A method for satellite payload temperature prediction and thermal parameter correction based on PINN

Abstract

The present invention discloses a method for satellite payload temperature prediction and thermal parameter correction based on PINN, which relates to the field of space thermal control technology. The key technical points are as follows: the present invention utilizes a thermophysical information neural network to effectively combine known temperature information data with thermophysical knowledge to achieve prediction of the satellite payload temperature field and correction of thermal model parameters; the present invention proposes a complete set of satellite payload temperature prediction and thermal parameter correction methods based on PINN, combining the neural network with thermophysical prior information to obtain thermal model parameters that meet physical laws and have high prediction accuracy, as well as a PINN model and thermal simulation model that accurately predict on-orbit / test temperatures, thereby improving the efficiency of thermal parameter correction and the accuracy of temperature prediction.

Description

Technical Field

[0001] The present invention relates to the field of space thermal control technology, and in particular to a method for satellite payload temperature prediction and thermal parameter correction based on PINN. Background Art

[0002] Satellite payload thermal models cannot accurately predict test and on-orbit temperature conditions due to inaccuracies in simulation models. Therefore, the thermal parameters of the thermal model must be corrected using test or on-orbit temperature data to ultimately obtain accurate thermal parameters and a thermal simulation model. Traditional correction methods involve manual trial and error by thermal control engineers, or Monte Carlo sampling followed by other optimization methods.

[0003] In recent years, to improve correction efficiency, some researchers have introduced neural network proxy models to replace simulation models. However, these methods lead to two problems. First, with the development of future spacecraft missions, the structural layout of satellite payloads and their thermal control subsystems is becoming increasingly complex, which leads to numerous uncertainties in thermal model correction. These thermal parameters may be mutually correlated and coupled, and temperature measurement points are often insufficient, often leading to underdetermined solutions. Second, both manual trial-and-error methods and Monte Carlo methods require repeatedly substituting groups of thermal parameters back into the simulation software, performing thousands of simulation calculations. This step is extremely time-consuming and the resulting data sets are sparse, making it impossible to meet the ever-shorter development cycles and increasingly high prediction accuracy requirements.

[0004] With the development of computers, some researchers have proposed using machine learning to solve the correction problem. Replacing time-consuming finite element calculations with neural networks can significantly improve correction efficiency. However, as neural networks are "black boxes," the thermal parameters they solve often lose their physical meaning, making it difficult to deeply explain and elaborate on the thermal mechanisms involved in thermal model correction. Furthermore, they are somewhat disconnected from the actual physical processes involved.

[0005] To this end, the present invention aims to provide a method for satellite payload temperature prediction and thermal parameter correction based on PINN, which adds thermal boundary conditions and physical constraints to the general proxy model to overcome the above shortcomings. Summary of the Invention

[0006] The purpose of the present invention is to solve the above problems and provide a method for satellite payload temperature prediction and thermal parameter correction based on PINN, which is used to solve the problems of prediction accuracy of thermal simulation models under sparse temperature measurement data, low efficiency of thermal parameter correction, and inconsistency between the corrected thermal parameters and the actual thermal physical phenomena, and the technical problems that the obtained thermal parameters are not in the real state, thereby improving the efficiency of thermal model correction and the accuracy of the simulation model.

[0007] In order to achieve the above object, the technical solution of the present invention is as follows:

[0008] The present invention provides a method for satellite payload temperature prediction and thermal parameter correction based on PINN, the method comprising the following steps:

[0009] S1. Establish an original thermal simulation model of the satellite payload, sample the thermal parameters therein and perform simulation calculations to obtain a training data set;

[0010] S2. Establish the thermophysical mechanism equation of the satellite payload, construct the PINN model and assign thermophysical constraints to the model based on the thermal equation;

[0011] S3, using the training data set of S1 as the first data driving information and the thermal physical constraints of S2 as the first physical driving information, train the PINN model, calculate the first loss function, and finally obtain an accurate PINN temperature prediction model;

[0012] S4. Using the heat transfer formula as the second physical driving information and the on-orbit telemetry test temperature as the second data driving information, a second loss function of the PINN model is obtained;

[0013] S5. Based on the PINN model of S3, adjust the thermal parameters to minimize the second loss function. The thermal parameters at this time are the corrected thermal parameters, which are substituted into the original thermal simulation model to obtain the corrected accurate thermal simulation model.

