Method and system for identifying three-dimensional eddy current field of converter transformer

Abstract

The invention discloses a three-dimensional eddy current field identification method and system for a converter transformer. The method comprises the following steps: acquiring target structure working condition parameters of a to-be-identified converter transformer; obtaining target magnetic vector potential distribution low-dimensional features corresponding to the target structure working condition parameters according to a preset reduced-order mapping model; wherein the reduced-order mapping model is used for representing a mapping relation between structure working condition parameters and magnetic vector potential distribution low-dimensional features; performing three-dimensional eddy current field identification on the target magnetic vector potential distribution low-dimensional features by using a pre-trained three-dimensional eddy current field identification network to obtain three-dimensional eddy current field distribution information of the to-be-identified converter transformer; wherein the three-dimensional eddy current field identification network is obtained by training a DGM network by using three-dimensional eddy current field finite element simulation data of the converter transformer under different structure working condition parameters. According to the method, the three-dimensional eddy current field identification efficiency of the converter transformer can be remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of converter transformer technology, and in particular to a method and system for identifying three-dimensional eddy current fields in converter transformers. Background Technology

[0002] Ultra-high voltage direct current (UHVDC) transmission has become an important means of power grid interconnection due to its advantages such as long transmission distance, large transmission capacity, and high stability. Among its components, the UHV converter transformer, as a core piece of equipment, undertakes key functions such as power transmission, AC / DC isolation, and voltage conversion; its operational safety and stability are particularly important. Compared to ordinary transformers, the internal electromagnetic environment of an UHV converter transformer is more complex. Improper design can lead to localized overheating, accelerating insulation aging and affecting its reliable operation.

[0003] Leakage flux is a particularly significant problem in converter transformers, significantly impacting their operation. Specifically, leakage flux entering metal structural components induces eddy current losses, which are particularly pronounced in large components and cannot be ignored. Eddy current losses increase the load losses of the converter transformer, reducing overall efficiency. Furthermore, dense leakage flux can lead to the concentration of eddy current losses within structural components, resulting in localized overheating and insulation aging, threatening the safe operation of the converter transformer.

[0004] To avoid eddy current losses affecting the safe operation of converter transformers, it is necessary to identify the three-dimensional eddy current field distribution within the transformer. Current techniques typically utilize the finite element method (FEM) to solve for this field. However, the induced current in the metal clamps of the converter transformer generates a skin effect. To accurately capture this effect, the mesh size on the metal clamp surface needs to be reduced, leading to an increase in the number of meshes. This increase in mesh size significantly raises the dimension of the stiffness matrix, resulting in increased computational load and decreased computational speed. Therefore, current techniques struggle to quickly identify the three-dimensional eddy current field of converter transformers. Summary of the Invention

[0005] This invention provides a method and system for identifying the three-dimensional eddy current field of a converter transformer, in order to solve the technical problem that existing technologies are unable to quickly identify the three-dimensional eddy current field of a converter transformer.

[0006] To address the aforementioned technical problems, a first aspect of this invention provides a method for identifying the three-dimensional eddy current field of a converter transformer, comprising: Obtain the target structural operating parameters of the converter transformer to be identified; According to the preset reduced-order mapping model, the low-dimensional features of the target magnetic vector potential distribution corresponding to the target structural operating parameters are obtained; wherein, the reduced-order mapping model is used to characterize the mapping relationship between the structural operating parameters and the low-dimensional features of the magnetic vector potential distribution. A pre-trained three-dimensional eddy current field identification network is used to identify the low-dimensional features of the target magnetic vector potential distribution in three dimensions, thereby obtaining the three-dimensional eddy current field distribution information of the converter transformer to be identified. The three-dimensional eddy current field identification network is obtained by training the DGM network with three-dimensional eddy current field finite element simulation data of the converter transformer under different structural operating parameters.

[0007] As a preferred embodiment, the method specifically trains the three-dimensional eddy current field recognition network through the following steps: Construct the physical residual loss function of the DGM network; wherein the physical residual loss function is a weighted sum of the control equation residual, the control boundary condition residual, and the initial magnetic vector potential condition residual; the control equation residual is determined based on Maxwell's equations, the control boundary condition residual is determined based on the output magnetic vector potential of the DGM network and the preset magnetic vector potential boundary conditions, and the initial magnetic vector potential condition residual is determined based on the preset initial magnetic vector potential distribution data and the output magnetic vector potential of the DGM network at the initial time; Based on the aforementioned three-dimensional vortex field finite element simulation data, a training dataset is constructed; Based on the physical residual loss function, the DGM network is trained using the training dataset. During the training process of the DGM network, the SARSA algorithm is used to iteratively optimize the network parameters of the DGM network to obtain the three-dimensional eddy current field recognition network.

[0008] As a preferred embodiment, the DGM network is trained using the training dataset based on the physical residual loss function, and the network parameters of the DGM network are iteratively optimized using the SARSA algorithm during the training process to obtain the three-dimensional eddy current field recognition network. Specifically, this includes: Based on the physical residual loss function, the DGM network is trained using the training dataset to obtain the current network output error; Based on the current network parameters of the DGM network and the network output error, determine the current state space; Based on the current state space, a greedy strategy is used to select actions, determine the current action space, and update the current network parameters of the DGM network; wherein, the action space includes at least one network parameter update action corresponding to the current state space of the DGM network. The updated DGM network is retrained using the training dataset to obtain an updated state space and reward value; wherein the reward value is determined based on the network output error and the total number of network parameters of the updated DGM network. Based on the updated state space, a greedy strategy is used to select actions and determine the updated action space; Based on the current state space, the current action space, the updated state space, the updated action space, the reward value, the preset learning rate, and the discount factor, the current Q-value table is updated; wherein, each Q-value in the Q-value table is used to represent the cumulative reward of different action spaces for different state spaces; Using the updated state space as the current state space and the updated action space as the current action space, the step of updating the current network parameters of the DGM network is re-executed until the Q-value table converges, thereby obtaining the three-dimensional eddy current field recognition network.

[0009] As a preferred embodiment, the step of selecting actions based on the current state space using a greedy strategy to determine the current action space specifically includes: Based on the current state space, determine the set of candidate actions corresponding to the current state space from the Q-value table; Based on random exploration probabilities, actions are selected from the set of candidate actions to determine the action space corresponding to the DGM network at present.

[0010] As a preferred embodiment, the method specifically constructs the reduced-order mapping model through the following steps: Based on several key structural parameters corresponding to several typical operating conditions of converter transformers, the magnetic vector potential distribution data corresponding to each typical operating condition of converter transformer is calculated using the three-dimensional eddy current field identification network. Extract several high-dimensional features of the magnetic vector potential distribution corresponding to each typical operating condition of the converter transformer from the magnetic vector potential distribution data; The high-dimensional features of the magnetic vector potential distribution are reduced in dimension to obtain several low-dimensional features of the magnetic vector potential distribution corresponding to each typical operating condition of the converter transformer. Based on the low-dimensional features of several magnetic vector potential distributions corresponding to the typical operating conditions of each converter transformer, the reduced-order mapping model is constructed.

[0011] As a preferred embodiment, the construction of the reduced-order mapping model based on several low-dimensional features of magnetic vector potential distribution corresponding to the typical operating conditions of each converter transformer specifically includes: Based on several low-dimensional features of magnetic vector potential distribution corresponding to each typical operating condition of the converter transformer, several sample data pairs are constructed to represent the correspondence between several key structural operating condition parameters and several low-dimensional features of magnetic vector potential distribution corresponding to the typical operating conditions of the converter transformer. The reduced-order mapping model is obtained by training a mapping model based on several sample data pairs; wherein the mapping model is a multinomial regression model, a support vector regression model, a Gaussian process regression model, a feedforward neural network, or a radial basis function neural network.

