A method for predicting reservoir bank slope deformation during construction based on the BVAR model

Abstract

The application discloses a construction period reservoir bank slope deformation prediction method based on a BVAR model, and is characterized in that the following steps are specifically implemented: step 1, original slope deformation monitoring data are obtained as initial samples, and the original data are subjected to ADF testing; step 2, a BVAR model is constructed; step 3, a Gibbs sampler is used to obtain BVAR model parameters to be estimated and model prediction values; and step 4, final slope deformation prediction values and prediction intervals are obtained. The application considers uncertainty factors in slope deformation monitoring based on prediction intervals, models and predicts analyzes a slope deformation monitoring sequence, has good effects, and solves the problems of less construction period slope deformation monitoring data and no environmental quantity monitoring data.

Description

Technical Field

[0001] This invention belongs to the field of deformation monitoring technology and relates to a method for predicting the deformation of reservoir bank slopes during the construction period based on the BVAR model. Background Technology

[0002] The safety and stability of reservoir bank slopes in water conservancy projects are crucial to the overall safety of the project; landslides can have severe consequences. Among various safety monitoring projects for reservoir bank slopes, deformation monitoring is the most intuitive and important, serving as the direct basis for judging slope safety and stability. Therefore, establishing a slope deformation monitoring model based on reservoir bank slope deformation monitoring data is extremely important for determining slope deformation trends and safety status. During the construction period, slope deformation monitoring data is limited, and environmental parameters (such as water level, rainfall, and temperature) are lacking, making it difficult to establish traditional statistical models using various environmental parameters as influencing factors or nonlinear models based on machine learning. Compared to traditional modeling methods, time series models predict future values ​​of monitoring data by leveraging the inherent structure and patterns within the data itself, exhibiting high efficiency and stability. Many scholars have already applied time series models to slope engineering, accurately predicting slope deformation trends. Traditional time series models often perform poorly when there is limited slope deformation monitoring data during the construction period. Among numerous time series models, the Bayesian Vector Autoregression (BVAR) model has a significant advantage in modeling with small sample data. Numerous studies have shown that the BVAR model offers higher accuracy. Using this method to model and analyze slope deformation monitoring data during construction can solve key problems associated with traditional methods, such as modeling difficulties and poor modeling results. In conclusion, modeling and predictive analysis of slope deformation monitoring data during construction is a pressing challenge that needs to be addressed. Summary of the Invention

[0003] The purpose of this invention is to provide a method for predicting the deformation of reservoir bank slopes during the construction period based on the BVAR model, which solves the problems of insufficient monitoring data and lack of environmental monitoring data for slope deformation during the construction period in existing methods.

[0004] The technical solution adopted in this invention is a method for predicting the deformation of reservoir bank slopes during the construction period based on the BVAR model, which is implemented according to the following steps:

[0005] Step 1: Obtain the original slope deformation monitoring data as the initial sample, and perform the ADF test on the original data;

[0006] Step 2, construct the BVAR model;

[0007] Step 3: Obtain the parameters to be estimated and the predicted values ​​of the BVAR model using the Gibbs sampler;

[0008] Step 4: Obtain the final predicted value and prediction range of slope deformation.

[0009] The invention is further characterized by:

[0010] Step 1 uses the adftest function built into MATLAB software to perform ADF test on the original deformation monitoring data of the slope, and judges the stability of the monitoring data through ADF test.

[0011] Step 2 is implemented in the following steps:

[0012] Step 2.1: Verify the optimal lag order of the BVAR model;

[0013] Step 2.1 specifically involves: using the order_var_d.m program in the MATLAB library BayVAR to calculate the BIC value of the monitoring data, and taking the minimum BIC value as the optimal lag order of the BVAR model; to avoid excessive model freedom, the maximum lag order in the order_var_d.m program is set to 5.

