Mechanism-data collaborative driven carbon dioxide mineralization process inversion method and system

Abstract

The present invention discloses a mechanism-data collaboratively driven carbon dioxide mineralization process inversion method and system, the method includes: determining the mineralization storage target area, based on an ultrasonic phased array monitoring system, an acoustic emission monitoring system, and an optical fiber micro-nano multi-parameter co-mining monitoring system, using the real-time monitoring data signal of the mineralization process as input data; using the physical field models of the mineralization process such as the fracture field, seepage field, and temperature field of the mineralization process as output data. And using simulation software to obtain the physical field model of the mineralization process to further supplement the input and output data; using an LSTM neural network to construct a fusion inversion model of the mineralization process, by inputting the real-time mineralization process reaction measured monitoring data, and fusion inversion output to obtain the physical fields of the fracture field, seepage field, and temperature field in the mineralization process; the method constructs and predicts the physical field model of the complex mineralization storage process through dual-drive of mechanism and data, which is conducive to improving the fusion inversion capability of the mineralization storage process.

Description

Technical Field

[0001] The present invention relates to the technical field of carbon dioxide mineralization, and in particular to a mechanism-data collaboratively driven carbon dioxide mineralization process inversion method and system. Background Art

[0002] Carbon dioxide storage, utilization and storage (CCUS) is a key technology to achieve deep emission reduction in the industrial field. Carbon dioxide geological storage is one of the footholds for emission reduction and carbon sink increase, and its core issue is the long-term safety and reliability of storage. Monitoring the physical state of carbon dioxide based on monitoring wells is the most common effect monitoring method at the geological storage project site, and how to monitor the storage process in real time has become a difficult problem.

[0003] With the continuous changes in the new generation of artificial intelligence technology, basic research paradigms such as mechanism models and data-driven are used in many fields of geophysics. In 2019, an important review was published in «Nature», proposing a combination of physical models and data-driven machine learning, which has important inspiration for the future development of artificial intelligence geophysics. CN117440404A proposed a data- and knowledge-driven intelligent and precise construction method for spectrum maps, which realized the construction of target frequency spectrum maps when target frequency spectrum data was missing; CN117371813A aimed at the real-time monitoring and prediction of tunnel gas, modeled the tunnel gas concentration from the perspectives of physical model prediction and big data prediction, and proposed a tunnel gas concentration prediction method based on model and big data dual drive; CN117372994A provided a collaborative perception method for intelligent connected vehicles based on model and data dual drive, and adopted a multi-point cloud robust registration model based on model and data dual drive to establish a vehicle collaborative detection and prediction solution for multiple intelligent connected vehicles. The machine learning research paradigm driven by both data and mechanism models has good applicability in multiple fields of geophysics. In this way, the physical field model of complex mineralization and storage processes can be constructed and predicted through the collaborative drive of mechanism and data. This method is conducive to improving the fusion inversion capability of the mineralization and storage process. Summary of the invention

[0004] In response to the problems and needs raised above, this solution proposes a mechanism-data collaboratively driven carbon dioxide mineralization process inversion method and system. Due to the adoption of the following technical features, it can achieve the above technical objectives and bring about many other technical effects.

[0005] One object of the present invention is to propose a mechanism-data collaboratively driven carbon dioxide mineralization process inversion method, comprising the following steps:

[0006] S10: Determine the target area for mineralization storage, and use the ultrasonic phased array monitoring system, acoustic emission monitoring system and optical fiber micro-nano multi-parameter co-mining monitoring system to obtain the stress-strain signal, acoustic emission signal, ultrasonic velocity signal and temperature signal obtained by monitoring the mineralization process, and use them as input data;

[0007] S20: Based on the acoustic emission inversion, the spatial position and volume of the fractures generated by the mineralization reaction are obtained, the permeability is obtained based on the fracture damage of the mineralized rock mass, and the physical field of the mineralization process including the fracture field, seepage field and temperature field of the mineralization process is established and used as the output data;

[0008] S30: Combined with the actual conditions of the mineralization storage site, the simulation software is used to obtain the physical field of the mineralization process and supplement the input data and output data; the input data and output data are combined using the K-fold cross-validation method to form a data set driven by data and mechanism, and the training set and validation set are divided respectively;

[0009] S40: Use LSTM neural network to build a fusion inversion model of mineralization process, train the fusion inversion model of mineralization process through training set, and finally use the real-time mineralization process reaction measured monitoring data to obtain the physical fields of fracture field, seepage field and temperature field in the mineralization process through fusion inversion;

[0010] S50: Improve the loss function in the error back-propagation algorithm, add the residual of the physical control equation to the loss function, construct a loss function driven by data and mechanism to measure the difference between the inversion output and the measured output, use the stochastic gradient descent algorithm to calculate the gradient of each weight parameter to update the weight matrix and bias vector of the network, use the validation set to verify the trained mineralization process fusion inversion model, repeat the above steps until the predetermined number of iterations is reached or the loss function value is less than the predetermined threshold, and stop the calculation step.