[0012] As a preferred embodiment, the step of using a pre-trained three-dimensional eddy current field recognition network to perform three-dimensional eddy current field recognition on the low-dimensional features of the target magnetic vector potential distribution, and obtaining the three-dimensional eddy current field distribution information of the converter transformer to be identified, specifically includes: The low-dimensional features of the target magnetic vector potential distribution are restored to the high-dimensional features of the target magnetic vector potential distribution; The target magnetic potential distribution data of the converter transformer to be identified is obtained by reconstructing the high-dimensional features of the target magnetic potential distribution using the three-dimensional eddy current field recognition network. Based on the conductivity of the converter transformer to be identified and the target magnetic vector potential distribution data, the three-dimensional eddy current field distribution information of the converter transformer to be identified is determined.

[0013] As a preferred embodiment, the method further includes: Obtain new structural parameters that do not belong to the training dataset and their corresponding finite element simulation magnetic vector potential reference data; The finite element simulation magnetic vector potential reference features are extracted from the finite element simulation magnetic vector potential reference data and dimensionality reduction is performed to obtain the low-dimensional features of the finite element simulation magnetic vector potential reference. Based on the reduced-order mapping model, obtain the low-dimensional features of the magnetic vector potential distribution to be verified corresponding to the new structural operating parameters; The magnetic vector potential is reconstructed from the low-dimensional features of the magnetic vector potential distribution to be verified using the three-dimensional eddy current field recognition network to obtain the magnetic vector potential distribution data to be verified. The magnetic vector potential distribution data to be verified and the low-dimensional features of the magnetic vector potential distribution to be verified are verified using the finite element simulation magnetic vector potential reference data and the low-dimensional features of the magnetic vector potential distribution to be verified, respectively, and the verification results are obtained. Based on the verification results, the reduced-order mapping model or the three-dimensional eddy field identification network is updated.

[0014] As a preferred embodiment, updating the reduced-order mapping model or the three-dimensional eddy current field recognition network based on the verification results specifically includes: When the verification result shows that the low-dimensional characteristics of the magnetic vector potential distribution to be verified are inconsistent with the low-dimensional characteristics of the finite element simulation magnetic vector potential reference, supplementary sample data pairs are generated using the low-dimensional characteristics of the finite element simulation magnetic vector potential reference and the new structural operating parameters. The reduced-order mapping model is updated using the supplementary sample data; When the verification result shows that the low-dimensional features of the magnetic vector potential distribution to be verified are consistent with the low-dimensional features of the finite element simulation magnetic vector potential reference, and the error between the magnetic vector potential distribution data to be verified and the finite element simulation magnetic vector potential reference data is greater than a preset error threshold, fine-tuning data is generated using the finite element simulation magnetic vector potential reference data and the new structural operating parameters. The coding layer parameters and hidden layer parameters of the three-dimensional eddy current field recognition network are fixed, and the three-dimensional eddy current field recognition network is fine-tuned using the fine-tuning data to update the three-dimensional eddy current field recognition network.

[0015] A second aspect of this invention provides a three-dimensional eddy current field identification system for converter transformers, comprising: The target structural operating condition parameter acquisition module is used to acquire the target structural operating condition parameters of the converter transformer to be identified. The order reduction mapping module is used to obtain the low-dimensional features of the target magnetic vector potential distribution corresponding to the target structural operating parameters according to a preset order reduction mapping model; wherein, the order reduction mapping model is used to characterize the mapping relationship between the structural operating parameters and the low-dimensional features of the magnetic vector potential distribution. The three-dimensional eddy current field identification module is used to identify the three-dimensional eddy current field of the target magnetic vector potential distribution using a pre-trained three-dimensional eddy current field identification network, thereby obtaining the three-dimensional eddy current field distribution information of the converter transformer to be identified; wherein, the three-dimensional eddy current field identification network is obtained by training the DGM network using three-dimensional eddy current field finite element simulation data of the converter transformer under different structural operating parameters.

[0016] Compared with the prior art, the beneficial effect of the embodiments of the present invention is that by using a reduced-order mapping model to characterize the mapping relationship between structural operating parameters and magnetic vector potential distribution characteristics, it is possible to quickly obtain the low-dimensional characteristics of the target magnetic vector potential distribution corresponding to the target structural operating parameters of the converter transformer to be identified, and use a three-dimensional eddy current field identification network to quickly complete the identification of the three-dimensional eddy current field without having to perform a computationally intensive finite element simulation solution process, which significantly improves the efficiency of three-dimensional eddy current field identification of converter transformers. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the three-dimensional eddy current field identification method for converter transformers in an embodiment of the present invention. Figure 2 This is a schematic diagram of the converter transformer simulation model in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the three-dimensional eddy current field identification system for the converter transformer in this embodiment of the invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 The first aspect of this invention provides a method for identifying a three-dimensional eddy current field in a converter transformer, comprising the following steps S1 to S3: Step S1: Obtain the target structural operating parameters of the converter transformer to be identified; Step S2: According to the preset reduced-order mapping model, obtain the low-dimensional features of the target magnetic vector potential distribution corresponding to the target structural operating parameters; wherein, the reduced-order mapping model is used to characterize the mapping relationship between the structural operating parameters and the low-dimensional features of the magnetic vector potential distribution. Step S3: Use a pre-trained three-dimensional eddy current field identification network to perform three-dimensional eddy current field identification on the low-dimensional features of the target magnetic vector potential distribution, and obtain the three-dimensional eddy current field distribution information of the converter transformer to be identified; wherein, the three-dimensional eddy current field identification network is obtained by training the DGM network with three-dimensional eddy current field finite element simulation data of the converter transformer under different structural operating parameters.

[0020] Specifically, since the magnetic vector potential and eddy current density may differ at different locations within the converter transformer, and these values ​​are affected by structural and operating parameters, this embodiment first obtains the target structural and operating parameters of the converter transformer to be identified. It can be understood that these target structural and operating parameters include the spatial coordinates to be identified, such as the coordinates of a certain point at the upper and lower clamps of the converter transformer to be identified, or the coordinates of a certain point at the core column, etc. They also include structural parameters, such as winding configuration, core thickness, window height, and housing dimensions, as well as operating parameters, such as operating voltage, operating current, operating frequency, and conductivity. This embodiment does not impose specific limitations on these parameters.

[0021] Furthermore, this embodiment pre-constructs a reduced-order mapping model to characterize the mapping relationship between structural operating parameters and low-dimensional features of the magnetic vector potential distribution. Through this reduced-order mapping model, the low-dimensional features of the target magnetic vector potential distribution corresponding to the target structural operating parameters can be obtained directly and quickly. It can be understood that the magnetic vector potential distribution features are statistically significant indicators that represent the magnetic vector potential distribution, including but not limited to the peak position, extreme value amplitude, mean square magnetic vector potential intensity of the magnetic vector potential component at a specific spatiotemporal location, energy distribution in different spatial regions, and field gradient variation trend on the equipotential surface. The set of all magnetic vector potential distribution features constitutes the original high-dimensional solution. The low-dimensional features of the magnetic vector potential distribution are the low-dimensional spatial representation of the above-mentioned magnetic vector potential distribution features, so as to transform the three-dimensional eddy current field problem in the converter transformer into a lower-dimensional computational problem, realizing a rapid connection between structural operating parameters and magnetic vector potential distribution features.

