Transformer three-dimensional electromagnetic simulation method based on NU-UNET and application

By combining the NU-UNET model with KD trees and U-Net neural networks, the problems of high computational cost and strong mesh dependence in three-dimensional electromagnetic field simulation are solved, realizing fast and stable electromagnetic field simulation and meeting the real-time simulation requirements of power equipment.

CN119066972APending Publication Date: 2024-12-03GUILIN UNIV OF ELECTRONIC TECH
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202411185163.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-27
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

Existing three-dimensional electromagnetic field simulation methods are computationally expensive, have long solution times, and are highly dependent on meshes, making it difficult to meet the requirements of real-time simulation in digital twins of power equipment.

Method used

The NU-UNET model combined with the KD tree algorithm is used for region decomposition. Different grid sizes are used for interpolation, and the U-Net neural network is used to extract the features of the input grid data, reducing the dependence on specific grid divisions and achieving fast and efficient prediction of electromagnetic field distribution.

Benefits of technology

It reduces computational costs and solution time, improves the stability and consistency of simulation results, and meets the real-time simulation requirements in digital twins of power equipment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066972A_ABST
    Figure CN119066972A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of three-phase transformer electromagnetic field simulation, in particular to a transformer three-dimensional electromagnetic simulation method based on NU-UNET and application, and the method comprises the following steps: S1, building a three-phase transformer model by using comsol; s2, acquiring and preprocessing three-dimensional electromagnetic field simulation data; s3, an NU-UNET model is built and trained; and S4, carrying out reverse processing and analysis on a model prediction result. According to the method, an NU-UNET model is used, a K-D tree algorithm is used for carrying out regional decomposition, then grids of different sizes are adaptively used for interpolation on different sub-domains, the sub-domain grids subjected to interpolation are used for training a grid-based model of a bottom layer, that is, input grid data features are extracted through a convolutional neural network, and the grid data features are extracted through the convolutional neural network. The method does not depend on specific grid division, and can quickly and efficiently carry out electromagnetic field distribution prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of electromagnetic field simulation of three-phase transformers, and in particular to a three-dimensional electromagnetic simulation method and application of transformers based on NU-UNET. Background Art

[0002] In the field of electromagnetic field simulation analysis, the existing technology mainly relies on traditional numerical simulation calculations for the simulation of the three-dimensional electromagnetic field of a three-phase transformer. These methods usually include numerical calculation techniques such as the finite element method (FEM), the finite difference method (FDM) and the finite volume method (FVM).

[0003] The prior art three-phase transformer three-dimensional electromagnetic field simulation method mainly includes the following four steps:

[0004] Establish a three-phase transformer three-dimensional geometric model and perform meshing: geometric modeling of the three-dimensional structure of the transformer is performed. Then, the geometric model is divided into several small units (such as tetrahedral units, hexahedral units, etc.) through meshing technology for numerical calculation.

[0005] Set physical parameters and boundary conditions: After establishing the geometric model, you need to set physical parameters (such as the material's electrical conductivity, magnetic permeability, etc.) and boundary conditions (such as voltage, current, magnetic field strength, etc.) for each small unit.

[0006] Solve the control equations: According to the electromagnetic field theory, the Maxwell equations are established, and the equations are discretized into a set of algebraic equations based on the material properties and boundary conditions. These algebraic equations are solved by numerical solution methods (such as matrix decomposition method, iteration method, etc.) to obtain the distribution of the electromagnetic field.

[0007] Post-processing and result analysis: Post-processing of the solved electromagnetic field distribution is mainly for the visualization and analysis of the magnetic field distribution.

[0008] Disadvantages of the existing technology:

[0009] High computational cost: Fine meshing and complex numerical solution processes lead to high computational costs, and these calculations need to be performed from scratch for each simulation, which is particularly significant when dealing with large-scale three-dimensional problems.

[0010] Long solution time: Due to the complexity and high computational complexity of electromagnetic field simulation, the solution time is usually long, which makes it difficult to meet the requirements of real-time simulation in the digital twin of power equipment.

[0011] Strong mesh dependence: The simulation results are highly dependent on the quality and distribution of the mesh. Different mesh divisions may lead to different simulation results, increasing the uncertainty of the results.