[0014] Step 2.2: Set the hyperparameter combination of the BVAR model;

[0015] First, given the possible values ​​of the hyperparameters, the attenuation parameter d = {0.01, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1}, the overall compactness γ = {0.01, 0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 1}, and the relative compactness w = {0.1, 0.2, 0.3, 0.4, 0.5, 1}.

[0016] Step 2.3: Determine the optimal combination of hyperparameters based on the Theil U statistic;

[0017] The err_d.m script in BayVAR was used to perform trial calculations on all the above hyperparameter combinations to obtain the model accuracy evaluation index, namely the Theil U statistic. The smaller the Theil U statistic, the smaller the prediction error. The Theil U statistic of all the hyperparameter combinations obtained from the trial calculations were compared, and the hyperparameter combination with the smallest Theil U statistic was taken as the optimal hyperparameter combination.

[0018] Step 3 specifically involves using the bvar_gs_d.m script in BayVAR to obtain the parameters to be estimated and the predicted values ​​of the BVAR model.

[0019] Step 3 sets the number of Gibbs pre-iterations to 5000 to ensure parameter convergence, then discards the original iterations and sets the number of Gibbs effective iterations to 1000; each element in the BVAR model coefficient matrix β and each element in the covariance matrix Σ are parameters to be estimated, and the Gibbs sampler estimates the parameters.

[0020] The specific process for estimating the parameters using the Gibbs sampler in step 3 is as follows:

[0021] Step 3.1: Specify the mean of the prior distribution as the initial value of the initial value parameters of the coefficient matrix β and the covariance matrix Σ;

[0022] Step 3.2: Given the initial values ​​and monitoring data, the posterior distribution of the coefficient matrix is ​​obtained by combining the prior information of the coefficients, and the coefficients are sampled on the basis of satisfying the stability condition to obtain the estimated values ​​of the coefficients.

[0023] Step 3.3: Given the updated coefficient matrix and the known monitoring data, calculate the posterior distribution of the parameters in the covariance matrix using the prior information, and sample the distribution based on the stability condition to obtain an estimate of the covariance matrix.

[0024] Step 3.4: Store the BVAR model parameters and the iterative sequence of the model parameters generated during the iteration process;

[0025] Step 3.5, repeat steps 3.2-3.3, and after a sufficient number of iterations, the iterative sequence of model parameters converges and becomes stationary;

[0026] Step 3.6: Obtain model prediction samples through the probability distribution of the Gibbs iteration sequence.

[0027] In step 4, 1000 slope deformation prediction value samples are obtained by Gibbs sampling. The final prediction value of the BVAR model is obtained by taking the median value of the slope deformation prediction value samples, and the maximum and minimum values ​​in the samples are used as the prediction interval of the BVAR model.

[0028] In step 1, if the monitoring data is non-stationary, the adftest function outputs variable h = 0; if the monitoring data is stationary, the adftest function outputs variable h = 1; if the original monitoring data is non-stationary, the original monitoring data is differencing by order n (n = 1, 2, 3...) and then the ADF test is performed again until the data becomes stationary.

[0029] The beneficial effects of this invention are: by considering the uncertainties in slope deformation monitoring based on the prediction interval, it effectively models and predicts slope deformation monitoring data, achieving good results. It effectively solves the key problem of modeling difficulties caused by the limited amount of slope deformation monitoring data and the lack of environmental monitoring data during the construction period in traditional modeling methods, providing a new and feasible method for slope modeling and analysis. The BVAR model solves the problems of difficult modeling and poor modeling results of traditional methods, demonstrating good predictive performance in slope modeling and predictive analysis, and advancing the analytical research on slope deformation prediction in water conservancy projects. Attached Figure Description

[0030] Figure 1 This is a flowchart of the construction-period reservoir bank slope deformation prediction method based on the BVAR model of the present invention;

[0031] Figure 2 This is the process line of the original monitoring data of the M07ZPR multi-point displacement meter in this invention;

[0032] Figure 3 This is the result of the hysteresis order test of measuring point M07ZPR(5m) in this invention;

[0033] Figure 4 This is the result of the hysteresis order test at measuring point M07ZPR (10m) in this invention;

[0034] Figure 5 This is the result of the hysteresis order test at measuring point M07ZPR (20m) in this invention;

[0035] Figure 6 This is the result of the hysteresis order test at measuring point M07ZPR (30m) in this invention;

[0036] Figure 7 This invention uses the BVAR model to model the monitoring data of measuring point M07ZPR (5m) and obtains the ten-step prediction value and prediction range.