[0011] In addition, the mechanism-data collaboratively driven carbon dioxide mineralization process inversion method according to the present invention may also have the following technical features:

[0012] In step S30, simulation software is used to obtain the physical field of the mineralization process, wherein the physical field includes flow field, seepage field and temperature field. The control equations and coupling equations involved in the physical field are as follows:

[0013] 1) Loss strain equation

[0014] Assuming that rock damage is elastic-brittle damage, the constitutive equation expressing stress in terms of strain is:

[0015] ;

[0016] In the formula, For prestressing; D is the elasticity matrix, For strain;

[0017] The damage is described by equivalent strain and determined by the exponential strain evolution relationship. When the strain threshold is reached, the damage begins to evolve and is defined as:

[0018] ;

[0019] In the formula, is the tensile strain of the elastic limit; is the ultimate tensile strain, , is the ultimate strain coefficient; is the equivalent strain; is the strain softening parameter;

[0020] The stress and strain of the solid are determined by adding the balanced differential equation, and after considering the inertial effect, it is expressed as:

[0021] ;

[0022] In the formula, f is the force per unit volume, N / m 3 ; is the stress vector, MPa; is the solid density, kg / m 3 , x is the solid displacement, m;

[0023] 2) Flow equation

[0024] The flow of carbon dioxide hydrate through the porous medium of the storage rock mass is described by the porous medium flow model. The flow equation of the fluid between the fractures in the rock mass medium is:

[0025] ;

[0026] In the formula, is the fluid density; is the porosity; u is the fluid velocity; p is the fluid pressure; I is the unit tensor; F is the volume force; t is the shear force; μ is the dynamic viscosity of the fluid; k is the permeability; β is the isothermal compressibility coefficient, Q m is the source term, g is the gravitational acceleration;

[0027] 3) Seepage equation

[0028] During fluid fracturing, permeability and damage have a nonlinear growth relationship. The permeability increases exponentially with the change of damage. The relationship based on damage to describe permeability is:

[0029] ;

[0030] In the formula, is the coefficient of porosity affected by damage; The permeability at the initial state; d is the loss variable;

[0031] 4) Temperature equation

[0032] In order to simulate the thermal environment around mineralized storage, the temperature field is introduced to reflect the real thermal impact fracture process of mineralized rock mass. By coupling the temperature field and the flow field, the non-isothermal flow equation is:

[0033] ;

[0034] Where T is temperature; is the heat capacity at constant pressure, J / (mol·K); is the thermal conductivity; Q is the heat source power.

[0035] In one example of the present invention, in step S40, a LSTM neural network is used to construct a fusion inversion model of the mineralization process, including the following steps:

[0036] S41: Initialize network parameters, update the forget gate, input gate, output gate and cell state through the initial weight matrix and bias vector;

[0037] S42: input data serialization, extract training set from data set, input training samples into time step and wait for training;

[0038] S43: Calculate the learning entities of the forget gate, input gate, cell state update and output gate in the LSTM neural network algorithm model in sequence, obtain memory information and then obtain the network algorithm fitting value.

[0039] In one example of the present invention, in step S43, the expression of the network algorithm fitting value is as follows:

[0040] ;

[0041] In the formula, For short-term memory information; are the neuron weight matrix and bias vector of the corresponding layers respectively; is the activation function, where the activation function includes the sigmoid activation function and the tanh activation function:

[0042] ;

[0043] .

[0044] In one example of the present invention, in step S43, the learning bodies of the forget gate, input gate, cell state update and output gate in the LSTM neural network algorithm model are calculated in sequence, including the following steps:

[0045] In step S43, the learning bodies of the forget gate, input gate, cell state update and output gate in the LSTM neural network algorithm model are calculated in sequence, including the following steps:

[0046] S431: Calculate the forget gate to determine the lost information of the cell state. The activation value of the forget gate is calculated by the following formula:

[0047] ;

[0048] In the formula, s is the sigmoid activation function, and is the weight matrix and bias vector of the forget gate, is the hidden state at the previous time step, is the input of the current time step;

[0049] S432: Calculate the input gate to determine the information in the stored cell state. The activation value of the input gate is calculated by the following formula:

[0050] ;

[0051] ;

[0052] In the formula, s is the sigmoid activation function, and is the weight matrix of the input gate; and is the bias vector of the input gate; and Mining the current time step information and the previous time step information respectively;

[0053] S433: Update the cell state. Update the cell memory information according to the output of the forget gate and the candidate cell state determined by the input gate. The activation value of the cell state is calculated by the following formula:

[0054] ；

[0055] In the formula, Training sample information for the previous time step;

[0056] S434: Calculate the output gate to determine the hidden state of the next time step , the activation value of the output gate is calculated as follows:

[0057] ;

[0058] ;

[0059] In the formula, and is the weight matrix of the output gate and the weight of the bias vector.

[0060] In one example of the present invention, in step S50, the residual of the physical control equation is added to the loss function, and a loss function driven by data and mechanism synergistically is constructed to measure the difference between the inversion output and the measured output, including the following steps:

[0061] S501: Calculate the residual of the control equation and construct the mechanism loss function as follows:

[0062] ;

[0063] In the formula, is the training sample index; The total number of sample points within the study area; is the differential term of the algorithm network fitting value;

[0064] S502: Calculate the data-driven residual and construct the data loss function as follows:

[0065] ;

[0066] In the formula, is the training sample index; The total number of sample points within the study area; is the measured value of the algorithm network; is the algorithm network fitting value;

[0067] S503: Calculate the sum of the residuals of the control equation and the data-driven, and construct a loss function that integrates the mechanism-driven and data-driven, as follows:

[0068] ;

[0069] In the formula, The mean square error loss calculated for the residuals of the control equations; is the mean square error loss calculated between the measured value and the algorithm network fitting value; α and β are the weight coefficients of the mechanism-driven term and the data-driven term, respectively.

