A flow heat transfer prediction method, device and medium based on adversarial neural network

Abstract

The present invention discloses a flow heat transfer prediction method, device and medium based on an adversarial neural network, which relates to the field of data prediction technology. Mass source terms and energy source terms are constructed through the geometric structure and physical characteristics of a target prediction device, and simulation is performed based on the mass source terms and energy source terms to obtain multi-physical field simulation results of the target prediction device. By generating mass source terms and energy source terms, the early data error can be reduced in the early parameter acquisition process. By constructing a reduced-order model, the multi-physical field simulation results of the target prediction device are reduced in order to reduce the data calculation complexity. By constructing a flow heat transfer prediction model through an adversarial neural network model for prediction, the prediction accuracy of the generated flow field cloud map and temperature field cloud map can be improved, the flow heat transfer characteristics in industrial equipment can be obtained in real time, data support can be provided for model development, and important analysis methods and tools can be provided for industrial digital twins.

Description

Technical Field

[0001] The present invention relates to the field of data prediction technology, and in particular to a flow heat transfer prediction method, device and medium based on an adversarial neural network. Background Art

[0002] Two-phase flow heat transfer is widely used in the industrial field. Two-phase flow heat transfer refers to the flow and heat transfer phenomenon that occurs in a fluid system where vapor and liquid phases coexist in the channels of certain devices. This process has a wide range of applications in chemical industry, nuclear energy, petroleum, environmental protection and other fields. In many industrial processes, such as air conditioners, refrigerators, boilers and nuclear power plants, efficient heat transfer technology is required to improve energy utilization efficiency. By studying the mechanism of vapor-liquid two-phase flow and heat transfer, and combining it with specific application requirements, the optimization and control of fluid flow and heat transfer can be achieved, which can improve the efficiency and quality of industrial production, while achieving energy conservation and environmental protection.

[0003] Therefore, to provide better fluid flow and heat transfer control in energy utilization research, it is necessary to predict the two-phase flow in the device. Numerical simulation of complex flow and heat transfer phenomena generally requires a large amount of computing resources. However, existing software for two-phase flow calculations is not accurate enough and suffers from large errors. Due to early data errors, the calculation complexity is high when building turbulence models and boundary layer transitions, resulting in reduced prediction accuracy, making it difficult to meet the high-precision requirements of industrial applications. Summary of the Invention

[0004] The technical problem to be solved by the present invention is that the existing prediction methods have high computational complexity and low prediction accuracy. The existing software is not accurate enough for the calculation of two-phase flow and has large errors. The purpose is to provide a flow heat transfer prediction method, equipment and medium based on an adversarial neural network. The mass source term and energy source term are constructed through the geometric structure and physical characteristics of the flow channel, and simulation is performed based on the mass source term and energy source term to obtain the multi-physical field simulation results of the target prediction device. By customizing the mass source term and energy source term, the early data error can be reduced in the early calculation and data acquisition process. By constructing a reduced-order model, the multi-physical field simulation results of the target prediction device are reduced in order to reduce the data calculation complexity. By constructing a flow heat transfer prediction model through an adversarial neural network model for prediction, the prediction accuracy of the generated flow field cloud map and temperature field cloud map can be improved.

[0005] The present invention is achieved through the following technical solutions:

[0006] A first aspect of the present invention provides a flow heat transfer prediction method based on an adversarial neural network, comprising the following specific steps:

[0007] According to the geometric structure and physical characteristics of the target prediction device, the target prediction device is meshed in a multi-physics field framework;

[0008] Based on the fluid phase change parameters, heat transfer parameters and phase change latent heat parameters of each grid, the mass source term and energy source term are obtained;

[0009] Perform simulation based on user-defined mass source terms and energy source terms to obtain multi-physics field simulation results of the target prediction device;

[0010] Construct a reduced-order model and generate temperature field cloud map data training samples and velocity field cloud map data training samples based on the multi-physics field simulation results of the target prediction device;

[0011] Construct an adversarial neural network model, use temperature field cloud map data training samples and velocity field cloud map data training samples to train the adversarial neural network model, and obtain a flow heat transfer prediction model;

[0012] The multi-physics field simulation results of the target prediction device are input into the flow heat transfer prediction model to predict and generate flow field cloud maps and temperature field cloud maps.

