Research On Improvement Of LSTM Model And Application In Prediction Of Landslide Vertical Displacement

In the past ten years,the average number of geological disasters in my country has been nearly 10,000 per year.Among them,geological landslides are the most frequent,most common,and most widespread geological disasters in my country.The number of casualties caused by landslides has reached more than 20,000.Frequent landslides seriously threaten people’s lives and property safety.And in recent years,frequent extreme weather has caused more and more dangerous landslides in my country,so it is extremely important to analyze the deformation of landslides and build a good prediction system.The accuracy of landslide prediction depends on the construction of high performance prediction model.There are many traditional landslide displacement prediction models.In recent years,the research on landslide displacement prediction has mainly focused on the development of multi-machine learning model fusion.A single model is easily affected by noise and has poor robustness.Therefore,in response to the above problems,this thesis improves the LSTM model to improve the accuracy of landslide displacement prediction.The research of this thesis is mainly as follows:(1)In this thesis,aiming at the inaccurate calculation of various parameters in the gross error detection of landslide monitoring data by the 3σ criterion and the interquartile range detection,by analyzing the original cumulative change and rainfall,it is found that the daily displacement due to rainfall in the early stage of the landslide Therefore,the original landslide monitoring data is divided into bins,and the data after binning is used for gross error detection,and linear interpolation and polynomial interpolation are used to analyze the data after gross errors are eliminated.Perform an interpolation operation.(2)A single long-short-term memory network(LSTM)model has high requirements on the quality and quantity of monitoring data,and noisy data will cause the problem of poor robustness of the LSTM prediction model.This thesis combines the wavelet threshold denoising algorithm to decompose the noise Based on the advantages of the Kalman filter model(KF)in dynamic space prediction and the strong nonlinear mapping and memory capabilities of the LSTM model,an improved LSTM model is designed,that is,the LSTM model that combines wavelet/KF.According to the deformation trend of Yangjiawan landslide,the landslide progress is divided into three categories: "accelerated deformation period","gradual change period" and "convergence period".The comparison proves that the prediction accuracy of the improved model has been improved.(3)In this thesis,the LSTM model fused with wavelet/KF is used to optimize BP neural network(PSO-BP),particle swarm optimization support vector machine(PSOSVM),and gray prediction model GM(1,1)respectively.The prediction results of each stage are compared,and the results show that: in this project,the LSTM model fused with wavelet/KF in this thesis has higher prediction accuracy than the above comparison model,and has strong stability and adaptability.