Landslide displacement prediction method based on fluctuation term and trend term decomposition
Through the combination of the fluctuation term, trend term decomposition and support vector regressor model of landslide displacement and the optimization parameters of gray wolf algorithm, the problem of low landslide displacement prediction accuracy in the existing technology is solved, and a more efficient landslide displacement prediction effect is achieved.
Patent Information
- Application Number
- CN202210258568.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-16
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2042-03-16
AI Technical Summary
The existing landslide displacement prediction methods have difficulties in accurately determining trend terms and period values, and the lack of parameter settings and generalization of artificial intelligence algorithms leads to low prediction accuracy.
The landslide displacement prediction method based on fluctuation term and trend term decomposition is adopted, and the landslide displacement is decomposed into trend term and fluctuation term through empirical modal decomposition. The prediction model is constructed and trained separately, combining gray correlation analysis and support vector regressor model, and using The Gray Wolf algorithm optimizes parameters to improve prediction accuracy.
The prediction performance of landslide displacement is significantly improved, the prediction accuracy of trend terms and fluctuation terms is improved, noise interference is reduced, and more ideal cumulative landslide displacement prediction results are achieved.
Smart Images

Figure CN114611948B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of landslide displacement prediction, and more specifically, relates to a landslide displacement prediction method based on the decomposition of fluctuation terms and trend terms. Background Art
[0002] Landslides are one of the most common geological disasters, with the characteristics of widespread distribution, frequent occurrence, great harm, and difficult control. If they are not properly handled, they will cause huge losses of life and property and have a significant social impact. Therefore, in order to achieve disaster prevention and mitigation, landslide displacement prediction has become a research hotspot for scholars today.
[0003] When most scholars decompose displacement, they usually separate the trend term displacement first. Common decomposition methods include: fitting methods based on historical displacement, such as gray model, polynomial trend fitting, etc., which are suitable for simple curves; filtering methods, such as sliding average method, are based on original displacement data and are easy to use, but the type and period value of trend term displacement are difficult to accurately determine; wavelet analysis method can extract multi-time scale information of landslide displacement, but the basis function and wavelet order need to be determined manually; smoothing prior method has a simple implementation process, but it requires multiple decomposition comparisons to determine the decomposition parameters, and the trend term displacement obtained in stages will still have step characteristics. In addition, in order to avoid the shortcomings of the above methods, series empirical mode decomposition (EMD) and ensemble empirical mode decomposition (EEMD) have also emerged. This method can thoroughly decompose the accumulated displacement of the landslide and significantly improve the prediction accuracy, but there are problems such as mode aliasing, low decomposition efficiency, and difficulty in completely eliminating noise. Therefore, Maria et al. proposed complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), adding adaptive white noise at each stage, which effectively solved the problems existing in the above-mentioned decomposition methods and showed very good decomposition effect on nonlinear time series.
[0004] In recent years, with the development of artificial intelligence and machine learning theory, more and more artificial intelligence algorithms have been applied to landslide displacement prediction, including BP neural network model (back propagation, BP), support vector regression model (support vector regression, SVR) and extreme learning machine model (extreme learning machine, ELM). BP neural network is easy to fall into local optimal, it is difficult to build a suitable network structure, the model convergence speed is slow, and it is often difficult to obtain ideal prediction results; in practical applications, extreme learning machines have simple parameter settings and good generalization, but the input of weights and biases is random, which reduces the prediction accuracy of the model. Summary of the invention
[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a landslide displacement prediction method based on the decomposition of fluctuation terms and trend terms. According to the internal and external factors that induce landslide deformation, the original landslide displacement is decomposed into trend term displacement and fluctuation term displacement respectively. Then, after respectively training the prediction models for prediction, the prediction results of the trend term displacement and the fluctuation term displacement are superimposed to obtain the prediction result of the landslide displacement, thereby improving the prediction performance of the landslide displacement.