[0037] Figure 8 This invention uses the BVAR model to model the monitoring data of measuring point M07ZPR (10m) to obtain ten-step predicted values ​​and prediction intervals;

[0038] Figure 9 This invention uses the BVAR model to model the monitoring data of measuring point M07ZPR (20m) to obtain ten-step predicted values ​​and prediction intervals;

[0039] Figure 10 This invention uses the BVAR model to model the monitoring data of measuring point M07ZPR (30m) and obtains the ten-step prediction value and prediction range. Detailed Implementation

[0040] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0041] The flowchart of the construction-phase reservoir bank slope deformation prediction method based on the BVAR model of this invention is as follows: Figure 1 As shown, please follow these steps:

[0042] Step 1: Obtain the original slope deformation monitoring data as the initial sample, and perform the ADF test on the original data;

[0043] The `adftest` function, built into MATLAB, was used to perform an AdF test on the original slope deformation monitoring data to determine the stationarity of the data. If the monitoring data is non-stationary, the `adftest` function outputs h = 0; if the monitoring data is stationary, the `adftest` function outputs h = 1. If the original monitoring data is non-stationary, it is then differxed by order n (n = 1, 2, 3, ...) and subjected to the ADF test again until the data becomes stationary.

[0044] Step 2, construct the BVAR model;

[0045] Step 2.1: Verify the optimal lag order of the BVAR model;

[0046] Step 2.1 specifically involves: using the order_var_d.m program in the MATLAB library BayVAR to calculate the BIC value of the monitoring data, and taking the minimum BIC value as the optimal lag order of the BVAR model; to avoid excessive model freedom, the maximum lag order in the order_var_d.m program is set to 5.

[0047] Step 2.2: Set the hyperparameter combination of the BVAR model;

[0048] First, given the possible values ​​of the hyperparameters, the attenuation parameter d = {0.01, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1}, the overall compactness γ = {0.01, 0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 1}, and the relative compactness w = {0.1, 0.2, 0.3, 0.4, 0.5, 1}.

[0049] Step 2.3: Determine the optimal combination of hyperparameters based on the Theil U statistic;

[0050] The BayVAR script err_d.m was used to perform trial calculations on all the above hyperparameter combinations to obtain the model accuracy evaluation index, namely the Theil U statistic. A smaller Theil U statistic indicates a smaller prediction error. The Theil U statistics of all the hyperparameter combinations obtained from the trial calculations were compared, and the hyperparameter combination with the smallest Theil U statistic was selected as the optimal hyperparameter combination.

[0051] Step 3: Obtain the parameters to be estimated and the predicted values ​​of the BVAR model through the Gibbs sampler.

[0052] Step 3 specifically involves using the bvar_gs_d.m script in BayVAR to obtain the parameters to be estimated and the model predictions. The Gibbs pre-iteration count is set to 5000 to ensure parameter convergence. Then, the original iterations are discarded, and the effective Gibbs iteration count is set to 1000.

[0053] Each element in the coefficient matrix β and each element in the covariance matrix Σ of the BVAR model are parameters to be estimated. The process of estimating these parameters using the Gibbs sampler includes:

[0054] Step 3.1: Specify the mean of the prior distribution as the initial value of the initial value parameters of the coefficient matrix β and the covariance matrix Σ;

[0055] Step 3.2: Given the initial values ​​and monitoring data, the posterior distribution of the coefficient matrix is ​​obtained by combining the prior information of the coefficients, and the coefficients are sampled on the basis of satisfying the stability condition to obtain the estimated values ​​of the coefficients.

