Electric field analysis method

Abstract

The invention provides an electric field analysis method. Firstly, a target model is generated through simulation modeling, continuous electric field distribution data is obtained, a first data set and a second data set are established, rasterized image feature extraction is performed on the first data set and the second data set through a graph neural network, corresponding first graph features and second graph features are obtained, and a cascade grid processing network is used to process the electric field distribution data. Firstly, a target power line is predicted according to the first graph feature, then the target power line is converted into a target electric field mode according to the second graph feature, and finally an electric field prediction result is generated and a preset analysis process is executed. According to the method, through the combination of the graph neural network and the cascade grid processing network, the continuity characteristic of electric field distribution is effectively reserved, and meanwhile, the noise problem caused by discretization is avoided. And through rasterized feature data processing, the data demand and the calculation complexity are greatly reduced, and the efficiency and the accuracy of modeling analysis are improved.

Description

Technical Field

[0001] The present application relates to the field of electrical detection, and in particular to an electric field analysis method. Background Art

[0002] During the manufacturing, installation, construction, and operation of high-voltage equipment, defects, surface burrs, and other anomalies can cause localized electric field distortion, leading to partial discharge. Therefore, electric field analysis of high-voltage equipment is a key tool for troubleshooting, defect elimination, and design optimization. Due to the complex structure and numerous components of high-voltage equipment, electromagnetic field theory equations are difficult to solve. Finite element methods are typically used for modeling and analysis, but this requires excessively large amounts of data and is inefficient. Therefore, deep learning methods can be combined to address these issues.

[0003] However, methods combined with deep learning still require a series of samples obtained through finite element simulation, allowing the deep learning algorithm to train the inherent physical connections, ultimately bypassing complex finite element calculations and obtaining the corresponding physical quantities. The accuracy of the calculation is affected by the quality of the samples. However, since deep learning algorithms can only process discrete data in this scenario, the data needs to be discretized when generating samples. Discretization may cause two problems. On the one hand, there is no longer a physical connection between the grids, but rather separate discrete points. The electric field distribution is mainly affected by the shape of the electrodes and boundaries. After the structure is discretized, it will cause a lot of noise in the prediction results. On the other hand, network training generally requires the input data matrix to be regular. Whether it is a grid or sampling, the area needs to remain unchanged, and the grid size needs to remain unchanged. For the sampling points, they also need to be evenly spaced and the point positions need to remain unchanged. Ultimately, only data of specific parameters can be obtained.

[0004] Therefore, an electric field analysis method is needed to achieve the output of electric field distribution and data analysis under the condition that the basic data retains the connection of physical structure and the continuity of electric field distribution, so as to reduce the amount of data required and improve the efficiency and effectiveness of modeling analysis. Summary of the Invention

[0005] The purpose of this application is to solve at least one of the above-mentioned technical defects, especially the technical defect in the prior art that the basic data in the electric field analysis process cannot retain the connection of physical structures and the continuity of electric field distribution.

[0006] In a first aspect, the present application provides an electric field analysis method, the method comprising:

[0007] Generate a target model through simulation modeling, obtain continuous electric field distribution data, and form a first data set and a second data set;

[0008] The first data set is used to indicate the physical size, physical parameters and excitation conditions of the target model, and the second data set is used to indicate the power line distribution and electric field norm of the target model;

[0009] Obtaining, by a graph neural network, a first graph feature corresponding to the first data set and a second graph feature corresponding to the second data set;

[0010] The first graph feature and the second graph feature include rasterized feature data, which are used to indicate the distribution of physical features within the grid and the information association between the grid and adjacent grids;

[0011] Through a cascaded grid processing network, target power lines are predicted based on the first graph features, and based on the second graph features, the target power lines are converted into target electric field models to generate electric field prediction results, and a preset analysis process is executed on the electric field prediction results.

[0012] As an optional implementation, obtaining the first graph feature corresponding to the first data set through a graph neural network includes:

[0013] Performing rasterized image feature extraction on the first data set using a graph neural network to obtain node features and boundary features corresponding to each first grid, and determining a first state vector corresponding to each first grid;

[0014] Aggregate the information indicated by the state vectors of adjacent grids through the attention mechanism, and update the state vector corresponding to each of the first grids;

[0015] The node feature types include circle, rectangle, triangle and their combined deformations, and the boundary features include geometric size, spatial position and material properties.

[0016] As an optional implementation manner, obtaining the second graph feature corresponding to the second data set through the graph neural network includes:

[0017] Performing rasterized image feature extraction on the second data set using a graph neural network to obtain node features, line segment features, and boundary features corresponding to each second grid, and determining a second state vector corresponding to each second grid;

[0018] Aggregate the information indicated by the state vectors of adjacent grids through the attention mechanism, and update the state vector corresponding to each second grid;

[0019] The line segment feature is used to indicate the distribution of power lines.

[0020] As an optional implementation, the cascaded grid processing network includes a cascaded UNET architecture;

[0021] The cascaded UNET architecture uses a learning rate warm-up mechanism to dynamically adjust the learning rate, optimizes the training process through an early stopping strategy, and dynamically adjusts hyperparameters based on the performance of the validation set;

[0022] The learning rate warm-up mechanism includes:

[0023] In the first stage of training, a linearly increasing learning rate warm-up phase is used, and the warm-up period is adaptively adjusted according to the sample set size, and the learning rate is dynamically reduced in the second stage of training.

[0024] As an optional implementation manner, after converting the target power lines into a target electric field norm according to the second graph feature, the method further includes:

[0025] Iteratively correct the prediction results based on the error feedback mechanism;

[0026] Wherein, the error feedback mechanism includes:

[0027] Marking abnormal areas based on the peak relative error and regional gradient difference between the target electric field norm distribution and the corresponding results of the target model;

[0028] The local prediction reliability is evaluated based on the power line direction cosine similarity, and secondary feature extraction and result fusion are performed on the low-reliability area;

[0029] An incremental training sample set is generated by combining historical error distribution data, the local parameters of the cascaded grid processing network are optimized through an elastic weight consolidation method, and the iteratively corrected target electric field norm is obtained based on the updated cascaded grid processing network.

[0030] As an optional implementation manner, generating a target model through simulation modeling and obtaining continuous electric field distribution data includes:

[0031] Through simulation modeling, a two-dimensional structural model is constructed according to the cross-sectional dimensions;

[0032] According to the cross-section material, the conductivity, dielectric constant and magnetic permeability parameters of each structural area are set, and the high potential area and ground area are set as boundary conditions;

[0033] Performing mesh division of differentiated sizes on the two-dimensional structural model to obtain a plurality of meshes;

[0034] Physical field calculation is performed on each of the divided grids obtained after division to generate continuous electric field distribution data.

