Reservoir bank landslide displacement prediction method based on main control characteristic factor ternary optimization strategy

Abstract

The invention provides a reservoir bank landslide displacement prediction method based on a master control characteristic factor ternary optimization strategy. The reservoir bank landslide displacement prediction method comprises the following steps: S1, collecting data, preprocessing the collected data, and constructing a multi-source heterogeneous variable characteristic set; s2, reconstructing the accumulated displacement of the landslide into a trend term component and a periodic term component by adopting a CEEMDAN algorithm; s3, predicting trend term displacement by using a double exponential smoothing algorithm DES; s4, providing a master control characteristic factor ternary cooperative characteristic optimization strategy, fusing RFE-XGBOOST-CV, CART and MIC to obtain an optimal combination of the master control factors, and then predicting periodic term displacement by using an LSSVM model, an RF model and a GRU model to verify a master control factor optimization result; and S5, finally, accumulating the predicted values of the trend term and the period term to obtain the prediction of accumulated displacement. According to the method, the limitation of traditional single feature screening is broken through, the model training efficiency and the complex working condition adaptability can be remarkably improved through the ternary collaborative optimization strategy, and a new technical thought is provided for reservoir bank landslide intelligent early warning.

Description

Technical Field

[0001] The present invention relates to the field of reservoir bank landslide displacement prediction, and particularly to a reservoir bank landslide displacement prediction method based on a ternary optimization strategy of main control characteristic factors. Background Technique

[0003] In recent years, scholars at home and abroad have conducted a large number of studies on the field of reservoir bank landslide displacement prediction and achieved remarkable progress. The landslide displacement prediction methods proposed in existing studies are divided into two categories: physically driven mechanism models and data-driven comprehensive models. The former mainly establishes finite element models (such as GeoStudio, etc.) based on the coupling of seepage field - stress field. However, its modeling is limited by the spatial variability of hydrogeological parameters and the uncertainty of boundary conditions, and has inherent defects such as low modeling efficiency and weak generalization ability; the latter reconstructs the cumulative displacement into meaningful trend terms and periodic terms through time series decoupling techniques such as empirical mode decomposition, and then uses machine learning models such as extreme learning machine (ELM) and support vector machine (SVR) for sub-item prediction. Although remarkable progress has been made, it still faces the challenge of insufficient dynamic feature representation ability. Especially in the prediction effect during the rapid step period of the periodic term displacement is poor. The fundamental crux lies in the following limitations of the existing factor screening methods: First, traditional correlation analysis methods (MIC, GRA) are difficult to effectively capture the time-delay attenuation effect of the coupling action of rainfall - reservoir water level, thus resulting in redundant screened features; second, single feature screening methods are difficult to comprehensively analyze the non-linear interaction between multiple factors, and each screening method has different criteria and unreliable results. The above deficiencies directly lead to a structural defect in which the input factors have both information redundancy and key feature loss, resulting in large prediction errors and poor generalization ability even when using deep learning models such as GRU and Informer, and it is still impossible to fully mine the spatio-temporal evolution law of landslide displacement evolution. Summary of the Invention

[0004] Aiming at the above technical defects, the present application proposes a reservoir bank landslide displacement prediction method based on a ternary optimization strategy of main control factors, aiming to alleviate the problems of insufficient representation of multi-source coupling action and dynamic feature distortion in the existing reservoir bank landslide displacement prediction field, and improve the accuracy of accurate dynamic prediction of reservoir bank landslide displacement, especially during the rapid step period of the periodic term, which is mainly achieved by constructing a "package - embedded - filtering" three - level feature optimization architecture.

[0005] The technical solution of the present application is: to provide a reservoir bank landslide displacement prediction method based on a ternary optimization strategy of main control factors, including the following steps:

[0006] S1. Data collection, preprocessing, and construction of a multi-source heterogeneous variable feature set: Collect high-precision GNSS historical displacement monitoring time-series data of the landslide deformation monitoring points in the study area. Synchronously observe the rainfall, reservoir water level, and other data of the inducing factors affecting the landslide displacement evolution within the landslide range and complete data cleaning. Based on the rainfall-reservoir water level coupling mechanism, construct a multi-source heterogeneous variable feature set, which includes but is not limited to various characteristic parameters such as effective rainfall in the early stage, average rainfall, reservoir water level fluctuation changes, rainfall and reservoir water level changes in the lagging n periods, and displacement changes.

