A Method for Simulating the Flow of CO2 Flooding in Fractured Reservoirs Using Neural Network Technology

Abstract

The present invention relates to a method for simulating the flow of CO2 flooding in a fractured reservoir, and particularly to a method for simulating the flow of CO2 flooding in a fractured reservoir using neural network technology. The method comprises the following steps: conducting a survey on the target reservoir, collecting reservoir geological parameters and fracture geometric parameters, and establishing a geometric model of the fractured reservoir; establishing a flash calculation model based on a neural network; establishing an embedded discrete fracture model to describe fractures; establishing a component model to describe the process of CO2 flooding; combining the flash calculation model based on the neural network model with the component model, and considering the embedded discrete fracture model, establishing a flow model for CO2 flooding in the fractured reservoir and performing numerical solution. The method of the present invention can utilize the advantages of the neural network to perform flash calculation quickly and stably, improve the convergence speed of the calculation, and accurately describe the influence of fractures on CO2 flooding.

Description

Technical Field

[0001] The present invention relates to a method for simulating the flow of CO2 flooding in a fractured reservoir, and particularly to a method for simulating the flow of CO2 flooding in a fractured reservoir using neural network technology. Background Art

[0002] In China, the resources of low-permeability and tight oil reservoirs are rich and have become the main growth point of reserves. However, the physical properties of the reservoirs in such oil and gas reservoirs are poor. Affected by factors such as deep burial, small well control range, and low abundance, conventional water flooding development is difficult to utilize, with no natural production capacity or low production; for the developed reserves, the single-well capacity is low and the recovery rate is low, and there is an urgent need to seek new methods to improve the recovery rate.

[0003] The technology of injecting CO2 for oil displacement is a relatively ideal method. Field gas injection experiments show that the gas injection capacity is more than 5 times that of water, and the reservoir energy can be quickly supplemented. Therefore, injecting CO2 underground can effectively supplement the formation energy and maintain the underground pressure; at the same time, CO2 is an excellent oil displacement agent, and injecting CO2 for oil displacement has been proven to be an effective method to improve the recovery rate. CO2 has the characteristics of low viscosity and easy injection. After contacting the crude oil in the formation, it can cause the volume of the crude oil to expand, reduce the viscosity of the crude oil, extract light components, reduce the interfacial tension between oil and water, and improve the mobility ratio, etc., which can effectively improve the crude oil recovery rate.

[0004] Numerical simulation technology is an important technology for understanding the dynamics of CO2 flooding and making many decisions such as formulating measures to improve the recovery rate of CO2. The component model can accurately describe the process of CO2 flooding. One of the key steps of this method is to calculate the phase state and component conditions using the flash calculation method. However, the flash calculation has a large amount of calculation, resulting in the component simulation often taking a long time. For this reason, experts and scholars at home and abroad have also proposed phase stability methods, etc. to increase the reliability of the calculation.

[0005] In recent years, the technology of artificial neural networks has made great progress and has been applied to many disciplines, achieving great achievements. Like biological nerves, artificial neural networks must go through learning to have intelligent characteristics. Its learning process is actually a process of adjusting the weights and thresholds. Deep learning is a branch of artificial neural network algorithms, and it has made remarkable research progress in many fields such as image recognition, aerospace, and computer translation. In recent years, deep learning technology has also been gradually applied to solving partial differential equations and the field of reservoir component numerical simulation.

[0006] Fractures, as the smallest geological structures, are widely distributed in the earth's crust, and fractures of different scales exist in almost all oil reservoirs. Moreover, during the exploitation of medium and low permeability oil reservoirs, in order to improve the oil production capacity of production wells or the water injection (gas injection) capacity of injection wells, artificial fracturing is often carried out, and artificial fracturing will generate fracturing fractures of different scales. Therefore, fractures will inevitably be encountered during the implementation of CO2 flooding. Fractures will significantly affect the pressure distribution, component distribution, saturation distribution, etc. during CO2 flooding. Therefore, in numerical simulation, it is necessary to accurately simulate fractures.

