Physical information neural network numerical simulation method suitable for fractured shale gas reservoir

Abstract

The invention relates to the technical field of reservoir engineering, and discloses a physical information neural network numerical simulation method suitable for a fractured shale gas reservoir, and the method comprises the steps: obtaining the basic data of the fractured shale gas reservoir; based on the dual-medium model and the embedded discrete model, establishing a matrix-microfracture-hydraulic fracture coupling model suitable for the fractured shale gas reservoir; establishing a neural network model which takes a full-connection network as a trunk network and two full-connection networks as branch networks, adding a network structure in the neural network model and setting related hyper-parameters; performing pre-training by taking the initial pressure as label data, and initializing a neural network model; and training all time steps by adopting a dynamic time step method based on the initialized neural network model to obtain pressure fields of all time steps of the shale gas reservoir. According to the method, the pressure change in the elastic exploitation process of the fractured shale gas reservoir can be solved through the physical information neural network under the condition of no label data, and the method has relatively high precision and efficiency.

Description

Technical Field

[0001] The present application relates to the field of oil reservoir engineering technology, and in particular to a physical information neural network numerical simulation method applicable to fractured shale gas reservoirs. Background Art

[0002] As global energy demand continues to grow and conventional oil and gas resources are gradually depleted, shale gas, as an important component of unconventional natural gas resources, has become a key area of ​​energy development. Fractured shale gas reservoirs exhibit significant multi-scale and multi-physics coupling characteristics due to their unique reservoir structure (such as the interaction between complex natural fracture networks and artificial fracturing induced fractures) and heterogeneity. In the development of such gas reservoirs, numerical simulation technology is the core tool for optimizing fracturing design, predicting production capacity and formulating development strategies. However, traditional numerical simulation methods (such as finite element method, finite volume method, etc.) have the following problems: the geometric complexity and multi-scale characteristics of the fracture network lead to difficulties in grid division and high computational costs; they are highly dependent on the accurate description of reservoir geological parameters and fracture distribution, which can easily lead to large errors when actual data is insufficient; at the same time, traditional data-driven models (such as pure machine learning methods) often ignore physical law constraints such as fluid mechanics constitutive equations, resulting in prediction results deviating from basic physical principles such as mass conservation and Darcy's law, limiting the reliability of engineering guidance.

[0003] In recent years, Physics-Informed Neural Networks (PINN) have provided new ideas for complex reservoir simulation because they can embed control equations into neural network training and take into account the advantages of both data-driven and physical law constraints. By constructing proxy models through physically meaningful neural network models, the problems of large computational complexity and long time consumption of traditional reservoir numerical simulators can be solved. However, for fractured shale gas reservoirs, since their control equations have strong nonlinear characteristics and their coefficients change significantly over time, the current PINN model is difficult to be directly applied to the numerical simulation of shale gas reservoirs. Existing proxy models often need to generate label data through traditional numerical solution methods, and these models can only predict the pressure distribution information at a certain moment. Summary of the invention

[0004] The purpose of this application is to provide a physical information neural network numerical simulation method suitable for fractured shale gas reservoirs, so as to solve the pressure changes during the elastic exploitation of fractured shale gas reservoirs through physical information neural networks without labeled data, and with high accuracy and efficiency.

[0005] In order to achieve the above purpose, the technical solutions adopted are as follows: The present application provides a physical information neural network numerical simulation method applicable to fractured shale gas reservoirs, the method comprising: Obtain basic data of fractured shale gas reservoirs; Based on the basic data of fractured shale gas reservoirs, a dual medium model is used to describe bedrock and microfractures, and an embedded discrete model is used to describe hydraulic fractures. The two are combined to establish a matrix-microfracture-hydraulic fracture coupling model suitable for fractured shale gas reservoirs, which is used for the training and prediction of neural network models. Establish a neural network model with a fully connected network as the main network and two fully connected networks as branch networks, and add a network structure to the neural network model and set relevant hyperparameters; wherein the network structure includes adjacent position anchoring, adaptive activation function, cross-layer connection and gated update; Use the initial pressure as label data for pre-training and initialize the neural network model; Based on the initialized neural network model, the dynamic time step method is used to perform training for all time steps. During the training process of each time step, the pressure predicted by the neural network model is substituted into the overall loss function, and back propagation is performed to continuously optimize the network parameters until the overall loss is reduced to a specified range, thereby obtaining the pressure field of the shale gas reservoir at all time steps; wherein the overall loss function includes the physical loss of the control equation based on finite volume discretization and the expert experience loss.

[0006] Furthermore, the basic data of the fractured shale gas reservoir include the size of the reservoir and the fracturing transformation area, the initial gas reservoir pressure, the bedrock porosity, the shale gas reservoir temperature, the Langmuir volume, the Langmuir pressure, the microcrack spacing and aperture, the hydraulic fracture distribution, the rock particle density, the rock compressibility and the gas adsorption-desorption curve.

