A pore-fracture dual-medium rock stratum water storage capacity prediction method

Abstract

The application discloses a pore-fracture dual-medium rock stratum water storage capacity prediction method in the technical field of coal mine "non-conventional water resource" deep storage and injection, and aims to solve the problem of inaccurate water storage capacity prediction. The method comprises the following steps: differentiating seepage modeling of different scale fractures, fusing sub-models to obtain an initial seepage model, considering the discrete differentiation effect of large and medium scale fractures and the local channeling mechanism of the pore-fracture interface, truly restoring the space-time distribution law of deep fluid in multi-scale media, and improving the accuracy of water storage capacity prediction; the water storage capacity predicted according to the dynamic seepage model fully considers the flow field and the dynamic feedback mechanism of permeability under the water-rock interaction, so that the distortion degree of the dynamic water storage capacity obtained and the calculated injection pressure prediction result is low; the physical law used in finite element evolution analysis is changed into a residual calculation rule of a physical information neural network, and then the high calculation amount is avoided, and the water storage capacity prediction in a specific time is realized.

Description

Technical Field

[0001] This invention relates to a method for predicting the water storage capacity of pore-fracture dual-medium rock formations, belonging to the field of deep storage and reinjection technology of unconventional water resources in coal mines. Background Technology

[0002] Intensive mining generates large amounts of mine water characterized by high mineralization. For environmental protection, mine water must be properly treated.

[0003] However, existing membrane separation technologies such as reverse osmosis have a treatment cost of 12-30 yuan per ton of water and face serious problems such as complex processes, easy clogging and corrosion of membrane modules, and difficulty in harmlessly treating concentrated wastewater. Therefore, a new low-cost solution has emerged: directly injecting high-TDS mine water into deep geological reservoirs, such as aquifers or old goaf water layers, under micro-treatment conditions for deep geological storage and reinjection. However, in actual deep storage and reinjection projects, existing water storage assessment models are generally based on the ideal assumption of "static permeability," failing to consider that high-salinity mine water is prone to violent dissolution and precipitation reactions with sandstone reservoir minerals under deep high temperature and pressure, leading to nonlinear dynamic decay or expansion of media permeability. Ignoring this dynamic feedback mechanism of flow field and permeability under water-rock interaction will result in serious distortion of the predicted dynamic water storage and injection pressure under long-term reinjection conditions. In addition, there is a lack of three-dimensional modeling technology that can accurately characterize the spatial heterogeneity of dual media. Traditional equivalent continuous medium models oversimplify the connectivity of fracture networks, completely obliterating the discrete differentiation effect of large and medium-scale fractures and the local flow mechanism at the pore-fracture interface, and failing to truly reproduce the spatiotemporal distribution of deep fluids in multi-scale media.

[0004] In summary, existing methods for predicting water storage capacity in pore-fracture dual-medium rock formations suffer from inaccurate predictions. Summary of the Invention

[0005] The purpose of this application is to overcome the shortcomings of the prior art and provide a method for accurately predicting water storage in pore-fracture dual-medium rock formations.

[0006] To achieve the above objectives, this application employs the following technical solution:

[0007] In a first aspect, this application provides a method for predicting the water storage capacity of pore-fracture dual-medium rock formations, including:

[0008] Hydrogeological data of the rock strata are obtained, and differential seepage models are performed on the pores and fractures of different scales in the rock strata based on the hydrogeological data to obtain multiple sub-models; the multiple sub-models are then merged into the initial seepage model.

[0009] Based on the aforementioned hydrogeological data, the porosity and permeability evolution mechanism is introduced into the initial seepage model to obtain a dynamic seepage model;

[0010] Collocation points are generated from the dynamic seepage model, a composite objective loss function is constructed based on the physical residuals, and the physical information neural network is trained using the collocation points and the composite objective loss function to obtain the trained physical information neural network.

[0011] The trained physical information neural network is used to predict water storage.

[0012] Furthermore, the process involves acquiring hydrogeological data of the rock strata, and then using this data to perform differentiated seepage modeling on the pores and fractures of different scales within the rock strata, resulting in multiple sub-models, including:

[0013] A fracture scale classification standard is established, which classifies all fractures from largest to smallest into large-scale fractures, medium-scale fractures, and small-scale fractures.

[0014] A three-dimensional seismic data volume of the rock strata is acquired. Fault planes are identified from the three-dimensional seismic data volume using variance volume and curvature volume techniques in attribute extraction. An ant tracking algorithm is used to filter noise from the fault planes through spatial multi-directional slicing, resulting in filtered fault planes. Based on the filtered fault planes and the fracture scale classification criteria, equivalent discrete surfaces matching large-scale fractures are generated in the model space, and these equivalent discrete surfaces are used as large-scale fracture sub-models. Based on the filtered fault planes and the fracture scale classification criteria, mesoscale fracture sub-models are generated in the model space. The filtered fault planes are binarized to obtain a binarized fault image with a resolution not lower than a resolution threshold. Different sized grids are used to extract noise from the binarized fault image. The computational space is divided; the range operation is performed on the maximum value of the dense region and the minimum value of the sparse region in the computational space to obtain fractal features; the fractal dimension of the fractal features is obtained, and a random number generator generates discrete fracture segments that satisfy the power-law distribution feature using the fractal dimension as input; the discrete fracture segments are dimensionally reduced using spherical coordinates and stereographic projection techniques to obtain dimension-reduced discrete fracture segments; the elbow rule is used to cluster the dimension-reduced discrete fracture segments to obtain clustered discrete fracture segments; Monte Carlo sampling is performed on the clustered discrete fracture segments using the Fisher probability distribution function to generate a discrete fracture network with random orientation and inclination angle with variance constraints, and the discrete fracture network is used as a small-scale fracture medium sub-model;

[0015] The hydrogeological data includes micro- and nano-scale Micro-CT scan images of the rock strata. Based on these images, characteristic parameters of the connecting pore throats are obtained. Matrix porosity, pore size distribution, and capillary pressure curves are extracted from these parameters. A characterization equation is constructed based on the matrix porosity, pore size distribution, and capillary pressure curves. A continuous three-dimensional volumetric unit is constructed based on the characterization equation. This continuous three-dimensional volumetric unit serves as a sub-model of the pores.

[0016] The process of fusing multiple sub-models into an initial seepage model includes:

[0017] A dual-flow medium spatial framework is established, and the large-scale fracture sub-model, the medium-scale fracture sub-model, the small-scale fracture sub-model, and the pore sub-model are nested in the same dual-flow medium spatial framework to obtain the initial flow model.

[0018] Further, the step of obtaining the fractal dimension of the fractal feature, and using the fractal dimension as input, generating discrete line segments of the crack that satisfy the power-law distribution feature by a random number generator, includes:

[0019] The length of the discrete segment of the crack is obtained based on the fractal dimension, and the formula for obtaining the length of the discrete segment of the crack is as follows:

[0020] ,

[0021] In the formula, The length of the discrete line segment of the crack. These are random numbers that are uniformly distributed within the interval. This represents the maximum crack length. This is the minimum crack length. For fractal dimension, The width of the box is the dimension used in the box-counting formula. To cover the number of box layout experiments, For the counting sequence number, For the first The maximum number of cracks obtained from the statistics within the covered box in this experiment. For the first The minimum number of cracks obtained from the statistics inside the covered box in this experiment;

[0022] The attitude of the fracture is generated using a Fisher distribution model, which includes the fracture dip direction and the fracture dip angle. The spatial orientation probability density functions of the fracture dip direction and the fracture dip angle are as follows:

[0023] ,

[0024] In the formula, The value of the spatial orientation probability density function. It is prone to fractures. The angle of the fracture dip. The model distribution parameters characterize the degree of concentration of the distribution. It is a hyperbolic sine function;

[0025] The discrete line segment of the crack is generated based on the spatial orientation distribution probability density function and the length of the discrete line segment of the crack.

[0026] Furthermore, the step of introducing the porosity and permeability evolution mechanism into the initial seepage model based on the hydrogeological data to obtain a dynamic seepage model includes:

[0027] The porosity and permeability evolution mechanism includes the flow field equation of the fracture system, the flow field equation of the pore system, and the dynamic coupling relationship of pore-fracture crossflow;

[0028] The flow field equation for the fracture system is:

[0029] ,

[0030] In the formula, To comprehensively compress the elastic water storage coefficient, The fluid pressure in the fracture. For time, For Hamiltonian operators, For dynamic permeability tensor, The porosity is a dynamic fissure that is subject to real-time changes due to chemical precipitation and dissolution. For fluid dynamic viscosity, For fluid density, It is the acceleration due to gravity. For depth, The pore-fracture flow rate, For source and sink items;

[0031] The flow field equation for the porous system is:

[0032] ,

[0033] In the formula, To comprehensively compress the elastic water storage coefficient, For the fluid pressure of the matrix, For dynamic permeability tensor, The porosity of the matrix is ​​dynamically altered in real time by chemical precipitation and dissolution. A non-Darcy initiation pressure gradient threshold correction term to account for the weak permeability characteristics of the pore matrix;

[0034] The dynamic coupling relationship of pore-fracture flow is as follows:

[0035] ,

[0036] In the formula, The shape exchange factor is used to reflect the dynamic updating of the geometric and physical characteristics of the pore-fracture interface as it undergoes dissolution and expansion.

[0037] Furthermore, the rock stratum is sandstone, and the porosity and permeability evolution mechanism also includes a sandstone permeability evolution formula, which is as follows:

[0038] ,

[0039] In the formula, for The sandstone permeability attenuation coefficient at any given time. The initial sandstone permeability decay coefficient. The permeability attenuation coefficient of sandstone under water quality control. To adjust the sandstone permeability growth coefficient at different temperatures, To fully acidify the sandstone permeability growth factor, and These are the initial response evolution coefficients of the permeability, and These represent the rates of change over time. and These are the coefficients of the power-law term. These are the fitting coefficients;

[0040] The core of the sandstone was obtained, the core was crushed and ground and sieved to below 60 mesh to obtain fine powder. An equal volume of high TDS mine water was added to the fine powder of equal mass to obtain a sample, and the diameter of the sandstone particles was recorded at the same time.

