Methods, devices, systems, and storage media for predicting the electromagnetic field of power transformer cores

By integrating the finite element method with a physical information neural network model, the problem of efficient prediction of electromagnetic field distribution in the core of a power transformer was solved. This model enables rapid and accurate prediction of electromagnetic field distribution under complex excitation conditions and has good generalization ability.

CN123116891APending Publication Date: 2026-11-17NORTH CHINA ELECTRIC POWER UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611635492.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-22
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies suffer from low computational efficiency and difficulty in achieving high-precision predictions when solving for the electromagnetic field distribution of power transformer cores, especially under complex excitation conditions. In particular, repeated iterations of high-dimensional sparse matrices lead to efficiency bottlenecks.

Method used

By integrating the physical supervision mechanism of the finite element method with the rapid reasoning capability of neural networks, a finite element-enhanced physical information neural network model is constructed, which combines physical constraints and automatic differentiation techniques to achieve rapid prediction of electromagnetic field distribution.

Benefits of technology

It achieves rapid and accurate prediction of the electromagnetic field distribution of transformer core under different complex excitation conditions, and has good generalization ability and computational efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116891A_ABST
    Figure CN123116891A_ABST
Patent Text Reader

Abstract

This invention discloses a method, device, system, and storage medium for predicting the electromagnetic field of a power transformer core, belonging to the field of power equipment condition monitoring technology. The method includes: obtaining the peak voltage, dynamic characteristics and spatial coordinates of the DC bias current, and static characteristics of the core's geometric dimensions and material properties corresponding to the excitation conditions based on the transformer core type and operating parameters; encoding the core control equations of the finite element method into differentiable physical residual modules embedded in the network according to the operating condition type, constructing a finite element-enhanced physical information neural network model, and constructing physical constraints based on the finite element solver; obtaining a prediction model based on the input features and the neural network model training; performing predictions for different operating conditions on the power transformer core model to be predicted; and calculating the electromagnetic field distribution using the automatic differentiation technology of the physical calculation layer module. This invention combines computational efficiency and accuracy in predicting the electromagnetic field distribution of transformer cores under complex excitation conditions and exhibits good generalization ability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of power equipment condition monitoring technology, specifically relating to a method, device, system, and storage medium for predicting the electromagnetic field of a power transformer core. Background Technology

[0002] Power transformers are the core hubs for energy conversion and transmission in power systems, and their operational reliability directly affects the safety and stability of the power grid. The core, as the magnetic circuit carrier of the transformer, has electromagnetic field distribution characteristics that are key factors determining the transformer's iron loss, temperature rise, and vibration and noise levels. With the increasing complexity of design parameters and structural models for transformers and other electrical equipment, quickly and accurately obtaining the electromagnetic field distribution of the transformer core under complex operating conditions is of paramount engineering significance for optimizing product design, assessing operating status, and preventing faults.

[0003] In the field of traditional numerical methods for solving electromagnetic field problems, the finite difference method, finite volume method, and finite element method can solve electromagnetic fields with high accuracy. However, for the periodic steady-state operating conditions of transformers, time-domain simulations require a long transient decay process, resulting in high computational costs. The harmonic balance finite element method achieves efficient solutions to steady-state fields through frequency domain transformation, but its iterative high-dimensional sparse matrix still faces efficiency bottlenecks in multi-condition scanning. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a method, device, system, and storage medium for predicting the electromagnetic field of a power transformer core. By integrating the physical supervision mechanism of the finite element method with the rapid reasoning capability of neural networks, it achieves both computational efficiency and accuracy in predicting the electromagnetic field distribution of the transformer core under complex excitation conditions, and exhibits good generalization ability.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A method for predicting the electromagnetic field of a power transformer core includes:

[0007] Based on the transformer core type and operating parameters, obtain the voltage peak value, dynamic characteristics and spatial coordinates of DC bias current, and static characteristics of core geometry and material properties corresponding to the excitation conditions.

[0008] According to the working condition type, the core control equations of the finite element method are encoded into differentiable physical residual modules and embedded into the network to construct a finite element enhanced physical information neural network model; among them, the physical constraints of the physical information neural network model are constructed based on the finite element solver.