[0007] Furthermore, based on the basic data of fractured shale gas reservoirs, a dual medium model is used to describe bedrock and microfractures, and an embedded discrete model is used to describe hydraulic fractures. The two are combined to establish a matrix-microfracture-hydraulic fracture coupling model suitable for fractured shale gas reservoirs, including: The matrix and microfracture system are characterized by a dual medium model; Characterization of hydraulic fracture systems through embedded discrete models; The dual medium model and the embedded discrete model are coupled to establish a matrix-microfracture-hydraulic fracture coupling model of shale gas reservoirs.

[0008] Furthermore, in the neural network model, the pressure data of the previous time step is trained in the main network and then input into the two branch networks to obtain the pressure data of the matrix system and the fracture system respectively, wherein the hyperparameter settings of each fully connected network are independent of each other; The fully connected network includes: Input layer, used to read and save input data; A hidden layer is located after the input layer and is set to be multiple, and the number of neurons in each hidden layer is set to be multiple. The hidden layer is used to perform feature extraction and nonlinear processing on the data of the previous layer; The output layer, located after the hidden layer, is used to generate the final training results and output them; Activation function, located in each neuron of the hidden layer and output layer, is used to perform nonlinear processing on neurons; Weights and biases are located at the connection point of each neuron. Weights are used to change the size of the input data, and biases are used to change the activation threshold of the neuron.

[0009] Furthermore, a network structure is added to the neural network model in the following manner: Adjacent position anchoring is to import the adjacency matrix representing the connection relationship between adjacent grids into the network while the neural network linearly processes the pressure field. The input pressure field is first encoded in the output layer to obtain the pressure matrix and the adjacency matrix representing the connection relationship between grids. After multiplying the pressure matrix with the input layer weight, the target pressure matrix is ​​obtained by adding the adjacency matrix. The adaptive activation function normalizes the target pressure matrix and then performs nonlinear processing on the normalized target pressure matrix in the hidden layer; Cross-layer connection and gated update are to directly transmit the input pressure field at the input layer to the output layer to achieve cross-layer connection, and combine the input pressure field and the nonlinearly processed pressure matrix for output through gated update.

[0010] Furthermore, the initial pressure is used as label data for pre-training to initialize the neural network model, including: The initial shale gas reservoir pressure data is used as pre-training label data; The initial pressure data of the shale gas reservoir is input into the constructed neural network model to obtain the pressure output data; The mean square error between the model's pressure output data and the label data is used as the loss value; Based on the set learning rate and ADAM optimizer, back propagation is performed to update the neural network model parameters; Iterate the training until the loss of the neural network model drops to the set range to obtain the initialized neural network model.

[0011] Furthermore, the calculation formula using the mean square error between the model's pressure output data and the label data as the loss value is: (4) in, is the loss value; For Grid , is the number of grids; Output data for the convolutional neural network model; is the label data.

[0012] Furthermore, based on the initialized neural network model, the dynamic time step method is used to train all time steps. During the training process of each time step, the pressure predicted by the neural network model is substituted into the overall loss function, and the network parameters are continuously optimized by back propagation until the overall loss is reduced to the specified range. The pressure field of the shale gas reservoir at all time steps is obtained, including: Dynamically set the time step according to pressure changes; The governing equation for determining fractured shale gas reservoirs is: (5) Among them, the superscript represents the previous time step, is the current time step; subscript Representative Grid , is the total number of grids; For Grid Volume, m 3 ; Representation Grid All adjacent grids ; For adjacent grids To Grid Mass flow rate, kg / s; is the rock porosity; is the gas viscosity; is the source-sink term, kg / mžs; is the gas viscosity, the mass of gas adsorbed on unit volume of rock particles, kg / m 3 ; Gas viscosity The expression is: (6) in, is pressure, Pa; is the molar mass, g / mol; is the gas compressibility factor; is the gas molecular constant, Pa·m 3 / (mol·K) ; is the absolute temperature, °C; The gas viscosity (the mass of gas adsorbed on a unit volume of rock particles) is determined by the following formula: : (12) in, , represents the first calculation coefficient, is the rock particle density, kg / m 3 ; is the gas density under standard conditions, kg / m 3 ; is the Langmuir volume, m 3 / kg; is the Langmuir pressure, Pa; Substituting equation (6) and equation (12) into equation (5), we obtain: (14) in, , represents the second calculation coefficient; Transforming equation (14), the residual of the shale gas reservoir control equation is obtained as: (twenty one) in, is the residual of the shale gas reservoir control equation at the current time step; t n+1 is the moment represented by the current time step, s; t n is the moment represented by the previous time step, s; is the neighboring grid at the current time step To Grid Mass flow rate, kg / s; The loss function of the shale gas reservoir control equation based on the residual construction of the shale gas reservoir control equation is: (twenty two) in, is the number of grids; MSE PDE is the loss function of the shale gas reservoir control equation; The grid in the shale gas reservoir control equation is The residual of During the elastic exploitation of shale gas reservoirs, the expert experience constraint consists of three parts. The first part is that the pressure of the shale gas reservoir is less than the initial reservoir pressure and greater than the bottom hole flow pressure; the second part is that the pressure of the next time step is less than the pressure of the current time step; the third part is that the fracture system pressure is less than the bedrock system pressure. Based on the first, second and third parts, the expert experience loss function is constructed. MSE exp : (twenty three) in, is the pressure range constraint, is the pressure decreasing constraint, Constraints for pressure systems; According to the loss function of the shale gas reservoir control equation and the expert experience loss function, the overall loss function is constructed MSE : (27) According to the overall loss function of formula (27), back propagation is performed to continuously optimize the model parameters until the loss value is reduced to the set range, and the pressure field of the shale gas reservoir at the current time step is output.