[0007] S2. Data decomposition and reconstruction: Use CEEMDAN to reconstruct the cumulative displacement into trend term and periodic term displacement components for the multi-source heterogeneous variable feature set constructed in step S1, where the decomposition parameters are adaptively determined by the GWO optimization algorithm with the minimum envelope entropy as the objective function.

[0008] S3. Trend term displacement prediction modeling: Use the double exponential smoothing algorithm DES to predict the trend term displacement component.

[0009] S4. Selection and prediction of the main control characteristic factors of the periodic term: Conduct time-series correlation analysis on the multi-source heterogeneous variable feature set constructed in S1 and the periodic term displacement component results obtained in S2. Subsequently, use RFE-XGBOOST-CV (wrapper) and CART (embedded) to perform feature selection on the multi-source heterogeneous variable feature set to obtain the key influencing factor set, and on this basis, calculate the MIC (filter) values of the remaining features. Set the MIC correlation coefficient threshold, and screen to obtain the optimal combination of the final main control factors according to the MIC numerical sizes of each input feature. Finally, use three different machine learning models, namely LSSVM, RF, and GRU, to predict the periodic term displacement respectively to verify the results of the main control factor selection.

[0010] S5. Cumulative displacement prediction: Add the trend term displacement predicted in step S3 and the periodic term displacement predicted in step S4 to obtain the final landslide cumulative displacement prediction result and evaluate the accuracy through the coefficient of determination (R 2 ) and root mean square error (RMSE) indicators.

[0011] The machine learning model can also adopt other learning models with the same or similar functions and effects as the LSSVM, RF, and GRU learning models.

[0012] Furthermore, calculate the effective rainfall in the early stage, and the expression is:

[0013] API t =λAPI t-Δt +PΔt Formula 1

[0014] Where: API t is the previous rainfall index at time t, λ is the rainfall influence attenuation coefficient, and P Δt is the rainfall from t - Δt to t.

[0015] Furthermore, the expression of the cumulative displacement is as follows:

[0016] D t = φ t + η t Formula 2

[0017] Where: t is the time step; D t is the cumulative displacement; φ t is the trend term displacement; η t is the periodic term displacement.

[0018] Furthermore, the expressions of the trend term and the periodic term obtained by decomposing through CEEMDAN are as follows:

[0019]

[0020] Where and represent the trend term displacement and the periodic term displacement after decomposition respectively; GWO and CEEMDAN represent the Grey Wolf Optimization algorithm (abbreviation: GWO) and the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise algorithm (abbreviation: CEEMDAN) respectively; D t represents the cumulative displacement; k1 and k2 represent the decomposition parameters to be optimized, and their values are determined by the Grey Wolf Optimization algorithm with the minimum envelope entropy as the objective function.

[0021] Furthermore, the expression of the deterministic prediction of the trend term displacement by the Double Exponential Smoothing method DES is as follows:

[0022]

[0023] Where, Y t+1 is the predicted value of the trend term displacement at time t - 1; S t and S t-1 represent the exponential smoothing values at time t and t - 1 respectively; b t and b t-1 represent the best estimates of the gradients at time t and t - 1; α and γ both represent smoothing factors; φ t is the trend term displacement.

[0024] The combination of the GWO and CEEMDAN algorithms realizes the adaptive decomposition of the cumulative displacement of the landslide, effectively overcomes the problem of mode mixing in decomposition algorithms such as empirical mode decomposition in traditional methods, realizes the high-precision separation of the trend term and the periodic term. Therefore, the trend term is an approximately monotonic curve with simple characteristics, and a good prediction effect can be obtained by setting appropriate values for the smoothing factor.

[0025] Furthermore, the MIC (filter type) described in step S4 is a correlation analysis method based on the mutual information coefficient. By calculating the maximum mutual information value between different data subsets, it measures the linear or non-linear correlation between variables. This method can quantify the complex dependence relationship between variables, retain the information of strongly correlated variables, and perform well in dealing with the non-linear relationship of weakly correlated variables. Compared with existing correlation evaluation methods such as grey relational degree, MIC has stronger universality and equivalence, can adapt to more diverse data patterns and relationship types. With this advantage, MIC can provide a more accurate measure of variable correlation. Especially when dealing with complex data sets, it can more comprehensively reveal potential association patterns, and its expression is:

[0026]

[0027] where: x represents each influencing feature; η t represents the displacement of the landslide periodic term; P(x, η t ) is the joint probability density distribution function between the influencing feature and the periodic term displacement; P(x) and P(η t ) are the marginal probability density functions of the two variables; MIC(x; η t ) is the maximum mutual information coefficient between variables; p and q are the grid distributions in the x and η t directions; B is a variable, b < n 0.6 .

