Deeply buried multi-structural plane underground powerhouse side wall pre-support intelligent design method and system
By combining an improved multilayer perceptron model with a particle swarm optimization algorithm, a multi-structure rock deformation prediction model was constructed. The optimal anchor bolt inclination and tilt angle were automatically solved, enabling earlier pre-support timing and precise directional adjustment. This solved the problems of lag and poor accuracy in the support of the surrounding rock of the underground powerhouse sidewall, ensuring construction safety.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA HYDROELECTRIC ENGINEERING CONSULTING GROUP CHENGDU RESEARCH HYDROELECTRIC INVESTIGATION DESIGN AND INSTITUTE
- Filing Date
- 2026-04-03
- Publication Date
- 2026-06-09
AI Technical Summary
Existing underground powerhouse sidewall rock support technology suffers from problems of delayed support and poor precision. Especially under conditions of multiple structural planes, it is difficult to control the deformation of the surrounding rock in a timely manner and accurately match the occurrence of the structural planes.
An intelligent design method combining an improved multilayer perceptron model and a particle swarm optimization algorithm is adopted. A dataset is constructed through discrete element numerical simulation, and a multi-structure rock deformation prediction model is trained. The optimal anchor bolt inclination and anchor bolt tilt angle are automatically solved, enabling early pre-support and precise adjustment of support direction.
It significantly improves the scientific nature and precision of support design, effectively controls the deformation of surrounding rock in high sidewalls, ensures construction safety, and solves the problems of outdated and mismatched traditional support technologies.
Smart Images
Figure CN121997437B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, specifically to an intelligent design method and system for pre-support of sidewalls in deeply buried multi-structure underground powerhouses. Background Technology
[0002] In geologically complex areas (such as the Jinsha River suture zone and the eastern tectonic junction), the rock mass structure is complex and well-developed, with widespread deep and large faults, resulting in poor rock mass integrity and stability. Under such conditions, the excavation and construction of underground caverns for hydropower stations faces significant challenges. During excavation, the rock mass experiences intense unloading and the structural planes open, making the surrounding rock prone to deformation, collapse, and instability, threatening construction safety.
[0003] Hydropower station underground powerhouses typically feature high, steep sidewalls, exceeding 30 meters in height. Excavation is generally carried out in layers, with each excavation layer typically ranging from 5 to 13 meters in height. Due to the complex excavation process and long construction period, the surrounding rock undergoes continuous unloading and stress adjustment, leading to structural surface cracking and displacement. This results in persistent deformation of the surrounding rock, ultimately causing large deformations that threaten the safety and stability of the cavern. Therefore, for deeply buried underground powerhouses with complex multi-surface structures, controlling the deformation and stability of the surrounding rock during the excavation of high sidewalls is a critical challenge restricting the project's construction.
[0004] The main difficulties of this problem lie in two aspects: First, deformation control is difficult to implement in a timely manner. Since shotcrete and bolt support technology is generally implemented after the excavation of the current layer of the sidewall is completed, the surrounding rock has already been unloaded, and the structural surface undergoes tensile cracking along with the unloading. The damage to the surrounding rock gradually expands from the surface inwards. Applying support measures at this point is too late compared to the unloading deformation. Second, deformation control is difficult to be precise. Because the structural surface has a certain orientation and is at a certain angle to the excavation face, conventional bolt support uses a single-direction application and does not consider the orientation of the structural surface, making it difficult to achieve targeted support effects based on the characteristics of the structural surface.
[0005] Regarding the deformation control of surrounding rock in high sidewalls of deeply buried multi-structure underground powerhouses, there are currently three main related or similar support technologies: First, shotcrete and ordinary anchor bolt support; this is the most commonly used support method in underground powerhouses, using ordinary mortar anchor bolts to support the excavated surrounding rock and then sealing it with shotcrete. Second, anchor cable or prestressed anchor cable support; anchor cable support is also widely used in underground powerhouses, especially in the sidewalls. Sometimes, depending on actual needs, prestressed anchor cables or through anchor cables may be used. Third, yielding anchor bolt support. This is a support method mainly for large compression deformations, based on the principle of yielding support and prioritizing resistance; it has relatively good support effects in soft rock tunnels.
[0006] However, the existing support technologies mentioned above all have the following shortcomings: First, the timing of support is delayed. Existing high sidewall support technologies all begin to be implemented only after the current excavation layer has been completed. However, multi-faceted rock masses have unique characteristics. Under the strong unloading action of high ground stress excavation, the densely distributed structural faces will tensile cracks as the load is unloaded, leading to the propagation of rock mass cracks and damage. This propagation continues from the surface inwards from the excavation face, causing the surrounding rock deformation to develop continuously and without convergence. This is a special mechanical property not found in intact rock masses. If the method of excavating first and then supporting is still used for intact rock masses, the support will be significantly delayed, making it difficult to control the development of deformation in a timely manner. Second, the support accuracy for the orientation of structural faces is poor. The concept of support methods such as anchor bolts and anchor cables is to limit deformation, apply confining pressure, and control opening displacement. Generally, the deformation control effect is best when the support force is consistent with the direction of deformation or displacement. The existing anchor bolts and anchor cables are generally laid perpendicular to the excavation face, but this is not suitable for rock mass conditions with multiple structural faces. When there are many structural faces, the direction perpendicular to the excavation face is not necessarily the optimal direction for deformation control. If this direction is still used, deformation control will appear to be rather blind. Summary of the Invention
[0007] This invention aims to solve the problems of delayed support and poor accuracy in existing support methods for the sidewalls of underground powerhouses, and proposes an intelligent design method and system for pre-support of sidewalls of deeply buried multi-structured underground powerhouses.
