Research On Slope Time Series Prediction Based On LSTM

Landslides are the world’s second largest geological disaster after earthquakes.At the beginning of the new century,it has brought serious economic losses and casualties to our country,and it poses a major threat to national safety and social modernization.Slope deformation prediction can feed back early warning information before the slope crash and damage.It is a vital part of the slope monitoring and early warning system and has great significance for disaster prevention and mitigation.Based on the two-year deformation data of the two measuring points of the No.2 mountain beam slope of the Xiaowan Power Station,this paper uses the python programming language to apply the traditional time series model,machine learning model and deep learning model to the slope deformation time series data analysis and forecast,To provide a reference for the disaster prevention and mitigation measures of the slope early warning system.The main content of this paper is as follows.(1)Construct an autoregressive moving average(ARIMA)model.The stationarity test(ADF test)of the original slope deformation time series data is performed,and the data is subjected to differential processing to meet the modeling requirements.The grid search method is used to optimize the parameters of the BIC information of each model,and the order of the model is determined as ARIMA(1,1,1),and finally use the Ljung-Box test to judge the significant validity of the model.Data preprocessing mainly uses Pandas and Numpy data analysis packages.ARIMA model modeling is mainly based on the ARIMA model package contained in the Statsmodels library.(2)Construct a support vector machine(SVM)model.Use the Min Max Scaler module in the Sklearn tool to normalize the data,combine the model prediction results under multiple kernel functions,select the Gaussian kernel function(RBF)as the kernel function,and call the Grid Search CV method to perform a grid search on the penalty factor C and the parameter gamma After optimization,it is finally determined that the C and gamma of sample one are 8.47 and 0.70,respectively,and the sample two are 7.81 and 0.67.SVM model modeling is mainly based on the SVR model package in the Sklearn library.(3)Construct a long and short-term memory neural network(LSTM)model.Based on the theory of the Recurrent Neural Network(RNN)model,a standard one-dimensional fully connected layer is used to build a single hidden layer network.After multiple adjustments,the main hyperparameters of the network are determined as follows: the number of hidden layer neurons is 20;The optimization algorithm selects the SGD algorithm;the number of iterations is 1200;the number of training batches is 64.The LSTM model construction is based on the Keras framework and Tensorflow backend.(4)Set up multiple experiments for model comparison.Two indicators of MAE(mean absolute error)and RMSE(root mean square error)are selected to evaluate the prediction performance of the above model,and after a comprehensive analysis of the characteristics of each model and the experimental results,it is concluded that in the two samples,the MAE and RMSE of the LSTM model are both It is the smallest,sample one is 0.44 and 0.52,and sample two is 0.29 and 0.34,verifying that the LSTM model has more advantages than the traditional ARIMA and SVM models in dealing with the slope time series data prediction problem.