[0070] In one example of the present invention, in step S501, the differential term of the mechanism loss function includes the equilibrium differential equation, the inter-fracture fluid flow equation, and the temperature equation, and the residuals of each term are summed up to obtain the expression:

[0071] ;

[0072] In the formula, f is the force per unit volume, N / m 3 ; is the stress vector, MPa; is the solid density, kg / m 3 , x is the solid displacement, m; is the fluid density, kg / m 3 ; is the porosity; u is the fluid velocity, m / s; p is the fluid pressure, MPa; F is the volume force, N; t is the shear force, MPa; k is the permeability; β is the isothermal compressibility coefficient, Q m is the source term; T is the temperature; is the heat capacity at constant pressure, J / (mol·K); is the thermal conductivity; Q is the heat source power.

[0073] In one example of the present invention, in step S50, the gradient of each weight parameter is calculated using a stochastic gradient descent algorithm to update the weight matrix and bias vector of the network, including the following steps:

[0074] Each iteration calculates the algorithm network parameter gradient, updates the parameters in the opposite direction of the gradient, and obtains the global optimal solution:

[0075] ;

[0076] In the formula, is the gradient of each parameter of the loss function; η is the update rate of the control step size; I t+1 and I t are the algorithm network parameters to be updated at the t+1th and tth iteration steps respectively.

[0077] Another object of the present invention is to propose a mechanism-data collaboratively driven carbon dioxide mineralization process inversion system, comprising:

[0078] An input data module is configured to determine a mineralized storage target area, and to obtain stress-strain signals, acoustic emission signals, ultrasonic velocity signals, and temperature signals obtained by monitoring the mineralization process using an ultrasonic phased array monitoring system, an acoustic emission monitoring system, and an optical fiber micro-nano multi-parameter co-mining monitoring system, and use these signals as input data;

[0079] An output data module is configured to obtain the spatial position and volume of the fractures generated by the mineralization reaction based on acoustic emission inversion, obtain the permeability based on the fracture damage of the mineralized rock mass, establish the physical field of the mineralization process regarding the fracture field, seepage field, and temperature field of the mineralization process, and use it as output data;

[0080] The data partitioning module is configured to combine the actual conditions of the mineralization storage site, use simulation software to obtain the physical field of the mineralization process and supplement the input data and output data; use the K-fold cross-validation method to combine the input data and output data to form a data set driven by the data and mechanism, and divide the training set and the validation set respectively;

[0081] The physical field calculation module is configured to use the LSTM neural network to build a fusion inversion model of the mineralization process, train the fusion inversion model of the mineralization process through a training set, and finally use the real-time mineralization process reaction measured monitoring data to obtain the physical fields of the fracture field, seepage field, and temperature field in the mineralization process through fusion inversion;

[0082] The loss calculation module is configured to improve the loss function in the error back propagation algorithm, add the residual of the physical control equation to the loss function, construct a loss function driven by data and mechanism to measure the difference between the inversion output and the measured output, use the stochastic gradient descent algorithm to calculate the gradient of each weight parameter to update the weight matrix and bias vector of the network, use the verification set to verify the trained mineralization process fusion inversion model, repeat the above steps until the predetermined number of iterations is reached or the loss function value is less than the predetermined threshold, and stop the calculation step.

[0083] In one example of the present invention, the loss calculation module includes:

[0084] The control equation residual calculation unit is configured to calculate the control equation residual and construct the mechanism loss function as follows:

[0085] ;

[0086] In the formula, is the training sample index; The total number of sample points within the study area; is the differential term of the algorithm network fitting value;

[0087] The data-driven residual calculation unit is configured to calculate the data-driven residual and construct the data loss function as follows:

[0088] ;

[0089] In the formula, is the training sample index; The total number of sample points within the study area; is the measured value of the algorithm network; is the algorithm network fitting value;

[0090] The residual sum calculation unit is configured to calculate the sum of the residuals of the control equation and the data-driven, and to construct a loss function integrated by the mechanism drive and the data drive, as follows:

[0091] ;

[0092] In the formula, The mean square error loss calculated for the residuals of the control equations; is the mean square error loss calculated between the measured value and the algorithm network fitting value; α and β are the weight coefficients of the mechanism-driven term and the data-driven term, respectively.

[0093] The beneficial effects of the present invention are as follows:

[0094] This method combines the actual conditions of the mineralization storage site and uses simulation software to obtain the physical field model of the mineralization process to further supplement the data set. The obtained data set is more in line with the actual mineralization situation, effectively solving the problem of insufficient mineralization monitoring data and significantly improving the accuracy of mineralization fusion inversion.

[0095] This method uses the LSTM neural network to construct a fusion inversion model of the mineralization process. The neural network can capture the state changes of the carbon dioxide mineralization process at long intervals in the time series, better handle long-distance dependencies, improve the prediction level of mineralization effects, and provide reliable data support for mineralization evaluation.

[0096] This method establishes the relationship between the reaction data of the monitored mineralization process and the fused inversion output physical field based on the LSTM neural network, extracts different types of monitoring data (such as images, text data, etc.) for fusion, and can extract the nonlinear characteristics of the mineralization process, more comprehensively characterizing the complexity of the inversion physical field.

[0097] This method improves the loss function in the error back propagation algorithm, adds the residual of the physical control equation into the loss function, and constructs a loss function driven by both data and mechanism to measure the difference between the inversion output and the measured output, thereby improving the accuracy and robustness of the model and expanding the application scope of the model to adapt to the changing monitoring environment.

[0098] The best embodiment for carrying out the present invention will be described in more detail below with reference to the accompanying drawings so that the features and advantages of the present invention can be easily understood. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings of the embodiments of the present invention, wherein the drawings are only used to illustrate some embodiments of the present invention, but not to limit all embodiments of the present invention thereto.

