Method, system and equipment for acquiring convective heat transfer coefficient of transformer based on PINN, and medium

Abstract

The invention discloses a PINN-based transformer convective heat transfer coefficient acquisition method, system and device and a medium, and relates to the technical field of transformers, and the method comprises the steps: building a three-dimensional thermal field simulation model based on transformer structure parameters, obtaining temperature field data through multi-physics field coupling simulation, building a sample set, training a physical information neural network model, and obtaining the convective heat transfer coefficient of a transformer. A loss function is defined, through weight adjustment and optimization algorithm training, loss change is dynamically monitored, convective heat transfer coefficient distribution is inversed, an inversion coefficient is substituted into a thermal model for forward calculation, iterative optimization is performed through comparison with independent measurement data, and accuracy and applicability are verified through actually measured temperature data. According to the method, the dependence on experimental data is reduced, the high-precision and high-efficiency acquisition of the convective heat transfer coefficient of the spatial change of the transformer is realized, and a reliable technical means is provided for thermal design and state evaluation of the transformer.

Description

Technical Field

[0001] This invention relates to the field of transformer technology, and specifically to a method, system, device, and medium for obtaining the convective heat transfer coefficient of a transformer based on PINN. Background Technology

[0002] Transformers are among the most important electrical devices in power systems, generating significant amounts of heat during operation due to factors such as iron and copper losses. Insufficient heat dissipation can lead to increased internal temperatures, resulting in insulation aging, decreased material performance, and even overheating faults, severely impacting the transformer's operational safety and service life. Therefore, accurately obtaining the heat transfer characteristics of different areas of the transformer, especially the convective heat transfer coefficient, is crucial for thermal design, operational status assessment, and lifespan prediction.

[0003] Existing methods for analyzing transformer heat dissipation mainly fall into two categories: one is based on empirical formulas, such as establishing correlations based on dimensionless criteria like Reynolds number and Nusselt number; these methods are simple but usually rely on idealized assumptions and are difficult to reflect the real heat transfer situation under complex transformer structures and non-uniform boundary conditions; the other is based on numerical simulation, such as using computational fluid dynamics (CFD) or finite element software to calculate the internal temperature and flow fields of the transformer; although these methods have high accuracy, they are complex to model, time-consuming to calculate, and sensitive to boundary conditions, often requiring a large amount of experimental data for correction, which limits their widespread application in engineering practice.

[0004] With the development of artificial intelligence and deep learning technologies, neural networks have demonstrated powerful capabilities in modeling nonlinear problems and approximating functions. However, traditional neural network methods rely solely on data-driven approaches and lack physical constraints, which may lead to predictions that do not conform to physical laws. Physical information neural networks, by directly incorporating physical laws such as governing equations and boundary conditions into the loss function, combine data-driven approaches with physical priors, enabling high-precision modeling of complex physical processes even with limited data. Summary of the Invention

[0005] In view of the above-mentioned existing problems, the present invention provides a method, system, device and medium for obtaining the convective heat transfer coefficient of transformers based on PINN, in order to solve the problems in the prior art that it is difficult to reflect the real heat transfer situation under the complex structure and non-uniform boundary conditions of transformers, the modeling is complicated, the calculation time is long, the prediction results are sensitive to boundary conditions and do not conform to physical laws.

[0006] To address the aforementioned technical issues, a method for obtaining the transformer convective heat transfer coefficient based on PINN is proposed, including: A three-dimensional thermal field simulation model was established based on the transformer's structural parameters. Multiphysics coupling simulation was performed by setting heat sources and boundary conditions to obtain temperature field data of the transformer in different regions, and a sample set was constructed. A physical information neural network model was built based on this sample set. A loss function was defined, and the loss was balanced through a weight adjustment mechanism. The physical information neural network model was trained using an optimization algorithm, and the changes in the loss function were dynamically monitored. Network parameters and training strategies were adjusted based on convergence to invert the convective heat transfer coefficient distribution in different regions of the transformer. The inverted convective heat transfer coefficient was substituted into the thermal model for forward calculation and compared with independent measurement data. Iterative optimization was performed based on error indicators. Transformer temperature data was measured using an experimental platform, and the measured results were compared with the neural network-predicted temperature to verify the accuracy and applicability of the convective heat transfer coefficient.

[0007] As a preferred embodiment of the PINN-based transformer convective heat transfer coefficient acquisition method of the present invention, the multi-physics field coupling simulation includes: calculating the temperature distribution of the transformer under operating conditions through a three-dimensional thermal field simulation model, dividing the data into training set and test set, and generating temperature field samples covering different regions of the transformer based on the interaction of heat conduction, convection and radiation.

