论文部分内容阅读
支持向量回归机(Support Vector Regression,SVR)在滑坡位移预测研究中已得到广泛应用,但SVR具有模型可解释性差的缺陷,即无法直接获得并筛选最佳预测变量,从而影响预测精度。为此,将较广泛应用于评价神经网络模型变量影响大小的平均影响值(Mean Impact Value,MIV)方法与SVR模型相结合,实现基于SVR-MIV的变量筛选,该方法不但能对所有预测模型初始变量影响大小进行排序,还可以进一步结合反向逐变量剔除分析实现变量筛选。为验证该方法的有效性,选择三峡库区两类典型水库滑坡代表的累积位移监测数据,在采用移动平均法将位移分解为趋势项和波动项的基础上,重点针对波动项位移,选择包括降雨及库水位变动特征在内的12项初始变量,采用SVR-MIV方法进行变量筛选分析。结果表明,该方法筛选出的变量理论上符合对应滑坡变形影响机理分析结论,且可以提高滑坡位移实际预测精度。
Support Vector Regression (SVR) has been widely used in the study of landslide displacement prediction. However, SVR has the disadvantage of poor interpretability of the model, that is, the best predictor can not be directly obtained and screened, which affects the prediction accuracy. For this reason, the mean impact value (MIV) method which is widely used to evaluate the influence of neural network model variables is combined with the SVR model to implement the variable screening based on SVR-MIV. This method can not only predict all the prediction models The size of the initial variable to sort, you can also further combined with the reverse by the variable elimination analysis to achieve variable selection. In order to verify the effectiveness of this method, the cumulative displacement monitoring data of two types of typical reservoir landslides in the Three Gorges Reservoir Area are selected. Based on the moving average method, the displacement is decomposed into trend items and fluctuating items. Focusing on the displacement of fluctuating items, 12 initial variables, including rainfall and the change of reservoir water level, were used to carry out variable screening analysis using SVR-MIV method. The results show that the variables screened by this method are in accordance with the theoretical analysis of the mechanism of landslide deformation, and can improve the actual prediction accuracy of landslide displacement.