Grouting effect prediction method for underground water-sealed cave depot based on integrated learning model

Abstract

The invention relates to an integrated learning model-based underground water-sealed cave depot grouting effect prediction method, which comprises the following steps of: obtaining environmental parameters and construction parameters of an underground water-sealed cave depot, and inputting the environmental parameters and the construction parameters into an integrated learning-based grouting effect prediction model to obtain a grouting effect prediction result; the grouting effect prediction model is trained by using a training set and is obtained by synchronously optimizing trained model parameters by using a simulated annealing algorithm; the training set comprises geological parameters, hydrological parameters and historical construction parameters; wherein the grouting effect prediction model comprises a random forest model, a gradient boosting tree model, an AdaBoost model and an ANN model. According to the method, a machine learning technology is combined, especially hyper-parameter optimization of integrated learning and a simulated annealing algorithm is adopted, and a scientific basis is provided for rock mass grouting seepage control construction decision in an underground water-sealed cave depot project, so that the overall construction quality and energy efficiency of a grouting project are improved.

Description

Technical Field

[0001] The invention relates to the technical field of rock mass grouting seepage control construction in underground water-sealed cavern engineering, and in particular to a method for predicting the grouting effect of underground water-sealed cavern based on an integrated learning model. Background Art

[0002] As an important part of the national energy reserve system, underground water-sealed caverns are not only used to store crude oil, but also other petroleum products, such as refined oil and LPG. It is a way to change and utilize natural rock mass through artificial intervention. It has the advantages of safety, environmental protection, strong concealment, low construction and operation costs, and almost no occupation of fertile farmland and commercial land. The construction and operation of underground water-sealed caverns have brought huge benefits to the national economy and social environment, especially playing an important role in improving national energy security and ensuring the stability of energy supply.

[0003] In the process of building underground water-sealed caverns, the rock mass is affected by the intertwined influence of various factors such as blasting and excavation, and various cracks are formed on the surface and inside of the cavern surrounding rock, allowing groundwater to seep through them. The seepage effect will reduce the stability of the cavern surrounding rock and increase the pumping and drainage costs of the cavern operation. Therefore, it is very necessary to carry out seepage control treatment on the cavern.

[0004] At present, grouting is the main means of controlling the seepage field of caverns. The integrity of the cavern rock structure is improved by injecting slurry into rock joints, fissures, alteration zones and other adverse geological structures. After the slurry solidifies, it can significantly improve the strength and integrity of the rock mass and effectively prevent leakage. However, due to the concealment of rock grouting, it is difficult to intuitively evaluate the distribution and solidification of the slurry, making it difficult to accurately evaluate the construction quality and effect. Therefore, if the grouting quality of each hole sequence can be predicted in advance conveniently and effectively during the grouting construction process, it will be of great significance to the overall control of the grouting construction quality and the improvement of grouting energy efficiency.

[0005] With the development of computer technology, numerical simulation analysis methods such as finite element method can be used to simulate the grouting process. However, since the grouting project is a complex and hidden system engineering, there are many factors that affect the grouting effect, so it is difficult to play a role in the evaluation of the grouting effect. Moreover, numerical simulation methods are often accompanied by the disadvantages of complex modeling process and high modeling time cost, which cannot meet the needs of actual on-site construction. Therefore, it is necessary to establish a fast and accurate intelligent model based on data-driven to predict the grouting effect and provide assistance for grouting construction decision-making.

[0006] With the development of artificial intelligence technology, the use of machine learning methods has provided feasibility for grouting effect prediction. This method can effectively solve the shortcomings of grouting effect evaluation caused by the concealment of rock grouting through data-driven. Early predictions of rock grouting effects were mostly composed of simple weak learners, such as decision trees. These prediction models will lead to unsatisfactory prediction results when processing insufficient sample data or too many input factors. Summary of the invention

[0007] In order to solve the problems existing in the above-mentioned prior art in predicting the effect of rock grouting, such as great evaluation difficulty, limited simulation methods, high complexity of prediction models, difficulty in hyperparameter optimization, insufficient generalization ability, and lack of real-time prediction and decision support, the present invention provides a method for predicting the grouting effect of underground water-sealed caverns based on an integrated learning model to solve the existing technical problems. The method can provide a reliable method support for estimating the grouting volume of the area to be irrigated, and has important engineering application value. Moreover, due to its high requirements for controlling the amount of rock seepage, the model is not only suitable for specific projects, but is also expected to be extended to other underground grouting projects.

