| Landslide disasters can lead to a large number of casualties and economic losses.In southwest China,there are many mountains and abundant rainfall,resulting in frequent landslides.Under the influence of seasonal precipitation and changes in reservoir water level,the cumulative displacement time curve of some landslides exhibits obvious "step type" dynamic deformation characteristics,with two states of accelerated deformation and stable slow deformation alternating.In order to effectively carry out landslide prevention and early warning,various localities have carried out mass measurement and prevention work,and a large number of landslide monitoring equipment have been deployed at potential landslide points,providing an effective data source for landslide early warning research.However,due to the complex and sudden triggering mechanism of landslides,landslide displacement prediction has always been a hot and difficult issue.This thesis focuses on the research of step type landslide displacement prediction.In order to solve the problem that it is difficult to accurately predict the displacement due to the step type growth of the landslide displacement curve,the displacement component splitting method is used to perform partial prediction.Firstly,the methods of eliminating abnormal values and filling in missing values of landslide displacement data are explored;Then,the exponential smoothing method is used to predict the cumulative displacement of a step type landslide by dividing it into trend terms and periodic terms,respectively;Then,Fourier curve fitting is used to predict the trend term displacement,and machine learning methods are used to analyze the response relationship between the impact factors and the landslide periodic term displacement,establishing a periodic term displacement prediction model;Finally,parameter optimization methods are used to optimize the parameters of the prediction model.The main research results are as follows:(1)Landslide data preprocessing based on median method and set empirical mode decompositionAccording to the characteristics of step type landslide displacement time series data,statistical analysis method and signal processing method are combined to achieve landslide data preprocessing.Firstly,starting from the complete set of data,based on the statistical analysis characteristics of the data,the abnormal outliers of landslide data are preliminarily eliminated;Then,the set empirical mode decomposition is used to split the step type landslide time series data into different feature scales for fine detection of local outliers;Finally,fill in the data after eliminating abnormal values.(2)Selection of optimal landslide impact factors based on recursive feature eliminationBased on the deformation characteristics of step type landslides,the landslide displacement is split and the SVR-RFE method is used to screen the impact factors of periodic term displacement.The exponential smoothing method is used to divide the cumulative displacement of landslides into trend terms and periodic terms,and Fourier curve is used to fit and predict the trend term displacement that is only affected by internal geological conditions;For periodic term displacements affected by external factors,fully explore the response relationship between the impact factors and displacements,construct an initial set of impact factors,and filter out impact factors that are highly correlated with periodic term displacements through a recursive feature elimination algorithm based on support vector regression models(SVR-RFE),making necessary preparations for building a prediction model.By substituting the impact factor screening results of this thesis and the impact factors extracted by the other two methods into the displacement prediction model,it has been verified that the impact factors screened by this method can more effectively improve the accuracy of the prediction model.(3)Parameter Optimization of Displacement Prediction Model Based on Particle Swarm Optimization AlgorithmUsing the influence factors selected based on the SVR-RFE feature selection algorithm as input variables,a support vector machine regression model is established to predict the periodic term displacement.At the same time,the particle swarm optimization algorithm is used to select the optimal parameters of the SVR model,improving the accuracy of the periodic term prediction model.Compared with the model based on grid search and genetic algorithm optimization,the prediction effect of the model in this thesis is better during the period of drastic changes in the displacement of stepwise landslides.Finally,the periodic term displacement prediction results are superimposed on the trend term displacement prediction results to obtain the final landslide cumulative displacement prediction value,with a goodness of fit R ^ 2 of 0.999,a root mean square error RMSE of 9.974 mm,and an average absolute error MAE of 7.037 mm. |