[0022] Furthermore, since the acquired low-dimensional features of the target magnetic vector potential distribution directly reflect the various characteristics of the magnetic vector potential distribution, by restoring these low-dimensional features to the original high-dimensional features, these high-dimensional features can be directly used to equivalently draw three-dimensional magnetic vector potential distribution maps, calculate isosurfaces, eddy current density distributions, and other engineering-related physical quantities. Since the high-dimensional features of the magnetic vector potential distribution are extracted from the magnetic vector potential distribution data output by the three-dimensional eddy current field recognition network during training, the three-dimensional eddy current field recognition network can quickly reconstruct the magnetic vector potential distribution based on these high-dimensional features, and rapidly calculate the corresponding three-dimensional eddy current field distribution information after clarifying the magnetic vector potential distribution data. In this embodiment, the three-dimensional eddy current field identification network is obtained by training a DGM (Deep Generative Models) network using three-dimensional eddy current field finite element simulation data of the converter transformer under different structural operating parameters. It can be understood that the three-dimensional eddy current field finite element simulation data characterizes the magnetic vector potential distribution finite element simulation data and the three-dimensional eddy current field distribution finite element simulation data of the converter transformer under different structural operating parameters and spatial coordinates. Therefore, it can be used as training data for supervised training of the DGM network. Furthermore, to ensure the accuracy of network training, the structure of the converter transformer corresponding to the selected three-dimensional eddy current field distribution finite element simulation data can be similar to or the same as the structure of the converter transformer to be identified. Alternatively, the three-dimensional eddy current field distribution finite element simulation data corresponding to the structural operating parameters under the historical operating conditions of the converter transformer to be identified can be directly used for DGM network training. This embodiment does not impose specific limitations here.

[0023] It is worth noting that before conducting finite element simulation, it is necessary to obtain the key structural and electrical parameters of the converter transformer to ensure the accuracy and reliability of the simulation results. Material parameters mainly include the magnetization characteristics of the core, which describe the core saturation and hysteresis characteristics. The electrical parameters of the converter transformer include operating voltage, current, and operating frequency. Simultaneously, the geometric dimensions of the converter transformer simulation model need to be obtained, including core thickness, window height, number of winding turns, and tank dimensions, to ensure that the simulation model can realistically reflect the structure of the converter transformer. To accurately simulate the actual operating state of the converter transformer, external circuit excitation needs to be set in the simulation model according to actual needs, connecting voltage or current sources. Considering the harmonic effects, to ensure the accuracy and efficiency of the eddy current field simulation calculation of the converter transformer, it is necessary to reasonably select the solver type, mesh generation method, and boundary condition settings: a transient solver (time-domain analysis) is used, and the mesh of the converter transformer components is refined during mesh generation.

[0024] The three-dimensional eddy current field identification method for converter transformers provided in this invention utilizes a reduced-order mapping model to characterize the mapping relationship between structural operating parameters and magnetic vector potential distribution characteristics. This enables the rapid acquisition of low-dimensional features of the target magnetic vector potential distribution corresponding to the target structural operating parameters of the converter transformer to be identified. Furthermore, the three-dimensional eddy current field identification is quickly completed using a three-dimensional eddy current field identification network, eliminating the need for computationally intensive finite element simulation solutions and significantly improving the efficiency of three-dimensional eddy current field identification for converter transformers.

[0025] As a preferred embodiment, the method specifically trains the three-dimensional eddy current field recognition network through the following steps: Construct the physical residual loss function of the DGM network; wherein the physical residual loss function is a weighted sum of the control equation residual, the control boundary condition residual, and the initial magnetic vector potential condition residual; the control equation residual is determined based on Maxwell's equations, the control boundary condition residual is determined based on the output magnetic vector potential of the DGM network and the preset magnetic vector potential boundary conditions, and the initial magnetic vector potential condition residual is determined based on the preset initial magnetic vector potential distribution data and the output magnetic vector potential of the DGM network at the initial time; Based on the aforementioned three-dimensional vortex field finite element simulation data, a training dataset is constructed; Based on the physical residual loss function, the DGM network is trained using the training dataset. During the training process of the DGM network, the SARSA algorithm is used to iteratively optimize the network parameters of the DGM network to obtain the three-dimensional eddy current field recognition network.

[0026] Specifically, the DGM network is a meshless solver that does not require mesh generation in the FEM (Finite Element Method). In this embodiment, the DGM network constructs a multi-layer neural network and directly optimizes the Maxwell equation residuals as part of the physical residual loss function. Therefore, it does not require the generation of discrete elements (such as tetrahedrons or hexahedrons), the use of weak integral formulas, or the assembly and solution of stiffness matrices. Thus, it does not require the mesh refinement of high-gradient regions such as clamps as in the FEM method.

[0027] Furthermore, this embodiment employs a DGM network to solve the eddy current field control equations through a deep neural network, thereby eliminating the efficiency bottleneck caused by the mesh dependency in traditional finite element method calculations. The DGM network, as the main electromagnetic field numerical solver in this embodiment, replaces all the steps of mesh generation, weak integration, and boundary treatment in the traditional finite element method for three-dimensional eddy current field problems with a neural network-driven end-to-end function approximation process. It is worth noting that the DGM network includes an input layer, a pre-encoding layer, a hidden layer, and an output layer. The input layer receives spatial coordinates (x, y, z) and time t; the hidden layer adopts a fully connected layer structure, and the activation functions used in the hidden layer include tanh, ReLU, and swish types; the output layer outputs the three directional components of the magnetic vector potential corresponding to the spatial coordinates. Then, the physical residual loss function of the DGM network is constructed. The physical residual loss function consists of three parts: the control equation residual, the control boundary condition residual, and the initial magnetic vector potential condition residual, expressed as: , This represents the physical residual loss value. Represents the residuals of the governing equations. Represents the residuals of the control boundary conditions. This represents the residual of the initial magnetic vector potential condition. The weights represent the residuals of the governing equations. This represents the weight of the residuals controlling the boundary conditions. This represents the weight of the initial magnetic vector potential conditional residual.

[0028] The residuals of the governing equations are determined based on Maxwell's equations. Specifically, the three-dimensional eddy current field distribution is controlled by Maxwell's equations, and in this embodiment, it is expressed in vector potential form: ; in, For curl operator, For the magnetic loss to be determined, Permeability, Resistivity For electrical conductivity, To obtain the magnetic vector potential distribution data, a DGM network is used to solve Maxwell's equations to obtain the excitation current density. The residuals of the governing equations are then expressed as follows: ; Based on the residual of this governing equation, it can be ensured that the magnetic vector potential distribution data output by the DGM network during training satisfies Maxwell's equations, thus ensuring the accuracy of the output results.

[0029] Furthermore, the control boundary condition residual is determined based on the output magnetic vector potential of the DGM network and the preset magnetic vector potential boundary conditions, thereby ensuring that the magnetic vector potential distribution data output by the DGM network during training is within a reasonable boundary range.

[0030] Furthermore, the initial magnetic vector potential condition residual is determined based on the preset initial magnetic vector potential distribution data and the output magnetic vector potential of the DGM network at the initial time. This residual term is used to limit the output magnetic vector potential of the DGM network at t=0, that is, the output magnetic vector potential at t=0 needs to be equal to the preset initial magnetic vector potential distribution data.

[0031] For the aforementioned physical residual loss function, this embodiment calculates the gradient using automatic differentiation technology to achieve physical consistency constraints during training. This ensures that the output of the DGM network satisfies the governing equations, control boundary conditions, and initial magnetic vector potential conditions, thereby achieving an approximate solution to high-dimensional complex electromagnetic field problems. The advantage of meshless computation lies in avoiding traditional meshing errors and distortions, allowing sampling at arbitrary points, handling arbitrary boundary shapes, and reducing preprocessing time and memory usage.

