Reliability Analysis Of Heterogeneous Reservoir Slopes Using Machine Learning Algorithms

Landslide disasters caused by rainfall and changes in reservoir water level occur frequently.Reservoir landslide accidents have caused huge losses to people’s lives and properties.The change of seepage field and soil pore water pressure is mainly due to the change of rainfall and reservoir water level,which in turn causes the increase of dynamic and hydrostatic loads acting on the bank slope and the decrease of soil shear strength.Therefore,it is of great practical significance to carry out research on the stability and reliability of the reservoir slope under the conditions of rainfall and reservoir water level changes for ensuring the safety of reservoir operation.At present,Monte Carlo simulation method is often used to analyze the reliability of complex reservoir slopes,but the calculation of this method is very time-consuming,especially for the reliability of low-probability horizontal slopes.In order to improve the calculation efficiency,this paper carried out a research on the reliability analysis of the heterogeneous reservoir slope based on the machine learning algorithm.The main research work and related conclusions are as follows:(1)A method for reliability analysis of reservoir slopes based on machine learning algorithms(back-propagation neural network BPNN,random forest RF and convolutional neural network CNN)is proposed,and an interface framework between slope reliability analysis and general finite element software is established.A batch program for finite element analysis of slope stability based on GEOSTUDIO and Fl AC3 D software interface was written in Matlab language,which expanded the application of machine learning algorithm in reservoir slope engineering.(2)Taking undrained saturated clay slope and friction/cohesive soil slope as examples,the advantages,disadvantages and applicability of three commonly used machine learning algorithms(BPNN,RF and CNN)in slope reliability analysis were quantitatively evaluated.The results show that: BPNN has the best effect and the strongest applicability;although RF is simple in theory and easy to use,it cannot solve the reliability problem of horizontal slope with low failure probability due to the imbalance of the training set;the effect of CNN is slightly inferior to that of BPNN.Although the model building process is complex,it has great potential to deal with high-dimensional problems.(3)Taking the two-dimensional unsaturated reservoir slope model under rainfall conditions as an example,the variation law of the failure probability of the reservoir slope with the rainfall duration considering both the shear strength parameter and the seepage parameter spatial variability is discussed.The results show that: BPNN has a good fitting ability for the transient unsaturated seepage analysis and stability safety factor calculation process of multi-parameter spatial variation of reservoir slopes;in addition,the spatial variability of shear strength parameters and seepage parameters is reliable for reservoir slopes and the slope failure probability increases gradually with the continuous rainfall.(4)Taking the Baishuihe landslide in the Three Gorges Reservoir area as an example,the failure probability of a heterogeneous reservoir slope under extreme conditions such as heavy rain and reservoir abruptness was quantitatively evaluated.The results show that the Baishuihe landslide will have a 19.5% probability of local instability failure under the rainstorm condition,while the failure probability of the local area below 175 m under the water level sudden drop condition is relatively small;in addition,the spatial variability of the saturated permeability coefficient of the landslide body still has a great influence on the safety factor of local stability.