A method for determining development depth of high-energy rock mass in valley based on analog inversion

Abstract

This invention relates to the field of valley rock mass development detection technology, and discloses a method for determining the development depth of high-energy-storage rock masses in valleys based on simulation inversion. The method includes: acquiring geological information and measured geostress data; constructing a three-dimensional geological model and unifying the coordinate system to complete stress transformation; using multiple linear regression to screen boundary loading parameters and ranges; generating multiple sets of boundary loading combinations through orthogonal design and performing valley downcutting numerical simulations to obtain corresponding geostress data; constructing a training sample set and training a neural network model; inputting the measured data into the model to invert the target boundary loading parameters, and substituting them back into the model for comparison and verification; finally, determining the development depth of high-energy-storage rock masses in valleys based on the verified geostress results. This application improves the inversion accuracy for steep valley areas.

Description

Technical Field

[0001] This invention relates to the field of valley rock mass development detection technology, and more specifically, to a method for determining the development depth of high-energy-storage rock masses in valleys based on simulation inversion. Background Technology

[0002] The development of the geostress field determines both the site selection of an engineering project and the design and construction of structures. For modern deep engineering projects, the development of high geostress imposes a large amount of unreleased strain energy on the surrounding rock. Such high-energy-storage rock masses are highly susceptible to rockbursts and other disasters under the disturbance of engineering excavation. Therefore, at the initial stage of engineering design, inverting and analyzing the overall geostress field of the region using measured stress data, and determining the strain energy development value of the regional rock mass based on this, can optimize the engineering layout, avoid disaster-prone areas, and optimize the construction plan.

[0003] Currently, in engineering practice, the initial geostress field inversion is mostly based on the principle of linear superposition of tectonic stress components, combined with multiple linear regression. Traditional multiple linear regression inversion of geostress treats the initial geostress as the linear superposition of stresses generated by various factors within the geological body. Using statistical methods, based on the principle of minimizing the sum of squared residuals, it selects the undetermined model boundary loading modes and obtains a unique solution for each undetermined coefficient, thus characterizing the geostress field. This method is computationally efficient and has a simple principle, but the causes of geostress are complex, the selection of the tectonic stress components to be inverted is uncertain, and for steep river valleys, the geostress field is affected by topographic relief and slope unloading, exhibiting significant nonlinear characteristics. Simple linear superposition is insufficient to achieve ideal inversion accuracy.

[0004] Therefore, it is necessary to design a method for determining the development depth of high-energy-storage rock masses in river valleys based on simulation inversion in order to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a method for determining the development depth of high-energy-storage rock masses in river valleys based on simulation inversion, aiming to solve the problem of low inversion accuracy in steep river valley areas by existing technologies.

[0006] This invention proposes a method for determining the development depth of high-energy-storage rock masses in river valleys based on simulation inversion, including: Acquire geological information data and measured geostress data; A three-dimensional geological model is constructed based on the geological information data, and stress coordinate transformation is performed on the measured geostress data according to the coordinate system of the three-dimensional geological model. Filter boundary loading parameters and boundary loading range data; select gradient values ​​within the boundary loading range data and generate several boundary loading combinations; input the several boundary loading combinations into the three-dimensional geological model to obtain the geostress simulation data corresponding to each boundary loading combination; construct a training sample set based on the boundary loading combinations and geostress simulation data; and train a neural network model based on the training sample set. The measured geostress data is input into the trained neural network model to invert and obtain the target boundary loading parameters; the target boundary loading parameters are input into the three-dimensional geological model to determine the geostress data to be compared; and the development depth of the high-energy-storage rock mass in the valley is determined based on the geostress data to be compared.

[0007] Furthermore, when performing stress coordinate transformation on the measured geostress data, the following steps are included: Extract the principal stress values, principal stress azimuths, and principal stress dip angles from the measured geostress data; determine the direction cosines between the principal stress vectors and each axis of the standard initial coordinate system based on the principal stress azimuths and the principal stress dip angles; and determine the stress component values ​​in the standard initial coordinate system based on the direction cosines and the principal stress values.

[0008] Furthermore, when performing stress coordinate transformation on the measured geostress data, the process also includes: When the coordinate system of the three-dimensional geological model is inconsistent with the standard initial coordinate system, the direction cosines between the coordinate axes of the three-dimensional geological model and each axis of the standard initial coordinate system are obtained, and the stress component values ​​are subjected to a second coordinate transformation based on the stress transformation matrix to obtain stress data consistent with the three-dimensional geological model.

[0009] Furthermore, when filtering boundary loading parameters and boundary loading range data, the following are included: Construct a multiple linear relationship between the initial stress field and influencing factors; calculate the test statistics of the regression coefficients corresponding to each boundary loading mode; Set a critical value; compare the test statistics with the critical value; introduce the boundary loading patterns corresponding to the test statistics greater than the critical value into the regression model, and eliminate the boundary loading patterns less than the critical value to obtain the boundary loading parameters; determine the boundary loading range data based on the results of the multiple linear regression analysis.

[0010] Furthermore, when selecting a gradient value within the boundary loading range data and generating several boundary loading combinations, the process includes: Several gradient levels are set within the boundary loading range data; several boundary loading combinations are generated by orthogonally designing and combining the gradient levels.

[0011] Furthermore, when constructing the training sample set, the following steps are included: Several of the aforementioned boundary loading combinations are paired with the geostress simulation data to form sample data pairs; the sample data pairs are randomly sorted; and the randomly sorted sample data pairs are divided into training sets and test sets. The training set is used for weight updates of the neural network model, and the test set is used for validation of the neural network model.

[0012] Furthermore, training the neural network model based on the training sample set includes: The network structure parameters of the neural network model are set, and the root mean square error is used as the loss function. The neural network model is trained based on the training sample set, and the neural network model is updated through an optimization algorithm until the loss function is less than the loss threshold, thus obtaining the trained neural network model.

[0013] Furthermore, when updating the neural network model through the optimization algorithm, the following steps are included: Convergence optimization is performed based on the Levenberg-Marquardt algorithm; a Jacobian matrix of the first derivatives of the error with respect to all parameters is defined, and a damping factor is introduced to adjust the parameter update step size; when the error decreases, the damping factor is decreased; when the error does not decrease, the damping factor is increased; the weight coefficients are repeatedly adjusted until the loss function is less than the loss threshold. The error is the difference between the output value of the neural network model and the training sample set.

[0014] Furthermore, when determining the development depth of high-energy-storage rock masses in the valley based on the aforementioned geostress data to be compared, the following steps are included: The stress data to be compared is compared with the measured stress data. Based on the comparison results, it is determined whether the inversion results meet the preset error requirements. If the preset error requirements are not met, the target boundary loading parameters and the corresponding stress simulation data are added as new samples to the training sample set, and the neural network model is retrained. If the preset error requirements are met, the development depth of the high-energy-storage rock mass in the valley is determined.