[0007] The simulation methods of fractures mainly include three flow mathematical models based on dual media, equivalent continuous media, and discrete fractures. The dual media model assumes that the fracture system is evenly distributed in a network shape and cannot truly reflect the randomness and multi-scale nature of fractures; the equivalent continuous media model regards the entire medium as a continuous system and characterizes its heterogeneity through equivalent parameters. The above two models are not applicable to the fine flow simulation of fractured oil reservoirs. The discrete fracture model performs dimensionality reduction processing on fracture units. Taking 2D fracture units as an example, based on the equivalent principle, the discrete fracture model uses 1D line units combined with fracture apertures to represent 2D fracture units. Therefore, no network meshing is required inside the fractures during discretization, saving computational resources. However, for oilfield-level models, the discrete fracture model still has problems such as complex mesh generation and large computational volume. Summary of the Invention

[0008] In order to overcome the difficulties faced by the existing CO2 flooding methods for fractured oil reservoirs, the present invention provides a CO2 flooding simulation method for fractured oil reservoirs using neural network technology. The method of the present invention can utilize the advantages of neural networks to perform flash calculations quickly and stably, improve the convergence speed of calculations, and accurately describe the influence of fractures on CO2 flooding.

[0009] To achieve the above object, the present invention adopts the following technical solutions:

[0010] The present invention provides a CO2 flooding flow simulation method for fractured oil reservoirs using neural network technology, which includes the following steps: conducting a survey of the target oil reservoir, collecting oil reservoir geological parameters and fracture geometric parameters, and establishing a geometric model of the fractured oil reservoir; establishing a flash calculation model based on neural networks; establishing an embedded discrete fracture model to describe fractures; establishing a component model to describe the CO2 flooding process; combining the flash calculation model based on the neural network model with the component model, and considering the embedded discrete fracture model, establishing a CO2 flooding flow model for fractured oil reservoirs and performing numerical solution.

[0011] Preferably, the oil reservoir geological parameters include the oil reservoir scale, shape, permeability, and porosity; the fracture geometric parameters include the number of fractures, azimuth, strike, as well as fracture length, aperture, and permeability.

[0012] Preferably, the method for establishing a neural network-based flash calculation model includes the following steps:

[0013] Establish a neural network model;

[0014] Use a large amount of data to train the neural network to predict the phase state;

[0015] Use a large amount of data to train the neural network to calculate the component mole fraction and capillary force parameters;

[0016] Establish a neural network-based flash calculation model: Initialize the equilibrium ratio K with the neural network model trained with a large amount of data i , and use this K i to solve the Rachford-Rice equation; if the neural network model predicts a single phase, the prediction result of the neural network can be directly used as the result of the flash calculation; if the neural network prediction result is the existence of gas-liquid two phases, the equilibrium ratio K predicted by the neural network needs to be used i for flash calculation.

[0017] Compared with initializing the equilibrium ratio K with the Wilson equation in the traditional method using the neural network trained with a large amount of data i , the prediction accuracy of the equilibrium ratio K i is higher.

[0018] Further preferably, the neural network model includes one input layer, one output layer, and four hidden layers, with 60 neurons in each layer. Neurons between adjacent layers are interconnected and transmit information through a non-linear function.

[0019] Further preferably, use a large amount of data to train the neural network to predict the phase state: Calculate and store a large amount of data through the conventional flash calculation method as the data for training the neural network; the neural network includes one input layer, one output layer, and four hidden layers, with 60 neurons in each layer;

[0020] If the sample contains a total of N c components, then let the number of input parameters be N c +3, and the input parameter X includes the component concentration z i , i = 1,..., N c , pressure P, temperature T, and pore radius r. X can be written as X = [P, T, r, z1,..., z Nc T ; the dimension of the output layer is 3. Therefore, the result output by the neural network has three possibilities, namely gas phase, liquid phase, and coexistence of gas phase and liquid phase;

[0021] The activation function used during phase state recognition training is the ReLU function: ​

[0022]

[0023] The activation function of the output layer is the Softmax function:

[0024]

[0025] Further preferably, the steps of training the neural network with a large amount of data to calculate the component mole fraction and the capillary force parameter specifically include:

[0026] Calculate a large amount of data through the conventional flash calculation method and store it as the data for training the neural network; the neural network includes one input layer, one output layer, and four hidden layers, with 60 neurons in each layer; the sample contains a total of N c components, then let the number of input parameters be N c +3, and the input parameter X includes the component concentration z i , i = 1, …, N c , the pressure P, the temperature T, and the pore radius r. X can be written as X = [P, T, r, z1, …, z Nc T ; the dimension of the output parameter is 2N c +3, and the output parameters include the capillary pressure P C , the gas-phase mole fraction F v , the liquid-phase mole fraction F L , and the component mole fraction y i in the gas phase, i = 1, …N c , the component mole fraction x i in the liquid phase, i = 1, …, N c ,