[0008] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:
[0009] In a first aspect, the present invention provides an intelligent design method for pre-support of sidewalls in a deeply buried multi-structure underground powerhouse, the method comprising:
[0010] Step 1: Set the value range of structural surface parameters and pre-support direction parameters. The structural surface parameters include structural surface inclination, structural surface dip angle and structural surface spacing. The pre-support direction parameters include anchor bolt inclination and anchor bolt dip angle.
[0011] Step 2: Based on the range of values, construct a three-dimensional numerical model under different working conditions, and use a discrete element numerical simulation program to simulate the excavation and inclined pre-support process of the underground powerhouse sidewall. Obtain the maximum deformation of the sidewall under different structural surface parameters and different pre-support direction parameters, and construct a dataset including structural surface parameters, pre-support direction parameters and maximum deformation of the sidewall.
[0012] Step 3: Construct an improved multilayer perceptron model. Take the structural surface parameters and pre-support direction parameters in the dataset as input parameters, and take the corresponding maximum deformation of the sidewall as the output target. Train the improved multilayer perceptron model to obtain a multi-structural surface surrounding rock deformation prediction model for predicting the maximum deformation of the sidewall.
[0013] Step 4: Build an optimization framework based on particle swarm optimization algorithm, and use the multi-structure surface surrounding rock deformation prediction model as the fitness function of particle swarm optimization algorithm. The output of the fitness function is the predicted value of the maximum deformation of the sidewall.
[0014] Step 5: Obtain the structural surface parameters of the rock mass in the area to be supported. The particle swarm optimization algorithm iteratively searches within the range of values of the pre-support direction parameters based on the structural surface parameters of the rock mass in the area to be supported, and solves for the combination of anchor bolt inclination and anchor bolt tilt angle that minimizes the output of the fitness function, which is taken as the optimal pre-support direction parameters.
[0015] Step 6: Before excavating the sidewall of the underground powerhouse, perform oblique pre-support on the rock mass of the next excavation layer that has not yet been excavated, according to the optimal pre-support direction parameters.
[0016] Furthermore, the value range of the structural surface inclination is 0°~360°, and the value range of the structural surface tilt angle is 0°~90°.
[0017] Furthermore, in step 2, the discrete element numerical simulation program used is the 3DEC discrete element program;
[0018] Before step 3, a preprocessing procedure for the dataset is included, specifically including:
[0019] Outlier data in the dataset is removed and the data is processed to have a unified dimension. The preprocessed dataset includes no less than 200 sets of data. The preprocessed dataset is divided into training set, validation set and test set in a ratio of 7:1:2, and the data is converted into a matrix form containing feature matrix and target vector.
[0020] Furthermore, in step 3, the improved multilayer perceptron model adopts a shallow network structure including an input layer, 2 to 3 hidden layers, and an output layer;
[0021] The improved multilayer perceptron model is trained using the training set, and the mean square error of the training and validation sets is monitored in real time during the training process to ensure that the two sets have the same trend.
[0022] Furthermore, in step 3, after obtaining the multi-structure surface surrounding rock deformation prediction model, the method further includes a step of verifying and optimizing the model using the test set:
[0023] The predictive ability of the model is evaluated by at least one of the following indicators: mean absolute error, coefficient of determination, or root mean square error. When the error exceeds a preset threshold, the network structure or hyperparameters of the multi-structure surface surrounding rock deformation prediction model are optimized.
[0024] Further, in step 4, the fitness function is defined as:
[0025] ;
[0026] in, For structural surface tendency, For the inclination angle of the structural surface, For the spacing between structural surfaces, For anchor bolt inclination, The anchor bolt inclination angle is given, and MLP is a multi-structure surface surrounding rock deformation prediction model; for a set of fixed structural surface parameters The output value of the fitness function varies with the pre-support direction parameter. The predicted maximum deformation of the sidewall changes accordingly;
[0027] The particle swarm optimization algorithm has 20-50 particles and 100-200 iterations.
[0028] Furthermore, in step 5, the automated process of the iterative search includes:
[0029] The structural plane dip, structural plane dip angle, and structural plane spacing of the rock mass in the area to be supported are obtained. Different anchor bolt dips and anchor bolt dip angles are iteratively searched using the particle swarm optimization algorithm. The maximum deformation prediction value of the sidewall is predicted iteratively using the multi-structural-plane surrounding rock deformation prediction model until the anchor bolt dip and anchor bolt dip angle that minimize the maximum deformation prediction value of the sidewall are found.
[0030] Furthermore, after step 5 and before step 6, a step of verifying the optimal pre-support direction parameters is also included:
[0031] Multiple sets of structural surface parameters of the rock mass in the area to be supported are selected. The corresponding structural surface parameters and the optimal pre-support direction parameters obtained by the solution are input into the discrete element numerical simulation program for verification calculation. The maximum deformation value of the sidewall calculated by numerical simulation under the optimal pre-support direction parameters is compared and verified to see if it is less than the maximum deformation value of the sidewall when other pre-support direction parameters are used.
[0032] Furthermore, in step 6, the side wall of the underground plant is a high side wall with a height greater than 30m, where the height refers to the vertical distance from the lower machine hole position to the arch line of the upper top arch;
[0033] The next excavation layer that has not yet been excavated refers to rock mass with a single excavation layer height of 5 to 13 meters;
[0034] The inclined pre-support refers to the operation of applying anchor bolts or anchor cables to the rock mass of the next excavation layer before the current excavation layer is excavated, according to the optimal pre-support direction parameters, during the layered excavation process of the sidewall.