[0006] In order to achieve the above-mentioned object of the invention, the landslide displacement prediction method based on the decomposition of fluctuation term and trend term of the present invention comprises the following steps:
[0007] S1: Determine N landslide displacement influencing factors according to the actual situation of the landslide displacement monitoring area, and then periodically collect the monitoring value F of each landslide influencing factor at T historical moments n (t) and the cumulative landslide displacement X(t) at each historical moment relative to the starting moment, where n = 1, 2, …, N, t = 1, 2, …, T;
[0008] S2: Perform empirical mode decomposition on the data sequence composed of the cumulative landslide displacement X(t) at T historical moments, and record the decomposition result as Among them, IMF k (t) represents the kth modal component obtained by decomposition, k = 1, 2, ..., K, K represents the number of modal components, and R (t) represents the residual component. Then the K modal component IMF is obtained by analysis k The fluctuation degree of the residual component R(t) and the component with a fluctuation degree less than the preset threshold is taken as the trend term and summed to obtain the trend term displacement C(t). The fluctuation term displacement W(t) is obtained by subtracting the trend term displacement C(t) from the cumulative landslide displacement X(t).
[0009] S3: Set the time deviation value D, and convert the D+1 monitoring values F of the N landslide displacement influencing factors n(td) and D fluctuation item displacements W(td′) are selected as candidate fluctuation item influencing factors, where d=0,1,…,D, d′=1,2,…,D, and then the (N+2)D candidate fluctuation item influencing factors and fluctuation item displacements W(t) are respectively subjected to correlation analysis, and the candidate fluctuation item influencing factors with correlation greater than a preset threshold are selected as fluctuation item influencing factors, and the number of fluctuation item influencing factors is recorded as K;
[0010] S4: for each fluctuation term displacement W(t), the value of the corresponding fluctuation term impact factor is used as the input in the training sample, and the fluctuation term displacement W(t) is used as the output of the training sample, thereby obtaining the fluctuation term training sample;
[0011] S5: construct a fluctuation item prediction model according to actual needs, and train it using the fluctuation item training samples obtained in step S4 to obtain a trained fluctuation item prediction model;
[0012] S6: Set the time deviation value R, and use the trend item displacement C(tr) of the previous R moments as the influencing factor of the trend item displacement C(t); for each trend item displacement C(t), use the corresponding trend item displacement C(tr) of the previous R moments as the input in the training sample, and use the trend item displacement C(t) as the output of the training sample, so as to obtain the trend item training sample;
[0013] S7: construct a trend item prediction model according to actual needs, and train it using the trend item training samples obtained in step S6 to obtain a trained trend item prediction model;
[0014] S8: When it is necessary to predict the cumulative sliding displacement of the landslide displacement monitoring area at time T′, the values of K fluctuation item influencing factors corresponding to time T′ are collected and input into the trained fluctuation item prediction model to obtain the fluctuation item displacement prediction value W(T′), and then the trend item displacement C(T′-r) of the previous R moments corresponding to time T′ is input into the trained trend item prediction model to obtain the trend item displacement prediction value C(T′), and finally the cumulative sliding displacement X(T′)=W(T′)+C(T′) at time T′ is calculated.
[0015] The landslide displacement prediction method based on the decomposition of fluctuation term and trend term of the present invention determines the landslide displacement influencing factors according to the actual situation of the landslide displacement monitoring area, collects the historical data of the landslide displacement influencing factors and the cumulative landslide displacement, decomposes the cumulative landslide displacement to obtain the trend term displacement and the fluctuation term displacement, screens the influencing factors of the fluctuation term displacement, constructs and trains to obtain the fluctuation term prediction model, then sets the influencing factors of the trend term displacement, constructs and trains to obtain the trend term prediction model, performs fluctuation term prediction and trend term prediction respectively, and superimposes the prediction results to obtain the prediction result of the cumulative sliding displacement. The present invention can effectively improve the prediction performance of landslide displacement. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 It is a flow chart of a specific implementation method of the landslide displacement prediction method based on the decomposition of fluctuation term and trend term of the present invention;
[0017] Figure 2 is a geological structure map of the landslide displacement monitoring area in this embodiment;
[0018] Figure 3 is a data graph of cumulative landslide displacement, rainfall, and reservoir water level in this embodiment;
[0019] Figure 4 is a decomposition result diagram of the cumulative landslide displacement in this embodiment;
[0020] Figure 5 is an example diagram of the displacement of the trend term and the displacement of the fluctuation term in this embodiment;
[0021] Figure 6 is a comparison diagram of fluctuation term displacement prediction of the fluctuation term prediction model in this embodiment;
[0022] Figure 7 is a comparison diagram of fluctuation term displacement prediction of the fluctuation term prediction model in this embodiment;
[0023] Figure 8 It is a comparison chart of the cumulative landslide displacement prediction in this embodiment. DETAILED DESCRIPTION
[0024] The specific implementation of the present invention is described below in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when the detailed description of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.