[0008] To achieve the above object, the present invention provides the following solutions:

[0009] A method for predicting the grouting effect of underground water-sealed caverns based on an integrated learning model, comprising:

[0010] Obtaining environmental parameters and construction parameters of an underground water-sealed cavern, inputting the environmental parameters and the construction parameters into a grouting effect prediction model based on ensemble learning, and obtaining a grouting effect prediction result; the grouting effect prediction model is trained using a training set, and the trained model parameters are synchronously optimized using a simulated annealing algorithm; the training set includes: geological parameters, hydrological parameters, and historical construction parameters;

[0011] The grouting effect prediction model includes: a random forest model, a gradient boosting tree model, an AdaBoost model and an ANN model; the random forest model, the gradient boosting tree model and the AdaBoost model are all connected to the ANN model.

[0012] Optionally, the geological parameters include: surrounding rock grade, joint surface density and alteration zone density;

[0013] The hydrological parameters include: water permeability before irrigation;

[0014] The construction parameters include: grouting pressure, hole sequence and elevation parameters.

[0015] Optionally, obtaining the grouting effect prediction result includes:

[0016] The environmental parameters and the construction parameters are input into a random forest model, a gradient boosting tree model, and an AdaBoost model to obtain a first prediction result; the random forest model is a random forest model that uses a decision tree as a base learner to perform a Bagging method integration, and the gradient boosting tree model is a gradient boosting tree model that uses a decision tree as a base learner to perform a Boosting method integration;

[0017] The first prediction result is input as input data into an ANN model to obtain the grouting effect prediction result; the ANN model is an ANN model integrated by the Stacking method using the base learner.

[0018] Optionally, synchronously optimizing the trained model parameters using a simulated annealing algorithm includes:

[0019] Step 1: Take the model parameters as the initial solution and set the initial temperature and cooling coefficient;

[0020] Step 2: updating the initial temperature according to the cooling plan and randomly selecting a new solution in the neighborhood of the initial solution;

[0021] Step 3: Calculate the fitness difference of the new solution, and judge the fitness difference. If the fitness difference is less than a preset value, accept the new solution; if the fitness difference is greater than the preset value, accept the new solution according to the Metropolis criterion.

[0022] Step 4: Take the new solution as the current solution and repeat steps 1 to 3 until the updated temperature is lower than the target threshold or reaches a preset number of iterations to obtain the optimal solution.

[0023] Optionally, the model parameters include: the learning rate and number of weak learners of the AdaBoost model, the learning rate, maximum tree depth, number of trees, and minimum number of samples for each leaf node of the gradient boosting tree model, the number of trees, maximum tree depth, and minimum number of leaf node samples of the random forest model, and the neuron weights, number of hidden layers, and activation function of the ANN model.

[0024] Optionally, calculating the fitness difference of the new solution includes:

[0025] Δf=f(i')-f(i);

[0026] Among them, Δf is the fitness difference, f(i') is the fitness of the new solution, and f(i) is the fitness difference of the current solution.

[0027] Optionally, accepting the new solution according to the Metropolis criterion includes:

[0028]

[0029] Wherein, P(Δf) is the new solution accepted according to the Metropolis criterion, T is the temperature, and e is the base of the natural logarithm.

[0030] Optionally, the method further comprises:

[0031] The determination coefficient, root mean square error, mean square error, mean absolute error and mean absolute percentage error were used to verify the model after simultaneous optimization.

