Complex dynamic reservoir simulation method based on adaptive physical meaning neural network

Abstract

The invention relates to the field of reservoir engineering technology, and discloses a complex dynamic reservoir simulation method based on an adaptive physical meaning neural network, the method comprising: step 1, obtaining complex production dynamic reservoir development data; step 2, establishing an adaptive physical meaning neural network model, initializing model parameters, and setting a production control system table; step 3, responding to the input initial pressure data of the reservoir, adjusting the internal boundary conditions according to the production system table, forward propagation mapping and back propagation calculation gradient; step 4, updating the saturation according to the pressure data output by the model; step 5, iteratively executing step 3 and step 4, using the pressure field output at each step and the neural network parameters obtained by training as the input of the next step, and continuously solving the simulation process of the entire reservoir. The invention can quickly adapt to the drastic changes of the reservoir under complex production dynamics.

Description

Technical Field

[0001] The invention relates to the technical field of oil reservoir engineering, and in particular to a complex dynamic oil reservoir simulation method based on an adaptive physical meaning neural network. Background Art

[0002] In the study of oil and gas field development, numerical simulation is an important technology that can be used to describe and predict the distribution of remaining oil in the reservoir and the effect of injection and production development. Traditional reservoir numerical simulation methods usually use the finite volume method to construct a numerical format that is discrete in time and space, and perform iterative calculations through numerical solution methods to obtain simulation results such as pressure and saturation distribution at different discrete time steps, as well as injection and production well dynamics, to guide the design and adjustment of reservoir development plans. However, traditional reservoir numerical simulation technology must strictly implement the numerical iteration format for cyclic calculations in the solution of each time step. When simulating complex dynamic reservoirs, a large number of iterative calculations may be required to achieve convergence. Each development plan simulation requires a large amount of computing resources and is expensive to use.

[0003] Recently, with the continuous development of computational physics and artificial intelligence technology, solving distributed parameter systems controlled by partial differential equations based on neural networks has gradually become a cutting-edge technology and has been widely studied and applied in underground flow simulation problems. This technology can use partial differential equations to construct physical constraints, thereby using neural networks for physically driven data-free training. However, the flow field of complex dynamic reservoirs changes violently, is a strong nonlinear problem, and has non-stationary changes. At present, the efficiency of solving complex dynamic reservoir numerical models based on physical meanings of neural networks is low, and it is even impossible to solve the problem of frequent changes in development and control systems. The convergence speed of the solution is slow and the convergence accuracy is low. Summary of the invention

[0004] The purpose of the present invention is to provide a complex dynamic reservoir simulation method based on an adaptive physical meaning neural network, which uses the pressure distribution and saturation distribution of complex dynamic reservoirs to solve the problem, and designs a point-by-point adaptive learning mechanism and vectorized discrete parameter gating, so that the model can quickly adapt to the dramatic changes of the reservoir under complex production dynamics through a physically driven training process without label data. This method shows excellent solution speed and convergence accuracy when dealing with complex oil and water well operations and changing production strategies, and is particularly suitable for reservoir conditions with significant heterogeneity and dynamic changes. The trained neural network can achieve an approximate format of the partial differential equation in the corresponding simulation scenario while solving each time step, and obtains good solution speed and convergence accuracy in complex oil and water well injection and production conversion, periodic injection and production, and continuous variable injection and production embodiments.

[0005] In order to achieve the above purpose, the technical solutions adopted are as follows:

[0006] The present invention provides a complex dynamic reservoir simulation method based on an adaptive physical meaning neural network, the method comprising:

[0007] Step 1: Obtaining complex production dynamic reservoir development data, including reservoir permeability distribution, fluid properties and complex production dynamic data;

[0008] Step 2: Establish an adaptive physical meaning neural network model, wherein the adaptive physical meaning neural network model includes a vectorized discrete parameter gating structure so that the model can be adjusted independently at each data point; initialize the model parameters, and set the production control system table;

[0009] Step 3: In response to the inputted initial pressure data of the reservoir, the internal boundary conditions are adjusted according to the production system table, the forward propagation mapping and the back propagation gradient calculation are performed, and the point-by-point adaptive update of the model parameters is achieved through the vectorized discrete parameter gating structure;

[0010] Step 4: Update the saturation according to the pressure data output by the model;

[0011] Step 5: Based on the set simulation time, iteratively execute steps 3 and 4, use the pressure field output in each step and the trained neural network parameters as the input for the next step, and continue to solve the simulation process of the entire reservoir until the set simulation time is reached.

[0012] Preferably, the complex production dynamic data includes the positions of the injection well and the production well at each time step, the injection rate of the injection well at each time step, and the production rate of the production well at each time step.

[0013] Preferably, the adaptive physical meaning neural network model includes an input fully connected layer, a hidden fully connected layer, an output fully connected layer, an adjacent position anchoring layer, a regularization layer, an adaptive activation function layer and a vectorized discrete parameter gating update layer; wherein the input fully connected layer is located at the front end of the neural network, the adjacent position anchoring layer is located after the input fully connected layer, the regularization layer is located after the adjacent position anchoring layer, the hidden fully connected layer is located after the adjacent position anchoring layer, the adaptive activation function layer is located after the hidden fully connected layer, the output fully connected layer is located after the adaptive activation layer, and the vectorized discrete parameter gating update layer is located after the output fully connected layer; the adaptive physical meaning neural network model learns the mapping of pressure to pressure between adjacent time steps.