[0100] Figure 1 A logic diagram of a carbon dioxide mineralization process inversion method driven by mechanism-data collaboration according to an embodiment of the present invention;

[0101] Figure 2 A schematic diagram of the relationship between the mechanism control equation and the coupling equation according to an embodiment of the present invention;

[0102] Figure 3 Schematic diagram of the structure of the LSTM neural network algorithm model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0103] In order to make the purpose, technical solution and advantages of the technical solution of the present invention clearer, the technical solution of the embodiment of the present invention will be clearly and completely described in conjunction with the drawings of specific embodiments of the present invention. The same figure marks in the drawings represent the same parts. It should be noted that the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0104] Unless otherwise defined, the technical terms or scientific terms used herein shall have the usual meanings understood by persons of ordinary skill in the field to which the present invention belongs. The words "first", "second" and similar words used in the patent application specification and claims of the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "one" or "one" do not necessarily indicate a quantity limitation. Words such as "include" or "comprise" mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Words such as "connect" or "connected" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0105] According to a first aspect of the present invention, a mechanism-data collaboratively driven carbon dioxide mineralization process inversion method is provided. Figure 1 and Figure 2 As shown, the following steps are included:

[0106] S10: Determine the target area for mineralization storage, and use the ultrasonic phased array monitoring system, acoustic emission monitoring system and optical fiber micro-nano multi-parameter co-mining monitoring system to obtain the stress-strain signal, acoustic emission signal, ultrasonic velocity signal and temperature signal obtained by monitoring the mineralization process, and use them as input data;

[0107] S20: Based on the acoustic emission inversion, the spatial position and volume of the fractures generated by the mineralization reaction are obtained, the permeability is obtained based on the fracture damage of the mineralized rock mass, and the physical field of the mineralization process including the fracture field, seepage field and temperature field of the mineralization process is established and used as the output data;

[0108] S30: Combined with the actual conditions of the mineralization storage site, the simulation software is used to obtain the physical field of the mineralization process and supplement the input data and output data; the input data and output data are combined using the K-fold cross-validation method to form a data set driven by data and mechanism, and the training set and validation set are divided respectively;

[0109] S40: Use LSTM neural network to build a fusion inversion model of mineralization process, train the fusion inversion model of mineralization process through training set, and finally use the real-time mineralization process reaction measured monitoring data to obtain the physical fields of fracture field, seepage field and temperature field in the mineralization process through fusion inversion;

[0110] S50: Improve the loss function in the error back-propagation algorithm, add the residual of the physical control equation to the loss function, construct a loss function driven by data and mechanism to measure the difference between the inversion output and the measured output, use the stochastic gradient descent algorithm to calculate the gradient of each weight parameter to update the weight matrix and bias vector of the network, use the validation set to verify the trained mineralization process fusion inversion model, repeat the above steps until the predetermined number of iterations is reached or the loss function value is less than the predetermined threshold, and stop the calculation step.

[0111] This method combines the actual conditions of the mineralization storage site, uses simulation software to obtain the physical field model of the mineralization process to further supplement the data set, and obtains a data set that is more in line with the actual mineralization situation, effectively solves the problem of insufficient mineralization monitoring data, and significantly improves the accuracy of mineralization fusion inversion. This method uses the LSTM neural network to construct a fusion inversion model of the mineralization process. The neural network can capture the state changes of the carbon dioxide mineralization process with a long interval in the time series, can better handle long-distance dependencies, improve the prediction level of the mineralization effect, and provide reliable data support for mineralization evaluation. Based on the LSTM neural network, this method establishes the relationship between the monitoring mineralization process reaction data and the fusion inversion output physical field, extracts different types of monitoring data (such as images, text data, etc.) for fusion, and can extract the nonlinear characteristics of the mineralization process, which more comprehensively characterizes the complexity of the inversion physical field. This method improves the loss function in the error back propagation algorithm, adds the residual of the physical control equation to the loss function, and constructs a loss function driven by data and mechanism to measure the difference between the inversion output and the measured output, which improves the accuracy and robustness of the model and expands the application scope of the model to adapt to the changing monitoring environment.

[0112] In one example of the present invention, in step S30, simulation software is used to obtain the physical field of the mineralization process, wherein the physical field includes a flow field, a seepage field, and a temperature field. The control equations and coupling equations involved in the physical field are as follows:

[0113] 1) Loss strain equation

[0114] Assuming that rock damage is elastic-brittle damage, the constitutive equation expressing stress in terms of strain is:

[0115] ;

[0116] In the formula, For prestressing; D is the elasticity matrix, For strain.

[0117] The damage is described by equivalent strain and determined by the exponential strain evolution relationship. When the strain threshold is reached, the damage begins to evolve and is defined as:

[0118] ;

[0119] In the formula, is the tensile strain of the elastic limit; is the ultimate tensile strain, , is the ultimate strain coefficient; is the equivalent strain; is the strain softening parameter.

[0120] The stress and strain of the solid are determined by adding the balanced differential equation, and after considering the inertial effect, it is expressed as:

[0121] ;

[0122] In the formula, f is the force per unit volume, N / m 3 ; is the stress vector, MPa; is the solid density, kg / m 3 , x is the solid displacement, m.

[0123] 2) Flow equation

[0124] The flow of carbon dioxide hydrate through the porous medium of the storage rock mass is described by the porous medium flow model. The flow equation of the fluid between the fractures in the rock mass medium is:

[0125] ;

[0126] In the formula, is the fluid density; is the porosity; u is the fluid velocity; p is the fluid pressure; I is the unit tensor; F is the volume force; t is the shear force; μ is the dynamic viscosity of the fluid; k is the permeability; β is the isothermal compressibility coefficient, Q m is the source term and g is the gravitational acceleration.