[0028] Furthermore, the RFE-XGBOOST-CV (wrapper type) described in step S4 is a method that combines the RFECV method and the XGBoost method;

[0029] The RFE-CV is a wrapper feature selection method that combines feature importance evaluation and model validation. Its core idea is to evaluate the importance of each feature through the base model and gradually remove the least important features until the predetermined number of features is reached or all features have been evaluated; finally, RFE-CV evaluates different feature subsets through cross-validation, calculates the average scores of each subset, and thus determines the best feature combination;

[0030] The XGBoost is an efficient ensemble learning algorithm that improves model performance by constructing and combining multiple decision trees. Its built-in feature importance scoring mechanism can assist RFECV in identifying features that have a relatively small impact on the model but still have strong predictive power. Combining RFE-CV with XGBoost can give full play to the advantages of both. Especially when dealing with high-dimensional data, XGBoost can focus on key variables through the features screened by RFE-CV, accurately screen out the optimal features, significantly reduce the risk of overfitting, and at the same time enhance the generalization ability and interpretability of the model.

[0031] Furthermore, the CART tree algorithm is a classic embedded feature screening algorithm that can effectively handle linear and non-linear relationships between features and target variables. This algorithm constructs a binary decision tree recursively and selects features using the least squares error minimization criterion. The basic steps are as follows:

[0032] ① Select the optimal splitting variable j and splitting point s that satisfy the following formula:

[0033]

[0034] where j and s are the feature and the splitting point of the feature respectively, and c1 and c2 represent the predicted values that minimize (y i -c1) 2 for this interval. Traverse all input features j to find the smallest (j, s) combination;

[0035] ② Divide the parent node according to the smallest (j, s) combination to obtain the output values corresponding to the two divided regions:

[0036]

[0037] In the formula: R1(j, s) and R2(j, s) are the divided regions; x j is the value of the sample on the j-th feature; N m is the number of samples in the region; is the predicted value;

[0038] ③ Repeat the above process in the child nodes until all leaf nodes are constructed. Each time a feature is split, select the feature and splitting point that can reduce the square error to the greatest extent. Finally, measure the importance of the feature by calculating the reduction of the average square error for each feature and normalizing it.

[0039] Furthermore, a system for predicting the displacement of a reservoir bank landslide based on a ternary optimization strategy of main control feature factors is provided, including:

[0040] The first processing module cleans and preprocesses multi-source heterogeneous data, and constructs a multi-source heterogeneous variable feature set according to the landslide displacement evolution law under the coupling action of reservoir water level and rainfall;

[0041] The second processing module is used to reconstruct the cumulative displacement of the monitoring point into a trend term and a periodic term by using the integrated CEEMDAN and GWO optimization algorithms according to the data preprocessing results;

[0042] The third processing module is used to realize the prediction of the trend term displacement by using the double exponential smoothing method DES;

[0043] The fourth processing module is used to select the main control factors of the periodic term and complete the prediction of the periodic term displacement. A three-way collaborative main control feature factor screening strategy integrating wrapped RFE-XGBOOST-CV, embedded CART, and filter MIC is proposed to obtain the best combination of main control factors from multi-source features; and on this basis, three different machine learning models, LSSVM, RF, and GRU, are used to predict the periodic term displacement respectively, and the superiority of the main control factor screening strategy is evaluated according to the prediction indexes;

[0044] The fifth processing module is used to add the prediction results of the trend term and the periodic term displacement to obtain the cumulative displacement prediction value, and calculate relevant indexes according to the true value and the prediction value of the cumulative displacement to evaluate the prediction results.

[0045] Beneficial effects: The present invention adopts a full-chain displacement prediction system from data cleaning and preprocessing, construction of a multi-source heterogeneous variable feature set, time series decomposition, optimization of main control factors, and sub-item modeling, and organically integrates the landslide displacement evolution mechanism and data-driven methods.