[0013] Furthermore, the pressure field data of the previous time step is used as the input of the current time step to obtain the pressure field data of the current time step, and the above steps are repeated to obtain the pressure field data of any time step.

[0014] The beneficial effects of this application are: The present invention can realize the forward solution of the pressure distribution of the shale gas reservoir coupling model (matrix + microfractures + hydraulic fractures) at any time step without labeled data, and significantly improves the solution speed while ensuring accuracy. A core of the present invention is to construct a loss function through the finite volume method and expert experience, and to improve the solution efficiency by combining expert experience constraints while considering the strong nonlinear problem of shale gas; to establish a neural network model with a fully connected network as the main and two fully connected networks as the auxiliary, and to add adjacent position anchoring, adaptive activation function, cross-layer connection and gated update network structures on this basis; to use a dynamic adjustment method for the time step so that the network can train the early pressure changes more finely; to introduce pre-training, initialize the network, reduce the randomness of the network output, and ensure that the network output is positive; to use the transfer learning method in the pressure training process to further improve the training efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A flowchart of a physical information neural network numerical simulation method applicable to fractured shale gas reservoirs provided in an embodiment of the present application; Figure 2 A specific calculation flow chart of the neural network model provided in the embodiment of the present application; Figure 3 A shale gas reservoir fracture distribution map provided in an embodiment of the present application; Figure 4 The real pressure field diagram for 100 days provided in the embodiment of the present application; Figure 5 A 100-day predicted pressure field diagram provided in the embodiment of the present application; Figure 6 The 100 days provided in the embodiment of the present application are relative error diagrams of the actual pressure field and the predicted pressure field; Figure 7 A comparison chart of the actual value and predicted value of the cumulative gas production of shale gas reservoirs provided in the embodiment of the present application; Figure 8 A comparison chart of the actual and predicted daily gas production of shale gas reservoirs provided in the embodiments of the present application. DETAILED DESCRIPTION

[0016] The following describes the embodiments of the present application through specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the contents disclosed in this specification. The present application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0017] The specific implementation of the present application is further described in detail below in conjunction with the drawings and examples.

[0018] See also Figure 1 , which is a flow chart of a physical information neural network numerical simulation method applicable to fractured shale gas reservoirs provided in an embodiment of the present application. The present application embodiment provides a physical information neural network numerical simulation method applicable to fractured shale gas reservoirs, which uses an improved fully connected neural network model to embed physical meanings to solve the pressure of shale gas reservoirs. Figure 1 As shown, it includes the following steps 1 to 6.

[0019] Step 1: Collect relevant data of shale gas reservoirs, including the size of reservoir and fracturing transformation area, initial gas reservoir pressure, bedrock porosity, shale gas reservoir temperature, Langmuir volume, Langmuir pressure, number and location of wells, bottom hole flowing pressure, gas molecular mass, gas constant, gas standard volume, micro-fracture spacing and opening, hydraulic fracture distribution, rock grain density, rock compression coefficient and gas adsorption and desorption curve, etc. The reservoir and fracturing transformation area information is used to establish the matrix-micro-fracture-hydraulic fracture coupling model of fractured shale gas reservoirs in step 2, including the size of reservoir and fracturing transformation area, bedrock porosity, number and location of wells, micro-fracture spacing and opening, hydraulic fracture distribution, rock grain density, rock compression coefficient; the rest of the information is mainly used to construct the loss function of the neural network in step 5.