[0035] In a second aspect, the present application provides an electric field analysis device, comprising:

[0036] An acquisition module is used to generate a target model through simulation modeling, acquire continuous electric field distribution data, and form a first data set and a second data set;

[0037] The first data set is used to indicate the physical size, physical parameters and excitation conditions of the target model, and the second data set is used to indicate the power line distribution and electric field norm of the target model;

[0038] a processing module, configured to obtain, through a graph neural network, a first graph feature corresponding to the first data set and a second graph feature corresponding to the second data set;

[0039] The first graph feature and the second graph feature include rasterized feature data, which are used to indicate the distribution of physical features within the grid and the information association between the grid and adjacent grids;

[0040] The processing module is also used to obtain target power lines based on the first graph feature prediction through a cascaded grid processing network, and convert the target power lines into target electric field models based on the second graph feature, generate electric field prediction results, and execute a preset analysis process on the electric field prediction results.

[0041] As an optional implementation manner, the processing module may obtain the first graph feature corresponding to the first data set through a graph neural network in a specific manner including:

[0042] Performing rasterized image feature extraction on the first data set using a graph neural network to obtain node features and boundary features corresponding to each first grid, and determining a first state vector corresponding to each first grid;

[0043] Aggregate the information indicated by the state vectors of adjacent grids through the attention mechanism, and update the state vector corresponding to each of the first grids;

[0044] The node feature types include circle, rectangle, triangle and their combined deformations, and the boundary features include geometric size, spatial position and material properties;

[0045] Furthermore, the specific manner in which the processing module obtains the second graph feature corresponding to the second data set through the graph neural network includes:

[0046] Performing rasterized image feature extraction on the second data set using a graph neural network to obtain node features, line segment features, and boundary features corresponding to each second grid, and determining a second state vector corresponding to each second grid;

[0047] Aggregate the information indicated by the state vectors of adjacent grids through the attention mechanism, and update the state vector corresponding to each second grid;

[0048] The line segment feature is used to indicate the distribution of power lines.

[0049] In a third aspect, the present application provides a computer device comprising one or more processors and a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the one or more processors, the steps of the method described in the first aspect are performed.

[0050] In a fourth aspect, the present application provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the method described in the first aspect.

[0051] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:

[0052] Based on any of the above-mentioned embodiments, the electric field analysis method proposed in this application generates a target model through simulation modeling and obtains continuous electric field distribution data, forming a first dataset containing the target model's physical dimensions, physical parameters, and excitation conditions, and a second dataset containing the distribution of power lines and the electric field norm. Subsequently, a graph neural network is used to perform gridded image feature extraction on the first and second datasets to obtain corresponding first and second graph features. These graph features not only intuitively indicate the distribution of physical features within the grid but also effectively characterize the information association between the grid and adjacent grids. On this basis, a cascaded grid processing network is used to first predict target power lines based on the first graph features. The target power lines are then converted into target electric field norms based on the second graph features. Finally, the electric field prediction results are generated and a pre-set analysis process is executed. This method, through the organic combination of simulation modeling, graph neural networks, and cascaded grid processing networks, ensures the continuity of the electric field distribution and the integrity of the physical structure, while avoiding the noise issues associated with traditional discretization methods. Furthermore, the gridded feature data processing significantly reduces data requirements and computational complexity, improving the efficiency and accuracy of modeling and analysis. During the training process, the introduction of a learning rate warm-up mechanism and an early stopping strategy further optimized the model's training efficiency and stability. Through iterative correction of the error feedback mechanism, the accuracy and reliability of the prediction results were effectively improved. In summary, the method provided in this application solves the problems of large data volumes and low processing efficiency in electric field analysis of high-voltage equipment, providing efficient and reliable technical support for troubleshooting, defect elimination, and design optimization, and has significant application value and promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0054] Figure 1 A schematic diagram of a flow chart of an electric field analysis method provided in one embodiment of the present application;

[0055] Figure 2 A schematic diagram showing the effect of an electric field analysis method provided in one embodiment of the present application;

[0056] Figure 3 A schematic diagram showing the effect of an electric field analysis method provided in one embodiment of the present application;

[0057] Figure 4 A schematic diagram showing the effect of an electric field analysis method provided in one embodiment of the present application;

[0058] Figure 5 A schematic diagram showing the effect of an electric field analysis method provided in one embodiment of the present application;

[0059] Figure 6 A schematic diagram showing the effect of an electric field analysis method provided in one embodiment of the present application;

[0060] Figure 7 This is a diagram of the internal structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0062] During the manufacturing, installation, construction, and operation of high-voltage equipment, defects, surface burrs, and other anomalies can cause localized electric field distortion, leading to partial discharge. Therefore, electric field analysis of high-voltage equipment is a key tool for troubleshooting, defect elimination, and design optimization. Due to the complex structure and numerous components of high-voltage equipment, electromagnetic field theory equations are difficult to solve. Finite element methods (FEM) are typically used for modeling and analysis. The core of this method is to discretize the target object into a finite number of meshes. Each mesh is then solved according to boundary conditions, excitations, and constraint equations, ultimately forming a complete physical field distribution. However, due to the complexity of actual target structures, the number of meshes can reach millions or even tens of millions. The large number of degrees of freedom required for the FEM method results in lengthy computational times. Furthermore, as model complexity increases, even subtle changes in the target parameters often require re-modeling and analysis. To meet practical production needs and improve the efficiency of electric field analysis, it is necessary to overcome the limitations of the FEM method and rapidly calculate the target object's electric field.

[0063] Currently, the main approach is to combine deep learning methods with physical field numerical simulation to achieve rapid electric field simulation. Taking the UNET network as an example, a dataset is first constructed based on the results of the finite element method solution. This data is then fed into the UNET for training. Through network parameter training, the deep nonlinear relationships contained in the data are obtained. These relationships are then stored as UNET weight parameters and can be re-input. Similarly, the UNET network can be replaced with other types of algorithms, such as PINN, and the basic process remains the same. The core of the existing technology is to generate a series of samples through finite element simulation, allowing the deep learning algorithm to train the underlying physical connections, ultimately bypassing the finite element calculations and obtaining the corresponding physical quantities. The accuracy of the calculation is affected by the quality of the samples. Since deep learning algorithms can only process discrete data, the data must be discretized when generating samples. For example, the UNET network requires gridded data input, while the PINN network requires scattered data. If the model obtained from finite element simulation is to be input into the UNET network, the corresponding grid must be set to transform the model into a "mosaic". If the model is to be input into the PINN network, the data must be interpolated and sampled. When handling this process, existing technologies, whether averaging the grid values ​​or performing interpolation sampling, generate discrete data matrices to characterize the dielectric constant or electric field norm of the material.