[0014] Preferably, the mapping functions of the input fully connected layer, the adjacent position anchor layer, the hidden fully connected layer, the adaptive activation function layer, the output fully connected layer, and the vectorized discrete parameter gated update layer are determined by the following method:

[0015] The reservoir pressure distribution matrix is ​​used as the input pressure matrix of the input fully connected layer, and the mapping function of the input fully connected layer is designed as follows:

[0016] (1)

[0017] in is the input fully connected layer mapping function, is the input fully connected layer pressure matrix, is the input fully connected layer weight, is the input fully connected layer bias;

[0018] The pressure matrix output by the input fully connected layer is used as the input pressure matrix of the adjacent position anchor layer, and the mapping function of the adjacent position anchor layer is designed as follows:

[0019] (2)

[0020] in is the adjacent position anchor layer mapping function, is the pressure matrix of the adjacent anchor layer, is the anchor weight, is the adjacency position matrix, is the mean of the anchor pressure matrix, is the standard deviation of the anchor pressure matrix, is a small value that prevents the denominator from reaching zero, The weight of the adjacent position anchor layer, is the anchor layer deviation of the adjacent position;

[0021] The pressure matrix output by the adjacent anchor layer is used as the input pressure matrix of the hidden fully connected layer, and the mapping function of the hidden fully connected layer is designed as follows:

[0022] (3)

[0023] in is the hidden fully connected layer mapping function, is the hidden fully connected layer pressure matrix, is the hidden fully connected layer weight, is the hidden fully connected layer bias;

[0024] The pressure matrix output by the hidden fully connected layer is used as the input pressure matrix of the adaptive activation layer, and the mapping function of the adaptive activation layer is designed as follows:

[0025] (4)

[0026] in is the adaptive activation layer mapping function, is the adaptive activation layer pressure matrix, are the adaptive activation weights, is the maximum value function;

[0027] The output pressure matrix of the adaptive activation layer is used as the input pressure matrix of the output fully connected layer, and the mapping function of the output fully connected layer is designed as follows:

[0028] (5)

[0029] in is the output fully connected layer mapping function, Output the fully connected layer pressure matrix, is the output fully connected layer weight, is the output fully connected layer bias;

[0030] The output pressure matrix of the output fully connected layer and the reservoir pressure matrix of the current time step are used as the input pressure matrix of the vectorized discrete parameter gating update layer, and the mapping function of the vectorized discrete parameter gating update layer is designed as follows:

[0031] (6)

[0032] in is the vectorized discrete parameter gated update layer mapping function, is the gated update candidate pressure matrix, is the gating update weight matrix.

[0033] Preferably, in step 2, initializing model parameters and setting a production control system table include:

[0034] If the current moment is the initial moment, initialize the learnable parameters in the neural network: Initialize Xavier and set The default initial value is 0.01, set The default initial value is 0.1, set The default initial value of all elements is 0.9; divide the time step according to the actual production dynamics and set the production control system table.

[0035] Preferably, in step 3, in response to the inputted initial pressure data of the reservoir, the internal boundary conditions are adjusted according to the production system table, the forward propagation mapping and the back propagation gradient calculation are performed, and the point-by-point adaptive update of the model parameters is achieved through the vectorized discrete parameter gating structure, including:

[0036] Step 3.1: Input the pressure matrix of the current time step into the adaptive physical meaning neural network model for forward propagation mapping, and output the predicted pressure matrix of the next time step. The complete mapping process is:

[0037] (7)

[0038] in is the reservoir pressure matrix at the current moment, The reservoir pressure matrix for the next time step predicted by the output neural network;

[0039] Step 3.2: Construct the finite volume method physical constraint loss function:

[0040] The residual formula for partial differential equation control in the context of slightly compressible two-phase flow is:

[0041] (8)

[0042] in is the residual of the partial differential equation, is the porosity, and are the saturations of the non-wetting phase and the wetting phase, respectively. and are the compressibility coefficients of the non-wetting phase and the wetting phase, is the compressibility coefficient of the flowing medium, and are the source and sink flow rates of the non-wetting phase and the wetting phase, respectively; To control the volume, for The pressure of the moment, for The pressure of the moment, represents the neighboring cells of the grid, and is the grid number, is the median value of the total flow of adjacent grids, is the conductivity, For Grid The wetting phase flow potential, For Grid The wetting phase flow potential, is the median value of the non-wetting phase fluidity between adjacent grids, For Grid The two-phase flow potential difference, For Grid The two-phase flow potential difference;

[0043] Read the source and sink terms of the next time step from the production and control system table, and use the following Peaceman well model to process the source and sink terms:

[0044] (9)

[0045] in is the source-sink flow, is the fluidity, is the permeability, is the formation perforation thickness, is the equivalent formation radius, is the wellbore radius, is the skin coefficient, is the grid pressure, is the bottom hole pressure;

[0046] The inner and outer boundary conditions are processed by hard constraints, and the residual of the physical constraint is:

[0047] (10)

[0048] in, Predicted grid The wetting phase flow potential, Predicted grid The wetting phase flow potential;

[0049] The calculation formula for predicting flow potential is:

[0050] (11)

[0051] in, is the predicted flow potential, is the predicted pressure, is the density, is the acceleration due to gravity, It is depth;

[0052] The loss value is calculated using the mean square error, and the loss function of the physical constraint is:

[0053] (12)

[0054] in, is the loss value at time t+1, is the number of discrete grids, i is the grid number, It is predicted Time pressure;

[0055] Step 3.3: If the current moment is the initial moment, set the default initial learning rate of the optimizer to 0.01 and the default loss threshold for model convergence to 1.0×10 -18 ; The model parameter gradient calculation and update formula is:

[0056] (13)

[0057] in, and They are Iteration step and Neural network model parameters at the iteration step, is the learning rate, is a small value that prevents the denominator from reaching zero, is the residual calculated from the predicted value, is the output pressure of the hidden layer of the neural network, is the gating weight;

[0058] Step 3.4, repeat the forward propagation prediction and backward propagation update process of step 3.1 to step 3.3 until the loss value is reduced to below the preset model convergence loss threshold, the pressure solution is completed, and the pressure data for the next time step is output.

[0059] Preferably, the step 4, updating the saturation according to the pressure data output by the model, comprises:

[0060] According to the output pressure data of the next time step and the saturation data of the current time step, the saturation distribution of the next time step is calculated by the following formula:

[0061] (14)

[0062] in, is the saturation distribution at the next time step, is the saturation data of the current time step, is the wetting phase fluidity.

[0063] Preferably, the step 5, based on the set simulation time, iteratively executes the steps 3 and 4, uses the pressure field outputted in each step and the trained neural network parameters as the input for the next step, and continuously solves the simulation process of the entire reservoir until the set simulation time is reached, including:

[0064] In the solution at the next moment, all the parameters of the neural network trained in the previous time step are used as the initialization parameters of the next time step; the pressure data solved in the previous time step is used as the input pressure data for forward propagation in the next time step; steps 3 to 4 are repeated to solve the pressure distribution and saturation distribution of all discrete time steps.

[0065] The beneficial effects of the present invention are:

[0066] The present invention designs a point-by-point adaptive learning mechanism and vectorized discrete parameter gating, so that the model can quickly adapt to the dramatic changes of the reservoir under complex production dynamics through a physically driven training process without labeled data. This method shows excellent solution speed and convergence accuracy when dealing with complex oil and water well operations and changing production strategies. It is particularly suitable for reservoir conditions with significant heterogeneity and dynamic changes. The trained neural network can achieve an approximate format of the partial differential equation in the corresponding simulation scenario while solving each time step, and obtains good solution speed and convergence accuracy in complex oil and water well injection and production conversion, periodic injection and production, and continuous variable injection and production embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a flow chart of a complex dynamic reservoir simulation method based on an adaptive physical meaning neural network in an embodiment of the present invention.

[0068] Figure 2 This is a permeability distribution diagram of a heterogeneous oil reservoir in an embodiment of the present invention.

[0069] Figure 3 It is a relative permeability curve diagram of a complex dynamic reservoir in an embodiment of the present invention.

[0070] Figure 4 This is the pressure distribution diagram of the reservoir on the 500th day of injection development in the embodiment of the present invention. Among them, (a) is the predicted pressure field, (b) is the reference pressure field, and (c) is the relative error between the predicted pressure field and the reference pressure field.

[0071] Figure 5 This is the saturation distribution diagram of the reservoir on the 500th day of injection development in the embodiment of the present invention. Among them, (a) is the predicted saturation field, (b) is the reference saturation field, and (c) is the relative error between the predicted saturation field and the reference saturation field.

[0072] Figure 6 It is a comparison chart of the predicted value and the reference value of the bottom hole pressure of the injection well of the transfer and development oil reservoir in the embodiment of the present invention.

[0073] Figure 7 It is a comparison chart of the predicted value and the reference value of the bottom hole pressure of the production well of the transferred oil reservoir in the embodiment of the present invention.

[0074] Figure 8 It is a comparison chart of the predicted value and reference value of the water content of the production wells of the transferred oil reservoir in the embodiment of the present invention.

[0075] Fig. 9 This is the pressure distribution diagram of the 500th day of the periodic injection and production development of the reservoir in the embodiment of the present invention. Among them, (a) is the predicted pressure field, (b) is the reference pressure field, and (c) is the relative error between the predicted pressure field and the reference pressure field.

[0076] Fig.10 This is the saturation distribution diagram of the 500th day of the periodic injection and production development of the reservoir in the embodiment of the present invention. Among them, (a) is the predicted saturation field, (b) is the reference saturation field, and (c) is the relative error between the predicted saturation field and the reference saturation field.

[0077] Fig.11 It is a comparison chart of the predicted value and the reference value of the bottom hole pressure of the injection well of the periodic injection and production development oil reservoir in the embodiment of the present invention.

[0078] Fig.12 It is a comparison chart of the predicted value and the reference value of the bottom hole pressure of the production well in the periodic injection and production development of the oil reservoir in the embodiment of the present invention.

[0079] Fig.13 It is a comparison chart of the predicted value and the reference value of the water content of the production wells in the periodic injection and production development of the oil reservoir in the embodiment of the present invention.