[0032] Furthermore, this embodiment constructs a training dataset based on three-dimensional vortex field finite element simulation data. It can be understood that this embodiment can set several training points {(xi, yi, zi, ti)} based on the three-dimensional vortex field finite element simulation data, where some training points can be labeled with specific three-dimensional vortex field finite element simulation data to improve the training efficiency of the DGM network.

[0033] Furthermore, this embodiment introduces the SARSA algorithm to iteratively optimize the network parameters of the DGM network. The SARSA algorithm treats each adjustment of the grid parameters of the DGM network as an action, the current network parameters and network output error as states, and the decrease in network output error and network complexity as reward signals, so as to minimize model complexity and computation time while maintaining computational accuracy.

[0034] As a preferred embodiment, the DGM network is trained using the training dataset based on the physical residual loss function, and the network parameters of the DGM network are iteratively optimized using the SARSA algorithm during the training process to obtain the three-dimensional eddy current field recognition network. Specifically, this includes: Based on the physical residual loss function, the DGM network is trained using the training dataset to obtain the current network output error; Based on the current network parameters of the DGM network and the network output error, determine the current state space; Based on the current state space, a greedy strategy is used to select actions, determine the current action space, and update the current network parameters of the DGM network; wherein, the action space includes at least one network parameter update action corresponding to the current state space of the DGM network. The updated DGM network is retrained using the training dataset to obtain an updated state space and reward value; wherein the reward value is determined based on the network output error and the total number of network parameters of the updated DGM network. Based on the updated state space, a greedy strategy is used to select actions and determine the updated action space; Based on the current state space, the current action space, the updated state space, the updated action space, the reward value, the preset learning rate, and the discount factor, the current Q-value table is updated; wherein, each Q-value in the Q-value table is used to represent the cumulative reward of different action spaces for different state spaces; Using the updated state space as the current state space and the updated action space as the current action space, the step of updating the current network parameters of the DGM network is re-executed until the Q-value table converges, thereby obtaining the three-dimensional eddy current field recognition network.

[0035] Specifically, in this embodiment, when training the DGM network using the training dataset, the output magnetic vector potential of the DGM network in the current network parameter optimization process is obtained. This output magnetic vector potential is the output result of the DGM network under the current network parameters. By analyzing the output result, it is possible to determine whether the current network parameters are optimal.

[0036] Furthermore, based on the physical residual loss function and the output magnetic vector potential, the network output error of the DGM network during the current network parameter optimization process can be directly calculated. Then, the current state space of the DGM network can be determined, including the current network parameters, network output error, and network complexity. The network parameters include the number of layers, the number of nodes, and the activation function type. The number of nodes refers to the number of neurons in each hidden layer, and this parameter is closely related to network capacity, approximation accuracy, and computational resource consumption. Within the reinforcement learning framework, by using the number of nodes as an adjustable variable and combining it with a reward function to balance the evaluation of simulation error and network complexity, dynamic optimization and adjustment of the DGM network structure can be achieved.

[0037] Furthermore, based on the current state space, a greedy strategy is used to select actions, determining the action space corresponding to the DGM network. It's worth noting that network parameter update actions include adding or removing layers, adding or removing nodes, and modifying the activation function type. Then, the determined action space is executed to update the current network parameters of the DGM network. At this point, the current network parameters need to be re-evaluated; therefore, the DGM network with updated parameters is retrained using the training dataset. The updated DGM network will then output a magnetic vector potential under the updated network parameters. Based on this output magnetic vector potential, the network output error under the updated network parameters can be determined using the physical residual loss function, thus obtaining the updated state space. Furthermore, in this embodiment, the reward function of the SARSA algorithm is the difference between the network output error reduction rate and the network complexity reduction rate. It can be understood that the network output error reduction rate is the rate at which the network output error under the updated network parameters decreases compared to the network output error before the parameter update. Network complexity is the result of evaluating the total number of network parameters in the DGM network after the parameter update; the total number of network parameters is the sum of the number of layers and the number of nodes. The reward function is shown in the following expression: ; in, As a reward value, The network output error under the updated network parameters. This represents the network output error before the network parameters were updated. This represents the total number of updated network parameters. This represents the total number of network parameters before the update. This is a preset network complexity penalty coefficient.

[0038] Furthermore, based on the current updated state space, the greedy strategy is re-adopted for action selection to determine the updated action space corresponding to the DGM network after the network parameters are updated, and the current Q-value table is updated based on the following expression: ; in, Let be the state space at time t. Let t be the action space at the current time. Let t+1 be the update state space. For the update action space at time t+1, Let Q be the Q-value corresponding to the state space and action space at time t. Let Q be the value corresponding to the update state space and update action space at time t+1. The preset learning rate, This is a preset discount factor. It's worth noting that the Q-value table stores the Q-values ​​corresponding to different state spaces and action spaces, with the initial value of the Q-value for each state space and action space set to 0. By continuously updating the network parameters of the DGM network until the Q-value table converges (i.e., the change in Q-value is less than a preset threshold), the strategy gradually approaches the optimal value, resulting in the final three-dimensional vortex field recognition network.

[0039] As a preferred embodiment, the step of selecting actions based on the current state space using a greedy strategy to determine the current action space specifically includes: Based on the current state space, determine the set of candidate actions corresponding to the current state space from the Q-value table; Based on random exploration probabilities, actions are selected from the set of candidate actions to determine the action space corresponding to the DGM network at present.

[0040] Specifically, the Q-value table stores several optional actions corresponding to each state, as well as the Q-value corresponding to each state and each optional action. Therefore, based on the current state space, this embodiment can directly determine the candidate action set from the Q-value table. Then, based on the current random exploration probability, it is worth noting that this random exploration probability can be a preset probability value, or a preset initial probability value that changes with the number of iterations of the SARSA algorithm. For example, a decay coefficient less than 1 can be set. After each iteration of the SARSA algorithm, the current random exploration probability is multiplied by the decay coefficient to obtain the random exploration probability corresponding to the next round of action selection. This embodiment does not make specific limitations here. Assuming the random exploration probability is represented by ε, and ε is less than 1, when selecting an action from the candidate action set, a random number is generated from the interval [0, 1]. If the random number is less than ε, an action is randomly selected from the candidate action set as the current action space; if the random number is greater than or equal to ε, the action corresponding to the largest Q-value in the candidate action set is selected as the current action space.

[0041] As a preferred embodiment, the method specifically constructs the reduced-order mapping model through the following steps: Based on several key structural parameters corresponding to several typical operating conditions of converter transformers, the magnetic vector potential distribution data corresponding to each typical operating condition of converter transformer is calculated using the three-dimensional eddy current field identification network. Extract several high-dimensional features of the magnetic vector potential distribution corresponding to each typical operating condition of the converter transformer from the magnetic vector potential distribution data; The high-dimensional features of the magnetic vector potential distribution are reduced in dimension to obtain several low-dimensional features of the magnetic vector potential distribution corresponding to each typical operating condition of the converter transformer. Based on the low-dimensional features of several magnetic vector potential distributions corresponding to the typical operating conditions of each converter transformer, the reduced-order mapping model is constructed.