[0012] In view of the above problems, the present invention proposes a transformer three-dimensional electromagnetic simulation method based on NU-UNET, which is used to efficiently simulate the electromagnetic field distribution of a three-dimensional transformer. By combining deep learning technology with traditional numerical simulation methods, this patent aims to solve the problems of high computational cost, long solution time and strong grid dependence in the prior art. Summary of the invention

[0013] The purpose of the present invention is to provide a NU-UNET-based transformer three-dimensional electromagnetic simulation method and application, using the NU-UNET model, using the KD tree algorithm to perform regional decomposition, and then adaptively using grids of different sizes for interpolation on different subdomains, and then using the interpolated subdomain grids to train the underlying grid-based model, that is, extracting the input grid data features through a convolutional neural network, not relying on specific grid division, and being able to quickly and efficiently predict the electromagnetic field distribution.

[0014] The present invention innovatively combines KD tree space partitioning and U-Net neural network algorithm to process transformer three-dimensional electromagnetic field simulation. This method can effectively realize accurate simulation of the transformer electromagnetic field and provide important support for rapid simulation of power equipment systems.

[0015] In order to achieve the above technical objectives and the above technical effects, the present invention is implemented through the following technical solutions:

[0016] A three-dimensional electromagnetic simulation method of a transformer based on NU-UNET, comprising the following steps:

[0017] S1. Use comsol to build a three-phase transformer model;

[0018] S2, 3D electromagnetic field simulation data acquisition and preprocessing;

[0019] S3, NU-UNET model construction and training;

[0020] S4. Back-processing and analysis of model prediction results.

[0021] Furthermore, the step S1 specifically includes: using the transformer case in COMSOL Multiphysics software to establish an electromagnetic field simulation model; using symmetrical simplification calculation, the transformer model is reduced to one quarter. Since the left and right physical fields of the model are different, especially for the coil, the phase difference is 120 degrees, so the symmetrical simplification is adopted in the up-down and front-back symmetric simplification. The model adopts a transient analysis method, taking into account the nonlinear BH curve characteristics of the multi-turn coil domain and the magnetic core in the transformer. The transformer is divided into 1528 grid units, and the finite element method is used to solve the equations of each unit, and finally the equations of the entire system are obtained.

[0022] Furthermore, the step S2 specifically includes the following sub-steps:

[0023] S2.1, Data loading and preprocessing: Load transformer point cloud data and physical quantities (simulation time and secondary resistance). Input the point cloud data into the KD-Tree algorithm and recursively divide it into two parts according to the spatial distribution of the data until the preset number of subdomains is reached;

[0024] S2.2, Subdomain data processing: For each subdomain, obtain its bounding box and the index of the points it contains. Calculate the length, width and height of its three-dimensional space. According to the number of points in the subdomain and the spatial scale, calculate its appropriate uniform grid size and shape so that it can effectively cover the data in the entire subdomain. And record the maximum grid shape to ensure that the shape is the maximum value of each dimension (x, y, z) in all subdomains;

[0025] S2.3, interpolation to uniform grid: use linear interpolation and nearest neighbor interpolator to interpolate the simulation time and secondary resistance in the point cloud data to the uniform grid. The interpolation results within each subdomain are mapped to a uniform grid that matches the feature size of its data. This process ensures the regularity and consistency of the data within each subdomain;

[0026] S2.4. Prepare to input the U-Net operator model: adjust the shape of the interpolation result of each subdomain, and map all subdomains to a three-dimensional grid with a uniform size of the maximum grid shape, so that they can be input into the same model for processing. For the part with less data after plasticity, the value of the 0th data point of the original data can be used for filling. The input of U-Net is regular grid data after interpolation, and each subdomain contains two input channels (simulation time, secondary resistance). Its input size is: number of channels × number of subdomains × maximum value of each dimension;

[0027] S2.5, training and evaluating the U-Net model: The model takes the interpolated regular grid data (simulation time, secondary resistance) as input to predict the target variable electromagnetic field value;

[0028] S2.6, interpolation and deinterpolation: Use the interpolation method to reversely interpolate the grid data output by the neural operator model to the grid of the original point cloud data to obtain the final predicted field distribution.

[0029] Furthermore, step S3 includes: constructing a U-Net model, including 4 3D convolutional layers and 4 3D deconvolutional layers, retaining feature information through jump connections; using a transient analysis method, considering the nonlinear BH curve characteristics of the multi-turn coil domain and the magnetic core in the transformer. The transformer is divided into 1528 grid units, and the finite element method is used to solve the equations of each unit, and finally the equations of the entire system are obtained.