[0008] As a preferred embodiment of the PINN-based transformer convective heat transfer coefficient acquisition method of the present invention, the construction of the physical information neural network model includes: embedding physical prior knowledge into the network structure, representing the convective heat transfer coefficient through learnable parameters, mapping spatial coordinates to the temperature field, performing data fitting under physical constraints, determining the number of network layers and neuron connection methods, and establishing a nonlinear mapping relationship from spatial coordinates to the temperature field.

[0009] As a preferred embodiment of the method for obtaining the transformer convection heat transfer coefficient based on PINN described in this invention, the definition of the loss function includes: combining data-driven terms and physical constraint terms, balancing each loss through a dynamic weight adjustment mechanism, determining the relative importance of each loss, and adjusting the balance between data fitting and physical constraints through weight coefficients.

[0010] As a preferred embodiment of the PINN-based transformer convection heat transfer coefficient acquisition method of the present invention, the step of training the physical information neural network model through optimization algorithms includes: pre-training with the Adam optimizer, adaptively adjusting the learning rate using the estimation characteristics of the first and second moments, and converging to a preliminary solution; when the change in the loss function is less than a preset threshold, switching to the L-BFGS optimizer for fine-tuning, using an approximate Hessian matrix to reflect the curvature characteristics of the loss function; monitoring the change in the loss function during training, and adjusting the network structure parameters or reallocating the proportion of training data when the loss decreases to a standstill; The method of training the physical information neural network model by optimizing the algorithm also includes, during the training process, when the root mean square error between the predicted temperature and the measured temperature exceeds a preset threshold, increasing the number of hidden layers or neurons in the neural network, reducing the learning rate to half of the original value, increasing the weight of the partial differential equation residual loss and boundary condition loss terms, and expanding the training set. The distribution of convective heat transfer coefficients in different regions of the inversion transformer includes obtaining the predicted temperature value of each point in space through a neural network, calculating the temperature gradient field, calculating the convective heat transfer coefficient of each point based on the material thermal conductivity, the dot product of the temperature gradient and the surface normal vector, and the difference between the surface temperature and the ambient temperature, and continuously executing the calculation throughout the entire solution domain to obtain the spatial distribution of the convective heat transfer coefficients.

[0011] As a preferred embodiment of the PINN-based transformer convective heat transfer coefficient acquisition method of the present invention, the forward calculation includes: substituting the inverted convective heat transfer coefficient into the heat conduction equation for forward verification calculation, solving the temperature field using the finite element method, and calculating the root mean square error between the predicted temperature and the measured temperature; when the root mean square error is greater than a preset threshold, the residual is introduced as a new constraint term into the loss function of the neural network, the network parameters are re-optimized, and the convective heat transfer coefficient distribution is updated until the error meets the accuracy requirements; The formula for solving the temperature field using the finite element method is expressed as: in, For the density of the material, The specific heat capacity of the material, For temperature, For time, The thermal conductivity of the material For a heat source per unit volume, For gradient operators; The updated formula for the distribution of convective heat transfer coefficient is expressed as follows: in, For convective heat flux density, The convective heat transfer coefficient is... For surface temperature, The ambient temperature; The formula for calculating the error is expressed as follows: in, The root mean square error, Let i be the number of temperature data points, and i be the index of the data point. This represents the temperature value predicted by the model for the i-th data point. Let be the temperature value measured in the experiment at the i-th data point.

[0012] As a preferred embodiment of the PINN-based transformer convective heat transfer coefficient acquisition method of the present invention, the verification of the accuracy and applicability of the convective heat transfer coefficient includes: building a transformer experimental platform, deploying distributed thermocouple temperature sensors on the core, windings and shell surfaces, measuring the temperature distribution under different operating conditions, comparing the measured temperature data with the neural network predicted temperature point by point, calculating the temperature error of each region, and verifying the accuracy and applicability of the obtained convective heat transfer coefficient in engineering applications.

[0013] The beneficial effects of this preferred technical solution are as follows: by continuously calculating the convective heat transfer coefficient at each point in space based on the temperature gradient field, material thermal conductivity, and temperature difference, a physically consistent mapping from the temperature field to the convective heat transfer coefficient field is realized, a quantitative relationship between temperature measurement and heat transfer characteristics is directly established, and a heat transfer coefficient with clear physical meaning and continuous spatial distribution is obtained, providing a foundation for accurate thermal analysis.

