A transformer direct current bias prediction method, device, equipment and medium

Abstract

The application discloses a transformer DC bias prediction method, device, equipment and medium. The method comprises the following steps: determining target training data according to a DC bias simulation result of a target transformer; the DC bias simulation result comprises multiple groups of electromagnetic field quantity simulation data under different DC bias excitation conditions; performing supervised training on a pre-constructed neural network model by using the target training data to obtain a DC bias prediction model; the neural network model is embedded with physical constraints, the physical constraints comprise a first constraint and a second constraint, the first constraint is described based on electromagnetic field quantity limitations corresponding to transformer material properties, and the second constraint is described based on physical loss in a loss function; and performing DC bias prediction on a to-be-tested transformer based on the DC bias prediction model. The scheme embeds the physical constraints in the training process of the neural network model, can reduce data dependence, ensures the rationality and generalization ability of model prediction, and realizes high-precision and high-efficiency DC bias prediction.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic simulation technology for power equipment, and in particular to a method, device, equipment, and medium for predicting DC bias magnetism in transformers. Background Technology

[0002] Throughout the entire lifecycle of a power transformer, DC bias is a significant electromagnetic interference factor threatening its safe and reliable operation. This phenomenon manifests as a DC magnetomotive force appearing in the transformer's magnetic circuit, inducing a corresponding steady-state magnetic flux, causing a shift in the core's operating point. Its causes mainly fall into three categories: quasi-DC intrusions caused by geomagnetic storms, DC components injected by high-voltage DC transmission operating in monopolar or asymmetrical bipolar modes, and DC disturbances generated by the non-ideal characteristics of power electronic equipment in the power grid. These DC disturbances force the core into the semi-saturation region, leading to a series of anomalies such as increased excitation current distortion, significantly enhanced mechanical vibration, and deteriorated operating noise, posing a substantial threat to the equipment and system operation.

[0003] Given the complexity and unpredictability of DC bias faults, relying solely on theoretical calculations or empirical judgments is insufficient to fully assess their impact on transformers. Traditional finite element analysis methods depend on iterative solutions using physical simulation models, which suffers from limitations such as high computational complexity and low computational efficiency. A single simulation can take several hours to several days, making real-time simulation impossible.

[0004] In related technologies, deep learning technology is used to replace traditional finite element analysis to achieve rapid simulation of the electromagnetic field of three-phase transformers, enabling predictions within seconds. However, purely data-driven deep learning models cannot guarantee the physical rationality of the solution and rely on a large amount of high-quality labeled data, resulting in low prediction reliability and limited generalization ability. Summary of the Invention

[0005] This invention provides a method, apparatus, device, and medium for predicting DC bias in transformers. By embedding physical constraints, which combine soft and hard constraints, into the training process of a neural network model, the invention reduces data dependence, ensures the physical rationality and generalization ability of the model's predictions, and thus achieves high-precision and high-efficiency DC bias prediction.

[0006] According to one aspect of the present invention, a method for predicting DC bias magnetism in a transformer is provided, the method comprising:

[0007] The target training data is determined based on the DC bias simulation results of the target transformer; wherein, the DC bias simulation results include multiple sets of electromagnetic field quantity simulation data under different DC bias excitation conditions;

[0008] A DC bias prediction model is obtained by supervising the training of a pre-constructed neural network model using the target training data. The neural network model incorporates physical constraints, including a first constraint and a second constraint. The first constraint is described based on electromagnetic field limitations corresponding to transformer material properties, while the second constraint is described based on physical losses in a loss function. The loss function is constructed based on a weighted sum of data loss and physical loss. The data loss is constructed based on the DC bias simulation results and the prediction results of the neural network model, while the physical loss is constructed based on Maxwell's equations, field-circuit coupling equations, and nonlinear magnetization characteristics.

[0009] DC bias prediction is performed on the transformer under test based on the DC bias prediction model.