[0041] The samples were placed in a high-pressure reactor for 100 days of soaking. The soaking process was divided into three phases: the first phase at 25°C and the second and third phases at 50°C. At the set time points during the soaking process, a rapid water quality analyzer, ICP-OES, and ion chromatograph were used to sample and test the solution of the soaked samples to obtain permeability evolution time-series data under single water quality conditions and permeability evolution time-series data under temperature control conditions.

[0042] The samples were classified according to different pore characteristics to obtain samples with different pore characteristics; the samples with different pore characteristics were placed in a core displacement experimental device to conduct long-term continuous water injection displacement experiments on the cores under different pressure and temperature conditions; after the core displacement experiment, Micro-CT was used to obtain the pore three-dimensional structure variation map of the samples with different pore characteristics, and the sandstone porosity was obtained based on the pore three-dimensional structure variation map.

[0043] Based on the sampling and testing results and the three-dimensional pore structure variation map, the theoretical mapping between sandstone permeability and sandstone porosity is constructed using the following formula:

[0044] ,

[0045] In the formula, For sandstone permeability, For sandstone porosity, The diameter of the sandstone grains;

[0046] The sandstone permeability-sandstone porosity theoretical mapping and the permeability evolution time series data under the single water quality action condition are fitted to obtain the sandstone permeability attenuation coefficient under water quality control. The initial response evolution coefficient of permeability and the time change rate are obtained by fitting the sandstone permeability-sandstone porosity theoretical mapping and the permeability evolution time series data under the temperature control condition.

[0047] The fitting coefficients and the power-law coefficients are obtained by fitting the sandstone permeability-sandstone porosity theoretical mapping, the permeability evolution time series data under the single water quality action condition, and the permeability evolution time series data under the temperature control condition.

[0048] Further, the step of generating collocation points from the dynamic seepage model, constructing a composite objective loss function based on the physical residuals, and training a physical information neural network using the collocation points and the composite objective loss function to obtain a trained physical information neural network includes:

[0049] The Latin hypercube sampling algorithm is used to randomly distribute points throughout the dynamic seepage model.

[0050] The spatial coordinates and time series of the collocation points are input into the initially untrained physical information neural network to obtain the output fracture pressure, matrix pressure and target ion concentration.

[0051] The first and second derivative biases of the fracture pressure, matrix pressure, and target ion concentration output can be obtained using TensorFlow or PyTorch.

[0052] Substituting the first-order derivative partial and the second-order derivative partial into the Darcy equation and the convection-diffusion equation, the physical residual is obtained;

[0053] The composite objective loss function is constructed based on the physical residual and the boundary head loss. The expression of the composite objective loss function is as follows:

[0054] ,

[0055] In the formula, For a composite objective loss function, , , and These are the loss weight hyperparameters that are dynamically and adaptively adjusted during training through sensitivity analysis or neural tangent nuclei. The mean square error between the network prediction output calculated at the point of reference and the field measured or seismic inversion hydrological data is given. The physical residuals of the flow field control equations are... This is the physical residual term of the solute transport reaction equation. The sum of the mean square errors of the predicted solution at the domain boundary and at the initial time point, which are the deviations from the set physical boundary conditions and initial conditions.

[0056] When training the physical information neural network, the Adam optimization algorithm is used to perform global coarse adjustment of the physical information neural network with a large step size.

[0057] If the number of training epochs in which the composite objective loss function value is lower than a preset threshold exceeds a threshold, or if the smoothness of the decrease in the composite objective loss function value reaches a smoothness threshold, the L-BFGS quasi-Newton method is used to perform a second-order fine local search on the physical information neural network until the composite objective loss function value is lower than the loss threshold, thus obtaining the trained physical information neural network.

[0058] Furthermore, the step of using the trained physical information neural network to predict water storage includes:

[0059] By inputting a set time and three-dimensional coordinate matrix into the trained physical information neural network, the predicted fracture pressure, matrix pressure, and target ion concentration are obtained.

[0060] The water storage capacity, dynamic water-bearing porosity increment, and water pressure rise range within the effective influence radius of the three-dimensional coordinates were calculated using the water balance method and the integral rule.

[0061] The predicted water storage is obtained based on the water storage within the effective influence radius corresponding to multiple three-dimensional coordinates. The critical time of curtain flow or the area of ​​severe scaling is determined based on the dynamic water-bearing porosity increment and the water pressure rise range. A safety assessment report for optimizing single-well injection flow rate and injection pressure is generated based on the critical time of curtain flow or the area of ​​severe scaling.

[0062] Secondly, this application also provides a computer device, including a processor and a memory connected to the processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, it performs the steps of the water storage prediction method for pore-fracture dual-medium rock layers as described in any embodiment of the first aspect.

[0063] Thirdly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the water storage prediction method for pore-fracture dual-medium rock layers as described in any embodiment of the first aspect.

[0064] Fourthly, this application also provides a computer program product, including a computer program / instructions, characterized in that, when the computer program / instructions are executed by a processor, they implement the steps of the water storage prediction method for pore-fracture dual-medium rock layers as described in any embodiment of the first aspect.

[0065] Compared with the prior art, the beneficial effects achieved by this application are as follows:

[0066] The water storage prediction method for pore-fracture dual-medium rock formations provided in this application performs differentiated seepage modeling based on fractures of different scales in hydrogeological data, and then merges the sub-models to obtain an initial seepage model. It takes into account the discrete differentiation effect of large and medium-scale fractures and the local channeling mechanism of the pore-fracture interface, and realistically restores the spatiotemporal distribution law of deep fluids in multi-scale media, thereby improving the accuracy of water storage prediction. The water storage predicted based on the dynamic seepage model fully considers the dynamic feedback mechanism of flow field and permeability under water-rock interaction, resulting in low distortion of the dynamic water storage and calculated injection pressure prediction results. The physical laws used in finite element evolution analysis are transformed into residual calculation rules of physical information neural networks, thereby avoiding excessive computation and realizing the prediction of water storage at a specified time. Attached Figure Description

[0067] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0068] Figure 1 This is a flowchart of the steps of the water storage prediction method for pore-fracture dual-medium rock formations provided in the embodiments of this application;

[0069] Figure 2 This is a flowchart illustrating the overall technical process of the water storage prediction method for pore-fracture dual-medium rock formations provided in this application embodiment.

[0070] Figure 3 This is a schematic diagram illustrating the principle of fractal extraction and fractal discretization modeling fusion in the water storage prediction method for pore-fracture dual-medium rock layers provided in this application embodiment;

[0071] Figure 4 This is a topology diagram of the physical information neural network architecture used in the water storage prediction method for pore-fracture dual-medium rock layers provided in the embodiments of this application;

[0072] Figure 5 This is a schematic block diagram of the computer device provided in the embodiments of this application. Detailed Implementation

[0073] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments and specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations thereof. In the absence of conflict, the embodiments and technical features in the embodiments can be combined with each other.

[0074] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0075] Example 1:

[0076] Figure 1 This is a flowchart illustrating a method for predicting water storage in a pore-fracture dual-medium rock formation according to Embodiment 1 of the present invention. This flowchart merely shows the logical sequence of the method described in this embodiment. Without conflict, in other possible embodiments of the present invention, different methods may be used. Figure 1 Complete the steps shown or described in the order indicated.

[0077] See Figure 1 The method implemented in this way specifically includes the following steps:

[0078] Hydrogeological data of the rock strata are obtained, and differential seepage models are performed on the pores and fractures of different scales in the rock strata based on the hydrogeological data to obtain multiple sub-models; the multiple sub-models are then merged into the initial seepage model.

[0079] Based on the aforementioned hydrogeological data, the porosity and permeability evolution mechanism is introduced into the initial seepage model to obtain a dynamic seepage model;

[0080] Collocation points are generated from the dynamic seepage model, a composite objective loss function is constructed based on the physical residuals, and the physical information neural network is trained using the collocation points and the composite objective loss function to obtain the trained physical information neural network.

[0081] The trained physical information neural network is used to predict water storage.

[0082] Traditional equivalent continuous medium models oversimplify the connectivity of fracture networks, completely obliterating the discrete differentiation effect of large and medium-scale fractures and the local flow mechanism at the pore-fracture interface, and failing to truly reproduce the spatiotemporal distribution of deep fluids in multi-scale media.

[0083] The water storage prediction method for pore-fracture dual-medium rock strata provided in this embodiment performs differentiated seepage modeling based on fractures of different scales in hydrogeological data, and then merges the sub-models to obtain an initial seepage model. It takes into account the discrete differentiation effect of large and medium-scale fractures and the local channeling mechanism of pore-fracture interface, and truly restores the spatiotemporal distribution law of deep fluids in multi-scale media, thereby improving the accuracy of water storage prediction.

[0084] Based on this, this embodiment also considers the evolution mechanism of porosity and permeability. Using hydrogeological data as the data source and physical laws as the rules, the initial seepage model is supplemented with rules to obtain a dynamic seepage model. The water storage predicted by the dynamic seepage model fully considers the dynamic feedback mechanism of the flow field and permeability under the interaction of water and rock, so that the distortion of the predicted results of dynamic water storage and calculated injection pressure obtained from it is lower than the distortion threshold.