[0015] Furthermore, when determining the development depth of high-energy-storage rock masses in the valley based on the aforementioned geostress data to be compared, the following steps are included: Based on the stress data to be compared, the elastic strain energy density of each unit in the three-dimensional geological model is determined; the elastic strain energy density of each unit is compared with the preset energy storage criterion threshold. Units with elastic strain energy density greater than the preset energy storage criterion threshold are identified as high-energy-storage rock mass development areas; the development depth of the high-energy-storage rock mass is determined based on the spatial coordinates of the high-energy-storage rock mass development area in the three-dimensional geological model.

[0016] Compared with existing technologies, the advantages of this invention are as follows: By utilizing the nonlinear characteristics of the backpropagation neural network, the influence of surface alteration on the ground stress in the training samples can be incorporated into the construction process of the relationship between the model boundary loading conditions and the measured stress. Furthermore, by expanding the training set through iterative calculations, the inversion accuracy is ensured. Combined with energy calculation principles, the elastic strain energy density of present-day river valleys can be calculated, and the development depth of high-energy-storage rock masses can be obtained based on the energy criteria for high-energy-storage rock masses. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of a method for determining the development depth of high-energy-storage rock masses in river valleys based on simulation inversion provided in an embodiment of the present invention; Figure 2 A three-dimensional model of a river valley provided in this embodiment of the invention for determining the development depth of high-energy-storage rock masses in river valleys based on simulation inversion; Figure 3 The maximum principal stress distribution of the valley at different incision stages is shown in the simulation inversion-based method for determining the development depth of high-energy-storage rock masses in valleys, as provided in this embodiment of the invention. Figure 4 The diagram shows the distribution of elastic strain energy density in a valley, based on a simulation-based inversion method for determining the development depth of high-energy-storage rock masses in a valley, as provided in this embodiment of the invention. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] In some embodiments of this application, see Figure 1 As shown, a method for determining the development depth of high-energy-storage rock masses in river valleys based on simulation inversion is proposed, including: S100: Acquire geological information data and measured geostress data; S200: Construct a three-dimensional geological model based on geological information data, and perform stress coordinate transformation on the measured geostress data according to the coordinate system of the three-dimensional geological model; S300: Filter boundary loading parameters and boundary loading range data; select gradient values ​​within the boundary loading range data and generate several boundary loading combinations; input several boundary loading combinations into the three-dimensional geological model to obtain the geostress simulation data corresponding to each boundary loading combination; construct a training sample set based on the boundary loading combinations and geostress simulation data; and train the neural network model based on the training sample set. S400: Input the measured geostress data into the trained neural network model to invert and obtain the target boundary loading parameters; input the target boundary loading parameters into the three-dimensional geological model to determine the geostress data to be compared; determine the development depth of the high-energy-storage rock mass in the valley based on the geostress data to be compared.

[0020] Specifically, in step S100, geological information data and measured geostress data of the study area are collected through on-site geological surveys, borehole tests and existing engineering data. The geological information data includes at least stratigraphic distribution, lithological characteristics, fault structures, valley landforms and terrace development, etc., to characterize the multi-stage downcutting evolution of the valley. The measured geostress data includes parameters such as principal stress values, principal stress azimuth angles and dip angles at each measuring point.

[0021] In step S200, a three-dimensional geological model is established based on geological information data. The model is spatially discretized and each element is assigned physical and mechanical parameters such as elastic modulus, Poisson's ratio, and unit weight. At the same time, the measured principal stress data is converted into stress components in the standard coordinate system through the direction cosine relationship. When the coordinate system of the three-dimensional model is inconsistent with the standard coordinate system, a stress transformation matrix is ​​constructed to perform a secondary coordinate transformation on the stress components, thereby obtaining geostress data consistent with the three-dimensional geological model, providing a unified data foundation for subsequent inversion.

[0022] In step S300, based on the principle of linear superposition, a significance analysis of different types of boundary loading modes is performed using a multiple linear stepwise regression method to screen out boundary loading parameters that have a significant impact on the geostress field and determine the value range of each parameter. On this basis, an orthogonal design method is used to set several gradient levels within the value range to construct multiple sets of boundary loading combination schemes. Each combination is applied to the boundary of the three-dimensional geological model, and the multi-stage downcutting process of the valley is simulated using numerical simulation methods to obtain geostress response data at each measuring point under different loading combinations. The boundary loading combination parameters are used as outputs, and the corresponding measuring point stress components are used as inputs to construct paired training sample data. After randomization, the samples are divided into training and test sets according to a preset ratio for training and performance verification of the neural network model.

[0023] In step S400, a backpropagation neural network model is constructed. The number of input layer nodes is set to the number of stress components at all measurement points, and the number of output layer nodes is set to the number of boundary loading parameters obtained through screening. One or more hidden layers are set according to the complexity of the actual problem. The mean square error is used as the loss function, and the network weights and biases are iteratively updated using the Levenberg-Marquardt optimization algorithm to make the model gradually approximate the nonlinear mapping relationship between the stress at the measurement points and the boundary loading parameters. After the model training is completed, the actual measured geostress data is input into the neural network model to invert the target boundary loading parameters. These parameters are then input into a three-dimensional geological model to calculate the geostress field and obtain the geostress data to be compared. The calculation results are compared with the measured data point by point, and the error index is calculated. When the error is greater than a preset threshold, the boundary loading parameters obtained in this inversion and the corresponding simulated stress data are added to the training sample set as new samples to retrain the neural network model and improve the model accuracy. The above iterative process is repeated until the error meets the preset requirements. After meeting the accuracy requirements, based on the geostress field results obtained from the final inversion, the stress state of each unit in the three-dimensional model is extracted, and the elastic strain energy density of each unit is calculated according to the elastic strain energy calculation method. By comparing it with the preset energy storage criterion threshold, the development area of ​​high energy storage rock mass is identified, and then the development depth of high energy storage rock mass in the valley is determined by combining its spatial location in the three-dimensional model.

[0024] Understandably, by introducing a combination of multiple linear stepwise regression and neural network inversion, key boundary loading factors can be screened and a nonlinear mapping relationship between the stress field and loading parameters can be established, thereby improving the accuracy and stability of geostress inversion. By combining orthogonal design with numerical simulation, the number of parameter combinations is reduced, and computational efficiency is improved.

[0025] In some embodiments of this application, stress coordinate transformation is performed on measured geostress data, including: Extract the principal stress values, principal stress azimuths, and principal stress dip angles from the measured geostress data; determine the direction cosines between the principal stress vectors and each axis of the standard initial coordinate system based on the principal stress azimuths and principal stress dip angles; and determine the stress component values ​​in the standard initial coordinate system based on the direction cosines and principal stress values.

[0026] The formula for calculating the direction cosines between the principal stress vector and each axis of the standard initial coordinate system is as follows: ; In the formula, The azimuth of the measured principal stress is 0° for true north, positive for clockwise and negative for counterclockwise. The angle of inclination is the measured value of the principal stress, with elevation being positive and depression being negative; , , These are the direction cosines of the principal stress axes and the x, y, and z axes, respectively.