[0010] According to another aspect of the present invention, a transformer DC bias prediction device is provided, the device comprising:

[0011] The training data determination module is used to determine the target training data based on the DC bias simulation results of the target transformer; wherein, the DC bias simulation results include multiple sets of electromagnetic field quantity simulation data under different DC bias excitation conditions;

[0012] The model training module is used to supervise the training of a pre-constructed neural network model using the target training data to obtain a DC bias prediction model. The neural network model embeds physical constraints, including a first constraint and a second constraint. The first constraint is described based on electromagnetic field limitations corresponding to transformer material properties, and the second constraint is described based on physical losses in a loss function. The loss function is constructed based on a weighted sum of data losses and physical losses. The data losses are constructed based on the DC bias simulation results and the prediction results of the neural network model, while the physical losses are constructed based on Maxwell's equations, field-circuit coupling equations, and nonlinear magnetization characteristics.

[0013] The model prediction module is used to predict the DC bias of the transformer under test based on the DC bias prediction model.

[0014] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising:

[0015] At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the transformer DC bias prediction method according to any embodiment of the present invention.

[0016] According to another aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to execute and implement the transformer DC bias prediction method according to any embodiment of the present invention.

[0017] The technical solution of this invention first determines target training data based on the DC bias simulation results of the target transformer. The DC bias simulation results include multiple sets of electromagnetic field simulation data under different DC bias excitation conditions. Then, the target training data is used to supervise the training of a pre-constructed neural network model to obtain a DC bias prediction model. The neural network model embeds physical constraints, including a first constraint and a second constraint. The first constraint is described based on electromagnetic field limitations corresponding to the transformer material properties, and the second constraint is described based on physical losses in a loss function. The loss function is constructed based on a weighted sum of data losses and physical losses. The data losses are constructed based on the DC bias simulation results and the prediction results of the neural network model, while the physical losses are constructed based on Maxwell's equations, field-circuit coupling equations, and nonlinear magnetization characteristics. Finally, the DC bias prediction model is used to predict the DC bias of the transformer under test. This technical solution embeds physical constraints, combining soft and hard constraints, into the training process of a neural network model. This allows the trained neural network model (i.e., the DC bias prediction model) to predict the DC bias of transformers. Compared to the traditional finite element analysis method, the neural network model replaces iterative solutions, achieving a second-level response and significantly improving prediction efficiency. Compared to the pure data-driven method, the embedded physical constraints reduce data dependence, ensure the physical rationality of the model's prediction, and improve the model's generalization ability, thereby achieving high-precision and high-efficiency DC bias prediction.

[0018] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of a transformer DC bias prediction method provided by an embodiment of the present invention;

[0021] Figure 2This is a flowchart of another method for predicting DC bias of a transformer according to an embodiment of the present invention;

[0022] Figure 3 This is a schematic diagram of the overall process for predicting DC bias of a transformer according to an embodiment of the present invention;

[0023] Figure 4 This is a schematic diagram of the structure of a transformer DC bias prediction device according to an embodiment of the present invention;

[0024] Figure 5 This is a schematic diagram of the structure of an electronic device for implementing a transformer DC bias prediction method according to an embodiment of the present invention. Detailed Implementation

[0025] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0026] It should be noted that the terms "first," "second," "target," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0027] Example 1

[0028] Figure 1 This is a flowchart of a transformer DC bias prediction method provided in Embodiment 1 of the present invention. This embodiment is applicable to situations where high-precision and high-efficiency prediction of transformer DC bias is required. The method can be executed by a transformer DC bias prediction device, which can be implemented in hardware and / or software. The transformer DC bias prediction device can be configured in an electronic device with data processing capabilities.

[0029] First, a brief introduction to the physical scenario of DC bias involved in this invention will be given. The simplified physical equations involved in the two-dimensional DC bias physical scenario include:

[0030] (1) Maxwell's equations - Current law: ;

[0031] (2) Maxwell's equations - Gauss's law for magnetic fields: ;

[0032] (3) Field-circuit coupling equations: .