[0013] Furthermore, the steps for obtaining the quality source item specifically include:

[0014] Get the grid volume and obtain the characteristic length of the grid unit;

[0015] Obtain the phase change adjustment coefficient, liquid phase change latent heat, gas density, bubble temperature, pressure inside the bubble, gas molar mass and molar gas constant, and obtain the condensation and evaporation heat transfer coefficients of the phase change interface;

[0016] Obtain the liquid's phase change latent heat, saturation temperature, and grid unit temperature, and determine the vapor-liquid mass transfer rate per unit volume of each grid by combining the condensation and evaporation heat transfer coefficients at the phase change interface.

[0017] Obtain the mass transfer rate constraint of the vapor-liquid two-phase per unit volume of the grid, and combine the mass transfer rate of the vapor-liquid two-phase per unit volume of each grid to obtain the mass source term of each grid.

[0018] Furthermore, obtaining the condensation and evaporation heat transfer coefficients of the phase change interface specifically includes:

[0019] ;

[0020] in, represents the phase change adjustment coefficient, that is, the non-ideal effect caused by the presence of non-condensable gas molecules at the vapor-liquid interface. h lg is the latent heat of phase change of liquid, r v represents the gas density,T v is the bubble temperature, represents the molar mass of the gas, is the molar gas constant, p v is the pressure inside the bubble.

[0021] Furthermore, determining the mass transfer rate of the vapor-liquid two-phase per unit volume of each grid specifically includes:

[0022] ;

[0023] in, m lg represents the mass transfer rate of vapor-liquid two-phase per unit volume in each grid, T sat represents the saturation temperature, T l represents the grid cell temperature, h ev represents the condensation and evaporation heat transfer coefficients at the phase change interface, h lg represents the latent heat of phase change of liquid, L Represents the characteristic length of the grid cell, which is the cube root of the cell volume.

[0024] Furthermore, the mass transfer rate constraint conditions specifically include:

[0025] When condensing,

[0026] ;

[0027] When evaporating,

[0028] ;

[0029] in, r v represents the steam density, a v represents the gas phase volume fraction, △ t Indicates the current calculation time step.

[0030] Furthermore, the steps for obtaining the energy source term specifically include:

[0031] The constant-pressure specific heat capacity of the liquid phase, the constant-pressure specific heat capacity of the vapor phase, the time step number, the temperature corresponding to the time step number, and the reference temperature in each grid are obtained, and the energy source phase is obtained by combining the mass transfer rate of the vapor-liquid two-phase per unit volume in each grid.

[0032] Furthermore, the specific calculation steps for obtaining the energy source phase include:

[0033] ;

[0034] in, represents the energy source phase, m lg represents the mass transfer rate of vapor-liquid two-phase per unit volume in each grid, c pl represents the specific heat capacity of the liquid phase at constant pressure, c pg represents the constant-pressure specific heat capacity of the vapor phase, T n Indicates the number of time steps n temperature, n represents the number of time steps, T ref Indicates the reference temperature.

[0035] Furthermore, the construction of the reduced-order model generates temperature field cloud map data training samples and velocity field cloud map data training samples based on the multi-physics field simulation results of the target prediction device, specifically including:

[0036] Extract the raw data of temperature field and velocity field from multi-physics simulation for preprocessing;

[0037] Build a reduced-order model, input the preprocessed data into the reduced-order model, and generate a learning matrix;

[0038] Perform singular value decomposition on the learning matrix to obtain the eigenvalues of the decomposed data;

[0039] Before using decomposed data K The eigenvalues are used as the order reduction basis to project the original data into the low-dimensional space to obtain the temperature field cloud map data training samples and the velocity field cloud map data training samples.

[0040] Furthermore, the process of obtaining the eigenvalues of the decomposed data also includes calculating the average projection error, and optimizing the number of eigenvalues of the decomposed data according to the average projection error. K .

[0041] Furthermore, the step of calculating the average projection error includes:

[0042] ;

[0043] in, relativeError represents the average projection error, M represents the learning matrix, K represents the number of eigenvalues, s Indicates deviation, i Indicates from 1 to K +1.