[0035] In a second aspect, the present invention provides an intelligent design system for pre-support of sidewalls of deeply buried multi-structure underground powerhouses, used to execute the intelligent design method for pre-support of sidewalls of deeply buried multi-structure underground powerhouses as described in the first aspect, the system comprising:
[0036] The parameter range setting module is used to set the value range of structural surface parameters and pre-support direction parameters. The structural surface parameters include structural surface inclination, structural surface dip angle and structural surface spacing. The pre-support direction parameters include anchor bolt inclination and anchor bolt dip angle.
[0037] The dataset construction module is used to construct three-dimensional numerical models under different working conditions based on the value range, and to simulate the excavation and inclined pre-support process of the underground powerhouse sidewall using a discrete element numerical simulation program. It obtains the maximum deformation of the sidewall under different structural surface parameters and different pre-support direction parameters, and constructs a dataset including structural surface parameters, pre-support direction parameters and maximum deformation of the sidewall.
[0038] The prediction model training module is used to construct an improved multilayer perceptron model. The structural surface parameters and pre-support direction parameters in the dataset are used as input parameters, and the corresponding maximum deformation of the sidewall is used as the output target. The improved multilayer perceptron model is trained to obtain a multi-structural surface surrounding rock deformation prediction model for predicting the maximum deformation of the sidewall.
[0039] The optimization framework construction module is used to build an optimization framework based on the particle swarm optimization algorithm. The multi-structure surface surrounding rock deformation prediction model is used as the fitness function of the particle swarm optimization algorithm, and the output of the fitness function is the predicted value of the maximum deformation of the sidewall.
[0040] The optimal parameter solution module is used to obtain the structural surface parameters of the rock mass in the area to be supported. The particle swarm optimization algorithm is based on the structural surface parameters of the rock mass in the area to be supported, and iteratively searches within the range of values of the pre-support direction parameters to find the combination of anchor bolt inclination and anchor bolt tilt angle that minimizes the output of the fitness function, which is taken as the optimal pre-support direction parameters.
[0041] The pre-support execution guidance module is used to perform oblique pre-support on the rock mass of the next excavation layer that has not yet been excavated, according to the optimal pre-support direction parameters, before the excavation of the sidewall of the underground powerhouse.
[0042] The beneficial effects of this invention are as follows: The intelligent design method and system for pre-support of sidewalls in deeply buried multi-structured underground powerhouses provided by this invention advances the pre-support timing to before excavation and combines an improved multilayer perceptron and particle swarm optimization algorithm to achieve intelligent optimization of the anchor bolt support direction under multi-structured rock mass conditions. This invention not only solves the problem that traditional support technology is difficult to control the tensile deformation of structural surfaces in a timely manner due to the lag behind excavation unloading, but also overcomes the defects of conventional anchor bolts having a fixed and singular direction and being unable to adapt to the orientation characteristics of structural surfaces. It can automatically generate the optimal anchor bolt inclination and inclination angle that minimizes the maximum deformation of the sidewall for specific geological conditions, thereby significantly improving the scientificity and accuracy of support design, effectively controlling the deformation of the surrounding rock of high sidewalls, and ensuring the safety of underground powerhouse construction. Attached Figure Description
[0043] Figure 1 A flowchart illustrating the intelligent design method for pre-support of sidewalls in a deeply buried multi-structure underground powerhouse, provided as an example.
[0044] Figure 2 This is a structural schematic diagram of the intelligent design system for pre-support of the sidewalls of a deeply buried multi-structure underground powerhouse, provided as an example. Detailed Implementation
[0045] Currently, the main support methods for controlling the deformation of high sidewalls in deeply buried underground powerhouses include ordinary mortar anchors, prestressed anchor cables, or pressure-relief anchors. However, these methods fail to fundamentally address the issues of support lagging behind the excavation and unloading process, and the mismatch between the support direction and the orientation of the structural planes. This results in poor deformation control of high sidewalls under multi-structural-plane conditions, leading to significant risks to the stability of the surrounding rock.
[0046] Based on this, the technical solution of this invention is proposed. In this invention, firstly, a dataset of maximum sidewall deformation is constructed using discrete element numerical simulation, covering combinations of different structural surface parameters (structural surface dip, structural surface inclination, and structural surface spacing) and different pre-support direction parameters (anchor bolt dip, anchor bolt inclination). Then, an improved multilayer perceptron model is trained using this dataset to establish a nonlinear mapping relationship from structural surface parameters and pre-support direction parameters to the maximum sidewall deformation, obtaining a multi-structural surface surrounding rock deformation prediction model. Next, this prediction model is used as the fitness function of a particle swarm optimization algorithm, with the minimum predicted maximum sidewall deformation as the optimization objective, and an iterative search is performed within the range of values for the pre-support direction parameters. Finally, for the actual structural surface parameters of the rock mass in the area to be supported, the optimal anchor bolt dip and anchor bolt inclination angle that minimize the maximum sidewall deformation are automatically solved using this optimization framework, and based on this, oblique pre-support is implemented on the unexcavated next layer of rock mass before excavation.
[0047] The technical solutions in this embodiment will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0048] Figure 1 A flowchart illustrating an intelligent design method for pre-support of sidewalls in a deeply buried multi-structure underground powerhouse is shown. Please refer to [link / reference]. Figure 1 The method includes the following steps:
[0049] Step 1: Set the value range of structural surface parameters and pre-support direction parameters. The structural surface parameters include structural surface inclination, structural surface dip angle and structural surface spacing. The pre-support direction parameters include anchor bolt inclination and anchor bolt dip angle.
[0050] This step aims to determine the boundary conditions for the input parameters in order to construct the subsequent numerical simulation dataset.