[0014] In step S1, the original thermal simulation model of the satellite payload is established, thermal parameters thereof are sampled and simulated to obtain a training data set, including the following steps:

[0015] S101. Establishing an original thermal simulation model in simulation modeling software according to a physical model of a satellite payload;

[0016] S102. Use finite element analysis software to simplify the geometry and divide the mesh;

[0017] S103, performing temperature field analysis to obtain the temperature field distribution and temperature level of key parts of the satellite payload;

[0018] S104, performing parameter identification on thermal parameters of the thermal model;

[0019] S105 , determining the value range of the thermal parameters in step S104 , performing Latin hypercube uniform sampling, and substituting the values ​​into the original thermal simulation model in step S101 for simulation calculation to obtain a data matrix of temperature and coordinates.

[0020] In step S2, the establishment of the thermophysical mechanism equation of the satellite payload, the construction of the PINN model, and the assignment of thermophysical constraints to the model based on the thermal equation include the following steps:

[0021] S201, establish the general thermal control equations for each component,

[0022] Where ρ is the density; φ is the general solution function, which is represented by the temperature matrix T in this problem; Γ φ is the generalized diffusion coefficient; λ is the thermal conductivity; S is the generalized source term, which is represented by the radiation heat transfer between the component and the space environment;

[0023] The first term of the heat control equation is a non-steady-state term, the second term is a convection term, the third term is a diffusion term, and the fourth term is a source term. Since the satellite payload is located in a space environment and the heat transfer mode is thermal radiation and heat conduction, the non-steady-state term and the convection term can be ignored, and the first and second terms are 0.

[0024] The control equation should be further expanded as follows:

[0025]

[0026] Where a is the thickness of the component, ε is the infrared hemispherical emissivity of the component surface, σ is the Boltzmann constant, and T u is the matrix of the absolute surface temperature distribution of the component and its position, T0 is the heat sink temperature, x is the abscissa in the Cartesian coordinate system, and y is the ordinate;

[0027] S202. The PINN model should be built on the basis of a general neural network model framework. The input unit includes thermal parameters in the satellite payload thermal model, including the surface infrared hemispherical emissivity ε, solar absorptivity α, thermal resistance r, heat transfer coefficient k, thermal conductivity λ, heat capacity C, heat loss h and other thermal parameters, spatial coordinate vector (x, y), and the simulation temperature matrix T required for data driving. (x,y) , referred to as T.

[0028] The PINN model output unit should include the predicted temperature at the (x,y) position Hereinafter referred to as The PINN model should be able to calculate the first-order partial differential and second-order partial differential equations between temperature and coordinates. Therefore, the output unit should also include

[0029] The PINN model should have a data-driven first loss function Loss data , the physical drive first loss function Loss PDE ;

[0030] The PINN model should have a hidden layer and its weight vector w ij , bias vector b ij and activation functions;

[0031] S203, when the heat sink is 4K cold air, T0 in the control equation of S201 4 Relative to T u 4 can be neglected, so the thermophysical constraint should be

[0032]

[0033] In step S3, the training data set of S1 is used as the first data driving information, the thermal physical constraints of S2 are used as the first physical driving information, the PINN model is trained, the first loss function is calculated, and finally an accurate PINN temperature prediction model is obtained, which includes the following steps:

[0034] S301, the data matrix of simulated temperature and coordinates obtained in S1 is used as a training data set, with coordinate data as input and temperature calculation value as output, to provide data-driven information for the training of the PINN model. The data-driven first loss function should be:

[0035]

[0036] S302, the thermophysical constraints obtained in S2 are given to PINN in the form of the first loss function, and the output unit The first loss function for calculating physical constraints should be

[0037]

[0038] S303: Perform weighted summation of the data-driven first loss function and the physical constraint first loss function to obtain the total loss function.

[0039] Loss f =αLoss data +βLoss PDE ;

[0040] Among them, α is the weight coefficient of the data-driven first loss function, and β is the weight coefficient of the physical constraint first loss function;

[0041] S304, to minimize Loss f That is min(Loss f) as the goal, the PINN model built in S202 is trained based on the temperature data obtained in S105, and the optimization algorithm is used to optimize the network structure, solver settings and hyperparameter optimization, as well as feature extraction and data enhancement, so that the prediction results of the PINN model can accurately replace the calculation results of the simulation model. At this time, the PINN model is an accurate PINN temperature prediction model.