[0046] First, high-precision separation of the trend term and the periodic term is realized based on GWO-CEEMDAN.

[0047] Secondly, a three-way collaborative main control feature factor optimization strategy of RFE-XGBOOST-CV (wrapped), CART (embedded), and MIC (filtered) is effectively integrated, breaking through the limitations of a single method. Through a three-dimensional collaborative mechanism of dynamic interaction analysis, time-delay quantification analysis, and redundancy elimination, a main control factor set with physical interpretability is constructed, which can significantly improve the characterization ability of the coupling action of reservoir water level and rainfall and its lagging attenuation response.

[0048] Finally, verified by multiple different models such as LSSVM, RF, and GRU, the main control factor feature screening strategy proposed by the present invention can break through the algorithm type limitation, effectively improve the matching degree between the input factors and the landslide dynamic evolution law under different modeling paradigms, provide a general solution for the characterization of multi-source coupling actions, and especially significantly improve the problem of large prediction errors of the existing methods during the rapid deformation period of the reservoir bank landslide displacement. Description of the Drawings

[0049] Figure 1 It is the flowchart of the method for predicting the displacement of the reservoir bank landslide in this application;

[0050] Figure 2 It is the flowchart for optimizing the main control factor of the periodic term and predicting the displacement of the periodic term in this application;

[0051] Figure 3 It is the graph of cumulative displacement, rainfall, and reservoir water level data in the embodiment of this application;

[0052] Figure 4 is Figure 3 the predicted graph of the cumulative displacement of the X-03 monitoring point in

[0053] Figure 5 is Figure 3 the predicted graph of the cumulative displacement of the X-06 monitoring point in Specific implementation manners

[0054] The following is a detailed description of this application with reference to specific drawings.

[0055] As Figure 1 shown, the flowchart of a method for predicting the displacement of a reservoir bank landslide based on a ternary optimization strategy of main control characteristic factors proposed by the present invention includes the following steps:

[0056] Step S1: Data collection, preprocessing, and construction of a multi-source heterogeneous variable feature set: Collect high-precision GNSS historical displacement monitoring time series data of reservoir bank landslide deformation monitoring points in the study area, synchronously observe rainfall, reservoir water level, and other data of factors inducing the evolution of landslide displacement within the landslide range and complete data cleaning, and construct a multi-source heterogeneous variable feature set based on the rainfall-reservoir water level coupling mechanism, including but not limited to various characteristic parameters such as previous effective rainfall, average rainfall, reservoir water level fluctuation changes, rainfall and reservoir water level changes and displacement changes in the lag n periods. The calculation formula of the previous effective rainfall index is as follows:

[0057] API t = λAPI t-Δt + P Δt ; Formula 1

[0058] In the formula: API t is the previous rainfall index at time t, λ is the rainfall influence attenuation coefficient, and P Δt is the rainfall from t - Δt to t.

[0059] The expression of the cumulative displacement is as follows:

[0060] D t = φ t + η t Formula 2

[0061] Where: t is the time step; D t is the cumulative displacement; φ t is the trend term displacement; η t is the periodic term displacement;

[0062] Step S2: The CEEMDAN algorithm is used to reconstruct the landslide cumulative displacement sequence to obtain the trend term component representing the long-term evolution law and the periodic term component reflecting the periodic fluctuation, and the Grey Wolf Optimization algorithm (abbreviation: GWO) is used to optimize the decomposition parameters of the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise algorithm (abbreviation: CEEMDAN); The expressions for reconstructing the landslide cumulative displacement into the trend term and periodic term displacement components by the CEEMDAN decomposition algorithm are as follows:

[0063]

[0064] where, imf φ and imf η represent the decomposed trend term displacement and periodic term displacement respectively; GWO and CEEMDAN represent the Grey Wolf Optimization algorithm and the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise algorithm respectively; D represents the cumulative displacement; k1 and k2 represent the decomposition parameters to be optimized;

[0065] Step S3: Based on the Double Exponential Smoothing algorithm DES, deterministic prediction is performed on the trend term displacement component; The expression is as follows:

[0066]

[0067] where, Y t+1 is the predicted value of the trend term displacement at time t-1; S t and S t-1 represent the exponential smoothing values at time t and t-1 respectively; b t and b t-1 represent the best estimates of the gradients at time t and t-1; α and γ both represent the smoothing factors; φ t is the trend term displacement;