[0025] Example
[0026] Figure 1 FIG. 1 is a flow chart of a specific implementation method of the landslide displacement prediction method based on the decomposition of fluctuation term and trend term of the present invention. Figure 1As shown, the specific steps of the landslide displacement prediction method based on the decomposition of the fluctuation term and the trend term of the present invention include:
[0027] S101: Get raw data:
[0028] According to the actual situation of the landslide displacement monitoring area, N landslide displacement influencing factors are determined, and then the monitoring value F of each landslide influencing factor at T historical moments is periodically collected. n (t) and the cumulative landslide displacement X(t) at each historical moment relative to the starting moment, where n = 1, 2, …, N and t = 1, 2, …, T.
[0029] S102: Decomposition of accumulated landslide displacement:
[0030] The influencing factors of landslide displacement mainly include internal factors caused by its own geological conditions and external factors caused by changes in the external environment. Therefore, in the present invention, the cumulative displacement of the landslide is decomposed into fluctuation term displacement and trend term displacement. The former is mainly determined by external factors, and the latter is mainly determined by internal factors. They will be processed and predicted separately later.
[0031] The empirical mode decomposition is performed on the data sequence composed of the cumulative landslide displacement X(t) at T historical moments, and the decomposition result is recorded as Among them, IMF k (t) represents the kth modal component obtained by decomposition, k = 1, 2, ..., K, K represents the number of modal components, and R (t) represents the residual component. Then the K modal component IMF is obtained by analysis k The fluctuation degree of the residual component R(t) and the component with a fluctuation degree less than the preset threshold is taken as the trend term and summed to obtain the trend term displacement C(t). The fluctuation term displacement W(t) is obtained by subtracting the trend term displacement C(t) from the cumulative landslide displacement X(t).
[0032] Empirical mode decomposition (EMD) is a signal decomposition method that can decompose the original signal according to the fluctuation scale to obtain different intrinsic mode components (IMF). Ensemble empirical mode decomposition (EEMD) adds Gaussian white noise that obeys normal distribution on the basis of EMD decomposition, solves the modal aliasing problem of EMD decomposition, and realizes effective modal component separation. In order to further solve the problems existing in the decomposition of the original signal, the CEEMDAN algorithm is proposed based on the development of the EEMD algorithm. This method adds adaptive Gaussian white noise to the signal at each stage of decomposition, effectively reduces the reconstruction error problem in EEMD, and improves the decomposition efficiency. Therefore, in this embodiment, the empirical mode decomposition adopts the CEEMDAN algorithm, and its specific principles and steps can be found in the document "TORRES ME, COLOMINAS MA, SCHLOTTHAUER G, et al. A complete ensemble empirical mode decomposition with adaptive noise [C] / / IEEE International Conference on Acoustics, Speech and Signal Processing, IEEE, 2011: 4144-4147.".
[0033] As for the fluctuation degree of the component, the variance can be used to represent it, or other characterization parameters can be selected according to the actual situation.
[0034] S103: Determine the impact factor of the fluctuation term:
[0035] Set the time deviation value D, and convert the D+1 monitoring values F of the N landslide displacement influencing factors into n (td) and D fluctuation item displacements W(td′) are used as candidate fluctuation item influencing factors, where d=0,1,…,D and d′=1,2,…,D. Then, correlation analysis is performed on the (N+2)D candidate fluctuation item influencing factors and the fluctuation item displacement W(t) respectively, and the candidate fluctuation item influencing factors with correlation greater than the preset threshold are selected as the fluctuation item influencing factors. The number of fluctuation item influencing factors is recorded as K.