[0027] Based on the coordinate transformation formula, the values ​​of six stress components at the measuring point in the standard initial coordinate system can be obtained; the formula for calculating the stress component values ​​in the standard initial coordinate system is as follows: ; In the formula, L 1. M 1. N 1 represents the direction cosine, which respectively represents the cosine value of the angle between the x′ axis in the new coordinate system and the x, y, and z axes of the original coordinate system; L 2. M 2. N 2. Here, represents the direction cosine, which is the cosine of the angle between the y′ axis in the new coordinate system and the x, y, and z axes of the original coordinate system, respectively. L 3. M 3. N 3 represents the direction cosine, which respectively represents the cosine values ​​of the angles between the z′ axis in the new coordinate system and the x, y, and z axes of the original coordinate system; σ 1. σ 2. σ 3 represents the three stress components of the principal stress vector.

[0028] Specifically, the principal stress values, principal stress azimuths, and principal stress dip angles for each measuring point are first extracted from the measured geostress data. The principal stress azimuth is defined as 0° north, with clockwise directions being positive and counterclockwise directions negative. The principal stress dip angles are based on the horizontal plane, with elevation angles being positive and depression angles being negative. Based on the azimuth and dip angles, and according to the definition of the direction cosine of spatial vectors, the cosine values ​​of the angles between the principal stress directions and the x-axis (eastward), y-axis (northward), and z-axis (vertical to the ground surface upward) of the standard initial coordinate system are calculated, thus obtaining the direction cosine parameters of the principal stress vector in the standard coordinate system. After obtaining the direction cosines, the principal stress values ​​and direction cosines are combined to construct a stress tensor expression, thereby converting the measured data originally expressed in principal stress form into six independent stress components (including three normal stress components and three shear stress components) in the standard initial coordinate system, to meet the requirements of numerical simulation calculations for a complete stress state input. Furthermore, in practical applications, when the coordinate system used in the 3D geological model differs from the aforementioned standard initial coordinate system, a secondary transformation between the coordinate systems is required. Specifically, this involves obtaining the spatial orientation relationship between the coordinate axes (x′, y′, z′) of the 3D geological model and the standard coordinate system (x, y, z), and calculating the direction cosines between each coordinate axis. L 1. M 1. N1 represents the direction cosine between the x′ axis and the axes of the original coordinate system. L 2. M 2. N 2. Represents the direction cosine of the y′ axis. L 3. M 3. N 3 represents the direction cosine of the z′ axis, and a coordinate transformation matrix is ​​constructed based on this. Subsequently, the stress components, which have been represented in the standard initial coordinate system, are subjected to tensor transformation through the direction cosine matrix to obtain the corresponding stress component values ​​in the three-dimensional geological model coordinate system, thus realizing a unified transformation of stress data from the measured coordinate system to the numerical model coordinate system.

[0029] Understandably, by adopting a stress coordinate transformation method based on direction cosines to uniformly convert the measured principal stress data into stress component forms in the model coordinate system, not only is the applicability and accuracy of geostress data in numerical simulation improved, but the problem of error accumulation caused by different coordinate systems is also avoided. By introducing a coordinate quadratic transformation mechanism, the method's adaptability to different modeling coordinate systems is enhanced.

[0030] In some embodiments of this application, when performing stress coordinate transformation on measured geostress data, the method further includes: When the coordinate system of the 3D geological model is inconsistent with the standard initial coordinate system, the direction cosines between the coordinate axes of the 3D geological model and each axis of the standard initial coordinate system are obtained. Based on the stress transformation matrix, a second coordinate transformation is performed on the stress component values ​​to obtain stress data consistent with the 3D geological model.

[0031] If the coordinates of the constructed 3D model differ from the standard initial coordinate system, a secondary coordinate transformation is required for the obtained stress components. Assuming the coordinate axes of the new coordinate system are x′, y′, and z′, based on the stress transformation matrix, the expression between the initial coordinate system (x, y, z) and the current coordinate system (x′, y′, z′) is as follows: ; In the formula, , , The directions are the cosines of the x′, y′, and z′ axes, respectively.

[0032] Among them, the stress in the initial coordinate system (x, y, z) is known. The stress calculation formula in the current coordinate system (x′, y′, z′) is: ; in, ; is the stress transformation matrix, used to transform stress components or coordinates from one coordinate system to another.

[0033] Specifically, when the coordinate system used in the 3D geological model is (x′, y′, z′) and its spatial orientation differs from the standard initial coordinate system, the directional relationships of the model coordinate axes x′, y′, and z′ relative to the standard coordinate axes x, y, and z are first determined through geometric relationships. Then, the direction cosine parameters between each coordinate axis are calculated, where the direction cosines of the x′ axis relative to the x, y, and z axes are denoted as follows: l 1 、m 1 、n 1 The direction cosine corresponding to the y′ axis is l 2 、m 2 、n 2 The direction cosine corresponding to the z′ axis is l 3 、m 3 、n 3 Based on the aforementioned direction cosine parameters, a coordinate transformation matrix is ​​constructed. T σ, this matrix describes the rotational relationship between the two coordinate systems. Based on this, the stress tensor σ, already represented in the standard initial coordinate system, is transformed using the stress transformation matrix to obtain the stress tensor in the model coordinate system (x′, y′, z′). σ′ Its transformation process satisfies the tensor transformation law, that is, through matrix transformation. T The stress tensor is transformed bilaterally by σ and its transpose matrix to obtain six stress component values ​​in the new coordinate system. This transformation ensures a strict correspondence between stress components in different coordinate systems, maintaining both the physical meaning of each stress component and its spatial orientation in the 3D geological model. In practice, the transformation process can be embedded in the data preprocessing module, processing the stress components at each measurement point sequentially and mapping the transformed stress data to the corresponding grid nodes or cell locations in the 3D geological model, thus forming a geostress data field completely consistent with the model.

[0034] Understandably, by introducing a stress tensor quadratic coordinate transformation method based on direction cosine, the measured geostress data can be accurately converted between different coordinate systems, thereby ensuring the consistency between the stress direction and the spatial direction of the three-dimensional geological model and improving the accuracy and stability of the numerical simulation results.

[0035] In some embodiments of this application, filtering boundary loading parameters and boundary loading range data includes: Construct a multiple linear relationship between the initial stress field and influencing factors; calculate the test statistics of the regression coefficients corresponding to each boundary loading mode; Set a critical value; compare the test statistics with the critical value; introduce the boundary loading patterns corresponding to the test statistics greater than the critical value into the regression model, and eliminate the boundary loading patterns less than the critical value to obtain the boundary loading parameters; determine the boundary loading range data based on the results of the multiple linear regression analysis.