[0064] This discretization process presents two problems. The grids no longer have a physical connection, but instead represent discrete points. The electric field distribution is primarily influenced by the shape of the electrodes and boundaries, so discretizing the structure can result in significant noise in the prediction results. Furthermore, network training generally requires a regular input data matrix. Whether using the grid or sampling, the area and grid size must remain constant. The sampling points must also be evenly spaced and fixed. Under these conditions, when generating a sample set through parameter sweeping, the parameter change step size is also fixed, typically based on the grid size or the sampling point spacing. Otherwise, two types of media may appear within the grid, or sampling points may be located at the interface between two types of media, leading to confusion in the values. Therefore, both sample generation and prediction are performed with discrete parameters, meaning that only data for specific parameters can be obtained. These two issues are inherent to deep learning algorithms. To mitigate them, the primary approach is to increase the number of sampling points or the grid density. However, this significantly increases the computational complexity and does not significantly improve the computational speed compared to finite element methods. Therefore, simply using existing technologies cannot address the series of issues arising from data and structural discretization.

[0065] This application mainly adopts graph neural network combined with grid processing network, such as UNET network. First, the data is processed using graph neural network to retain the connection of physical structure and the continuity of electric field distribution. The formed data set is then used for subsequent UNET network training, thereby eliminating the corresponding limitations.

[0066] The technical concept of this application lies in the electric field analysis method provided herein. This method generates a target model through simulation modeling and obtains continuous electric field distribution data, forming a first dataset containing the target model's physical dimensions, physical parameters, and excitation conditions, and a second dataset containing the distribution of power lines and the electric field norm. Subsequently, a graph neural network is used to perform gridded image feature extraction on the first and second datasets to obtain corresponding first and second graph features. These graph features not only intuitively indicate the distribution of physical features within the grid but also effectively characterize the information association between the grid and adjacent grids. Based on this, a cascaded grid processing network is used to first predict target power lines based on the first graph features. The target power lines are then converted into target electric field norms based on the second graph features. Ultimately, the electric field prediction results are generated and a pre-defined analysis process is executed. This method, through the organic combination of simulation modeling, graph neural networks, and cascaded grid processing networks, ensures the continuity of the electric field distribution and the integrity of the physical structure, while avoiding the noise issues associated with traditional discretization methods. Furthermore, the gridded feature data processing significantly reduces data requirements and computational complexity, improving the efficiency and accuracy of modeling and analysis. During the training process, the introduction of a learning rate warm-up mechanism and an early stopping strategy further optimized the model's training efficiency and stability. Through iterative correction of the error feedback mechanism, the accuracy and reliability of the prediction results were effectively improved. In summary, the method provided in this application solves the problems of large data volumes and low processing efficiency in electric field analysis of high-voltage equipment, providing efficient and reliable technical support for troubleshooting, defect elimination, and design optimization, and has significant application value and promotion prospects.

[0067] The method provided in this application is described in detail below based on corresponding implementation methods in some actual application scenarios.

[0068] See also Figure 1 , Figure 1 A flow chart of an electric field analysis method provided in one embodiment of the present application is shown in FIG. Figure 1 As shown, the method includes:

[0069] S101, generating a target model through simulation modeling, obtaining continuous electric field distribution data, and forming a first data set and a second data set;

[0070] The first data set is used to indicate the physical size, physical parameters and excitation conditions of the target model, and the second data set is used to indicate the power line distribution and electric field norm of the target model;

[0071] S102. Obtain, through a graph neural network, a first graph feature corresponding to the first data set and a second graph feature corresponding to the second data set;

[0072] The first graph feature and the second graph feature include rasterized feature data, which are used to indicate the distribution of physical features within the grid and the information association between the grid and adjacent grids;

[0073] S103. Predicting target power lines based on the first graph features through a cascaded grid processing network, converting the target power lines into target electric field models based on the second graph features, generating electric field prediction results, and executing a preset analysis process on the electric field prediction results.

[0074] The electric field analysis method provided in this application first generates a target model through simulation modeling and obtains continuous electric field distribution data, forming a first data set containing the physical size, physical parameters and excitation conditions of the target model, and a second data set containing the distribution of power lines and electric field norm conditions. Subsequently, the first data set and the second data set are subjected to rasterized image feature extraction through a graph neural network to obtain corresponding first graph features and second graph features. These graph features can not only intuitively indicate the distribution of physical features within the grid, but also effectively characterize the information association between the grid and the adjacent grids, thereby ensuring the continuity of data processing and the integrity of the physical structure. On this basis, through a cascaded grid processing network, the target power lines are first predicted based on the first graph features, and then the target power lines are converted into the target electric field norm based on the second graph features, finally generating the electric field prediction results and executing the preset analysis process. This method effectively retains the continuity characteristics of the electric field distribution through the combination of a graph neural network and a cascaded grid processing network, while avoiding the noise problem caused by discretization. In addition, through rasterized feature data processing, the data demand and computational complexity are greatly reduced, and the efficiency and accuracy of modeling analysis are improved, thus providing reliable technical support for troubleshooting, defect elimination and design optimization of high-voltage equipment.

[0075] As an optional implementation manner, generating a target model through simulation modeling and obtaining continuous electric field distribution data includes:

[0076] Through simulation modeling, a two-dimensional structural model is constructed according to the cross-sectional dimensions;

[0077] According to the cross-section material, the conductivity, dielectric constant and magnetic permeability parameters of each structural area are set, and the high potential area and ground area are set as boundary conditions;

[0078] Performing mesh division of differentiated sizes on the two-dimensional structural model to obtain a plurality of meshes;

[0079] Physical field calculation is performed on each of the divided grids obtained after division to generate continuous electric field distribution data.

[0080] This embodiment describes in detail the specific steps of generating a target model through simulation modeling. First, a two-dimensional structural model is constructed according to the cross-sectional dimensions, and the conductivity, dielectric constant and magnetic permeability parameters of each structural area are set according to the cross-sectional material, and the high potential area and the grounding area are set as boundary conditions. This process ensures that the physical parameters of the model are highly matched with the actual application scenario. Subsequently, the two-dimensional structural model is meshed with differentiated sizes to obtain multiple meshes. This meshing method can flexibly adjust the meshing accuracy according to actual needs, while ensuring computational efficiency and taking into account the detailed expression of the model. Finally, physical field calculations are performed on each mesh obtained after meshing to generate continuous electric field distribution data. In this way, this embodiment effectively solves the problems of excessive data volume and low processing efficiency in the traditional modeling process, ensures the continuity and integrity of the electric field distribution data, and provides a reliable data basis for subsequent graph feature extraction and electric field analysis, thereby improving the efficiency and application value of the entire electric field analysis process.