[0080] Fig.14 This is the pressure distribution diagram of the 500th day of the continuous variable injection and production development of the reservoir in the embodiment of the present invention. Among them, (a) is the predicted pressure field, (b) is the reference pressure field, and (c) is the relative error between the predicted pressure field and the reference pressure field.

[0081] Fig.15 This is the saturation distribution diagram of the 500th day of the continuous variable injection and production development of the reservoir in the embodiment of the present invention. Among them, (a) is the predicted saturation field, (b) is the reference saturation field, and (c) is the relative error between the predicted saturation field and the reference saturation field.

[0082] Fig.16 It is a comparison chart between the predicted value and the reference value of the bottom hole pressure of the injection well of the continuous variable injection and production development oil reservoir in the embodiment of the present invention.

[0083] Fig.17 It is a comparison chart of the predicted value and the reference value of the bottom hole pressure of the production well in the continuous variable injection and production development of the oil reservoir in the embodiment of the present invention.

[0084] Fig.18 It is a comparison chart of the predicted value and the reference value of the water content of the production wells in the continuous variable injection and production development of the oil reservoir in the embodiment of the present invention. DETAILED DESCRIPTION

[0085] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention 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 invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0086] The specific implementation of the present invention is further described in detail below in conjunction with the drawings and examples.

[0087] The embodiment of the present invention provides a complex dynamic reservoir simulation method based on an adaptive physical meaning neural network, such as Figure 1 As shown, the method includes steps 1 to 5, which are described in detail as follows.

[0088] Step 1: Obtain complex production dynamic reservoir development data, including reservoir permeability distribution, fluid properties and complex production dynamic data.

[0089] In this embodiment, the purpose of step 1 is to collect complex production dynamic reservoir development data, collect the permeability distribution, fluid properties and complex production dynamic data of the reservoir, and provide basic data support for subsequent model training.

[0090] In some embodiments, the complex production performance data includes the positions of the injection well and the production well at each time step, the injection rate of the injection well at each time step, and the production rate of the production well at each time step.

[0091] Step 2: Establish an adaptive physical meaning neural network model, which includes a vectorized discrete parameter gating structure to enable the model to be independently adjusted at each data point; initialize model parameters and set a production control system table.

[0092] In this embodiment, the adaptive physical meaning neural network model established in step 2 allows the model to be adjusted independently at each data point to adapt to drastic changes in the local physical field by introducing an innovative vectorized discrete parameter gating structure; initialize the model parameters and set the production control system table to adapt to the dynamic changes of the reservoir.

[0093] In some embodiments, the adaptive physical meaning neural network model has three fully connected layers, including an input fully connected layer, a hidden fully connected layer, and an output fully connected layer; in addition, the adaptive physical meaning neural network model has innovative functional layers, including a vectorized discrete parameter gating update layer; the hierarchical order of each layer is as follows: the input fully connected layer is located at the front end of the neural network, the adjacent position anchoring layer is located after the input fully connected layer, the regularization layer is located after the adjacent position anchoring layer, the hidden fully connected layer is located after the adjacent position anchoring layer, the adaptive activation function layer is located after the hidden fully connected layer, the output fully connected layer is located after the adaptive activation layer, and the vectorized discrete parameter gating update layer is located after the output fully connected layer; the adaptive physical meaning neural network learns the pressure-to-pressure mapping between adjacent time steps.

[0094] In some embodiments, the specific mapping functions of each layer in the adaptive physical meaning neural network model are involved, and a series of network structures are innovatively designed to improve the adaptability of the neural network to flow field changes, enhance the convergence ability of the neural network in complex production dynamic simulations, and construct a pressure mapping using an adaptive physical meaning neural network. In this embodiment, the design steps of the specific mapping function of each layer include steps 2.1 to 2.6, which are described in detail as follows.

[0095] Step 2.1: Take the reservoir pressure distribution matrix as the input pressure matrix of the input fully connected layer, and design the mapping function of the input fully connected layer as follows:

[0096] (1)

[0097] in is the input fully connected layer mapping function, is the input fully connected layer pressure matrix, is the input fully connected layer weight, is the input fully connected layer bias.

[0098] Step 2.2: Use the pressure matrix output by the input fully connected layer as the input pressure matrix of the adjacent position anchor layer, and design the mapping function of the adjacent position anchor layer as follows:

[0099] (2)

[0100] in is the adjacent position anchor layer mapping function, is the pressure matrix of the adjacent anchor layer, is the anchor weight, is the adjacency position matrix, is the mean of the anchor pressure matrix, is the standard deviation of the anchor pressure matrix, is a small value that prevents the denominator from reaching zero, The weight of the adjacent position anchor layer, is the anchor layer deviation of the adjacent position.

[0101] Step 2.3: Use the pressure matrix output by the adjacent anchor layer as the input pressure matrix of the hidden fully connected layer, and design the mapping function of the hidden fully connected layer as follows:

[0102] (3)

[0103] in is the hidden fully connected layer mapping function, is the hidden fully connected layer pressure matrix, is the hidden fully connected layer weight, is the hidden fully connected layer bias.