[0032] The beneficial effects of the present invention are:

[0033] The advantages and positive effects of the present invention are as follows: the present invention adopts a prediction model based on ensemble learning, and by combining the diversity and stability of base learners such as random forest, gradient boosting tree, AdaBoost, and the nonlinear relationship processing ability of ANN neural network meta-learner, the accuracy and stability of grouting volume prediction can be effectively improved. This ensemble method makes full use of the advantages of different learners. The random forest has good robustness due to the randomness of its construction process and is not easy to overfit; the gradient boosting tree can adapt to a variety of loss functions through iterative optimization and has high accuracy; AdaBoost improves model performance by focusing on misclassified samples. At the same time, the ANN neural network, as a meta-learner, can effectively deal with the uncertainty problem of the prediction results and prevent overfitting due to its good numerical approximation ability. In addition, the present invention also adopts parameter optimization techniques such as simulated annealing to further improve the performance of the model. The simulated annealing algorithm is selected because it can maintain population diversity during the search process and avoid falling into the local optimum.

[0034] The present invention solves the following technical problems of the existing machine learning grouting effect prediction model: hydrological structural geological parameters are not fully considered, the time cost of numerical simulation is too high, the small sample size leads to lack of persuasiveness, and the sample size is sufficient but the accuracy is not high. Compared with the weak learner model, the present invention significantly improves the performance of predicting the grouting effect, can quickly and accurately guide on-site construction decisions, and ensure the economic benefits and engineering stability of grouting; in actual grouting projects, the present invention can provide reliable method support for estimating the grouting volume of the area to be grouted, and has important engineering application value. Moreover, due to its high requirements for the control of rock seepage, the model is not only suitable for specific projects, but is also expected to be extended to other underground grouting projects. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0036] Figure 1 It is a flow chart of a method for predicting the grouting effect of underground water-sealed caverns based on an integrated learning model according to an embodiment of the present invention;

[0037] Figure 2 It is a schematic diagram of the flow chart of the underground water-sealed cavern grouting effect optimization model based on the integrated learning model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0040] This example uses bagging, boosting, and stacking as the three main methods of ensemble learning in machine learning. They improve prediction performance by combining multiple models, and each method has its own unique characteristics and applicable scenarios. By innovatively integrating these three ensemble learning methods, a powerful prediction model is constructed that can not only capture complex patterns and nonlinear relationships in the data, but also improve the accuracy and robustness of predictions.

[0041] In addition, the parameter configuration of each model has a crucial impact on the prediction results. Accurate parameter settings can significantly improve the accuracy of model predictions. Therefore, it is particularly important to choose an efficient optimization algorithm to tune the model's hyperparameters. The simulated annealing algorithm, as a probability-based global optimization method, imitates the random search mechanism in the material annealing process, can effectively avoid local optimal solutions, and then explore the global optimal solution, providing a powerful strategy for model hyperparameter optimization.

[0042] In summary, this embodiment combines machine learning technology, especially ensemble learning and hyperparameter optimization of simulated annealing algorithm, to provide a fast and accurate intelligent model for predicting the effect of rock grouting, and provide a scientific basis for construction decision-making of rock grouting seepage control in underground water-sealed cavern projects, thereby improving the overall construction quality and energy efficiency of grouting projects.

[0043] like Figure 1 As shown, this embodiment discloses a method for predicting the grouting effect of an underground water-sealed cavern based on an integrated learning model, including: obtaining environmental parameters and construction parameters of an underground water-sealed cavern, inputting the environmental parameters and construction parameters into a grouting effect prediction model based on integrated learning, and obtaining a grouting effect prediction result; the grouting effect prediction model is trained using a training set, and the trained model parameters are synchronously optimized using a simulated annealing algorithm; the training set includes: geological parameters, hydrological parameters, and historical construction parameters; the grouting effect prediction model includes: a random forest model, a gradient boosting tree model, an AdaBoost model, and an ANN model; the random forest model, the gradient boosting tree model, and the AdaBoost model are all connected to the ANN model.

[0044] Furthermore, geological parameters include: surrounding rock grade, joint surface density and alteration zone density; hydrological parameters include: pre-grouting permeability; construction parameters include: grouting pressure, hole sequence and elevation parameters.