[0030] The model was trained using the Adam optimizer with a learning rate of 0.001 and a weight decay of 10. -4 , and combined with ReduceLROnPlateau learning rate scheduler for training. The loss function of the model consists of two parts: the first part is the deviation between the network output and the true subdomain grid, and the second part is the regularization term, that is, the interpolation error between the network output and the true point cloud. The weight of the regularization term is 5×10 -3 The training dataset contains 6,000 sets of data and the test dataset contains 600 sets of data, which are used to evaluate the accuracy of the model.

[0031] Further, the step S4 specifically includes: after model training, the transformer secondary load resistance and simulation time, and the three-dimensional point cloud in the test data set are input into the trained model to obtain the simulation result of the electromagnetic field. Then the optimal model is screened out after the accuracy evaluation using the K-fold cross validation method. Finally, the simulation results are reversed, and the electromagnetic field data output by the model are interpolated to restore to the 1528 grid cells generated by the original COMSOL simulation. After the interpolation is completed, a mean smoothing operation is performed to determine the adjacent cells around the grid cell, calculate the data mean of the adjacent cells, and use the mean as the value of the current grid cell to obtain the final simulation result of the transformer magnetic field.

[0032] On the other hand, the present invention proposes the application of the above method in the simulation of the three-dimensional electromagnetic field of a three-phase transformer.

[0033] Beneficial effects of the present invention:

[0034] Traditional numerical simulation methods rely on detailed grid division and complex numerical solution processes, and the computational cost is high. By using the NU-UNET model, the present invention can quickly and efficiently complete electromagnetic field simulation with less computing resource investment. In the preprocessing, the present invention uses the KD tree algorithm to sub-domain the data, and interpolates the point cloud data onto a regular grid through linear interpolation and nearest neighbor interpolation methods to generate a data set for model training and testing. This process reduces the reliance on high-precision grid division, thereby reducing computational costs. The NU-UNET model extracts input grid data features through a convolutional neural network and does not rely on specific grid division. In this way, the simulation process does not need to start from scratch every time to perform high-computation grid division and numerical solution, further reducing computational costs.

[0035] Due to its complexity and high computational complexity, traditional numerical simulation methods take a long time to solve, making it difficult to meet the requirements of real-time simulation in digital twins of power equipment. The method of the present invention can greatly shorten the solution time. After training, the NU-UNET model of the present invention can quickly predict the distribution of electromagnetic fields. Since the reasoning process of the deep learning model is relatively fixed, the solution time depends only on the processing of the input data and the reasoning speed of the model, which is usually faster than traditional numerical simulation methods. In the model training process, the Adam optimizer and the ReduceLROnPlateau scheduler are used, combined with the LpLoss and MSELoss loss functions, to optimize the model, which effectively accelerates the convergence of the model and improves the training efficiency.

[0036] Traditional numerical simulation methods are highly dependent on the quality and distribution of the grid. Different grid divisions may lead to different simulation results, increasing the uncertainty of the results. The simulation method based on NU-UNET can reduce this dependence. The point cloud data is regionally decomposed through the KD tree algorithm to ensure that the data in each subdomain is evenly distributed and normalized to adapt to the network input. This method can reduce the dependence on specific grid divisions and improve the stability and consistency of simulation results. The NU-UNET model uses a multi-layer 3D convolution and deconvolution network structure, combined with jump connections to retain feature information, and can extract stable electromagnetic field features under different grid divisions, reducing the dependence of simulation results on grid quality.

[0037] In terms of verifying the accuracy and efficiency of the model, the method proposed in this invention has been compared with the traditional numerical simulation results, proving its superiority in accuracy and efficiency; the electromagnetic field distribution output by the model is restored to the actual coordinate system through the inverse normalization process to obtain the final simulation results. By comparing the results with those of the traditional numerical simulation method, the accuracy of the model in different simulation scenarios is verified; through multiple experimental comparisons, it is proved that the method based on NU-UNET can greatly improve the simulation efficiency at the same accuracy, meeting the real-time simulation requirements in the digital twin of power equipment.

[0038] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0040] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0041] Figure 2 Schematic diagram of the KD tree construction process;

[0042] Figure 3 This is a schematic diagram of test loss comparison;

[0043] Figure 4 This is a schematic diagram for comparing training losses;

[0044] Figure 5 This is a schematic diagram for comparing simulation results. DETAILED DESCRIPTION

[0045] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] Example 1

[0047] like Figure 1 , 2 As shown, the transformer three-dimensional electromagnetic simulation method based on NU-UNET described in this embodiment includes the following steps:

[0048] S1. Use comsol to build a three-phase transformer model;

[0049] S2, 3D electromagnetic field simulation data acquisition and preprocessing;

[0050] S3, NU-UNET model construction and training;

[0051] S4. Back-processing and analysis of model prediction results.