[0033] in, Resistivity magnetic vector potential In the vertical direction The amount, Current density In the vertical direction The amount, and These represent the horizontal and vertical directions, respectively. Magnetic flux density The AC voltage is used to excite the ports of each phase winding. This represents the equivalent DC voltage drop across each phase winding. For each phase winding in Length in direction, The number of turns in each phase winding. This is the cross-sectional area of ​​the winding. For the winding area, This represents the corresponding DC resistance of the winding. For the winding region .

[0034] It is important to note the applicable scope of the above physical equations and the calculation methods for the physical quantities in different material properties. Equations (1) and (2) are applicable to the spatial regions corresponding to the three materials: air, winding, and iron core. The air region and the iron core region... And the winding area Equation (3) applies only to the winding region. The reluctance varies depending on the material region. There are also significant differences, specifically between the air region and the winding region. Given a fixed value; for the core material, the magnetic reluctance is... It is the magnitude of magnetic induction intensity A nonlinear function can be expressed as Among them, the magnetic reluctance of the core material It can be determined through the magnetization curve of the core material, where... In the physical equations (1)-(3), , , , , Since the magnetization curve data of the iron core is known, and For simulation input variables, , and It is an unknown physical quantity.

[0035] In summary, the two-dimensional DC bias simulation can be described as: excitation with a known voltage ( and ) and transformer-specific parameters ( , , , , Using the core magnetization curve data as input, and based on physical information such as the current law, Gauss's law for magnetic fields, and field-circuit coupling equations, the key electromagnetic field distributions at different times within the simulation period are solved. , and ).

[0036] like Figure 1 As shown, the method includes:

[0037] S110, determine the target training data based on the DC bias simulation results of the target transformer; wherein, the DC bias simulation results include multiple sets of electromagnetic field quantity simulation data under different DC bias excitation conditions.

[0038] The target transformer can refer to the transformer object participating in the model training, and can be one or more transformers. DC bias excitation can refer to voltage excitation under DC bias conditions. and Electromagnetic field quantities include , and The target training data can refer to the DC bias simulation data used in model training.

[0039] In this embodiment, numerical simulation methods (such as finite element analysis) can be used to simulate the DC bias magnetization of the target transformer, obtaining DC bias magnetization simulation results including multiple sets of electromagnetic field simulation data under different DC bias magnetization excitation conditions, and determining the target training data accordingly. It should be noted that one set of DC bias magnetization excitation conditions corresponds to multiple sets of electromagnetic field simulation data at different excitation times. The DC bias magnetization simulation results under the same set of DC bias magnetization excitation conditions are considered as one data sample. Ultimately, multiple data samples can be obtained as the dataset for model training, and the data in this dataset is the target training data.

[0040] In this embodiment, optionally, the target training data is determined based on the DC bias simulation results of the target transformer, including: constructing a target geometric model based on the geometric structure and material properties of the target transformer; performing DC bias simulation on the target geometric model using the finite element analysis method to obtain the DC bias simulation results of the target transformer; and performing data preprocessing on the DC bias simulation results to obtain the target training data.

[0041] Specifically, firstly, a geometric model suitable for finite element simulation is constructed based on the target transformer's geometry and material properties as the target geometric model. The geometry describes the geometric configuration and dimensions, while material properties include air, the core, and the windings. For example, taking a two-dimensional DC bias scenario, since the transformer's laminated core is symmetrical along two axes, a quarter-laminated core can be selected as the simulation object. A 2D finite element model matching the simulation object's geometry is constructed in COMSOL Multiphysics, and material parameters (including material properties and the core magnetization curve) are set, thus obtaining the target geometric model.

[0042] Then, the finite element method is used to simulate the DC bias of the constructed target geometric model under multiple sets of DC bias excitation conditions, thereby obtaining the DC bias simulation results of the target transformer. For example, 300 sets of DC bias excitation conditions can be simulated (…). and The output includes 100 mesh data points corresponding to each excitation condition, where the mesh data includes mesh spatial coordinates, mesh area (area), excitation time (t), material property type, and electromagnetic field simulation data. , and It should be noted that this embodiment does not specifically limit the shape of the grid; for example, it can be a rectangle, square, or triangle. Furthermore, the grid spatial coordinates... The average value of the coordinates of each vertex of the mesh can be used for characterization. For example, the DC bias simulation result of a certain mesh at a certain excitation moment can be expressed as: , t=0, area=0.036, type=air, , , , , Wherein, "air" represents an area of ​​air. This represents the vacuum permeability. It should be noted that, due to... , , , , It has a wide range of applications, including data such as magnetization curves, and can be built into the training code.