[0042] Specifically, to obtain the mapping relationship between the structural operating parameters of the converter transformer and the low-dimensional characteristics of the magnetic vector potential distribution, this embodiment selects N typical operating conditions from the structural parameter space using Latin hypercube sampling or orthogonal experimental design. Each typical operating condition covers several key structural operating parameters, such as frequency, conductivity, core size, and voltage amplitude. This embodiment does not impose specific limitations on these parameters; it is understood that each typical operating condition is defined by a series of key parameters, including but not limited to: geometric parameters such as core thickness, number of winding turns, window height, and tank size; material parameters such as the magnetization characteristics of the core and the conductivity of metal components; and operating frequency, etc. Electrical parameters such as voltage amplitude, current magnitude, load conditions, harmonic interference, and DC bias magnetization result in different eddy current field distributions under different operating conditions. Therefore, it is necessary to obtain several key structural operating parameters corresponding to multiple different typical operating conditions. Then, a trained three-dimensional eddy current field identification network is used to calculate the magnetic vector potential distribution data corresponding to each converter transformer under typical operating conditions. It can be understood that the three-dimensional eddy current field identification network has already embedded Maxwell's control equations and related physical constraints during the training process. It can directly solve for the corresponding magnetic vector potential distribution data based on the permeability, reluctance, conductivity, and excitation current density in the key structural operating parameters.

[0043] Furthermore, several high-dimensional features of the magnetic vector potential distribution corresponding to the typical operating conditions of each converter transformer are extracted from the magnetic vector potential distribution data. The set of these high-dimensional features is the original high-dimensional solution corresponding to the typical operating conditions of the converter transformer. Then, the high-dimensional features of the magnetic vector potential distribution are dimensionality reduced using principal component analysis or stacked autoencoders. If principal component analysis is used for linear dimensionality reduction, the process is as follows: the original feature matrix composed of several high-dimensional features of the magnetic vector potential distribution is zero-mean normalized, and then the covariance matrix is ​​decomposed into eigenvalues. The first k principal component direction vectors are extracted, satisfying that the cumulative variance contribution rate exceeds a threshold (e.g., 95%). Then, the dimensionality-reduced feature representation is constructed, which is a low-dimensional feature space that retains the key magnetic vector potential distribution characteristics after dimensionality reduction, used for subsequent training and rapid prediction of the mapping model. If... Nonlinear dimensionality reduction is performed using stacked autoencoders, as follows: A symmetric stacked autoencoder model is constructed, comprising: an encoder part, which takes the original high-dimensional feature vector as input and compresses it into a low-dimensional latent representation through a multi-layer nonlinear neural network; and a decoder part, which reconstructs the original features from the low-dimensional latent variables. The stacked autoencoder model is trained using a mean squared error loss function with the goal of minimizing the reconstruction error. After training, the encoder output layer is extracted as a low-dimensional magnetic vector potential distribution feature representation, which is used to establish the mapping relationship between structural operating parameters and the low-dimensional features of the magnetic vector potential distribution.

[0044] Furthermore, after obtaining several low-dimensional features of the magnetic vector potential distribution corresponding to the typical operating conditions of each converter transformer, it is possible to construct a reduced-order mapping model based on interpolation or fitting methods. This model can then be used to quickly predict the mapping results under other operating conditions, thereby clarifying the mapping relationship between structural operating parameters and low-dimensional features of the magnetic vector potential distribution.

[0045] As a preferred embodiment, the construction of the reduced-order mapping model based on several low-dimensional features of magnetic vector potential distribution corresponding to the typical operating conditions of each converter transformer specifically includes: Based on several low-dimensional features of magnetic vector potential distribution corresponding to each typical operating condition of the converter transformer, several sample data pairs are constructed to represent the correspondence between several key structural operating condition parameters and several low-dimensional features of magnetic vector potential distribution corresponding to the typical operating conditions of the converter transformer. The reduced-order mapping model is obtained by training a mapping model based on several sample data pairs; wherein the mapping model is a multinomial regression model, a support vector regression model, a Gaussian process regression model, a feedforward neural network, or a radial basis function neural network.

[0046] Specifically, this embodiment first combines several key structural operating parameters corresponding to the typical operating conditions of each converter transformer with several low-dimensional features of the magnetic vector potential distribution to form several sample data pairs. The key structural operating parameters and low-dimensional features of the magnetic vector potential distribution within these sample data pairs represent data with a mapping relationship. Then, a mapping model is constructed using models such as multinomial regression, support vector regression, Gaussian process regression, feedforward neural networks, or radial basis function neural networks. This model is then trained using the aforementioned sample data pairs. Optionally, cross-validation, mean squared error, and coefficient of determination R0 can be used. 2 The generalization performance of the mapping function is evaluated using various metrics, and the final reduced-order mapping model is determined under the premise that the error range is controllable.

[0047] As a preferred embodiment, the step of using a pre-trained three-dimensional eddy current field recognition network to perform three-dimensional eddy current field recognition on the low-dimensional features of the target magnetic vector potential distribution, and obtaining the three-dimensional eddy current field distribution information of the converter transformer to be identified, specifically includes: The low-dimensional features of the target magnetic vector potential distribution are restored to the high-dimensional features of the target magnetic vector potential distribution; The target magnetic potential distribution data of the converter transformer to be identified is obtained by reconstructing the high-dimensional features of the target magnetic potential distribution using the three-dimensional eddy current field recognition network. Based on the conductivity of the converter transformer to be identified and the target magnetic vector potential distribution data, the three-dimensional eddy current field distribution information of the converter transformer to be identified is determined.

[0048] Specifically, in order to reconstruct and obtain the target magnetic vector potential distribution data of the converter transformer to be identified, this embodiment first needs to restore the low-dimensional features of the target magnetic vector potential distribution to the high-dimensional features of the target magnetic vector potential distribution. It is worth noting that for the low-dimensional features of the target magnetic vector potential distribution that are linearly reduced using principal component analysis, the high-dimensional features of the target magnetic vector potential distribution are restored by linear combination of principal components; for the low-dimensional features of the target magnetic vector potential distribution that are nonlinearly reduced using stacked autoencoders, the decoder is used to nonlinearly map the low-dimensional features of the target magnetic vector potential distribution back to the high-dimensional features of the target magnetic vector potential distribution.

[0049] After reconstructing the target magnetic vector potential distribution data using a three-dimensional eddy current field identification network, the three-dimensional eddy current density data of the converter transformer to be identified can be directly calculated based on the conductivity of the converter transformer to be identified and the target magnetic vector potential distribution data, thus forming three-dimensional eddy current field distribution information.

[0050] As a preferred embodiment, the method further includes: Obtain new structural parameters that do not belong to the training dataset and their corresponding finite element simulation magnetic vector potential reference data; The finite element simulation magnetic vector potential reference features are extracted from the finite element simulation magnetic vector potential reference data and dimensionality reduction is performed to obtain the low-dimensional features of the finite element simulation magnetic vector potential reference. Based on the reduced-order mapping model, obtain the low-dimensional features of the magnetic vector potential distribution to be verified corresponding to the new structural operating parameters; The magnetic vector potential is reconstructed from the low-dimensional features of the magnetic vector potential distribution to be verified using the three-dimensional eddy current field recognition network to obtain the magnetic vector potential distribution data to be verified. The magnetic vector potential distribution data to be verified and the low-dimensional features of the magnetic vector potential distribution to be verified are verified using the finite element simulation magnetic vector potential reference data and the low-dimensional features of the magnetic vector potential distribution to be verified, respectively, and the verification results are obtained. Based on the verification results, the reduced-order mapping model or the three-dimensional eddy field identification network is updated.

[0051] Specifically, for new structural parameters not involved in training, this embodiment uses high-precision finite element simulation data to verify the output results of the three-dimensional eddy current field identification network and the reduced-order mapping model. This allows for assessment of the generalization performance of the three-dimensional eddy current field identification network and the reduced-order mapping model for the new structural parameters. This embodiment uses the reduced-order mapping model to obtain the low-dimensional features of the magnetic vector potential distribution to be verified corresponding to the new structural parameters. Then, the three-dimensional eddy current field identification network is used to reconstruct the magnetic vector potential from these low-dimensional features, obtaining the magnetic vector potential distribution data to be verified.