[0020] Step 2: Based on the basic data of fractured shale gas reservoirs, a dual medium model is used to describe bedrock and microfractures, and an embedded discrete model is used to describe hydraulic fractures. The two are combined to establish a matrix-microfracture-hydraulic fracture coupling model suitable for fractured shale gas reservoirs for training and prediction of the neural network model. The specific model construction process is as follows: Step 2.1: Use the dual medium model to characterize the matrix system (bedrock + microfractures) and describe the preferential flow characteristics of microfractures.

[0021] Step 2.2: Use the embedded discrete fracture model to characterize the fracture system (hydraulic fracture). While describing the high conductivity of the hydraulic fracture, the unstructured grid design based on the hydraulic fracture can be avoided.

[0022] Step 2.3: According to the quasi-steady-state cross-flow between the microfracture system and the hydraulic fracture system, the two models are coupled to establish a matrix-microfracture-hydraulic fracture coupling model of the shale gas reservoir.

[0023] Step 3: Establish a neural network model with a fully connected network as the backbone network and two fully connected networks as branch networks, add network structures such as adjacent position anchoring, adaptive activation function, cross-layer connection and gated update, and set relevant hyperparameters.

[0024] A fully connected neural network usually consists of five parts: (1) the input layer, whose main function is to read and save input data; (2) the hidden layer, which is located after the input layer. Its number can be flexibly changed. The number of neurons in each hidden layer is also flexible, making the grid fully flexible. Its main function is to extract features and perform nonlinear processing on the data in the previous layer; (3) the output layer, which is located after the hidden layer. Its main function is to generate the final training results and output them; (4) the activation function, which is located in each neuron in the hidden layer and the output layer. Its main function is to perform nonlinear processing on neurons, thereby improving the neural network's ability to extract features and fit; (5) weights and biases, which are located at the location where each neuron is connected. The former is used to change the size of the input data, and the latter is used to change the activation threshold of the neuron. On this basis, network structures such as adjacent position anchoring, adaptive activation function, cross-layer connection and gated update are added.

[0025] The steps for constructing the neural network model for shale gas reservoirs are as follows: Step 3.1, first establish a main fully connected neural network, and then establish two branch fully connected neural networks. The input of the first fully connected neural network is the pressure data of the previous time step, and the output is the input of the two branch fully connected neural networks. The outputs of the two neural networks are the pressure of the matrix system and the fracture system, respectively.

[0026] Step 3.2: Add adjacent position anchoring based on the above neural network model, and import the adjacency matrix representing the connection relationship between adjacent grids into the network while the neural network linearly processes the pressure field. First, the input pressure field is processed in the output layer. Encode and get The pressure matrix , which represents the adjacency matrix of the connection relationship between grids , the same size . and After multiplication, add Then we get the target pressure matrix , the specific process is shown in formula (1): (1) in, is the input layer weight; Bias for the input layer; is the adjacency matrix.

[0027] Step 3.3: Add an adaptive activation function based on the above neural network model. Normalized, and then nonlinearly processed in the hidden layer, as shown in formula (2): (2) in, is the pressure matrix for nonlinear processing; is the hidden layer weight; is the hidden layer bias; ReLU is the linear rectification activation function; is the learnable parameter of the adaptive activation function; BatchNorm For standardization, the pressure features of different dimensions are normalized to keep them in the same order of magnitude to maintain the stability of gradient back propagation; Step 3.4: Add cross-layer connections and gate updates based on the above neural network model, and convert the initial input pressure field in the input layer Directly transmit to the output layer to realize cross-layer connection, and update the input pressure field through the gated update method of formula (3). and the pressure matrix for nonlinear processing Combined output: (3) in, is the pressure matrix at time t+1; is the updateable weight; is the output layer weight; is the bias for the output layer.

[0028] Step 4: Use the initial pressure as label data for pre-training and initialize the neural network model.

[0029] Since the solution of shale gas pressure is a strong nonlinear problem, the coefficients in the control equation are all functions of pressure. Since the initial training of the neural network is highly random, if the initial training result is not a positive value, some coefficients in the control equation cannot be solved, and the loss value in step 5 cannot be calculated, resulting in simulation failure. Therefore, the initial pressure data is used as input and label for pre-training, and the relevant parameters of the model are updated to reduce the error between the output and the true value of the model at the zeroth time step, ensuring that the model output is positive, so that the coefficients of the control equation can be solved, and then the loss value in step 5 can be calculated. This step is only to reasonably output the neural network. The solution of subsequent time steps does not require label data. It only needs to be forward-solved according to the control equation and expert constraints as usual to obtain the pressure of the next time step.

[0030] The specific steps of pre-training are as follows: Step 4.1, using the initial pressure data of the shale gas reservoir as label data; Step 4.2, set the appropriate number of iterations and loss value, input the initial pressure data of the shale gas reservoir into the neural network model constructed in step 3, and obtain output data with the same size as the input data; Step 4.3, substitute the output data and label data into the formula, and calculate the mean square error between the two as the training loss value; (4) in, is the loss value; For Grid , is the number of grids; Output data for the convolutional neural network model; is the label data.