[0051] Among them, structural plane parameters are key geological indicators describing the spatial distribution characteristics of discontinuities in rock masses. Specifically, they include: structural plane dip direction (the projection direction of the structural plane's dip direction line onto the horizontal plane); structural plane dip angle (the maximum acute angle between the structural plane and the horizontal plane); and structural plane spacing (the normal distance between the same group of structural planes). These three parameters together constitute the attitude characteristics of multi-structural-plane rock masses and determine the deformation and failure mode of the surrounding rock after excavation and unloading.
[0052] The pre-support direction parameters are the design variables to be optimized, specifically including the anchor bolt inclination and anchor bolt tilt angle, which respectively characterize the projection direction of the anchor bolt on the horizontal plane and the angle with the horizontal plane.
[0053] In practical applications, the range of values for the above parameters must cover various working conditions that may occur in engineering practice: the range of structural plane dip is 0° to 360° in all directions, the range of structural plane dip angle is 0° to 90° covering steep to gentle dips, and the range of structural plane spacing is set with reasonable discrete values (such as 2m, 4m, 6m, 8m, 10m, etc.) based on actual geological survey data; the anchor bolt dip and dip angle are also taken within the complete ranges of 0° to 360° and 0° to 90°, respectively. By uniformly combining the values of each parameter within their respective ranges, it can be ensured that the subsequently constructed dataset can fully reflect the response law of the surrounding rock under different geological conditions and support schemes, providing sufficient training samples for the neural network model.
[0054] Step 2: Based on the value range, construct a three-dimensional numerical model under different working conditions, and use a discrete element numerical simulation program to simulate the excavation and inclined pre-support process of the underground powerhouse sidewall. Obtain the maximum deformation of the sidewall under different structural surface parameters and different pre-support direction parameters, and construct a dataset including structural surface parameters, pre-support direction parameters and maximum deformation of the sidewall.
[0055] This step aims to generate a large-sample dataset covering different geological conditions and support schemes through discrete element numerical simulation, providing support for the training of subsequent neural network models.
[0056] In practical applications, three-dimensional numerical models under different working conditions are constructed based on the parameter value ranges set in step 1. Specifically, the values of structural surface inclination (0°~360°), structural surface inclination angle (0°~90°), and structural surface spacing (e.g., 2m, 4m, 6m, 8m, 10m, etc.) need to be uniformly combined. At the same time, for each combination of structural surface parameters, different pre-support direction parameters are selected and combined within the value ranges of anchor bolt inclination (0°~360°) and anchor bolt inclination angle (0°~90°) to form a calculation scheme covering various possible working conditions.
[0057] In selecting the numerical simulation program, this embodiment adopts the 3DEC discrete element method. This is because the deformation and failure of multi-structure rock masses are mainly controlled by the structural planes, and 3DEC, based on discrete element theory, can effectively simulate discontinuous deformation behaviors such as the opening and slippage of structural planes. Compared with the continuous medium mechanics method, it can better reflect the real mechanical response of structural plane cracking and rock mass damage propagation under high ground stress excavation unloading.
[0058] For each parameter combination condition, a three-dimensional numerical model of the underground powerhouse sidewall was established in 3DEC to simulate the layered excavation process, and oblique pre-support was applied according to the set anchor bolt inclination and angle. Through numerical calculation, the maximum deformation value of the sidewall under this condition was extracted as a control index for evaluating the support effect. Each (total) The structural plane inclination used in the calculation (time) Structural surface inclination angle Structural surface spacing Anchor bolt inclination Anchor bolt inclination angle And the calculated maximum deformation value of the sidewall As a data sample, it eventually forms a collection of The dataset of the group samples is shown in Table 1 (in the table...) (This is a sample number; the values of each parameter correspond to different operating conditions).
[0059] Table 1. Data format obtained from numerical simulation
[0060]
[0061] After the dataset is constructed, data preprocessing is required to ensure data quality and model training effectiveness. The preprocessing process includes the following steps:
[0062] First, remove outlier data from the dataset, such as abrupt changes in calculation results caused by non-convergence of numerical simulation, or samples with parameter combinations that are clearly inconsistent with engineering practice, to ensure the rationality of the input data.
[0063] Secondly, the data should be standardized in terms of units to eliminate the impact of different parameters' different units on model training. For example, although angle parameters (degrees) and spacing parameters (meters) belong to different physical units, in practical applications, normalization or standardization can be omitted, and the original values can be used directly. However, it must be ensured that the numerical range of all parameters is within the acceptable input range of the model. At the same time, it is necessary to ensure that the number of samples in the preprocessed dataset is no less than 200 to guarantee that the training of the neural network has sufficient statistical significance and generalization ability.
[0064] Then, the preprocessed dataset was divided into training, validation, and test sets in a 7:1:2 ratio. 70% of the data was used as the training set for model parameter learning, 10% as the validation set for hyperparameter tuning and overfitting monitoring during training, and 20% as the test set for independent evaluation of the final model performance.
[0065] Finally, the data is converted into a matrix form suitable for input to Python machine learning libraries (such as TensorFlow, PyTorch, etc.), i.e., the feature matrix is constructed. ( The sample size is [number], and the five features are as follows: ) and target vector (corresponding maximum deformation value of the sidewall) This prepares the data for training the subsequent improved multilayer perceptron model.
[0066] Step 3: Construct an improved multilayer perceptron model. Use the structural surface parameters and pre-support direction parameters in the dataset as input parameters, and the corresponding maximum deformation of the sidewall as the output target. Train the improved multilayer perceptron model to obtain a multi-structural surface surrounding rock deformation prediction model for predicting the maximum deformation of the sidewall.