[0044] Furthermore, the construction of the adversarial neural network model specifically includes:

[0045] Build the generator;

[0046] Obtain random noise data of temperature field cloud map data training samples and velocity field cloud map data training samples, and input them into the generator;

[0047] The generator trains the random noise data of the multi-physics field simulation results of the target prediction device to learn the probability distribution of the physics field cloud map until the training reaches a set threshold, thereby obtaining a generated physics field cloud map;

[0048] Build the discriminator;

[0049] Extracting real physical field cloud maps based on temperature field cloud map data training samples and velocity field cloud map data training samples;

[0050] The generated physical field cloud map and the real physical field cloud map are distinguished, and the generated physical field cloud map is optimized according to the discrimination result until the discrimination result reaches the set threshold, and the flow heat transfer prediction model is obtained.

[0051] Furthermore, the computational steps for constructing the adversarial neural network model include:

[0052] ;

[0053] in, Represents the random noise data of the temperature field cloud map data training samples and the velocity field cloud map data training samples, G represents a generator, D represents the discriminator, G ( z ) represents the generated physical field cloud map data, D ( x ) represents the real physical field cloud map, x represents the real physical field cloud map sample, z represents the random noise vector, E Indicates the expected value, Pdata(x) Represents the actual physical field cloud distribution.

[0054] A second aspect of the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a flow heat transfer prediction method based on an adversarial neural network is implemented.

[0055] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a flow heat transfer prediction method based on an adversarial neural network.

[0056] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0057] The mass source term and energy source term are constructed through the geometric structure and physical characteristics of the target prediction device, and simulation is performed based on the mass source term and energy source term to obtain the multi-physical field simulation results of the target prediction device. By customizing the mass source term and energy source term, the early data error can be reduced in the early data acquisition process. By constructing a reduced-order model, the multi-physical field simulation results of the target prediction device are reduced in order to reduce the complexity of data calculation. The flow heat transfer prediction model is constructed by using an adversarial neural network model for prediction, which can improve the prediction accuracy of the generated flow field cloud map and temperature field cloud map, and can obtain the flow heat transfer characteristics in industrial equipment in real time, provide data support for model development, and provide important analysis methods and tools for industrial digital twins. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the examples. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be considered as limiting the scope. A person of ordinary skill in the art can also derive other relevant drawings based on these drawings without inventive effort. In the drawings:

[0059] Figure 1 This is the prediction process in the embodiment of the present invention. DETAILED DESCRIPTION

[0060] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with examples and drawings. The exemplary embodiments of the present invention and their descriptions are only used to explain the present invention and are not intended to limit the present invention.

[0061] As a possible implementation, Figure 1As shown, this embodiment provides a flow heat transfer prediction method based on an adversarial neural network, comprising the following specific steps: meshing the target prediction device in a multi-physics field framework according to the geometric structure and physical properties of the target prediction device; obtaining mass source terms and energy source terms based on the fluid phase change parameters, heat transfer parameters, and phase change latent heat parameters of each grid; performing simulation based on the mass source terms and energy source terms to obtain multi-physics field simulation results of the target prediction device; constructing a reduced-order model, and generating temperature field cloud map data training samples and velocity field cloud map data training samples according to the multi-physics field simulation results of the target prediction device; constructing an adversarial neural network model, and using the temperature field cloud map data training samples and the velocity field cloud map data training samples to train the adversarial neural network model to obtain a flow heat transfer prediction model; inputting the multi-physics field simulation results of the target prediction device into the flow heat transfer prediction model to predict and generate a flow field cloud map and a temperature field cloud map. In this embodiment, mass source terms and energy source terms are constructed through the geometric structure and physical characteristics of the target prediction device, and simulation is performed based on the mass source terms and energy source terms to obtain multi-physical field simulation results of the target prediction device. By customizing the mass source terms and energy source terms, it is possible to reduce early data errors in the early data acquisition process. By constructing a reduced-order model, the multi-physical field simulation results of the target prediction device are reduced in order to reduce the complexity of data calculation. By constructing a flow heat transfer prediction model through an adversarial neural network model for prediction, the prediction accuracy of the generated flow field cloud map and temperature field cloud map can be improved.

[0062] In some possible implementations, because commercial CFD software requires user-defined source terms, inaccurate definitions can lead to inaccurate boiling two-phase heat transfer and phase change calculations. Therefore, the present invention employs a self-compiled UDF (User Defined Function) to correct the source terms in commercial CFD software. Customized mass and energy source terms are added to the commercial CFD software via the UDF interface, defining the mass and energy source terms in the commercial CFD software. The mass and energy source terms in this application are derived using the following method:

[0063] In some possible implementations, the step of obtaining the mass source term specifically includes:

[0064] Get the grid volume and obtain the characteristic length of the grid unit;

[0065] Obtain the phase change adjustment coefficient, liquid phase change latent heat, gas density, bubble temperature, pressure inside the bubble, gas molar mass and molar gas constant, and obtain the condensation and evaporation heat transfer coefficients of the phase change interface;

[0066] Obtain the liquid's phase change latent heat, saturation temperature, and grid unit temperature, and determine the vapor-liquid mass transfer rate per unit volume of each grid by combining the condensation and evaporation heat transfer coefficients at the phase change interface.