[0052] In this embodiment, the step S1 specifically includes: using the transformer case in COMSOL Multiphysics software to establish an electromagnetic field simulation model; using symmetrical simplification calculation, the transformer model is reduced to one quarter. Since the left and right physical fields of the model are different, especially for the coil, the phase difference is 120 degrees, so the symmetrical simplification is adopted in the up-down and front-back symmetric simplification. The model adopts a transient analysis method, taking into account the nonlinear BH curve characteristics of the multi-turn coil domain and the magnetic core in the transformer. The transformer is divided into 1528 grid units, and the finite element method is used to solve the equations of each unit, and finally the equations of the entire system are obtained.

[0053] In this embodiment, the convolutional neural operator can better learn highly non-uniform transformer electromagnetic field data. The present invention uses the non-uniform framework (NU) proposed by Liu. The framework processing process is as follows Figure 2 As shown, it specifically includes the following sub-steps:

[0054] S2.1, Data loading and preprocessing: Load transformer point cloud data and physical quantities (simulation time and secondary resistance). Input the point cloud data into the KD-Tree algorithm and recursively divide it into two parts according to the spatial distribution of the data until the preset number of subdomains is reached;

[0055] S2.2, Subdomain data processing: For each subdomain, obtain its bounding box and the index of the points it contains. Calculate the length, width and height of its three-dimensional space. According to the number of points in the subdomain and the spatial scale, calculate its appropriate uniform grid size and shape so that it can effectively cover the data in the entire subdomain. And record the maximum grid shape to ensure that the shape is the maximum value of each dimension (x, y, z) in all subdomains;

[0056] S2.3, interpolation to uniform grid: use linear interpolation and nearest neighbor interpolator to interpolate the simulation time and secondary resistance in the point cloud data to the uniform grid. The interpolation results within each subdomain are mapped to a uniform grid that matches the feature size of its data. This process ensures the regularity and consistency of the data within each subdomain;

[0057] S2.4. Prepare to input the U-Net operator model: adjust the shape of the interpolation result of each subdomain, and map all subdomains to a three-dimensional grid with a uniform size of the maximum grid shape, so that they can be input into the same model for processing. For the part with less data after plasticity, the value of the 0th data point of the original data can be used for filling. The input of U-Net is regular grid data after interpolation, and each subdomain contains two input channels (simulation time, secondary resistance). Its input size is: number of channels × number of subdomains × maximum value of each dimension;

[0058] S2.5, training and evaluating the U-Net model: The model takes the interpolated regular grid data (simulation time, secondary resistance) as input to predict the target variable electromagnetic field value;

[0059] S2.6, interpolation and deinterpolation: Use the interpolation method to reversely interpolate the grid data output by the neural operator model to the grid of the original point cloud data to obtain the final predicted field distribution.

[0060] In this embodiment, step S3 includes: constructing a U-Net model, including 4 3D convolutional layers and 4 3D deconvolutional layers, and retaining feature information through skip connections;

[0061] Table 1 UNet structure and channel number change table

[0062] Layer InputChannels OutputChannels Additional Information Input 2 - Linear n_channels+3 32 Down1 32 64 MaxPool3D+DoubleConv Down2 64 128 MaxPool3D+DoubleConv Down3 128 256 MaxPool3D+DoubleConv Down4 256 512 MaxPool3D+DoubleConv Up1 512 256 Upsample+DoubleConv Up2 256 128 Upsample+DoubleConv Up3 128 64 Upsample+DoubleConv Up4 64 32 Upsample+DoubleConv Output 32 1 Linear

[0063] In the table, n_channels is the product of the input feature dimension and the number of subdomains, and n_classes is the product of the output feature dimension and the number of subdomains. The transient analysis method is used to consider the nonlinear BH curve characteristics of the multi-turn coil domain and the core in the transformer. The transformer is divided into 1528 grid units, and the finite element method is used to solve the equations of each unit, and finally the equations of the entire system are obtained.

[0064] Data preprocessing

[0065] In the FEM simulation, set the constant frequency F = 50 Hz, AC voltage Vac = 110 kV, primary resistance Rp = 100 ohms, primary coil turns Np = 1100, and secondary coil turns Ns = 105. Set the value of Rs to 1 to 10 ohms, with a step size of 1 ohm, and the value of T to 0.001 to 0.06 seconds, with a step size of 0.001 seconds. A total of 6000 sets of data are obtained as training sets. Set Rs = [5.5, 15.5, 25.5, 35.5, 45.5, 55.5, 65.5, 75.5, 85.5, 95.5], the value of T to 0.001 to 0.06 seconds, with a step size of 0.001 seconds, and a total of 600 sets of data are obtained as test sets, each set of data consists of 1528 numbers.

