Research On Forecasting Step-like Displacement Of Hydrodynamic Landslide Based On Deep Learning

As one of the most serious geological disasters,landslides cause a large number of casualties and economic losses every year.Hydrodynamic landslides are widely distributed in the Three Gorges Reservoir area.Taking effective means to monitor such landslides and then making real-time forecasts is an important step to guide disaster prevention and mitigation.According to the investigation and analysis of typical hydrodynamic landslide data,the monitoring curve of step-like displacement in unstable hydrodynamic landslides accounts for a large proportion,so forecasting of step-like displacement is the key to hydrodynamic landslide prediction.In order to accurately forecast the step-like displacement,the Gated Recurrent Unit(GRU)was introduced to dynamically forecast step-like displacement after the comparison of multiple deep learning models.Combined with time series theory,Empirical Mode Decomposition(EMD)and Error Correction(EC),this paper presented a hybrid model(EMDGRU-EC)for forecasting landslide displacement.Then,taking the Baishuihe landslide as an example,step-like displacement was expressed as the trend and periodic displacements.Analyses of landslide evolution and monitoring data were used to select multiple influencing factors as input for periodic displacement.Finally,step-like displacement was accurately forecasted.The main research results obtained in this paper are as follows:(1)Investigate typical hydrodynamic landslides and systematically summarize the reasons for the formation of hydrodynamic landslides.According to the classification of cumulative displacement curves,the types of displacement curves of hydrodynamic landslides in different states are counted,and the landslides with step-liked characteristics are analyzed.(2)The deep learning model comparison research was carried out,and the results showed that GRU is suitable for the forecast of step-like displacements.Step-like displacement was expressed as the trend and periodic displacements based on time series theory,which respectively reflected the growth trend and the periodic changes of displacements.Through comparison and verification,EMD was used to separate the trend item of step-like displacement,and EC was introduced to improve the forecast accuracy of the periodic displacement.Then,a model suitable for step-like displacement forecast was proposed.(3)The Baishuihe landslide was selected as the research object,after analyzing the relationship between the statistics of different scales of rainfall,reservoir water level and steplike displacement,the influencing factors were found as the forecasting input of the periodic displacement.The analysis results show that the cumulative rainfall,maximum continuous rainfall,reservoir water level elevation,reservoir water level change rate,landslide displacement increment and displacement growth rate considering the hysteresis effect can well characterize rainfall,reservoir water level and landslide deformation state,considering the delayed effect,can well reflect the relationship of rainfall,reservoir water level,the state of the landslide and step-like displacements.(4)The hybrid model was applied to the forecast of the step-like displacement of Baishuihe landslide.The application results show that cumulative displacement of Baishuihe landslide is nonlinear and non-stationary,so EMD can be implemented.GRU has excellent performance in forecasting the trend and periodic displacements,and the addition of EMD and EC enhanced the performance of forecast model on different levels.The hybrid forecast model,EMD-GRUEC,proposed in this paper can provide a new way for timely forecasting of hydrodynamic landslides in Three Gorges Reservoir area.