[0068] Step S4: As Figure 2 shown, it shows the process of optimizing the main control characteristic factors of the periodic term displacement and predicting the periodic term displacement. First, a multi-source heterogeneous variable feature set is constructed through the analysis of the landslide displacement evolution mechanism, and then the proposed triple optimization cooperation strategy of the main control characteristic factors based on RFE-XGBOOST-CV (wrapper), CART (embedded) and MIC (filter) is used to obtain the best combination of the main control factors of the periodic term from the multi-source features; and on this basis, three different machine learning models (such as LSSVM, RF and GRU) are used to predict the periodic term displacement respectively and verify the optimization results of the main control factors according to the prediction indexes;

[0069] Step S5: Linearly superpose the predicted values of the trend term and the periodic term to obtain the predicted result of the cumulative displacement, and evaluate the accuracy of the finally obtained predicted result of the landslide cumulative displacement through the coefficient of determination R 2 (English: coefficient of determination) and the root mean square error RMSE (English: root mean square error) index.

[0070] In this embodiment, the test area takes the Xinpu Landslide, a typical landslide on the Three Gorges reservoir bank, as an example.

[0071] The Xinpu Landslide is located on the south bank of the Yangtze River in Fengjie County, the Three Gorges Reservoir Area, 182 km away from the Three Gorges Dam. It is a super-large landslide group of bedding - slightly cutting layers. The overall oblique length of the landslide is about 2 km, the relative height difference is 634.4 m, and the slope is 15° - 20°. The landform is complex, and the multi-level terrace landform is significant. The strata are mainly composed of the Xujiahe Formation of the upper Triassic series, the artificial accumulations of the Quaternary system, and landslides, etc. The terrain in the landslide area has a gentle - steep transition. A steep bank and a steep slope are formed at the front edge, and the slope at the rear edge exceeds 35°, gradually becoming gentle to 24°. The height of the steep bank formed along the resettlement highway is 5 - 10 m, the slope exceeds 60°, and the slope of the lower part is still above 30°. In addition, there are also gullies with a depth of 2 m - 11.5 m, which are consistent with the slope trend. A total of 10 high-precision GNSS ground displacement monitoring points are arranged in the Xinpu Landslide area.

[0072] To better illustrate the technical effects of the present invention, landslide data at different spatial positions are selected for analysis. In this embodiment, the typical landslide in the Three Gorges - the Xinpu Landslide is taken as an example. To minimize the mutual interference between monitoring points as much as possible, two monitoring points, X - 03 and X - 06 stations, which are far apart and have the most complete data, are selected. The daily - scale data from January 10, 2020 to March 26, 2024 are used as the research object to conduct experimental verification on the present invention.

[0073] As Figure 3 shown, it is the data graph of the cumulative displacement, rainfall, and reservoir water level of the Xinpu Landslide in this embodiment. The analysis shows that the cumulative displacement curve of the Xinpu Landslide presents an obvious stepped shape. From May to September every year, with the sharp drop of the reservoir water level and the arrival of the rainy season, the "displacement step" phenomenon appears in the landslide; while in other time periods, with the rise of the reservoir water level and the decrease of rainfall, the cumulative displacement curve of the landslide tends to be flat. This indicates that the changes in rainfall and reservoir water level are the main triggering factors for the evolution of landslide displacement, and their influence on the displacement of the reservoir bank landslide has certain hysteresis and attenuation.

[0074] First, based on the landslide displacement monitoring data and the concurrent rainfall, reservoir water level, temperature, and wind speed data, a landslide displacement coupling and modeling prediction analysis is carried out. Considering the noise signals mixed in the original data, in this embodiment, the data is preprocessed and standardized, and then a multi-source heterogeneous variable feature set is constructed, including rainfall factors: such as the average rainfall in the previous n days / months, the lagged rainfall in the previous n days / months, and the rainfall on the current day / month; reservoir water level factors: such as the average reservoir water level in the previous n days / months, the lagged reservoir water level in the previous n days / months, and the reservoir water level on the current day / month; and key features such as API factors. Among them, the effective rainfall index in the early stage is obtained through Formula 1. Through investigation, analysis, research, and combining the achievements of predecessors, the maximum lag period of the Xinpu landslide is set to 60 days and incremented from 0 in sequence.