[0042] In step S4, the heat transfer formula is used as the second physical driving information, and the on-orbit telemetry test temperature is used as the second data driving information to obtain the second loss function of the PINN model. The difference from the first driving information is that the first driving information is the field quantity of the temperature field, expressed as a temperature matrix, while the second driving information is the predicted temperature value, expressed as a scalar. Specifically, the following steps are included:

[0043] S401. Since there are thermal interfaces between components, the temperature fields of different components are coupled with each other and should satisfy the heat transfer formula:

[0044] Steady-state heat conduction formula for the wall

[0045]

[0046] Where k is the heat transfer coefficient between components, A is the contact area, d is the thickness, and T i is the temperature of component i, T j is the temperature of component j;

[0047] The second loss function of steady-state heat conduction on the wall is

[0048]

[0049] in, Predicted temperature for the PINN of component i, Predict the temperature of PINN for component j;

[0050] Heat balance formula for a single component

[0051]

[0052] Where j is the component adjacent to component i, D ji is the inter-component conduction network coefficient, R ji is the radiation network coefficient between components, Q i is the total heat source of the component;

[0053] The second loss function of the thermal balance formula in steady state

[0054]

[0055] Heat conduction formula for a single component

[0056]

[0057] Among them, T a is the temperature of temperature measuring point a, T b is the temperature of temperature measuring point b;

[0058] The second loss function driven by physics should be

[0059] Loss py =mLoss1+nLoss2+sLoss3;

[0060] Among them, m, n, and s are the proportions of the second loss function of the three physical constraints respectively;

[0061] The PINN model obtained in S402 and S3 no longer adjusts the network structure, hyperparameters and other information, but only adjusts the input parameters, namely the thermal parameters, and uses the on-orbit telemetry or experimental measurement temperature as the new data-driven information. The data-driven first loss function should be

[0062]

[0063] in, To predict the temperature value of the PINN model after adjusting the thermal parameters, Measuring temperature for on-orbit telemetry or testing;

[0064] S403: Perform weighted summation on the new data-driven second loss function and the new physical constraint second loss function to obtain a new total second loss function:

[0065] Loss g =aLoss ms +bLoss py ;

[0066] Among them, a is the weight coefficient of the new data-driven second loss function, and b is the weight coefficient of the new physical constraint second loss function.

[0067] In step S5, the thermal parameters are adjusted based on the PINN model obtained in step S3 so that the second loss function of step S4 is minimized. The thermal parameters at this time are the corrected thermal parameters, which are substituted into the simulation model to obtain a corrected accurate thermal simulation model, including the following steps:

[0068] S501, to minimize Loss f That is min(Loss f) as the goal, the precise PINN model obtained in S3 is trained based on a temperature dataset matrix of experimentally measured on-orbit telemetry temperature values. No adjustments are made to the network structure, solver settings, or hyperparameters. Only the input parameters, namely thermal parameters such as infrared hemispherical emissivity ε, solar absorptivity α, thermal resistance r, heat transfer coefficient k, thermal conductivity λ, heat capacity C, and heat loss h, are adjusted. This ensures that the PINN model's predictions accurately match the on-orbit telemetry / experimental temperature results. The thermal parameters at this point are precisely corrected thermal parameters.

[0069] S502. Substitute the corrected thermal parameters obtained in S501 into the simulation software in sequence to obtain a thermal simulation prediction model that can accurately predict the on-orbit / test temperature, and obtain a PINN model that can accurately predict the on-orbit / test temperature.

[0070] Compared with the existing technology, this solution has the following beneficial effects:

[0071] 1. The present invention uses the PINN model instead of the simulation model for temperature field prediction and parameter optimization, rationally utilizing sparse simulation data and measured temperature data, greatly improving the efficiency of designing and revising thermal models, making them less expensive, enabling efficient processing and optimization, significantly reducing the demand for computing resources, and better meeting the increasingly shorter development cycle requirements of satellite payload thermal control subsystems.