[0036] In some embodiments of this application, when selecting a gradient value within the boundary loading range data and generating several boundary loading combinations, the process includes: Several gradient levels are set within the boundary loading range data; several boundary loading combinations are generated by orthogonally designing and combining the gradient levels.

[0037] The initial stress field is expressed as: ; In the formula: The initial stress values ​​represent three stress components in a two-dimensional problem and six stress components in a three-dimensional problem; x, y, and z are the spatial coordinates of the terrain and geological bodies; E, , The elastic modulus, Poisson's ratio, and unit weight of the rock mass are not necessarily consistent at different locations, but their values ​​are generally assumed to be independent of stress magnitude and loading process; G is the self-weight factor, U, V, and W are geological structural factors, and T is the temperature factor.

[0038] Typically, it is assumed that the model's ground stress consists of horizontal compression in the x-direction, horizontal compression in the y-direction, self-weight in the z-direction, horizontal shear in the xy-plane, vertical shear in the xz-plane, and vertical shear in the yz-plane. In numerical simulation software, different types of unit quantity boundary conditions are applied to the existing valley model (which has undergone multi-stage downcutting). Based on elastic constitutive relations, the stress values ​​at the measured stress points can be obtained. Using a multiple linear regression method based on the principle of linear superposition, the simulated stress values ​​at the measured stress points are used as the independent variables in the regression, and the measured ground stress values ​​are used to verify the regression accuracy. This constructs a multiple linear relationship between the initial stress field and its influencing factors, thus expressing the regional ground stress field as: ; In the formula: k refers to the measurement point number; The dependent variable is the value of the stress component in the j direction at the k measurement point in the numerical simulation results (j=1, 2, ..., 6). , where is the independent variable, refers to the calculated result of the j-th stress component at the k-th measuring point under the i-th type of structural boundary loading condition; represents the multiple regression coefficients of each structural boundary; n represents various stress boundary condition loading schemes.

[0039] Assuming there are m measured data points, the expression for the sum of squared residuals is: ; In the formula: Let j be the observed value of the stress component in the j-th direction at the k-th measured point. The simulated value of the j-direction stress component at the k-th measured point under the i-th type of structural boundary loading condition.

[0040] Among them, the normal equation based on the least squares method yields the expression for the optimal solution when the sum of squared residuals Sresidual is minimized: ; Among them, multiple linear stepwise regression analysis adds an evaluation of the significance of the regression components to the basic multiple linear regression analysis, enabling the elimination of insignificant influencing factors. Based on existing measured stress data, the least squares method is used to perform multiple linear stepwise regression analysis to calculate and obtain the boundary condition types and their corresponding regression coefficients. Compared with the conventional multiple linear regression inversion method for geostress, multiple linear stepwise regression analysis, through the F-test, introduces the significantly influential model boundary loading mode into the regression analysis, thereby determining the boundary loading mode.

[0041] Assuming different model boundary loading modes , , … The stress values ​​obtained from the measurement points are used as the independent variables in the regression. , , … Given the measured stress value as the dependent variable Y, construct a univariate regression model with each of the n independent regression variables and the dependent variable Y as follows: ; For a constructed regression model, the expression for calculating the F-test statistic Fi of the regression coefficients corresponding to the independent variables is: ; In the formula: RSS is the regression sum of squares, and ESS is the residual sum of squares. To simulate the stress values, This represents the average value of the measured stress values. denoted as , where n is the number of stress boundary modes with significant influence, and p is the number of measured stress values.

[0042] For a given significance level, calculate the corresponding critical value F1, and take the maximum value among the F-test statistics in the same group as Fimax. If Fimax ≥ F1, then the regression independent variable corresponding to the Fimax value is... This is introduced into the regression model. Then, the dependent variable Y and the subset of independent variables { are established. , },…,{ , }, { , },…,{ , For a binary regression model, calculate the statistical value of the regression coefficient F-test and the critical value F2 corresponding to a given significance level. If the maximum statistical value of the regression coefficient F-test is not less than F2, then the corresponding independent variable is... Introduce it into the regression model. Repeat the above steps, each time selecting one independent variable from those not introduced into the regression model, until the test shows that no variable has been introduced.

[0043] Based on the obtained stress values ​​at measurement points under various unit boundary conditions, the stress field regression expression of the current model in the valley is calculated using the least squares method. Furthermore, based on the results of stepwise regression analysis, the combination type of boundary loading and the range of loading values ​​are determined.

[0044] Specifically, firstly, based on the formation mechanism of geostress, a multiple linear relationship between the initial geostress field and its influencing factors is constructed. These influencing factors include at least horizontal compression in the x-direction, horizontal compression in the y-direction, self-weight in the z-direction, and shearing action in the xy, xz, and yz planes. The effects of geological tectonic activity and temperature on the geostress field are also considered. Subsequently, in numerical simulation software, different types of unit boundary loading conditions are applied to a three-dimensional geological model that has undergone multi-stage valley incision. Calculations are performed under elastic constitutive constraints to obtain the stress response values ​​at each measuring point. The stress calculation results under various loading modes are used as the independent variables in the regression, and the measured stress values ​​are used as the dependent variable to establish a multiple linear regression model. In the regression analysis, a residual sum of squares function is constructed based on the least squares method, and the initial estimates of the regression coefficients are obtained by solving the normal equation. Based on this, a stepwise regression method is further used to test the significance of each boundary loading mode. The F-test statistics for each regression variable are calculated and compared with the critical values ​​at a set significance level. Variables with test statistics greater than the critical values ​​are introduced into the regression model, while insignificant variables are removed, thus identifying the boundary loading parameters that have a significant impact on the geostress field. Specifically, a univariate regression model is first established for each single loading mode, and the significant variable with the largest F-value is selected for inclusion in the model. Then, new variables are gradually introduced to construct a multivariate regression model, and the significance test is repeated until no new variables meet the inclusion conditions, thus obtaining the final regression model structure and corresponding regression coefficients. After determining the boundary loading parameter types, reasonable value ranges for each boundary loading parameter are determined based on the value range of each regression coefficient and engineering experience, forming boundary loading range data. Based on this, to improve the representativeness and computational efficiency of parameter combinations, an orthogonal design method is used to set multiple gradient levels within the loading range, and several representative boundary loading combination schemes are generated through orthogonal combination. Each combination of schemes is loaded into the three-dimensional geological model as input conditions to carry out numerical simulation calculations of the valley incision process and obtain stress component values ​​at each measuring point under different loading combinations.

[0045] Understandably, by introducing multiple linear stepwise regression analysis to screen various boundary loading modes, factors with little impact on the geostress field can be eliminated, thereby improving the simplicity and interpretability of the model. The combination of least squares method and F-test improves the accuracy and stability of boundary loading parameter identification. The orthogonal design method reduces the number of combinations while ensuring parameter space coverage, lowers computational costs, improves modeling efficiency, and achieves efficient screening and reasonable combination of boundary loading parameters.