[0031] Step 4.4, set the appropriate learning rate, select the ADAM optimizer, back propagate, and update the model related parameters; Step 4.5: Repeat steps 4.2 to 4.4 until the loss value reaches the expected value. At this time, the model-related parameters are initialized, and the initialized neural network model is obtained.

[0032] Step 5: Use the dynamic time step method to train all time steps, that is, the time step length gradually increases over time until it reaches the specified maximum value. During the training process of each time step, the pressure predicted by the neural network is substituted into the overall loss function (physical loss of the control equation based on finite volume discretization + expert experience loss), and back propagation is performed to continuously optimize the network parameters until the overall loss is reduced to the specified range; During the elastic exploitation of fractured shale gas reservoirs, the early pressure will change greatly in a short period of time, so the time step cannot be too large, otherwise the pressure change cannot be accurately simulated. The later pressure does not change much in a long time, so a larger time step can be used to simulate the pressure, reduce the number of training times, and improve training efficiency.

[0033] The specific steps for setting the dynamic time step are: Step 5.1, set the time step to 0.001 day, simulate training 10 times to 0.01 day, and output the pressure data for each time step; Step 5.2, change the time step to 0.01 days, simulate training 9 times, the total time is 0.1 days, and output the pressure data for each time step; Step 5.3, change the time step to 0.1 day, simulate training 9 times, the total time is 1 day, and output the pressure data for each time step; Step 5.4, change the time step to 1 day, and keep it unchanged until the end of the simulation, and output the pressure data for each time step; According to the pressure and model parameters of the previous time step, forward propagation is performed, the loss of the control equation is calculated by the finite volume method, and reverse propagation is performed with the assistance of expert constraints to continuously optimize the model parameters until the loss value is reduced to the set range, and the pressure field of the shale gas reservoir in the current time step is output; No label data is required for the time step after pre-training. It is only necessary to perform forward propagation based on the pressure data and model parameters output by the model, construct the loss function using the finite volume method and expert loss, and calculate the loss error output by the model. The specific steps are as follows: Step 5.5, input the initial pressure data of the shale gas reservoir into the convolutional neural network model initialized in step 4, perform a forward propagation, and obtain output data of the same size as the input data; Step 5.6: Construct the residual of the control equation taking into account the adsorption / desorption effect, real gas and slippage effect. The calculation formula construction process is as follows: The governing equation of shale gas reservoir is shown in formula (5): (5) Among them, the superscript represents the previous time step, is the current time step; subscript Representative Grid , is the total number of grids; For Grid Volume, m 3 ; Representation Grid All adjacent grids ; For adjacent grids To Grid Mass flow rate, kg / s; is the rock porosity; is the gas viscosity, and its expression is: (6) in, is pressure, Pa; is the molar mass, g / mol; is the gas compressibility factor; is the gas molecular constant, Pa·m 3 / (mol·K) ; is the absolute temperature, ℃.

[0034] In formula (5), is the source and sink term, kg / mžs. Since there is a large pressure gradient near the well, it is difficult to solve it directly by automatic differentiation. Therefore, a well model is introduced to reduce the pressure gradient near the well. The specific definition is as follows: (7) in is the well index, defined as: (8) in is the oil layer thickness, m; is the wellbore radius, m; is the skin factor; is the gas reservoir radius, and its expression is: (9) in is the length of the grid in the x direction, m; is the length of the grid in the y direction, m.

[0035] Formula (7) is the total fluidity, 1 / (Pa·s), defined as: (10) in, is the viscosity of the fluid, Pa·s.

[0036] In formula (5), is the gas viscosity, the mass of gas adsorbed on unit volume of rock particles, kg / m 3 , whose expression is: (11) in, is the rock particle density, kg / m 3 ; is the gas density under standard conditions, kg / m 3 ; is the Langmuir volume, m 3 / kg; is the Langmuir pressure, Pa.

[0037] Simplifying, we get: (12) in, , represents the first calculation coefficient.

[0038] Substituting equation (6) and equation (12) into equation (5), we obtain: (13) Simplifying, we get: (14) in, , represents the second calculation coefficient; Neighboring grids at the current time step To Grid Mass flow rate The expression is: (15) in, is the geometric conductivity term of the conductivity, and The current time step grid j and Grid i Pressure, Pa; is the fluidity term of the conductivity coefficient, expressed as: (16) In formula (15), is the geometric conductivity term of the conductivity. For an orthogonal grid, its expression is: (17) in, For Grid and Grid Contact area, m 2 ; For Grid Center to grid and Grid The distance of the contact surface, m; For Grid Center to grid and Grid The distance of the contact surface, m; For Grid and Grid The intrinsic permeability The average value of , generally the harmonic mean is used.