[0072] 2. This invention improves the lack of interpretability of neural networks as "black boxes" by imposing thermal physical information constraints and utilizing physical information. By combining data information, physical mechanism information, and required expert prior knowledge, the thermal parameters obtained have certain physical meanings, conform to physical laws, and establish a thermal connection with the actual physical process of thermal science. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 1 is a flow chart of a method for satellite payload temperature prediction and thermal parameter correction based on PINN in an embodiment of the present invention;

[0074] Figure 2 is a schematic diagram of a cryogenic optical link in an embodiment of the present invention;

[0075] Figure 3 is a plan view of the heat dissipation surface of a cryogenic optical link according to an embodiment of the present invention;

[0076] Figure 4 FIG. 1 is a diagram of a PINN network structure in an embodiment of the present invention.

[0077] In the figure: 1. Heat dissipation surface; 2. Cold end of heat pipe; 3. Heat pipe; 4. Hot end of heat pipe; 5. Mirror; 6. Cold box; 7. Bottom plate; 8. Mirror frame. DETAILED DESCRIPTION

[0078] In order to enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described in detail below in conjunction with the embodiments of the present invention and the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0079] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the embodiments.

[0080] Example:

[0081] (1) Establishing an original thermal simulation model of the satellite payload, sampling the thermal parameters and performing simulation calculations to obtain a training data set;

[0082] Thermal simulation modeling was performed in simulation modeling software based on the physical model of the low-temperature optical link. Finite element analysis software was used to simplify the geometry and divide the mesh, and the temperature field was solved and analyzed to obtain the temperature field distribution of the heat dissipation surface 1 and the temperature level of the cold end 2 of the heat pipe.

[0083] Parameter identification is performed on the thermal parameters of all thermal models, including the infrared hemispherical emissivity of the heat dissipation surface 1, the solar absorptivity, the heat transfer coefficient between the cold end 2 of the heat pipe and the heat dissipation surface 1, the thermal resistance of the heat pipe, and other thermal parameters. The value range of the thermal parameters is determined, and Latin hypercube uniform sampling is performed. The parameters are substituted into the thermal simulation model for simulation calculation to obtain the data matrix of the temperature and coordinates of the heat dissipation surface 1.

[0084] (2) Establish the thermal physics mechanism equation of the satellite payload, construct the PINN model and assign thermal physics constraints to the model based on the thermal equation; establish the thermal control equation of the heat dissipation surface 1,

[0085]

[0086] Where a is the thickness of the heat dissipation surface 1, ε is the infrared hemispherical emissivity of the heat dissipation surface 1, σ is the Boltzmann constant, T u is a matrix of the absolute temperature distribution on the surface of the heat dissipation surface 1 and its position, where x is the abscissa in the Cartesian coordinate system and y is the ordinate.

[0087] like Figure 4 As shown in the figure, a PINN model is built on the basis of a general neural network model framework. The input units are the infrared hemispherical emissivity ε of the heat dissipation surface 1, the solar absorptivity α, the heat transfer coefficient k between the cold end 2 of the heat pipe and the heat dissipation surface 1, the heat conduction resistance r of the heat pipe, and other thermal parameters, as shown in the figure. Figure 3, data driven simulation temperature matrix T (x,y) , referred to as T. The output unit includes the predicted temperature of the heat dissipation surface 1 at the (x, y) position Hereinafter referred to as Calculate the first-order partial differential and second-order partial differential equations between temperature and coordinates, so the output unit also includes

[0088] Set the data-driven first loss function Loss data , the physical drive first loss function Loss PDE , hidden layer and its weight vector w ij , bias vector b ij and activation functions.

[0089] (3) Using the training data set of (1) as data-driven information and the thermal physical constraints of (2) as physical-driven information, the PINN model is trained and the first loss function is calculated to finally obtain an accurate PINN temperature prediction model;

[0090] The sparse simulation temperature calculation results and the coordinate (x, y) matrix are used as training data sets, with the coordinate data as input and the temperature calculation value as output, providing data-driven information for the training of the PINN model. The data-driven first loss function should be:

[0091]

[0092] The thermophysical constraints obtained by S2 are given to PINN in the form of the first loss function, which is determined by the output unit Calculate the first loss function of physical constraints as

[0093]

[0094] The total feeding loss function is obtained by weighted summing the data-driven first loss function and the physical constraint first loss function.