[0085] To accelerate prediction speed and ensure the feasibility of engineering implementation, this embodiment does not employ the finite element evolution analysis method with its potentially massive grid size (tens of millions of pixels). Instead, it constructs a physical information neural network that converges based on physical residuals. This transforms the physical laws used in finite element evolution analysis into residual calculation rules for the physical information neural network, thereby avoiding excessive computational load. By using the trained physical information neural network as a prediction solver, it is possible to predict water storage volume at a specified time with relatively low computational load.

[0086] Therefore, the water storage prediction method for pore-fracture dual-medium rock strata provided in this embodiment performs differentiated seepage modeling based on fractures of different scales in hydrogeological data, and then merges the various sub-models to obtain an initial seepage model. It takes into account the discrete differentiation effect of large and medium-scale fractures and the local crossflow mechanism of the pore-fracture interface, and realistically restores the spatiotemporal distribution law of deep fluids in multi-scale media, thereby improving the accuracy of water storage prediction. The water storage predicted based on the dynamic seepage model fully considers the dynamic feedback mechanism of flow field and permeability under water-rock interaction, so that the predicted results of dynamic water storage and calculated injection pressure obtained from it have low distortion. The physical laws used in finite element evolution analysis are transformed into residual calculation rules of physical information neural network, thereby avoiding excessive computation and realizing the prediction of water storage at a specified time.

[0087] Example 2:

[0088] This embodiment provides a method for predicting the water storage capacity of pore-fracture dual-medium rock formations. This embodiment is an optimization based on Embodiment 1 to improve the technical effect and refine the technical solution. For details not described in this embodiment, please refer to Embodiment 1.

[0089] As one embodiment, the process involves acquiring hydrogeological data of the rock strata, and then performing differentiated seepage modeling on the pores and fractures of different scales within the rock strata based on the hydrogeological data, resulting in multiple sub-models, including:

[0090] A fracture scale classification standard is established, which classifies all fractures into large-scale fractures, medium-scale fractures, and small-scale fractures in descending order of fracture length.

[0091] A three-dimensional seismic data volume of the rock strata is acquired. Fault planes are identified from the three-dimensional seismic data volume using variance volume and curvature volume techniques in attribute extraction. An ant tracking algorithm is used to filter noise from the fault planes through spatial multi-directional slicing, resulting in filtered fault planes. Based on the filtered fault planes and the fracture scale classification criteria, equivalent discrete surfaces matching large-scale fractures are generated in the model space, and these equivalent discrete surfaces are used as large-scale fracture sub-models. Based on the filtered fault planes and the fracture scale classification criteria, mesoscale fracture sub-models are generated in the model space. The filtered fault planes are binarized to obtain a binarized fault image with a resolution not lower than a resolution threshold. Different sized grids are used to extract noise from the binarized fault image. The computational space is divided; the range operation is performed on the maximum value of the dense region and the minimum value of the sparse region in the computational space to obtain fractal features; the fractal dimension of the fractal features is obtained, and a random number generator generates discrete fracture segments that satisfy the power-law distribution feature using the fractal dimension as input; the discrete fracture segments are dimensionally reduced using spherical coordinates and stereographic projection techniques to obtain dimension-reduced discrete fracture segments; the elbow rule is used to cluster the dimension-reduced discrete fracture segments to obtain clustered discrete fracture segments; Monte Carlo sampling is performed on the clustered discrete fracture segments using the Fisher probability distribution function to generate a discrete fracture network with random orientation and inclination angle with variance constraints, and the discrete fracture network is used as a small-scale fracture medium sub-model;

[0092] The hydrogeological data includes micro- and nano-scale Micro-CT scan images of the rock strata. Based on these images, characteristic parameters of the connecting pore throats are obtained. Matrix porosity, pore size distribution, and capillary pressure curves are extracted from these parameters. A characterization equation is constructed based on the matrix porosity, pore size distribution, and capillary pressure curves. A continuous three-dimensional volumetric unit is constructed based on the characterization equation. This continuous three-dimensional volumetric unit serves as a sub-model of the pores.

[0093] The process of fusing multiple sub-models into an initial seepage model includes:

[0094] A dual-flow medium spatial framework is established, and the large-scale fracture sub-model, the medium-scale fracture sub-model, and the small-scale fracture sub-model are coordinately embedded with the pore sub-model and placed into the same dual-flow medium spatial framework to obtain the initial flow model.

[0095] As one embodiment, obtaining the fractal dimension of the fractal feature, and using the fractal dimension as input, generating discrete crack segments that satisfy the power-law distribution characteristics by a random number generator includes:

[0096] The length of the discrete segment of the crack is obtained based on the fractal dimension, and the formula for obtaining the length of the discrete segment of the crack is as follows:

[0097] ,

[0098] In the formula, The length of the discrete line segment of the crack. These are random numbers that are uniformly distributed within the interval. This represents the maximum crack length. This is the minimum crack length. For fractal dimension, The width of the box is the dimension used in the box-counting formula. To cover the number of box layout experiments, For the counting sequence number, For the first The maximum number of cracks obtained from the statistics within the covered box in this experiment. For the first The minimum number of cracks obtained from the statistics inside the covered box in this experiment;

[0099] The attitude of the fracture is generated using a Fisher distribution model, which includes the fracture dip direction and the fracture dip angle. The spatial orientation probability density functions of the fracture dip direction and the fracture dip angle are as follows:

[0100] ,

[0101] In the formula, The value of the spatial orientation probability density function. It is prone to fractures. The angle of the fracture dip. The model distribution parameters characterize the degree of concentration of the distribution. It is a hyperbolic sine function;

[0102] This formula characterizes the rotationally symmetric distribution of fracture poles around the average azimuth, and the fracture dip angle. It follows a uniform distribution in the interval [0, 2π], therefore the core density function is determined only by the fracture inclination. Decide.

[0103] The discrete line segment of the crack is generated based on the spatial orientation distribution probability density function and the length of the discrete line segment of the crack.

[0104] As one embodiment, the step of introducing the porosity and permeability evolution mechanism into the initial seepage model based on the hydrogeological data to obtain a dynamic seepage model includes:

[0105] The porosity and permeability evolution mechanism includes the flow field equation of the fracture system, the flow field equation of the pore system, and the dynamic coupling relationship of pore-fracture crossflow;

[0106] The flow field equation for the fracture system is:

[0107] ,

[0108] In the formula, To comprehensively compress the elastic water storage coefficient, The fluid pressure in the fracture. For time, For Hamiltonian operators, For dynamic permeability tensor, The porosity is a dynamic fissure that is subject to real-time changes due to chemical precipitation and dissolution. For fluid dynamic viscosity, For fluid density, It is the acceleration due to gravity. For depth, The pore-fracture flow rate, For source and sink items;

[0109] The flow field equation for the porous system is:

[0110] ,

[0111] In the formula, To comprehensively compress the elastic water storage coefficient, For the fluid pressure of the matrix, For dynamic permeability tensor, The porosity of the matrix is ​​dynamically altered in real time by chemical precipitation and dissolution. A non-Darcy initiation pressure gradient threshold correction term to account for the weak permeability characteristics of the pore matrix;

[0112] The dynamic coupling relationship of pore-fracture flow is as follows:

[0113] ,

[0114] In the formula, The shape exchange factor is used to reflect the dynamic updating of the geometric and physical characteristics of the pore-fracture interface as it undergoes dissolution and expansion.

[0115] As one embodiment, when the rock stratum is the most common sandstone, the porosity and permeability evolution mechanism also includes a sandstone permeability evolution formula, which is as follows:

[0116] ,

[0117] In the formula, for The sandstone permeability attenuation coefficient at any given time. The initial sandstone permeability decay coefficient. The permeability attenuation coefficient of sandstone under water quality control. To adjust the sandstone permeability growth coefficient at different temperatures, To fully acidify the sandstone permeability growth factor, and These are the initial response evolution coefficients of the permeability, and These represent the rates of change over time. and These are the coefficients of the power-law term. These are the fitting coefficients;

[0118] The core of the sandstone was obtained, the core was crushed and ground and sieved to below 60 mesh to obtain fine powder. An equal volume of high TDS mine water was added to the fine powder of equal mass to obtain a sample, and the diameter of the sandstone particles was recorded at the same time.

[0119] The sample was placed in a high-pressure reactor for 100 days of soaking. The soaking process was divided into three phases: Phase 1, Phase 2, and Phase 3, with the three phases preferably having equal durations. The soaking temperature was 25°C in Phase 1 and 50°C in Phases 2 and 3. At the set time points during the soaking process, a rapid water quality analyzer, ICP-OES, and ion chromatograph were used to sample and analyze the solution of the soaked sample to obtain permeability evolution time-series data under single water quality conditions and permeability evolution time-series data under temperature control conditions.

[0120] The samples were classified according to different pore characteristics to obtain samples with different pore characteristics; the samples with different pore characteristics were placed in a core displacement experimental device to conduct long-term continuous water injection displacement experiments on the cores under different pressure and temperature conditions; after the core displacement experiment, Micro-CT was used to obtain the pore three-dimensional structure variation map of the samples with different pore characteristics, and the sandstone porosity was obtained based on the pore three-dimensional structure variation map.

[0121] Based on the sampling and testing results and the three-dimensional pore structure variation map, the theoretical mapping between sandstone permeability and sandstone porosity is constructed using the following formula:

[0122] ,

[0123] In the formula, For sandstone permeability, For sandstone porosity, The diameter of the sandstone grains;

[0124] The sandstone permeability-sandstone porosity theoretical mapping and the permeability evolution time series data under the single water quality action condition are fitted to obtain the sandstone permeability attenuation coefficient under water quality control. The initial response evolution coefficient of permeability and the time change rate are obtained by fitting the sandstone permeability-sandstone porosity theoretical mapping and the permeability evolution time series data under the temperature control condition.

[0125] The fitting coefficients and the power-law coefficients are obtained by fitting the sandstone permeability-sandstone porosity theoretical mapping, the permeability evolution time series data under the single water quality action condition, and the permeability evolution time series data under the temperature control condition.