[0067] This step aims to build and train an improved multilayer perceptron (MLP) model. Using the dataset generated in step 2, a nonlinear mapping relationship is established between structural surface parameters and pre-support direction parameters and the maximum deformation of the sidewall, thereby obtaining a multi-structural surface surrounding rock deformation prediction model that can accurately predict the maximum deformation of the sidewall.
[0068] In practical applications, the dimensions of the input and output layers of the improved MLP model are determined based on the data format obtained after preprocessing in step 2. The input layer corresponds to five feature parameters, namely the structural surface tendency. Structural surface inclination angle Structural surface spacing Anchor bolt inclination and anchor bolt inclination angle Therefore, the number of nodes in the input layer is set to 5. The maximum deformation value of the corresponding sidewall in the output layer... With 1 node, a linear activation function is used to output continuous predicted values of the maximum deformation of the sidewall.
[0069] In terms of hidden layer design, this embodiment adopts a shallow network structure with 2 to 3 hidden layers. The number of nodes in each hidden layer can be dynamically adjusted according to the training situation (e.g., 64, 128, etc.), but the principle of "avoiding overfitting" must be followed, that is, the network complexity should not be too high. The activation function of the hidden layers can be a non-linear function such as ReLU to enhance the expressive power of the model. This improved shallow structure design aims to balance the model's fitting ability and generalization performance, and prevent overfitting under limited samples due to the network being too deep or too wide.
[0070] The training process is implemented using a deep learning framework in Python (such as TensorFlow / Keras or PyTorch). The specific steps are as follows:
[0071] Model compilation: Mean squared error (MSE) is chosen as the loss function because this task is a regression problem, and MSE can effectively measure the deviation between the predicted and the true values. An adaptive learning rate algorithm such as Adam can be used as the optimizer to accelerate convergence.
[0072] Model Training: The model is trained using the training set, while validation set data is input into the model. The MSE (Mean Sequence Equation) of both the training and validation sets is calculated at the end of each training cycle. By monitoring the trends of the two MSE curves in real time, it can be determined whether the model is underfitting or overfitting. If the training set MSE continues to decrease while the validation set MSE tends to stabilize or begins to rise, it indicates overfitting, and the model structure needs to be adjusted or training stopped promptly.
[0073] Hyperparameter and structure tuning: During training, hyperparameters such as the number of hidden layers, the number of nodes per layer, the learning rate, and the batch size can be adjusted based on the performance of the validation set until the MSE of both the training and validation sets tends to be stable and low, ensuring that the two trends are consistent. At this point, the model has good fitting and generalization capabilities.
[0074] After training, the final multi-structure rock deformation prediction model needs to be independently evaluated using a test set to verify its prediction accuracy. The evaluation metrics used include:
[0075] Mean Absolute Error (MAE): Measures the average level of the absolute error between the predicted and the actual values, and can intuitively reflect the accuracy of the prediction.
[0076] Root mean square error (RMSE): The square root of the average of the squares of the error. It is more sensitive to larger errors and can be used to assess the stability of the model.
[0077] Coefficient of determination ( ): Reflects the degree to which the model interprets the data, and its value ranges from 1 to 1. The closer it is to 1, the better the fit.
[0078] If any of the above indicators exceeds the preset threshold (e.g., MAE is greater than a certain engineering allowable error), the model needs to be further optimized, such as adjusting the hidden layer structure, modifying the activation function, introducing regularization methods, or increasing training data, and then retraining and validating until the model performance meets the requirements.
[0079] The final multi-structure rock deformation prediction model can predict deformation based on any given structural parameters. and pre-support direction parameters This model quickly predicts the maximum deformation value of the corresponding sidewall. It will serve as the fitness function for subsequent particle swarm optimization algorithms, providing a quantitative target for intelligent optimization of the optimal pre-support direction.
[0080] Step 4: Build an optimization framework based on particle swarm optimization algorithm, and use the multi-structure surface surrounding rock deformation prediction model as the fitness function of particle swarm optimization algorithm. The output of the fitness function is the predicted value of the maximum deformation of the sidewall.
[0081] This step aims to build an optimization framework based on the particle swarm optimization (PSO) algorithm. This framework uses the multi-structure surface surrounding rock deformation prediction model trained in step 3 as the core component for intelligent solution of the optimal pre-support direction parameters.
[0082] In particle swarm optimization (PSO) algorithms, the fitness function is the criterion for evaluating the quality of each particle's position. In this embodiment, the multi-structure surface surrounding rock deformation prediction model (MLP) obtained in step 3 is directly used as the fitness function, and its mathematical expression is:
[0083] ;
[0084] in, For structural surface tendency, For the inclination angle of the structural surface, For the spacing between structural surfaces, For anchor bolt inclination, The anchor bolt inclination angle is denoted by MLP, which is a multi-structure surface surrounding rock deformation prediction model.
[0085] The physical meaning of this formula is: for a fixed set of structural surface parameters The output value of the fitness function varies with the pre-support direction parameter. The predicted maximum deformation of the sidewall changes accordingly. Because... When dealing with a specific area to be supported, the value is a known fixed value, therefore It is actually a parameter only related to the direction of pre-support. The goal of the particle swarm optimization algorithm is to find a binary function within the domain of this binary function (i.e., the reasonable range of values for anchor dip and anchor tilt angle) that satisfies the function value. smallest combination.
[0086] The particle swarm optimization algorithm operates in a two-dimensional optimization space, where the two dimensions correspond to the anchor dip. and anchor bolt inclination angle At the start of the algorithm, a population of multiple particles needs to be initialized. The current position of each particle represents a set of pre-support orientation parameters to be evaluated. The velocity of each particle determines its direction of movement and step size in the search space. Based on a balance between engineering experience and algorithm efficiency, this embodiment sets the number of particles to 20 to 50 to ensure sufficient diversity in the population to cover the search space without causing excessive computation due to too many particles. The number of iterations is set to 100 to 200 to ensure the algorithm has enough iterations to converge to near the global optimum.