[0095] Loss f =αLoss data +βLoss PDE ;

[0096] Among them, α is the weight coefficient of the data-driven first loss function, and β is the weight coefficient of the physical constraint first loss function.

[0097] To minimize Loss f That is min(Loss f) as the goal, the PINN model built in S202 is trained based on the temperature data set matrix obtained in S105, and the optimization algorithm is used to optimize the network structure, solver settings and hyperparameter optimization, as well as feature extraction and data enhancement, so that the prediction results of the PINN model can accurately replace the calculation results of the simulation model. At this time, the PINN model is an accurate PINN temperature prediction model.

[0098] (4) Using the heat transfer formula as the new physical driving information and the on-orbit telemetry / test temperature as the new data driving information, the second loss function of the PINN model is obtained;

[0099] The second loss function of steady-state heat conduction is

[0100]

[0101] Where k is the heat transfer coefficient between the heat dissipation surface 1 and the cold end 2 of the heat pipe, A is the contact area, d is the thickness, Predicted temperature for PINN of heat sink 1, Predict the temperature for the PINN at the cold end 2 of the heat pipe.

[0102] The heat balance formula of heat dissipation surface 1 should also be satisfied

[0103] The second loss function of the thermal balance formula in steady state is

[0104]

[0105] The heat conduction formula of heat pipe 3 is

[0106]

[0107] Among them, T a is the temperature of the cold end 2 of the heat pipe, T b is the temperature of the heat pipe hot end 4, and r is the heat conduction resistance of the heat pipe itself. It is believed that the temperature of the heat pipe hot end can represent the temperature of the cold box 6 and its optical components 5 and 7.

[0108] The second loss function driven by physics should be

[0109] Loss py =mLoss1+nLoss2+sLoss3;

[0110] Among them, m, n, and s are the proportions of the second loss function of the three physical constraints respectively.

[0111] Taking the on-orbit telemetry or experimental measurement temperature as the new data-driven information, the data-driven second loss function is

[0112]

[0113] in, The temperature value predicted by the PINN model after adjusting the thermal parameters, T′ (x,y) Measure temperature values ​​for on-orbit telemetry or testing.

[0114] The data-driven second loss function and the physical constraint second loss function are weighted summed to obtain the new total second loss function:

[0115] Loss g =aLoss ms +bLoss py

[0116] Among them, a is the weight coefficient of the new data-driven second loss function, and b is the weight coefficient of the new physical constraint second loss function.

[0117] (5) Based on the PINN model obtained in (3), the thermal parameters are adjusted to minimize the second loss function in (4). The thermal parameters at this time are the corrected thermal parameters, which are substituted into the simulation model to obtain the corrected accurate thermal simulation model.

[0118] To minimize Loss f That is min(Loss f ) as the goal, the precise PINN model obtained in (3) is trained based on the temperature data set matrix of the experimental measurement / on-orbit telemetry temperature value, and the input parameters, namely the thermal parameters, are adjusted, including the infrared hemispherical emissivity ε of the heat dissipation surface 1, the solar absorptivity α, the heat transfer coefficient k between the cold end 2 of the heat pipe and the heat dissipation surface 1, the thermal resistance r of the heat pipe, and other thermal parameters, so that the prediction results of the PINN model can accurately match the temperature results of the on-orbit telemetry / experimental measurement. At this time, the thermal parameters are the precise corrected thermal parameters.

[0119] The corrected thermal parameters obtained in S501 are sequentially re-entered into the simulation software to obtain a thermal simulation prediction model that can accurately predict on-orbit / test temperatures, and a PINN model that can accurately predict on-orbit / test temperatures is obtained. The above specific embodiments are merely explanations of the present invention and are not intended to limit the present invention. After reading this specification, those skilled in the art may make modifications to the present embodiment as needed without contributing any creative ideas. However, as long as they fall within the scope of the claims of the present invention, they are protected by patent law.