[0075] The CEEMDAN algorithm is used to reconstruct the landslide cumulative displacement sequence to obtain the trend term component representing the long-term evolution law and the periodic term component reflecting the periodic fluctuations, and the Grey Wolf Optimization algorithm (abbreviation: GWO) is used to optimize the decomposition parameters of the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise algorithm (abbreviation: CEEMDAN); the expressions of the trend term and the periodic term reconstructed by CEEMDAN are shown in Formula 3.

[0076] In this embodiment, since the trend term displacement is mainly driven by internal factors such as geological conditions and mainly shows long-term stable growth with an approximately smooth curve, the Double Exponential Smoothing algorithm DES is used to realize the deterministic prediction of the trend term displacement, and the specific method is shown in Formula 4.

[0077] In this embodiment, the method for selecting the main control factors for the periodic term displacement is as follows: First, a multi-source heterogeneous variable feature set is constructed through the analysis of the landslide displacement evolution mechanism, and the ternary collaborative strategy based on RFE-XGBOOST-CV (wrapper), CART (embedded), and MIC (filter) proposed by the present invention is adopted to obtain the optimal combination of the main control factors from the multi-source features; and on this basis, three different machine learning models (such as LSSVM, RF, and GRU) are used to predict the periodic term displacement respectively, and the superiority of the main control factor screening strategy is evaluated according to the prediction indexes.

[0078] The process of the ternary optimization strategy for the main control feature factors is as follows:

[0079] S401. Construct a wrapper feature screening (RFE-XGBOOST-CV) model; define a recursive feature elimination strategy: eliminate the 10% features with the lowest contribution rate in each round; calculate the feature importance scores using 5-fold cross-validation, retain the feature subset with a cumulative contribution of ≥80%, and dynamically optimize the feature combination to finally obtain the optimal subset;

[0080] S402. Construct an embedded feature screening (CART) model and calculate the feature importance weights to quantify the non-linear interaction effects between factors and identify the features that have a significant cumulative impact on the periodic term displacement;

[0081] The CART tree algorithm is a classic embedded feature screening algorithm that can effectively handle the linear and non-linear relationships between features and target variables. This algorithm recursively constructs a binary decision tree and selects features using the criterion of minimizing the squared error. The basic steps are as follows:

[0082] ① Select the optimal splitting variable j and splitting point s that satisfy the following formula:

[0083]

[0084] where j and s are the feature and the splitting point of the feature respectively, and c1 and c2 represent the predicted values that minimize (y i -c1) 2 in this interval. Traverse all input features j to find the smallest (j, s) combination;

[0085] ② Divide the parent node according to the smallest (j, s) combination to obtain the output values corresponding to the two divided regions:

[0086]

[0087] In the formula: R1(j, s) and R2(j, s) are the divided regions; x j is the value of the sample on the j-th feature; N m is the number of samples in the region; is the predicted value;

[0088] ③ Repeat the above process in the child nodes until all leaf nodes are constructed. Each time a feature is split, select the feature and splitting point that can reduce the squared error to the greatest extent. Finally, measure the importance of features by calculating the reduction in the mean squared error for each feature and normalizing it.

[0089] S403. Construct a filtering feature screening (MIC) model to quantify the non-linear correlation between influencing factors and set a threshold to eliminate redundant features. Finally, fuse the RFE-XGBOOST-CV optimized subset, the CART feature importance ranking, and the MIC redundant elimination results to construct a master factor set with both physical significance (the reservoir water level-rainfall coupling mechanism) and statistical significance, as shown in Table 1.

[0090]

[0091] (Table 1)

[0092] MIC calculates the maximum mutual information value between different data subsets to measure the linear or non-linear correlation between variables. The expression is:

[0093]

[0094] where: x represents each influencing feature; η t represents the displacement of the landslide periodic term; P(x, η t ) is the joint probability density distribution function between the influencing feature and the periodic term displacement; P(x) and P(η t ) are the marginal probability density functions of the two variables; MIC(x; η t ) is the maximum mutual information coefficient between the variables; p and q are the grid distributions in the x and η t directions; B is a variable, b < n 0.6 .

[0095] S404. Input the main control factor sets into the LSSVM, RF, and GRU models respectively, and verify the scientificity and rationality of the proposed main control factor feature screening strategy through the prediction results of the periodic term displacement and indicators.