[0087] During the iterative search process, each particle updates its velocity and position based on the following two extreme values: one is the best position found by the particle in its own history (individual extreme value), that is, the position that the particle has found that minimizes the predicted maximum deformation of the sidewall. The combination; secondly, the optimal position found throughout the entire population history (global extremum), that is, the position found by all particles that minimizes the predicted maximum deformation of the sidewall. combination.
[0088] With the above settings, the MLP-PSO joint optimization framework is established. The working logic of this framework is: embedding the MLP model trained in step 3 into the fitness function of the PSO, for any given structural surface parameters... The PSO algorithm can automatically adjust the anchor bolt inclination. and anchor bolt inclination angle Iterative search is performed within the range of values, and each time a set is found... The MLP model is called to predict the maximum deformation value of the corresponding sidewall until the iteration ends. Finally, the set of values that minimizes the fitness function value (i.e., the predicted maximum deformation value of the sidewall) is output. This is the optimal pre-support direction parameter.
[0089] Step 5: Obtain the structural surface parameters of the rock mass in the area to be supported. The particle swarm optimization algorithm iteratively searches within the range of values for the pre-support direction parameters based on the structural surface parameters of the rock mass in the area to be supported, and solves for the combination of anchor bolt inclination and anchor bolt tilt angle that minimizes the output of the fitness function, which is taken as the optimal pre-support direction parameters.
[0090] This step is the application stage of this method in actual engineering. Its core task is to automatically solve the optimal pre-support direction parameters for a specific area to be supported using the MLP-PSO joint optimization framework built in step 4.
[0091] In practical applications, the actual structural parameters of the rock mass in the area to be supported are obtained through on-site geological surveys or laboratory tests, specifically including the dip direction of the structural planes. Structural surface inclination angle Structural surface spacing The obtained structural surface parameters The input is then fed into the MLP-PSO joint optimization framework established in step 4. At this point, the structure surface parameters in the fitness function... As a known constant, the function value depends only on the anchor bolt inclination. and anchor bolt inclination angle It changes with the changes.
[0092] Then, the particle swarm optimization algorithm is initiated for automated iterative search. The algorithm operates within the preset range of pre-support direction parameters (anchor bolt inclination). : 0°~360°, anchor bolt inclination angle (0°~90°) Initialize a group of particles, where the current position of each particle represents a set of candidate pre-support direction parameters. In each iteration, the algorithm will... The input is fed into the multi-structure surface surrounding rock deformation prediction model (MLP) trained in step 3, which predicts the maximum deformation value of the corresponding sidewall in real time. This value is the fitness value of the particle. Based on the fitness value, the algorithm continuously updates the individual optimal position of each particle and the global optimal position of the entire population, and adjusts the velocity and position of all particles accordingly, driving the particle swarm to gradually move towards the region where the fitness value is smaller.
[0093] After a preset number of iterations (e.g., 100-200), the particle swarm will converge to the vicinity of the global optimum. At this point, the position corresponding to the globally optimal particle will be determined. This refers to the combination of anchor bolt inclination and anchor bolt angle that minimizes the predicted maximum deformation of the sidewall. This set of parameters is output as the optimal pre-support direction parameters for subsequent construction.
[0094] Step 6: Before excavating the sidewall of the underground powerhouse, perform oblique pre-support on the rock mass of the next excavation layer that has not yet been excavated, according to the optimal pre-support direction parameters.
[0095] This step aims to implement the optimal pre-support direction parameters obtained in step 5, thereby achieving effective control over the deformation of high sidewalls in multi-structure rock masses.
[0096] In this embodiment, the underground powerhouse sidewall refers to a vertical sidewall with a height greater than 30m in a large underground powerhouse of a hydropower station. This height is defined as the vertical distance from the lower machine hole to the arching line of the upper top arch. Due to their large height and complex excavation process, these high sidewalls are prone to unloading tension cracking and large deformation problems under multi-structure rock mass conditions.
[0097] High sidewalls in underground powerhouses are typically excavated in layers, with each excavation layer usually ranging from 5 to 13 meters in height. This layered construction method provides space for the pre-support technology described in this embodiment: that is, during the construction of the current excavation layer, the rock mass of the next layer that has not yet been excavated can be pre-reinforced.
[0098] In practical applications, before the excavation of the current layer begins, the optimal pre-support direction parameters (i.e., the optimal anchor bolt inclination) calculated in step 5 are used. and optimal anchor bolt inclination angle ), and apply anchor bolts or anchor cables to the rock mass of the next excavation layer in advance.
[0099] By applying the optimal pre-support direction parameters obtained from theoretical calculations to actual construction, the next layer of rock mass is supported by inclined anchor bolts or anchor cables before excavation. This effectively reinforces the surrounding rock before the multi-structure rock mass is unloaded during excavation, thereby achieving the goal of timely control of sidewall deformation and ensuring construction safety.
[0100] Since the support direction in this embodiment is the optimal direction obtained through intelligent optimization and matches the orientation of the structural surface of the area to be supported, it is called "oblique pre-support". "Oblique" has two meanings: first, unlike traditional single-direction support perpendicular to the excavation face, the anchor bolt direction in this embodiment is inclined; second, this inclined direction is not arbitrarily set, but is the optimal direction obtained through intelligent algorithm optimization, which can maximally suppress tensile deformation of the structural surface. Through this technical means of "pre-support before excavation and precise adaptation of support direction", this embodiment achieves two core innovations: first, it advances the support timing from the traditional "after excavation" to "before excavation", solving the problem of support lagging behind unloading deformation; second, it changes the support direction from a fixed vertical direction to an oblique direction that can be adaptively adjusted according to the orientation of the structural surface, solving the problem of blind and difficult-to-precise control of traditional support directions.