[0127] 3) Seepage equation

[0128] During fluid fracturing, permeability and damage have a nonlinear growth relationship. The permeability increases exponentially with the change of damage. The relationship based on damage to describe permeability is:

[0129] ;

[0130] In the formula, is the coefficient of porosity affected by damage; The permeability at the initial state; d is the loss variable.

[0131] 4) Temperature equation

[0132] In order to simulate the thermal environment around mineralized storage, the temperature field is introduced to reflect the real thermal impact fracture process of mineralized rock. By coupling the temperature field and the flow field, the non-isothermal flow equation is:

[0133] ;

[0134] Where T is temperature; is the heat capacity at constant pressure, J / (mol·K); is the thermal conductivity; Q is the heat source power.

[0135] In one example of the present invention, in step S40, as Figure 3 As shown in the figure, the LSTM neural network is used to construct a fusion inversion model of the mineralization process, which includes the following steps:

[0136] S41: Initialize network parameters, update the forget gate, input gate, output gate and cell state through the initial weight matrix and bias vector;

[0137] S42: input data serialization, extract training set from data set, input training samples into time step and wait for training;

[0138] S43: Calculate the learning entities of the forget gate, input gate, cell state update and output gate in the LSTM neural network algorithm model in sequence, obtain memory information and then obtain the network algorithm fitting value.

[0139] In one example of the present invention, in step S43, the expression of the network algorithm fitting value is as follows:

[0140] ;

[0141] In the formula, For short-term memory information; are the neuron weight matrix and bias vector of the corresponding layers respectively; is the activation function, including sigmoid activation function and tanh activation function;

[0142] ;

[0143] .

[0144] In one example of the present invention, in step S43, the learning bodies of the forget gate, input gate, cell state update and output gate in the LSTM neural network algorithm model are calculated in sequence, including the following steps:

[0145] In step S43, the learning bodies of the forget gate, input gate, cell state update and output gate in the LSTM neural network algorithm model are calculated in sequence, including the following steps:

[0146] S431: Calculate the forget gate to determine the lost information of the cell state. The activation value of the forget gate is calculated by the following formula:

[0147] ;

[0148] In the formula, s is the sigmoid activation function, and is the weight matrix and bias vector of the forget gate, is the hidden state at the previous time step, is the input of the current time step;

[0149] S432: Calculate the input gate to determine the information in the stored cell state. The activation value of the input gate is calculated by the following formula:

[0150] ;

[0151] ;

[0152] In the formula, s is the sigmoid activation function, and is the weight matrix of the input gate; and is the bias vector of the input gate; and Mining the current time step information and the previous time step information respectively;

[0153] S433: Update the cell state. Update the cell memory information according to the output of the forget gate and the candidate cell state determined by the input gate. The activation value of the cell state is calculated by the following formula:

[0154] ；

[0155] In the formula, Training sample information for the previous time step;

[0156] S434: Calculate the output gate to determine the hidden state of the next time step , the activation value of the output gate is calculated as follows:

[0157] ;

[0158] ;

[0159] In the formula, and is the weight matrix of the output gate and the weight of the bias vector.

[0160] In one example of the present invention, in step S50, the residual of the physical control equation is added to the loss function, and a loss function driven by data and mechanism synergistically is constructed to measure the difference between the inversion output and the measured output, including the following steps:

[0161] S501: Calculate the residual of the control equation and construct the mechanism loss function as follows:

[0162] ;

[0163] In the formula, is the training sample index; The total number of sample points within the study area; is the differential term of the algorithm network fitting value;

[0164] S502: Calculate the data-driven residual and construct the data loss function as follows:

[0165] ;

[0166] In the formula, is the training sample index; The total number of sample points within the study area; is the measured value of the algorithm network; is the algorithm network fitting value;

[0167] S503: Calculate the sum of the residuals of the control equation and the data-driven, and construct a loss function that integrates the mechanism-driven and data-driven, as follows:

[0168] ;

[0169] In the formula, The mean square error loss calculated for the residuals of the control equations; is the mean square error loss calculated between the measured value and the algorithm network fitting value; α and β are the weight coefficients of the mechanism-driven term and the data-driven term, respectively.

[0170] In one example of the present invention, in step S501, the differential term of the mechanism loss function includes the equilibrium differential equation, the inter-fracture fluid flow equation, and the temperature equation, and the residuals of each term are summed up to obtain the expression:

[0171] ;

[0172] In the formula, f is the force per unit volume, N / m 3 ; is the stress vector, MPa; is the solid density, kg / m 3 , x is the solid displacement, m; is the fluid density, kg / m 3 ; is the porosity; u is the fluid velocity, m / s; p is the fluid pressure, MPa; F is the volume force, N; tis the shear force, MPa; k is the permeability; β is the isothermal compressibility coefficient, Q m is the source term; T is the temperature; is the heat capacity at constant pressure, J / (mol·K); is the thermal conductivity; Q is the heat source power.

[0173] In one example of the present invention, in step S50, the gradient of each weight parameter is calculated using a stochastic gradient descent algorithm to update the weight matrix and bias vector of the network, including the following steps:

[0174] Each iteration calculates the algorithm network parameter gradient, updates the parameters in the opposite direction of the gradient, and obtains the global optimal solution:

[0175] ;

[0176] In the formula, is the gradient of each parameter of the loss function; η is the update rate of the control step size; I t+1 and I t are the algorithm network parameters to be updated at the t+1th and tth iteration steps, including the weight matrix and bias vector in each learning body, , as well as C t and ℎ 𝑡 .