[0126] As one embodiment, the step of generating collocation points from the dynamic seepage model, constructing a composite objective loss function based on physical residuals, and training a physical information neural network using the collocation points and the composite objective loss function to obtain a trained physical information neural network includes:

[0127] The Latin hypercube sampling algorithm is used to randomly distribute points throughout the dynamic seepage model.

[0128] The spatial coordinates and time series of the collocation points are input into the initially untrained physical information neural network to obtain the output fracture pressure, matrix pressure and target ion concentration.

[0129] The first and second derivative biases of the fracture pressure, matrix pressure, and target ion concentration are obtained using TensorFlow or PyTorch; TensorFlow is TensorFlow, and PyTorch is PyTorch.

[0130] Substituting the first-order derivative partial and the second-order derivative partial into the Darcy equation and the convection-diffusion equation, the physical residual is obtained;

[0131] The composite objective loss function is constructed based on the physical residual and the boundary head loss. The expression of the composite objective loss function is as follows:

[0132] ,

[0133] In the formula, For a composite objective loss function, , , and These are the loss weight hyperparameters that are dynamically and adaptively adjusted during training through sensitivity analysis or neural tangent kernels, to prevent any one loss component from dominating gradient updates. The mean square error between the network prediction output calculated at the point of reference and the field measured or seismic inversion hydrological data is given. For the physical residual terms of the flow field control equations, the physical information neural network uses its inherent automatic differentiation logic to accurately solve for the second-order partial derivatives of the network output terms with respect to the spatial coordinate vector and the first-order partial derivatives with respect to the time scalar. After substituting these into the control equations, the sum of squares of deviations from zero is calculated. This is the physical residual term of the solute transport reaction equation. The sum of the mean square errors of the predicted solution at the domain boundary and at the initial time point, which are the deviations from the set physical boundary conditions and initial conditions.

[0134] When training the physical information neural network, the Adam optimization algorithm is used to perform global coarse adjustment of the physical information neural network with a large step size.

[0135] If the number of training epochs in which the composite objective loss function value is lower than a preset threshold exceeds a threshold, or if the smoothness of the decrease in the composite objective loss function value reaches a smoothness threshold, the L-BFGS quasi-Newton method is used to perform a second-order fine local search on the physical information neural network until the composite objective loss function value is lower than the loss threshold, thus obtaining the trained physical information neural network.

[0136] As one embodiment, the step of using the trained physical information neural network to predict water storage includes:

[0137] By inputting a set time and three-dimensional coordinate matrix into the trained physical information neural network, the predicted fracture pressure, matrix pressure, and target ion concentration are obtained.

[0138] The water storage capacity, dynamic water-bearing porosity increment, and water pressure rise range within the effective influence radius of the three-dimensional coordinates were calculated using the water balance method and the integral rule.

[0139] The predicted water storage is obtained based on the water storage within the effective influence radius corresponding to multiple three-dimensional coordinates. The critical time of curtain flow or the area of ​​severe scaling is determined based on the dynamic water-bearing porosity increment and the water pressure rise range. A safety assessment report for optimizing single-well injection flow rate and injection pressure is generated based on the critical time of curtain flow or the area of ​​severe scaling.

[0140] The water storage prediction method for pore-fracture dual-medium rock formations provided in this embodiment improves the technical effect and refines the technical solution compared to Embodiment 1.

[0141] Example 3:

[0142] This embodiment provides a method for predicting the water storage capacity of pore-fracture dual-medium rock formations. This embodiment is an optimization based on Embodiment 2 to improve the technical effect and refine the technical solution. For details not described in this embodiment, please refer to Embodiment 2.

[0143] As one embodiment, the process involves acquiring hydrogeological data of the rock strata, and then performing differentiated seepage modeling on the pores and fractures of different scales in the rock strata based on the hydrogeological data to obtain multiple sub-models. Alternatively, another approach can be selected, which includes:

[0144] A fracture scale classification standard is established, which classifies all fractures from largest to smallest into large-scale fractures, medium-scale fractures, and small-scale fractures.

[0145] A three-dimensional seismic data volume of the rock strata is acquired. Fault planes are identified from the three-dimensional seismic data volume using variance and curvature volume techniques in attribute extraction. An ant tracking algorithm is used to filter noise in the fault planes through spatial multi-directional slicing, resulting in filtered fault planes. Based on the filtered fault planes and the fracture scale classification criteria, an equivalent discrete surface matching large-scale fractures is generated in the model space, and this equivalent discrete surface is used as a large-scale fracture sub-model. A medium-scale fracture sub-model is generated in the model space based on the filtered fault planes and the fracture scale classification criteria. The filtered fault planes are binarized to obtain a binarized fault image with a resolution not lower than a resolution threshold. A computational space is divided from the binarized fault image using grids of different sizes. The maximum value of dense regions and the maximum value of sparse regions in the computational space are then analyzed. The minimum value of the sparse region is used to perform range calculation to obtain fractal features; the fractal dimension of the fractal features is obtained, and a random number generator generates discrete fracture segments that satisfy the power-law distribution feature using the fractal dimension as input; the discrete fracture segments are dimensionally reduced using spherical coordinates and stereographic projection techniques to obtain dimension-reduced discrete fracture segments; the elbow rule is used to cluster the dimension-reduced discrete fracture segments to obtain clustered discrete fracture segments; the dominant strike and dominant dip angle of the fractures under a specific geological background are extracted as expected mean values, and Monte Carlo sampling is performed on the clustered discrete fracture segments using the Fisher probability distribution function; the generated fracture poles are aligned to the expected mean values ​​using a three-dimensional coordinate rotation matrix, thereby generating a random discrete fracture network sub-model that is distributed around the specified dominant strike and dip angle and has variance constraints;

[0146] The hydrogeological data includes micro- and nano-scale Micro-CT scan images of the rock strata. Based on these images, characteristic parameters of the connecting pore throats are obtained. Matrix porosity, pore size distribution, and capillary pressure curves are extracted from these parameters. A characterization equation is constructed based on the matrix porosity, pore size distribution, and capillary pressure curves. A continuous three-dimensional volumetric unit is constructed based on the characterization equation. This continuous three-dimensional volumetric unit serves as a sub-model of the pores.

[0147] The process of fusing multiple sub-models into an initial seepage model includes:

[0148] A dual-flow medium spatial framework is established, and the large-scale fracture sub-model, the medium-scale fracture sub-model, the small-scale fracture sub-model, and the pore sub-model are nested in the same dual-flow medium spatial framework to obtain the initial flow model.

[0149] The process of generating collocation points from the dynamic seepage model, constructing a composite objective loss function based on physical residuals, and training a physical information neural network using the collocation points and the composite objective loss function to obtain a trained physical information neural network also includes another scheme, which includes:

[0150] The Latin hypercube sampling algorithm is used to randomly distribute points throughout the dynamic seepage model.

[0151] The spatial coordinates and time series of the collocation points are input into the initially untrained physical information neural network to obtain the output fracture pressure, matrix pressure and target ion concentration.

[0152] The first and second derivative biases of the fracture pressure, matrix pressure, and target ion concentration output can be obtained using TensorFlow or PyTorch.

[0153] Substituting the first-order and second-order derivative partial values ​​into the Darcy equation and the convection-diffusion equation, the physical residual is obtained; the spatiotemporal collocation set is extracted from the dynamic seepage model. The physical residuals are specifically composed of the flow field physical residuals. With chemical field physical residual The composition and its calculation formulas are as follows:

[0154] Formula for calculating physical residuals of flow field: ;

[0155] Formula for calculating chemical field physical residuals: ;

[0156] In the formula, The total number of points. This is the Hamiltonian operator, also known as the spatial gradient operator. The pressure of network prediction, For dynamic permeability tensor, The ion concentration is predicted by the network. For Darcy velocity, To effectively disperse the tensor, The source and sink terms of the water-rock chemical reaction are represented. The physical information neural network obtains partial derivatives through an automatic differentiation mechanism, making the above residual terms approach zero, thereby implicitly satisfying the fluidization coupling partial differential equation.

[0157] The composite objective loss function is constructed based on the physical residual and the boundary head loss. The expression of the composite objective loss function is as follows:

[0158] ,

[0159] In the formula, For a composite objective loss function, , , and These are the loss weight hyperparameters that are dynamically and adaptively adjusted during training through sensitivity analysis or neural tangent nuclei. The mean square error between the network prediction output calculated at the point of reference and the field measured or seismic inversion hydrological data is given. The physical residuals of the flow field control equations are... This is the physical residual term of the solute transport reaction equation. The sum of the mean square errors of the predicted solution at the domain boundary and at the initial time point, which are the deviations from the set physical boundary conditions and initial conditions.

[0160] During model training, due to significant differences in the data magnitudes of the pressure field, concentration field, and boundary conditions, the Neural Tangent Kernel (NTK) theory was used to adjust the loss weight hyperparameters. Dynamic adaptive adjustment is performed. In each training iteration (Epoch), the weights... The dynamic update formula is defined as follows:

[0161] ;

[0162] Alternatively, based on the multi-task learning gradient matching criterion, the update formula is as follows:

[0163] ;

[0164] In the formula, This represents the computation of the trace of a matrix. To approximate the neural tangent kernel matrix of the observed data, For the first The neural tangent kernel matrix corresponding to each physical residual or conditional loss term; The data loss term is relative to the network parameters. gradient, For the first The average absolute value of the gradients of each loss term.

[0165] The above formula automatically amplifies the slow-converging physical field weights during backpropagation and suppresses loss terms that easily dominate the gradient, thereby ensuring absolute stability and high-precision convergence in solving the pore-fracture dual-medium water-chemistry strongly coupled equation.