[0039] In shale gas reservoirs, in order to consider the Klinkenberg effect, equation (16) is modified as follows: (18) in, is the Klinkenberg coefficient, expressed as: (19) in, is the effective Knudsen diffusion coefficient of gas migration in porous media, in m 2 / s; is the intrinsic permeability of the porous medium, in m 2 ; is the gas compressibility coefficient, the unit is 1 / Pa, and the expression is: (20) Transforming equation (14), the residual of the shale gas reservoir control equation is obtained as: (twenty one) in, is the residual of the shale gas reservoir control equation at the current time step; t n+1 is the moment represented by the current time step, s; t n is the moment represented by the previous time step, s; The loss function of the shale gas reservoir control equation is constructed as follows: (twenty two) in, is the number of grids.

[0040] Step 5.7, construct the expert experience loss function. During the elastic exploitation of shale gas reservoirs, the expert experience constraints consist of three parts: the pressure of the shale gas reservoir must be less than the initial reservoir pressure and greater than the bottom hole flow pressure; the pressure of the next time step must be less than the pressure of the current time step; the fracture system pressure must be less than the bedrock system pressure. The expert experience loss function is constructed as follows: (twenty three) in, is the constraint of the pressure range, that is, the pressure in the shale gas reservoir must be between the initial reservoir pressure and the bottom hole flowing pressure. The expression is: (twenty four) in, is the bottom hole flowing pressure, Pa; is the initial pressure of shale gas reservoir, Pa; ReLU is the linear rectification activation function.

[0041] In formula (23), is the constraint of decreasing pressure, that is, the pressure in the gas reservoir in the next time step must be less than the pressure in the current time step. The expression is: (25) in, is the pressure of the shale gas reservoir at the next time step, Pa; is the pressure of the shale gas reservoir at the current time step, Pa.

[0042] In formula (23), is the constraint of the pressure system, that is, the pressure in the gas reservoir in the next time step must be less than the pressure in the current time step, and the expression is: (26) in, is the pressure of the fracture system, Pa; is the matrix grid pressure adjacent to the fracture system, Pa; is the pressure of the remaining matrix system, Pa.

[0043] Step 5.8, combining the control equation loss function and the expert experience loss function, the complete loss function is established as follows: (27) Step 5.7: Perform back propagation according to the loss function of formula (27), continuously optimize the model parameters until the loss value is reduced to a certain range, and output the pressure field of the shale gas reservoir at the current time step.

[0044] Step 6: Use the pressure data of the previous time step as the input for the current time step training, and repeat step 5 to obtain the pressure data of the shale gas reservoir at any time step, and then obtain a series of data such as the internal pressure and production well output at any time step.

[0045] In order to prove the feasibility of the present invention, the following specific example is given. The process of numerical simulation of shale gas reservoir in this example is as follows: The gas reservoir model size is 440m×250m×10m, and the grid size is 10m×10m×10m. The initial gas reservoir pressure is 25MPa, the initial porosity is 0.08, the shale gas reservoir temperature is 70℃, and the rock particle density is 2850kg / m 3 , the rock compression coefficient is 1e -10 1 / Pa, Langmuir volume is 0.015m 3 / kg, Langmuir pressure is 2.5MPa, and the distribution of fractures in shale gas reservoirs after hydraulic fracturing is as follows Figure 3 As shown, the production well locations are marked in the figure (corresponding to the dot locations in the figure). The simulation time is 100 days and the time step is 1 day.

[0046] Construct a fully connected neural network model, such as Figure 2 As shown, there are three fully connected neural networks in total. Each fully connected neural network includes an input layer, an output layer and a hidden layer. The matrix system pressure and the fracture system pressure are solved respectively through two branch networks.

[0047] The initial pressure data of shale gas reservoirs is used as label data, and the number of iterations is set to 5000 and the loss value is set to 1×10 -18 , the initial pressure data of shale gas reservoirs is input into the neural network model, and the output data with the same size as the input data is obtained. The mean square error between the output data and the label data is calculated as the training loss value. The learning rate is set to 0.001, the ADAM optimizer is selected, back propagation is performed, and the model related parameters are updated. The above process is repeated until the loss value reaches 1×10 -18 , and get the initialized neural network model.

[0048] Set the dynamic time step. The time step is first set to 0.001 days, and simulate training 10 times to 0.01 days; then change the time step to 0.01 days, simulate training 9 times, and the total time is 0.1 days; then change the time step to 0.1 days, simulate training 9 times, and the total time is 1 day, and output the pressure data for each time step; finally change the time step to 1 day, and keep it unchanged until the end of the simulation.