[0009] Dynamic and static features are input into the physical information neural network model to train a prediction model. Based on the prediction model, different operating conditions are predicted for the core model of the power transformer to be predicted, and the electromagnetic field distribution is calculated using the automatic differentiation technology of the physical computing layer module.

[0010] Preferably, the construction of physical constraints based on the finite element solver includes:

[0011]

[0012]

[0013] in, Let be the magnetic vector potential loss function and the magnetic induction intensity loss function for the i-th operating condition, respectively. This is the magnetic vector potential result of the j-th coordinate point under the i-th working condition, predicted by the finite element-enhanced physical information neural network. This refers to the magnetic vector result of the j-th coordinate point under the i-th working condition, calculated by the physical residual module. This is the magnetic field strength at the j-th coordinate point under the i-th working condition, predicted by a finite element-enhanced physical information neural network. The magnetic field strength at the j-th coordinate point under the i-th operating condition, calculated by the physical computing layer. For parameters related to the neural network, c j = (x, y) are the coordinates of the points.

[0014] The present invention also provides a device for predicting the electromagnetic field of a power transformer core, comprising:

[0015] The first processing module is used to obtain the voltage peak value, dynamic characteristics and spatial coordinates of DC bias current, and static characteristics of core geometry and material properties corresponding to the excitation conditions, based on the transformer core type and operating parameters.

[0016] The second processing module is used to encode the core control equations of the finite element method into differentiable physical residual modules and embed them into the network according to the working condition type, thereby constructing a finite element-enhanced physical information neural network model; wherein, the physical constraints of the physical information neural network model are constructed based on the finite element solver.

[0017] The third processing module is used to input dynamic and static features into the physical information neural network model, train it to obtain a prediction model, predict different operating conditions of the core model of the power transformer to be predicted based on the prediction model, and use the automatic differentiation technology of the physical calculation layer module to calculate the electromagnetic field distribution.

[0018] Preferably, the construction of physical constraints based on the finite element solver includes:

[0019]

[0020]

[0021] in, Let be the magnetic vector potential loss function and the magnetic induction intensity loss function for the i-th operating condition, respectively. This is the magnetic vector potential result of the j-th coordinate point under the i-th working condition, predicted by the finite element-enhanced physical information neural network. This refers to the magnetic vector result of the j-th coordinate point under the i-th working condition, calculated by the physical residual module. This is the magnetic field strength at the j-th coordinate point under the i-th working condition, predicted by a finite element-enhanced physical information neural network. The magnetic field strength at the j-th coordinate point under the i-th operating condition, calculated by the physical computing layer. For parameters related to the neural network, c j = (x, y) are the coordinates of the points.

[0022] The present invention also provides a power transformer core electromagnetic field prediction system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a power transformer core electromagnetic field prediction method when executed by the processor.

[0023] The present invention also provides a storage medium storing a computer program, which executes a method for predicting the electromagnetic field of a power transformer core when running.

[0024] This invention first obtains dynamic characteristics such as peak voltage and DC bias current, as well as static characteristics such as spatial coordinates, core geometry, and material properties, corresponding to the excitation conditions based on the transformer core type and operating parameters. Then, according to the operating condition type, the core control equations of the finite element method are encoded into differentiable physical residual modules and embedded into the network to construct a finite element-enhanced physical information neural network model. Physical constraints are then constructed based on the finite element solver. These physical constraints consist of two parts: a magnetic vector potential loss term constraining the spatial distribution consistency between the network output and the finite element solution, and a magnetic induction intensity loss term constraining the physical calculation layer output and the physical quantity agreement between the physical calculation layer output and the finite element solution. Finally, a prediction model is obtained based on the input features and the neural network model training. This model is used to predict the electromagnetic field distribution of the transformer core under different operating conditions, and the automatic differentiation technology of the physical calculation layer module is used to calculate the electromagnetic field distribution. This technical solution, by integrating the physical supervision mechanism of the finite element method with the fast reasoning capability of the neural network, can achieve rapid prediction of the electromagnetic field distribution of the transformer core under different complex excitation conditions, exhibiting good generalization ability while maintaining both computational efficiency and accuracy. Attached Figure Description

[0025] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the 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.

