Study On Multiple-field Information Monitoring And Ensemble Forecasting Model Optimization Of Landslides In The Three Gorges Reservoir Area

As one of the most widely distributed and frequently occurring geological disasters in nature,landslides pose a greater threat to the environment,natural resources,hydraulic engineering,etc.Taking the Three Gorges Project as an example,since the first impoundment in 2003,large-scale ancient landslides such as the Huangtupo landslide and Huanglashi landslide,etc.show significant deformation signs again,resulting in the forced relocation of nearby cities and 1.2 million people.Among the 5000 landslides or potential landslides in the Three Gorges Reservoir area(TGRA).the slow-moving landslide,which accounts for the majority of them,are move at the speed of 6 cm to 1.5 m/year,and some even less than 6 cm/year.Owing to its characteristic of slowly moving downslope for decades or even centuries until the rapid acceleration of deformation,the slow-moving reservoir landslides are important objects for the study of the landslide prediction.At present,there are two main requirements for the study of the slow-moving reservoir landslide in monitoring methods and displacement prediction methods.On the one hand,because the appearance of landslide deformation and failure is a spatiotemporal dynamic evolution process,it is often accompanied by characteristics such as multi-field coupling and large deformation.However,it has not yet been possible to propose a monitoring method for the whole process of gradual landslide evolution.The existing deep monitoring technology cannot adapt to the large deformation characteristics of the landslide,and there is also a lack of multi-information parameter monitoring technology to achieve "one hole,multiple measurements" under in-situ conditions,etc..On the other hand,with the big explosion of algorithms and models in the information age,a large number of new algorithms and new model frameworks are used to develop new landslide prediction models based on the evolutionary law of gradual landslide deformation and mining the frequency components in the historical information of inducing factors.It is particularly urgent.In this thesis,four typical slow-moving landslides were selected as the case study.The main research contents and results are as follows:(1)The impoundment of the TGRA is an important cause of these landslides.The slow-moving reservoir landslide experience accelerated deformation many times in deformation history.The monitoring data of slow-moving landslides in the TGRA generally shows a step-like shape.From May to August each year,the landslide moves rapidly,with little movement in the remaining months.This displacement pattern is consistent with a seasonal increase in monthly rainfall in the reservoir area and a corresponding increase in the groundwater pressures in the landside.The intensification of sliding mass deformation is closely related to the external inducing factors like rainfall and reservoir water level changes.Under the combined action of these external inducing factors,ancient landslides began to show a slow acceleration of deformation,and new deformations and failures are generated locally.Due to the differences of topography,slope structure and external factors,landslides at different locations usually show different deformation rates.Based on the field survey,after the start of sliding,obvious deformation features such as cracks and collapsed areas will appear on the surface of the sliding mass,and it will continue to slowly evolve and expand.Such landslides have a large volume and a wide range of impacts,which often cause huge economic losses and have the potential to cause secondary disasters.(2)Three landslide monitoring equipment are designed to solve the problems of large deformation and multi-field information monitoring of landslides.A device designed based on the concept of the prefabricated magnetic field can effectively use the magnetic field to realize non-contact,long-term monitoring of landslide deformation.The deformation coordination problem caused by the coupling of inclinometer pipe and soil under large deformation of landslide mass is avoided,and the precision is relative to the traditional inclinometer.Devices for multi-field geological parameters monitoring outside borehole in sliding mass were designed.Integrated sensors were inserted into the rock and soil environment to be monitored outside the hole,which can realize the multi geological parameter information monitoring of the rock and soil mass of surrounding borehole in the sliding mass,and that makes the monitoring result is closer to the real underground environment.Meantime,considering the complexity of the underground environment,the obvious disturbance of rock and soil caused by the layout of the anti-collapse trough is avoided.The whole system adopts wireless power supply mode,with simple structure,reasonable design,economic efficiency and convenient promotion.In order to provide a safe and stable operating environment for the aforementioned multiple geological parameters monitoring devices,a device for tamping the sidewall of the borehole was designed,which can solve the pipe soil coupling problem and the reinforcement problem of the borehole at the same time during the inclinometer pipeline implantation stage.It has important practical significance for engineering applications in the field of geological hazard monitoring and prevention.