[0046] In some embodiments of this application, constructing the training sample set includes: Several boundary loading combinations were paired with geostress simulation data to form sample data pairs; the sample data pairs were randomly sorted; and the randomly sorted sample data pairs were divided into training sets and test sets. The training set is used for weight updates in the neural network model, while the test set is used to validate the neural network model.

[0047] Specifically, based on the orthogonal design principle, gradient values ​​are selected within the range of boundary condition loading values, and different boundary loading combinations are designed. These loading combinations correspond to the output layer data in the backpropagation neural network training samples. Based on nonlinear constitutive relations, the designed boundary loading combinations are applied to the valley incision model, and different stages of river incision are simulated through multi-stage valley model excavation, thereby calculating the current stress field distribution of the valley under different boundary loading combinations. By recording and organizing the stress component values ​​at measured stress points under different boundary loading combinations, the input layer data corresponding to each output layer in the neural network training samples are obtained. The neural network training samples are randomly sorted, and 70% of the data is selected as the training set and 30% as the test set for training and validation of the neural network.

[0048] Specifically, based on the orthogonal design principle, several gradient levels are set within the range of values ​​for each boundary loading parameter. Multiple representative boundary loading combination patterns are generated through orthogonal combination. Each boundary loading combination corresponds to a set of defined loading parameter values, which serve as the target data for the output layer of the subsequent neural network model. Then, each boundary loading combination is applied to the three-dimensional geological model one by one, and numerical simulations are conducted under nonlinear constitutive constraints. By simulating the valley incision process in stages, the current stress field distribution of the valley under different loading combinations is obtained. During the simulation, stress component values ​​at each measured stress point are extracted and used as input data corresponding to that boundary loading combination. Thus, the "measured stress component data" and the "corresponding boundary loading parameter combination" are paired one-to-one to form sample data pairs, thereby constructing a complete training sample set. After the sample construction is completed, to avoid the sample order from biasing the neural network training process, all sample data pairs are randomly sorted to ensure that different loading combinations and their corresponding stress responses are evenly distributed during training. Furthermore, the randomly sorted sample data is divided into a training set and a test set according to a preset ratio. Preferably, approximately 70% of the samples are selected as the training set for updating and iteratively optimizing the weight parameters of the neural network model, while the remaining approximately 30% are used as the test set to independently verify the model's generalization ability and prediction accuracy during or after training. In practice, the performance of the neural network model can be evaluated by calculating the prediction error index on the test set samples, and the network structure or training parameters can be adjusted accordingly.

[0049] Understandably, by combining orthogonal design with numerical simulation to construct the training sample set, the samples can cover the boundary loading parameter space with a smaller number of combinations, improving sample representativeness and modeling efficiency. By constructing sample data through input-output pairs, the correspondence between stress response and boundary loading parameters is established, providing a reliable foundation for neural networks to learn nonlinear mappings. By random sorting and dividing the training and test sets, the overfitting problem in the model training process is avoided, improving the model's generalization ability and prediction accuracy.

[0050] In some embodiments of this application, training a neural network model based on a training sample set includes: Define the network structure parameters of the neural network model and use the root mean square error as the loss function; train the neural network model based on the training sample set, and update the neural network model through optimization algorithms until the loss function is less than the loss threshold, thus obtaining the trained neural network model.

[0051] In some embodiments of this application, updating the neural network model through an optimization algorithm includes: Convergence optimization is performed based on the Levenberg-Marquardt algorithm; the Jacobian matrix of the first derivative of the error with respect to all parameters is defined, and a damping factor is introduced to adjust the parameter update step size; when the error decreases, the damping factor is decreased; when the error does not decrease, the damping factor is increased; the weight coefficients are repeatedly adjusted until the loss function is less than the loss threshold. Here, the error is the difference between the output value of the neural network model and the training sample set.

[0052] The backpropagation neural network is a multi-layer feedforward neural network that uses an error backpropagation algorithm for model training. Its network structure mainly consists of an input layer, hidden layers, and an output layer. Each layer comprises several neurons. The hidden layers can be configured according to the network's performance requirements. The expression for multiple hidden layers is as follows: ; In the formula: This represents the input layer of a neural network. For the expression of the output layer; For the constructed neural network structure, where These represent the number of nodes in the input layer, the 1st hidden layer, ..., the pth hidden layer, and the output layer, respectively. The number of nodes in the input layer is the set of six stress component values ​​for all measured points, and the number of nodes in the output layer is the set of boundary loading condition values ​​for the selected model.

[0053] In this network, neurons in different layers are propagated through weight coefficients. During the forward propagation of the neural network, the expressions for the input and output of hidden layer neuron j are: ; The expressions for the input and output of neuron l in the output layer are as follows: ; In the formula: netj is the input of hidden layer neuron j. Let be the weights from the input layer neuron to the hidden layer neuron j. Here, netk is the input to neuron l in the output layer, and yk is the output of neuron l in the output layer. The weights from the input layer neurons to the hidden layer neurons are denoted as l. This is the bias for the corresponding hidden layer.

[0054] Among them, for the error function, the expression for choosing the root mean square error (MSE) as the evaluation index is: ; The calculated error is backpropagated through a neural network, updating the weights and biases of each neuron layer by layer. Through multiple iterations, the error is reduced to a set threshold, resulting in a fully trained neural network model. The Levenberg-Marquardt algorithm, combining gradient descent and Gauss-Newton methods, is applied for convergence optimization. The Jacobian matrix J is defined, containing the first derivative of the error with respect to all parameters.

[0055] The second-order Taylor expansion of the loss function E(w) around the parameter w is expressed as: ; In the formula: g is the gradient vector, H is the Hessian matrix, is the second derivative matrix, and e is the error vector of all samples.

[0056] For the sum of squared errors, the expression for the Hessian matrix is: ; Introducing damping factor This allows for adjustments to the parameter update step size. The expression to be determined is: ; In the formula: Let be the approximate Hessian matrix of the Gauss-Newton method. This approximation is accurate when the algorithm approaches the optimal solution and can provide a second-order convergence speed; I is the identity matrix. When the value is very large, the equation is approximated as ,Right now This degenerates into gradient descent with a very small step size, suitable for the early stages of iteration or when errors increase, ensuring the stability of the neural network; when When the value is very small, the equation is approximately: It can provide fast, near-second-order convergence, suitable for approaches to the optimal solution. If the calculated error decreases after adjustment, the weights are updated, and the value is reduced. If the error does not decrease, then keep the weight unchanged and increase it. Then, the weights are recalculated. The weight coefficients are repeatedly adjusted until the loss function E(w) is less than the threshold or the gradient is small enough, thus obtaining a fully trained neural network model.

[0057] Specifically, based on the backpropagation neural network principle and training set sample data, the number of nodes in the input and output layers of the neural network is set according to the number of measurement points and the number of model boundary loading groups obtained by stepwise regression. Training parameters such as the hidden layer structure and learning rate of the neural network are adjusted and determined, and the neural network is trained. The root mean square error is used as the loss function to verify the performance of the neural network, thereby realizing the construction of a nonlinear mapping between the stress data of the measurement points and the model loading boundary conditions, and obtaining a neural network model that meets the performance requirements.