[0056] Step 3.3: Given the updated coefficient matrix and the known monitoring data, calculate the posterior distribution of the parameters in the covariance matrix using the prior information, and sample the distribution based on the stability condition to obtain an estimate of the covariance matrix.

[0057] Step 3.4: Store the BVAR model parameters and the iterative sequence of the model parameters generated during the iteration process;

[0058] Step 3.5, repeat steps 3.2-3.3, and after a sufficient number of iterations, the iterative sequence of model parameters converges and becomes stationary;

[0059] Step 3.6: Obtain model prediction samples through the probability distribution of the Gibbs iteration sequence.

[0060] Step 4: Obtain the final predicted value and prediction range of the slope deformation monitoring data.

[0061] 1000 slope deformation prediction samples were obtained by Gibbs sampling. The final prediction value of the BVAR model was obtained by taking the median value of the slope deformation prediction sample, and the maximum and minimum values ​​in the sample were used as the prediction interval of the BVAR model.

[0062] The principle of the BVAR model in this invention is as follows: Assume a VAR model with p-order lag contains m variables: {y1, y2, ..., y m If the model does not contain a constant term, then the VAR model system can be represented in the following form:

[0063]

[0064] In the formula, ai,j,τ Let y be the variable in the i-th equation. j The τ-th order lag term y j-τ coefficient, u it Let be the random perturbation vector in the i-th equation. In the above VAR model, 2m needs to be determined. 2 p hyperparameters: including m 2 p prior mean δ i,j,τ and m 2 p prior variances

[0065] Bayesian methods can incorporate prior information into VAR models, significantly reducing the number of parameters that need to be estimated in the model system. This invention employs the Minnesota prior and includes the following basic assumptions:

[0066] (1) Normality: random perturbation vector u t =(u1,u2,…,u m ) T Follows a multivariate normal distribution N m (0,Σ);

[0067] (2) Independence: covariance matrix Σ and model coefficients a i,j,τ They are independent of each other;

[0068] (3) The prior distribution of the covariance matrix Σ is taken as the diffusion prior distribution, i.e., π(Σ)∝|Σ| -(m+1) / 2 ;

[0069] (4) Model coefficients a i,j,τ They are independent of each other, and a i,j,τ Follows a normal distribution

[0070] (5)a i,j,τ The prior expectation is δ i,j,τ Its value is determined by formula (2);

[0071]

[0072] (6) Standard deviation S i,j,τ Determined by formula (3), where γ is the overall compactness;

[0073]

[0074] Wherein, g(τ) is the tightness of the τ-order lagged variable relative to the first-order variable. The g(τ) function in this invention adopts the harmonic tightness function shown in formula (4).

[0075] g(τ)=τ -d ,d>0 (4)

[0076] f(i,j) is the relative compactness function, and its value is determined according to formula (5), where w ij It is a constant between 0 and 1. Therefore, choosing the parameter f(i,j) is transformed into determining the size of the hyperparameter w;

[0077]

[0078] s i / s j s is the ratio of the standard deviations of the autoregressive residuals of the i-th equation to those of the j-th equation. i The standard deviation of the residuals is obtained from the least squares regression of the i-th equation with respect to the constant term and its p-th lag term. Therefore, the prior variance of all parameters is obtained through formula (3). The magnitude of the value is transformed into the magnitude of the three hyperparameters (γ, d, and w).

[0079] The selection criteria for the magnitudes of hyperparameters (γ, d, and w) are determined by the prediction accuracy. In the BVAR model system, prediction accuracy is determined by the Theil U statistic. The Theil U statistic is the ratio of the root mean square error of the prediction model under study to the root mean square error of the random walk-based prediction model, i.e.:

[0080]

[0081] In the formula, A t F represents the actual value of the monitoring data at point t. M,t This represents the predicted value based on prediction method M, while F... RW,t These are predictions based on a random walk model. If U = 1.0, it means that the prediction accuracy of method M is basically the same as that of the random walk model; if U < 1.0, it means that the prediction accuracy of method M is higher than that of the random walk model; if U > 1.0, it means that the prediction accuracy of method M is lower than that of the random walk model. The closer the U value is to 0, the better the prediction effect. Therefore, when the U value reaches its minimum, the optimal hyperparameter combination can be determined.