(3)A prediction method based on CEEMD-DTW and ACO-SVR was proposed with the study of the Baishuihe landslide and Shuping landslide.CEEMD with high decomposition accuracy and high operation efficiency can better highlight the local fluctuation characteristics of the inducing factors time series.The DTW can be adopted for the selection of the most relevant inducing factors of periodic displacement.The results of the proposed hybrid model show that after considering the frequency components of landslide-induced factors,the prediction accuracy of the displacement prediction model based on ACO-SVR is better than that of other models based on SVR and GA-SVR.(4)Chapter 5 Forecasting model with CEEMD-LCSS reconstruction and ABC-SVR method was proposed based on the study of Baishuihe landslide.The data of inducing factors were preprocessed using CEEMD to identify their high-frequency,low-frequency,and residue components.The application of LCSS enables the identification of the time-series data that have the strongest influence on the measured landslide displacements at each monitoring point.The prediction precision of the combined forecasting model that is based on ABC-SVR exceeds those of the models that are based on GA-SVR and PSO-SVR.which can increase the prediction accuracy of landslide displacements by 54% at most in the evaluation of RMSE.(5)An ensemble prediction model based on EDR selection and multiple SIs was proposed based on the study of the Shiliushubao landslide.The deformation of the Shiliushubao landslide is closely associated with the related factors including reservoir level fluctuations,rainfall intensity,and previous deformations.The EDR method can identify the most relevant factors influencing a landslide's movements to use as input variables for an SVR model.The DA and GWO based SVR model provided the best prediction of periodic displacement and trend displacement,respectively.compared to the prediction model proposed by Deng et al.(2017),the ensemble prediction model based on EDR selection and multiple SIs was slightly improved.(6)A comparative test has been conducted from four aspects to find the optimal combination of an ensemble prediction model with four nodes for predicting slow-moving landslide displacement in the reservoir area.The EMD series model can be applied to the decomposition of displacement data and the frequency component extraction of inducing factors.Among them,the EEMD method is suitable for the decomposition of landslide displacement and the small dispersion,well-fitted prediction results can therefore be obtained.In the aspect of inducing factors decomposition,the combination of CEEMDAN with t-test can better highlight the frequency component of inducing factors.Compared with other methods,the use of CEEMDAN in the node II of the prediction model shows competitiveness.The GRD and the trajectory similarity judgment model can be used for the selection of the input variable of the landslide prediction model.The application of the LCSS method can lead to the best prediction results and should be the preferred method in node III,followed by GRD,DTW and EDR.Based on LCSS,the R2(the cumulated rainfall in the previous two months)and W5(the low-frequency components of the original reservoir water level time series)are the two external inducing factors that affect the Outang landslide most.Among all the nine SI optimization algorithms tested in this chapter,the application of PSO in SVR based prediction model gives the best prediction result with the minimum value of MAPE,RMSE,MAE and the maximum value of R2.Thus,the best combination for this kind of prediction framework with four nodes is EEMD-CEEMDAN-LCSS-PSO.The innovative points in this thesis are summarized as follows:(1)The theoretical design of three devices for the landslide multi-field information monitoring was developed in two aspects with high precision,high reliability and high coupling degree,including the monitoring of landslide's deformation in the depths and the monitoring of multi-field information in the sliding mass.A device for laying out a prefabricated magnetic field and methods of responding state of a slip mass,devices for multi-field geological parameters monitoring outside borehole in sliding mass and a device for tamping the side wall of the borehole were developed.(2)Based on the monitoring data of slow-moving reservoir landslides,the frequency components of landslide inducing factors are considered,and the application of CEEMD in landslide displacement decomposition and frequency components extraction of inducing factors is discussed.For the first time,the similarity evaluation methods of time series(such as DTW,LCSS,EDR)are taken as the most relevant variable selection method for the optimization.For the first time,multi swarm intelligent optimization algorithms(such as sparrow search SSA,ant lion optimization algorithm,etc.)are applied to the optimization of the landslide displacement prediction model,and the optimization performance is compared with the common classical algorithms.Based on the above research,three combined prediction models are proposed,and their generalization performance is tested through the application and comparison of landslide cases in the TGRA.(3)The commonly used framework of the ensemble forecasting model was summarized into four steps,and the comparative experiments of the above application methods in each step were carried out.The generalization performance of different methods in the same step was compared,and the optimal combination of the SVR-based forecasting model in the four step framework is obtained.