[0043] To ensure the accuracy and validity of the training data, it is necessary to preprocess the DC bias simulation results and determine the target training data based on the preprocessed DC bias simulation results. For example, preprocessing may include outlier removal, data standardization, and data normalization. Furthermore, to accommodate the training requirements of neural network models, the target training data can be divided into training, validation, and test sets according to a preset ratio (e.g., 8:1:1).

[0044] S120, a DC bias prediction model is obtained by supervising the training of a pre-constructed neural network model using target training data; wherein, the neural network model embeds physical constraints, which include a first constraint and a second constraint. The first constraint is described based on the electromagnetic field quantity limit corresponding to the transformer material properties, and the second constraint is described based on the physical loss in the loss function. The loss function is constructed based on the weighted sum of the data loss and the physical loss. The data loss is constructed based on the DC bias simulation results and the prediction results of the neural network model, and the physical loss is constructed based on Maxwell's equations, field-circuit coupling equations, and nonlinear magnetization characteristics.

[0045] The DC bias prediction model can refer to a neural network model capable of predicting the DC bias of a transformer, specifically comprising an input layer, multiple hidden layers, and an output layer. The input layer consists of six input nodes, namely... t, area, type and The output layer consists of three output nodes, namely... , and For example, the hidden layers can be 8-10 fully connected layers, each containing 256-512 neurons, using Tanh or Swish activation functions to ensure that the neural network model has sufficient non-linear fitting capability.

[0046] The physical constraints embedded in the neural network model can be used to reflect physical laws, including a first constraint (i.e., hard constraint) and a second constraint (i.e., soft constraint). The first constraint is described based on the electromagnetic field quantity restrictions corresponding to the transformer material properties. Optionally, the electromagnetic field quantity restrictions corresponding to the transformer material properties include zero transformer current density in the air region and iron core region, and a known constant transformer reluctance in the air region and winding region. For example, the first constraint can be set to... and .in, and These represent the air region and the iron core region, respectively. The second constraint is described based on the physical loss in the loss function, which is constructed by weighted summation of data loss and physical loss. The data loss is constructed based on DC bias simulation results and neural network model prediction results, while the physical loss is constructed based on Maxwell's equations, field-circuit coupling equations, and nonlinear magnetization characteristics.

[0047] It should be noted that hard constraints, through mathematical construction, directly ensure that the output of the neural network model strictly satisfies key physical laws within a specific material region, thereby fundamentally avoiding constraint violations. These explicit boundary constraints effectively guide the training process, accelerate convergence, and enhance training stability. Soft constraints, on the other hand, "fine-tune" and optimize complex physical relationships that cannot be precisely expressed using hard constraints. Through this optimization process, the solution gradually approximates the physical laws, ensuring the accuracy and physical rationality of the solution. Therefore, the hybrid strategy of "soft constraints + hard constraints" proposed in this invention provides an innovative technical path for achieving high accuracy, high efficiency, and high reliability in transformer DC bias prediction.

[0048] For example, loss function It can be represented as ,in, and These are data loss and physical loss, respectively. and These are the data loss weights and the physical loss weights, respectively. For example, the data loss can be expressed as: ,in, This represents the number of data points included in a data sample (i.e., (equal to the product of the number of excitation times and the number of grids in the data sample), subscript Indicates the prediction result of the neural network model, subscript This represents the simulation results of DC bias. and They represent the first and second data samples, respectively. The grid space coordinates and excitation time corresponding to each data point.