[0052] Furthermore, the low-dimensional features of the magnetic vector potential reference from the finite element simulation are compared with the low-dimensional features of the magnetic vector potential distribution to be verified to determine whether they are consistent. If they are inconsistent, it indicates that there is a deviation in the output of the reduced-order mapping model. Then, the reference data of the magnetic vector potential from the finite element simulation are compared with the data of the magnetic vector potential distribution to be verified to determine whether the error between them is within a reasonable range. If not, it indicates that the three-dimensional eddy current field identification network has poor identification performance for new structural working parameters and needs to be updated.

[0053] It is worth noting that this embodiment calculates the error between the finite element simulation magnetic vector potential reference data and the magnetic vector potential distribution data to be verified using the following expression: ; in, To account for the error between the reference data of the magnetic vector potential in the finite element simulation and the magnetic vector potential distribution data to be verified, For the magnetic vector potential distribution data to be verified, This is the reference data for the magnetic vector potential in finite element simulation.

[0054] As a preferred embodiment, updating the reduced-order mapping model or the three-dimensional eddy current field recognition network based on the verification results specifically includes: When the verification result shows that the low-dimensional characteristics of the magnetic vector potential distribution to be verified are inconsistent with the low-dimensional characteristics of the finite element simulation magnetic vector potential reference, supplementary sample data pairs are generated using the low-dimensional characteristics of the finite element simulation magnetic vector potential reference and the new structural operating parameters. The reduced-order mapping model is updated using the supplementary sample data; When the verification result shows that the low-dimensional features of the magnetic vector potential distribution to be verified are consistent with the low-dimensional features of the finite element simulation magnetic vector potential reference, and the error between the magnetic vector potential distribution data to be verified and the finite element simulation magnetic vector potential reference data is greater than a preset error threshold, fine-tuning data is generated using the finite element simulation magnetic vector potential reference data and the new structural operating parameters. The coding layer parameters and hidden layer parameters of the three-dimensional eddy current field recognition network are fixed, and the three-dimensional eddy current field recognition network is fine-tuned using the fine-tuning data to update the three-dimensional eddy current field recognition network.

[0055] Specifically, when the verification result shows that the low-dimensional features of the magnetic vector potential distribution to be verified are inconsistent with the low-dimensional features of the magnetic vector potential reference in the finite element simulation, it indicates that the error mainly comes from the fitting deviation between the reduced-order mapping model and the low-dimensional features of the magnetic vector potential distribution for the structural operating parameters. Therefore, supplementary sample data pairs can be generated using the low-dimensional features of the magnetic vector potential reference in the finite element simulation and the new structural operating parameters. These supplementary sample data pairs can be added to the original sample data pairs, and the reduced-order mapping model can be retrained to achieve the purpose of updating. Alternatively, the weights of the reduced-order mapping model can be updated through online incremental training without reconstructing the entire reduced-order mapping model.

[0056] Furthermore, when the verification result shows that the low-dimensional features of the magnetic vector potential distribution to be verified are consistent with the low-dimensional features of the finite element simulation magnetic vector potential reference, but the error between the magnetic vector potential distribution data to be verified and the finite element simulation magnetic vector potential reference data is greater than the preset error threshold, it indicates that the three-dimensional eddy current field recognition network has poor recognition performance for new structural parameters and needs to be updated. At this time, the coding layer parameters and hidden layer parameters of the three-dimensional eddy current field recognition network are fixed, and only the parameters of its output layer or the last few layers near the output layer are updated. The loss function still adopts the physical residual loss function mentioned above. Then, fine-tuning is performed using fine-tuning data composed of finite element simulation magnetic vector potential reference data and new structural parameters. A small learning rate (e.g., ...) is used during the fine-tuning process. To prevent disruption of the original parameter distribution.

[0057] To fully illustrate the beneficial effects of the three-dimensional eddy current field identification method for converter transformers provided in the embodiments of the present invention, the embodiments of the present invention will be compared and analyzed with the traditional finite element simulation method below.

[0058] like Figure 2 The figure shows a converter transformer simulation model, in which eight spatial points to be identified are set: points 1 and 2 located in the upper clamping member, points 3 and 4 located in the lower clamping member, and points 5, 6, 7, and 8 located at the core column. Based on the structural operating parameters of these eight points, the eddy current density is identified using both the three-dimensional eddy current field identification method of this invention and the traditional finite element simulation method. The identification results are shown in Table 1 below.

[0059] Table 1 Comparison of Eddy Current Density Identification Results Therefore, it can be seen that the eddy density identification results obtained by the three-dimensional eddy field identification method of the present invention are close to those obtained by the finite element simulation method. That is, the three-dimensional eddy field identification method of the present invention has high identification accuracy. However, compared with the finite element simulation method, the present invention does not require the generation of discrete elements, the use of weak integral formulas, or the assembly and solution of stiffness matrices. Thus, it does not require the densification of the mesh in high gradient regions such as clamps as in the finite element simulation method, which significantly improves the identification efficiency of the three-dimensional eddy field.

[0060] Please see Figure 3 The second aspect of this invention provides a three-dimensional eddy current field identification system for converter transformers, comprising: The target structural operating condition parameter acquisition module 101 is used to acquire the target structural operating condition parameters of the converter transformer to be identified. The order reduction mapping module 102 is used to obtain the low-dimensional features of the target magnetic vector potential distribution corresponding to the target structural operating parameters according to a preset order reduction mapping model; wherein, the order reduction mapping model is used to characterize the mapping relationship between the structural operating parameters and the low-dimensional features of the magnetic vector potential distribution. The three-dimensional eddy current field identification module 103 is used to identify the three-dimensional eddy current field of the target magnetic vector potential distribution low-dimensional features using a pre-trained three-dimensional eddy current field identification network, and obtain the three-dimensional eddy current field distribution information of the converter transformer to be identified; wherein, the three-dimensional eddy current field identification network is obtained by training the DGM network using three-dimensional eddy current field finite element simulation data of the converter transformer under different structural operating parameters.

[0061] As a preferred embodiment, the system further includes a three-dimensional eddy current field recognition network training module, used for: Construct the physical residual loss function of the DGM network; wherein the physical residual loss function is a weighted sum of the control equation residual, the control boundary condition residual, and the initial magnetic vector potential condition residual; the control equation residual is determined based on Maxwell's equations, the control boundary condition residual is determined based on the output magnetic vector potential of the DGM network and the preset magnetic vector potential boundary conditions, and the initial magnetic vector potential condition residual is determined based on the preset initial magnetic vector potential distribution data and the output magnetic vector potential of the DGM network at the initial time; Based on the aforementioned three-dimensional vortex field finite element simulation data, a training dataset is constructed; Based on the physical residual loss function, the DGM network is trained using the training dataset. During the training process of the DGM network, the SARSA algorithm is used to iteratively optimize the network parameters of the DGM network to obtain the three-dimensional eddy current field recognition network.

[0062] As a preferred embodiment, the three-dimensional eddy current field recognition network training module is used to train the DGM network based on the physical residual loss function and the training dataset. During the training process, the SARSA algorithm is used to iteratively optimize the network parameters of the DGM network to obtain the three-dimensional eddy current field recognition network. Specifically, this includes: Based on the physical residual loss function, the DGM network is trained using the training dataset to obtain the current network output error; Based on the current network parameters of the DGM network and the network output error, determine the current state space; Based on the current state space, a greedy strategy is used to select actions, determine the current action space, and update the current network parameters of the DGM network; wherein, the action space includes at least one network parameter update action corresponding to the current state space of the DGM network. The updated DGM network is retrained using the training dataset to obtain an updated state space and reward value; wherein the reward value is determined based on the network output error and the total number of network parameters of the updated DGM network. Based on the updated state space, a greedy strategy is used to select actions and determine the updated action space; Based on the current state space, the current action space, the updated state space, the updated action space, the reward value, the preset learning rate, and the discount factor, the current Q-value table is updated; wherein, each Q-value in the Q-value table is used to represent the cumulative reward of different action spaces for different state spaces; Using the updated state space as the current state space and the updated action space as the current action space, the step of updating the current network parameters of the DGM network is re-executed until the Q-value table converges, thereby obtaining the three-dimensional eddy current field recognition network.