[0036] In this embodiment, the correlation analysis between the candidate fluctuation item influencing factors and the fluctuation item displacement adopts the grey correlation analysis, which is a method of measuring the correlation degree between factors based on the similarity or difference of the development trends between factors, that is, "grey correlation". The grey correlation analysis method is a commonly used correlation analysis method, and its principle and specific steps will not be repeated here.
[0037] S104: Determine the fluctuation item training sample:
[0038] For each fluctuation term displacement W(t), the value of the corresponding fluctuation term impact factor is used as the input in the training sample, and the fluctuation term displacement W(t) is used as the output of the training sample, thereby obtaining the fluctuation term training sample.
[0039] S105: Build and train a volatility prediction model:
[0040] A fluctuation term prediction model is constructed according to actual needs, and the fluctuation term training sample obtained in step S104 is used to train the model to obtain a trained fluctuation term prediction model.
[0041] S106: Construct trend item training samples:
[0042] The time deviation value R is set, and the trend item displacement C(tr) of the previous R moments is used as the influencing factor of the trend item displacement C(t), r = 1, 2, ..., R. For each trend item displacement C(t), the corresponding trend item displacement C(tr) of the previous R moments is used as the input in the training sample, and the trend item displacement C(t) is used as the output of the training sample, thereby obtaining the trend item training sample.
[0043] S107: Build and train trend item prediction model:
[0044] A trend item prediction model is constructed according to actual needs, and is trained using the trend item training samples obtained in step S106 to obtain a trained trend item prediction model.
[0045] S108: Landslide displacement prediction:
[0046] When it is necessary to predict the cumulative sliding displacement of the landslide displacement monitoring area at time T′, the values of K fluctuation item influencing factors corresponding to time T′ are collected and input into the trained fluctuation item prediction model to obtain the fluctuation item displacement prediction value W(T′), and then the trend item displacement C(T′-r) of the previous R moments corresponding to time T′ is input into the trained trend item prediction model to obtain the trend item displacement prediction value C(T′), and finally the cumulative landslide displacement X(T′)=W(T′)+C(T′) at time T′ is calculated.
[0047] According to the above description, the fluctuation term prediction model and the trend term prediction model have a great influence on the accuracy of the final landslide displacement prediction. According to the data characteristics in the present invention, this embodiment preferably uses the support vector regression machine model (SVR) as the fluctuation term prediction model and the trend term prediction model. The main idea of the support vector regression machine model is to map the low-dimensional input vector to the high-dimensional feature space through the selected mapping function, and search for the regression function in the high-dimensional feature space, thereby converting the low-dimensional nonlinear regression problem into a high-dimensional linear regression problem. The support vector regression machine model effectively solves the problems of small samples, high dimensions, and nonlinearity, has strong robustness, and has been widely used in prediction and other aspects. When using the support vector regression machine model in this embodiment, the kernel function is set to the radial basis function. However, in the application process, the values of the penalty factor C, the kernel function parameter g, and the insensitive parameter ε of the support vector regression machine model have a great influence on the prediction ability of the model, and directly determine the generalization ability of the model.
[0048] In order to achieve optimal performance, the present invention uses the grey wolf optimizer (GWO) to optimize the penalty factor C, kernel function parameter g and insensitive parameter ε of the support vector regression model. The grey wolf optimizer (GWO) is a new heuristic algorithm proposed by Mirjalili et al. in 2014 that imitates the grey wolf's hunting of prey in nature. The main contents of this method include five parts: social hierarchy, surrounding prey, hunting, attacking prey, and finding prey. The wolf pack is initialized to search for prey, and the three wolves α, β, and δ with the best adaptability in the population are retained each time, and the positions of other grey wolves are updated accordingly, and the optimal solution is finally found through continuous iteration. Compared with traditional algorithms such as particle swarm optimization algorithm PSO, genetic algorithm GA, artificial bee colony algorithm ABC, the grey wolf algorithm has the advantages of simple parameter setting, strong convergence performance, and high optimization accuracy. At the same time, the wolf pack conducts random search during the optimization process to avoid falling into the local optimum. Therefore, the use of the Grey Wolf Algorithm to optimize the parameters of the support vector regression machine model can greatly improve the prediction accuracy of the support vector regression machine model, and thus achieve accurate prediction of landslide displacement.