[0104] Step 2.4: Use the pressure matrix output by the hidden fully connected layer as the input pressure matrix of the adaptive activation layer, and design the mapping function of the adaptive activation layer as follows:

[0105] (4)

[0106] in is the adaptive activation layer mapping function, is the adaptive activation layer pressure matrix, are the adaptive activation weights, is the maximum function:

[0107] Step 2.5: Use the output pressure matrix of the adaptive activation layer as the input pressure matrix of the output fully connected layer, and design the mapping function of the output fully connected layer as follows:

[0108] (5)

[0109] in is the output fully connected layer mapping function, Output the fully connected layer pressure matrix, is the output fully connected layer weight, is the output fully connected layer bias.

[0110] Step 2.6: Use the output pressure matrix of the output fully connected layer and the reservoir pressure matrix of the current time step as the input pressure matrix of the vectorized discrete parameter gating update layer. The mapping function of the vectorized discrete parameter gating update layer is innovatively designed as follows:

[0111] (6)

[0112] in is the vectorized discrete parameter gated update layer mapping function, is the gated update candidate pressure matrix, is the gating update weight matrix.

[0113] In some embodiments, in step 2, initializing model parameters and setting the production control system table specifically includes:

[0114] If the current moment is the initial moment, initialize the learnable parameters in the neural network: Initialize Xavier and set The default initial value is 0.01, set The default initial value is 0.1, set The default initial value of all elements is 0.9; divide the time step according to the actual production dynamics and set the production control system table.

[0115] Step 3: In response to the input initial reservoir pressure data, the internal boundary conditions are adjusted according to the production system table, the forward propagation mapping and the back propagation gradient calculation are performed, and the point-by-point adaptive update of the model parameters is achieved through the vectorized discrete parameter gating structure.

[0116] In this embodiment, based on the adaptive physical meaning neural network model, after the initial pressure data of the reservoir is input, the internal boundary conditions are adjusted according to the production system table, the forward propagation mapping and the back propagation calculation gradient are used, and the point-by-point adaptive update of the model parameters is achieved through the vectorized gating structure, which significantly improves the response speed and accuracy of the model to complex dynamic changes.

[0117] In some embodiments, step 3 specifically includes steps 3.1 to 3.4, which are described in detail as follows.

[0118] Step 3.1: Input the pressure matrix of the current time step into the neural network for forward propagation mapping, and output the predicted pressure matrix of the next time step. The complete mapping process is:

[0119] (7)

[0120] in is the reservoir pressure matrix at the current moment, The output of the neural network predicts the reservoir pressure matrix for the next time step; when the current moment is the initial moment, the initial values ​​of all elements of the initial reservoir pressure matrix are set to 1.0.

[0121] Step 3.2: Construct the finite volume method physical constraint loss function:

[0122] Combining the implicit pressure explicit saturation solution method, the residual formula of partial differential equation control in the context of slightly compressible two-phase flow is constructed as follows:

[0123] (8)

[0124] in is the residual of the partial differential equation, is the porosity, and are the saturations of the non-wetting phase and the wetting phase, respectively. and are the compressibility coefficients of the non-wetting phase and the wetting phase, is the compressibility coefficient of the flowing medium, and are the source and sink flow rates of the non-wetting phase and the wetting phase, respectively; To control the volume, for The pressure of the moment, for The pressure of the moment, represents the neighboring cells of the grid, and is the grid number, is the median value of the total flow of adjacent grids, is the conductivity, For Grid The wetting phase flow potential, For Grid The wetting phase flow potential, is the median value of the non-wetting phase fluidity between adjacent grids, For Grid The two-phase flow potential difference, For Grid The two-phase flow potential difference.

[0125] Read the source and sink terms of the next time step from the production and control system table, and use the following Peaceman well model to process the source and sink terms:

[0126] (9)

[0127] in is the permeability, is the formation perforation thickness, is the equivalent formation radius, is the wellbore radius, is the skin coefficient, is the bottom hole pressure.

[0128] The inner and outer boundary conditions are processed by hard constraints, and the residual of the physical constraint is:

[0129] (10)

[0130] in, Predicted grid The wetting phase flow potential, Predicted grid The wetting phase flow potential.

[0131] The calculation formula for the predicted flow potential is:

[0132] (11)

[0133] in, is the predicted flow potential, is the predicted pressure, is the density, is the acceleration due to gravity, It's depth.

[0134] The loss value is calculated using the mean square error, and the loss function of the physical constraint is:

[0135] (12)

[0136] in, is the loss value at time t+1, is the number of discrete grids, i is the grid number, It is predicted Constant pressure.

[0137] Step 3.3: Combined with vectorized discrete parameter gating, the innovative design uses the gradient accumulation adaptive optimizer for back propagation and updating of neural network parameters; if the current moment is the initial moment, set the default initial learning rate of the optimizer to 0.01, and set the default loss threshold for model convergence to 1.0×10 -18 ; The model parameter gradient calculation and update formula is:

[0138] (13)

[0139] in, and They are Iteration step and Neural network model parameters at the iteration step, is the learning rate, is a small value that prevents the denominator from reaching zero, is the residual calculated from the predicted value, is the output pressure of the hidden layer of the neural network, is the gating weight.

[0140] Step 3.4: Repeat the forward propagation prediction and backward propagation update process of steps 3.1 to 3.3 above until the loss value is reduced to below the preset model convergence loss threshold, the pressure solution is completed, and the pressure data for the next time step is output.

[0141] Step 4: Update the saturation according to the pressure data output by the model.