[0067] Obtain the mass transfer rate constraints of the vapor-liquid two-phase per unit volume of the grid unit, and combine them with the mass source term to obtain the mass transfer rate of the vapor-liquid two-phase per unit volume of each grid.

[0068] In some possible implementations, obtaining the condensation and evaporation heat transfer coefficients of the phase change interface specifically includes:

[0069] ;

[0070] in, represents the phase change adjustment coefficient, that is, the non-ideal effect caused by the presence of non-condensable gas molecules at the vapor-liquid interface. h lg is the latent heat of phase change of liquid, r v represents the gas density, T v is the bubble temperature, represents the molar mass of the gas, is the molar gas constant, p v is the pressure inside the bubble.

[0071] In some possible implementations, determining the mass transfer rate of the vapor-liquid two-phase per unit volume of each grid specifically includes:

[0072] ;

[0073] in, m lg represents the mass transfer rate of vapor-liquid two-phase per unit volume in each grid, T sat represents the saturation temperature, T l represents the grid cell temperature, h ev represents the condensation and evaporation heat transfer coefficients at the phase change interface, h lg represents the latent heat of phase change of liquid, L Represents the characteristic length of the grid unit. The characteristic length of the unit cell is the cube root of the unit cell volume. For liquids with a temperature greater than a certain superheat, the negative sign is used in front of this formula; for liquids with a temperature below the saturation temperature, the negative sign is used in front of this formula. T sat The steam in this formula takes a positive sign. That is, in this embodiment, the UDF setting is only when the liquid superheat is greater than the set value (the present invention is set to 10 o C), it will evaporate and turn into steam; and when the steam temperature is lower than the saturation temperature, it will condense and turn into liquid.

[0074] In some possible implementations, the mass transfer rate constraint specifically includes:

[0075] When condensing,

[0076] ;

[0077] When evaporating,

[0078] ;

[0079] in, r v represents the steam density, a v represents the gas phase volume fraction, △ t Indicates the current calculation time step.

[0080] In some possible implementations, the step of obtaining the energy source term specifically includes:

[0081] The constant-pressure specific heat capacity of the liquid phase, the constant-pressure specific heat capacity of the vapor phase, the time step number, the temperature corresponding to the time step number, and the reference temperature in each grid are obtained, and the energy source phase is obtained by combining the mass transfer rate of the vapor-liquid two-phase per unit volume in each grid.

[0082] In some possible implementations, the specific calculation steps for obtaining the energy source phase include:

[0083] ;

[0084] in, represents the energy source phase, m lg represents the mass transfer rate of vapor-liquid two-phase per unit volume in each grid, c pl represents the specific heat capacity of the liquid phase at constant pressure, c pg represents the constant-pressure specific heat capacity of the vapor phase, T n Indicates the number of time steps n temperature, n represents the number of time steps, T ref Indicates the reference temperature.

[0085] In some possible implementations, a reduced-order model is constructed to generate temperature field cloud map data training samples and velocity field cloud map data training samples based on the multi-physics field simulation results of the target prediction device, specifically including:

[0086] Extract the raw data of temperature field and velocity field from multi-physics simulation for preprocessing;

[0087] Build a reduced-order model, input the preprocessed data into the reduced-order model, and generate a learning matrix;

[0088] Perform singular value decomposition on the learning matrix to obtain the eigenvalues of the decomposed data;

[0089] Before using decomposed data K The eigenvalues are used as the order reduction basis to project the original data into the low-dimensional space to obtain the temperature field cloud map data training samples and the velocity field cloud map data training samples.

[0090] In some possible implementations, the reduced-order model may use singular value decomposition (SVD) to perform reduced-order simulation of the physical field.

[0091] In some possible implementations, the process of obtaining the eigenvalues of the decomposed data further includes calculating the average projection error, and optimizing the number of eigenvalues of the decomposed data according to the average projection error. K .