[0014] As a preferred embodiment of the transformer convection heat transfer coefficient acquisition system based on PINN described in this invention, it is characterized by including a multiphysics simulation and sample generation module, a physical information neural network construction module, a hybrid optimization training module, and a coefficient inversion and verification feedback module.

[0015] The multiphysics simulation and sample generation module is used to perform magnetic-thermal coupling simulation using finite element software based on the actual physical structure and material properties of the transformer, and to calculate the steady-state and transient temperature field distribution inside and on the surface of the transformer under given loss and boundary conditions.

[0016] The physical information neural network construction module is used to take spatial coordinates as input, temperature as output, and use the unknown quantity to be solved, namely the spatially varying convective heat transfer coefficient, as a learnable parameter inside the network. The heat conduction equation describing energy conservation and the physical rules of convective heat transfer boundary conditions are encoded into the network's loss function in the form of residual loss.

[0017] The hybrid optimization training module is used to pre-train using the Adam optimizer, switch to the L-BFGS optimizer based on second-order derivative information for fine-tuning, and dynamically monitor and adjust the weight balance between data loss and loss of each physical constraint during training.

[0018] The coefficient inversion and verification feedback module is used to extract the converged convective heat transfer coefficients from the network and substitute them into the complete heat model for forward calculation to obtain the predicted temperature field. By calculating the root mean square error between the predicted temperature and the measured temperature that did not participate in the training, the accuracy of the inversion coefficients is quantitatively evaluated, and iterative optimization is performed until the prediction results meet the engineering accuracy requirements.

[0019] A computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method for obtaining the transformer convective heat transfer coefficient based on PINN.

[0020] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for obtaining the PINN-based transformer convective heat transfer coefficient.

[0021] The beneficial effects of this invention are as follows: By establishing a three-dimensional thermal field simulation model and performing multi-physics coupled simulation, this invention generates temperature field sample data covering different regions of a transformer, overcoming the dependence of traditional methods on a large amount of experimental data; by embedding physical prior knowledge into the network structure and using learnable parameters to represent the convective heat transfer coefficient, it achieves a deep integration of physical laws and data-driven approaches, and can still obtain physically reasonable solutions under limited data conditions; by balancing the loss functions of data-driven terms and physical constraint terms through a dynamic weight adjustment mechanism, it improves training stability and convergence efficiency; by adopting a hybrid training strategy of Adam and L-BFGS optimizers, it balances training speed and convergence accuracy; by monitoring prediction errors and adaptively adjusting the network structure, learning rate, and physical constraint weights, it enhances the robustness and adaptability of the method to operating conditions; by continuously inverting the convective heat transfer coefficients at each point in space based on the temperature gradient field and physical relationships, it obtains a heat transfer characteristic distribution with clear physical meaning; by constructing a self-correcting closed-loop verification mechanism through forward computation and error iterative optimization, it ensures that the results conform to both physical constraints and measured data; and by building an experimental platform for distributed temperature measurement verification, it confirms the accuracy and practicality in engineering applications. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the overall process of obtaining the convective heat transfer coefficient of a transformer based on PINN, as provided in one embodiment of the present invention.

[0024] Figure 2 A schematic diagram of a single-phase two-column transformer simulation model for a method of obtaining the convective heat transfer coefficient of a transformer based on PINN, provided in an embodiment of the present invention.

[0025] Figure 3 This is a schematic diagram of the thermal field calculation results of a single-phase two-column transformer, based on a method for obtaining the convective heat transfer coefficient of a transformer using PINN, as provided in an embodiment of the present invention.

[0026] Figure 4 This is a schematic diagram of the placement of distributed thermocouples in a method for obtaining the convective heat transfer coefficient of a transformer based on PINN, provided in an embodiment of the present invention.

[0027] Figure 5 The flowchart shows a system scheme for obtaining the convective heat transfer coefficient of a transformer based on PINN, as provided in one embodiment of the present invention. Detailed Implementation

[0028] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0029] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for obtaining the convective heat transfer coefficient of a transformer based on PINN is provided, comprising: S100: Establish a three-dimensional thermal field simulation model based on the transformer's structural parameters. Perform multi-physics field coupling simulation by setting heat sources and boundary conditions to obtain temperature field data of the transformer in different regions and construct a sample set. Based on the sample set, construct a physical information neural network model.

[0030] S200: Defines the loss function and balances the loss through a weight adjustment mechanism. It trains the physical information neural network model through an optimization algorithm, dynamically monitors the changes in the loss function, and adjusts the network parameters and training strategy according to the convergence status to invert the convective heat transfer coefficient distribution in different regions of the transformer.