[0081] As an optional implementation, obtaining the first graph feature corresponding to the first data set through a graph neural network includes:

[0082] Performing rasterized image feature extraction on the first data set using a graph neural network to obtain node features and boundary features corresponding to each first grid, and determining a first state vector corresponding to each first grid;

[0083] Aggregate the information indicated by the state vectors of adjacent grids through the attention mechanism, and update the state vector corresponding to each of the first grids;

[0084] The node feature types include circle, rectangle, triangle and their combined deformations, and the boundary features include geometric size, spatial position and material properties.

[0085] This embodiment further clarifies the specific process of obtaining the first graph feature through the graph neural network. First, the first data set is subjected to rasterized image feature extraction to obtain the node features and boundary features of each first grid, and the corresponding first state vector is determined. Among them, the node feature types cover circles, rectangles, triangles and their combined deformations, and the boundary features describe the geometric dimensions, spatial positions and material properties in detail. Through the extraction of these feature data, the physical structure information of the target model can be fully characterized. Subsequently, the information indicated by the state vectors of adjacent grids is aggregated through the attention mechanism, and the state vectors corresponding to each first grid are updated. This process can not only effectively integrate the correlation information between adjacent grids, but also further improve the accuracy and continuity of feature expression by dynamically updating the state vector. In this way, this embodiment ensures the efficiency and reliability of the first graph feature in dealing with the problem of physical structure continuity, provides a more accurate data basis for the subsequent target power line prediction, and thus improves the efficiency and accuracy of the entire electric field analysis process.

[0086] As an optional implementation manner, obtaining the second graph feature corresponding to the second data set through the graph neural network includes:

[0087] Performing rasterized image feature extraction on the second data set using a graph neural network to obtain node features, line segment features, and boundary features corresponding to each second grid, and determining a second state vector corresponding to each second grid;

[0088] Aggregate the information indicated by the state vectors of adjacent grids through the attention mechanism, and update the state vector corresponding to each second grid;

[0089] The line segment feature is used to indicate the distribution of power lines.

[0090] This embodiment describes in detail the specific process of obtaining the second graph features through a graph neural network. First, rasterized image feature extraction is performed on the second data set to obtain the node features, line segment features, and boundary features of each second grid, and the corresponding second state vector is determined. Among them, the line segment features are specifically used to indicate the distribution of power lines, which enables the second graph features to more intuitively reflect the spatial information and distribution characteristics of the power lines. Subsequently, the information indicated by the state vectors of adjacent grids is aggregated through the attention mechanism, and the state vector corresponding to each second grid is updated. This process not only enhances the information correlation between adjacent grids, but also ensures the continuous expression of the power line distribution data. In this way, this embodiment effectively solves the problem of missing physical associations in traditional discretization methods and ensures the accuracy and completeness of the power line distribution data. In addition, through the comprehensive extraction of node features, line segment features, and boundary features, the information dimension of the second graph features is further enriched, providing more reliable data support for the subsequent conversion of the target power line into the target electric field model, thereby improving the accuracy and efficiency of the entire electric field analysis process.

[0091] As an optional implementation, the cascaded grid processing network includes a cascaded UNET architecture;

[0092] The cascaded UNET architecture uses a learning rate warm-up mechanism to dynamically adjust the learning rate, optimizes the training process through an early stopping strategy, and dynamically adjusts hyperparameters based on the performance of the validation set;

[0093] The learning rate warm-up mechanism includes:

[0094] In the first stage of training, a linearly increasing learning rate warm-up phase is used, and the warm-up period is adaptively adjusted according to the sample set size, and the learning rate is dynamically reduced in the second stage of training.

[0095] This embodiment further optimizes the training process of the cascaded grid processing network. Specifically, the cascaded grid processing network adopts a cascaded UNET architecture, which enhances the ability of feature extraction and fusion through multi-layer cascading. During the training process, the learning rate is dynamically adjusted by introducing a learning rate warm-up mechanism, that is, a linearly growing learning rate warm-up phase is adopted in the first stage of training, and the warm-up period is adaptively adjusted according to the size of the sample set, while the learning rate is dynamically reduced in the second stage of training. This learning rate adjustment strategy effectively avoids the oscillation problem in the early stage of training, while ensuring stable convergence in the later stage of training. In addition, the training process is terminated by the early stopping strategy to prevent the occurrence of overfitting, and the hyperparameters are dynamically adjusted according to the performance of the validation set, which further improves the generalization ability and prediction accuracy of the model. Through the above optimization measures, this embodiment significantly improves the training efficiency and stability of the cascaded grid processing network, ensuring its reliability and accuracy in complex electric field analysis tasks.

[0096] As an optional implementation manner, after converting the target power lines into a target electric field norm according to the second graph feature, the method further includes:

[0097] Iteratively correct the prediction results based on the error feedback mechanism;

[0098] Wherein, the error feedback mechanism includes:

[0099] Marking abnormal areas based on the peak relative error and regional gradient difference between the target electric field norm distribution and the corresponding results of the target model;

[0100] The local prediction reliability is evaluated based on the power line direction cosine similarity, and secondary feature extraction and result fusion are performed on the low-reliability area;

[0101] An incremental training sample set is generated by combining historical error distribution data, the local parameters of the cascaded grid processing network are optimized through an elastic weight consolidation method, and the iteratively corrected target electric field norm is obtained based on the updated cascaded grid processing network.

[0102] This embodiment introduces an error feedback mechanism after generating the target electric field modulus, and further improves the accuracy of the analysis by iteratively correcting the prediction results. Specifically, based on the peak relative error and regional gradient difference between the target electric field modulus distribution and the corresponding result of the target model, the abnormal area is marked; then, the local prediction credibility is evaluated based on the power line direction cosine similarity, and secondary feature extraction and result fusion are performed on the low-credibility area; finally, an incremental training sample set is generated based on the historical error distribution data, and the local parameters of the cascade grid processing network are optimized by the elastic weight consolidation method, and the iteratively corrected target electric field modulus is obtained based on the updated network. This process can not only effectively identify and correct abnormal areas in the prediction, but also improve the prediction credibility of local areas through secondary feature extraction and result fusion. At the same time, the elastic weight consolidation method ensures the stability and continuity of model optimization. Through this iterative correction mechanism, this embodiment significantly improves the accuracy and reliability of the electric field prediction results, providing more accurate data support for troubleshooting and design optimization of high-voltage equipment.

[0103] The following describes in detail the application of various embodiments of the present application in combination with an actual application scenario. The corresponding method of the present application can be divided into four parts: model simulation, data generation, network training and prediction results. The main creative point lies in the data generation part.