[0049] In this embodiment, optionally, the physical loss construction process includes: constructing a first loss based on the current law equation in Maxwell's equations; constructing a second loss based on the Gaussian law equation for the magnetic field in Maxwell's equations; constructing a third loss based on the field-circuit coupling equation; constructing a fourth loss based on the nonlinear magnetization characteristics; and determining the physical loss based on the weighted sum of the first, second, third, and fourth losses.

[0050] For example, the first loss can be expressed as The second loss can be expressed as The third loss can be expressed as The fourth loss can be expressed as For example, physical loss can be expressed as: ,in, , , and These represent the first loss weight, the second loss weight, the third loss weight, and the fourth loss weight, respectively.

[0051] In this embodiment, based on the pre-built neural network model, the DC bias prediction model can be obtained through supervised training using the target training data. Specifically, the training process of the neural network model includes: (1) Data loading and model initialization: Read the target training data, initialize the weights of the neural network model using the Xavier method, load the magnetization curve data and create an interpolator, set the optimizer (such as Adam) and the initial learning rate (such as 0.001); (2) Model training and optimization: Under the physical constraints embedded in the neural network model, iterative supervised training is performed using the training set, and the model performance is evaluated once every 10 rounds on the validation set, and the parameter model with the minimum validation loss is saved; (3) Model validation: Evaluate whether the neural network model meets the engineering accuracy requirements on the test set. If it does, the neural network model obtained from the last training is determined as the DC bias prediction model; if it does not, the average relative error of the prediction is determined by comparing the electromagnetic field prediction results with the finite element simulation results, and the model training parameters are adjusted based on the error result and the model training continues until the engineering accuracy requirements are met.

[0052] Furthermore, efficient deployment and system integration of DC bias prediction models can be achieved by designing a unified data interface and modular architecture. Specifically, the standardized data interface design can include the following: (1) Defining input and output data formats: The input interface supports JSON or CSV formats, including excitation parameters ( and ), geometric parameters (such as core size, number of turns), mesh data (such as excitation time, mesh area, mesh material properties); the output interface returns electromagnetic field quantities in structured data (such as HDF5 or JSON). , and (2) Provide API encapsulation: Develop a lightweight API based on RESTful or gRPC protocol, allowing users to send simulation parameters via HTTP requests and receive responses within seconds; (3) Data verification and error handling: The interface has a built-in data verification module to check the physical rationality of input parameters (such as voltage range, geometric constraints) and return standardized error codes (such as HTTP 400 when illegal input is received). The standardized interface lowers the barrier to entry for users, supports rapid integration into existing design processes, and avoids redundant development.

[0053] For modular deployment schemes, the following can be included: (1) Independent model encapsulation: DC bias prediction models can be packaged into Docker containers or Python packages, including model parameters, configuration files (such as network structure, loss function weights) and a list of dependent environments (such as Python runtime environment); (2) Support for multi-platform integration: The module provides plug-in interfaces with commonly used power simulation software (such as ANSYS Maxwell, MATLAB), and achieves seamless coupling through standard data exchange formats (such as FMI / FMU); (3) Scalability and maintainability: Adopting a microservice architecture, training, inference and monitoring functions can be separated, supporting cloud deployment (such as Kubernetes clusters), and realizing dynamic resource scheduling and version management (such as model iteration tracking through Git). Modular deployment ensures the portability and maintainability of the model, which can be applied to different hardware environments (such as edge devices or cloud servers), and meets the needs of real-time monitoring and batch prediction.

[0054] S130, based on the DC bias prediction model, performs DC bias prediction on the transformer under test.

[0055] In this embodiment, after obtaining the DC bias prediction model, it can be encapsulated as an API interface and integrated into the transformer design software. Users only need to input new excitation parameters to generate a visualized electromagnetic field distribution map within seconds, which can be used for transformer DC bias risk assessment. Specifically, the DC bias excitation of the transformer under test is obtained (…). and The DC bias prediction model is used to predict the DC bias of the transformer under test, and finally outputs the DC bias prediction result of the transformer under test. , and The system visualizes the predicted electromagnetic field distribution for users by processing the DC bias prediction results. The transformer under test can refer to the transformer object that requires DC bias simulation.