[0177] According to the second aspect of the present invention, a mechanism-data collaboratively driven carbon dioxide mineralization process inversion system comprises:

[0178] An input data module is configured to determine a mineralized storage target area, and to obtain stress-strain signals, acoustic emission signals, ultrasonic velocity signals, and temperature signals obtained by monitoring the mineralization process using an ultrasonic phased array monitoring system, an acoustic emission monitoring system, and an optical fiber micro-nano multi-parameter co-mining monitoring system, and use these signals as input data;

[0179] An output data module is configured to obtain the spatial position and volume of the fractures generated by the mineralization reaction based on acoustic emission inversion, obtain the permeability based on the fracture damage of the mineralized rock mass, establish the physical field of the mineralization process regarding the fracture field, seepage field, and temperature field of the mineralization process, and use it as output data;

[0180] The data partitioning module is configured to combine the actual conditions of the mineralization storage site, use simulation software to obtain the physical field of the mineralization process and supplement the input data and output data; use the K-fold cross-validation method to combine the input data and output data to form a data set driven by the data and mechanism, and divide the training set and the validation set respectively;

[0181] The physical field calculation module is configured to use the LSTM neural network to build a fusion inversion model of the mineralization process, train the fusion inversion model of the mineralization process through a training set, and finally use the real-time mineralization process reaction measured monitoring data to obtain the physical fields of the fracture field, seepage field, and temperature field in the mineralization process through fusion inversion;

[0182] The loss calculation module is configured to improve the loss function in the error back propagation algorithm, add the residual of the physical control equation to the loss function, construct a loss function driven by data and mechanism to measure the difference between the inversion output and the measured output, use the stochastic gradient descent algorithm to calculate the gradient of each weight parameter to update the weight matrix and bias vector of the network, use the verification set to verify the trained mineralization process fusion inversion model, repeat the above steps until the predetermined number of iterations is reached or the loss function value is less than the predetermined threshold, and stop the calculation step.

[0183] The system combines the actual conditions of the mineralization storage site and uses simulation software to obtain the physical field model of the mineralization process to further supplement the data set, so that the data set is more in line with the actual mineralization situation, effectively solves the problem of insufficient mineralization monitoring data, and significantly improves the accuracy of mineralization fusion inversion. The system uses the LSTM neural network to construct a fusion inversion model of the mineralization process. The neural network can capture the state changes of the carbon dioxide mineralization process with a long interval in the time series, can better handle long-distance dependencies, improve the prediction level of the mineralization effect, and provide reliable data support for mineralization evaluation. Based on the LSTM neural network, the system establishes the relationship between the reaction data of the monitoring mineralization process and the fusion inversion output physical field, extracts different types of monitoring data (such as images, text data, etc.) for fusion, and can extract the nonlinear characteristics of the mineralization process, which more comprehensively characterizes the complexity of the inversion physical field. The system improves the loss function in the error back propagation algorithm, adds the residual of the physical control equation to the loss function, and constructs a loss function driven by data and mechanism to measure the difference between the inversion output and the measured output, which improves the accuracy and robustness of the model and expands the application scope of the model to adapt to the changing monitoring environment.

[0184] In one example of the present invention, the physical field calculation module includes:

[0185] A parameter initialization unit, configured to initialize network parameters, and update a forget gate, an input gate, an output gate, and a cell state through an initial weight matrix and a bias vector;

[0186] A data training unit is configured to serialize input data, extract training sets from the data set, and input training samples into time steps to wait for training;

[0187] The network algorithm fitting unit is configured to sequentially calculate the learning bodies of the forget gate, input gate, cell state update and output gate in the LSTM neural network algorithm model, obtain memory information and then obtain the network algorithm fitting value.

[0188] In one example of the present invention, the expression of the network algorithm fitting value is as follows:

[0189] ;

[0190] In the formula, For short-term memory information; are the neuron weight matrix and bias vector of the corresponding layers respectively; is the activation function, including sigmoid activation function and tanh activation function;

[0191] ;

[0192] .

[0193] In one example of the present invention, the loss calculation module includes:

[0194] The control equation residual calculation unit is configured to calculate the control equation residual and construct the mechanism loss function as follows:

[0195] ;

[0196] In the formula, is the training sample index; The total number of sample points within the study area; is the differential term of the algorithm network fitting value;

[0197] The data-driven residual calculation unit is configured to calculate the data-driven residual and construct the data loss function as follows:

[0198] ;

[0199] In the formula, is the training sample index; The total number of sample points within the study area; is the measured value of the algorithm network; is the algorithm network fitting value;

[0200] The residual sum calculation unit is configured to calculate the sum of the residuals of the control equation and the data-driven, and to construct a loss function integrated by the mechanism drive and the data drive, as follows:

[0201] ;

[0202] In the formula, The mean square error loss calculated for the residuals of the control equations; is the mean square error loss calculated between the measured value and the algorithm network fitting value; α and β are the weight coefficients of the mechanism-driven term and the data-driven term, respectively.

[0203] The exemplary implementation scheme of the mechanism-data collaboratively driven carbon dioxide mineralization process inversion method and system proposed in the present invention is described in detail above with reference to the preferred embodiments. However, those skilled in the art will appreciate that, without departing from the concept of the present invention, various modifications and variations can be made to the above-mentioned specific embodiments, and various technical features and structures proposed in the present invention can be combined in various ways without exceeding the protection scope of the present invention, which is determined by the appended claims.