[0066] The model was trained using the Adam optimizer with a learning rate of 0.001 and a weight decay of 10. -4 , and combined with ReduceLROnPlateau learning rate scheduler for training. The loss function of the model consists of two parts: the first part is the deviation between the network output and the true subdomain grid, and the second part is the regularization term, that is, the interpolation error between the network output and the true point cloud. The weight of the regularization term is 5×10 -3 The training dataset contains 6,000 sets of data and the test dataset contains 600 sets of data, which are used to evaluate the accuracy of the model.

[0067] In this embodiment, the step S4 specifically includes: after model training, the transformer secondary load resistance and simulation time, and the three-dimensional point cloud in the test data set are input into the trained model to obtain the simulation result of the electromagnetic field. Then the optimal model is screened out after the accuracy evaluation using the K-fold cross validation method. Finally, the simulation results are reversed, and the electromagnetic field data output by the model are interpolated to restore to the 1528 grid cells generated by the original COMSOL simulation. After the interpolation is completed, a mean smoothing operation is performed to determine the adjacent cells around the grid cell, calculate the data mean of the adjacent cells, and use the mean as the value of the current grid cell to obtain the final simulation result of the transformer magnetic field.

[0068] On the other hand, the present invention proposes the application of the above method in the simulation of the three-dimensional electromagnetic field of a three-phase transformer.

[0069] Example 2

[0070] Load data: Use numpy to load point cloud data files (data_xyz.npy and merged_data.npy) and store the data in memory;

[0071] KD tree construction: Use the KD tree algorithm to divide the point cloud data into subdomains, decompose the point cloud data into regions, generate multiple subdomains, and ensure that the data in each subdomain is evenly distributed; the data in each subdomain is normalized by UnitGaussianNormalizer to adapt to the network input;

[0072] Interpolation method: Linear interpolation and nearest neighbor interpolation methods are used to interpolate the point cloud data in the subdomain onto a regular grid to generate a data set for model training and testing.

[0073] Optimization algorithm: Use the Adam optimizer and ReduceLROnPlateau scheduler, combined with the LpLoss and MSELoss loss functions, to train the model and optimize the parameters.

[0074] The simulation time T and the secondary load resistance Rs are taken as input and input into the trained UNet3d model after preprocessing, and then the model outputs the electromagnetic field distribution for output prediction;

[0075] Furthermore, the step S4 specifically includes:

[0076] The electromagnetic field distribution output by the model is restored to the actual coordinate system, and the final simulation results are obtained through the inverse normalization process; the results are compared with those of traditional numerical simulation methods to verify the accuracy and efficiency of the model.

[0077] Example 3

[0078] like Figure 3-5 Shown

[0079] After model training, the transformer secondary load resistance, simulation time, and three-dimensional point cloud in the test data set are input into the trained model to obtain the simulation results of the electromagnetic field. The optimal model is screened out after accuracy evaluation using the K-fold cross validation method. Finally, the simulation results are reversed and the electromagnetic field data output by the model is interpolated to restore it to the 1528 grid cells generated by the original COMSOL simulation. After the interpolation is completed, a mean smoothing operation is performed to determine the adjacent cells around the grid cell, calculate the data mean of the adjacent cells, and use the mean as the value of the current grid cell to obtain the final simulation results of the transformer magnetic field.

[0080] The simulation results show that the magnetic induction intensity simulated by the FEM model, U-Net model and NU-Une model is compared when the secondary resistance is set to 91.5Ω and the simulation time is 0.04s. From the comparison of the simulation result graph and the test set loss graph, it can be seen that the model combined with the non-uniform framework (NU) is more advantageous than the baseline model in processing the three-dimensional electromagnetic field of the transformer. In addition, the proposed three-dimensional electromagnetic simulation method of the transformer based on NU-UNET shortens the simulation time from 3.094s required by FEM to 0.103s in the electromagnetic field simulation of the three-phase power transformer, which meets the requirements of real-time simulation of digital twins.

[0081] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A three-dimensional electromagnetic simulation method for transformer based on NU-UNET, characterized in that: The following steps are involved: S1. Use comsol to build a three-phase transformer model; S2, 3D electromagnetic field simulation data acquisition and preprocessing; S3, NU-UNET model construction and training; S4. Back-processing and analysis of model prediction results.