[0031] S300: The convective heat transfer coefficient obtained by inversion is substituted into the thermal model for forward calculation and compared with independent measurement data. Iterative optimization is performed based on the error index. The transformer temperature data is measured through the experimental platform, and the measured results are compared with the temperature predicted by the neural network to verify the accuracy and applicability of the convective heat transfer coefficient.

[0032] It should be noted that by embedding prior physical knowledge into the network structure and dynamically balancing the loss function, a deep integration of data-driven approaches and physical mechanisms is achieved, ensuring the physical rationality of the inversion results. With the help of a hybrid optimization strategy of Adam and L-BFGS and adaptive training adjustment, the convective heat transfer coefficients of the transformer with continuous distribution across the entire domain are inverted. Through closed-loop iterative optimization of inversion-verification-correction, and verified by experimental platform, the accuracy and reliability of the coefficients are significantly improved, providing high-fidelity spatialized parameters for transformer thermal management.

[0033] Example 2, refer to Figures 1-4 This is a second embodiment of the present invention, which provides a method for obtaining the convective heat transfer coefficient of a transformer based on PINN, including: In this embodiment of the application, step S100, the multiphysics coupling simulation includes steps S101 to S104: S101: Based on the actual structure of the transformer, the main components such as the core and windings are reasonably simplified to construct a three-dimensional simulation model of the transformer. Each component in the model is given precise material physical properties, including density, conductivity, specific heat capacity and relative permeability, laying the foundation for subsequent electromagnetic-thermal coupling calculations.

[0034] S102: The core loss of the transformer under no-load conditions is analyzed through magnetic field simulation, and the DC resistance loss of the winding is obtained by analytical calculation. In order to accurately convert the loss into the input conditions for thermal simulation, the loss density per unit volume (internal heat source density) is calculated as the heat source boundary condition for temperature field simulation. The formula is expressed as: in, To determine the internal heat source density in the corresponding solution domain, The corresponding electromagnetic loss value in the solution domain obtained from the simulation. The volume of the corresponding solution domain.

[0035] S103: To accurately simulate heat transfer, a multi-physics field non-isothermal flow interface technology is adopted to establish a two-way coupling mechanism between the solid domain and the fluid domain (air). Taking into account the three heat transfer modes of heat conduction, heat convection and heat radiation, after setting the ambient temperature consistent with the experiment, the finite element software is used to solve and obtain the full-domain temperature field distribution data of the transformer under the operating conditions. The heat conduction formula is expressed as: in, For temperature, , and Let x be the thermal conductivity of the material in the x, y, and z directions. The internal heat source is the heat generated per unit volume inside the transformer. For the density of the material, The specific heat capacity of the material, For time; The heat convection formula is expressed as: in, Heat flux density is the amount of heat exchanged per unit time through a unit surface area. The convective heat transfer coefficient is... The temperature on the high-temperature side is the temperature of the solid surface. This refers to the temperature on the low-temperature side, i.e., the mainstream temperature of the fluid. The formula for thermal radiation is expressed as: in, This refers to the heat flow transferred between two radiating surfaces. For Stefan-Boltzmann constant, and Let be the thermodynamic temperatures of the two radiating surfaces. and Given the emissivity of the two surfaces, and Let be the radii of the two radiating surfaces.

[0036] S104: In the heat transfer process of an oil-immersed transformer, the heat generated by each component is transferred to the insulating oil and then dissipated to the external environment through the tank wall. The Nusselt number is introduced for calculation, and the formula is expressed as: in, The Nusselt number is the ratio of the intensity of convective heat transfer to that of pure conduction. For characteristic length, The convective heat transfer coefficient is... The thermal conductivity of the fluid; For dry-type transformers, the fluid acting as the insulating medium is air, and the generated heat is dissipated through the air. The formulas for calculating the Nusselt number in the horizontal and vertical directions are as follows: in, The Nusselt number of the horizontal surface. The Nusselt number is the number of the perpendicular surface. The Rayleigh number is a key dimensionless number that describes the intensity of natural convection. Prandtl number is a dimensionless number that describes the properties of fluids. For Grashof numbers, It is the acceleration due to gravity. is the coefficient of volumetric expansion of the fluid. The temperature difference between the wall and the external fluid. The dynamic viscosity of the fluid. The specific heat capacity at constant pressure of the fluid. The thermal conductivity of the fluid; Fluid flow follows the continuity equation, the energy conservation equation, and the momentum conservation equation, expressed as follows: in, For fluid density, For Hamiltonian operators, The velocity vector of the fluid. The temperature of the fluid, The thermal conductivity of the fluid. The specific heat capacity at constant pressure of the fluid. This is a viscous dissipation term. , and velocity vector Components in the x, y, and z directions, The pressure of the fluid. , and Force per unit mass volume, , , , , , , , and These are the components of the viscous stress tensor.