[0027] The activation function is the ReLU function

[0028]

[0029] The activation function of the output layer is:

[0030]

[0031] Further preferably, the specific steps of using flash calculation include:

[0032] There are gas-liquid two phases in the CO2 flooding process, and the equilibrium of component i in the system satisfies the thermo-dynamic equilibrium:

[0033]

[0034] Where It represents the fugacity coefficient, which can be calculated from data such as component composition, pressure, temperature, and volume. First, the initial value K of the gas-liquid equilibrium constant is predicted by the trained neural network. i , and substitute K i into the Rachford-Rice equation to solve for the mole fraction F of the gas phase components. v

[0035]

[0036] Then, further solve to obtain the mole fraction parameters of the components. The temperature T, pressure P, and volume V parameters can be calculated by the PR equation of state:

[0037]

[0038] During the calculation process, iteration is achieved by detecting whether the equations satisfy the mass conservation condition and the thermodynamic equilibrium condition. The equilibrium ratio is updated during the iteration process until the conservation condition is met and the iteration ends.

[0039] Preferably, an embedded discrete fracture model is established as follows: The embedded discrete fracture model directly embeds the fracture system into the grid. After grid meshing, the fractures are embedded into the grid, and then the intersection relationship between the fracture system and the grid system is calculated; the fracture system and the bedrock are related according to the conservation law; the mass conservation equation of the embedded discrete model is:

[0040]

[0041] Among them, k F is the permeability of the large fracture, p F is the pressure of the large fracture, μ is the fluid viscosity, q F is the source-sink term of the large fracture, q Ff is the cross-flow rate between the two systems, q FF represents the cross-flow rate between the intersecting fracture elements, δ FF represents whether the large fracture element intersects with other fracture elements. If the large fracture element intersects with other fracture elements, δ FF = 1, otherwise δ FF = 0, V F are the volumes of the fracture elements respectively;

[0042] Further preferably, the steps of grid meshing specifically include: First, for the entire reservoir, since the fractures are embedded and do not need to be used as the inner boundary of the bedrock grid, the reservoir can be directly meshed, and then the fracture system is directly embedded into the reservoir grid to calculate the intersection information between the fractures and the reservoir grid and the geometric information of the fractures after being meshed by the reservoir grid;

[0043] Further preferably, for a reservoir with regular shape, orthogonal grid discretization is adopted. According to the size of the research area and the required calculation accuracy for the research, the grid step size and quantity in each spatial direction are determined. The grid shape of the orthogonal grid is generally a regular shape, which is convenient for calculation;

[0044] Further preferably, the intersection relationship between the fracture system and the grid system includes the intersection coordinates of the fractures and the grid, and the situation of the fractures being cut by the grid boundaries;

[0045] Further preferably, calculate the parameters of the embedded discrete fracture model: determined according to the actual data of the oilfield. For natural fractures, it is necessary to determine their fracture spatial coordinates, number, and spacing. Then, on this basis, calculate the grid cells passed through by each fracture and the connection information between the fractures and the grid cells, including the contact area between the fractures and the grid, and the average distance from the grid where the fracture is located to the fracture surface, so as to prepare for the subsequent simulation of component flow in the hybrid model.

[0046] Preferably, the method for establishing a CO2 flooding flow model for a fractured reservoir includes the following steps:

[0047] Establish the conservation equations for each component, and combine the embedded discrete fracture model and the flash calculation technology based on neural network to form a CO2 flooding flow model for a fractured reservoir; among them, establish the mass conservation equation for each component

[0048]

[0049] Where i represents the component, i = 1,..., nc, and nc represents the number of components. F is the mass flow rate, q represents the source-sink term, and N is the mass change; N can be expressed by the phase molar density ρ, saturation S, and the mole fraction x of component i in the oil phase i , and the mole fraction y of component i in the gas phase i is expressed as

[0050]

[0051] F can be written as

[0052]

[0053] Where v β is the Darcy velocity of phase β

[0054] Each component should satisfy the component mass conservation condition

[0055]

[0056] Preferably, the specific steps for numerically solving the CO2 flooding model for a fractured reservoir based on neural network include:

[0057] For numerical solution, it is first necessary to discretize the control values. The time term is discretized using the backward Euler method. To ensure the stability of the calculation, an implicit discretization scheme is adopted, a Jacobian matrix is constructed, and the iterative method is used for solution. To reduce the matrix dimension, the equations in the CO2 flooding mathematical model are divided into main equations and auxiliary equations, and the variables are divided into main variables and auxiliary variables. During the numerical solution process, the main variables are first obtained by solving the main equations, and then the auxiliary variables are obtained through the auxiliary equations.