[0166] When training the physical information neural network, the Adam optimization algorithm is used to perform global coarse adjustment of the physical information neural network with a large step size.

[0167] If the number of training epochs in which the composite objective loss function value is lower than a preset threshold exceeds a threshold, or if the smoothness of the decrease in the composite objective loss function value reaches a smoothness threshold, the L-BFGS quasi-Newton method is used to perform a second-order fine local search on the physical information neural network until the composite objective loss function value is lower than the loss threshold, thus obtaining the trained physical information neural network.

[0168] As one embodiment, the Fisher distribution model is used to generate the fracture attitude, which includes the fracture dip direction and the fracture dip angle. The spatial orientation probability density functions of the fracture dip direction and the fracture dip angle are as follows:

[0169] ,

[0170] In the formula, The value of the spatial orientation probability density function. It is prone to fractures. The dip angle represents the rotational symmetry of the fracture poles around the average azimuth. It follows a uniform distribution in [0, 2π], therefore the core density function is determined only by the deviation angle. Decide, The model distribution parameters characterize the degree of concentration of the distribution. It is a hyperbolic sine function;

[0171] Example 4:

[0172] refer to Figures 1 to 4 This embodiment provides a method for predicting the water storage capacity of pore-fracture dual-medium rock formations. This embodiment is an optimization based on embodiments two and three to improve the technical effect and refine the technical solution. For details not described in this embodiment, please refer to embodiments two and three.

[0173] As one embodiment, the water storage prediction method for pore-fracture dual-medium rock formations provided in this embodiment can also be described as follows:

[0174] We acquired on-site hydrogeological data and constructed a multi-scale pore-fracture dual-medium three-dimensional visualization geometric model based on it. This model divides the rock strata into large-scale fractures, medium-scale fractures, small-scale fractures, and pore matrix from macroscopic to microscopic levels.

[0175] Based on the aforementioned three-dimensional visualization geometric model, a dual-medium multi-physics field coupled control equation covering water-rock chemical interaction and nonlinear dynamic evolution of permeability is established. A dynamic permeability feedback mechanism is introduced into the equation, taking the injection water quality, temperature, and ion dissolution-precipitation volume changes as control factors to establish a nonlinear dynamic mathematical equation for porosity and permeability evolution. The dynamic changes of permeability over time are fed back to the flow field equations of the fracture system and pore matrix in real time.

[0176] A physical information neural network is constructed as a global proxy solver to transform the dual-medium multi-physics field coupling control equations into embedded physical constraints of the neural network. By defining a composite objective loss function that includes data loss terms, conditional loss terms, and physical residual loss terms, an optimization algorithm is used to iteratively update the network weights, thereby solving the corresponding fracture pressure, matrix pressure, and chemical ion concentrations in a continuous spatiotemporal domain without mesh.

[0177] The evolution data of seepage field and chemical field output by the physical information neural network are extracted and processed. Combined with the dynamic permeability tensor, the dynamic water storage of a single well is calculated and the evolution law of groundwater level rise and overflow is identified for safety strategy feedback.

[0178] As one embodiment, the multi-scale pore-fracture dual-medium three-dimensional spatial structure characterization and refined modeling based on multi-source data fusion includes: collecting geophysical logging data, 3D seismic datasets, and core scanning data of the target injection well or area; extracting geological modeling information for the target reservoir, such as the Liujiagou Formation sandstone; following the objective laws of the hydrogeological properties of multiple media, dividing the target media into four levels from macroscopic to microscopic: large-scale fractures, mesoscale fractures, small-scale fractures, and pore matrix; constructing equivalent discrete surfaces for the large-scale fracture media sub-model by combining 3D seismic attributes and manual interpretation methods; automatically picking and extracting the spatial distribution and connectivity of the mesoscale fracture media sub-model using an ant tracking algorithm; and constructing characterization equations for the pore matrix sub-model by using homogenization equivalence theory, combined with digital core CT scanning, mercury intrusion porosimetry, and core permeability testing to obtain matrix porosity, pore size distribution, and capillary pressure curves. Specifically, the small-scale fracture media sub-model uses a method combining fractal geometry and Fisher statistical distribution functions to generate a spatial three-dimensional discrete fracture network (Discrete Fracture). Networ, DFN; Finally, by integrating the above four sub-models, a three-dimensional visualization geometric model of the pore-fracture dual medium reflecting multi-scale spatial heterogeneity is established.

[0179] As one embodiment, a dual-medium multi-physics field coupled control equation is constructed, encompassing water-rock chemical interactions and the nonlinear dynamic evolution of permeability. This includes: defining the flow field control equation for fluids in the dual media and the transport control equation for chemical substances in the liquid phase based on the three-dimensional visualization geometric model; the flow field control equation uses a dual-pore model to describe rapid flow, Darcy or non-Darcy flow in the fracture system, and slow seepage in the porous matrix, respectively, and describes the fluid mass exchange between the two through an interfacial crossflow equation; a permeability dynamic feedback mechanism is introduced into the flow field and chemical field control equations, using injection water quality, temperature, and ion dissolution-precipitation volume changes as control factors to establish a nonlinear dynamic mathematical equation for porosity and permeability evolution; the dynamic changes in permeability over time are fed back in real-time to the flow field equations of the fracture system and the porous matrix, enabling adaptive updates of reservoir hydrogeological parameters over time.

[0180] For the first, second, and third sections of the Liujiagou Formation, considering the main controlling factors such as water quality, temperature, and acid corrosion in the water-rock coupling, a theoretical decision-making and control model for permeability during on-site construction was established. Based on the evolutionary fitting trends of the attenuation and growth coefficients of the above three controlling factors, the evolution formula for the permeability of the Liujiagou Formation sandstone can be derived as follows:

[0181] ,

[0182] In the formula, for The sandstone permeability attenuation coefficient at any given time. The initial sandstone permeability attenuation coefficient. The permeability attenuation coefficient of sandstone under water quality control. To adjust the sandstone permeability growth coefficient at different temperatures, To fully acidify the sandstone permeability growth factor, and These are the initial response evolution coefficients of the permeability, and These represent the rates of change over time. and These are the coefficients of the power-law term. These are the fitting coefficients;

[0183] This decision-making model can be used to guide on-site water injection projects. Under the premise of appropriate artificial control of water quality, temperature increase and sufficient acidification, the permeability of the Liujiagou sandstone layer can be changed, the single-well water injection flow rate and the dynamic water storage capacity of the Liujiagou sandstone can be increased, and long-term sustainable water storage potential can be obtained.

[0184] As one embodiment, a Physical Information Neural Network (PINN) is employed to achieve meshless and efficient numerical solutions to the co-equations of the flow field and chemical field, including:

[0185] A deep physical information neural network architecture is constructed as a global proxy solver to transform the dual-medium multi-physics field coupled control equations into embedded physical prior constraints of the neural network. A composite objective loss function of the physical information neural network is defined, which includes a data loss term approximating the observed data, a conditional loss term constraining the initial state of the system and the given head / flow boundary, and physical residual loss terms of each partial differential equation calculated by automatic differentiation and AD technique from the flow field control equations and chemical transport control equations.

[0186] In a continuous spatiotemporal domain, coordinate points are randomly generated. The spatial coordinates and time series of these coordinate points are used as input features of the physical information neural network. A gradient-based optimization algorithm is used to iteratively update the network weights and biases to minimize the composite objective loss function. This results in the direct output of high-precision predicted solutions for fracture pressure, matrix pressure, and chemical ion concentrations at each spatiotemporal node.

[0187] As one embodiment, the dynamic water storage accurate assessment and reinjection engineering safety strategy feedback includes: extracting and post-processing the spatiotemporal evolution data of the global seepage field and chemical field output by the physical information neural network trained in the previous step; combining the real-time updated reservoir water storage coefficient and dynamic permeability tensor; and using the spatial volume integral and water balance method to calculate the dynamic water storage of a single well under continuous water injection conditions; based on the chemical field characteristics such as ion concentration distribution and chemical reaction rate, identifying the fracture channel dissolution and expansion zone and matrix pore scaling and blockage zone in the long-term deep storage and reinjection process of high-salinity mine water, and outputting a safety assessment report for optimizing the single-well injection flow rate and injection pressure.

[0188] As one embodiment, to effectively handle the high-dimensional error surface stagnation phenomenon that may be caused by complex dual-medium models, the physical information neural network adopts a two-stage accelerated convergence training strategy: the first stage uses the Adam optimization algorithm with momentum mechanism to perform global large-step coarse adjustment to overcome local minima; when the decrease of the composite objective loss function in several consecutive epochs is lower than the preset threshold or tends to level off, the second stage is entered; the second stage uses L-BFGS, Broyden-Fletcher-Goldfarb-Shanno, and quasi-Newton methods to perform second-order fine local search until the loss function value converges to the range allowed by physical engineering tolerance, thereby realizing the smooth solution of the fast flow field in the fracture and the slow flow field in the matrix under meshless conditions.

[0189] As one embodiment, in Embodiment 2, after obtaining the orientation of the fracture, the fracture can be generated according to the dip direction, dip angle and fracture length, thus avoiding the increase in computational load caused by scanning a large number of large-scale fractures.

[0190] As one embodiment, the physical information neural network in Embodiment 2 adopts the following two-stage accelerated convergence training strategy for network training optimization: the first stage uses the Adam optimization algorithm with momentum mechanism to perform global coarse adjustment to overcome local minima; when the decrease of the composite objective loss function is lower than the preset threshold, the second stage is entered, and the L-BFGS quasi-Newton method is used for local search until convergence.

[0191] As one embodiment, the water storage prediction method for pore-fracture dual-medium rock formations provided in this embodiment can be deployed in a prediction system composed of multiple modules, which includes:

[0192] The data acquisition and preprocessing module is used to receive field geological and well logging data and generate multi-scale spatial pore-fracture discrete grids and collocation models.