[0096] Finally, superimpose the predicted values of the trend term and the periodic term to obtain the cumulative displacement prediction result, and based on the true value and the predicted value of the cumulative displacement, obtain the error value to complete the prediction of the reservoir bank landslide displacement.

[0097] As Figure 4 and Figure 5 shown, to verify the effectiveness of this application, the embodiments select two monitoring points, X-03 and X-06, of the Xinpu landslide, and the monitoring data from January 10, 2020 to March 26, 2024. Use the main control factors determined by the proposed triple optimization strategy for main control factors in this application. On this basis, use three different models, LSSVM, RF, and GRU, to achieve landslide displacement prediction respectively. From Figure 4 and Figure 5 it can be seen that the main control factor strategy determined by the method described in this application can be applied to a variety of different models, and the overall prediction accuracy is relatively high. Even the prediction during the rapid step period of the reservoir bank landslide is very accurate, and the prediction error is small. It shows that using the proposed triple optimization strategy for main control factors to screen the main control factors in this application can effectively improve the prediction error caused by the redundant coupling of multi-source feature factors in the traditional method, and at the same time verify the universality of this application to different prediction algorithms.

[0098] In summary, this application adopts a full-chain displacement prediction system from data cleaning and preprocessing, construction of factor sets, time series decomposition, screening of main control factors, and sub-item modeling, which organically integrates the landslide displacement evolution mechanism and data-driven methods.

[0099] The above are only feasible implementation examples of this application, and do not limit the scope of rights of this application. Any equivalent changes made using this application are included in the protection scope of this application.

Claims

1. A method for predicting the displacement of a reservoir bank landslide based on a ternary optimization strategy of main control feature factors, characterized in that It includes the following steps: S1. Data collection, preprocessing, and construction of a multi-source heterogeneous variable feature set: Collect high-precision GNSS historical displacement monitoring time-series data of the landslide deformation monitoring points in the reservoir bank study area, synchronously observe rainfall, reservoir water level, and other data of the inducing factors affecting the landslide displacement evolution within the landslide range, and complete data cleaning. Based on the rainfall-reservoir water level coupling mechanism, construct a multi-source heterogeneous variable feature set, where the multi-source heterogeneous variable feature set includes various characteristic parameters such as antecedent effective rainfall, average rainfall, reservoir water level fluctuation change, lag-n period rainfall and reservoir water level change, and displacement change; S2. Data decomposition and reconstruction: Use CEEMDAN to reconstruct the cumulative displacement into trend-term and periodic-term displacement components for the multi-source heterogeneous variable feature set constructed in step S1, where the decomposition parameters are adaptively determined by the GWO optimization algorithm with the minimum envelope entropy as the objective function; S3. Trend-term displacement prediction modeling: Use the double exponential smoothing algorithm DES to predict the trend-term displacement component; S4. Selection and prediction of the main control characteristic factors of the periodic term: Conduct time-series correlation analysis on the multi-source heterogeneous variable feature set constructed in S1 and the periodic-term displacement component results obtained in S2; Subsequently, use the wrapper RFE-XGBOOST-CV and the embedded CART to perform feature selection on the multi-source heterogeneous variable feature set in sequence to obtain the key influencing factor set, and calculate the remaining feature filter MIC value based on this; Set the MIC correlation coefficient threshold, and screen to obtain the best combination of the final main control factors according to the MIC numerical values of each input feature; Finally, use three different machine learning models, namely LSSVM, RF, and GRU, to predict the periodic-term displacement respectively to verify the results of the main control factor selection; S5. Cumulative displacement prediction: Add the trend-term displacement predicted in step S3 and the periodic-term displacement predicted in step S4 to obtain the final predicted result of the landslide cumulative displacement, and evaluate the accuracy of the finally obtained predicted result of the landslide cumulative displacement by the coefficient of determination R 2 and the root mean square error RMSE index.

2. The reservoir bank landslide displacement prediction method based on the ternary optimization strategy of the main control feature factors according to claim 1, characterized in that, Calculate the antecedent effective rainfall, and the expression is: API t = λAPI t-Δt + P Δt Formula 1 Where: API t is the antecedent precipitation index at time t, λ is the rainfall influence attenuation coefficient, and P Δt is the rainfall from time t - Δt to time t.