[0082] If we convert the model in equation (1) into an unrelated model of the following form:

[0083] Y i =Zβ i +ε i ,i=1,2,…,m (7)

[0084] in:

[0085]

[0086]

[0087] Therefore, the Minnesota prior distribution of the model parameters can be abbreviated as:

[0088] (β|Σ)~N(μ0,V0),π(Σ)∝|Σ| -(m+1) / 2 (8)

[0089] Here, the components of μ0 are either 0 or 1, and the covariance matrix V0 is composed of n diagonal matrices, where the elements of each diagonal matrix are... The composition consists of off-diagonal blocks where all elements are zero. According to Bayes' theorem, the joint posterior probability density function of the parameters (β, Σ) is:

[0090]

[0091] in:

[0092] The aforementioned posterior distribution contains information about the parameters, such as the mean and variance, but the density function is too complex. Therefore, the posterior estimation of each parameter is achieved through Gibbs sampling, updating only one component at a time, thus transforming the high-dimensional problem into a low-dimensional one.

[0093] A sufficiently large number of Gibbs sampling iterations is required for the iterative sequence to reach a stationary distribution, i.e., removing sample points from the initially unstable iterative sequence. Therefore, it is necessary to set the pre-iteration number and the effective iteration number to ensure the convergence of the iterative sequence. The probability distribution of the Gibbs iterative sequence yields the predicted value samples for the BVAR model. The median of these predicted value samples is then used as the final predicted value of the BVAR model, and the maximum and minimum values ​​in the samples are used as the prediction intervals for the BVAR model.

[0094] Example

[0095] This invention presents an embodiment of a method for predicting reservoir bank slope deformation during construction based on a BVAR model. A BVAR model is established using monitoring data from the M07ZPR multi-point displacement gauge on the left bank slope of the Huangjinxia Hydropower Project. The original monitoring data process line is shown below. Figure 2 As shown. Figure 2 The M07ZPR multi-point displacement meter includes four measuring points: M07ZPR(5m), M07ZPR(10m), M07ZPR(20m), and M07ZPR(30m).

[0096] Typically, the existing sample data T is divided into two periods, T1 and T~T1. Data within period T1 is used to estimate BVAR model parameters and make predictions, while data within period T~T1 is used to calculate and compare prediction errors and determine the final hyperparameter values. To ensure sufficient data for model estimation and prediction, the first 122 monitoring data points for each monitoring point are selected as the initial estimation interval for the model, predicting forward for ten periods.

[0097] Performing the Augmented Dickey-Fuller (ADF) test on the original data is to prevent the "spurious regression" phenomenon from occurring in the slope deformation monitoring data during the modeling process, and to ensure the validity and robustness of the results.

[0098] The `adftest` function, built into MATLAB, was used to perform an AdF test on the raw monitoring data from four measuring points: M07ZPR(5m), M07ZPR(10m), M07ZPR(20m), and M07ZPR(30m). Since the raw data failed the ADF test, first-order differencing was performed on all data. The ADF test results are shown in Table 1.

[0099] Table 1 shows the ADF test results of the slope deformation monitoring data in this embodiment.

[0100]

[0101] Note: In the ADF test, c is the constant term, t is the trend term, and k is the lag order; the lag order k is selected based on the principle of minimizing the Bayesian Information Criterion (BIC).

[0102] The data after first-order differencing are all stationary at the 5% significance level and can be used to build a BVAR model.

