and test the developed model, to enable prediction of the magnitude ground movements with the seven input helpful of variables that have regard to direct physical significance. According to monitoring date picked 12 set event of typhoon or rainstorm. 3 situation of train versus simulation and test is introduced in the network architecture apart from 6,8,10 set-batch. The developed BPNN optimal model by 8 set-batch demonstrates that have good potential accurately for predicting error 13.1% to thee slope movement. This promising result can offer the reference of building of monitoring and utilizing management value at the landslide areas.