[0058] Compared with the prior art, the present invention has the following advantages:

[0059] Aiming at the problems of large computational consumption when the flash calculation technology is used for gas-liquid equilibrium calculation in the component model and inaccurate description of fractures in the CO2 flooding numerical simulation of the present invention, a flash calculation model based on neural network technology is established to reduce the number of iterations in the flash calculation, improve the prediction accuracy of the equilibrium ratio, and improve the calculation efficiency and stability of the calculation. The present invention uses an embedded discrete fracture model to characterize large fractures, performs dimensionality reduction processing on the fractures, reduces the computational complexity, and directly embeds the fractures into the grid system, avoiding complex grid meshing and reducing the computational amount.

[0060] The present invention provides an efficient simulation method for implementing CO2 flooding in oilfields, makes full use of the advantages of big data and artificial intelligence, efficiently realizes the numerical simulation of CO2 flooding in fractured reservoirs, and provides technical support for the efficient development of oil and gas reservoirs. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The accompanying drawings forming a part of this invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention.

[0062] Figure 1 It is a flowchart of an embodiment of the CO2 flooding flow simulation method for fractured reservoirs provided by the present invention;

[0063] Figure 2 It is a schematic diagram of the neural network model provided by the present invention;

[0064] Figure 3 It is a grid schematic diagram of the embedded discrete fracture model of a CO2 flooding flow simulation method provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0065] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0066] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, and / or combinations thereof.

[0067] In order to enable those skilled in the art to more clearly understand the technical solution of the present invention, the technical solution of the present invention will be described in detail below with reference to specific embodiments.

[0068] As Figure 1 shown, the CO2 flooding flow simulation method for fractured reservoirs using neural network technology includes the following steps:

[0069] S1. Conduct a survey of the target reservoir, collect reservoir geological parameters and fracture geometric parameters, and establish a geometric model of the fractured reservoir; the reservoir geological parameters include reservoir scale, shape, permeability, and porosity; the fracture geometric parameters include the number, orientation, strike of fractures, as well as fracture length, aperture, and permeability.

[0070] S2. Establish a flash calculation model based on a neural network, as Figure 2 shown, including the following steps:

[0071] (1) Establish a neural network model: The neural network model includes one input layer, one output layer, and four hidden layers, with 60 neurons in each layer. Neurons between adjacent layers are interconnected and transmit information through a non-linear function.

[0072] (2) Use a large amount of data to train the neural network to predict the phase state: Calculate and store a large amount of data through a conventional flash calculation method as the data for training the neural network; the neural network includes one input layer, one output layer, and four hidden layers, with 60 neurons in each layer;

[0073] The sample contains a total of N c components, so let the number of input parameters be N c +3. The input parameter X includes component concentration z i , i = 1,..., N c , pressure P, temperature T, and pore radius r. X can be written as X = [P, T, r, z1,..., z Nc T ; the output layer dimension is 3. Therefore, the result output by the neural network has three possibilities, namely gas phase, liquid phase, and coexistence of gas phase and liquid phase;

[0074] The activation function used during phase state recognition training is the ReLU function: ​

[0075]

[0076] The activation function of the output layer is the Softmax function:

[0077]

[0078] (3) Train the neural network using a large amount of data to calculate parameters such as component mole fractions and capillary pressures: Calculate and store a large amount of data through conventional flash calculation methods as the data for training the neural network; the neural network includes one input layer, one output layer, and four hidden layers, with 60 neurons in each layer; the sample contains a total of N c components, then let the number of input parameters be N c +3, and the input parameter X includes the component concentration z i , i = 1, …, N c , pressure P, temperature T, and pore radius r. X can be written as X = [P, T, r, z1, …, z Nc T ; the dimension of the output parameter is 2N c +3, and the output parameters include capillary pressure P C , gas-phase mole fraction F v , liquid-phase mole fraction F L , and the component mole fractions y i in the gas phase, i = 1, …N c , the component mole fractions x i in the liquid phase, i = 1, …, N c ,