[0142] In this embodiment, the saturation is updated in combination with the pressure data output by the neural network to ensure that the data at each time step accurately reflects the state of the reservoir.

[0143] In some embodiments, step 4 specifically includes:

[0144] (14)

[0145] in, is the saturation distribution at the next time step, is the saturation data of the current time step, is the wetting phase fluidity.

[0146] Step 5: Based on the set simulation time, iteratively execute steps 3 and 4, use the pressure field output in each step and the trained neural network parameters as the input for the next step, and continue to solve the simulation process of the entire reservoir until the set simulation time is reached.

[0147] In some embodiments, step 5 specifically includes:

[0148] In the solution at the next moment, all the parameters of the neural network trained in the previous time step are used as the initialization parameters of the next time step; the pressure data solved in the previous time step is used as the input pressure data for forward propagation in the next time step; steps 3 to 4 are repeated to obtain the pressure distribution and saturation distribution of all discrete time steps, thus completing the numerical simulation solution of the reservoir based on the adaptive physical meaning neural network.

[0149] Therefore, the complex dynamic reservoir simulation method based on adaptive physical meaning neural network provided in this embodiment realizes the design of point-by-point adaptive learning mechanism and vectorized discrete parameter gating, so that the model can quickly adapt to the dramatic changes of reservoirs under complex production dynamics through a physically driven training process without labeled data. This method shows excellent solution speed and convergence accuracy when dealing with complex oil and water well operations and changing production strategies. It is particularly suitable for reservoir conditions with significant heterogeneity and dynamic changes. The trained neural network can achieve the approximate format of the partial differential equation in the corresponding simulation scenario while solving each time step, and obtains good solution speed and convergence accuracy in complex oil and water well injection and production conversion, periodic injection and production, and continuous variable injection and production embodiments. It has a significantly improved convergence effect in the complex production dynamics embodiment.

[0150] In order to prove the feasibility of the present invention, a specific example is given for a complex production dynamic model. The specific process of numerical simulation of complex production dynamic reservoir in the example is as follows:

[0151] Step 1: Collect the permeability distribution of complex dynamic reservoirs such as Figure 2 As shown, the permeability difference is 10; fluid related data are as follows Figure 3 As shown, the water phase is the wetting phase; well location data, including an injection well located at the center of the grid in the lower left corner and an injection well located at the center of the grid in the upper right corner; reservoir development and control system data, including three production dynamic data: transfer injection, periodic injection and production, and continuous variable injection and production. The simulated production time is 1000 days, and the time step is 0.01 day.

[0152] Step 2: Construct an adaptive physical meaning neural network. The model hyperparameters are set as follows: the input dimension of the input fully connected layer is 1D, the output dimension is 625D, and the number of neurons is 625; the input dimension of the hidden fully connected layer is 625D, the output dimension is 625D, and the number of neurons is 625; the input dimension of the output fully connected layer is 625D, the output dimension is 1D, and the number of neurons is 1; the initial value of the anchor weight is 0.01; the initial value of the activation weight is 0.1; the initial value of the discrete parameterized gating is 0.9; the initial value of the adaptive learning rate is 0.01; the model convergence loss threshold is 1×10 -18 .

[0153] Step 3: Input the pressure data of the current time step into the adaptive physical meaning neural network initialized in step 2, forward propagate the output pressure distribution data, read the production dynamic data from the development and control system table, calculate the residual of each grid through the physical constraint loss function, calculate the loss value using the mean square error loss function, calculate the gradient of each learnable parameter through back propagation, and update the neural network parameters using the adaptive optimizer. By performing forward propagation and back propagation multiple times until the loss value is reduced to below the convergence loss threshold, the pressure solution for the next time step is completed.

[0154] Step 4: Calculate the saturation of the next time step explicitly through the implicit pressure and explicit saturation decoupling formula, calculate and record the bottom hole pressure data of the source and sink items using the well model, and calculate and record the water content data of the production well.

[0155] Step 5. Repeat steps 3 and 4 to obtain the saturation distribution and pressure distribution of the complex production dynamic reservoir at all time steps. At this point, the pressure field distribution, saturation field distribution and source-sink production dynamics of all time steps are solved.

[0156] In this example, the pressure distribution of the injection reservoir obtained by the adaptive physical meaning neural network method of the present invention is compared with the reference value. The results are as follows: Figure 4 As shown; the saturation distribution of the reservoir solved by the new method is compared with the reference value, and the results are as follows Figure 5 As shown; the bottom hole pressure of the injection well of the reservoir solved by the new method is compared with the reference value, and the results are as follows Figure 6 As shown; the bottom hole pressure of the production well of the oil reservoir solved by the new method is compared with the reference value, and the results are as follows Figure 7 As shown; the water content of the production wells of the reservoir solved by the new method is compared with the reference value, and the results are as follows Figure 8 shown.

[0157] In this example, the pressure distribution of the periodic injection and production reservoir obtained by the adaptive physical meaning neural network method of the present invention is compared with the reference value. The results are as follows: Fig. 9 As shown; the saturation distribution of the reservoir solved by the new method is compared with the reference value, and the results are as follows Fig.10 As shown; the bottom hole pressure of the injection well of the reservoir solved by the new method is compared with the reference value, and the results are as follows Fig.11 As shown; the bottom hole pressure of the production well of the oil reservoir solved by the new method is compared with the reference value, and the results are as follows Fig.12 As shown; the water content of the production wells of the reservoir solved by the new method is compared with the reference value, as shown in Fig.13 shown.