[0049] The parameter optimization process of the support vector regression model by the gray wolf algorithm in this embodiment is as follows:
[0050] (1) Taking the fluctuation item training samples or the trend item training samples as the data set, the data set is divided into a test set and a training set.
[0051] (2) Set the parameters to be optimized of the support vector regression model: the value range of the penalty factor C, the kernel function parameter g, and the insensitive parameter ε, and select the prediction evaluation index of the support vector regression model as the fitness function.
[0052] (3) Initialize the parameters of the gray wolf algorithm, set the population size and the maximum number of iterations, use the parameters to be optimized of the support vector regression model as the gray wolf individuals, and then initialize the positions of each gray wolf individual.
[0053] (4) Calculate the fitness value of each individual in the wolf pack, rank their fitness, retain the three gray wolf individuals with the strongest fitness, α, β, and δ, and record their positions.
[0054] (5) Calculate the distance between other gray wolf individuals and the three best gray wolves retained, update the position of each searched gray wolf, and update the corresponding parameters at the same time.
[0055] (6) Determine whether the maximum number of iterations has been reached, output the position of the gray wolf α with the best fitness, and obtain the optimized support vector regression machine model parameters C, g, and ε. Use the obtained optimal parameters to establish a displacement prediction model corresponding to the data set.
[0056] In order to better illustrate the technical effect of the present invention, a specific example is used to experimentally verify the present invention.
[0057] The landslide displacement monitoring area selected in this example is located in the reservoir area. The plane shape is round chair-shaped. The landslide is about 800m long, 360-1200m wide, with an average thickness of 30m and an area of 620,000m. 2 , volume 18.5 million m 3 . Figure 2 is the geological structure map of the landslide displacement monitoring area in this embodiment. Figure 2 As shown in the figure, the micro-topography type of the landslide displacement monitoring area is a steep slope-gentle slope on the left side of the Yangtze River. The landslide front edge section below the elevation of 180-250m has a large terrain slope, generally 30-45°, and is mostly in the reservoir water drawdown zone and its impact area; the landslide middle and rear edge area at an elevation of 250-300m is mainly gentle slope, with a terrain slope of generally 10-15°, and 3 platforms of varying sizes are developed, which are the main farmland areas and villagers' residential areas; above the elevation of 300m, it is mainly steep slope, which is the rear edge boundary of the landslide; the gully on the left side of the landslide, with a cutting depth of more than 300m, is the left boundary of the landslide. The plane shape of the landslide is "triangular dustpan shape", and the cross-section shape is "convex shape", with an average slope of about 20°. The front edge of the landslide displacement monitoring area is located on the southeast wing of the Wushan slope. The surface of the landslide area is mainly Quaternary loose accumulation layer, which is easy to soften. The sliding material structure is loose to slightly dense and has good permeability. The lower bedrock is the middle Triassic Badong Formation (T2b) marl, siltstone, silty mudstone and the upper Triassic Xujiahe Formation (T3xj) medium-coarse-grained feldspar quartz sandstone, gravel-containing feldspar sandstone, siltstone. The rock formation has a dip of 355°∠15°, and the rock surface is mostly in a strongly weathered state, which is a slippery stratum.
[0058] Through investigation, analysis and research, the landslide in the landslide monitoring area is a traction-type super-large soil landslide with rainfall-reservoir as the main inducing factor, so rainfall and reservoir water level are selected as factors affecting landslide displacement. In this embodiment, the data of the landslide monitoring area from July 2016 to November 2020 are used to obtain the cumulative landslide displacement, rainfall and reservoir water level of the landslide monitoring area on a weekly basis, wherein the cumulative landslide displacement takes the average value of the cumulative displacement this week, the reservoir water level takes the change in the reservoir water level this week, and the rainfall takes the sum of the rainfall this week, and finally 230 sets of data are obtained. 190 sets of data from July 2016 to February 2020 are selected as the training sample set, and 40 sets of data from February 2020 to November 2020 are taken as the test sample set. Figure 3 It is a data graph of cumulative landslide displacement, rainfall, and reservoir water level in this embodiment.