[0037] In an optional implementation, step S100 further includes establishing a transformer fluid-solid coupling model, setting flow boundary conditions for insulating oil or air, simulating fluid flow and heat transfer processes by solving the Navier-Stokes equations and energy equations, discretizing the computational domain using the finite volume method, iteratively calculating the temperature field and flow field, and outputting temperature distribution data for subsequent analysis.

[0038] In another optional implementation, in step S100, the multiphysics coupling simulation may further include decomposing the transformer into multiple thermal nodes (such as the core, windings, and casing), constructing an equivalent circuit based on thermal resistance and thermal capacity, calculating the heat flow and temperature between nodes using empirical formulas, and solving for the node temperature using circuit simulation software to generate simplified temperature field data.

[0039] Furthermore, in this embodiment of the application, step S100, the multiphysics coupling simulation includes steps S111~S114: S111: Define the neural network input as spatial coordinates (x, y, z) and the output as temperature value.

[0040] S112: The convective heat transfer coefficient is embedded as a learnable parameter into the hidden layer of the network, and the temperature is predicted by forward propagation.

[0041] S113: The heat conduction equation and convective heat transfer boundary conditions are directly integrated into the network training process as physical constraints.

[0042] S114: Adjust network weights and learnable parameters through backpropagation so that the output simultaneously satisfies data fitting and physical laws.

[0043] In an optional implementation, step S100 further includes converting the temperature field data into an image format, inputting it into a CNN for feature extraction, and learning spatial patterns through convolutional and pooling layers to output a temperature prediction.

[0044] In another optional implementation, in step S100, the multiphysics coupling simulation may further include extracting the transformer spatial coordinates and heat source features as input, mapping them to a high-dimensional space using a kernel function, predicting the temperature distribution by optimizing the support vector regression loss function, and calibrating the model parameters by combining a small amount of experimental data.

[0045] In this embodiment of the application, in step S200, defining the loss function includes steps S201 to S203: S201: Combined data loss term (calculates the error between predicted and measured temperatures), partial differential equation residual loss term (ensures the temperature field satisfies the heat conduction equation), boundary condition loss term (forced convection heat transfer boundary conforms to physical laws), and initial value loss term (matches initial temperature conditions). Mass conservation loss: For fluid mass conservation, the divergence must be zero. Using the residual square constraint, the formula is expressed as: in, For loss due to conservation of mass, Let j be the number of sampling points used for mass conservation loss, and j be the index of the sampling point. For fluid density, and Let x be the components of the fluid velocity vector in the x and y directions. and Let j be the spatial coordinates of the sampling point. The time for sampling point j; The momentum loss due to conservation of momentum, the momentum equation in the x-direction is expressed as: in, The pressure of the fluid. Loss due to conservation of momentum The number of sampling points for momentum conservation loss. The dynamic viscosity of the fluid. For time; Energy conservation losses, considering heat conduction and convection in the air, are expressed by the following formula: in, Loss due to energy conservation The number of sampling points is the energy conservation loss. The thermal conductivity of the fluid. is the specific heat capacity at constant volume of the fluid. For temperature; The heat loss due to convective boundary conditions at the outer wall is due to the heat exchange between the outer wall and the ambient temperature of 293.15 K via a heat transfer rate h, expressed by the formula: in, The thermal conductivity of the solid wall surface. Loss due to convection boundary conditions on the outer wall. The number of sampling points for the convection boundary condition loss on the outer wall. Let be the temperature gradient along the direction normal to the wall. The convective heat transfer coefficient; The loss in the heat conduction equation is expressed as: in, Specific heat capacity of solid materials The internal heat source is the heat generated per unit volume inside the transformer. Let the divergence of the heat conduction term be denoted as . For the loss in the heat conduction equation, The number of sampling points lost in the heat conduction equation.

[0046] S202: Dynamically adjust the weight coefficients of each item, and balance the data fitting and physical constraints according to the training phase. S203: The total loss is obtained through weighted summation to guide network optimization. The total loss is expressed as: in, For data loss, For the loss in the heat conduction equation, For boundary condition loss, This is the initial residual loss. , , and The weights of the loss terms are dynamically adjusted based on the training phase.

[0047] In an optional implementation, in step S200, defining the loss function further includes calculating the mean squared error between the predicted temperature and the target temperature, minimizing the error through gradient descent, and relying on a large amount of data to avoid overfitting.