3. The method for predicting the displacement of the reservoir bank landslide based on the ternary optimization strategy of the main control feature factors as described in claim 2, wherein, The expression of the cumulative displacement is as follows: D t = φ t + η t Formula 2 where: t is the time step; D t is the cumulative displacement; φ t is the trend-term displacement; η t is the periodic-term displacement.

4. The method for predicting the displacement of the reservoir bank landslide based on the ternary optimization strategy of the main control feature factors according to claim 3, wherein The expressions for decomposing into trend term and periodic term by CEEMDAN are as follows: In the formula: and respectively represent the displacement of the trend term and the displacement of the periodic term after decomposition; GWO and CEEMDAN respectively represent the GWO optimization algorithm and the CEEMDAN decomposition algorithm; D t represents the cumulative displacement; k1 and k2 represent the decomposition parameters to be optimized, and their values are determined by the GWO optimization algorithm with the minimum envelope entropy as the objective function.

5. The method for predicting the displacement of the reservoir bank landslide based on the ternary optimization strategy of the main control feature factors according to claim 4, wherein, The expression for the deterministic prediction of the trend-term displacement by the double exponential smoothing method DES is as follows: Where Y t+1 is the predicted value of the trend term displacement at time t-1; S t and S t-1 represent the exponential smoothing values at time t and t-1 respectively; b t and b t-1 represent the best estimates of the gradients at time t and t-1; both α and γ represent smoothing factors; φ t is the trend term displacement.

6. The method for predicting the displacement of the reservoir bank landslide based on the ternary optimization strategy of the main control feature factors as claimed in claim 5, wherein, The filter MIC described in step S4 is a correlation analysis method based on the mutual information coefficient, which measures the linear or non-linear correlation between variables by calculating the maximum mutual information value between different data subsets, and its expression is: In the formula: x represents each influencing feature; η t represents the displacement of the landslide periodic term; P(x, η t ) is the joint probability density distribution function between the influencing characteristics and the periodic term displacement; P(x) and P(η t ) are the marginal probability density functions of the two variables; MIC(x; η t ) is the maximum mutual information coefficient between the variables; p and q are the grid distributions in the x and η t directions; B is a variable, where b < n 0.6 .

7. The method for predicting the displacement of the reservoir bank landslide based on the ternary optimization strategy of the main control feature factors as described in claim 5, characterized in that The wrapper RFE-XGBOOST-CV described in step S4 is a method that combines the RFE-CV method and the XGBoost method; The RFE-CV is a wrapper feature selection method that combines feature importance evaluation and model validation. It evaluates the importance of each feature through the base model and gradually removes the least important features until the predetermined number of features is reached or all features have been evaluated; Finally, RFE-CV evaluates different feature subsets through cross-validation, calculates the average scores of each subset, and thus determines the best feature combination; The XGBoost is an efficient ensemble learning algorithm that improves the model performance by constructing and combining multiple decision trees. The XGBoost can focus on key variables through the features screened by RFE-CV and accurately screen out the optimal features.

8. The method for predicting the displacement of the reservoir bank landslide based on the ternary optimization strategy of the main control feature factors according to claim 5, characterized in that, The CART tree algorithm is a classic embedded feature screening algorithm that can effectively handle the linear and non-linear relationships between features and target variables. This algorithm constructs a binary decision tree recursively and selects features using the least squares error minimization criterion. The basic steps are as follows: ① Select the optimal splitting variable j and splitting point s that satisfy the following formula: Where: j and s are the feature and the segmentation point of the feature respectively, and c1 and c2 represent the predicted values that minimize (y i - c1) 2 Traverse all input features j to find the smallest (j, s) combination by taking the predicted value that minimizes (y - c1) for this interval; ② Divide the parent node according to the smallest (j, s) combination to obtain the output values of the two divided regions: R1(j, s) = {x | x (j) ≤ s}, R2(j, s) = {x | x (j) ≤ s} Equation 7 where: R1(j, s) and R2(j, s) are the divided regions; x j is the value of the sample on the j-th feature; N m is the number of samples in the region; is the predicted value; ③ Repeat the above process in the child nodes until all leaf nodes are constructed. Each time a feature is split, select the feature and splitting point that can reduce the squared error to the greatest extent. Finally, measure the importance of the features by calculating the reduction in the mean squared error for each feature and normalizing it.

Images