[0045] Furthermore, obtaining the grouting effect prediction result includes: inputting the environmental parameters and construction parameters into the random forest model, the gradient boosting tree model, and the AdaBoost model to obtain the first prediction result; the random forest model is a random forest model that uses the decision tree as the base learner for the Bagging method integration, and the gradient boosting tree model is a gradient boosting tree model that uses the decision tree as the base learner for the Boosting method integration; the first prediction result is input as input data into the ANN model to obtain the grouting effect prediction result; the ANN model is an ANN model that uses the base learner for the Stacking method integration.

[0046] Specifically, the grouting effect prediction model based on ensemble learning includes a random forest model with a Bagging method integrated with a decision tree as the base learner, a gradient boosting tree model and an AdaBoost model with a Boosting method integrated with a decision tree as the base learner, and an ANN model with a Stacking method integrated with the above base learners. Among them, the data set is randomly divided into a training set and a validation set according to a certain ratio, and the final result is finally predicted by the validation set.

[0047] The method of constructing the data set is to collect historical data of geological parameters, hydrological parameters and actual construction parameters: the geological parameters are the surrounding rock grade that quantifies the stability of the surrounding rock and the complexity of the geological conditions, the density of the joint surface that affects the permeability and stability of the rock mass, and the density of the altered zone that reflects the spatial distribution of the degree of rock alteration; the hydrological parameter is the pre-grouting permeability that measures the permeability of the rock mass before grouting; the actual construction parameters are the hole sequence and grouting pressure of each grouting hole, and the unit ash consumption is used as the indicator for evaluating the grouting effect.

[0048] Furthermore, the simulated annealing algorithm is used to synchronously optimize the parameters of the trained model, including: step 1, taking the model parameters as the initial solution and setting the initial temperature and cooling coefficient; step 2, updating the initial temperature according to the cooling plan, and randomly selecting a new solution in the neighborhood of the initial solution; step 3, calculating the fitness difference of the new solution, judging the fitness difference, if the fitness difference is less than a preset value, accepting the new solution, if the fitness difference is greater than the preset value, accepting the new solution according to the Metropolis criterion; step 4, taking the new solution as the current solution, repeating steps 1 to 3 until the updated temperature is lower than the target threshold or reaches the preset number of iterations to obtain the optimal solution.

[0049] Furthermore, the model parameters include: learning rate and number of weak learners for the AdaBoost model, learning rate, maximum tree depth, number of trees, minimum number of samples per leaf node for the gradient boosting tree model, number of trees, maximum tree depth, minimum number of leaf node samples for the random forest model, and neuron weights, number of hidden layers, and activation function for the ANN model.

[0050] Specifically:

[0051] like Figure 2 As shown, the method for synchronously optimizing model parameters using a simulated annealing algorithm includes the following steps:

[0052] Initialization: Select an initial solution i and set the initial temperature T and cooling coefficient α (where 0<α<1);

[0053] Temperature update formula: The temperature decrease usually follows a predetermined cooling schedule, such as linear or exponential decrease. k+1 =α·T k , where T k is the temperature of the kth iteration, T k+1 is the temperature of the k+1th iteration;

[0054] Neighborhood solution generation: randomly select a new solution i' in the neighborhood of the current solution i;

[0055] Fitness calculation: Calculate the fitness difference between the new solution i' and the current solution i Δf = f(i') - f(i);

[0056] Acceptance criteria: If Δf<0, accept the new solution i'; if Δf>0, accept the new solution according to the Metropolis criterion;

[0057]

[0058] Where P(Δf) is the new solution accepted according to the Metropolis criterion, and T is the temperature. In the simulated annealing algorithm, temperature is a control parameter used to determine the probability of accepting a new solution; e is the base of the natural logarithm, which is approximately equal to 2.71828. In the Metropolis criterion, e is used to calculate the probability formula for accepting a new solution.

[0059] Iterative update: Set i = i', that is, update the current solution with the new solution;

[0060] Termination condition: If T k If the value falls below a certain threshold or reaches the maximum number of iterations, the algorithm is terminated;

[0061] Output result: Return the current solution i as the approximate optimal solution.