[0079] The activation function is the ReLU function

[0080]

[0081] The activation function of the output layer is:

[0082]

[0083] (4) Establish a flash calculation model based on the neural network: Initialize the equilibrium ratio K using the neural network model trained with a large amount of data i , and use this K i to solve the Rachford-Rice equation; if the neural network model predicts that the phase state is single-phase, the prediction result of the neural network can be directly used as the result of the flash calculation; if the neural network prediction result is that there are gas-liquid two phases, the equilibrium ratio K predicted by the neural network needs to be used i for flash calculation.

[0084] The specific steps for flash calculation include: ​

[0085] During the CO2 flooding process, there are gas-liquid two phases, and the equilibrium of component i in the system satisfies thermo-dynamic equilibrium:

[0086]

[0087] where represents the fugacity coefficient, which can be calculated from data such as component composition, pressure, temperature, and volume. First, the initial value K of the gas-liquid equilibrium constant is predicted by the trained neural network i , and K i is substituted into the Rachford-Rice equation to solve for the mole fraction F of the gas phase components v

[0088]

[0089] Then, the mole fraction parameters of the components are further solved. The temperature T, pressure P, and volume V parameters can be calculated by the PR equation of state:

[0090]

[0091] During the calculation process, iteration is achieved by detecting whether the equations satisfy the mass conservation condition and the thermo-dynamic equilibrium condition. During the iteration process, the equilibrium ratio is updated until the conservation condition is met and the iteration ends.

[0092] S3. Establish an embedded discrete fracture model to describe fractures; the embedded discrete fracture model directly embeds the fracture system into the grid, as Figure 3 shown. After grid meshing, the fractures are embedded into the grid, and then the intersection relationship between the fracture system and the grid system is calculated, including the intersection coordinates of the fractures and the grid, and the cutting situation of the fractures by the grid boundaries; the fracture system is connected to the bedrock according to the conservation law.

[0093] The mass conservation equation of the embedded discrete model is:

[0094]

[0095] where k F is the permeability of the large fracture, p F is the pressure of the large fracture, μ is the fluid viscosity, q F is the source-sink term of the large fracture, q Ff is the cross-flow rate between the two systems, q FF represents the cross-flow rate between the intersecting fracture elements, δ FF indicates whether the large fracture element intersects with other fracture elements. If the large fracture element intersects with other fracture elements, δ FF = 1, otherwise δ FF = 0, V F are the volumes of the fracture elements respectively;

[0096] The steps of grid generation specifically include: First, for the entire reservoir, since the fractures are embedded and do not need to be used as the inner boundary of the bedrock grid, the reservoir can be directly grid-generated, and then the fracture system is directly embedded into the reservoir grid to calculate the intersection information between the fractures and the reservoir grid and the geometric information of the fractures after being grid-generated by the reservoir grid.

[0097] For reservoirs with regular shapes, orthogonal grid generation is adopted. According to the size of the research area and the required calculation accuracy for the research, the grid step sizes and quantities in each spatial direction are determined. The grid shapes of orthogonal grids are generally regular shapes, which are convenient for calculation.

[0098] Calculate the parameters of the embedded discrete fracture model: Determine according to the actual data of the oilfield. For natural fractures, it is necessary to determine their fracture spatial coordinates, numbers, and spacings. Then, based on this, calculate the grid cells traversed by each fracture and the connection information between the fractures and the grid cells, including the contact area between the fractures and the grid, and the average distance from the grid where the fracture is located to the fracture surface, so as to prepare for the subsequent simulation of component flow in the hybrid model.

[0099] S4. Establish a component model to describe the CO2 flooding process.

[0100] S5. Combine the flash calculation model based on the neural network model with the component model, and consider the embedded discrete fracture model to establish a CO2 flooding flow model for fractured reservoirs and perform numerical solution.