[0059] This embodiment uses the CEEMDAN algorithm to decompose the accumulated landslide displacement. Figure 4 is the decomposition result diagram of the cumulative landslide displacement in this embodiment. Figure 4 As shown in the figure, four significantly different IMF characteristic components and one residual component are obtained in this embodiment. It can be seen from the analysis that the fluctuation degree of the component IMF4(t) and the residual component R(t) is small, which can be considered as the actual displacement deformation under the dynamic action of the landslide. Therefore, the component IMF4(t) and the residual component R(t) are summed as the trend term displacement C(t), and the fluctuation term displacement C(t) is subtracted from the accumulated landslide displacement X(t) to obtain the fluctuation term displacement W(t). Figure 5 is an example diagram of the displacement of the trend term and the fluctuation term in this embodiment. Figure 5 As shown in the figure, the displacement of the trend term accounts for a large proportion, and the highest displacement is close to the total landslide displacement at the corresponding moment, so the accuracy of the trend term prediction is particularly important. Although the fluctuation term displacement accounts for a small proportion of the cumulative displacement, the accuracy of its predicted value will also affect the accuracy of the prediction results to a certain extent. Therefore, it is necessary to select an appropriate method based on the deformation characteristics of the landslide displacement to make a correct prediction in order to obtain better prediction results.
[0060] Next, it is necessary to determine the fluctuation item influencing factors. Set the time deviation value D=3, that is, take the rainfall F1(t), F1(t-1), F1(t-2), F1(t-3), the reservoir water level F2(t), F2(t-1), F2(t-2), F2(t-3), and the fluctuation item displacement W(t-1), W(t-2), W(t-3) as the alternative fluctuation item influencing factors, calculate the gray correlation between the alternative fluctuation item influencing factors and the fluctuation item displacement W(t), and take the alternative fluctuation item influencing factors with a gray correlation greater than 0.6 as the fluctuation item influencing factors, and finally obtain 8 fluctuation item influencing factors, including rainfall F1(t), F1(t-1), reservoir water level F2(t), F2(t-1), F2(t-2), and fluctuation item displacement W(t-1), W(t-2), W(t-3). Table 1 is the gray correlation between the 8 fluctuation item influencing factors and the fluctuation item displacement in this embodiment.
[0061]
[0062] Table 1
[0063] According to the selection results of the volatility influencing factors, relevant data are obtained from the training sample set to obtain the volatility training sample set. Then a support vector regression machine model is constructed as the volatility prediction model, whose input is 8 volatility influencing factors and output is the volatility displacement. The gray wolf algorithm is used to optimize the parameters of the volatility prediction model in this embodiment, where the population size is set to 40 and the maximum number of iterations is 100. Because the penalty factor C, kernel function parameter g and insensitive parameter ε of the SVR model are optimized, the dimension is 3, and the search intervals of the three parameters are C=[10 -6 , 10 6 ],g=[10 -6 , 10 6 ],ε=[10 -6 , 10 6 ], set the fitness function, and the fitness function is the MAPE (mean absolute percentage error) value of the test set. The final optimized parameter values are C = 802455.4068, g = 1.4072×10 -5 ,ε=1.2246×10 -6 .
[0064] The trained volatility prediction model is tested using the test sample set. Figure 6 : is a comparison diagram of the fluctuation term displacement prediction of the fluctuation term prediction model in this embodiment. Figure 6As shown in the figure, the root mean square error RMSE of the fluctuation displacement prediction results is 0.9900, the mean absolute error MAE is 0.6856, the mean absolute percentage error MAPE is 16.5153%, and the discrimination coefficient R2 is 0.9835. Compared with the existing related methods, the prediction results are better and more accurate.