[0062] The model parameters for hyperparameter optimization using the simulated annealing optimization algorithm include: the learning rate and number of weak learners for the AdaBoost model, the learning rate, maximum tree depth, number of trees, and minimum number of samples for each leaf node for the gradient boosting tree model, the number of trees, maximum tree depth, and minimum number of leaf node samples for the random forest model, and the neuron weights, number of hidden layers, and activation function for the ANN model.

[0063] Figure 2 Middle: X 1-n is the set of input variables, Y 1 is the set of output quantities, X 1-n Train is the input variable in the training set, Y 1 Train is the output in the training set, X 1-n Test is the input variable in the validation set, Y 1 Test is the output in the validation set, Bagging is a type of ensemble learning, Boosting is a type of ensemble learning, Stacking is a type of ensemble learning, SA is a type of heuristic search algorithm, M 1-n : Several models with different parameters, M best: optimal model parameters, decision tree is a basic regression method, AdaBoost is a boosting algorithm, random forest is a bagging algorithm, gradient boosting tree is a boosting algorithm, ANN is a neural network mathematical model algorithm, P 1 P is the output prediction value obtained by the AdaBoost model based on the training set or validation set. 2 P is the output prediction value obtained by the random forest model based on the training set or validation set. 3 It is the output prediction value obtained by using the gradient boosting tree model based on the training set or validation set.

[0064] Among them, Bagging: a type of ensemble learning, which generates multiple sub-datasets by performing multiple random samplings with replacement on the original data set, and trains a base learner on each sub-dataset, and finally averages or votes the prediction results of these base learners. Boosting: a type of ensemble learning, which trains multiple base learners sequentially, each base learner focuses on the samples misclassified by the previous learner, and finally weights the results of multiple base learners. Stacking: a meta-learning method, which trains multiple base learners and uses a meta-learner to combine the prediction results of these base learners. Simulated Annealing: a heuristic search algorithm used to find the global optimal solution in a large-scale search space. Decision Tree: a basic regression method that predicts the value of the target variable by learning simple decision rules. AdaBoost: a type of Boosting algorithm, which iteratively trains decision trees and adjusts sample weights to focus on mispredicted samples, and finally weights the prediction results of these decision trees into a strong learner. Random Forest: A type of bagging algorithm that improves the prediction accuracy of the model by building multiple decision trees and aggregating their predictions. Gradient Boosting Tree: A type of boosting algorithm that iteratively trains decision trees, with each tree predicting the residual of the previous round of predictions and accumulating the predictions of the new tree into the final prediction of the model, thereby gradually reducing the value of the loss function. ANN: A type of neural network mathematical model algorithm that learns and predicts complex patterns in data by forward propagating information, applying nonlinear activation functions, and adjusting weights through backpropagation.

[0065] Furthermore, the method also includes validating the synchronously optimized model using the coefficient of determination, root mean square error, mean square error, mean absolute error and mean absolute percentage error.

[0066] Specifically

[0067] The performance of the optimized ensemble learning model in predicting grouting volume was analyzed and compared with the unoptimized model to verify the effectiveness of the simulated annealing optimization algorithm. The prediction accuracy of different models, including the coefficient of determination (R 2 ), root mean square error (RMSE), mean square error (MSE), mean absolute error (MAE), mean absolute percentage error (MAE), and select the optimal model.

[0068]

[0069] Among them, R 2 is the coefficient of determination, RMSE is the root mean square error, MSE is the mean square error, MAE is the mean absolute error, and MAE is the mean absolute percentage error.

[0070] This embodiment discloses a method for predicting the grouting effect of underground water-sealed caverns based on an integrated learning model, comprising:

[0071] The grouting effect prediction model based on ensemble learning includes a random forest model with a Bagging method integrated using a decision tree as the base learner, a gradient boosting tree model and an AdaBoost model with a Boosting method integrated using a decision tree as the base learner, and an ANN model with a Stacking method integrated using the above base learners. The first layer includes three ensemble base learners for training and verification; the second layer includes a meta-learner that is an ANN artificial neural network; the data set of the ANN model comes from the prediction outputs of the three ensemble base learners.