[0101] A method for establishing a CO2 flooding flow model for fractured reservoirs includes the following steps:

[0102] Establish the conservation equations for each component, and combine the embedded discrete fracture model and the flash calculation technology based on the neural network to form a CO2 flooding flow model for fractured reservoirs; among them, establish a mass conservation equation for each component

[0103]

[0104] Where i represents the component, i = 1,..., nc, and nc represents the number of components. F is the mass flow rate, q represents the source-sink term, and N is the mass change; N can be expressed by the phase molar density ρ, saturation S, and the molar fraction x of component i in the oil phase i , and the molar fraction y of component i in the gas phase i is expressed as

[0105]

[0106] F can be written as

[0107]

[0108] where v β is the Darcy velocity of phase β

[0109] Each component should satisfy the component mass conservation condition

[0110]

[0111] The specific steps for numerically solving the CO2 flooding model of a fractured reservoir based on a neural network include:

[0112] For numerical solution, it is first necessary to discretize the control values. The time term is discretized using the backward Euler method. To ensure the stability of the calculation, an implicit discretization scheme is adopted, a Jacobian matrix is constructed, and the iterative method is used to solve it. To reduce the matrix dimension, the equations in the CO2 flooding mathematical model are divided into main equations and auxiliary equations, and the variables are divided into main variables and auxiliary variables. During the numerical solution process, the main variables are first obtained by solving the main equations, and then the auxiliary variables are obtained through the auxiliary equations.

[0113] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and are all included in the protection scope of the present invention.

Claims

1. A method for simulating the flow of CO2 flooding in fractured reservoirs using neural network technology, characterized in that, The method includes the following steps: surveying a target reservoir, collecting reservoir geological parameters and fracture geometric parameters, and establishing a fractured reservoir geometric model; establishing a flash calculation model based on a neural network; establishing an embedded discrete fracture model to describe fractures; establishing a component model to describe the CO2 flooding process; combining the flash calculation model based on the neural network model with the component model, and considering the embedded discrete fracture model, establishing a CO2 flooding flow model for the fractured reservoir and performing numerical solution; The method for establishing a flash calculation model based on a neural network includes the following steps: establishing a neural network model; using a large amount of data to train the neural network to predict phase conditions; using a large amount of data to train the neural network to calculate component mole fractions and capillary force parameters; Establish a neural network-based flash calculation model: Initialize the equilibrium ratio K with a neural network model trained with a large amount of data i , and use this K i to solve the Rachford-Rice equation; if the neural network model predicts a single phase, directly use the prediction result of the neural network as the result of the flash calculation; if the neural network prediction result is that there are gas-liquid two phases, it is necessary to use the equilibrium ratio K predicted by the neural network i to perform flash calculation; Establishing a neural network model includes one input layer, one output layer, and four hidden layers, with 60 neurons in each layer. Neurons between adjacent layers are interconnected and transmit information through a non-linear function; Using a large amount of data to train the neural network to predict phase conditions: through flash calculation, calculating a large amount of data and storing it as data for training the neural network; the neural network includes one input layer, one output layer, and four hidden layers, with 60 neurons in each layer; The sample contains a total of N c components. Let the number of input parameters be N c +3. The input parameter X includes the component concentration z i , where i = 1, …, N c , pressure P, temperature T, and pore radius r. X is written as X = [P, T, r, z1, …, z Nc T ; the dimension of the output layer is 3. Therefore, the results output by the neural network are of three types, namely gas phase, liquid phase, and coexistence of gas phase and liquid phase;​ The activation function used during phase identification training is the ReLU function: The activation function of the output layer is the Softmax function: The steps of using a large amount of data to train the neural network to calculate component mole fractions and capillary force parameters specifically include: Calculate and store a large amount of data through a conventional flash calculation method as data for training a neural network; the neural network includes one input layer, one output layer, and four hidden layers, with 60 neurons in each layer; the sample contains a total of N c components, so let the number of input parameters be N c +3. The input parameter X includes the component concentration z i , where i = 1, …, N c , the pressure P, the temperature T, and the pore radius r. X is written as X = [P, T, r, z1, …, z Nc T ; the dimension of the output parameter is 2N c +3. The output parameters include the capillary pressure P C , the gas-phase mole fraction F v , the liquid-phase mole fraction F L , and the mole fraction of the component in the gas phase y i , where i = 1, …, N c , the mole fraction of the component in the liquid phase x i , where i = 1, …, N c ,​ The activation function is the ReLU function The activation function of the output layer is:

2. The method according to claim 1, characterized in that Reservoir geological parameters include reservoir scale, shape, permeability, and porosity; fracture geometric parameters include fracture number, orientation, strike, as well as fracture length, aperture, and permeability.