[0026] Figure 1 This is a flowchart of the electromagnetic field prediction method for power transformer core according to an embodiment of the present invention;

[0027] Figure 2 A schematic diagram of a laminated core model of a power transformer provided in an embodiment of this application;

[0028] Figure 3 This is a schematic diagram of the structure of the calculation part of the power transformer laminated core model provided in the embodiments of this application;

[0029] Figure 4 A schematic diagram of the finite element-enhanced physical information neural network prediction architecture in the electromagnetic field prediction method for power transformer cores provided in this application embodiment;

[0030] Figure 5 A comparison diagram of the predicted electromagnetic field distribution of the laminated core model of a power transformer under DC bias excitation and the theoretical calculation results provided in the embodiments of this application;

[0031] Figure 6 This is a comparison diagram of the predicted electromagnetic field distribution at a certain point within a period of a power transformer laminated core model provided in this application under DC bias excitation, and the theoretical calculation results. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] Example 1

[0035] like Figure 1 As shown, this invention provides a method for predicting the electromagnetic field of a power transformer core, comprising:

[0036] S101: Based on the transformer core type and operating parameters, obtain the voltage peak value, dynamic characteristics and spatial coordinates of the DC bias current, and static characteristics of the core geometry and material properties corresponding to the excitation conditions.

[0037] In this step, for example, the following is measured experimentally: Figure 2 The physical model of the laminated core of the power transformer shown is used to obtain the material property m, i.e., the BH curve of the core material. Considering the symmetry of the model in terms of geometry and excitation, to reduce computational costs, a quarter of the model is used for research. The computational domain is set to 350 mm × 350 mm, the core corners are the active regions with coil winding, and the air parts are the passive regions without coil winding. Based on this, the following is selected... Figure 3 The quarter-model shown is partitioned to extract the spatial coordinates c=(x,y) and core geometric parameters g from the computational domain. g includes boundary information, mesh information, and material type. The parameters obtained above are all static input features, inherent parameters that define the model itself and do not change with excitation conditions, denoted as P. s =[c, g, m]. The dynamic input characteristics are determined by the model's operating conditions, including the peak voltage U of the excitation winding. max and DC current I dc Let P be the vector. d =[U max I dc ] T This is the main dimension reflecting the model's generalization ability. The DC bias excitation condition is selected based on the actual operating conditions, and the peak voltage U... max The range is set to [200V, 400V], with a step size of 20V, and a DC current I. dc The range is set to [0.2A, 0.4A], with a step size of 0.02A, for a total of 121 different operating conditions.

[0038] For the same operating condition i, the static input features and dynamic input features are concatenated into an input tuple P. (i) =[P s (i) P d (i) All input features are represented as P=[P s P d This input feature will be used as the input layer of the network for model training. The operating condition selected in this step is DC bias excitation, which belongs to nonlinear steady-state time harmonic field. The corresponding physical equations will be called according to this operating condition type in subsequent steps.

[0039] S102: According to the working condition type, the core control equations of the finite element are encoded into differentiable physical residual modules and embedded in the network to construct a finite element enhanced physical information neural network model, and physical constraints are constructed based on the finite element solver.

[0040] The core equations are derived and discretized for DC bias conditions. In the two-dimensional plane, the governing equations for the time-harmonic field are:

[0041]

[0042] Where c = (x, y) are spatial coordinates; A z (c,t;P d Let J be the magnetic vector in the z-direction at a certain moment. s (c,t;P d Let P be the source current density of the excitation winding at a certain moment, derived from the dynamic input characteristic P. d Obtained from, ν(P) s )=1 / μ is the nonlinear magnetoresistive density, given by P s The material property m in σ(P) is determined by σ(P) s ) represents conductivity, and t represents time.

[0043] Under steady-state DC bias excitation, both the magnetic vector potential and the source current density exhibit periodic changes. A truncated Fourier series expansion can be directly performed on these two physical quantities:

[0044]

[0045]

[0046] in, Let be the magnetic vector amplitude of the nth harmonic in the z-direction. The current density amplitude of the nth harmonic source; n is the harmonic order. t The harmonic order is truncated in the calculation; ω0=2πf is the fundamental angular frequency.