[0072] Furthermore, the method for constructing a data set includes: according to the seepage control characteristics of the underground water-sealed cavern, the two construction processes, pre-grouting and post-grouting, are divided into two construction processes to construct data sets respectively, and the corresponding surrounding rock grade, joint surface density, alteration zone density, pre-grouting permeability, grouting pressure, hole sequence, and elevation parameters are obtained by manually sorting and summarizing geological sketches, grouting construction record sheets, etc. The output parameter is unit ash consumption, wherein the surrounding rock grade, joint surface density, alteration zone density, pre-grouting permeability, grouting pressure, hole sequence, and elevation parameters are selected for the pre-grouting construction process; the surrounding rock grade, joint surface density, alteration zone density, pre-grouting permeability, grouting pressure, and hole sequence parameters are selected for the post-grouting construction process, and the two grouting construction data sets are randomly divided into a training set and a validation set according to a determined ratio.

[0073] Furthermore, the performance of the optimized ensemble learning model in predicting grouting volume was analyzed and compared with the unoptimized model to verify the effectiveness of the simulated annealing optimization algorithm. 2), root mean square error (RMSE), mean square error (MSE), mean absolute error (MAE), mean absolute percentage error (MAE), and select the optimal model.

[0074]

[0075] Among them, R 2 is the coefficient of determination, RMSE is the root mean square error, MSE is the mean square error, MAE is the mean absolute error, and MAE is the mean absolute percentage error.

[0076] Furthermore, MSE is adopted as the objective loss function of the simulated annealing algorithm because it can directly measure the average square error between the predicted value and the actual value, is easy to optimize and can effectively reflect the accuracy of the model prediction. It is a commonly used loss function in regression problems.

[0077] Furthermore, the specific implementation steps of the method for synchronously optimizing model parameters using the simulated annealing algorithm include:

[0078] Step 1, select an initial solution i and set the initial temperature T and cooling coefficient α (where 0<α<1);

[0079] Step 2: Develop a cooling plan and set the temperature update formula T k+1 =α·T k , where T k is the temperature of the kth iteration, T k+1 is the temperature of the k+1th iteration;

[0080] Step 3, randomly select a new solution i' in the neighborhood of the current solution i;

[0081] Step 4, calculate the MSE value of the current solution i and the new solution i' and calculate the fitness difference Δf=f(i')-f(i) between the new solution i' and the current solution i;

[0082] Step 5, compare the fitness of the new solution and the current solution. If the new solution is better, accept the new solution as the current solution. If the new solution is worse, there is still a certain probability of accepting the new solution, and this probability decreases as the "temperature" decreases. The acceptance principle is as follows: if Δf<0, accept the new solution i'; if Δf>0, accept the new solution according to the Metropolis criterion;

[0083]

[0084] Step 6, perform iterative update, set i = i', that is, update the current solution with the new solution;

[0085] Step 7, set the termination condition, if T k If the value falls below a certain threshold or reaches the maximum number of iterations, the algorithm is terminated;

[0086] Step 8: Output the result and return the current solution i as the approximate optimal solution.

[0087] Furthermore, the model parameters for hyperparameter optimization using the simulated annealing optimization algorithm include: the learning rate l of the AdaBoost model 1 and the number of weak learners n 1 , the learning rate l of the gradient boosted tree model 2 , the maximum depth of the tree D max1 , the number of trees n 2 , the minimum number of samples for each leaf node J min1 , the number of trees in the random forest model n 3 , the maximum depth of the tree D max2 , minimum number of leaf node samples J min2 And the neuron weight w and the number of hidden layers n of the ANN model hid , activation function f. The optimization process based on the simulated annealing algorithm can be transformed into the following formula:

[0088]

[0089] In the formula, y i and f(x i ; θ(φ), φ) represent the actual unit ash consumption value and the final unit ash consumption prediction value respectively, φ represents the set of parameters to be optimized, and H is the value range of the model parameters to be optimized. Thus, a grouting effect prediction model based on ensemble learning is constructed.