2. The transformer three-dimensional electromagnetic simulation method based on NU-UNET according to claim 1, characterized in that: The step S1 specifically includes: An electromagnetic field simulation model was established using the transformer case in COMSOL Multiphysics software. Symmetrical simplification calculation was used to reduce the transformer model to one-fourth, and symmetrical simplification was adopted in the top-bottom and front-back simplification. The transient analysis method was used to divide the transformer into 1528 grid units, and the finite element method was used to solve the equations for each unit, finally obtaining the equations for the entire system.

3. The transformer three-dimensional electromagnetic simulation method based on NU-UNET according to claim 1, characterized in that: The step S2 specifically includes the following sub-steps: S2.1, data loading and preprocessing: load transformer point cloud data and physical quantities; input the point cloud data into the KD-Tree algorithm, and recursively divide it into two parts according to the spatial distribution of the data until the preset number of subdomains is reached; S2.2, subdomain data processing: For each subdomain, obtain its bounding box and the index of the points it contains; Calculate the length, width and height of its three-dimensional space, and calculate its appropriate uniform grid size and shape according to the number of points and spatial scale in the subdomain, so as to effectively cover the data in the entire subdomain; and record the maximum grid shape to ensure that the shape is the maximum value of each dimension (x, y, z) in all subdomains; S2.3, interpolation to uniform grid: use linear interpolation and nearest neighbor interpolator to interpolate the simulation time and secondary resistance in the point cloud data to the uniform grid; the interpolation results within each subdomain are mapped to a uniform grid that matches the feature size of its data; this process ensures the regularity and consistency of the data within each subdomain; S2.4, prepare to input U-Net operator model: adjust the shape of the interpolation result of each subdomain, and map all subdomains to a three-dimensional grid with a unified size of the maximum grid shape, so that they can be input into the same model for processing; for the part with less data after plasticity, the value of the 0th data point of the original data can be used to fill it; the input of U-Net is the regular grid data after interpolation, and each subdomain contains two input channels: simulation time and secondary resistance; Its input size is: number of channels × number of subdomains × maximum value of each dimension; S2.5, training and evaluating the U-Net model: the model takes the interpolated regular grid data as input and predicts the electromagnetic field value of the target variable; S2.6, interpolation and deinterpolation: Use the interpolation method to reversely interpolate the grid data output by the neural operator model to the grid of the original point cloud data to obtain the final predicted field distribution.

4. The transformer three-dimensional electromagnetic simulation method based on NU-UNET according to claim 1, characterized in that: The step S3 specifically includes: The step S3 comprises: constructing a U-Net model, including 4 3D convolutional layers and 4 3D deconvolutional layers, and retaining feature information through skip connections; The model was trained using the Adam optimizer with a learning rate of 0.001 and a weight decay of 10. -4 , and combined with ReduceLROnPlateau learning rate scheduler for training; the loss function of the model consists of two parts: the first part is the deviation between the network output and the real subdomain grid, and the second part is the regularization term, that is, the interpolation error between the network output and the real point cloud; the weight of the regularization term is 5×10 -3 ; The training data set contains 6000 sets of data and the test data set contains 600 sets of data, which are used to evaluate the accuracy of the model.

5. The transformer three-dimensional electromagnetic simulation method based on NU-UNET according to claim 1, characterized in that: The step S4 specifically includes: After model training, the transformer secondary load resistance, simulation time and three-dimensional point cloud in the test data set are input into the trained model to obtain the simulation results of the electromagnetic field. The optimal model is screened out after accuracy evaluation using the K-fold cross-validation method. Finally, the simulation results are inversely processed and the electromagnetic field data output by the model are interpolated to restore them to the 1528 grid cells generated by the original COMSOL simulation. After the interpolation is completed, a mean smoothing operation is performed to determine the adjacent cells around the grid cell, calculate the data mean of the adjacent cells, and use the mean as the value of the current grid cell to obtain the final simulation results of the transformer magnetic field.

6. Application of the method according to any one of claims 1 to 5 in the simulation of the three-dimensional electromagnetic field of a three-phase transformer.

Citation Information

Cited By

  • Forging and pressing simulation method and related device

    CN119783488A

  • Three-dimensional space field prediction method and device based on large model and data compression, equipment and medium

    CN120317119A

  • Electromagnetic field near-field visualization method and device fusing machine vision and electromagnetic measurement

    CN120993054A