[0101] Based on the above technical solutions, this embodiment also provides an intelligent design system for pre-support of sidewalls in deeply buried multi-structure underground powerhouses, used to execute the intelligent design method for pre-support of sidewalls in deeply buried multi-structure underground powerhouses as described in the embodiment. Please refer to [link to relevant documentation]. Figure 2 The system includes:
[0102] The parameter range setting module is used to set the value range of structural surface parameters and pre-support direction parameters. The structural surface parameters include structural surface inclination, structural surface dip angle and structural surface spacing. The pre-support direction parameters include anchor bolt inclination and anchor bolt dip angle.
[0103] The dataset construction module is used to construct three-dimensional numerical models under different working conditions based on the value range, and to simulate the excavation and inclined pre-support process of the underground powerhouse sidewall using a discrete element numerical simulation program. It obtains the maximum deformation of the sidewall under different structural surface parameters and different pre-support direction parameters, and constructs a dataset including structural surface parameters, pre-support direction parameters and maximum deformation of the sidewall.
[0104] The prediction model training module is used to construct an improved multilayer perceptron model. The structural surface parameters and pre-support direction parameters in the dataset are used as input parameters, and the corresponding maximum deformation of the sidewall is used as the output target. The improved multilayer perceptron model is trained to obtain a multi-structural surface surrounding rock deformation prediction model for predicting the maximum deformation of the sidewall.
[0105] The optimization framework construction module is used to build an optimization framework based on the particle swarm optimization algorithm. The multi-structure surface surrounding rock deformation prediction model is used as the fitness function of the particle swarm optimization algorithm, and the output of the fitness function is the predicted value of the maximum deformation of the sidewall.
[0106] The optimal parameter solution module is used to obtain the structural surface parameters of the rock mass in the area to be supported. The particle swarm optimization algorithm is based on the structural surface parameters of the rock mass in the area to be supported, and iteratively searches within the range of values of the pre-support direction parameters to find the combination of anchor bolt inclination and anchor bolt tilt angle that minimizes the output of the fitness function, which is taken as the optimal pre-support direction parameters.
[0107] The pre-support execution guidance module is used to perform oblique pre-support on the rock mass of the next excavation layer that has not yet been excavated, according to the optimal pre-support direction parameters, before the excavation of the sidewall of the underground powerhouse.
[0108] It is understood that the intelligent design system for pre-support of sidewalls of deep-buried multi-structure underground powerhouses described in this embodiment is a system for implementing the intelligent design method for pre-support of sidewalls of deep-buried multi-structure underground powerhouses described in the embodiment. As the system disclosed in the embodiment corresponds to the method disclosed in the embodiment, the description is relatively simple. For relevant parts, please refer to the description of the method. It will not be repeated here.
Claims
1. A smart design method for pre-support of sidewalls in deeply buried multi-structure underground powerhouses, characterized in that, The method includes: Step 1: Set the value range of structural surface parameters and pre-support direction parameters. The structural surface parameters include structural surface inclination, structural surface dip angle and structural surface spacing. The pre-support direction parameters include anchor bolt inclination and anchor bolt dip angle. Step 2: Based on the range of values, construct a three-dimensional numerical model under different working conditions, and use a discrete element numerical simulation program to simulate the excavation and inclined pre-support process of the underground powerhouse sidewall. Obtain the maximum deformation of the sidewall under different structural surface parameters and different pre-support direction parameters, and construct a dataset including structural surface parameters, pre-support direction parameters and maximum deformation of the sidewall. Step 3: Construct an improved multilayer perceptron model. Take the structural surface parameters and pre-support direction parameters in the dataset as input parameters, and take the corresponding maximum deformation of the sidewall as the output target. Train the improved multilayer perceptron model to obtain a multi-structural surface surrounding rock deformation prediction model for predicting the maximum deformation of the sidewall. Step 4: Build an optimization framework based on particle swarm optimization algorithm, and use the multi-structure surface surrounding rock deformation prediction model as the fitness function of particle swarm optimization algorithm. The output of the fitness function is the predicted value of the maximum deformation of the sidewall. Step 5: Obtain the structural surface parameters of the rock mass in the area to be supported. The particle swarm optimization algorithm iteratively searches within the range of values of the pre-support direction parameters based on the structural surface parameters of the rock mass in the area to be supported, and solves for the combination of anchor bolt inclination and anchor bolt tilt angle that minimizes the output of the fitness function, which is taken as the optimal pre-support direction parameters. Step 6: Before excavating the sidewall of the underground powerhouse, perform oblique pre-support on the rock mass of the next excavation layer that has not yet been excavated, according to the optimal pre-support direction parameters.
2. The intelligent design method for pre-support of sidewalls of deeply buried multi-structure underground powerhouses according to claim 1, characterized in that, The value range of the structural surface inclination is 0° to 360°, and the value range of the structural surface tilt angle is 0° to 90°.
3. The intelligent design method for pre-support of sidewalls of deeply buried multi-structure underground powerhouses according to claim 1, characterized in that, In step 2, the discrete element numerical simulation program used is the 3DEC discrete element program; Before step 3, a preprocessing procedure for the dataset is included, specifically including: Outlier data in the dataset is removed and the data is processed to have a unified dimension. The preprocessed dataset includes no less than 200 sets of data. The preprocessed dataset is divided into training set, validation set and test set in a ratio of 7:1:2, and the data is converted into a matrix form containing feature matrix and target vector.