[0090] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for predicting the grouting effect of underground water-sealed caverns based on an integrated learning model, characterized in that: include: Obtaining environmental parameters and construction parameters of the underground water-sealed cavern, inputting the environmental parameters and the construction parameters into a grouting effect prediction model based on ensemble learning, and obtaining a grouting effect prediction result; The grouting effect prediction model is trained using a training set, and the trained model parameters are synchronously optimized using a simulated annealing algorithm; the training set includes: geological parameters, hydrological parameters and historical construction parameters; The grouting effect prediction model includes: a random forest model, a gradient boosting tree model, an AdaBoost model and an ANN model; the random forest model, the gradient boosting tree model and the AdaBoost model are all connected to the ANN model.

2. The method for predicting the grouting effect of underground water-sealed caverns based on an integrated learning model according to claim 1 is characterized in that: The geological parameters include: surrounding rock grade, joint surface density and alteration zone density; The hydrological parameters include: water permeability before irrigation; The construction parameters include: grouting pressure, hole sequence and elevation parameters.

3. The method for predicting the grouting effect of underground water-sealed caverns based on an integrated learning model according to claim 1 is characterized in that: Obtaining the grouting effect prediction result includes: The environmental parameters and the construction parameters are input into a random forest model, a gradient boosting tree model, and an AdaBoost model to obtain a first prediction result; the random forest model is a random forest model that uses a decision tree as a base learner to perform a Bagging method integration, and the gradient boosting tree model is a gradient boosting tree model that uses a decision tree as a base learner to perform a Boosting method integration; The first prediction result is input as input data into an ANN model to obtain the grouting effect prediction result; the ANN model is an ANN model integrated by the Stacking method using the base learner.

4. The method for predicting the grouting effect of underground water-sealed caverns based on an integrated learning model according to claim 1 is characterized in that: The simulated annealing algorithm is used to simultaneously optimize the parameters of the trained model, including: Step 1: Take the model parameters as the initial solution and set the initial temperature and cooling coefficient; Step 2: updating the initial temperature according to the cooling plan and randomly selecting a new solution in the neighborhood of the initial solution; Step 3: Calculate the fitness difference of the new solution, and judge the fitness difference. If the fitness difference is less than a preset value, accept the new solution; if the fitness difference is greater than the preset value, accept the new solution according to the Metropolis criterion. Step 4: Take the new solution as the current solution and repeat steps 1 to 3 until the updated temperature is lower than the target threshold or reaches a preset number of iterations to obtain the optimal solution.

5. The method for predicting the grouting effect of underground water-sealed caverns based on an integrated learning model according to claim 4 is characterized in that: The model parameters include: the learning rate and number of weak learners of the AdaBoost model, the learning rate, maximum tree depth, number of trees, and minimum number of samples for each leaf node of the gradient boosting tree model, the number of trees, maximum tree depth, and minimum number of leaf node samples of the random forest model, and the neuron weights, number of hidden layers, and activation function of the ANN model.

6. The method for predicting the grouting effect of underground water-sealed caverns based on an integrated learning model according to claim 4 is characterized in that: Calculating the fitness difference of the new solution includes: Δf=f(i')-f(i); Among them, Δf is the fitness difference, f(i') is the fitness of the new solution, and f(i) is the fitness difference of the current solution.

7. The method for predicting the grouting effect of underground water-sealed caverns based on an integrated learning model according to claim 4 is characterized in that: Accepting the new solution according to the Metropolis criterion includes: Wherein, P(Δf) is the new solution accepted according to the Metropolis criterion, T is the temperature, and e is the base of the natural logarithm.

8. The method for predicting the grouting effect of underground water-sealed caverns based on an integrated learning model according to claim 1 is characterized in that: The method further includes: The determination coefficient, root mean square error, mean square error, mean absolute error and mean absolute percentage error were used to verify the model after simultaneous optimization.

Images