Claims

1. A mechanism-data collaboratively driven carbon dioxide mineralization process inversion method, characterized in that: The steps include: S10: Determine the target area for mineralization storage, and use the ultrasonic phased array monitoring system, acoustic emission monitoring system and optical fiber micro-nano multi-parameter co-mining monitoring system to obtain the stress-strain signal, acoustic emission signal, ultrasonic velocity signal and temperature signal obtained by monitoring the mineralization process, and use them as input data; S20: Based on the acoustic emission inversion, the spatial position and volume of the fractures generated by the mineralization reaction are obtained, the permeability is obtained based on the fracture damage of the mineralized rock mass, and the physical field of the mineralization process including the fracture field, seepage field and temperature field of the mineralization process is established and used as the output data; S30: Combined with the actual conditions of the mineralization storage site, the simulation software is used to obtain the physical field of the mineralization process and supplement the input data and output data; the input data and output data are combined using the K-fold cross-validation method to form a data set driven by data and mechanism, and the training set and validation set are divided respectively; S40: Use LSTM neural network to build a fusion inversion model of mineralization process, train the fusion inversion model of mineralization process through training set, and finally use the real-time mineralization process reaction measured monitoring data to obtain the physical fields of fracture field, seepage field and temperature field in the mineralization process through fusion inversion; S50: Improve the loss function in the error back-propagation algorithm, add the residual of the physical control equation to the loss function, construct a loss function driven by data and mechanism to measure the difference between the inversion output and the measured output, use the stochastic gradient descent algorithm to calculate the gradient of each weight parameter to update the weight matrix and bias vector of the network, use the validation set to verify the trained mineralization process fusion inversion model, repeat the above steps until the predetermined number of iterations is reached or the loss function value is less than the predetermined threshold, and stop the calculation step.

2. The mechanism-data collaboratively driven carbon dioxide mineralization process inversion method according to claim 1 is characterized in that: In step S30, simulation software is used to obtain the physical field of the mineralization process, wherein the physical field includes flow field, seepage field and temperature field. The control equations and coupling equations involved in the physical field are as follows: Loss strain equation: Assuming that rock damage is elastic-brittle damage, the constitutive equation expressing stress in terms of strain is: ； In the formula, For prestressing; D is the elasticity matrix, For strain; The damage is described by equivalent strain and determined by the exponential strain evolution relationship; when the strain threshold is reached, the damage begins to evolve and is defined as: ； In the formula, is the tensile strain of the elastic limit; is the ultimate tensile strain, , is the ultimate strain coefficient; is the equivalent strain; is the strain softening parameter; The stress and strain of the solid are determined by adding the balanced differential equation, and after considering the inertial effect, it is expressed as: ； In the formula, f is the force per unit volume; is the stress vector; is the solid density, x is the solid displacement; flow equation: The flow of carbon dioxide hydrate through the porous medium of the storage rock mass is described by the porous medium flow model. The flow equation of the fluid between the fractures in the rock mass medium is: ； In the formula, is the fluid density; is the porosity; u is the fluid velocity; p is the fluid pressure; I is the unit tensor; F is the volume force; τ is the shear force; μ is the dynamic viscosity of the fluid; k is the permeability; β is the isothermal compressibility coefficient, Q m is the source term, g is the gravitational acceleration; Seepage equation: During fluid fracturing, permeability and damage have a nonlinear growth relationship. The permeability increases exponentially with the change of damage. The relationship based on damage to describe permeability is: ； In the formula, is the coefficient of porosity affected by damage; The permeability at the initial state; d is the loss variable; temperature equation: In order to simulate the thermal environment around mineralized storage, the temperature field is introduced to reflect the real thermal impact fracture process of mineralized rock mass. By coupling the temperature field and the flow field, the non-isothermal flow equation is: ； Where T is temperature; is the heat capacity at constant pressure; is the thermal conductivity; Q is the heat source power.

3. The mechanism-data collaboratively driven carbon dioxide mineralization process inversion method according to claim 1 is characterized in that: In step S40, a LSTM neural network is used to construct a fusion inversion model of the mineralization process, including the following steps: S41: Initialize network parameters, update the forget gate, input gate, output gate and cell state through the initial weight matrix and bias vector; S42: input data serialization, extract training set from data set, input training samples into time step and wait for training; S43: Calculate the learning entities of the forget gate, input gate, cell state update and output gate in the LSTM neural network algorithm model in sequence, obtain memory information and then obtain the network algorithm fitting value.

4. The mechanism-data collaboratively driven carbon dioxide mineralization process inversion method according to claim 3 is characterized in that: In step S43, the expression of the network algorithm fitting value is as follows: ； In the formula, For short-term memory information; are the neuron weight matrix and bias vector of the corresponding layers respectively; is the activation function, where the activation function includes the sigmoid activation function and the tanh activation function: ； 。 5. The mechanism-data collaboratively driven carbon dioxide mineralization process inversion method according to claim 3 is characterized in that: In step S43, the learning bodies of the forget gate, input gate, cell state update and output gate in the LSTM neural network algorithm model are calculated in sequence, including the following steps: S431: Calculate the forget gate to determine the lost information of the cell state. The activation value of the forget gate is calculated by the following formula: ； In the formula, σ is the sigmoid activation function, and is the weight matrix and bias vector of the forget gate, is the hidden state at the previous time step, is the input of the current time step; S432: Calculate the input gate to determine the information in the stored cell state. The activation value of the input gate is calculated by the following formula: ； ； In the formula, σ is the sigmoid activation function, and is the weight matrix of the input gate; and is the bias vector of the input gate; and Mining the current time step information and the previous time step information respectively; S433: Update the cell state. Update the cell memory information according to the output of the forget gate and the candidate cell state determined by the input gate. The activation value of the cell state is calculated by the following formula: ； In the formula, Training sample information for the previous time step; S434: Calculate the output gate to determine the hidden state of the next time step , the activation value of the output gate is calculated as follows: ； ； In the formula, and is the weight matrix of the output gate and the weight of the bias vector.