[0158] In this example, the pressure distribution of the continuously variable injection and production reservoir obtained by the adaptive physical meaning neural network method of the present invention is compared with the reference value. The results are as follows: Fig.14 As shown; the saturation distribution of the reservoir solved by the new method is compared with the reference value, and the results are as follows Fig.15 As shown; the bottom hole pressure of the injection well of the reservoir solved by the new method is compared with the reference value, and the results are as follows Fig.16 As shown; the bottom hole pressure of the production well of the oil reservoir solved by the new method is compared with the reference value, and the results are as follows Fig.17 As shown; the water content of the production wells of the reservoir solved by the new method is compared with the reference value, as shown in Fig.18 shown.

[0159] The above implementation modes are only used to illustrate the present invention, but not to limit the present invention. Ordinary technicians in the relevant technical field can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions also belong to the scope of the present invention. The patent protection scope of the present invention should be defined by the claims.

Claims

1. A complex dynamic reservoir simulation method based on adaptive physical meaning neural network, characterized in that: The method comprises: Step 1: Obtaining complex production dynamic reservoir development data, including reservoir permeability distribution, fluid properties and complex production dynamic data; Step 2: Establish an adaptive physical meaning neural network model, wherein the adaptive physical meaning neural network model includes a vectorized discrete parameter gating structure so that the model can be adjusted independently at each data point; initialize the model parameters, and set the production control system table; Step 3: In response to the inputted initial pressure data of the reservoir, the internal boundary conditions are adjusted according to the production system table, the forward propagation mapping and the back propagation gradient calculation are performed, and the point-by-point adaptive update of the model parameters is achieved through the vectorized discrete parameter gating structure; Step 4: Update the saturation according to the pressure data output by the model; Step 5: Based on the set simulation time, iteratively execute steps 3 and 4, use the pressure field output in each step and the trained neural network parameters as the input for the next step, and continuously solve the simulation process of the entire reservoir until the set simulation time is reached; The adaptive physical meaning neural network model includes an input fully connected layer, a hidden fully connected layer, an output fully connected layer, an adjacent position anchoring layer, a regularization layer, an adaptive activation function layer and a vectorized discrete parameter gating update layer; wherein the input fully connected layer is located at the front end of the neural network, the adjacent position anchoring layer is located after the input fully connected layer, the regularization layer is located after the adjacent position anchoring layer, the hidden fully connected layer is located after the adjacent position anchoring layer, the adaptive activation function layer is located after the hidden fully connected layer, the output fully connected layer is located after the adaptive activation layer, and the vectorized discrete parameter gating update layer is located after the output fully connected layer; the adaptive physical meaning neural network model learns the mapping of pressure to pressure between adjacent time steps.

2. The method according to claim 1, characterized in that The complex production dynamic data include the positions of the injection well and the production well at each time step, the injection rate of the injection well at each time step, and the production rate of the production well at each time step.

3. The method according to claim 1, characterized in that The mapping functions of the input fully connected layer, adjacent position anchor layer, hidden fully connected layer, adaptive activation function layer, output fully connected layer, and vectorized discrete parameter gated update layer are determined by the following method: The reservoir pressure distribution matrix is ​​used as the input pressure matrix of the input fully connected layer, and the mapping function of the input fully connected layer is designed as follows: f1(P0)=w1·P0+b1 (1) Where f1 is the input fully connected layer mapping function, P0 is the input fully connected layer pressure matrix, w1 is the input fully connected layer weight, and b1 is the input fully connected layer bias; The pressure matrix output by the input fully connected layer is used as the input pressure matrix of the adjacent position anchor layer, and the mapping function of the adjacent position anchor layer is designed as follows: Where f2 is the adjacent position anchor layer mapping function, P1 is the adjacent position anchor layer pressure matrix, ω1 is the anchor weight, m adj is the adjacency position matrix, μ B is the mean of the anchor pressure matrix, σ B is the standard deviation of the anchor pressure matrix, ∈1 is a small value to prevent the denominator from being zero, γ is the weight of the adjacent position anchor layer, and ζ is the adjacent position anchor layer bias; The pressure matrix output by the adjacent anchor layer is used as the input pressure matrix of the hidden fully connected layer, and the mapping function of the hidden fully connected layer is designed as follows: f3(P2)=w2·P2+b2 (3) Where f3 is the hidden fully connected layer mapping function, P2 is the hidden fully connected layer pressure matrix, w2 is the hidden fully connected layer weight, and b2 is the hidden fully connected layer bias; The pressure matrix output by the hidden fully connected layer is used as the input pressure matrix of the adaptive activation layer, and the mapping function of the adaptive activation layer is designed as follows: f4(P3)=max(0,α·P3) (4) Where f4 is the adaptive activation layer mapping function, P3 is the adaptive activation layer pressure matrix, α is the adaptive activation weight, and max(·) is the maximum value function; The output pressure matrix of the adaptive activation layer is used as the input pressure matrix of the output fully connected layer, and the mapping function of the output fully connected layer is designed as follows: f5(P4)=w3·P4+b3 (5) Among them, f5 is the output fully connected layer mapping function, P4 is the output fully connected layer pressure matrix, w3 is the output fully connected layer weight, and b3 is the output fully connected layer bias; The output pressure matrix of the output fully connected layer and the reservoir pressure matrix of the current time step are used as the input pressure matrix of the vectorized discrete parameter gating update layer, and the mapping function of the vectorized discrete parameter gating update layer is designed as follows: f6(P0,P5)=in g ⊙P0+(1-in g )⊙P5 (6) Where f6 is the vectorized discrete parameter gated update layer mapping function, P5 is the gated update candidate pressure matrix, and w g is the gating update weight matrix.