[0058] Specifically, the structural parameters of the neural network model are first set based on the input and output features of the training samples. The number of nodes in the input layer is set to the set dimension of the six stress components at all measurement points, and the number of nodes in the output layer is set to the number of boundary loading parameters obtained through multivariate linear stepwise regression. One or more hidden layers are set according to the complexity of the problem, and the weight coefficients and bias parameters of each layer's neurons are initialized. Subsequently, the mean squared error is used as the loss function to measure the deviation between the output of the neural network model and the true values ​​of the training samples. During model training, the training samples are input into the neural network, and forward propagation is performed first: the input layer data is passed to the hidden layers through the weight coefficients, and each hidden layer neuron obtains its output according to the weighted summation and activation function, which is then passed layer by layer to the output layer to obtain the predicted boundary loading parameter values. Then, based on the error between the predicted and true values, the loss function value is calculated, and the error signal is passed from the output layer to the input layer layer by layer through the error backpropagation algorithm. The gradient contribution of each layer's weights and biases to the error is calculated, thereby realizing parameter updates.

[0059] Specifically, in the parameter optimization process, the Levenberg-Marquardt algorithm is introduced to improve the traditional backpropagation process. This is achieved by constructing the Jacobian matrix, the first derivative of the error with respect to the model parameters, and combining it with an approximate Hessian matrix to achieve second-order optimization. Specifically, in each iteration, the error vector is calculated based on the current parameters, and the parameter update direction is solved using the Jacobian matrix. Simultaneously, a damping factor is introduced to adaptively adjust the update step size. When the error decreases after iteration, the damping factor is decreased to accelerate convergence; when the error increases, the damping factor is increased to ensure iterative stability. By continuously repeating the above forward propagation, error calculation, and parameter update process, the loss function gradually decreases. When the loss function value is below a preset threshold or the gradient change tends to stabilize, iterative training stops, resulting in a converged neural network model. After training, the model's performance can be validated using test set data. The model's generalization ability is evaluated by calculating the prediction error of the test samples. If the test error is large, the hidden layer structure, learning rate, or training parameters can be further adjusted, and the model can be retrained.

[0060] Understandably, by constructing a multi-layer neural network model based on the backpropagation mechanism, the complex nonlinear relationship between ground stress and boundary loading parameters can be characterized, thus improving the inversion accuracy. The introduction of the Levenberg-Marquardt optimization algorithm achieves an organic combination of gradient descent and the Gauss-Newton method, improving the convergence speed while ensuring computational stability. Through the adaptive damping factor adjustment mechanism, oscillation or divergence problems during training are avoided, thus improving the robustness and reliability of model training.

[0061] In some embodiments of this application, determining the development depth of high-energy-storage rock masses in a river valley based on the geostress data to be compared includes: The geostress data to be compared is compared with the measured geostress data. Based on the comparison results, it is determined whether the inversion results meet the preset error requirements. If the preset error requirements are not met, the target boundary loading parameters and the corresponding geostress simulation data are added as new samples to the training sample set, and the neural network model is retrained. If the preset error requirements are met, the development depth of the high-energy-storage rock mass in the valley is determined.

[0062] The process involves extracting the stress component values ​​of each measuring point in the geostress data to be compared and the corresponding stress component values ​​of each measuring point in the actual geostress data; determining the difference; generating an error index based on the difference; and comparing the error index with an error threshold. If the error index is less than or equal to the error threshold, the inversion result is determined to meet the preset error requirements; if the error index is greater than the error threshold, the inversion result is determined to not meet the preset error requirements.

[0063] Specifically, existing measured stress values ​​are input into the trained neural network model for calculation, and the output layer results become the corresponding model boundary loading conditions. These results are then substituted into the valley incision model, and based on nonlinear constitutive relations, the valley incision process is simulated to calculate the current valley stress field. The calculated results are then compared with the measured results. If the calculated stress values ​​at the measuring points differ significantly from the measured values, the results are used as a training sample and loaded into the neural network for retraining. The measured stress data is then substituted into the retrained neural network for calculation, and the boundary loading conditions obtained after recalculating the valley model are used to calculate the stress field. If the calculated stress values ​​still do not meet the error requirements, the above steps are repeated, and the recalculated data is used as a training set to input into the neural network for the next round of calculations until the valley stress calculation results meet the error requirements, thus achieving a valley stress field inversion simulation considering the river incision process.

[0064] Specifically, the measured geostress data is first input into a trained neural network model for forward computation to obtain the corresponding model boundary loading parameters. These boundary loading parameters are then applied to a three-dimensional geological model, and a multi-stage downcutting process of the valley is numerically simulated under nonlinear constitutive constraints. The calculated geostress values ​​at each measuring point in the current valley state are obtained as the geostress data to be compared. Stress component values ​​for each measuring point are extracted from the geostress data to be compared, and the measured geostress component values ​​for the corresponding measuring points are also extracted. The two are matched one-to-one, and the differences between the corresponding components are calculated. Based on this, an error index is constructed according to a preset error evaluation method. The error index can be in the form of the average value, root mean square error, or weighted error of each component error at each measuring point, to comprehensively reflect the overall inversion accuracy. Subsequently, the error index is compared with a preset error threshold. When the error index is less than or equal to the error threshold, the current inversion result is determined to meet the accuracy requirements; when the error index is greater than the error threshold, the current inversion result is determined to not meet the accuracy requirements.

[0065] Specifically, when the inversion results do not meet the error requirements, the boundary loading parameters obtained from this inversion and their corresponding numerical simulation stress data are used to construct new sample data pairs, which are then added to the original training sample set to expand the sample space and enhance the model's ability to fit the current working conditions. Subsequently, the neural network model is retrained based on the updated training sample set, updating the network weights and bias parameters. After retraining, the measured stress data is input into the neural network model again for inversion calculation, and the above loading, simulation, and error comparison process is repeated. Through continuous iteration of the closed-loop process of "inversion—simulation—verification—retraining," the neural network model gradually approximates the characteristics of the real geostress field. In actual implementation, a maximum number of iterations or a minimum error change threshold can be set as auxiliary termination conditions. When the error meets the preset accuracy requirements or the iteration process converges, the iterative calculation is stopped, and the final boundary loading parameters and corresponding geostress field results are output.

[0066] Understandably, by constructing an iterative verification mechanism based on error feedback, the neural network inversion results and numerical simulation results are validated in a closed loop, which can continuously correct model biases and improve inversion accuracy. By dynamically expanding the training samples, the model can continuously learn new stress response characteristics, enhance its adaptability to complex geological conditions, and avoid the uncertainty problem of single inversion results.