[0104] Model simulation:

[0105] For the object to be analyzed, a corresponding finite element model is established, and after meshing and physical field calculation, the relevant electric field norm data is obtained. In engineering applications, the electric field analysis of the cross section is generally the main focus. For this reason, the present invention only focuses on two-dimensional electric field calculation. The specific steps are: according to the cross-sectional size structure, a two-dimensional structural model is established in the simulation software; according to the cross-sectional material, the necessary parameters such as conductivity, dielectric constant, and magnetic permeability of different structures in the model are set; according to the actual situation, high potential and grounding areas are set; meshing is performed, and some areas can be differentiated according to needs during meshing to improve calculation accuracy; then physical field calculation is performed to obtain the electric field distribution on the cross section. Determine the structure and size of focus, and obtain the electric field distribution data at different sizes through parameter scanning to form a sample set. The patent of this invention has no restrictions on the parameter scanning step size, and does not need to be scanned according to the grid size or sampling point interval. According to the model simulation results, two sets of input and output data sets are formed, where the input data includes the model physical size, model physical parameters, and excitation, and the output data is the electric field norm and power line distribution. In this application, the first data set is used to refer to the input data set, and the second data set is used to refer to the output data set, which should be understood equivalently in specific application scenarios.

[0106] Data generation:

[0107] In the dataset generated by model simulation, the input data are the parameters of the physical model, which have a physical connection; the output data are the results of the finite element calculation, which are in the form of scattered points with coordinates determined by the meshing. In existing technologies, to obtain a regular input and output format, interpolation sampling or a fixed grid is required to achieve input and output unification, but this results in discretization of the model and data.

[0108] This application uses graph neural network combined with UNET network technology for processing, that is, the subsequent electric field calculation network uses UNET, and the physical connection between the front-end grids is extracted through graph neural network. The advantage of using UNET as the back-end electric field calculation is that the grid divides a complete area, while PINN and other methods discretize the model from the beginning through scattered input, and it is difficult to process the mutual physical connection later. Graph neural network is a type of deep learning model specifically used to process and analyze graph structure data. It can effectively capture the complex relationships in the graph structure, thereby comprehensively describing the physical connections within the grid.

[0109] For the graph neural network, the dataset is first rasterized. The grid size and area are determined according to actual needs. A 100*100 grid layout can be used, and the same grid settings are used for each set of inputs and outputs.

[0110] Next, graph feature extraction is performed on the input data, i.e., the first dataset. The input data includes the model's physical dimensions, model physical parameters, and excitations. This information is based on the physical dimensions, meaning that the graph features are consistent. Therefore, the same graph neural network is used for processing. Because physical models in finite element calculations are combinations of standard shapes, after rasterizing the input data, the images within each grid share the same characteristics. Therefore, the present invention employs a universal method to extract image features within each grid. Figure 2 This is a schematic diagram of the effect of the electric field analysis method provided by one embodiment of the present application, which is used to indicate the general form of the input data raster image, such as Figure 2 As shown, the general form consists of nodes and edges. Node shapes include circles (semicircles, full circles, arcs, and rings), rectangles, and triangles; edges are composed of rectangles, the number of which is not fixed. This setup allows complex physical models to be converted into a combination of general forms.

[0111] Figure 3 This is a schematic diagram of the effect of the electric field analysis method provided in one embodiment of the present application, which illustrates a two-dimensional model of a ball-plate electrode. Figure 3 After 10X10 grid division, Figure 4 This is a schematic diagram of the effect of the electric field analysis method provided by an embodiment of the present application. The grid image of a typical position 1 is shown as follows: Figure 4 As shown, the grid image features are: circular nodes, sized by the circle radius and center coordinates, with no edges. Furthermore, the grid image features for typical positions 2 and 3 are: 2 has no nodes, an edge at the top, sized to the grid width, and no edges on the remaining three sides; 3 has rectangular nodes, sized to the grid width, and no edges on the remaining four sides. Because the grid is densely divided in real models, it is extremely rare for a node to be located in the center. This general setting allows for feature extraction from raster images.

[0112] Figure 5 A schematic diagram of the effect of the electric field analysis method provided in one embodiment of the present application is used to illustrate the graph neural network of the present application, the Attentive FP network (AFPNet). The process is as follows Figure 5 As shown. First, in the graph featurization stage, information is extracted from a single grid graph and embedded into a single grid feature graph. This information includes node features (node ​​shape, size, position, material) and edge features (edge ​​size and position, material). Based on the above processing, a single grid graph G = (V, E) is obtained, where the set , represents the node, set E={e1, e2, }, representing edges. Each vertex and edge in the graph is assigned an initial feature vector to represent the structural state, encoded in a one-hot format. The encoded graph is then used as a single grid feature.

[0113] Through the above operations, the initial state of each grid can be obtained. To establish physical connections between adjacent grids, during the network message passing phase, the initial state vectors of a single grid and its neighboring grids are embedded through the graph attention layer according to Equation 1 below. This allows the single grid to gradually aggregate "messages" (including node and edge information) from its neighbors. Then, according to Equation 2 below, a new state vector is generated for each grid in each attention layer, allowing it to associate information with adjacent grids.

[0114] (1)

[0115] (2)

[0116] Where N(ν) is the set of all neighboring grids of node ν, and is the state vector of the m-1th layer neighboring grid u and the target grid ν, It is the graph attention mechanism of the m-1 layer, and the neighboring grids are Aggregate information on the target grid to obtain attention context , is the state vector of the target grid ν at the mth layer, and is updated by the update function at the m-1th layer. Will and the previous layer state vector Update to get the state vector of the mth layer Clearly, the state vector is continuously updated after passing through multiple stacking layers, containing richer neighborhood information and enhancing the feature representation of the target grid. By stacking the states of adjacent grids, we can effectively establish physical connections between different grids. The resulting state matrix for the entire grid serves as the input to the UNET.

[0117] Afterwards, the output data, i.e., the second data set, is subjected to graph feature extraction. The output data of the prior art is an electric field modulus or a magnetic field modulus, which is rasterized based on the physical field data, i.e., mosaicized, and has no physical association. Figure 6 This is a schematic diagram of the effect of the electric field analysis method provided in one embodiment of the present application, as shown in FIG. Figure 6 As shown, in the present invention, the output electric field norm data is not directly extracted, but the power line data output by the finite element calculation is first processed. The power line is a manually defined parameter. Its tangent direction represents the direction of the field strength at the point, and its density represents the magnitude of the field strength. The power line is closely coupled with the electric field strength. The advantage of analyzing the power line is that the power line is a continuous directed line segment and does not intersect with each other. Its characteristics are more obvious and more conducive to the processing of the graph neural network. The electric field norm data is also rasterized, mainly for the subsequent establishment of the connection between the power line and the electric field norm.