[0063] As a preferred embodiment, the three-dimensional eddy current field recognition network training module is used to select actions based on the current state space using a greedy strategy to determine the current action space, specifically including: Based on the current state space, determine the set of candidate actions corresponding to the current state space from the Q-value table; Based on random exploration probabilities, actions are selected from the set of candidate actions to determine the action space corresponding to the DGM network at present.

[0064] As a preferred embodiment, the system further includes a reduced-order mapping model construction module, used for: Based on several key structural parameters corresponding to several typical operating conditions of converter transformers, the magnetic vector potential distribution data corresponding to each typical operating condition of converter transformer is calculated using the three-dimensional eddy current field identification network. Extract several high-dimensional features of the magnetic vector potential distribution corresponding to each typical operating condition of the converter transformer from the magnetic vector potential distribution data; The high-dimensional features of the magnetic vector potential distribution are reduced in dimension to obtain several low-dimensional features of the magnetic vector potential distribution corresponding to each typical operating condition of the converter transformer. Based on the low-dimensional features of several magnetic vector potential distributions corresponding to the typical operating conditions of each converter transformer, the reduced-order mapping model is constructed.

[0065] As a preferred embodiment, the reduced-order mapping model construction module is used to construct the reduced-order mapping model based on several low-dimensional features of magnetic vector potential distribution corresponding to each typical operating condition of the converter transformer, specifically including: Based on several low-dimensional features of magnetic vector potential distribution corresponding to each typical operating condition of the converter transformer, several sample data pairs are constructed to represent the correspondence between several key structural operating condition parameters and several low-dimensional features of magnetic vector potential distribution corresponding to the typical operating conditions of the converter transformer. The reduced-order mapping model is obtained by training a mapping model based on several sample data pairs; wherein the mapping model is a multinomial regression model, a support vector regression model, a Gaussian process regression model, a feedforward neural network, or a radial basis function neural network.

[0066] As a preferred embodiment, the three-dimensional eddy current field identification module 103 is used to perform three-dimensional eddy current field identification on the low-dimensional features of the target magnetic vector potential distribution using a pre-trained three-dimensional eddy current field identification network, thereby obtaining the three-dimensional eddy current field distribution information of the converter transformer to be identified, specifically including: The low-dimensional features of the target magnetic vector potential distribution are restored to the high-dimensional features of the target magnetic vector potential distribution; The target magnetic potential distribution data of the converter transformer to be identified is obtained by reconstructing the high-dimensional features of the target magnetic potential distribution using the three-dimensional eddy current field recognition network. Based on the conductivity of the converter transformer to be identified and the target magnetic vector potential distribution data, the three-dimensional eddy current field distribution information of the converter transformer to be identified is determined.

[0067] As a preferred embodiment, the system further includes a model update module, used for: Obtain new structural parameters that do not belong to the training dataset and their corresponding finite element simulation magnetic vector potential reference data; The finite element simulation magnetic vector potential reference features are extracted from the finite element simulation magnetic vector potential reference data and dimensionality reduction is performed to obtain the low-dimensional features of the finite element simulation magnetic vector potential reference. Based on the reduced-order mapping model, obtain the low-dimensional features of the magnetic vector potential distribution to be verified corresponding to the new structural operating parameters; The magnetic vector potential is reconstructed from the low-dimensional features of the magnetic vector potential distribution to be verified using the three-dimensional eddy current field recognition network to obtain the magnetic vector potential distribution data to be verified. The magnetic vector potential distribution data to be verified and the low-dimensional features of the magnetic vector potential distribution to be verified are verified using the finite element simulation magnetic vector potential reference data and the low-dimensional features of the magnetic vector potential distribution to be verified, respectively, and the verification results are obtained. Based on the verification results, the reduced-order mapping model or the three-dimensional eddy field identification network is updated.

[0068] As a preferred embodiment, the model update module is used to update the reduced-order mapping model or the three-dimensional eddy current field recognition network based on the verification results, specifically including: When the verification result shows that the low-dimensional characteristics of the magnetic vector potential distribution to be verified are inconsistent with the low-dimensional characteristics of the finite element simulation magnetic vector potential reference, supplementary sample data pairs are generated using the low-dimensional characteristics of the finite element simulation magnetic vector potential reference and the new structural operating parameters. The reduced-order mapping model is updated using the supplementary sample data; When the verification result shows that the low-dimensional features of the magnetic vector potential distribution to be verified are consistent with the low-dimensional features of the finite element simulation magnetic vector potential reference, and the error between the magnetic vector potential distribution data to be verified and the finite element simulation magnetic vector potential reference data is greater than a preset error threshold, fine-tuning data is generated using the finite element simulation magnetic vector potential reference data and the new structural operating parameters. The coding layer parameters and hidden layer parameters of the three-dimensional eddy current field recognition network are fixed, and the three-dimensional eddy current field recognition network is fine-tuned using the fine-tuning data to update the three-dimensional eddy current field recognition network.

[0069] The three-dimensional eddy current field identification system for converter transformers provided in this invention utilizes a reduced-order mapping model to characterize the mapping relationship between structural operating parameters and magnetic vector potential distribution characteristics. This enables the rapid acquisition of low-dimensional features of the target magnetic vector potential distribution corresponding to the target structural operating parameters of the converter transformer to be identified. Furthermore, the system rapidly completes the identification of the three-dimensional eddy current field using a three-dimensional eddy current field identification network, eliminating the need for computationally intensive finite element simulation solutions and significantly improving the efficiency of three-dimensional eddy current field identification for converter transformers.

[0070] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for identifying the three-dimensional eddy current field of a converter transformer, characterized in that, include: Obtain the target structural operating parameters of the converter transformer to be identified; According to the preset reduced-order mapping model, the low-dimensional features of the target magnetic vector potential distribution corresponding to the target structural operating parameters are obtained; wherein, the reduced-order mapping model is used to characterize the mapping relationship between the structural operating parameters and the low-dimensional features of the magnetic vector potential distribution. A pre-trained three-dimensional eddy current field identification network is used to identify the low-dimensional features of the target magnetic vector potential distribution in three dimensions, thereby obtaining the three-dimensional eddy current field distribution information of the converter transformer to be identified. The three-dimensional eddy current field identification network is obtained by training the DGM network with three-dimensional eddy current field finite element simulation data of the converter transformer under different structural operating parameters.

2. The three-dimensional eddy current field identification method for converter transformers as described in claim 1, characterized in that, The method specifically trains the three-dimensional eddy field recognition network through the following steps: Construct the physical residual loss function of the DGM network; wherein the physical residual loss function is a weighted sum of the control equation residual, the control boundary condition residual, and the initial magnetic vector potential condition residual; the control equation residual is determined based on Maxwell's equations, the control boundary condition residual is determined based on the output magnetic vector potential of the DGM network and the preset magnetic vector potential boundary conditions, and the initial magnetic vector potential condition residual is determined based on the preset initial magnetic vector potential distribution data and the output magnetic vector potential of the DGM network at the initial time; Based on the aforementioned three-dimensional vortex field finite element simulation data, a training dataset is constructed; Based on the physical residual loss function, the DGM network is trained using the training dataset. During the training process of the DGM network, the SARSA algorithm is used to iteratively optimize the network parameters of the DGM network to obtain the three-dimensional eddy current field recognition network.