[0067] In some embodiments of this application, determining the development depth of high-energy-storage rock masses in river valleys includes: Based on the geostress data to be compared, the elastic strain energy density of each unit in the three-dimensional geological model is determined; the elastic strain energy density of each unit is compared with the preset energy storage criterion threshold. Units with elastic strain energy density greater than the preset energy storage criterion threshold are identified as high-energy-storage rock mass development areas; the development depth of high-energy-storage rock mass is determined based on the spatial coordinates of the high-energy-storage rock mass development area in the three-dimensional geological model.

[0068] Among them, strain energy is a kind of potential energy stored in rock and soil in the form of strain and stress. Assuming that no heat exchange occurs during the deformation of rock under stress, the strain energy in the rock mass consists of dissipated energy and elastic strain energy. The former is the energy consumed by the rock mass under stress due to deformation and microcrack expansion, while the latter is the energy stored inside the rock mass, which will be released when the stress is unloaded.

[0069] Wherein, strain energy density is the energy stored per unit volume of rock mass. Based on the principle of elastic strain energy, the expression for elastic strain energy density is: ; In the formula: Principal stress, The strain corresponding to the principal stress direction For elastic strain energy density, The volumetric strain energy density, Let be the shape strain energy density, and its expression is: ; In the formula: E The elastic modulus of the soil and rock mass. It is Poisson's ratio.

[0070] Specifically, based on the calculated stress field of the valley, the stress conditions of each element within the model are extracted by traversing the model mesh parameters. Using the elastic energy calculation formula, the strain energy density of the current valley is calculated and analyzed to obtain the volumetric strain energy density and shape strain energy density of each element within the model. These are then added together with the model mesh values ​​to obtain the elastic strain energy density model of the current valley. Currently, there is no clear classification standard for high-energy-storage rock masses. Based on previous research, 40 kJ / m³ can be used as the classification criterion for high-energy-storage rock masses to identify the development of elastic strain energy density in the valley, thereby obtaining the development depth of high-energy-storage rock masses.

[0071] Specifically, the inverted geostress field results are first mapped to each grid cell of the three-dimensional geological model, and the stress state parameters corresponding to each cell are extracted, including principal stress values ​​and their spatial distribution characteristics. Based on this, and assuming no heat exchange occurs during the deformation of the rock mass according to elasticity theory, the internal energy of the rock mass is divided into dissipated energy and elastic strain energy, where elastic strain energy is recoverable energy and used to characterize the energy storage capacity of the rock mass. For each model cell, the elastic strain energy density is calculated based on its stress state and corresponding strain relationship. In practice, the elastic strain energy density can be decomposed into two parts: volumetric strain energy density and shape strain energy density. The volumetric strain energy density reflects the energy stored by the volume change of the rock mass, and the shape strain energy density reflects the energy stored by shear deformation. Combining the rock mass's elastic modulus and Poisson's ratio parameters, the stress state of each cell is calculated, and the volumetric strain energy density and shape strain energy density are obtained separately. The total elastic strain energy density of each cell is then obtained by superposition.

[0072] Specifically, in the implementation process, the three-dimensional geological model is traversed cell by cell, and the calculated elastic strain energy density is assigned to the corresponding grid cell, thereby constructing a complete elastic strain energy density distribution model of the valley area. Subsequently, the elastic strain energy density of each cell is compared with a preset energy storage criterion threshold. In this embodiment, 40 kJ / m³ can be selected as the discrimination threshold for high-energy-storage rock masses. When the elastic strain energy density of a cell is greater than this threshold, the cell is marked as a high-energy-storage rock mass development cell. By performing spatial clustering and connectivity analysis on all cells that meet the conditions, the high-energy-storage rock mass development area is identified, and based on the spatial coordinates of this area in the three-dimensional model, its vertical distribution range is extracted, thereby determining the development depth of the high-energy-storage rock mass in the valley.

[0073] Understandably, quantitative characterization of rock mass energy storage capacity using calculation methods based on elastic strain energy density reflects the energy accumulation state of rock mass under complex stress conditions, improving the scientific rigor and accuracy of identifying high-energy-storage rock masses; and refined determination of the development depth of high-energy-storage rock masses is achieved through unit-by-unit calculation and spatial positioning using a three-dimensional model.

[0074] See Figure 2-4 As shown, and illustrated through the following practical examples: The method of the present invention will be specifically described using a typical deep valley engineering project as an example.

[0075] Step 1: In this project, a rectangular area of ​​approximately 4000m × 8000m was selected within the horizontal plane as the boundary of the research model. The angle between the model's major axis and true north was 33°. The bottom interface of the model was set 2000m below the ground surface, and the top surface was constructed using high-precision DEM data to create a three-dimensional topographic surface. The model internally considered two main branch fault structures, modeled as solid elements, with a fault zone width of 100m. The lithology was divided into three categories: quartzite, pore-shaped metamorphic rock, and mélange, each assigned corresponding mechanical parameters. Considering the combined effects of tectonic uplift and rapid river downcutting in the valley, and combining DEM slope analysis results and terrace elevation information, a generalized model of the multi-stage downcutting process of the valley was created. Simultaneously, for the current valley topography, a strong unloading zone was set within 0–30m below the ground surface, and a weak unloading zone within 30–100m below the ground surface, to reflect the slope unloading effect.

[0076] Step Two: In determining the boundary loading mode, firstly, based on the multiple linear regression method, six types of unit boundary loading conditions were applied to the initial model (which did not consider valley incision) to obtain the stress response at each measuring point. Regression analysis and significance testing were then performed using the measured stress data. The analysis results show that the horizontal compressive stress in the x and y directions, as well as the self-weight term, have a significant impact on the regional geostress field, while the significance of shear-type boundary loading is low and can be ignored. Based on this, horizontal bidirectional compression and self-weight were selected as the main boundary loading parameters, and their value ranges were determined. Table 1 below shows the correspondence between the significance of the boundary loading modes and the regression coefficients for each model.

[0077] Table 1. Significance of Boundary Loading Modes and Correspondence with Regression Coefficients for Each Model

[0078] Step 3: An orthogonal design method was used to construct a boundary loading combination scheme. Multiple loading calculations were performed on the 3D model, and numerical simulations were conducted in conjunction with the gradual downcutting process of the valley. At different downcutting stages, the mechanical parameters of the strong unloading zone and the weak unloading zone were adjusted to reflect the evolution of the rock mass's mechanical properties. Stress component values ​​at measuring points under each loading combination were extracted to construct neural network training samples. Using the stress components as input and the boundary loading parameters as output, a backpropagation neural network model was established, and the training and test sets were divided at a ratio of 70% and 30% for model training and validation. After multiple iterations and optimizations, a neural network model meeting the error requirements was obtained. In the model application stage, the geostress data from six measured points were input into the trained neural network model to invert and obtain the boundary loading parameters of the valley-scale model. The x-direction loading was approximately 1.76, the y-direction loading was approximately 2.25, and the z-direction self-weight stress coefficient was approximately 0.915. This inversion result was loaded into the 3D model to simulate the multi-stage downcutting process of the valley, obtaining the geostress field distribution characteristics at different evolution stages. The table below shows the boundary loading conditions of the valley model based on BP neural network inversion.