[0118] Whether it is power lines or electric field strength, their distribution patterns are based on physical structures, but in the results, physical structures are more likely to be affected by boundaries. After rasterizing the power lines, the image features within the grid are more obvious, generally pointing from one side to the other. Since the shapes of the power lines within the same grid are basically the same after rasterization, the present invention sets a general form to extract image features within each grid. Figure 6 As shown, the general form consists of boundaries (nodes, edges) and line segments. The boundary settings are consistent with the previous steps, that is, the node shapes can be circles (semicircles, full circles, arcs, rings), rectangles, and triangles; the edges are composed of rectangles, and the number is not fixed.

[0119] This step still uses the Attentive FP network (AFPNet). First, in the graph featurization stage, information is extracted from a single grid graph and embedded into a single grid feature graph. This information includes boundary features and line segment information. The boundary information includes node shape, size, position, and the size and position of each edge. The influence of material and other factors has been reflected in the line segment features. In order to reduce the complexity of the network, this is no longer considered. The line segment information is the starting point, end point, minimum spacing, maximum spacing, average spacing, and concavity. According to the above processing, the graph G=(V, E, L) of a single grid is obtained, where the set V={v1, v2, }, representing nodes, set E={e1, e2, }, represents the edge, set L={l1, l2, }, representing a line segment.

[0120] Subsequent processing is consistent with the previous step. By continuously updating the state vector of a single grid, the neighborhood information contained therein is enriched, and the feature representation of the target grid is also enhanced. By stacking the states of adjacent grids, the physical connection between different grids can be effectively established. The final state matrix of the entire grid is used as the input of the UNET.

[0121] Network training:

[0122] After obtaining the network input dataset, network training is performed. Since the final prediction should be the electric field norm, the only quantity after graph neural network processing is the power lines. Therefore, two grids are trained: a front grid for power line prediction and a back grid for converting power lines to electric field norms. Both grids are UNET grids with the same settings.

[0123] Based on the UNET input processed by the graph neural network, a two-dimensional UNET network is constructed. Training hyperparameters are set based on model complexity. A learning rate warm-up can be used to dynamically adjust the learning rate. After the network is configured, 80% of the dataset is used for network training, and 20% is used for validation.

[0124] Prediction results:

[0125] After network training is completed, 20% of the data set is used for testing, and the electric field prediction results of the test samples are output. The network prediction results are compared with the finite element calculation results, and the model performance is evaluated and analyzed to achieve rapid calculation of the electric field.

[0126] The present application also provides an electric field analysis device, comprising:

[0127] An acquisition module is used to generate a target model through simulation modeling, acquire continuous electric field distribution data, and form a first data set and a second data set;

[0128] The first data set is used to indicate the physical size, physical parameters and excitation conditions of the target model, and the second data set is used to indicate the power line distribution and electric field norm of the target model;

[0129] a processing module, configured to obtain, through a graph neural network, a first graph feature corresponding to the first data set and a second graph feature corresponding to the second data set;

[0130] The first graph feature and the second graph feature include rasterized feature data, which are used to indicate the distribution of physical features within the grid and the information association between the grid and adjacent grids;

[0131] The processing module is also used to obtain target power lines based on the first graph feature prediction through a cascaded grid processing network, and convert the target power lines into target electric field models based on the second graph feature, generate electric field prediction results, and execute a preset analysis process on the electric field prediction results.

[0132] The electric field analysis method provided in this application first generates a target model through simulation modeling and obtains continuous electric field distribution data, forming a first data set containing the physical size, physical parameters and excitation conditions of the target model, and a second data set containing the distribution of power lines and electric field norm conditions. Subsequently, the first data set and the second data set are subjected to rasterized image feature extraction through a graph neural network to obtain corresponding first graph features and second graph features. These graph features can not only intuitively indicate the distribution of physical features within the grid, but also effectively characterize the information association between the grid and the adjacent grids, thereby ensuring the continuity of data processing and the integrity of the physical structure. On this basis, through a cascaded grid processing network, the target power lines are first predicted based on the first graph features, and then the target power lines are converted into the target electric field norm based on the second graph features, finally generating the electric field prediction results and executing the preset analysis process. This method effectively retains the continuity characteristics of the electric field distribution through the combination of a graph neural network and a cascaded grid processing network, while avoiding the noise problem caused by discretization. In addition, through rasterized feature data processing, the data demand and computational complexity are greatly reduced, and the efficiency and accuracy of modeling analysis are improved, thus providing reliable technical support for troubleshooting, defect elimination and design optimization of high-voltage equipment.

[0133] As an optional implementation manner, the processing module may obtain the first graph feature corresponding to the first data set through a graph neural network in a specific manner including:

[0134] Performing rasterized image feature extraction on the first data set using a graph neural network to obtain node features and boundary features corresponding to each first grid, and determining a first state vector corresponding to each first grid;

[0135] Aggregate the information indicated by the state vectors of adjacent grids through the attention mechanism, and update the state vector corresponding to each of the first grids;

[0136] The node feature types include circle, rectangle, triangle and their combined deformations, and the boundary features include geometric size, spatial position and material properties;

[0137] This embodiment further clarifies the specific process of obtaining the first graph feature through the graph neural network. First, the first data set is subjected to rasterized image feature extraction to obtain the node features and boundary features of each first grid, and the corresponding first state vector is determined. Among them, the node feature types cover circles, rectangles, triangles and their combined deformations, and the boundary features describe the geometric dimensions, spatial positions and material properties in detail. Through the extraction of these feature data, the physical structure information of the target model can be fully characterized. Subsequently, the information indicated by the state vectors of adjacent grids is aggregated through the attention mechanism, and the state vectors corresponding to each first grid are updated. This process can not only effectively integrate the correlation information between adjacent grids, but also further improve the accuracy and continuity of feature expression by dynamically updating the state vector. In this way, this embodiment ensures the efficiency and reliability of the first graph feature in dealing with the problem of physical structure continuity, provides a more accurate data basis for the subsequent target power line prediction, and thus improves the efficiency and accuracy of the entire electric field analysis process.

[0138] As an optional implementation manner, the specific manner in which the processing module obtains the second graph feature corresponding to the second data set through a graph neural network includes:

[0139] Performing rasterized image feature extraction on the second data set using a graph neural network to obtain node features, line segment features, and boundary features corresponding to each second grid, and determining a second state vector corresponding to each second grid;

[0140] Aggregate the information indicated by the state vectors of adjacent grids through the attention mechanism, and update the state vector corresponding to each second grid;

[0141] The line segment feature is used to indicate the distribution of power lines.

