Research On The Application Of Deep Learning Based Combined Models In Displacement Prediction Of Loess Landslides

Accurate loess landslide prediction can help people identify and prevent landslide disasters in advance,thus avoiding human casualties and property damage.Therefore,it is necessary to monitor and predict landslide displacements.There are many algorithms applied to landslide displacement early warning models,but the adaptation conditions and levels of the models vary,so it is of great importance in engineering practice to analyze the applicability conditions of the models based on the landslide displacement time series characteristics and select the suitable models for forecasting.In this paper,we use the monitoring data of Beidou/GNSS monitoring stations BZ01,BZ02 and FJ01 of the landslide in Zizang Village and Fangjia Village in Jishiyama County,Linxia City,Gansu Province as the research object,and use six intelligent forecasting models with MAE,MAPE and MSE as evaluation indexes to comprehensively evaluate and analyze the application performance of each intelligent model for forecasting in order to find the suitable forecasting model.The main research contents and results are as follows:1.The evolution patterns of periodicity,trend and abrupt change characteristics of monitoring stations BZ01,BZ02 and FJ01 were compared and analyzed using Morlet wavelet transform,Hurst index and M-K abrupt change test.The results show that the trend of displacement time series of monitoring stations BZ01,BZ02 and FJ01 shows decreasing changes,and there are different levels of regular transformation in periodicity;in addition,the sudden change time of displacement time series of the three stations varies due to the disturbance of observation environment,weather change,geological conditions and human activities.The results of the comprehensive three-station time series analysis help to reveal the diffraction mechanism and deformation characteristics of the landslide displacement time series,and provide strong basic theoretical support for the subsequent landslide displacement prediction research.2.Six artificial intelligence algorithms were constructed,namely,long and short term memory neural network(LSTM),gated recurrent unit(GRU),temporal convolutional network(TCN),BP neural network model,support vector machine model(SVM)and kernel extreme learning machine(KELM).Firstly,the performance of each prediction model was compared based on the monitoring data of Beidou/GNSS monitoring stations BZ01,BZ01 and FJ01,and the results indicated that the gated recurrent unit and temporal convolutional network deep learning algorithms were more efficient than BP,SVM and KELM classical machine learning algorithms,and the GRU and TCN models could both track the landslide displacement time series change pattern and accurately and rapidly predict the The GRU and TCN models can both track the landslide displacement time series change pattern and predict the landslide displacement time series trend accurately and rapidly.Secondly,by analyzing the monitoring data of the three stations longitudinally,it can be seen that station FJ01 has more differences in data characteristics than stations BZ01 and BZ02,and the difficulty of predicting the monitoring data of FJ01 is smaller compared with BZ01 and BZ02.3.Since the landslide displacement time series has the characteristics of nonlinearity and non-smoothness,it is difficult for a single model to globally track the trend changes of the landslide displacement time series,and a combined model of joint variational mode decomposition(VMD)and deep learning prediction is constructed in this paper.Firstly,the modal decomposition of the landslide displacement time series is performed based on the VMD decomposition technique,and then the sub-signals obtained by using the decomposition are input to the deep learning model,and finally training tests are conducted.The results show that the three decomposition methods coupled with BP,SVM,KELM,LSTM,GRU and TCN models are more adaptable,and the accuracy of the combined prediction model based on decomposition and deep learning algorithms can be more satisfactory compared with the prediction results of a single model.4.In this paper,a hybrid prediction model combining sparrow search algorithm(SSA),variational modal decomposition and GRU neural network is proposed.The sparrow search algorithm is used to obtain the optimal initial parameters of GRU to solve the problem that the random selection of the initial parameters of deep learning leads to its slow convergence and falls into local optimum.The model was compared with GRU and VMD-GRU prediction results,and the combined VMD-SSA-GRU model obtained better prediction results.