[0048] In another optional implementation, in step S200, defining the loss function may further include adding an L2 regularization term to the MSE loss to penalize excessively large network weights, controlling the regularization strength by fixing the weight coefficients to reduce the risk of overfitting, and focusing on data smoothness during the training process.

[0049] Furthermore, in this embodiment of the application, in step S200, the training of the physical information neural network model through the optimization algorithm includes steps S211~S213: S211: The Adam optimizer is used for pre-training, and its first and second moment estimation characteristics are used to adaptively adjust the learning rate to quickly converge to the initial solution; S212: When the change in the loss function is less than the preset threshold, switch to the L-BFGS optimizer for fine-tuning. The curvature characteristics of the loss function are reflected by an approximate Hessian matrix, which further reduces the loss value. S213: Continuously monitor the changes in the loss function during training. When the loss stops decreasing, adjust the network structure parameters or reallocate the proportion of training data.

[0050] In an optional implementation, in step S200, training the physical information neural network model using the optimization algorithm further includes iteratively updating the network weights using SGD, adding a momentum term to accelerate convergence, and gradually reducing the learning rate using a fixed learning rate scheduler. The training process relies on a large number of iterations to ensure stability.

[0051] In another optional implementation, in step S200, training the physical information neural network model through the optimization algorithm may further include treating the network parameters as particle positions, searching for the optimal solution through group collaboration, iteratively updating the particle velocity and position, minimizing the loss function, and combining global and local search to balance exploration and utilization.

[0052] It should be noted that in step S213, during the training process, when the root mean square error between the predicted temperature and the measured temperature exceeds 1.5K, an adjustment step is performed to increase the number of hidden layers or neurons in the neural network, reduce the learning rate to 0.5 of the original value, increase the weight coefficients of the partial differential equation residual loss and boundary condition loss terms, strengthen the physical constraint effect, expand the training dataset or enhance the sample space coverage. The formula for minimizing the difference between the predicted and measured temperatures is expressed as: in, The root mean square error, Let i be the number of temperature data points, and i be the index of the data point. This represents the temperature value predicted by the model for the i-th data point. Let be the temperature value measured experimentally at the i-th data point; when The model is considered optimal when the change is less than 0.01K for five consecutive iterations. If the model's prediction accuracy is insufficient, it needs to be readjusted by increasing the number of hidden layers or neurons in the neural network, reducing the learning rate to 0.5 of the original value, increasing the weight coefficients of the partial differential equation residual loss and boundary condition loss terms, strengthening the physical constraint effect, expanding the training dataset, or enhancing the sample space coverage.

[0053] The distribution of convective heat transfer coefficients in different regions of the inversion transformer includes obtaining the predicted temperature value of each point in space through a neural network, calculating the temperature gradient field, calculating the convective heat transfer coefficient of each point based on the material thermal conductivity, the dot product of the temperature gradient and the surface normal vector, and the difference between the surface temperature and the ambient temperature, and continuously executing the calculation throughout the entire solution domain to obtain the spatial distribution of the convective heat transfer coefficients. The inversion formula is: in, Let X be the convective heat transfer coefficient at a point X in space. These are spatial coordinates, representing a specific location point inside the transformer. Let Z be the thermal conductivity of the solid material at position Z. Let Z be the gradient of the temperature field at position Z. Let Z be the unit normal vector of the solid surface at position Z. The ambient temperature, This is the solid surface temperature predicted at position Z by a trained physical information neural network.

[0054] In step S300, the forward calculation includes substituting the inverted convective heat transfer coefficient into the heat conduction equation for forward verification calculation, solving the temperature field using the finite element method, and calculating the root mean square error between the predicted temperature and the measured temperature; when the root mean square error is greater than 1K, the residual is introduced as a new constraint term into the loss function of the neural network, the network parameters are re-optimized, and the convective heat transfer coefficient distribution is updated until the error meets the accuracy requirements; The formula for solving the temperature field using the finite element method is expressed as: in, For the density of the material, The specific heat capacity of the material, For temperature, For time, The thermal conductivity of the material For a heat source per unit volume, For gradient operators; The updated formula for the distribution of convective heat transfer coefficient is expressed as follows: in, For convective heat flux density, The convective heat transfer coefficient is... For surface temperature, The ambient temperature; The formula for calculating the error is expressed as follows: in, The root mean square error, Let i be the number of temperature data points, and i be the index of the data point. This represents the temperature value predicted by the model for the i-th data point. Let be the temperature value measured in the experiment at the i-th data point.

