Research On Prediction Of Landslide Displacement Time Series Based On Recurrent Neural Network

To be one kind of complex geological and natural disasters,landslide brings great harm to human beings and their living environment.The complicated geological environment,the different topography and the new tectonic movement such as reservoir impounding,which make there are many geological problems in the whole Three Gorges Reservoir area,especially landslide.Landslide has brought a lot of property losses and casualties,and even affected the full implementation of the whole emigration project.Therefore,it is of great significance to enhance the monitoring and forecasting of landslide and to reduce the hazard of landslide.The landslide is mainly due to the sliding geological phenomenon of slope rock and soil along the shear failure surface.There are many external factors that affect the landslide,such as natural conditions change,including earthquake,rainfall,freezing and thawing,tsunami,storm surge and human production activities.The evolution of landslide is considered as a nonlinear process.It is impossible to describe the relationship between the influence factors and the evolution of the landslide by using the existing linear method.The research on the prediction and control of landslide has been a hot topic in the field of engineering because of its great harm and the trend of increasing year by year.So,we use regression neural network(RNN)method,artificial intelligence method and combination of other nonlinear methods to the research of landslide prediction,taking Baishuihe(BSH)landslide and Liangshuijing(LSJ)landslide in Three Gorges area as experimental subject.The specific contents are included as follows.The external environmental factors that affect the stability of the landslide are studied.Based on idea of decision tree classification,random forest(RF)algorithm is proposed for stability analysis of landslide.There are two main characteristics of RF used for data processing:one is to avoid the use of traditional methods into the data over fitting,the other is the use of the existing data on the new stability analysis of landslide.In the same environment,compared with the traditional methods,such as support vector machine(SVM)algorithm,RF is more suitable for landslide stability analysis without consideration of multiple factors which impact the experimental results.Considering the prediction of landslide displacement time series,a model based on generalized regression neural networks(GRNNs)with K-fold cross-validation(K-GRNNs)is proposed.The main characteristic of GRNNs is that there is only one affected factor of human control.Therefore,this dissertation uses K-fold cross-validation to optimize the value of the threshold radial basis function(RBF)and the prediction performance of GRNNs.Take LSJ landslide and BSH landslide in the Three Gorges Reservoir area as examples to illustrate the predictive performance of our algorithm is higher than the traditional back propagation neural network(BPNN)and radial basis function(RBF).In order to meet the demand of the accuracy of landslide prediction,there are two main methods multi-step prediction and interval prediction used to extend the width of time of landslide prediction.In this dissertation,a method based on extended Kalman filter(EKF)and Back propagation through time(BPTT)is proposed to multi-step landslide displacement prediction.First of all,EKF is used to optimize the weights of the time reversal algorithm.And then the network parameters are adjusted through the latest observed value and model output values of the training results in real time to improve the reliability of prediction.Finally,the parameters of multi-step displacement prediction are determined by considering the relationship between the variation of the rainfall and the change of the reservoir level and the landslide displacement time series.Two case studies of LSJ and BSH landslide are presented and the experimental results show that the prediction using EKF-BPTT model are better than real-time recurrent learning(RTRL)algorithm and BPTT algorithm.Considering further optimizing the results of the width of time of landslide prediction,two methods are prepared for use in interval prediction in this dissertation.Firstly,the interval prediction model based on bootstrap and echo state network(ESN)with bifurcation iteration is proposed.The time series data is obtained by interpolation using ESN model.The training set is extended,and divided into B sub sequences with the bootstrap method corresponding to B ESNs.Then the value of the variance estimation of algorithm is calculated.Finally,an ESN is used to calculate the variance of error.The feasibility of this method is verified by two typical landslide data sets in the Three Gorges Reservoir area.Compared with the generalized regression neural network based on bootstrap(Bootstrap-GRNN)method,the proposed method can achieve better results.In this dissertation,based on the neural network method for interval prediction,a hybrid algorithm is proposed,combining particle swarm optimization(PSO)with gravitational search algorithm(GSA)(PSOGSA)used for ELMAN neural network.The main characteristic of PSOGSA-ELMAN is simple calculation with no need to deal with data over and again.The global search performance of PSO and local search performance of GSA are used to correct the weight of ELMAN neural network.Then the optimal RNN is constructed,getting the upper and lower bounds of the interval prediction.The experimental results show that the prediction using PSOGSA-ELMAN model are better than Bootstrap-ESN with bifurcation iteration in previous chapter.All the given experimental results and comparisons show that the method proposed in this dissertation can solve some practical problems and provide some new ideas and provide the necessary basis for the relevant departments and personnel to carry out corresponding treatment of the landslide hazard area.