[0047] Taking the time partial derivatives term by term of the expansion of the magnetic vector potential, we get:

[0048]

[0049] Substituting the harmonic expansions of the magnetic vector and source current density into the time-harmonic field control equation, the magnetic vector potential operator term is:

[0050]

[0051] The eddy term is:

[0052]

[0053] The right-hand source term is:

[0054]

[0055] Substituting into the equation, we get:

[0056]

[0057] Integrating both sides of the equation over one period, and utilizing orthogonality, we obtain the frequency domain partial differential equation for the nth harmonic as follows:

[0058]

[0059] Where, ν n This represents the equivalent magnetoresistive distribution corresponding to the nth frequency in the current iteration.

[0060] Take the finite element shape function N i (c) As the weight function and interpolation basis, the computational domain is denoted as Ω, which includes the iron core magnetic conductor region, the excitation winding region, and the surrounding air region. The computational domain is triangulated, and each triangular element obtained by the triangulation is a first-order triangle with 3 vertices:

[0061]

[0062] The total number of global nodes is N node Each node corresponds to coordinate c. j = (x, y), for any element e, the area coordinates are represented as λ1, λ2, λ3, which satisfy the following relationship:

[0063]

[0064] The formula for the area of ​​a triangular unit is:

[0065]

[0066] Among them, △ e The coefficient for the unit area is calculated from the coordinates of the three vertices, i.e.:

[0067]

[0068] By spatially discretizing the frequency domain partial differential equation using the first-order triangular element and the Galerkin finite element method, we can obtain:

[0069]

[0070] harmonic amplitude Substituting into the above equation and rearranging, we get:

[0071]

[0072] Where, N node To represent the total number of global nodes, the parts of the above formula can be written in matrix-vector form:

[0073]

[0074] Assembling all the nodal degrees of freedom yields the field-path coupling nonlinear algebraic matrix equation for the k-th harmonic:

[0075]

[0076] Where M is the highest harmonic order, S k For element S n,ij The assembled stiffness matrix is ​​related to the harmonic order k and the material property m; T is the stiffness matrix obtained from the elements T ij The mass matrix obtained from assembly, which is related to the eddy current effect, is related to the material property m, F k For element F n , i The excitation vector of the kth harmonic obtained by assembly is related to P. d Related to the winding definition, A k Let be the k-th harmonic magnetic vector potential node coefficient vector to be determined. By solving this system of equations using the Newton-Raphson iteration, a high-precision solution set A of the harmonic coefficients can be obtained. k ,Right now:

[0077] The matrix equation for the k-th harmonic can be written in residual form:

[0078]

[0079] Define the global residual vector as:

[0080]

[0081] The l-th step of Newton's iteration yields:

[0082]

[0083] The Jacobian matrix is ​​as follows:

[0084]

[0085] The diagonal pieces are:

[0086]

[0087] The third term on the right is the material nonlinear coupling term, and q represents the other harmonic orders involved in the coupling.

[0088] Iterative updates are represented as:

[0089]

[0090] Where α is the damping factor. When ΔA (l) After the convergence condition is met, the amplitude vectors of each harmonic node can be obtained, which is the high-precision harmonic coefficient solution set:

[0091]

[0092] Corresponding spatial field:

[0093]

[0094] The inverse transformation finite element reference time-domain solution, i.e., magnetic vector potential A, is obtained by using harmonic superposition. cal :

[0095]

[0096] The magnetic flux density B is then obtained by solving for the curl of the magnetic vector potential. cal :

[0097]

[0098] The process of solving for the magnetic vector potential and magnetic induction intensity using the harmonic balance finite element method described above is encoded as a differentiable physical residual module embedded in the network. The magnetic vector potential A is then calculated by this module. cal With magnetic induction intensity B cal This leads to the construction of physical constraints for a high-fidelity finite element solver. The physical constraints consist of two parts: a magnetic vector potential loss term constraining the spatial distribution consistency between the constraint network output and the finite element solution, and a magnetic induction intensity loss term constraining the physical computation layer output and the physical quantity agreement between the constraint and the finite element solution. These are expressed as follows:

[0099]

[0100]