[0079] Table 2. Correspondence of Boundary Loading Conditions for Valley Models Based on BP Neural Network Inversion

[0080] Step 4: Further, based on the elastic strain energy calculation method, the strain energy density of each element in the model is calculated, and a three-dimensional distribution cloud map is generated. Analysis results show that the strain energy density is low in the shallow area of ​​the valley, while in the longitudinal profile Y-Y', when the elevation is below 1200m, the elastic strain energy density in most areas exceeds 40kJ / m³, indicating that this depth range is a high-energy-storage rock mass development zone. Therefore, it can be determined that the high-energy-storage rock mass in this valley is mainly developed in the deep area below approximately 1200m.

[0081] In summary, by leveraging the nonlinear characteristics of the backpropagation neural network, the influence of surface alteration on the in-situ stress in the training samples can be incorporated into the construction process of the relationship between the model's boundary loading conditions and the measured stress. Furthermore, by iteratively expanding the training set, the inversion accuracy is ensured. Combined with energy calculation principles, the elastic strain energy density of the present-day river valley can be calculated. Based on the energy criterion for high-energy-storage rock masses, the development depth of these rock masses is obtained.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for determining the development depth of high-energy rock mass in a valley based on analog inversion, characterized in that, include: Acquire geological information data and measured geostress data; A three-dimensional geological model is constructed based on the geological information data, and stress coordinate transformation is performed on the measured geostress data according to the coordinate system of the three-dimensional geological model. Filter boundary loading parameters and boundary loading range data; select gradient values ​​within the boundary loading range data and generate several boundary loading combinations; input the several boundary loading combinations into the three-dimensional geological model to obtain the geostress simulation data corresponding to each boundary loading combination; construct a training sample set based on the boundary loading combinations and geostress simulation data; and train a neural network model based on the training sample set. The measured geostress data is input into the trained neural network model to invert and obtain the target boundary loading parameters; the target boundary loading parameters are input into the three-dimensional geological model to determine the geostress data to be compared; and the development depth of the high-energy-storage rock mass in the valley is determined based on the geostress data to be compared.

2. The method for determining the development depth of high-energy-storage rock masses in river valleys based on simulation inversion as described in claim 1, characterized in that, When performing stress coordinate transformation on the measured geostress data, the following steps are included: Extract the principal stress values, principal stress azimuths, and principal stress dip angles from the measured geostress data; determine the direction cosines between the principal stress vectors and each axis of the standard initial coordinate system based on the principal stress azimuths and the principal stress dip angles; and determine the stress component values ​​in the standard initial coordinate system based on the direction cosines and the principal stress values.

3. The method for determining the development depth of high-energy-storage rock masses in river valleys based on simulation inversion according to claim 2, characterized in that, When performing stress coordinate transformation on the measured geostress data, the method further includes: When the coordinate system of the three-dimensional geological model is inconsistent with the standard initial coordinate system, the direction cosines between the coordinate axes of the three-dimensional geological model and each axis of the standard initial coordinate system are obtained, and the stress component values ​​are subjected to a second coordinate transformation based on the stress transformation matrix to obtain stress data consistent with the three-dimensional geological model.

4. The method for determining the development depth of high-energy-storage rock masses in river valleys based on simulation inversion as described in claim 1, characterized in that, When filtering boundary load parameters and boundary load range data, the following should be included: Construct a multiple linear relationship between the initial stress field and influencing factors; calculate the test statistics of the regression coefficients corresponding to each boundary loading mode; Set a critical value; compare the test statistics with the critical value; introduce the boundary loading patterns corresponding to the test statistics greater than the critical value into the regression model, and eliminate the boundary loading patterns less than the critical value to obtain the boundary loading parameters; determine the boundary loading range data based on the results of the multiple linear regression analysis.

5. The method for determining the development depth of high-energy-storage rock masses in river valleys based on simulation inversion according to claim 4, characterized in that, When selecting a gradient value within the boundary loading range data and generating several boundary loading combinations, the following is included: Several gradient levels are set within the boundary loading range data; several boundary loading combinations are generated by orthogonally designing and combining the gradient levels.

6. The method for determining the development depth of high-energy-storage rock masses in river valleys based on simulation inversion according to claim 5, characterized in that, When constructing the training sample set, the following are included: Several of the aforementioned boundary loading combinations are paired with the geostress simulation data to form sample data pairs; the sample data pairs are randomly sorted; and the randomly sorted sample data pairs are divided into training sets and test sets. The training set is used for weight updates of the neural network model, and the test set is used for validation of the neural network model.

7. The method for determining the development depth of high-energy-storage rock masses in river valleys based on simulation inversion as described in claim 6, characterized in that, Training a neural network model based on the training sample set includes: The network structure parameters of the neural network model are set, and the root mean square error is used as the loss function. The neural network model is trained based on the training sample set, and the neural network model is updated through an optimization algorithm until the loss function is less than the loss threshold, thus obtaining the trained neural network model.

8. The method for determining the development depth of high-energy-storage rock masses in river valleys based on simulation inversion according to claim 7, characterized in that, When updating the neural network model using an optimization algorithm, the following are included: Convergence optimization is performed based on the Levenberg-Marquardt algorithm; a Jacobian matrix of the first derivatives of the error with respect to all parameters is defined, and a damping factor is introduced to adjust the parameter update step size; when the error decreases, the damping factor is decreased; when the error does not decrease, the damping factor is increased; the weight coefficients are repeatedly adjusted until the loss function is less than the loss threshold. The error is the difference between the output value of the neural network model and the training sample set.

9. The method for determining the development depth of high-energy-storage rock masses in river valleys based on simulation inversion according to claim 1, characterized in that, When determining the development depth of high-energy-storage rock masses in river valleys based on the aforementioned geostress data to be compared, the following are included: The stress data to be compared is compared with the measured stress data. Based on the comparison results, it is determined whether the inversion results meet the preset error requirements. If the preset error requirements are not met, the target boundary loading parameters and the corresponding stress simulation data are added as new samples to the training sample set, and the neural network model is retrained. If the preset error requirements are met, the development depth of the high-energy-storage rock mass in the valley is determined.

10. The method for determining the development depth of high-energy-storage rock masses in river valleys based on simulation inversion according to claim 9, characterized in that, Determining the development depth of high-energy-storage rock masses in river valleys includes: Based on the stress data to be compared, the elastic strain energy density of each unit in the three-dimensional geological model is determined; the elastic strain energy density of each unit is compared with the preset energy storage criterion threshold. Units with elastic strain energy density greater than the preset energy storage criterion threshold are identified as high-energy-storage rock mass development areas; the development depth of the high-energy-storage rock mass is determined based on the spatial coordinates of the high-energy-storage rock mass development area in the three-dimensional geological model.

Images