[0193] The physical equation builder module is used to define a set of partial differential governing equations for dual media, including dynamic permeability feedback, fluid flow, and solute transport.

[0194] The meshless solution engine includes a physical information neural network processor, which is used to transform the partial differential control equations into neural network loss function constraints and perform automatic differential gradient calculation and two-stage joint optimization training.

[0195] The dynamic assessment and feedback module is used to calculate the dynamic limit water storage capacity based on the output results, and output water injection assessment and safety alarm parameters.

[0196] To further clarify the operational specifications and systematic process of this embodiment in actual engineering, this embodiment takes a typical high TDS mine water, i.e., a high-mineralization mine water deep storage and reinjection project in the Ordos Basin as an example, stripping away the absolute numerical parameters of specific geology, and focusing on the standardized implementation and safe operation steps of the method and system provided in this embodiment throughout the entire life cycle.

[0197] The initial phase involved importing basic scenario data and constructing a multi-scale digital twin. At the project's inception, system operators imported the target reservoir's 3D seismic interpretation data, multi-dimensional logging curves, and core sampling and analysis data—such as thick, weakly permeable sandstone—into the "Data Acquisition and Preprocessing Module." At the macroscopic level, the system automatically extracted large faults and structural fracture surfaces. At the meso-microscopic level, it ran an embedded ant tracking and improved box-counting dimension fractal algorithm to adaptively generate a discrete fracture network (DFN). Subsequently, the equivalent parameters of the microporous matrix were matched with the DFN using spatial coordinate nesting, thereby generating a meshless 3D geometric skeleton of the target water injection area and surrounding aquifers. This ultimately completed the construction of a digital twin model of the initial physical state of the deep reservoir.

[0198] Secondly, the reinjection operation condition configuration and fluid-chemical coupling dynamic feedback mechanism are activated. After completing the construction of the multi-medium model, the operator configures the specific boundary conditions of the project in the "Physical Equation Builder Module". Characteristic water quality component parameters of high-TDS mine water are input, such as the initial concentration of characteristic anions and cations, and the predicted injection water temperature. The single-well operation mode is then set, for example, using high-pressure constant flow injection or constant pressure injection boundary. Subsequently, the system's underlying layer automatically activates the permeability dynamic evolution equation module. In long-term simulation settings, the system not only calculates conventional pressure conduction and solute convection diffusion, but also tracks in real time the "dissolution and expansion" phenomenon caused by high-speed water flow in the near-wellbore zone, and the "precipitation and scaling" phenomenon caused by supersaturated ion aggregation in the far-field low-speed seepage zone. In this closed-loop process, porosity, permeability tensor, and shape factor characterizing cross-flow at the dual-medium interface are automatically and nonlinearly updated as physicochemical reactions occur, completely replacing the tedious and highly error-prone manual segmented parameter correction operations required in traditional software.

[0199] Then, the configuration of the PINN meshless proxy model and the efficient multi-field collaborative solution enter the complex numerical solution stage. This system does not need to go through the extremely time-consuming and error-prone unstructured mesh generation process of traditional finite element software. The engine directly uses the above-mentioned three-dimensional digital space and employs sampling algorithms such as Latin hypercube to randomly distribute the computation points in the entire domain, and densifies the computation point density around the injection well and in known fault zones. After successfully constructing the physical information neural network, the computation engine transforms the non-Darcy flow equation, solute transport equation and boundary conditions of the dual medium into automatic differential constraint loss within the network. During network training, the system background automatically executes a two-stage accelerated convergence strategy: initially, it uses optimization algorithms with momentum mechanisms, such as Adam, to quickly reduce data residuals and global physical residuals with large step sizes; later, it smoothly transitions to higher-order quasi-Newton algorithms, such as L-BFGS, for local fine-tuning optimization. This process completely avoids the fatal bottleneck of ill-conditioned stiffness matrix and algebraic iteration divergence caused by the huge difference in permeability between the two media. It not only greatly reduces the hardware overhead and waiting period required for computation, but also ensures the absolute smoothness and high fidelity of the multi-field strongly coupled solution process.

[0200] Finally, the system outputs a lifecycle dynamic reserve projection and disaster early warning scheme. As the neural network training fully converges, operators can use the "Dynamic Evaluation Feedback Module" in the system backend to retrieve and visualize three-dimensional underground flow field and chemical field spatiotemporal evolution slices for any target time point with a single click. At the system's underlying level, through a spatial volume integral algorithm combined with real-time evolving effective water storage space parameters, the system accurately outputs the ultimate dynamic water storage evolution curve of the injection well throughout its entire lifecycle. More importantly, based on the damage index of the pressure contour's advancing front and the permeability of the distal medium, the system will automatically identify the critical time point when the high-pressure reinjection water body approaches the boundary impermeable curtain or mining right boundary. If the system determines that there is a risk of curtain overflow or end bypass during the predetermined water injection cycle, it will immediately trigger a safety red alarm and assist engineers in outputting strategy optimization suggestions, such as: dynamically lowering the maximum water injection pressure threshold in the later stage of the project operation, planning the acidizing and unblocking construction process cycle in advance, or suggesting the addition of pressure relief pumping wells at specific far-field boundaries, thereby building a closed-loop green, efficient, and safe digital assessment and control flow for deep mine water storage reinjection projects.

[0201] As one embodiment, the water storage prediction method for pore-fracture dual-medium rock formations provided in this embodiment constructs a multi-scale refined three-dimensional discrete fracture-pore matrix geometric model by fractal geometry and clustering algorithms. It also innovatively embeds the dynamic permeability evolution equation driven by water-rock interaction into the coupled control equation of seepage and chemical solute transport. At the same time, this embodiment is committed to introducing the Physical Information Neural Network (PINN) to replace the traditional iterative solver, so as to achieve meshless, high-efficiency, and high-precision collaborative solution of seepage field, chemical field, and medium evolution under the long-term deep storage and reinjection of high-salinity mine water. This provides a scientific and reliable calculation tool for the dynamic water storage assessment and solute transport of single-well water injection.

[0202] The water storage prediction method for porous-fracture dual-medium rock formations provided in this embodiment breaks through the traditional "static parameter" formula and realizes a real cross-scale coupled response of permeability and flow field, significantly improving the accuracy of dynamic water storage calculation. In traditional numerical simulations of porous media flow fields, permeability is often simply regarded as a spatial constant distribution that does not change with time. However, the complex salinity of high-TDS mine water is prone to significant dissolution expansion or precipitation blockage effects under long-term high-pressure injection, leading to nonlinear deformation of the rock formation skeleton. This embodiment innovatively embeds the porosity-permeability evolution equation, which is controlled by both water quality and temperature, into the control core of seepage and chemical solute transport, constructing a complete closed-loop positive / negative feedback mechanism of "flow field update → chemical reaction → pore size variation → permeability feedback → flow field redistribution". Compared with traditional simulation methods that ignore this feedback, this embodiment greatly eliminates the static assumption error caused by long-term water injection, which is often >30%, giving extremely high calculation accuracy to dynamic water storage assessment. Furthermore, this embodiment accurately characterizes multi-scale heterogeneous structures, overcoming the oversimplification inherent in traditional equivalent continuous medium models. For extremely thick, weakly permeable sandstone reservoirs, this embodiment avoids the spatial detail loss caused by traditional single-scale or uniform equivalent continuous medium models, such as the permeability tensor. Utilizing CT digital cores, ant tracking algorithms, and a unique improved box-dimensional fractal theory, this embodiment achieves cross-scale geometric fusion from micrometer-level pores to macroscopic kilometer-level fault fractures. Based on this, the defined dual system of fracture conduit flow and matrix laminar flow fully considers the non-Darcy effect of weakly permeable strata, thus accurately capturing the rapid penetration phenomenon in dominant water-conducting channels and the slow ion diffusion and exchange within the matrix blocks. This embodiment innovatively introduces the Physical Information Neural Network (PINN), overcoming the divergence bottleneck of the strong asymmetric stiffness matrix in dual-medium solutions in traditional numerical methods. Traditional mesh solvers such as finite element methods or finite difference methods must generate massive amounts of extremely irregular and disparate three-dimensional mesh elements when processing fracture network models. Under conditions where the permeability difference between pores and fractures can be several orders of magnitude, the stiffness matrix generated by discretization is highly asymmetric and severely ill-conditioned. This makes traditional algebraic iterative algorithms prone to iterative stagnation, numerical oscillations, or even complete divergence. This embodiment introduces PINN technology, transforming the PDE governing equations into constraints of the loss function, thus achieving completely meshless computation. It not only perfectly avoids the geometric distortion errors caused by mesh generation, but also, with the support of an automatic differentiation engine, finds physical extrema through gradient descent and L-BFGS quasi-Newton higher-order optimization. While ensuring physical conservation laws, it greatly reduces memory consumption and computation time, transforming multi-field strongly coupled computation from difficult-to-converge to smooth and efficient. Figure 2 It clearly outlines the logical steps from data fusion modeling to PINN deep learning solution and security assessment. Figure 4The spatial temporal input, hidden layer forward propagation, physical partial differential residual constraint calculation (i.e., automatic differentiation), and closed-loop control process of multi-loss weighted optimization are clearly annotated.