[0056] The technical solution of this invention first determines target training data based on the DC bias simulation results of the target transformer. The DC bias simulation results include multiple sets of electromagnetic field simulation data under different DC bias excitation conditions. Then, the target training data is used to supervise the training of a pre-constructed neural network model to obtain a DC bias prediction model. The neural network model embeds physical constraints, including a first constraint and a second constraint. The first constraint is described based on electromagnetic field limitations corresponding to the transformer material properties, and the second constraint is described based on physical losses in a loss function. The loss function is constructed based on a weighted sum of data losses and physical losses. The data losses are constructed based on the DC bias simulation results and the prediction results of the neural network model, while the physical losses are constructed based on Maxwell's equations, field-circuit coupling equations, and nonlinear magnetization characteristics. Finally, the DC bias prediction model is used to predict the DC bias of the transformer under test. This technical solution embeds physical constraints, combining soft and hard constraints, into the training process of a neural network model. This allows the trained neural network model (i.e., the DC bias prediction model) to predict the DC bias of transformers. Compared to the traditional finite element analysis method, the neural network model replaces iterative solutions, achieving a second-level response and significantly improving prediction efficiency. Compared to the pure data-driven method, the embedded physical constraints reduce data dependence, ensure the physical rationality of the model's prediction, and improve the model's generalization ability, thereby achieving high-precision and high-efficiency DC bias prediction.

[0057] Example 2

[0058] Figure 2 This is a flowchart of another transformer DC bias prediction method provided in Embodiment 2 of the present invention. This embodiment is based on the above embodiment and optimized. Specifically, the optimization is as follows: a DC bias prediction model is obtained by supervising the training of a pre-built neural network model using target training data, including: supervising the training of the neural network model based on target training data and physical constraints using a multi-stage training method; wherein the weight configuration of data loss and physical loss is different in different training stages; and the neural network model that first meets the preset accuracy requirements is determined as the DC bias prediction model.

[0059] like Figure 2 As shown, the method in this embodiment specifically includes the following steps:

[0060] S210, determine the target training data based on the DC bias simulation results of the target transformer; wherein, the DC bias simulation results include multiple sets of electromagnetic field quantity simulation data under different DC bias excitation conditions.

[0061] S220 employs a multi-stage training approach to supervise the training of a pre-built neural network model based on target training data and physical constraints; the weight configurations for data loss and physical loss differ in different training stages.

[0062] The neural network model incorporates physical constraints, including a first constraint and a second constraint. The first constraint is described based on the electromagnetic field quantity limit corresponding to the transformer material properties, while the second constraint is described based on the physical loss in the loss function. The loss function is constructed based on the weighted sum of the data loss and the physical loss. The data loss is constructed based on the DC bias simulation results and the prediction results of the neural network model, while the physical loss is constructed based on Maxwell's equations, the field-circuit coupling equations, and the nonlinear magnetization characteristics.

[0063] In this embodiment, to further improve model training efficiency while ensuring training accuracy, a multi-stage training approach can be adopted for neural network model training. Each training stage focuses on a different optimization objective and selects different weight configurations for the data loss and physical loss in the objective function. Optionally, the training stages include a first stage, a second stage, and a third stage. In each training stage, the sum of the physical loss weight and the data loss weight is the same. In the first and second stages, the physical loss weight is greater than the data loss weight. In the first stage, the difference between the physical loss weight and the data loss weight is greater than the difference in the second stage. In the third stage, the physical loss weight and the data loss weight are equal.