[0103] The BIC value of the monitoring data was calculated using the order_var_d.m script in BayVAR to verify the optimal lag order for four measuring points: M07ZPR(5m), M07ZPR(10m), M07ZPR(20m), and M07ZPR(30m). To avoid excessive model freedom, the maximum lag order was set to 5. The optimal lag order for the four measuring points was determined based on the minimum BIC value. Therefore, the optimal lag order for the M07ZPR(5m) measuring point is 2. Figure 3 As shown, the optimal lag order for the M07ZPR (10m) measuring point is 3. Figure 4 As shown, the optimal lag order for both measuring points M07ZPR(20m) and M07ZPR(30m) is 1. Figures 5-6 As shown.

[0104] A set of possible values ​​for the hyperparameters (attenuation parameter d, overall compactness γ, and relative compactness w) of the BVAR model is defined. Specifically, d = {0.01, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1}, γ = {0.01, 0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 1}, and w = {0.1, 0.2, 0.3, 0.4, 0.5, 1}. The hyperparameter combinations for the four measuring points M07ZPR (5m), M07ZPR (10m), M07ZPR (20m), and M07ZPR (30m) will be selected from this set.

[0105] The err_d.m script in BayVAR was used to perform trial calculations on all hyperparameter combinations for the four measurement points to verify the model accuracy, i.e., the Theil U statistic. A smaller Theil U statistic indicates a smaller prediction error. The Theil U statistics of all hyperparameter combinations obtained from the trial calculations were compared, and the hyperparameter combination with the smallest Theil U statistic was obtained, as shown in Table 2, and was taken as the optimal hyperparameter combination for the BVAR model.

[0106] Table 2 shows the hyperparameter combinations that minimize the TheilU statistic of the BVAR model in this invention.

[0107]

[0108] The bvar_gs_d.m script in BayVAR was used to obtain the parameters to be estimated and the predicted values ​​of the BVAR model. In this invention, the number of iterations was set to the default settings in the program. The Gibbs pre-iteration count was set to 5000 to ensure parameter convergence. Then, the original iterations were discarded, and the effective Gibbs iteration count was set to 1000. After the Gibbs iteration sequence converged and stabilized, ten-step predicted value samples of slope deformation monitoring data (a total of 10 × 1000 predicted sample data) were obtained. The median value of the sample data was calculated to obtain the ten-step predicted value of the slope deformation monitoring data, and the maximum and minimum values ​​of the sample data were calculated to obtain the prediction interval of the slope deformation monitoring data. The ten-step prediction results of the monitoring data at four measuring points M07ZPR (5m), M07ZPR (10m), M07ZPR (20m), and M07ZPR (30m) are shown below. Figures 7-10 As shown. By Figures 7-10 It can be seen that among the prediction results of the above four measurement points, 90% of the predicted values ​​are within the prediction interval of the BVAR model, indicating that the BVAR model has a good short-term prediction effect in terms of interval prediction.

[0109] Summary of Implementation Examples

[0110] To comprehensively evaluate the prediction performance of the established BVAR model, the sum of squared residuals (SSE), mean absolute error (MAE), and root mean square error (RMSE) of the ten-step prediction of the traditional VAR model and the BVAR model are compared, as shown in Table 3.

[0111] Table 3 shows the comparison results of the ten-step prediction errors of the VAR model and the BVAR model in this invention.

[0112]

[0113] Table 3 shows that the ten-step prediction error of the BVAR model at the four measuring points M07ZPR (5m), M07ZPR (10m), M07ZPR (20m), and M07ZPR (30m) is generally smaller than that of the traditional VAR model. All three error indices of the BVAR model (SSE, MAE, and RMSE) are below 0.1, indicating that the overall prediction effect of the BVAR model is good. The prediction results of the above examples demonstrate that using the BVAR model for short-term prediction of slope deformation monitoring data during construction is effective and can provide a basis for practical engineering.