[0055] The verification of the accuracy and applicability of the convective heat transfer coefficient includes building a transformer experimental platform, deploying distributed thermocouple temperature sensors on the core, windings and shell surfaces, measuring the temperature distribution under different operating conditions, comparing the measured temperature data with the neural network predicted temperature point by point, calculating the temperature error of each region, and verifying the accuracy and applicability of the obtained convective heat transfer coefficient in engineering applications.

[0056] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0057] Example 3, referring to Figure 5 The third embodiment of the present invention provides a transformer convection heat transfer coefficient acquisition system based on PINN, including a multiphysics simulation and sample generation module, a physical information neural network construction module, a hybrid optimization training module, and a coefficient inversion and verification feedback module.

[0058] The multiphysics simulation and sample generation module is used to perform magnetic-thermal coupling simulation using finite element software based on the actual physical structure and material properties of the transformer, and to calculate the steady-state and transient temperature field distribution inside and on the surface of the transformer under given loss and boundary conditions.

[0059] The physical information neural network construction module is used to take spatial coordinates as input, temperature as output, and use the unknown quantity to be solved, namely the spatially varying convective heat transfer coefficient, as a learnable parameter inside the network. The heat conduction equation describing energy conservation and the physical rules of convective heat transfer boundary conditions are encoded into the network's loss function in the form of residual loss.

[0060] The hybrid optimization training module is used to pre-train using the Adam optimizer, switch to the L-BFGS optimizer based on second-order derivative information for fine-tuning, and dynamically monitor and adjust the weight balance between data loss and loss of each physical constraint during training.

[0061] The coefficient inversion and verification feedback module is used to extract the converged convective heat transfer coefficients from the network and substitute them into the complete heat model for forward calculation to obtain the predicted temperature field. By calculating the root mean square error between the predicted temperature and the measured temperature that did not participate in the training, the accuracy of the inversion coefficients is quantitatively evaluated, and iterative optimization is performed until the prediction results meet the engineering accuracy requirements.

[0062] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0063] Example 4, the fourth embodiment of the present invention, differs from the previous three embodiments in that: If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0064] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0065] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0066] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

Claims

1. A method for obtaining the convective heat transfer coefficient of a transformer based on PINN, characterized in that: include, A three-dimensional thermal field simulation model is established based on the transformer's structural parameters. Multi-physics field coupling simulation is performed by setting heat sources and boundary conditions to obtain temperature field data of the transformer in different regions. A sample set is then constructed, and a physical information neural network model is built based on the sample set. Define a loss function and balance the loss through a weight adjustment mechanism. Train the physical information neural network model through an optimization algorithm, dynamically monitor the changes in the loss function, and adjust the network parameters and training strategy according to the convergence situation to invert the convective heat transfer coefficient distribution in different regions of the transformer. The convective heat transfer coefficient obtained by inversion is substituted into the thermal model for forward calculation and compared with independent measurement data. Iterative optimization is performed based on the error index. The transformer temperature data is measured through the experimental platform, and the measured results are compared with the temperature predicted by the neural network to verify the accuracy and applicability of the convective heat transfer coefficient.

2. The method for obtaining the transformer convective heat transfer coefficient based on PINN as described in claim 1, characterized in that: The multiphysics coupling simulation includes calculating the temperature distribution of the transformer under operating conditions using a three-dimensional thermal field simulation model, dividing the data into training and testing sets, and generating temperature field samples covering different regions of the transformer based on the interaction of heat conduction, convection, and radiation.

3. The method for obtaining the transformer convective heat transfer coefficient based on PINN as described in claim 2, characterized in that: The construction of the physical information neural network model includes embedding prior physical knowledge into the network structure, representing the convective heat transfer coefficient through learnable parameters, mapping spatial coordinates to the temperature field, performing data fitting under physical constraints, determining the number of network layers and the neuron connection method, and establishing a nonlinear mapping relationship from spatial coordinates to the temperature field.

4. The method for obtaining the transformer convective heat transfer coefficient based on PINN as described in claim 3, characterized in that: The defined loss function includes combining data-driven terms and physical constraint terms, balancing each loss through a dynamic weight adjustment mechanism, determining the relative importance of each loss, and adjusting the balance between data fitting and physical constraints through weight coefficients.