[0065] For the trend item displacement prediction, the time deviation value R is set to 3, that is, the trend item displacements C(t-1), C(t-2), and C(t-3) of the previous three moments are used as the influencing factors of the trend item displacement C(t). Relevant data are obtained from the training sample set to obtain the trend item training sample set. Similar to the fluctuation item prediction model, a support vector regression machine model is constructed as the trend item prediction model, whose input is the three trend item influencing factors and the output is the trend item displacement. The gray wolf algorithm is also used for parameter optimization, and the optimized parameter values are C=94093.2659, g=2.0655×10 -5 ,ε=1.9312×10 -6 .
[0066] The trained trend item prediction model is tested using the test sample set. Figure 7 : is a comparison diagram of the fluctuation term displacement prediction of the fluctuation term prediction model in this embodiment. Figure 7 As shown, the root mean square error RMSE is 0.0209, the mean absolute error MAE is 0.0176, the mean absolute percentage error MAPE is 0.0069%, and the discrimination coefficient R2 is 0.999996. Figure 7 It can be found that the actual displacement curve of the trend term and the predicted displacement curve overlap to a large extent. The prediction results show that the trend term prediction effect is better.
[0067] The predicted value of the cumulative landslide displacement can be obtained by summing the predicted fluctuation term displacement and trend term displacement. Figure 8 is a comparison chart of the cumulative landslide displacement prediction in this embodiment. Figure 8 As shown in the figure, the root mean square error RMSE of the cumulative landslide displacement prediction results is 0.9846 mm, the mean absolute error MAE is 0.6818, the mean absolute percentage error MAPE is 0.2719%, and the discrimination coefficient R 2 The final cumulative landslide displacement prediction results show that the overall fitting and prediction results are good and ideal, which is in line with the expectations of the model design.
[0068] Finally, the technical effects of the present invention and the existing landslide displacement prediction method are compared. First, a landslide displacement prediction method without decomposition of fluctuation term and trend term is used as a comparison method. In this method, the landslide displacement prediction model adopts a support vector regression model, so this method is abbreviated as SVR. The method of the present invention also adopts a support vector regression model, and adopts the CEEMDAN algorithm when decomposing the fluctuation term and trend term, which is abbreviated as CEEMDAN-SVR. Table 2 is a comparison table of the prediction performance of the present invention and the comparison method SVR in this embodiment.
[0069] method RMSE MAE MAPE(%) <![CDATA[R 2 ]]> SVR 1.8819 1.3200 0.5234 0.9868 CEEMDAN-SVR 1.5020 1.1214 0.4424 0.9916
[0070] Table 2
[0071] As shown in Table 2, the prediction accuracy of the CEEMDAN-SVR method is significantly improved compared with the SVR method. It can be seen that after the cumulative displacement of the landslide is decomposed into fluctuation terms and trend terms, the trend term displacement and fluctuation term displacement obtained are more regular than the original displacement data, which reduces the interference of random factors and noise on the prediction process and can greatly improve the prediction accuracy. Then, targeted differentiation processing is carried out according to the characteristics of the data, and predictions are made separately, which can significantly improve the prediction accuracy.
[0072] In order to illustrate the influence of the parameter optimization of the gray wolf algorithm on the prediction performance of the present invention, the prediction performance of the parameter optimization method using the gray wolf algorithm (referred to as CEEMDAN-GWO-SVR) is compared with the prediction performance without parameter optimization. Table 3 is a comparison table of the prediction performance before and after the gray wolf algorithm is used for parameter optimization.
[0073] method RMSE MAE MAPE(%) <![CDATA[R 2 ]]> CEEMDAN-SVR 1.5020 1.1214 0.4424 0.9916 CEEMDAN-GWO-SVR 0.9845 0.6818 0.2719 0.9964
[0074] Table 3
[0075] As shown in Table 3, after using the Gray Wolf Algorithm for parameter optimization, RMSE and MAE decreased by 0.5175 and 0.4396 respectively, and MAPE increased to 0.1705%. 2 The increase of 0.0048 indicates that the prediction effect can be significantly improved after the Grey Wolf Algorithm is used to optimize the parameters of the prediction model.
[0076] In summary, the present invention can effectively improve the prediction effect of landslide displacement by decomposing the fluctuation term and the trend term and predicting them separately. In addition, the Grey Wolf Algorithm can be used to optimize parameters to further improve the prediction effect.