[0142] This embodiment describes in detail the specific process of obtaining the second graph features through a graph neural network. First, rasterized image feature extraction is performed on the second data set to obtain the node features, line segment features, and boundary features of each second grid, and the corresponding second state vector is determined. Among them, the line segment features are specifically used to indicate the distribution of power lines, which enables the second graph features to more intuitively reflect the spatial information and distribution characteristics of the power lines. Subsequently, the information indicated by the state vectors of adjacent grids is aggregated through the attention mechanism, and the state vector corresponding to each second grid is updated. This process not only enhances the information correlation between adjacent grids, but also ensures the continuous expression of the power line distribution data. In this way, this embodiment effectively solves the problem of missing physical associations in traditional discretization methods and ensures the accuracy and completeness of the power line distribution data. In addition, through the comprehensive extraction of node features, line segment features, and boundary features, the information dimension of the second graph features is further enriched, providing more reliable data support for the subsequent conversion of the target power line into the target electric field model, thereby improving the accuracy and efficiency of the entire electric field analysis process.

[0143] As an optional implementation, the cascaded grid processing network includes a cascaded UNET architecture;

[0144] The processing module is further configured to: dynamically adjust the learning rate of the cascaded UNET architecture using a learning rate warm-up mechanism, optimize the training process using an early stopping strategy, and dynamically adjust hyperparameters based on validation set performance;

[0145] The learning rate warm-up mechanism includes:

[0146] In the first stage of training, a linearly increasing learning rate warm-up phase is used, and the warm-up period is adaptively adjusted according to the sample set size, and the learning rate is dynamically reduced in the second stage of training.

[0147] This embodiment further optimizes the training process of the cascaded grid processing network. Specifically, the cascaded grid processing network adopts a cascaded UNET architecture, which enhances the ability of feature extraction and fusion through multi-layer cascading. During the training process, the learning rate is dynamically adjusted by introducing a learning rate warm-up mechanism, that is, a linearly growing learning rate warm-up phase is adopted in the first stage of training, and the warm-up period is adaptively adjusted according to the size of the sample set, while the learning rate is dynamically reduced in the second stage of training. This learning rate adjustment strategy effectively avoids the oscillation problem in the early stage of training, while ensuring stable convergence in the later stage of training. In addition, the training process is terminated by the early stopping strategy to prevent the occurrence of overfitting, and the hyperparameters are dynamically adjusted according to the performance of the validation set, which further improves the generalization ability and prediction accuracy of the model. Through the above optimization measures, this embodiment significantly improves the training efficiency and stability of the cascaded grid processing network, ensuring its reliability and accuracy in complex electric field analysis tasks.

[0148] As an optional implementation manner, the processing module is further configured to iteratively correct the prediction result based on an error feedback mechanism after converting the target power line into a target electric field norm according to the second graph feature;

[0149] Wherein, the error feedback mechanism includes:

[0150] Marking abnormal areas based on the peak relative error and regional gradient difference between the target electric field norm distribution and the corresponding results of the target model;

[0151] The local prediction reliability is evaluated based on the power line direction cosine similarity, and secondary feature extraction and result fusion are performed on the low-reliability area;

[0152] An incremental training sample set is generated by combining historical error distribution data, the local parameters of the cascaded grid processing network are optimized through an elastic weight consolidation method, and the iteratively corrected target electric field norm is obtained based on the updated cascaded grid processing network.

[0153] This embodiment introduces an error feedback mechanism after generating the target electric field modulus, and further improves the accuracy of the analysis by iteratively correcting the prediction results. Specifically, based on the peak relative error and regional gradient difference between the target electric field modulus distribution and the corresponding result of the target model, the abnormal area is marked; then, the local prediction credibility is evaluated based on the power line direction cosine similarity, and secondary feature extraction and result fusion are performed on the low-credibility area; finally, an incremental training sample set is generated based on the historical error distribution data, and the local parameters of the cascade grid processing network are optimized by the elastic weight consolidation method, and the iteratively corrected target electric field modulus is obtained based on the updated network. This process can not only effectively identify and correct abnormal areas in the prediction, but also improve the prediction credibility of local areas through secondary feature extraction and result fusion. At the same time, the elastic weight consolidation method ensures the stability and continuity of model optimization. Through this iterative correction mechanism, this embodiment significantly improves the accuracy and reliability of the electric field prediction results, providing more accurate data support for troubleshooting and design optimization of high-voltage equipment.

[0154] As an optional implementation manner, the acquisition module generates a target model through simulation modeling, and the specific manner of acquiring continuous electric field distribution data includes:

[0155] Through simulation modeling, a two-dimensional structural model is constructed according to the cross-sectional dimensions;

[0156] According to the cross-section material, the conductivity, dielectric constant and magnetic permeability parameters of each structural area are set, and the high potential area and ground area are set as boundary conditions;

[0157] Performing mesh division of differentiated sizes on the two-dimensional structural model to obtain a plurality of meshes;

[0158] Physical field calculation is performed on each of the divided grids obtained after division to generate continuous electric field distribution data.

[0159] This embodiment describes in detail the specific steps of generating a target model through simulation modeling. First, a two-dimensional structural model is constructed according to the cross-sectional dimensions, and the conductivity, dielectric constant and magnetic permeability parameters of each structural area are set according to the cross-sectional material, and the high potential area and the grounding area are set as boundary conditions. This process ensures that the physical parameters of the model are highly matched with the actual application scenario. Subsequently, the two-dimensional structural model is meshed with differentiated sizes to obtain multiple meshes. This meshing method can flexibly adjust the meshing accuracy according to actual needs, while ensuring computational efficiency and taking into account the detailed expression of the model. Finally, physical field calculations are performed on each mesh obtained after meshing to generate continuous electric field distribution data. In this way, this embodiment effectively solves the problems of excessive data volume and low processing efficiency in the traditional modeling process, ensures the continuity and integrity of the electric field distribution data, and provides a reliable data basis for subsequent graph feature extraction and electric field analysis, thereby improving the efficiency and application value of the entire electric field analysis process.

[0160] It should be noted that it should be understood that the division of the various modules of the above device is merely a division of logical functions. In actual implementation, they can be fully or partially integrated into one physical entity, or they can be physically separated. Moreover, these modules can all be implemented in the form of software called by a processing element; or they can all be implemented in the form of hardware; or some modules can be implemented in the form of software called by a processing element, and some modules can be implemented in the form of hardware. For example, the processing module can be a separately established processing element, or it can be integrated into a chip of the above device. In addition, it can also be stored in the memory of the above device in the form of program code, and called by a processing element of the above device to perform the functions of the above-mentioned module. The implementation of other modules is similar. In addition, these modules can all or partly be integrated together, or they can be implemented independently. The processing element here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each of the above modules can be completed by the hardware integrated logic circuit in the processor element or by instructions in the form of software.