5. The method for obtaining the transformer convective heat transfer coefficient based on PINN as described in claim 4, characterized in that: The training of the physical information neural network model by the optimization algorithm includes: pre-training with the Adam optimizer, adaptively adjusting the learning rate using the estimation characteristics of the first and second moments, and converging to the preliminary solution; when the change in the loss function is less than a preset threshold, switching to the L-BFGS optimizer for fine adjustment, and reflecting the curvature characteristics of the loss function through an approximate Hessian matrix. During training, monitor changes in the loss function. When the loss decreases to a standstill, adjust the network structure parameters or reallocate the proportion of training data. The method of training the physical information neural network model by optimizing the algorithm also includes, during the training process, when the root mean square error between the predicted temperature and the measured temperature exceeds a preset threshold, increasing the number of hidden layers or neurons in the neural network, reducing the learning rate to half of the original value, increasing the weight of the partial differential equation residual loss and boundary condition loss terms, and expanding the training set. The distribution of convective heat transfer coefficients in different regions of the inversion transformer includes obtaining the predicted temperature value of each point in space through a neural network, calculating the temperature gradient field, calculating the convective heat transfer coefficient of each point based on the material thermal conductivity, the dot product of the temperature gradient and the surface normal vector, and the difference between the surface temperature and the ambient temperature, and continuously executing the calculation throughout the entire solution domain to obtain the spatial distribution of the convective heat transfer coefficients.

6. The method for obtaining the transformer convective heat transfer coefficient based on PINN as described in claim 5, characterized in that: The forward calculation includes substituting the inverted convective heat transfer coefficient into the heat conduction equation for forward verification calculation, solving the temperature field using the finite element method, and calculating the root mean square error between the predicted temperature and the measured temperature. When the root mean square error is greater than a preset threshold, the residual is introduced as a new constraint term into the loss function of the neural network, the network parameters are re-optimized, and the convective heat transfer coefficient distribution is updated until the error meets the accuracy requirements. The formula for solving the temperature field using the finite element method is expressed as: in, For the density of the material, The specific heat capacity of the material, For temperature, For time, The thermal conductivity of the material For a heat source per unit volume, For gradient operators; The updated formula for the distribution of convective heat transfer coefficient is expressed as follows: in, For convective heat flux density, The convective heat transfer coefficient is... Surface temperature, The ambient temperature; The formula for calculating the error is expressed as follows: in, The root mean square error, Let i be the number of temperature data points, and i be the index of the data point. This represents the temperature value predicted by the model for the i-th data point. Let be the temperature value measured in the experiment at the i-th data point.

7. The method for obtaining the transformer convective heat transfer coefficient based on PINN as described in claim 6, characterized in that: The verification of the accuracy and applicability of the convective heat transfer coefficient includes building a transformer experimental platform, deploying distributed thermocouple temperature sensors on the core, windings and shell surfaces, measuring the temperature distribution under different operating conditions, comparing the measured temperature data with the neural network predicted temperature point by point, calculating the temperature error of each region, and verifying the accuracy and applicability of the obtained convective heat transfer coefficient in engineering applications.

8. A transformer convective heat transfer coefficient acquisition system based on PINN, employing the transformer convective heat transfer coefficient acquisition method based on PINN as described in any one of claims 1 to 7, characterized in that, It includes a multiphysics simulation and sample generation module, a physical information neural network construction module, a hybrid optimization training module, and a coefficient inversion and verification feedback module; The multiphysics simulation and sample generation module is used to perform magnetic-thermal coupling simulation using finite element software based on the actual physical structure and material properties of the transformer, and to calculate the steady-state and transient temperature field distribution inside and on the surface of the transformer under given loss and boundary conditions. The physical information neural network construction module is used to take spatial coordinates as input, temperature as output, and take the unknown quantity to be solved, namely the spatially varying convective heat transfer coefficient, as a learnable parameter inside the network. It also encodes the heat conduction equation describing energy conservation and the physical rules of convective heat transfer boundary conditions into the network's loss function in the form of residual loss. The hybrid optimization training module is used to pre-train using the Adam optimizer, switch to the L-BFGS optimizer based on second derivative information for fine-tuning, and dynamically monitor and adjust the weight balance between data loss and loss of each physical constraint during training. The coefficient inversion and verification feedback module is used to extract the converged convective heat transfer coefficients from the network and substitute them into the complete heat model for forward calculation to obtain the predicted temperature field. By calculating the root mean square error between the predicted temperature and the measured temperature that did not participate in the training, the accuracy of the inversion coefficients is quantitatively evaluated, and iterative optimization is performed until the prediction results meet the engineering accuracy requirements.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method for obtaining the transformer convection heat transfer coefficient based on PINN as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for obtaining the PINN-based transformer convection heat transfer coefficient as described in any one of claims 1 to 7.

Images