[0064] Specifically, the first stage (e.g., 0-100 rounds) is the physical law-driven stage, where the optimization goal is to establish basic physical field mapping relationships. The focus is on ensuring that the model's prediction results satisfy Maxwell's equations and the field-path coupling equations, minimizing the equation residuals. For example, assuming the sum of the physical loss weight and the data weight loss is 1, the physical loss weight in the first stage can be set to 0.8, the data weight loss to 0.2, and the initial learning rate to 0.001. The second stage (e.g., 100-300 rounds) is the physical constraint reinforcement stage, where the optimization goal is to strengthen physical consistency and improve generalization ability. The focus is on optimizing magnetization curve consistency and field-path coupling accuracy, and the learning rate can be dynamically adjusted based on the validation loss. For example, the physical loss weight in the second stage can be set to 0.6, and the data weight loss to 0.4. The third stage (e.g., after 300 rounds) is the balancing and fine-tuning stage, where the optimization goal is to balance physical consistency and data fitting accuracy. The focus is on minimizing the total prediction error through fine-tuning while maintaining physical rationality. For example, the physical loss weight and data weight loss in the third stage can both be set to 0.5, and the learning rate can decay to 0.0001.

[0065] S230 identifies the neural network model that first meets the preset accuracy requirements as the DC bias prediction model.

[0066] In this embodiment, model performance verification and accuracy evaluation are required during model training. When the neural network model is detected to meet the preset accuracy requirements for the first time, the model training process ends, and the neural network model that meets the preset accuracy requirements is identified as the DC bias prediction model. The preset accuracy requirements can be flexibly set according to actual engineering needs.

[0067] S240, based on the DC bias prediction model, performs DC bias prediction on the transformer under test.

[0068] Figure 3 A schematic diagram illustrating the overall process of predicting DC bias in a transformer is presented. Figure 3 As shown, the overall process consists of three stages: high-fidelity simulation data preparation (stage one), physical information neural network construction and training (stage two), and online deployment and real-time simulation (stage three). Specifically, in stage one, a target geometric model is first constructed using parametric geometric modeling based on COMSOL. Then, multi-condition finite element simulation is performed on the target geometric model to obtain DC bias simulation results. Finally, high-dimensional data is parametrically organized to form coordinate-field quantity CSV format data as target training data, thereby constructing a standardized training database.

[0069] In Phase Two, target training data is first acquired from a standardized training database, and then a pre-built neural network model undergoes multi-stage adaptive training based on this data. It's important to note that because the neural network model embeds physical constraints (including hard and soft constraints), it can be considered a Physics-Informed Neural Network (PINN) model, which combines deep learning with physical modeling. During model training, model performance verification and accuracy evaluation are required. If the model accuracy meets the requirements, a high-performance PINN model is directly output as the DC bias prediction model. If the model accuracy does not meet the requirements, network update parameters are determined based on the model evaluation results, the network architecture is initialized, and the model undergoes multi-stage adaptive training again until the model accuracy meets the requirements, at which point a high-performance PINN model is output as the DC bias prediction model.

[0070] In Phase Three, the excitation parameters under the new DC bias condition are first obtained. and The model file parameters of the DC bias prediction model are read and loaded through the model inference engine. Based on the preprocessed excitation parameters, DC bias prediction is performed, and the electromagnetic field quantity prediction parameters are finally output (i.e., , and This enables rapid prediction of electromagnetic field quantities under DC bias conditions.

[0071] The technical solution of this invention employs a multi-stage training approach to supervise the training of a neural network model based on target training data and physical constraints. The weight configurations for data loss and physical loss differ in different training stages. The neural network model that first meets the preset accuracy requirements is identified as the DC bias prediction model, and DC bias prediction of the transformer under test is performed based on this model. This technical solution, by using a multi-stage training approach for model training and achieving different optimization objectives through different weight configurations in the loss function, can further improve model training efficiency while ensuring training accuracy.