[0092] In some possible implementations, singular value decomposition (SVD) is used to mesh the entire heat conduction structure or part of the target prediction device under the heat conduction problem to obtain N working conditions (s1, s2, ...s N ), each grid represents a working condition, and the working condition represents the combination of input conditions for each set of calculations.

[0093] Calculate data for N working conditions, specify the input parameters and output parameters of the entire data set, and divide it into a learning set q and a validation set m. The input parameters of the learning data and the temperature values of each grid constitute a learning matrix M.

[0094] Perform SVD decomposition on the learning matrix M, and then select the front K The feature values are used to obtain an approximate expression of the learning matrix M. Because the number of rows of the learning matrix M cannot be changed for the rapid simulation of the thermal conductive structure, column dimensionality reduction is required, and the approximate expression of M is obtained as follows:

[0095] ;

[0096] in, M Represents the learning matrix, whose dimension is n × m ,in n represents the number of samples, m Represents the number of features, X represents the approximately reconstructed data matrix, and its dimension is also n × m , a i Indicates the i The eigenvalue represents the iThe proportion of the eigenvectors in the data variance, U indicates that the column vector of the matrix is the left singular vector of the data matrix M, forming an orthogonal basis of the data, K Represents the number of eigenvalues, which is used to approximate the original data matrix M. K Usually much smaller than m , which can greatly reduce the dimension of the data.

[0097] In some possible implementations, the step of calculating the average projection error includes:

[0098] ;

[0099] in, relativeError represents the average projection error, M represents the learning matrix, K represents the number of eigenvalues, s Indicates deviation, i Indicates from 1 to K +1.

[0100] In some possible implementations, the standard average projection error is: each learning vector is removed from the learning set, projected using the mode calculated using the remaining vectors, and then the resulting projection errors are averaged (cross validation).

[0101] In some possible implementations, since the eigenvalues of a matrix have the characteristic of decreasing rapidly, the sum of the first 10% or even the first 1% eigenvalues accounts for more than 99% of the total, so as K As K increases, the average projection error will inevitably decrease, that is, the degree of dispersion will become smaller and smaller, and will even drop rapidly when it reaches a certain value. However, the average standard projection error will gradually stabilize as K increases and as K decreases, that is, the degree of deviation will no longer change. For small values of K, the average projection error and the average standard projection error are almost the same.

[0102] In some possible implementations, it also includes: using dozens of groups of learning sets α , the learning set is already known α i … α K (Each sample point has a group of q group), the validation set is also known α i … α K (Each sample point has a group, a total of m groups), the validation set is used to verify the optimal number of patterns K, and the response surface is generated within the given range using the genetic aggregation interpolation method (input parameters and α The relationship between αEquivalent to the output parameters of this response surface). When using the generated response surface for prediction, each sample point is predicted separately. α i（i=1，2,…K） After that, only one column of U is needed to reach the predicted learning matrix M, making dimensionality reduction faster. However, because the number of rows in the M matrix cannot be changed for fast simulation of thermal conductive structures, column dimensionality reduction is required.

[0103] In some possible implementations, a generative adversarial neural network (GAN) is composed of two neural network generators G and discriminators D with conflicting objectives, where:

[0104] Construct an adversarial neural network model, specifically including:

[0105] Build the generator;

[0106] Obtain random noise data of temperature field cloud map data training samples and velocity field cloud map data training samples, and input them into the generator;

[0107] The generator trains the random noise data of the multi-physics field simulation results of the target prediction device to learn the probability distribution of the physics field cloud map until the training reaches a set threshold, thereby obtaining a generated physics field cloud map;

[0108] Build the discriminator;

[0109] Extracting real physical field cloud maps based on temperature field cloud map data training samples and velocity field cloud map data training samples;

[0110] The generated physical field cloud map and the real physical field cloud map are distinguished, and the generated physical field cloud map is optimized according to the discrimination result until the discrimination result reaches the set threshold, and the flow heat transfer prediction model is obtained.