[0049] The training process after pre-training does not require labeled data. It performs forward propagation based on the pressure and model parameters of the previous time step, calculates the loss of the control equation through the finite volume method, and performs back propagation with the help of expert constraints to continuously optimize the model parameters until the loss value is reduced to 1×10 -18 , output the pressure field of the shale gas reservoir at the current time step. Repeat the above process to obtain the pressure field data of any time step, and the data simulation is completed. On this basis, the daily production and cumulative production of the production well can be further obtained.

[0050] Figure 4 The comparison between the predicted pressure field data and the actual pressure field data at 100 days, from left to right are the predicted pressure field map and the actual pressure field map. Figure 5 is the relative error between the two. It can be seen that the maximum relative error of pressure prediction is 0.07%. Figure 6 , Figure 7 The cumulative production and daily production of shale gas reservoirs are compared with the actual values. It can be seen that the simulation accuracy is very high.

[0051] The above implementation modes are only used to illustrate the present application, and are not intended to limit the present application. Ordinary technicians in the relevant technical field may make various changes and modifications without departing from the spirit and scope of the present application. Therefore, all equivalent technical solutions also belong to the scope of the present application, and the scope of patent protection of the present application shall be limited by the claims.

Claims

1. A physical information neural network numerical simulation method suitable for fractured shale gas reservoirs, characterized in that: The method comprises: Obtain basic data of fractured shale gas reservoirs; Based on the basic data of fractured shale gas reservoirs, a dual medium model is used to describe bedrock and microfractures, and an embedded discrete model is used to describe hydraulic fractures. The two are combined to establish a matrix-microfracture-hydraulic fracture coupling model suitable for fractured shale gas reservoirs, which is used for the training and prediction of neural network models. Establish a neural network model with a fully connected network as the main network and two fully connected networks as branch networks, and add a network structure to the neural network model and set relevant hyperparameters; wherein the network structure includes adjacent position anchoring, adaptive activation function, cross-layer connection and gated update; Use the initial pressure as label data for pre-training and initialize the neural network model; Based on the initialized neural network model, the dynamic time step method is used to perform training for all time steps. During the training process of each time step, the pressure predicted by the neural network model is substituted into the overall loss function, and back propagation is performed to continuously optimize the network parameters until the overall loss is reduced to a specified range, thereby obtaining the pressure field of the shale gas reservoir at all time steps; wherein the overall loss function includes the physical loss of the control equation based on finite volume discretization and the expert experience loss.

2. The physical information neural network numerical simulation method applicable to fractured shale gas reservoirs according to claim 1, characterized in that: The basic data of the fractured shale gas reservoir include the size of the reservoir and the fracturing transformation area, the initial gas reservoir pressure, the bedrock porosity, the shale gas reservoir temperature, the Langmuir volume, the Langmuir pressure, the micro-fracture spacing and aperture, the hydraulic fracture distribution, the rock particle density, the rock compressibility and the gas adsorption-desorption curve.

3. The physical information neural network numerical simulation method applicable to fractured shale gas reservoirs according to claim 1, characterized in that: Based on the basic data of fractured shale gas reservoirs, a dual medium model is used to describe bedrock and microfractures, and an embedded discrete model is used to describe hydraulic fractures. The two are combined to establish a matrix-microfracture-hydraulic fracture coupling model suitable for fractured shale gas reservoirs, including: The matrix and microfracture system are characterized by a dual medium model; Characterization of hydraulic fracture systems through embedded discrete models; The dual medium model and the embedded discrete model are coupled to establish a matrix-microfracture-hydraulic fracture coupling model of shale gas reservoirs.

4. The physical information neural network numerical simulation method applicable to fractured shale gas reservoirs according to claim 1, characterized in that: In the neural network model, the pressure data of the previous time step is trained in the main network and then input into the two branch networks to obtain the pressure data of the matrix system and the fracture system respectively, wherein the hyperparameter settings of each fully connected network are independent of each other; The fully connected network includes: Input layer, used to read and save input data; A hidden layer is located after the input layer and is set to be multiple, and the number of neurons in each hidden layer is set to be multiple. The hidden layer is used to perform feature extraction and nonlinear processing on the data of the previous layer; The output layer, located after the hidden layer, is used to generate the final training results and output them; Activation function, located in each neuron of the hidden layer and output layer, is used to perform nonlinear processing on neurons; Weights and biases are located at the connection point of each neuron. Weights are used to change the size of the input data, and biases are used to change the activation threshold of the neuron.