Claims

1. A method for satellite payload temperature prediction and thermal parameter correction based on PINN, characterized by: The method comprises the following steps: S1. Establish an original thermal simulation model of the satellite payload, sample the thermal parameters therein and perform simulation calculations to obtain a training data set; S2. Establish the thermophysical mechanism equation of the satellite payload, construct the PINN model and assign thermophysical constraints to the model based on the thermal equation; S3, using the training data set of S1 as the first data driving information and the thermal physical constraints of S2 as the first physical driving information, train the PINN model, calculate the first loss function, and finally obtain an accurate PINN temperature prediction model; S4. Using the heat transfer formula as the second physical driving information and the on-orbit telemetry test temperature as the second data driving information, a second loss function of the PINN model is obtained; S5. Based on the PINN model of S3, adjust the thermal parameters to minimize the second loss function. The thermal parameters at this time are the corrected thermal parameters, which are substituted into the original thermal simulation model to obtain the corrected accurate thermal simulation model.

2. The method for satellite payload temperature prediction and thermal parameter correction based on PINN as claimed in claim 1, wherein: In step S1, the original thermal simulation model of the satellite payload is established, thermal parameters thereof are sampled and simulated to obtain a training data set, including the following steps: S101. Establishing an original thermal simulation model in simulation modeling software according to a physical model of a satellite payload; S102. Use finite element analysis software to simplify the geometry and divide the mesh; S103, performing temperature field analysis to obtain the temperature field distribution and temperature level of key parts of the satellite payload; S104, identifying thermal parameters of the thermal model, including but not limited to infrared hemispherical emissivity, solar absorptivity, thermal resistance, heat transfer coefficient, thermal conductivity, heat capacity, heat consumption and other thermal parameters of the surface; S105 , determining the value range of the thermal parameters in step S104 , performing Latin hypercube uniform sampling, and substituting the values ​​into the original thermal simulation model in step S101 for simulation calculation to obtain a data matrix of temperature and coordinates.

3. The method for satellite payload temperature prediction and thermal parameter correction based on PINN as claimed in claim 1, characterized in that: In step S2, the establishment of the thermophysical mechanism equation of the satellite payload, the construction of the PINN model, and the assignment of thermophysical constraints to the model based on the thermal equation include the following steps: S201. Establish universal thermal control equations for each component Where ρ is the density; φ is the general solution function, which is represented by the temperature matrix T in this problem; Γ φ is the generalized diffusion coefficient; λ is the thermal conductivity; S is the generalized source term, which is represented by the radiation heat transfer between the component and the space environment; The first term of the heat control equation is a non-steady-state term, the second term is a convection term, the third term is a diffusion term, and the fourth term is a source term. Since the satellite payload is located in a space environment and the heat transfer mode is thermal radiation and heat conduction, the non-steady-state term and the convection term can be ignored, and the first and second terms are 0. The control equation should be further expanded as follows: Where a is the thickness of the component, ε is the infrared hemispherical emissivity of the component surface, σ is the Boltzmann constant, and T u is the matrix of the absolute surface temperature distribution of the component and its position, T0 is the heat sink temperature, x is the abscissa in the Cartesian coordinate system, and y is the ordinate; S202. The PINN model should be built on the basis of a general neural network model framework. The input unit includes thermal parameters in the satellite payload thermal model, including the surface infrared hemispherical emissivity ε, solar absorptivity α, thermal resistance r, heat transfer coefficient k, thermal conductivity λ, heat capacity C, heat loss h and other thermal parameters, spatial coordinate vector (x, y), and the simulation temperature matrix T required for data driving. (x,y) , referred to as T; The PINN model output unit should include the predicted temperature at the (x,y) position Hereinafter referred to as The PINN model should be able to calculate the first-order partial differential and second-order partial differential equations between temperature and coordinates. Therefore, the output unit should also include The PINN model should have a data-driven first loss function Loss data , the physical drive first loss function Loss PDE ; The PINN model should have a hidden layer and its weight vector w ij , bias vector b ij and activation functions; S203, when the heat sink is 4K cold air, T0 in the control equation of S201 4 Relative to T u 4 can be neglected, so the thermophysical constraint should be 4. The method for satellite payload temperature prediction and thermal parameter correction based on PINN as claimed in claim 1, characterized in that: In step S3, the training data set of S1 is used as the first data driving information, the thermal physical constraints of S2 are used as the first physical driving information, the PINN model is trained, the first loss function is calculated, and finally an accurate PINN temperature prediction model is obtained, which includes the following steps: S301, the data matrix of simulated temperature and coordinates obtained in S1 is used as a training data set, with coordinate data as input and temperature calculation value as output, to provide data-driven information for the training of the PINN model. The data-driven first loss function should be: S302, the thermophysical constraints obtained in S2 are given to PINN in the form of the first loss function, and the output unit The first loss function for calculating physical constraints should be S303: Perform weighted summation of the data-driven first loss function and the physical constraint first loss function to obtain the total first loss function. Loss f =αLoss data +βLoss PDE ; Among them, α is the weight coefficient of the data-driven first loss function, and β is the weight coefficient of the physical constraint first loss function; S304, to minimize Loss f That is min(Loss f ) as the goal, the PINN model built in S202 is trained based on the temperature data obtained in S105, and the optimization algorithm is used to optimize the network structure, solver settings and hyperparameter optimization, as well as feature extraction and data enhancement, so that the prediction results of the PINN model can accurately replace the calculation results of the simulation model. At this time, the PINN model is an accurate PINN temperature prediction model.