4. The method according to claim 3, characterized in that In step 2, the model parameters are initialized and the production control system table is set, including: If the current moment is the initial moment, initialize the learnable parameters in the neural network: perform Xavier initialization on w1, w2, w3, γ, b1, b2, b3, ζ, set the default initial value of ω1 to 0.01, set the default initial value of α to 0.1, set w g The default initial value of all elements is 0.9; divide the time step according to the actual production dynamics and set the production control system table.

5. The method according to claim 3, characterized in that The step 3, in response to the inputted initial pressure data of the reservoir, adjusts the internal boundary conditions according to the production system table, forward propagates the mapping and backward propagates the gradient, and realizes point-by-point adaptive updating of the model parameters through the vectorized discrete parameter gating structure, including: Step 3.1: Input the pressure matrix of the current time step into the adaptive physical meaning neural network model for forward propagation mapping, and output the predicted pressure matrix of the next time step. The complete mapping process is: Where P t is the reservoir pressure matrix at the current moment, The reservoir pressure matrix for the next time step predicted by the output neural network; Step 3.2: Construct the finite volume method physical constraint loss function: The residual formula for partial differential equation control in the context of slightly compressible two-phase flow is: Where R i is the residual of the partial differential equation, φ is the porosity, S n and S w are the saturation of the non-wetting phase and the wetting phase, respectively, and C n and C w are the compressibility coefficients of the non-wetting phase and the wetting phase, C r is the compressibility coefficient of the flowing medium, q n and q w are the source and sink flow rates of the non-wetting phase and the wetting phase respectively; V i To control the volume, is the pressure at time t+1, is the pressure at time t, M i represents the adjacent cells of the grid, i and j are the grid numbers, λ t,ij+1 / 2 is the median value of the total flow of adjacent grids, T ij is the conductivity, is the wetting phase flow potential of grid j, is the wetting phase flow potential of grid i, λ n,ij+1 / 2 is the median value of the non-wetting phase fluidity between adjacent grids, is the two-phase flow potential difference at grid j, is the two-phase flow potential difference at grid i; Read the source and sink terms of the next time step from the production and control system table, and use the following Peaceman well model to process the source and sink terms: Where q is the source-sink flow, λ is the mobility, K is the permeability, h is the formation penetration thickness, and r e is the equivalent formation radius, r w is the borehole radius, s is the skin factor, P is the grid pressure, P wf is the bottom hole pressure; The inner and outer boundary conditions are processed by hard constraints, and the residual of the physical constraint is: in, The predicted wetting phase flow potential of grid j is: Predicted wetting phase flow potential of grid i; The calculation formula for predicting flow potential is: in, is the predicted flow potential, is the predicted pressure, ρ is the density, g is the acceleration due to gravity, and D is the depth; The loss value is calculated using the mean square error, and the loss function of the physical constraint is: in, is the loss value at time t+1, N g is the number of discrete grids, i is the grid number, is the predicted pressure at time t+1; Step 3.3: If the current moment is the initial moment, set the default initial learning rate of the optimizer to 0.01 and the default loss threshold for model convergence to 1.0×10 -18 ; The model parameter gradient calculation and update formula is: Among them, θ τ+1 and θ τ are the neural network model parameters of the τ+1th iteration step and the τth iteration step, η is the learning rate, ∈2 is a small value to prevent the denominator from being zero, is the residual calculated from the predicted value, is the output pressure of the hidden layer of the neural network, w g,i,τ is the gating weight; Step 3.4, repeat the forward propagation prediction and backward propagation update process of step 3.1 to step 3.3 until the loss value is reduced to below the preset model convergence loss threshold, the pressure solution is completed, and the pressure data for the next time step is output.

6. The method according to claim 5, characterized in that The step 4, updating the saturation according to the pressure data output by the model, includes: According to the output pressure data of the next time step and the saturation data of the current time step, the saturation distribution of the next time step is calculated by the following formula: in, is the saturation distribution at the next time step, is the saturation data of the current time step, λ w is the wetting phase fluidity.

7. The method according to claim 5, characterized in that The step 5, based on the set simulation time, iteratively executes the steps 3 and 4, uses the pressure field outputted in each step and the trained neural network parameters as the input of the next step, and continuously solves the simulation process of the entire reservoir until the set simulation time is reached, including: In the solution at the next moment, all the parameters of the neural network trained in the previous time step are used as the initialization parameters of the next time step; the pressure data solved in the previous time step is used as the input pressure data for forward propagation in the next time step; steps 3 to 4 are repeated to solve the pressure distribution and saturation distribution of all discrete time steps.

Images