5. The physical information neural network numerical simulation method applicable to fractured shale gas reservoirs according to claim 4, characterized in that: Add a network structure to the neural network model as follows: Adjacent position anchoring is to import the adjacency matrix representing the connection relationship between adjacent grids into the network while the neural network linearly processes the pressure field. The input pressure field is first encoded in the output layer to obtain the pressure matrix and the adjacency matrix representing the connection relationship between grids. After multiplying the pressure matrix with the input layer weight, the target pressure matrix is ​​obtained by adding the adjacency matrix. The adaptive activation function normalizes the target pressure matrix and then performs nonlinear processing on the normalized target pressure matrix in the hidden layer; Cross-layer connection and gated update are to directly transmit the input pressure field at the input layer to the output layer to achieve cross-layer connection, and combine the input pressure field and the nonlinearly processed pressure matrix for output through gated update.

6. The physical information neural network numerical simulation method applicable to fractured shale gas reservoirs according to claim 1, characterized in that: Use the initial pressure as label data for pre-training and initialize the neural network model, including: The initial shale gas reservoir pressure data is used as pre-training label data; The initial pressure data of the shale gas reservoir is input into the constructed neural network model to obtain the pressure output data; The mean square error between the model's pressure output data and the label data is used as the loss value; Based on the set learning rate and ADAM optimizer, back propagation is performed to update the neural network model parameters; Iterate the training until the loss of the neural network model drops to the set range to obtain the initialized neural network model.

7. The physical information neural network numerical simulation method applicable to fractured shale gas reservoirs according to claim 6, characterized in that: The calculation formula for the loss value is the mean square error between the model's pressure output data and the label data: (4) in, is the loss value; For Grid , is the number of grids; Output data for the convolutional neural network model; is the label data.

8. The physical information neural network numerical simulation method applicable to fractured shale gas reservoirs according to claim 1, characterized in that: Based on the initialized neural network model, the dynamic time step method is used to train all time steps. During the training process of each time step, the pressure predicted by the neural network model is substituted into the overall loss function, and the network parameters are continuously optimized by back propagation until the overall loss is reduced to the specified range. The pressure field of the shale gas reservoir at all time steps is obtained, including: Dynamically set the time step according to pressure changes; The governing equation for determining fractured shale gas reservoirs is: (5) Among them, the superscript represents the previous time step, is the current time step; subscript Representative Grid , is the total number of grids; For Grid Volume, m 3 ; Representation Grid All adjacent grids ; For adjacent grids To Grid Mass flow rate, kg / s; is the rock porosity; is the gas viscosity; is the source-sink term, kg / mžs; is the gas viscosity, the mass of gas adsorbed on unit volume of rock particles, kg / m 3 ; Gas viscosity The expression is: (6) in, is pressure, Pa; is the molar mass, g / mol; is the gas compressibility factor; is the gas molecular constant, Pa·m 3 / (mol·K) ; is the absolute temperature, °C; The gas viscosity (the mass of gas adsorbed on a unit volume of rock particles) is determined by the following formula: : (12) in, , represents the first calculation coefficient, is the rock particle density, kg / m 3 ; is the gas density under standard conditions, kg / m 3 ; is the Langmuir volume, m 3 / kg; is the Langmuir pressure, Pa; Substituting equation (6) and equation (12) into equation (5), we obtain: (14) in, , represents the second calculation coefficient; Transforming equation (14), the residual of the shale gas reservoir control equation is obtained as: (21) in, is the residual of the shale gas reservoir control equation at the current time step; t n+1 is the moment represented by the current time step, s; t n is the moment represented by the previous time step, s; is the neighboring grid at the current time step To Grid Mass flow rate, kg / s; The loss function of the shale gas reservoir control equation based on the residual construction of the shale gas reservoir control equation is: (22) in, is the number of grids; MSE PDE is the loss function of the shale gas reservoir control equation; The grid in the shale gas reservoir control equation is The residual of During the elastic exploitation of shale gas reservoirs, the expert experience constraint consists of three parts. The first part is that the pressure of the shale gas reservoir is less than the initial reservoir pressure and greater than the bottom hole flow pressure; the second part is that the pressure of the next time step is less than the pressure of the current time step; the third part is that the fracture system pressure is less than the bedrock system pressure. Based on the first, second and third parts, the expert experience loss function is constructed. MSE exp : (23) in, is the pressure range constraint, is the pressure decreasing constraint, Constraints for pressure systems; According to the loss function of the shale gas reservoir control equation and the expert experience loss function, the overall loss function is constructed MSE : (27) According to the overall loss function of formula (27), back propagation is performed to continuously optimize the model parameters until the loss value is reduced to the set range, and the pressure field of the shale gas reservoir at the current time step is output.

9. The physical information neural network numerical simulation method applicable to fractured shale gas reservoirs according to claim 8, characterized in that: The pressure field data of the previous time step is used as the input of the current time step to obtain the pressure field data of the current time step. The above steps are repeated to obtain the pressure field data of any time step.

Images