6. The mechanism-data collaboratively driven carbon dioxide mineralization process inversion method according to claim 1 is characterized in that: In step S50, the residual of the physical control equation is added to the loss function, and a loss function driven by data and mechanism is constructed to measure the difference between the inversion output and the measured output, including the following steps: S501: Calculate the residual of the control equation and construct the mechanism loss function as follows: ； In the formula, is the training sample index; The total number of sample points within the study area; is the differential term of the algorithm network fitting value; S502: Calculate the data-driven residual and construct the data loss function as follows: ； In the formula, is the training sample index; The total number of sample points within the study area; is the measured value of the algorithm network; is the algorithm network fitting value; S503: Calculate the sum of the residuals of the control equation and the data-driven, and construct a loss function that integrates the mechanism-driven and data-driven, as follows: ； In the formula, The mean square error loss calculated for the residuals of the control equations; is the mean square error loss calculated between the measured value and the algorithm network fitting value; α and β are the weight coefficients of the mechanism-driven term and the data-driven term, respectively.

7. The mechanism-data collaboratively driven carbon dioxide mineralization process inversion method according to claim 6 is characterized in that: In step S501, the differential term of the mechanism loss function includes the equilibrium differential equation, the inter-fracture fluid flow equation, and the temperature equation, and the residuals of each term are summed up to obtain the expression: ； In the formula, f is the force per unit volume; is the stress vector; is the solid density, x is the solid displacement; is the fluid density; is the porosity; u is the fluid velocity; p is the fluid pressure; F is the volume force, N; τ is the shear force; k is the permeability; β is the isothermal compressibility coefficient, Q m is the source term; T is the temperature; is the heat capacity at constant pressure; is the thermal conductivity; Q is the heat source power.

8. The mechanism-data collaboratively driven carbon dioxide mineralization process inversion method according to claim 1 is characterized in that: In step S50, the gradient of each weight parameter is calculated using the stochastic gradient descent algorithm to update the weight matrix and bias vector of the network, including the following steps: Each iteration calculates the algorithm network parameter gradient, updates the parameters in the opposite direction of the gradient, and obtains the global optimal solution: ； In the formula, is the gradient of each parameter of the loss function; η is the update rate of the control step size; Θ t+1 and Θ t are the algorithm network parameters to be updated at the t+1th and tth iteration steps respectively.

9. A mechanism-data collaboratively driven carbon dioxide mineralization process inversion system, characterized in that: include: An input data module is configured to determine a mineralized storage target area, and to obtain stress-strain signals, acoustic emission signals, ultrasonic velocity signals, and temperature signals obtained by monitoring the mineralization process using an ultrasonic phased array monitoring system, an acoustic emission monitoring system, and an optical fiber micro-nano multi-parameter co-mining monitoring system, and use these signals as input data; An output data module is configured to obtain the spatial position and volume of the fractures generated by the mineralization reaction based on acoustic emission inversion, obtain the permeability based on the fracture damage of the mineralized rock mass, establish the physical field of the mineralization process regarding the fracture field, seepage field, and temperature field of the mineralization process, and use it as output data; The data partitioning module is configured to combine the actual conditions of the mineralization storage site, use simulation software to obtain the physical field of the mineralization process and supplement the input data and output data; use the K-fold cross-validation method to combine the input data and output data to form a data set driven by the data and mechanism, and divide the training set and the validation set respectively; The physical field calculation module is configured to use the LSTM neural network to build a fusion inversion model of the mineralization process, train the fusion inversion model of the mineralization process through a training set, and finally use the real-time mineralization process reaction measured monitoring data to obtain the physical fields of the fracture field, seepage field, and temperature field in the mineralization process through fusion inversion; The loss calculation module is configured to improve the loss function in the error back propagation algorithm, add the residual of the physical control equation to the loss function, construct a loss function driven by data and mechanism to measure the difference between the inversion output and the measured output, use the stochastic gradient descent algorithm to calculate the gradient of each weight parameter to update the weight matrix and bias vector of the network, use the verification set to verify the trained mineralization process fusion inversion model, repeat the above steps until the predetermined number of iterations is reached or the loss function value is less than the predetermined threshold, and stop the calculation step.

10. The mechanism-data collaboratively driven carbon dioxide mineralization process inversion system according to claim 9, characterized in that: The loss calculation module comprises: The control equation residual calculation unit is configured to calculate the control equation residual and construct the mechanism loss function as follows: ； In the formula, is the training sample index; The total number of sample points within the study area; is the differential term of the algorithm network fitting value; The data-driven residual calculation unit is configured to calculate the data-driven residual and construct the data loss function as follows: ； In the formula, is the training sample index; The total number of sample points within the study area; is the measured value of the algorithm network; is the algorithm network fitting value; The residual sum calculation unit is configured to calculate the sum of the residuals of the control equation and the data-driven, and to construct a loss function integrated by the mechanism drive and the data drive, as follows: ； In the formula, The mean square error loss calculated for the residuals of the control equations; is the mean square error loss calculated between the measured value and the algorithm network fitting value; α and β are the weight coefficients of the mechanism-driven term and the data-driven term, respectively.

Images