5. The method for satellite payload temperature prediction and thermal parameter correction based on PINN as claimed in claim 1, characterized in that: In step S4, the heat transfer formula is used as the second physical driving information, and the on-orbit telemetry test temperature is used as the second data driving information to obtain a new second loss function of the PINN model, including the following steps: S401. Since there are thermal interfaces between components, the temperature fields of different components are coupled with each other and should satisfy the heat transfer formula: Steady-state heat conduction formula for the wall Where k is the heat transfer coefficient between components, A is the contact area, d is the thickness, and T i is the temperature of component i, T j is the temperature of component j; The second loss function of steady-state heat conduction on the wall is in, Predicted temperature for the PINN of component i, Predict the temperature of PINN for component j; Heat balance formula for a single component Where j is the component adjacent to component i, D ji is the inter-component conduction network coefficient, R ji is the radiation network coefficient between components, Q i is the total heat source of the component; The second loss function of the thermal balance formula in steady state Heat conduction formula for a single component Among them, T a is the temperature of temperature measuring point a, T b is the temperature of temperature measuring point b; The second loss function driven by physics should be Loss py =mLoss1+nLoss2+sLoss3; Among them, m, n, and s are the proportions of the second loss function of the three physical constraints respectively; The PINN model obtained in S402 and S3 no longer adjusts the network structure, hyperparameters and other information, but only adjusts the input parameters, namely the thermal parameters, and uses the on-orbit telemetry or experimental measurement temperature as the new data-driven information. The data-driven first loss function should be in, The temperature value predicted by the PINN model after adjusting the thermal parameters, T′ (x,y) Measuring temperature for on-orbit telemetry or testing; S403: Perform weighted summation of the data-driven second loss function and the physical constraint second loss function to obtain the total second loss function: Loss g =aLoss ms +bLoss py ; Among them, a is the weight coefficient of the new data-driven second loss function, and b is the weight coefficient of the new physical constraint second loss function.

6. The method for satellite payload temperature prediction and thermal parameter correction based on PINN as claimed in claim 1, characterized in that: In step S5, the thermal parameters are adjusted based on the PINN model obtained in step S3 so that the second loss function of step S4 is minimized. The thermal parameters at this time are the corrected thermal parameters, which are substituted into the simulation model to obtain a corrected accurate thermal simulation model, including the following steps: S501, to minimize Loss f That is min(Loss f ) as the goal, the accurate PINN model obtained in S3 is trained based on the temperature dataset matrix of the on-orbit telemetry temperature values ​​measured experimentally. No adjustments are made to the network structure, solver settings, and hyperparameters. Only the input parameters, i.e., thermal parameters such as infrared hemispherical emissivity ε, solar absorptivity α, thermal resistance r, heat transfer coefficient k, thermal conductivity λ, heat capacity C, and heat loss h, are adjusted. This allows the prediction results of the PINN model to accurately match the temperature results of on-orbit telemetry / experimental measurements. The thermal parameters at this time are the accurate corrected thermal parameters. S502. Substitute the corrected thermal parameters obtained in S501 into the simulation software in sequence to obtain a thermal simulation prediction model that can accurately predict the on-orbit / test temperature, and obtain a PINN model that can accurately predict the on-orbit / test temperature.

Images