[0161] Schematically, as Figure 7 As shown, Figure 7 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of the present application. The computer device 300 can be provided as a server. Figure 7Computer device 300 includes a processing component 302, which further includes one or more processors, and a memory resource represented by memory 301 for storing instructions executable by processing component 302, such as an application. The application stored in memory 301 may include one or more modules, each corresponding to a set of instructions. In addition, processing component 302 is configured to execute the instructions to perform the method of any of the above embodiments.

[0162] The computer device 300 may further include a power supply component 303 configured to perform power management of the computer device 300, a wired or wireless network interface 304 configured to connect the computer device 300 to a network, and an input / output (I / O) interface 305. The computer device 300 may operate based on an operating system stored in the memory 301, such as Windows Server™, Mac OS X™, Unix™, Linux™, Free BSD™, or the like.

[0163] Those skilled in the art will understand that Figure 7 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0164] An embodiment of the present application provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute a method as provided in any embodiment.

[0165] Finally, it should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0166] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referenced to each other.

[0167] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An electric field analysis method, characterized in that: The method comprises: Generate a target model through simulation modeling, obtain continuous electric field distribution data, and form a first data set and a second data set; The first data set is used to indicate the physical size, physical parameters and excitation conditions of the target model, and the second data set is used to indicate the power line distribution and electric field norm of the target model; Obtaining, by a graph neural network, a first graph feature corresponding to the first data set and a second graph feature corresponding to the second data set; The first graph feature and the second graph feature include rasterized feature data, which are used to indicate the distribution of physical features within the grid and the information association between the grid and adjacent grids; Through a cascaded grid processing network, target power lines are predicted based on the first graph features, and based on the second graph features, the target power lines are converted into target electric field models to generate electric field prediction results, and a preset analysis process is executed on the electric field prediction results.

2. The method according to claim 1, characterized in that The obtaining, by using a graph neural network, a first graph feature corresponding to the first data set includes: Performing rasterized image feature extraction on the first data set using a graph neural network to obtain node features and boundary features corresponding to each first grid, and determining a first state vector corresponding to each first grid; Aggregate the information indicated by the state vectors of adjacent grids through the attention mechanism, and update the state vector corresponding to each of the first grids; The node feature types include circle, rectangle, triangle and their combined deformations, and the boundary features include geometric size, spatial position and material properties.

3. The method according to claim 2, characterized in that The obtaining, by using a graph neural network, a second graph feature corresponding to the second data set includes: Performing rasterized image feature extraction on the second data set using a graph neural network to obtain node features, line segment features, and boundary features corresponding to each second grid, and determining a second state vector corresponding to each second grid; Aggregate the information indicated by the state vectors of adjacent grids through the attention mechanism, and update the state vector corresponding to each second grid; The line segment feature is used to indicate the distribution of power lines.

4. The method according to claim 1, wherein The cascaded grid processing network includes a cascaded UNET architecture; The cascaded UNET architecture uses a learning rate warm-up mechanism to dynamically adjust the learning rate, optimizes the training process through an early stopping strategy, and dynamically adjusts hyperparameters based on the performance of the validation set; The learning rate warm-up mechanism includes: In the first stage of training, a linearly increasing learning rate warm-up phase is used, and the warm-up period is adaptively adjusted according to the sample set size, and the learning rate is dynamically reduced in the second stage of training.

5. The method according to claim 1, characterized in that After converting the target power lines into a target electric field norm according to the second graph feature, the method further includes: Iteratively correct the prediction results based on the error feedback mechanism; Wherein, the error feedback mechanism includes: Marking abnormal areas based on the peak relative error and regional gradient difference between the target electric field norm distribution and the corresponding results of the target model; The local prediction reliability is evaluated based on the power line direction cosine similarity, and secondary feature extraction and result fusion are performed on the low-reliability area; An incremental training sample set is generated by combining historical error distribution data, the local parameters of the cascaded grid processing network are optimized through an elastic weight consolidation method, and the iteratively corrected target electric field norm is obtained based on the updated cascaded grid processing network.

6. The method according to claim 1, characterized in that The target model is generated through simulation modeling to obtain continuous electric field distribution data, including: Through simulation modeling, a two-dimensional structural model is constructed according to the cross-sectional dimensions; According to the cross-section material, the conductivity, dielectric constant and magnetic permeability parameters of each structural area are set, and the high potential area and ground area are set as boundary conditions; Performing mesh division of differentiated sizes on the two-dimensional structural model to obtain a plurality of meshes; Physical field calculation is performed on each of the divided grids obtained after division to generate continuous electric field distribution data.

7. An electric field analysis device, characterized in that: The device comprises: An acquisition module is used to generate a target model through simulation modeling, acquire continuous electric field distribution data, and form a first data set and a second data set; The first data set is used to indicate the physical size, physical parameters and excitation conditions of the target model, and the second data set is used to indicate the power line distribution and electric field norm of the target model; a processing module, configured to obtain, through a graph neural network, a first graph feature corresponding to the first data set and a second graph feature corresponding to the second data set; The first graph feature and the second graph feature include rasterized feature data, which are used to indicate the distribution of physical features within the grid and the information association between the grid and adjacent grids; The processing module is also used to obtain target power lines based on the first graph feature prediction through a cascaded grid processing network, and convert the target power lines into target electric field models based on the second graph feature, generate electric field prediction results, and execute a preset analysis process on the electric field prediction results.

8. The device according to claim 7, characterized in that The specific manner in which the processing module obtains the first graph feature corresponding to the first data set through the graph neural network includes: Performing rasterized image feature extraction on the first data set using a graph neural network to obtain node features and boundary features corresponding to each first grid, and determining a first state vector corresponding to each first grid; Aggregate the information indicated by the state vectors of adjacent grids through the attention mechanism, and update the state vector corresponding to each of the first grids; The node feature types include circle, rectangle, triangle and their combined deformations, and the boundary features include geometric size, spatial position and material properties; Furthermore, the specific manner in which the processing module obtains the second graph feature corresponding to the second data set through the graph neural network includes: Performing rasterized image feature extraction on the second data set using a graph neural network to obtain node features, line segment features, and boundary features corresponding to each second grid, and determining a second state vector corresponding to each second grid; Aggregate the information indicated by the state vectors of adjacent grids through the attention mechanism, and update the state vector corresponding to each second grid; The line segment feature is used to indicate the distribution of power lines.

9. A computer device, characterized in that: The method comprises one or more processors and a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the one or more processors, the steps of the method according to any one of claims 1 to 6 are performed.

10. A storage medium, characterized in that: The storage medium stores computer-readable instructions, which, when executed by one or more processors, enable the one or more processors to perform the steps of the method according to any one of claims 1 to 6.

Images