[0101] in, Let be the magnetic vector potential loss function and the magnetic induction intensity loss function for the i-th operating condition, respectively. This is the magnetic vector potential result of the j-th coordinate point under the i-th working condition, predicted by the finite element-enhanced physical information neural network. This refers to the magnetic vector result of the j-th coordinate point under the i-th working condition, calculated by the physical residual module. This is the magnetic field strength at the j-th coordinate point under the i-th working condition, predicted by a finite element-enhanced physical information neural network. The magnetic field strength at the j-th coordinate point under the i-th operating condition, calculated by the physical computing layer. For parameters related to the neural network, c j = (x, y) represents the coordinates of the point. Total loss function Represented as:

[0102]

[0103] in, and To balance hyperparameters, optimization algorithms such as Adam are used to minimize them. To train network parameters This enables the network not only to predict the magnetic field distribution under different excitation conditions, but also to internalize the laws of physical conservation during the training process.

[0104] like Figure 4 As shown, the complete physical information neural network model architecture of the fused finite element method obtained according to step S102 is presented. This network uses a multilayer perceptron as the core mapping learner. The input layer consists of static input features such as material properties and dynamic input features composed of different working conditions. After mapping through 4 hidden layers, each layer contains 512 neurons. The SiLU activation function is used between layers to ensure gradient smoothness. Finally, the magnetic vector potential prediction value A is output. z Output A in the network. z The magnetic flux density B is then obtained by connecting a parameterless physical calculation layer.

[0105] S103: Based on the input features and the neural network model training, a prediction model is obtained. The model of the core of the power transformer to be predicted is predicted under different operating conditions, and the electromagnetic field distribution of the model is obtained by using the physical computing layer.

[0106] In this step, for example, a prediction model is trained based on the neural network architecture in step S102, and a DC bias condition U that is not within the training data range is selected. max =480 V, I dc =0.35A is used for prediction, and the curl of the magnetic vector potential output by the network is obtained by automatic differentiation through a parameterless physical computing layer, thus obtaining the magnetic flux density, i.e.:

[0107]

[0108] For each working condition, the root mean square error (RMSE) and mean absolute error (MAE) are used as quantitative indicators for comparative analysis.

[0109]

[0110]

[0111] Among them, B pred,i For the predicted magnetic flux density, B cal,i This is the calculated result of the magnetic flux density.

[0112] like Figure 5 , Figure 6 As shown, a comparison chart of the prediction results obtained by the method of this embodiment and the theoretical calculation results under the same training conditions is presented, which verifies the accuracy and good generalization ability of the physical information neural network after integrating the finite element method in this embodiment to predict the electromagnetic field distribution of the power transformer core.

[0113] In summary, the electromagnetic field prediction method for power transformer cores provided in this embodiment can quickly predict the distribution of electromagnetic fields in transformer cores under different complex excitation conditions. It exhibits good generalization ability while maintaining both computational efficiency and accuracy.

[0114] Example 2

[0115] This invention also provides a system for predicting the electromagnetic field of a power transformer core, comprising:

[0116] The first processing module is used to obtain the voltage peak value, dynamic characteristics and spatial coordinates of DC bias current, and static characteristics of core geometry and material properties corresponding to the excitation conditions, based on the transformer core type and operating parameters.

[0117] The second processing module is used to encode the core control equations of the finite element method into differentiable physical residual modules and embed them into the network according to the working condition type, thereby constructing a finite element-enhanced physical information neural network model; wherein, the physical constraints of the physical information neural network model are constructed based on the finite element solver.

[0118] The third processing module is used to input dynamic and static features into the physical information neural network model, train it to obtain a prediction model, predict different operating conditions of the core model of the power transformer to be predicted based on the prediction model, and use the automatic differentiation technology of the physical calculation layer module to calculate the electromagnetic field distribution.

[0119] As one embodiment of the present invention, the construction of physical constraints based on the finite element solver includes:

[0120]

[0121]

[0122] in, Let be the magnetic vector potential loss function and the magnetic induction intensity loss function for the i-th operating condition, respectively. This is the magnetic vector potential result of the j-th coordinate point under the i-th working condition, predicted by the finite element-enhanced physical information neural network. This refers to the magnetic vector result of the j-th coordinate point under the i-th working condition, calculated by the physical residual module. This is the magnetic field strength at the j-th coordinate point under the i-th working condition, predicted by a finite element-enhanced physical information neural network. The magnetic field strength at the j-th coordinate point under the i-th operating condition, calculated by the physical computing layer. For parameters related to the neural network, c j = (x, y) are the coordinates of the points.