[0203] The water storage prediction method for pore-fracture dual-medium rock strata provided in this embodiment can realize hierarchical modeling of multi-scale dual-medium structures within complex geometric domains. Extracting fractures from deeply buried geological bodies often faces the data challenge of extremely large scale spans, from millimeter-level microcracks to hundred-meter-level faults. This embodiment employs an original "top-down-bottom-up" bidirectional fusion database construction method. For large and medium-scale fractures longer than 5 meters, this embodiment first identifies fault planes based on the acquired 3D seismic data volume using variance and curvature volume techniques in attribute extraction. It then utilizes the industry-standard ant tracking algorithm to filter noise through spatial multi-directional slicing, directly generating a defined discrete fracture grid plane in the model space. For invisible micro-scale fractures in the 0.1~1 m range, their actual number is enormous and exhibits significant local clustering and self-similarity. This embodiment introduces an improved box-counting dimension algorithm for characterization: traditional box-counting algorithms typically perform a binary count (either 0 or 1) to determine the presence of features within the grid. However, due to the extremely high heterogeneity of fractures, this embodiment performs high-precision binarization processing on the target slice image. For example, each image is divided into a high-resolution pixel grid of 1024×1024, with the smallest unit being a single pixel. The computational space is then divided according to different grid sizes. In the improved formula, the range operation between the maximum value in dense regions and the minimum value in sparse regions at the same iteration scale is introduced. This innovative method greatly amplifies the complexity reflected by the local unevenness of the fracture network, and can more accurately capture the fractal characteristics of the real reservoir compared to conventional calculation methods. After obtaining the accurate fractal dimension, for example, a fractal dimension of 1.4013 for a certain segment indicates extremely low regularity of fracture trajectories and complex damage. A random number generator is used to generate fracture lengths that satisfy the power-law distribution characteristics. To enable these discrete line segments to be rationally assembled into a mesh in three-dimensional space, this embodiment further utilizes spherical coordinates and stereographic projection techniques for dimensionality reduction, combines the elbow rule for clustering, and finally generates random orientations and inclinations with variance constraints through Monte Carlo sampling of the Fisher probability distribution function, thereby constructing a highly realistic discrete fracture network. Simultaneously, for the porous matrix, characteristic parameters of the connecting pore throats are extracted based on micro-nano-scale Micro-CT scan images. Finally, the discrete line-surface network fractures and the continuous three-dimensional volumetric matrix are nested in coordinates to establish a dual-flow medium spatial framework.

[0204] This embodiment also realizes the establishment of nonlinear multi-field coupled PDE and the integration of dynamic mechanisms. On this framework, in order to break the constraints of static assumptions, this embodiment constructs a multivariate partial differential equation system with time series evolution as the core.

[0205] In the hydrodynamic evolution of groundwater, fluids exhibit completely different mechanisms in fractures and the matrix. As set in the equations of the examples, the fracture system is considered as a Darcy fast channel, while the matrix permeability is extremely low. The most critical coupling hub lies in the mass exchange term between them. The traditional classical two-porosity Warren-Root model assumes that it is merely a combination function of fluid viscosity, matrix permeability, and a constant shape factor. However, during long-term injection of high-TDS water rich in sulfate and carbonate concentrations, dissolution and precipitation constantly occur between the fluid and the fracture walls.

[0206] Calcite dissolution: ,

[0207] Gypsum precipitation: ,

[0208] Such reaction source-sink terms continuously alter the effective pore volume. This embodiment utilizes a dynamically modified Kozeny-Carman theory, defining porosity as a function that changes in real-time with increasing solution volume and decreasing precipitation volume. Furthermore, permeability and the shape factor describing the interfacial contact area are both endowed with time-dependent properties. This allows the dual-medium model to not only handle spatial differentiation but also realistically present the dynamic physical nature of damage / expansion over time.

[0209] This embodiment realizes the meshless optimization solution control flow of the Physical Information Neural Network (PINN). After completing the complex physical description of the PDE, how to perform efficient numerical computation is the core challenge of this invention. If the traditional finite element method is used, to match the scale jump between the extremely fine cracks and the huge matrix generated above, the model needs to be divided into tens of millions of unstructured meshes, resulting in an exponential explosion in computation time. Furthermore, the asymmetric stiffness matrix easily leads to non-convergence of Newton iterations. This embodiment makes deep use of the revolutionary mathematical tool, the Physical Information Neural Network (PINN):

[0210] First, a fully connected feedforward deep network with 7 hidden layers, each containing 128 neurons, is constructed. The input consists only of spatial coordinates and timestamps; the output channels are designed to represent fracture pressure, matrix pressure, and target ion concentration. The entire computational space does not require meshing; only tens of thousands of collocation points need to be distributed within its domain using Latin hypercube sampling.

[0211] Secondly, a composite loss function, LTotal, is defined, where LTotal represents the total loss. Unlike traditional black-box AI that relies solely on data label-supervised learning, PINN's brilliance lies in its "physical information constraints." Utilizing the inherent automatic differentiation mechanism of underlying computing frameworks such as TensorFlow or PyTorch, the network can directly calculate absolutely accurate first and second-order derivative partial quantities for the output terms. Substituting these predicted derivatives into the Darcy equation and convection-diffusion equation defined in step S2, if the equations do not hold, residual losses LPDE,seepage and LPDE,chem are generated. LPDE,seepage represents the residual loss of the seepage partial differential equation, and LPDE,chem represents the residual loss of the chemical solute transport partial differential equation. These, along with the boundary head loss LBC, are summed to form the target loss function. This essentially transforms the problem of solving partial differential equations into an optimization and dimensionality reduction problem in deep learning.

[0212] Finally, to avoid potential cancellation and learning stagnation in gradient calculations due to multi-physics fields, i.e., pressure differences at the MPa level and concentration differences at the mg / L level, the embodiment employs an adaptive loss weight adjustment strategy based on the Neural Tangent Kernel and NTK theory. Furthermore, during the initial 10,000 epochs of training, the Adam optimization algorithm is used for rapid optimization with large step sizes; once the gradient of the loss function becomes more gradual, it seamlessly switches to the L-BFGS second-order quasi-Newton optimization algorithm, which calculates the approximation of the full Hessian matrix, significantly accelerating the convergence speed while maintaining physical accuracy.

[0213] This embodiment can also achieve safety evaluation output. After network convergence, any time and three-dimensional coordinate matrix can be input, and the network can perform forward inference instantly without any step-by-step cumulative calculation. This embodiment will extract these inference results and use the water balance method and integral rule to calculate the dynamic water-bearing porosity increment and water pressure rise range within the effective influence radius in real time, directly determining the critical time of curtain flow or the area of ​​severe scaling.

[0214] Example 5:

[0215] This embodiment provides a computer device, including a processor and a memory connected to the processor. The memory stores a computer program. When the computer program is executed by the processor, it performs the steps of the water storage prediction method for pore-fracture dual-medium rock layers as provided in Embodiment 1 or 2.

[0216] The computer device may be a server or an electronic terminal, as one embodiment, see reference. Figure 5The computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores data acquired and generated in the method for predicting the water storage capacity of porous-fractured dual-medium rock formations. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the method for predicting the water storage capacity of porous-fractured dual-medium rock formations provided in Embodiment 1 or 2.

[0217] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0218] The computer device provided in this embodiment has the same technical effects as that in Embodiment 1 or 2, and will not be described again here.

[0219] Example 6:

[0220] This embodiment provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the water storage prediction method for pore-fracture dual-medium rock layers provided in Embodiment 1 or Embodiment 2.

[0221] The computer-readable storage medium provided in this embodiment has the same technical effects as that in Embodiment 1 or 2, and will not be described again here.

[0222] Example 7:

[0223] This embodiment provides a computer program product storing a computer program that, when executed by a processor, implements the steps of the water storage prediction method for pore-fracture dual-medium rock formations provided in Embodiment 1 or Embodiment 2. The computer program product provided in this embodiment can be transmitted, distributed, and downloaded via the Internet in the form of signals.

[0224] The computer program product provided in this embodiment has the same technical effects as that in Embodiment 1 or 2, and will not be described again here.

[0225] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0226] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0227] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0228] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0229] Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0230] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0231] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting water storage capacity in pore-fracture dual-medium rock formations, characterized in that, include: Hydrogeological data of the rock strata are obtained, and differential seepage modeling is performed on the pores and fractures of different scales in the rock strata based on the hydrogeological data to obtain multiple sub-models; Multiple sub-models are merged into an initial seepage model; Based on the aforementioned hydrogeological data, the porosity and permeability evolution mechanism is introduced into the initial seepage model to obtain a dynamic seepage model; Collocation points are generated from the dynamic seepage model, a composite objective loss function is constructed based on the physical residuals, and the physical information neural network is trained using the collocation points and the composite objective loss function to obtain the trained physical information neural network. The trained physical information neural network is used to predict water storage.

2. The method for predicting water storage capacity in pore-fracture dual-medium rock formations according to claim 1, characterized in that, The process involves acquiring hydrogeological data of the rock strata, and then using this data to perform differentiated seepage modeling on the pores and fractures of different scales within the rock strata, resulting in multiple sub-models, including: A fracture scale classification standard is established, which classifies all fractures from largest to smallest into large-scale fractures, medium-scale fractures, and small-scale fractures. A three-dimensional seismic data volume of the rock strata is acquired. Fault planes are identified from the three-dimensional seismic data volume using variance volume and curvature volume techniques in attribute extraction. An ant tracking algorithm is used to filter noise from the fault planes through spatial multi-directional slicing, resulting in filtered fault planes. Based on the filtered fault planes and the fracture scale classification criteria, equivalent discrete surfaces matching large-scale fractures are generated in the model space, and these equivalent discrete surfaces are used as large-scale fracture sub-models. Based on the filtered fault planes and the fracture scale classification criteria, mesoscale fracture sub-models are generated in the model space. The filtered fault planes are binarized to obtain a binarized fault image with a resolution not lower than a resolution threshold. Different sized grids are used to extract noise from the binarized fault image. The computational space is divided; the range operation is performed on the maximum value of the dense region and the minimum value of the sparse region in the computational space to obtain fractal features; the fractal dimension of the fractal features is obtained, and a random number generator generates discrete fracture segments that satisfy the power-law distribution feature using the fractal dimension as input; the discrete fracture segments are dimensionally reduced using spherical coordinates and stereographic projection techniques to obtain dimension-reduced discrete fracture segments; the elbow rule is used to cluster the dimension-reduced discrete fracture segments to obtain clustered discrete fracture segments; Monte Carlo sampling is performed on the clustered discrete fracture segments using the Fisher probability distribution function to generate a discrete fracture network with random orientation and inclination angle with variance constraints, and the discrete fracture network is used as a small-scale fracture medium sub-model; The hydrogeological data includes micro- and nano-scale Micro-CT scan images of the rock strata. Based on these images, characteristic parameters of the connecting pore throats are obtained. Matrix porosity, pore size distribution, and capillary pressure curves are extracted from these parameters. A characterization equation is constructed based on the matrix porosity, pore size distribution, and capillary pressure curves. A continuous three-dimensional volumetric unit is constructed based on the characterization equation. This continuous three-dimensional volumetric unit serves as a sub-model of the pores. The process of fusing multiple sub-models into an initial seepage model includes: A dual-flow medium spatial framework is established, and the large-scale fracture sub-model, the medium-scale fracture sub-model, the small-scale fracture sub-model, and the pore sub-model are nested in the same dual-flow medium spatial framework to obtain the initial flow model.