3. The three-dimensional eddy current field identification method for converter transformers as described in claim 2, characterized in that, The process of training the DGM network using the training dataset based on the physical residual loss function, and iteratively optimizing the network parameters of the DGM network using the SARSA algorithm during the training process to obtain the three-dimensional eddy current field recognition network, specifically includes: Based on the physical residual loss function, the DGM network is trained using the training dataset to obtain the current network output error; Based on the current network parameters of the DGM network and the network output error, determine the current state space; Based on the current state space, a greedy strategy is used to select actions, determine the current action space, and update the current network parameters of the DGM network; wherein, the action space includes at least one network parameter update action corresponding to the current state space of the DGM network. The updated DGM network is retrained using the training dataset to obtain an updated state space and reward value; wherein the reward value is determined based on the network output error and the total number of network parameters of the updated DGM network. Based on the updated state space, a greedy strategy is used to select actions and determine the updated action space; Based on the current state space, the current action space, the updated state space, the updated action space, the reward value, the preset learning rate, and the discount factor, the current Q-value table is updated; wherein, each Q-value in the Q-value table is used to represent the cumulative reward of different action spaces for different state spaces; Using the updated state space as the current state space and the updated action space as the current action space, the step of updating the current network parameters of the DGM network is re-executed until the Q-value table converges, thereby obtaining the three-dimensional eddy current field recognition network.

4. The three-dimensional eddy current field identification method for converter transformers as described in claim 3, characterized in that, The step of selecting actions based on the current state space using a greedy strategy to determine the current action space specifically includes: Based on the current state space, determine the set of candidate actions corresponding to the current state space from the Q-value table; Based on random exploration probabilities, actions are selected from the set of candidate actions to determine the action space corresponding to the DGM network at present.

5. The three-dimensional eddy current field identification method for converter transformers as described in claim 1, characterized in that, The method specifically constructs the reduced-order mapping model through the following steps: Based on several key structural parameters corresponding to several typical operating conditions of converter transformers, the magnetic vector potential distribution data corresponding to each typical operating condition of converter transformer is calculated using the three-dimensional eddy current field identification network. Extract several high-dimensional features of the magnetic vector potential distribution corresponding to each typical operating condition of the converter transformer from the magnetic vector potential distribution data; The high-dimensional features of the magnetic vector potential distribution are reduced in dimension to obtain several low-dimensional features of the magnetic vector potential distribution corresponding to each typical operating condition of the converter transformer. Based on the low-dimensional features of several magnetic vector potential distributions corresponding to the typical operating conditions of each converter transformer, the reduced-order mapping model is constructed.

6. The three-dimensional eddy current field identification method for converter transformers as described in claim 5, characterized in that, The reduced-order mapping model, constructed based on several low-dimensional features of magnetic vector potential distributions corresponding to typical operating conditions of each converter transformer, specifically includes: Based on several low-dimensional features of magnetic vector potential distribution corresponding to each typical operating condition of the converter transformer, several sample data pairs are constructed to represent the correspondence between several key structural operating condition parameters and several low-dimensional features of magnetic vector potential distribution corresponding to the typical operating conditions of the converter transformer. The reduced-order mapping model is obtained by training a mapping model based on several sample data pairs; wherein the mapping model is a multinomial regression model, a support vector regression model, a Gaussian process regression model, a feedforward neural network, or a radial basis function neural network.

7. The three-dimensional eddy current field identification method for converter transformers as described in claim 2, characterized in that, The step of using a pre-trained three-dimensional eddy current field recognition network to perform three-dimensional eddy current field recognition on the low-dimensional features of the target magnetic vector potential distribution, and obtaining the three-dimensional eddy current field distribution information of the converter transformer to be identified, specifically includes: The low-dimensional features of the target magnetic vector potential distribution are restored to the high-dimensional features of the target magnetic vector potential distribution; The target magnetic potential distribution data of the converter transformer to be identified is obtained by reconstructing the high-dimensional features of the target magnetic potential distribution using the three-dimensional eddy current field recognition network. Based on the conductivity of the converter transformer to be identified and the target magnetic vector potential distribution data, the three-dimensional eddy current field distribution information of the converter transformer to be identified is determined.

8. The three-dimensional eddy current field identification method for converter transformers as described in claim 7, characterized in that, The method further includes: Obtain new structural parameters that do not belong to the training dataset and their corresponding finite element simulation magnetic vector potential reference data; The finite element simulation magnetic vector potential reference features are extracted from the finite element simulation magnetic vector potential reference data and dimensionality reduction is performed to obtain the low-dimensional features of the finite element simulation magnetic vector potential reference. Based on the reduced-order mapping model, obtain the low-dimensional features of the magnetic vector potential distribution to be verified corresponding to the new structural operating parameters; The magnetic vector potential is reconstructed from the low-dimensional features of the magnetic vector potential distribution to be verified using the three-dimensional eddy current field recognition network to obtain the magnetic vector potential distribution data to be verified. The magnetic vector potential distribution data to be verified and the low-dimensional features of the magnetic vector potential distribution to be verified are verified using the finite element simulation magnetic vector potential reference data and the low-dimensional features of the magnetic vector potential distribution to be verified, respectively, and the verification results are obtained. Based on the verification results, the reduced-order mapping model or the three-dimensional eddy field identification network is updated.

9. The three-dimensional eddy current field identification method for converter transformers as described in claim 8, characterized in that, The step of updating the reduced-order mapping model or the three-dimensional eddy current field recognition network based on the verification results specifically includes: When the verification result shows that the low-dimensional characteristics of the magnetic vector potential distribution to be verified are inconsistent with the low-dimensional characteristics of the finite element simulation magnetic vector potential reference, supplementary sample data pairs are generated using the low-dimensional characteristics of the finite element simulation magnetic vector potential reference and the new structural operating parameters. The reduced-order mapping model is updated using the supplementary sample data; When the verification result shows that the low-dimensional features of the magnetic vector potential distribution to be verified are consistent with the low-dimensional features of the finite element simulation magnetic vector potential reference, and the error between the magnetic vector potential distribution data to be verified and the finite element simulation magnetic vector potential reference data is greater than a preset error threshold, fine-tuning data is generated using the finite element simulation magnetic vector potential reference data and the new structural operating parameters. The coding layer parameters and hidden layer parameters of the three-dimensional eddy current field recognition network are fixed, and the three-dimensional eddy current field recognition network is fine-tuned using the fine-tuning data to update the three-dimensional eddy current field recognition network.

10. A three-dimensional eddy current field identification system for a converter transformer, characterized in that, include: The target structural operating condition parameter acquisition module is used to acquire the target structural operating condition parameters of the converter transformer to be identified. The order reduction mapping module is used to obtain the low-dimensional features of the target magnetic vector potential distribution corresponding to the target structural operating parameters according to a preset order reduction mapping model; wherein, the order reduction mapping model is used to characterize the mapping relationship between the structural operating parameters and the low-dimensional features of the magnetic vector potential distribution. The three-dimensional eddy current field identification module is used to identify the three-dimensional eddy current field of the target magnetic vector potential distribution using a pre-trained three-dimensional eddy current field identification network, thereby obtaining the three-dimensional eddy current field distribution information of the converter transformer to be identified; wherein, the three-dimensional eddy current field identification network is obtained by training the DGM network using three-dimensional eddy current field finite element simulation data of the converter transformer under different structural operating parameters.

Images