Claims

1. A method for predicting deformation of a bank slope during construction based on a BVAR model, characterized in that, Specifically, the following steps are implemented: Step 1, obtaining the original data of slope deformation monitoring as initial samples, and performing ADF test on the original data; Step 2, constructing a BVAR model; Step 3, obtaining the estimated parameters of the BVAR model and the model prediction value through a Gibbs sampler; Step 4, obtaining the final prediction value and prediction interval of the slope deformation; In step 1, the adftest function in MATLAB software is used to perform ADF test on the original slope deformation monitoring data, and the stationarity of the monitoring data is determined through ADF test; Step 2 is implemented according to the following steps: Step 2.1, testing the optimal lag order of the BVAR model; Step 2.1 is specifically: using the order_var_d.m program in the MATLAB library BayVAR to calculate the BIC value of the monitoring data, and taking the minimum BIC value as the optimal lag order of the BVAR model; in order to avoid too many degrees of freedom of the model, the maximum lag order in the order_var_d.m program is taken as 5 orders; Step 2.2, setting the BVAR model hyperparameter combination; First, given the possible values of the hyperparameters, among which the decay parameter = {0.01, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1}, the total tightness = {0.01, 0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 1}, the relative tightness = {0.1, 0.2, 0.3, 0.4, 0.5, 1}; Step 2.3, determining the optimal hyperparameter combination according to the Theil U statistic; The err_d.m script program in BayVAR is used to perform trial calculation on the BVAR model hyperparameter combination set in step 2.1, and the model precision evaluation index, i.e. Theil U statistic, is obtained; the smaller the Theil U statistic, the smaller the prediction error; all the Theil U statistics of the hyperparameter combinations obtained by trial calculation are compared, and the hyperparameter combination with the smallest Theil U statistic is taken as the optimal hyperparameter combination; Step 3 is specifically: using the bvar_gs_d.m script program in BayVAR to obtain the estimated parameters of the BVAR model and the model prediction value; The step 3 sets the Gibbs pre-iteration number to 5000 to ensure the convergence of the parameters, and then discards the original iteration, and sets the effective Gibbs iteration number to 1000; the BVAR model coefficient matrix Each element in the matrix is a to-be-estimated parameter, and the Gibbs sampler estimates the parameters; Each element in the matrix is a to-be-estimated parameter, and the Gibbs sampler estimates the parameters; The specific process of estimating parameters by the Gibbs sampler in step 3 is as follows: Step 3.1, specify the mean of the prior distribution as the initial value of the coefficient matrix and the initial value of the initial value parameter of the covariance matrix ​ Step 3.2, under the condition that the initial value and the monitoring data are known, the posterior distribution of the coefficient matrix is obtained by combining the prior information of the coefficient, and the coefficient is sampled on the basis of meeting the stability condition to obtain the estimated value of the coefficient; Step 3.3, under the condition that the updated coefficient matrix and the monitoring data are known, the posterior distribution of the parameter in the covariance matrix is calculated by combining the prior information of the parameter, and the parameter is sampled on the basis of meeting the stability condition to obtain the estimated value of the covariance matrix; Step 3.4, storing the BVAR model parameters and the iteration sequence of the model parameters generated in the iteration process; Step 3.5, repeating step 3.2-step 3.3, after a sufficient number of iterations, the iteration sequence of the model parameters converges and is stable; Step 3.6, obtaining the model prediction value sample through the probability distribution of the Gibbs iteration sequence.

2. The BVAR model-based construction period bank slope deformation prediction method according to claim 1, characterized in that, In step 4, 1000 slope deformation prediction value samples are obtained by Gibbs sampling, the final prediction value of the BVAR model is obtained by taking the median of the slope deformation prediction value samples, and the maximum value and the minimum value in the sample are taken as the prediction interval of the BVAR model.

3. The BVAR model-based construction period bank slope deformation prediction method according to claim 1, characterized in that, If the monitoring data is non-stationary in step 1, the adftest function outputs variable h = 0; if the monitoring data is stationary, the adftest function outputs variable h = 1; if the original monitoring data is non-stationary, the original monitoring data is differentiated by n orders and then subjected to the ADF test again until the data becomes stationary.

Images