3. The method for predicting water storage capacity in pore-fracture dual-medium rock formations according to claim 2, characterized in that, The step of obtaining the fractal dimension of the fractal feature, and using the fractal dimension as input, generating discrete crack segments that satisfy the power-law distribution characteristics by a random number generator, includes: The length of the discrete segment of the crack is obtained based on the fractal dimension, and the formula for obtaining the length of the discrete segment of the crack is as follows: ， In the formula, The length of the discrete line segment of the crack. These are random numbers that are uniformly distributed within the interval. This represents the maximum crack length. This is the minimum crack length. For fractal dimension, The width of the box is the dimension used in the box-counting formula. To cover the number of box layout experiments, For the counting sequence number, For the first The maximum number of cracks obtained from the statistics within the covered box in this experiment. For the first The minimum number of cracks obtained from the statistics inside the covered box in this experiment; The attitude of the fracture is generated using a Fisher distribution model, which includes the fracture dip direction and the fracture dip angle. The spatial orientation probability density functions of the fracture dip direction and the fracture dip angle are as follows: ， In the formula, The value of the spatial orientation probability density function. It is prone to fractures. The angle of the fracture dip. The model distribution parameters characterize the degree of concentration of the distribution. It is a hyperbolic sine function; The discrete line segment of the crack is generated based on the spatial orientation distribution probability density function and the length of the discrete line segment of the crack.

4. The method for predicting water storage capacity in pore-fracture dual-medium rock formations according to claim 3, characterized in that, The process of introducing the porosity and permeability evolution mechanism into the initial seepage model based on the hydrogeological data to obtain a dynamic seepage model includes: The porosity and permeability evolution mechanism includes the flow field equation of the fracture system, the flow field equation of the pore system, and the dynamic coupling relationship of pore-fracture crossflow; The flow field equation for the fracture system is: ， In the formula, To comprehensively compress the elastic water storage coefficient, The fluid pressure in the fracture. For time, For Hamiltonian operators, For dynamic permeability tensor, The porosity is a dynamic fissure that is subject to real-time changes due to chemical precipitation and dissolution. For fluid dynamic viscosity, For fluid density, It is the acceleration due to gravity. For depth, The pore-fracture flow rate, For source and sink items; The flow field equation for the porous system is: ， In the formula, To comprehensively compress the elastic water storage coefficient, The fluid pressure of the matrix, For dynamic permeability tensor, The porosity of the matrix is ​​dynamically altered in real time by chemical precipitation and dissolution. A non-Darcy initiation pressure gradient threshold correction term to account for the weak permeability characteristics of the pore matrix; The dynamic coupling relationship of pore-fracture flow is as follows: ， In the formula, The shape exchange factor is used to reflect the dynamic updating of the geometric and physical characteristics of the pore-fracture interface as it undergoes dissolution and expansion.

5. The method for predicting water storage capacity in pore-fracture dual-medium rock formations according to claim 4, characterized in that, The rock stratum is sandstone, and the porosity and permeability evolution mechanism also includes a sandstone permeability evolution formula, which is as follows: ， In the formula, for The sandstone permeability attenuation coefficient at any given time. The initial sandstone permeability decay coefficient. The permeability attenuation coefficient of sandstone under water quality control. To adjust the sandstone permeability growth coefficient at different temperatures, To fully acidify the sandstone permeability growth factor, and These are the initial response evolution coefficients of the permeability, and These represent the rates of change over time. and These are the coefficients of the power-law term, These are the fitting coefficients; The core of the sandstone was obtained, the core was crushed and ground and sieved to below 60 mesh to obtain fine powder. An equal volume of high TDS mine water was added to the fine powder of equal mass to obtain a sample, and the diameter of the sandstone particles was recorded at the same time. The samples were placed in a high-pressure reactor for 100 days of soaking. The soaking process was divided into three phases: the first phase at 25°C and the second and third phases at 50°C. At the set time points during the soaking process, a rapid water quality analyzer, ICP-OES, and ion chromatograph were used to sample and test the solution of the soaked samples to obtain permeability evolution time-series data under single water quality conditions and permeability evolution time-series data under temperature control conditions. The samples are classified according to different pore characteristics to obtain samples with different pore characteristics; Samples with different pore characteristics were placed in a core displacement experimental device and subjected to long-term continuous water injection displacement experiments under different pressure and temperature conditions. After the core displacement experiment, Micro-CT was used to obtain the pore three-dimensional structure variation map of the samples with different pore characteristics, and the sandstone porosity was obtained based on the pore three-dimensional structure variation map. Based on the sampling and testing results and the three-dimensional pore structure variation map, the theoretical mapping between sandstone permeability and sandstone porosity is constructed using the following formula: ， In the formula, For sandstone permeability, For sandstone porosity, The diameter of the sandstone grains; The sandstone permeability-sandstone porosity theoretical mapping and the permeability evolution time series data under the single water quality action condition are fitted to obtain the sandstone permeability attenuation coefficient under water quality control. The initial response evolution coefficient of permeability and the time change rate are obtained by fitting the sandstone permeability-sandstone porosity theoretical mapping and the permeability evolution time series data under the temperature control condition. The fitting coefficients and the power-law coefficients are obtained by fitting the sandstone permeability-sandstone porosity theoretical mapping, the permeability evolution time series data under the single water quality action condition, and the permeability evolution time series data under the temperature control condition.

6. The method for predicting water storage capacity in pore-fracture dual-medium rock formations according to claim 5, characterized in that, The process of generating collocation points from the dynamic seepage model, constructing a composite objective loss function based on physical residuals, and training a physical information neural network using the collocation points and the composite objective loss function to obtain a trained physical information neural network includes: The Latin hypercube sampling algorithm is used to randomly distribute points throughout the dynamic seepage model. The spatial coordinates and time series of the collocation points are input into the initially untrained physical information neural network to obtain the output fracture pressure, matrix pressure and target ion concentration; The first and second derivative biases of the fracture pressure, matrix pressure, and target ion concentration in the output can be obtained using TensorFlow or PyTorch. Substituting the first-order derivative partial and the second-order derivative partial into the Darcy equation and the convection-diffusion equation, the physical residual is obtained; The composite objective loss function is constructed based on the physical residual and the boundary head loss. The expression of the composite objective loss function is as follows: ， In the formula, For a composite objective loss function, , , and These are the loss weight hyperparameters that are dynamically and adaptively adjusted during training through sensitivity analysis or neural tangent nuclei. The mean square error between the network prediction output calculated at the point of reference and the field measured or seismic inversion hydrological data is given. The physical residuals of the flow field control equations are... This is the physical residual term of the solute transport reaction equation. The sum of the mean square errors of the predicted solution at the domain boundary and at the initial time point, which are the deviations from the set physical boundary conditions and initial conditions. When training the physical information neural network, the Adam optimization algorithm is used to perform global coarse adjustment of the physical information neural network with a large step size. If the number of training epochs in which the composite objective loss function value is lower than a preset threshold exceeds a threshold, or if the smoothness of the decrease in the composite objective loss function value reaches a smoothness threshold, the L-BFGS quasi-Newton method is used to perform a second-order fine local search on the physical information neural network until the composite objective loss function value is lower than the loss threshold, thus obtaining the trained physical information neural network.

7. The method for predicting water storage capacity in pore-fracture dual-medium rock formations according to claim 6, characterized in that, The method of using the trained physical information neural network to predict water storage includes: By inputting a set time and three-dimensional coordinate matrix into the trained physical information neural network, the predicted fracture pressure, matrix pressure, and target ion concentration are obtained. The water storage capacity, dynamic water-bearing porosity increment, and water pressure rise range within the effective influence radius of the three-dimensional coordinates were calculated using the water balance method and the integral rule. The predicted water storage is obtained based on the water storage within the effective influence radius corresponding to multiple three-dimensional coordinates. The critical time of curtain flow or the area of ​​severe scaling is determined based on the dynamic water-bearing porosity increment and the water pressure rise range. A safety assessment report for optimizing single-well injection flow rate and injection pressure is generated based on the critical time of curtain flow or the area of ​​severe scaling.

8. A computer device, characterized in that, The device includes a processor and a memory connected to the processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the steps of the water storage prediction method for pore-fracture dual-medium rock formations as described in any one of claims 1 to 7 are performed.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method for predicting the water storage capacity of pore-fracture dual-medium rock formations as described in any one of claims 1 to 7.

10. A computer program product comprising a computer program / instructions, characterized in that, When executed by a processor, the computer program / instruction implements the steps of the method for predicting the water storage capacity of pore-fracture dual-medium rock formations as described in any one of claims 1 to 7.

Images