[0111] In some possible implementations, the generator models the joint probability, representing the distribution of data from a statistical perspective and describing how the data is generated; the discriminator models the conditional probability, with the main goal of finding the optimal classification surface between different categories. Figure 1As shown in Figure 1, in a GAN, the generator G takes random noise z as input and continuously learns the probability distribution of real data (physical field cloud maps) in the training set. Its goal is to transform the input random noise z into a physical field cloud map, that is, to generate a physical field cloud map that is as similar as possible to the physical field cloud maps in the training set. The discriminator D takes the real physical field cloud map generated by the generator as input and outputs the probability value of the physical field cloud map being real. Its function is to determine whether a physical field cloud map is real, with the goal of distinguishing the physical field cloud map generated by the generator G from the physical field cloud maps in the input training set. The existence of the discriminator D enables G to learn to approximate real data well without extensive prior knowledge or prior distribution, ultimately making the data generated by the model indistinguishable from the real one.

[0112] The original generative adversarial network GAN's discriminator D and generator G both use multi-layer perceptrons. GAN defines a noise As a priori, used to learn the generator G in the training data x The probability distribution on G( z ) represents the noise input z Generated physical field cloud data , D( x ) represents the probability from the true data distribution, so the optimization objective is as follows:

[0113] ;

[0114] in, Represents the random noise data of the temperature field cloud map data training samples and the velocity field cloud map data training samples, G represents a generator, D represents the discriminator, G ( z ) represents the generated physical field cloud map data, D ( x ) represents the real physical field cloud map, x represents the real physical field cloud map sample, z represents the random noise vector, E Indicates the expected value, Pdata(x) Represents the actual physical field cloud distribution.

[0115] The optimization objective function can be interpreted as: when updating the parameters of the discriminator D, for the Pdata Sample x , the goal of GAN is the discriminator D( x ) is close to 1, that is, logD( x ) is larger; for passing noise z The generated data G( z ), the goal of GAN is the output D(G(z )) is close to 0, that is, log(1- D ( G (z))) is larger, so D needs to be maximized. When updating the parameters of the generator G, the goal of the generator is to generate a physical field cloud map as close to the real physical field cloud map as possible, that is, D(G( z )) is close to 1, log(1- D ( G The smaller (z))) the better, so G needs to be minimized.

[0116] As a possible implementation, the target prediction device could be a reactor core, where flow and heat transfer predictions are performed within the reactor core channels based on the core's geometry and physical properties. Alternatively, the target prediction device could be a heat exchanger, high-power electronic device, or boiler.

[0117] As a possible implementation method, this embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, a flow heat transfer prediction method based on an adversarial neural network is implemented.

[0118] As a possible implementation, this embodiment provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, a flow heat transfer prediction method based on an adversarial neural network is implemented.

[0119] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A flow heat transfer prediction method based on adversarial neural network, characterized in that: The specific steps include: According to the geometric structure and physical characteristics of the target prediction device, the target prediction device is meshed in a multi-physics field framework; Based on the fluid phase change parameters, heat transfer parameters and phase change latent heat parameters of each grid, the mass source term and energy source term are obtained; Based on the mass source term and the energy source term, a multi-physics field simulation result of the target prediction device is obtained; Construct a reduced-order model and generate temperature field cloud map data training samples and velocity field cloud map data training samples based on the multi-physics field simulation results of the target prediction device; Construct an adversarial neural network model, use temperature field cloud map data training samples and velocity field cloud map data training samples to train the adversarial neural network model, and obtain a flow heat transfer prediction model; Input the multi-physics field simulation results of the target prediction device into the flow heat transfer prediction model to predict and generate flow field cloud maps and temperature field cloud maps; The steps for obtaining quality source items include: Get the grid volume and obtain the characteristic length of the grid unit; Obtain the phase change adjustment coefficient, liquid phase change latent heat, gas density, bubble temperature, pressure inside the bubble, gas molar mass and molar gas constant, and obtain the condensation and evaporation heat transfer coefficients of the phase change interface; Obtain the liquid's phase change latent heat, saturation temperature, and grid unit temperature, and determine the vapor-liquid mass transfer rate per unit volume of each grid by combining the condensation and evaporation heat transfer coefficients at the phase change interface. Obtain the mass transfer rate constraint of the vapor-liquid two-phase per unit volume of the grid, and combine the mass transfer rate of the vapor-liquid two-phase per unit volume of each grid to obtain the mass source term of each grid; The steps for obtaining the energy source item include: Obtain the constant-pressure specific heat capacity of the liquid phase, the constant-pressure specific heat capacity of the vapor phase, the time step number, the temperature corresponding to the time step number, and the reference temperature in each grid, and combine the mass transfer rate of the vapor-liquid two-phase per unit volume in each grid to obtain the energy source term; The constructing of the adversarial neural network model specifically includes: Build the generator; Obtain random noise data of temperature field cloud map data training samples and velocity field cloud map data training samples, and input them into the generator; The generator trains the random noise data of the multi-physics field simulation results of the target prediction device to learn the probability distribution of the physics field cloud map until the training reaches a set threshold, thereby obtaining a generated physics field cloud map; Build the discriminator; Extracting real physical field cloud maps based on temperature field cloud map data training samples and velocity field cloud map data training samples; The generated physical field cloud map and the real physical field cloud map are distinguished, and the generated physical field cloud map is optimized according to the discrimination result until the discrimination result reaches the set threshold, and the flow heat transfer prediction model is obtained.