[0123] Example 3

[0124] The present invention also provides a power transformer core electromagnetic field prediction system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a power transformer core electromagnetic field prediction method when executed by the processor.

[0125] Example 4

[0126] The present invention also provides a storage medium storing a computer program, which executes a method for predicting the electromagnetic field of a power transformer core when running.

[0127] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting the electromagnetic field of a power transformer core, characterized in that, include: Based on the transformer core type and operating parameters, obtain the voltage peak value, dynamic characteristics and spatial coordinates of DC bias current, and static characteristics of core geometry and material properties corresponding to the excitation conditions. According to the working condition type, the core control equations of the finite element method are encoded into differentiable physical residual modules and embedded into the network to construct a finite element enhanced physical information neural network model; among them, the physical constraints of the physical information neural network model are constructed based on the finite element solver. Dynamic and static features are input into the physical information neural network model to train a prediction model. Based on the prediction model, different operating conditions are predicted for the core model of the power transformer to be predicted, and the electromagnetic field distribution is calculated using the automatic differentiation technology of the physical computing layer module.

2. The method for predicting the electromagnetic field of a power transformer core as described in claim 1, characterized in that, The construction of physical constraints based on the finite element solver includes: in, Let be the magnetic vector potential loss function and the magnetic induction intensity loss function for the i-th operating condition, respectively. This is the magnetic vector potential result of the j-th coordinate point under the i-th working condition, predicted by the finite element-enhanced physical information neural network. This refers to the magnetic vector result of the j-th coordinate point under the i-th working condition, calculated by the physical residual module. This is the magnetic field strength at the j-th coordinate point under the i-th working condition, predicted by a finite element-enhanced physical information neural network. The magnetic field strength at the j-th coordinate point under the i-th operating condition, calculated by the physical computing layer. For parameters related to the neural network, c j = (x, y) are the coordinates of the points.

3. A device for predicting the electromagnetic field of a power transformer core, characterized in that, include: The first processing module is used to obtain the voltage peak value, dynamic characteristics and spatial coordinates of DC bias current, and static characteristics of core geometry and material properties corresponding to the excitation conditions, based on the transformer core type and operating parameters. The second processing module is used to encode the core control equations of the finite element method into differentiable physical residual modules and embed them into the network according to the working condition type, thereby constructing a finite element-enhanced physical information neural network model; wherein, the physical constraints of the physical information neural network model are constructed based on the finite element solver. The third processing module is used to input dynamic and static features into the physical information neural network model, train it to obtain a prediction model, predict different operating conditions of the core model of the power transformer to be predicted based on the prediction model, and use the automatic differentiation technology of the physical calculation layer module to calculate the electromagnetic field distribution.

4. The electromagnetic field prediction device for power transformer cores as described in claim 3, characterized in that, The construction of physical constraints based on the finite element solver includes: in, Let be the magnetic vector potential loss function and the magnetic induction intensity loss function for the i-th operating condition, respectively. This is the magnetic vector potential result of the j-th coordinate point under the i-th working condition, predicted by the finite element-enhanced physical information neural network. This refers to the magnetic vector result of the j-th coordinate point under the i-th working condition, calculated by the physical residual module. This is the magnetic field strength at the j-th coordinate point under the i-th working condition, predicted by a finite element-enhanced physical information neural network. The magnetic field strength at the j-th coordinate point under the i-th operating condition, calculated by the physical computing layer. For parameters related to the neural network, c j = (x, y) are the coordinates of the points.

5. A power transformer core electromagnetic field prediction system, characterized in that, include: A memory and a processor, wherein the memory stores a computer program executed by the processor, the computer program executing the electromagnetic field prediction method for the core of a power transformer as described in any one of claims 1-2 when run by the processor.

6. A storage medium, characterized in that, The storage medium stores a computer program, which executes the electromagnetic field prediction method for the core of a power transformer as described in any one of claims 1-2 when the computer program is run.