[0077] Although the above describes the illustrative specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.
Claims
1. A landslide displacement prediction method based on the decomposition of fluctuation term and trend term, characterized in that: The following steps are involved: S1: Determine N landslide displacement influencing factors according to the actual situation of the landslide displacement monitoring area, and then periodically collect the monitoring value F of each landslide influencing factor at T historical moments n (t) and the cumulative landslide displacement X(t) at each historical moment relative to the starting moment, where n = 1, 2, …, N, t = 1, 2, …, T; S2: Perform empirical mode decomposition on the data sequence composed of the cumulative landslide displacement X(t) at T historical moments, and record the decomposition result as Among them, IMF k (t) represents the kth modal component obtained by decomposition, k = 1, 2, ..., K, K represents the number of modal components, R (t) represents the residual component; then the K modal component IMF is obtained by analysis k The fluctuation degree of the residual component R(t) and the component with a fluctuation degree less than the preset threshold is taken as the trend term and summed to obtain the trend term displacement C(t). The fluctuation term displacement W(t) is obtained by subtracting the trend term displacement C(t) from the cumulative landslide displacement X(t). S3: Set the time deviation value D, and convert the D+1 monitoring values F of the N landslide displacement influencing factors n (td) and D fluctuation item displacements W(td′) are selected as candidate fluctuation item influencing factors, where d=0,1,…,D, d′=1,2,…,D, and then the (N+2)D candidate fluctuation item influencing factors and fluctuation item displacements W(t) are respectively subjected to correlation analysis, and the candidate fluctuation item influencing factors with correlation greater than a preset threshold are selected as fluctuation item influencing factors, and the number of fluctuation item influencing factors is recorded as K; S4: for each fluctuation term displacement W(t), the value of the corresponding fluctuation term impact factor is used as the input in the training sample, and the fluctuation term displacement W(t) is used as the output of the training sample, thereby obtaining the fluctuation term training sample; S5: construct a fluctuation item prediction model according to actual needs, and train it using the fluctuation item training samples obtained in step S4 to obtain a trained fluctuation item prediction model; S6: Set the time deviation value R, and use the trend item displacement C(tr) of the previous R moments as the influencing factor of the trend item displacement C(t), r = 1, 2, ..., R; for each trend item displacement C(t), use the corresponding trend item displacement C(tr) of the previous R moments as the input in the training sample, and use the trend item displacement C(t) as the output of the training sample, so as to obtain the trend item training sample; S7: construct a trend item prediction model according to actual needs, and train it using the trend item training samples obtained in step S6 to obtain a trained trend item prediction model; S8: When it is necessary to predict the cumulative sliding displacement of the landslide displacement monitoring area at time T′, the values of K fluctuation item influencing factors corresponding to time T′ are collected and input into the trained fluctuation item prediction model to obtain the fluctuation item displacement prediction value W(T′), and then the trend item displacement C(T′-r) of the previous R moments corresponding to time T′ is input into the trained trend item prediction model to obtain the trend item displacement prediction value C(T′), and finally the cumulative sliding displacement X(T′)=W(T′)+C(T′) at time T′ is calculated.
2. The landslide displacement prediction method according to claim 1, characterized in that: In step S2, the empirical mode decomposition adopts the CEEMDAN algorithm.
3. The landslide displacement prediction method according to claim 1, characterized in that: The correlation analysis in step S3 adopts grey correlation analysis.
4. The landslide displacement prediction method according to claim 1, characterized in that: The fluctuation item prediction model in step S5 and the trend item prediction model in step S7 adopt a support vector regression model.
5. The landslide displacement prediction method according to claim 4, characterized in that: The kernel function in the support vector regression model is a radial basis function.
6. The landslide displacement prediction method according to claim 4, characterized in that: The penalty factor C, kernel function parameter g and insensitive parameter ε in the support vector regression model are optimized using the grey wolf algorithm.
Citation Information
Patent Citations
Landslide displacement dynamic prediction method based on multiple influence factors
CN112270400A