2. The flow heat transfer prediction method based on adversarial neural network according to claim 1, characterized in that: The method of obtaining the condensation and evaporation heat transfer coefficients of the phase change interface specifically includes: ； in, represents the phase change adjustment coefficient, that is, the non-ideal effect caused by the presence of non-condensable gas molecules at the vapor-liquid interface. is the latent heat of phase change of liquid, represents the gas density, is the bubble temperature, represents the molar mass of the gas, is the molar gas constant, is the pressure inside the bubble.

3. The flow heat transfer prediction method based on adversarial neural network according to claim 1, characterized in that: Determining the mass transfer rate of the vapor-liquid two-phase per unit volume of each grid specifically includes: ； in, represents the mass transfer rate of vapor-liquid two-phase per unit volume in each grid, represents the saturation temperature, represents the grid cell temperature, represents the condensation and evaporation heat transfer coefficients at the phase change interface, represents the latent heat of phase change of liquid, Represents the characteristic length of the grid cell, which is the cube root of the cell volume.

4. The flow heat transfer prediction method based on adversarial neural network according to claim 1, characterized in that: The mass transfer rate constraint conditions specifically include: When condensing, ； When evaporating, ； in, represents the steam density, Represents the gas phase volume fraction, and △t represents the current calculation time step.

5. The flow heat transfer prediction method based on adversarial neural network according to claim 1, characterized in that: The specific calculation steps for obtaining the energy source term include: ； in, represents the energy source term, represents the mass transfer rate of vapor-liquid two-phase per unit volume in each grid, represents the specific heat capacity of the liquid phase at constant pressure, represents the constant-pressure specific heat capacity of the vapor phase, represents the temperature at time step n, where n represents the number of time steps, Indicates the reference temperature.

6. The flow heat transfer prediction method based on adversarial neural network according to claim 1, characterized in that: The construction of the reduced-order model, generating temperature field cloud map data training samples and velocity field cloud map data training samples according to the multi-physics field simulation results of the target prediction device, specifically includes: Extract the raw data of temperature field and velocity field from multi-physics simulation for preprocessing; Build a reduced-order model, input the preprocessed data into the reduced-order model, and generate a learning matrix; Perform singular value decomposition on the learning matrix to obtain the eigenvalues of the decomposed data; The first K eigenvalues of the decomposed data are used as the order reduction basis to project the original data into a low-dimensional space to obtain the training samples of the temperature field cloud map data and the velocity field cloud map data.

7. The flow heat transfer prediction method based on adversarial neural network according to claim 6, characterized in that: The process of obtaining the eigenvalues of the decomposed data also includes calculating an average projection error, and optimizing the number K of eigenvalues of the decomposed data according to the average projection error.

8. The flow heat transfer prediction method based on adversarial neural network according to claim 7, characterized in that: The calculation step of the average projection error includes: ； in, represents the average projection error, M represents the learning matrix, K represents the number of eigenvalues, σ represents the deviation, and i represents from 1 to K+1.

9. The flow heat transfer prediction method based on adversarial neural network according to claim 1, characterized in that: The computational steps for constructing the adversarial neural network model include: ； in, represents the random noise data of the temperature field cloud map data training samples and the velocity field cloud map data training samples, G represents the generator, D represents the discriminator, G(z) represents the generated physical field cloud map data, D(x) represents the real physical field cloud map, x represents the real physical field cloud map sample, z represents the random noise vector, and E represents the expected value. Represents the actual physical field cloud distribution.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the flow heat transfer prediction method based on the adversarial neural network as described in any one of claims 1 to 9 is implemented.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the flow heat transfer prediction method based on the adversarial neural network as described in any one of claims 1 to 9 is implemented.

Images