3. The method according to claim 1, wherein The specific steps of using flash calculation include: During the CO2 flooding process, there are gas-liquid two phases, and the equilibrium of component i in the system satisfies thermo-dynamic equilibrium: Among them represents the fugacity coefficient, which can be calculated from the component composition, pressure, temperature, and volume data; first, the initial value K of the gas-liquid equilibrium constant is predicted by a neural network obtained through training i , and K i is substituted into the Rachford-Rice equation to solve for the mole fraction F of the gas-phase components v Then further solving to obtain component mole fraction parameters; while the temperature T, pressure P, and volume V parameters are calculated through the PR equation of state: During the calculation process, iteration is realized by detecting whether the equation satisfies the mass conservation condition and the thermo-dynamic equilibrium condition. During the iteration process, the equilibrium ratio is updated until the conservation condition is satisfied and the process ends.

4. The method according to claim 1, wherein Establishing an embedded discrete fracture model: The embedded discrete fracture model directly embeds the fracture system into the grid. After grid meshing, the fractures are embedded into the grid, and then the intersection relationship between the fracture system and the grid system is calculated, and the fracture system and the bedrock are connected according to the conservation law; The mass conservation equation of the embedded discrete model is: where k F is the permeability of the large fracture, p F is the pressure of the large fracture, μ is the fluid viscosity, q F is the source-sink term of the large fracture, q Ff is the cross-flow rate between the two systems, q FF represents the cross-flow rate between intersecting fracture elements, δ FF indicates whether the large fracture element intersects with other fracture elements. If the large fracture element intersects with other fracture elements, δ FF = 1, otherwise δ FF = 0, V F is the volume of the fracture element; The steps of grid meshing specifically include: First, for the overall reservoir, since the fractures are embedded and do not need to be used as the inner boundary of the bedrock grid, the reservoir is directly meshed, and then the fracture system is directly embedded into the reservoir grid to calculate the intersection information between the fractures and the reservoir grid and the geometric information of the fractures after being meshed by the reservoir grid; For reservoirs with regular shapes, orthogonal grid meshing is adopted. According to the size of the research area and the required calculation accuracy of the research, the grid step sizes and numbers in each spatial direction are determined. The grid shape of the orthogonal grid is a regular shape, which is convenient for calculation; The intersection relationship between the fracture system and the grid system includes the intersection coordinates of fractures and the grid, and the situation of fractures being cut by the grid boundaries; Calculating the parameters of the embedded discrete fracture model: determined according to the actual data of the oilfield. For natural fractures, it is necessary to determine their fracture spatial coordinates, number, and spacing. Then, on this basis, calculate the grid cells passed through by each fracture and the connection information between the fracture and the grid cells, including the contact area between the fracture and the grid, and the average distance from the grid where the fracture is located to the fracture surface, to prepare for the subsequent simulation of component flow in the hybrid model.

5. The method according to claim 1, wherein A method for establishing a CO2 flooding flow model for fractured reservoirs includes the following steps: Establishing the conservation equations for each component, and combining the embedded discrete fracture model and the flash calculation technology based on neural networks to form a CO2 flooding flow model for fractured reservoirs; among them, a mass conservation equation is established for each component where \(i\) represents the component, \(i = 1,\ldots,n_c\), \(n_c\) represents the number of components, \(F\) is the mass flow rate, \(q\) represents the source-sink term, and \(N\) is the mass change; \(N\) is expressed through the phase molar density \(\rho\), the saturation \(S\), the mole fraction \(x\) of component \(i\) in the oil phase i , and the mole fraction \(y\) of component \(i\) in the gas phase i as F can be written as where v β is the Darcy velocity of phase β Each component should satisfy the component mass conservation condition 6. The method according to claim 1, wherein The steps for numerically solving the CO2 flooding model for fractured reservoirs based on neural networks include: Numerical solution first requires discretizing the control numerically. The time term is discretized using the backward Euler method. To ensure the stability of the calculation, an implicit discretization scheme is adopted, a Jacobian matrix is constructed, and the iterative method is used to solve it; to reduce the matrix dimension, the equations in the CO2 flooding mathematical model are divided into main equations and auxiliary equations, and the variables are divided into main variables and auxiliary variables. During the numerical solution process, the main variables are obtained by solving the main equations first, and then the auxiliary variables are obtained through the auxiliary equations.

Images