4. The intelligent design method for pre-support of sidewalls of deeply buried multi-structure underground powerhouses according to claim 3, characterized in that, In step 3, the improved multilayer perceptron model adopts a shallow network structure including an input layer, 2 to 3 hidden layers, and an output layer. The improved multilayer perceptron model is trained using the training set, and the mean square error of the training and validation sets is monitored in real time during the training process to ensure that the two sets have the same trend.
5. The intelligent design method for pre-support of sidewalls of deeply buried multi-structure underground powerhouses according to claim 4, characterized in that, In step 3, after obtaining the multi-structure surface surrounding rock deformation prediction model, the method further includes verifying and optimizing the model using the test set. The predictive ability of the model is evaluated by at least one of the following indicators: mean absolute error, coefficient of determination, or root mean square error. When the error exceeds a preset threshold, the network structure or hyperparameters of the multi-structure surface surrounding rock deformation prediction model are optimized.
6. The intelligent design method for pre-support of sidewalls of deeply buried multi-structure underground powerhouses according to claim 1, characterized in that, In step 4, the fitness function is defined as: ； in, For structural surface tendency, For the inclination angle of the structural surface, For the spacing between structural surfaces, For anchor bolt inclination, The anchor bolt inclination angle is given, and MLP is a multi-structure surface surrounding rock deformation prediction model; for a set of fixed structural surface parameters The output value of the fitness function varies with the pre-support direction parameter. The predicted maximum deformation of the sidewall changes accordingly; The particle swarm optimization algorithm has 20-50 particles and 100-200 iterations.
7. The intelligent design method for pre-support of sidewalls of deeply buried multi-structure underground powerhouses according to claim 6, characterized in that, In step 5, the automated process of the iterative search includes: The structural plane dip, structural plane dip angle, and structural plane spacing of the rock mass in the area to be supported are obtained. Different anchor bolt dips and anchor bolt dip angles are iteratively searched using the particle swarm optimization algorithm. The maximum deformation prediction value of the sidewall is predicted iteratively using the multi-structural-plane surrounding rock deformation prediction model until the anchor bolt dip and anchor bolt dip angle that minimize the maximum deformation prediction value of the sidewall are found.
8. The intelligent design method for pre-support of sidewalls of deeply buried multi-structure underground powerhouses according to claim 1, characterized in that, After step 5 and before step 6, the method further includes a step of verifying the results of the optimal pre-support direction parameters: Multiple sets of structural surface parameters of the rock mass in the area to be supported are selected. The corresponding structural surface parameters and the optimal pre-support direction parameters obtained by the solution are input into the discrete element numerical simulation program for verification calculation. The maximum deformation value of the sidewall calculated by numerical simulation under the optimal pre-support direction parameters is compared and verified to see if it is less than the maximum deformation value of the sidewall when other pre-support direction parameters are used.
9. The intelligent design method for pre-support of sidewalls of deeply buried multi-structure underground powerhouses according to claim 1, characterized in that, In step 6, the side wall of the underground plant is a high side wall with a height greater than 30m. The height refers to the vertical distance from the lower machine nest position to the arch line of the upper top arch. The next excavation layer that has not yet been excavated refers to rock mass with a single excavation layer height of 5 to 13 meters; The inclined pre-support refers to the operation of applying anchor bolts or anchor cables to the rock mass of the next excavation layer before the current excavation layer is excavated, according to the optimal pre-support direction parameters, during the layered excavation process of the sidewall.
10. An intelligent design system for pre-support of sidewalls in deeply buried multi-structure underground powerhouses, characterized in that: The system is used to execute the intelligent design method for pre-support of sidewalls of deeply buried multi-structure underground powerhouses as described in any one of claims 1 to 9, the system comprising: The parameter range setting module is used to set the value range of structural surface parameters and pre-support direction parameters. The structural surface parameters include structural surface inclination, structural surface dip angle and structural surface spacing. The pre-support direction parameters include anchor bolt inclination and anchor bolt dip angle. The dataset construction module is used to construct three-dimensional numerical models under different working conditions based on the value range, and to simulate the excavation and inclined pre-support process of the underground powerhouse sidewall using a discrete element numerical simulation program. It obtains the maximum deformation of the sidewall under different structural surface parameters and different pre-support direction parameters, and constructs a dataset including structural surface parameters, pre-support direction parameters and maximum deformation of the sidewall. The prediction model training module is used to construct an improved multilayer perceptron model. The structural surface parameters and pre-support direction parameters in the dataset are used as input parameters, and the corresponding maximum deformation of the sidewall is used as the output target. The improved multilayer perceptron model is trained to obtain a multi-structural surface surrounding rock deformation prediction model for predicting the maximum deformation of the sidewall. The optimization framework construction module is used to build an optimization framework based on the particle swarm optimization algorithm. The multi-structure surface surrounding rock deformation prediction model is used as the fitness function of the particle swarm optimization algorithm, and the output of the fitness function is the predicted value of the maximum deformation of the sidewall. The optimal parameter solution module is used to obtain the structural surface parameters of the rock mass in the area to be supported. The particle swarm optimization algorithm is based on the structural surface parameters of the rock mass in the area to be supported, and iteratively searches within the range of values of the pre-support direction parameters to find the combination of anchor bolt inclination and anchor bolt tilt angle that minimizes the output of the fitness function, which is taken as the optimal pre-support direction parameters. The pre-support execution guidance module is used to perform oblique pre-support on the rock mass of the next excavation layer that has not yet been excavated, according to the optimal pre-support direction parameters, before the excavation of the sidewall of the underground powerhouse.