Study Of Stepwise Displacement Prediction For Reservoir Landslides Based On Ensemble Empirical Mode Decomposition

The reservoir landslide is one of the most important engineering problems in water conservancy and hydropower projects.The Three Gorges reservoir area is characterized by high mountains,steep slopes,and deep valleys.The engineering geological conditions are complex and diverse.In particular,the annual rainfall in the rainy season is large and typical of landslide accumulation.Disaster-hit area.Due to the seasonality of rainfall and the periodicity of reservoir water level scheduling,the cumulative displacement versus time curve of many reservoir landslides in the Three Gorges reservoir area has a upstairs stepwise pattern,thus this type of reservoir landslide is called “stepwise landslide”.On the one hand,in-depth study of the inducing mechanism of stepwise landslides and the deformation and destruction characteristics of landslides after reservoir impoundment,as well as the mining of monitoring time series can make more effective use of data;on the other hand,choosing what kind of prediction method is also the key to the forecasting results,which plays an important role in predicting the landslide displacement trend in advance.This paper systematically summarizes the types and mechanisms of reservoir landslide,and discusses the advantages and disadvantages of the current methods for stepwise displacement decomposition based on the theory of displacement component response model for stepwise landslides.The mechanism of groundwater action in the process of lifting and lowering is selected,and typical stepwise landslides in the Three Gorges reservoir area are selected to analyze the deformation evolution characteristics during rainfall and reservoir water operation cycles.The causal relationship and lags of impact between inducing factors(monthly rainfall and reservoir water level)and the displacement is testedby Granger test.After analyzing the shortcomings of the current displacement decomposition methods,the ensemble empirical mode decomposition(EEMD)method is used to decompose the stepwise displacements of landslides into trending displacement and fluctuating displacement in the Three Gorges Reservoir area;EEMD and the t-test methods are used to reconstruct the time series for rainfall and reservoir level.These two time series are reconstructed into high frequency components and low frequency components,and three nonlinear methods are adopted to predict the displacement of the fluctuating displacement,the polynomial fitting method is adopted to predict the trending displacement,after which the total predicted displacement is obtained.The main research results obtained are as follows:(1)Combining with previous work,the main internal control factors of the reservoir slope and the role of external inducing factors are explained,and the main mechanisms of various factors are analyzed.On this basis,using typical stepwise landslides,i.e.Baishuihe landslide and Bazimen landslide,based on Eviews software totest the Granger causality between inducing factors time series with 1~2 lags.It turns out that there is a strong causal relationship between the two factors(monthly average rainfall and the monthly average reservoir level)and the deformation rate of the stepwise reservoir landslide,and the impact lag period of rainfall on the two landslides may reach two months or even longer,the effect lag period of reservoir water level change on deformation is one month.This lag period also depends on the monitoring interval.Under the condition that the monitoring interval is smaller,the lag period could be closer to the actual lags.(2)Collecting historical monitoring data of multiple landslides in the Three Gorges reservoir area.According to the displacement component response theory,withensemble empirical mode decomposition that can highlight the characteristics of local fluctuations,the stepwise displacement of the reservoir landslide is decomposedinto trending displacement and the fluctuating displacement,the trending displacement change year rising trend,and the later,the trending displacement change more slowly,and the overall actual trend of monitoring displacement is the same,but the magnitude of the fluctuating displacement are greatly influenced by the intensity of the external inducing factors.(3)By empirical modal decomposition of the set,the displacement component response model is used as a theoretical basis to perform empirical mode decomposition of the precipitating factors.Without considering the trend terms,multiple high-frequency and low-frequency component reconstructions of the multiple intrinsic mode functions(IMFs)are performed based on the t-test.The high-frequency components can be intenseinducing factors.For example,multiple heavy rainfalls in short-term will play a major role in promoting the deformation of landslides,and may be the major contribution to the occurrence of surge increase of displacement;low-frequency components reflect the slow changes in inducing factors with low intensity and low frequency,which is also one of the important factors for landslide deformation.(4)Combining the results of time series decomposition and reconstruction for the 6 stepwise displacement landslides in this paper,using grey relational analysis to select the factors that have a largest degree of correlationas predominant factors from various inducing factors of the same type.It can be found that there are 4 landslides with high frequency rainfall as the best advantage factor in the rainfall factor group,accounting for 4/6,and 2 low frequency rainfalls as dominant factors,accounting for 2/6,high frequency rainfall and low-frequency rainfall accounting for 1.It is proved that the rain component factors obtained by EEMD decomposition and reconstruction are more correlated than other rainfall factors.In the reservoir water level factor group,the monthly variation of the water level in the reservoir accounts for a large proportion,which to some extent proves that the change rate of the reservoir water has a great influence on the deformation.The mining resultof reservoir level by the empirical mode decomposition and reconstruction method is unsatisfactory,but successful on constructing a new rainfall component with a higher correlation with the fluctuatingdisplacement.(5)Taking inducing factors with highest grey relationship degree as the input variable into the nonlinear predictive models and comparing the particle swarm optimized SVR method,RBF neural network and Elman neural network to predict the fluctuating displacement.The results by the Elman neural network and RBF neural network are more consistent with the actual value of the displacement at local surge fluctuation.But from the accuracy index,the three methods can reach the accuracy of the actual demand,and SVR method is better;the trending displacement can be predicted by polynomial fitting,and the fitting accuracy R-Square is 1;the total prediction can be obtained by superimposing the two displacement prediction values.The error range ofprediction is very small.