[0072] Example 3

[0073] Figure 4 This is a schematic diagram of a transformer DC bias prediction device provided in Embodiment 3 of the present invention. This device can execute the transformer DC bias prediction method provided in any embodiment of the present invention, and possesses the corresponding functional modules and beneficial effects of the method. Figure 4 As shown, the device includes:

[0074] The training data determination module 310 is used to determine the target training data based on the DC bias simulation results of the target transformer; wherein, the DC bias simulation results include multiple sets of electromagnetic field quantity simulation data under different DC bias excitation conditions;

[0075] The model training module 320 is used to supervise the training of a pre-constructed neural network model using the target training data to obtain a DC bias prediction model. The neural network model embeds physical constraints, including a first constraint and a second constraint. The first constraint is described based on electromagnetic field limitations corresponding to transformer material properties, and the second constraint is described based on physical loss in a loss function. The loss function is constructed based on a weighted sum of data loss and physical loss. The data loss is constructed based on DC bias simulation results and the prediction results of the neural network model, while the physical loss is constructed based on Maxwell's equations, field-circuit coupling equations, and nonlinear magnetization characteristics.

[0076] The model prediction module 330 is used to predict the DC bias of the transformer under test based on the DC bias prediction model.

[0077] Optionally, the training data determination module 310 is specifically used for:

[0078] Construct a target geometric model based on the target transformer's geometry and material properties;

[0079] The DC bias magnetization simulation results of the target transformer were obtained by performing DC bias magnetization simulation on the target geometric model using the finite element analysis method.

[0080] The DC bias simulation results are preprocessed to obtain target training data.

[0081] Optionally, the electromagnetic field quantity restrictions corresponding to the transformer material properties include zero transformer current density in the air region and core region, and a known constant transformer reluctance in the air region and winding region.

[0082] Optionally, the process of constructing the physical loss includes:

[0083] The first loss is constructed based on the total current law equation in Maxwell's equations;

[0084] The second loss is constructed based on Gauss's theorem equation for the magnetic field in Maxwell's equations;

[0085] A third loss is constructed based on the field-path coupling equation;

[0086] A fourth loss is constructed based on the aforementioned nonlinear magnetization characteristics;

[0087] The physical loss is determined by a weighted sum of the first loss, the second loss, the third loss, and the fourth loss.

[0088] Optionally, the model training module 320 is used for:

[0089] The neural network model is trained in a supervised manner using a multi-stage training approach based on the target training data and the physical constraints; wherein the weight configurations of the data loss and the physical loss are different in different training stages.

[0090] The neural network model that first meets the preset accuracy requirements is identified as the DC bias prediction model.

[0091] Optionally, the training phase includes a first phase, a second phase, and a third phase. The sum of the physical loss weight and the data weight loss is the same in each training phase. The physical loss weight in the first phase and the second phase is greater than the data weight loss. The difference between the physical loss weight and the data weight loss in the first phase is greater than the difference between the physical loss weight and the data weight loss in the second phase. The physical loss weight in the third phase is equal to the data weight loss.

[0092] The transformer DC bias prediction device provided in this embodiment of the invention can execute the transformer DC bias prediction method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method.

[0093] Example 4

[0094] Figure 5A schematic diagram of an electronic device 10, which can be used to implement embodiments of the present invention, is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0095] like Figure 5 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0096] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0097] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as the transformer DC bias prediction method.

[0098] In some embodiments, the transformer DC bias prediction method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the transformer DC bias prediction method described above may be performed. Alternatively, in other embodiments, processor 11 may be configured to perform the transformer DC bias prediction method by any other suitable means (e.g., by means of firmware).

[0099] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0100] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0101] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0102] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0103] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0104] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0105] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0106] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for predicting DC bias magnetism in a transformer, characterized in that, The method includes: The target training data is determined based on the DC bias simulation results of the target transformer; wherein, the DC bias simulation results include multiple sets of electromagnetic field quantity simulation data under different DC bias excitation conditions; A DC bias prediction model is obtained by supervising the training of a pre-constructed neural network model using the target training data. The neural network model incorporates physical constraints, including a first constraint and a second constraint. The first constraint is described based on electromagnetic field limitations corresponding to transformer material properties, while the second constraint is described based on physical losses in a loss function. The loss function is constructed based on a weighted sum of data loss and physical loss. The data loss is constructed based on the DC bias simulation results and the prediction results of the neural network model, while the physical loss is constructed based on Maxwell's equations, field-circuit coupling equations, and nonlinear magnetization characteristics. Based on the DC bias prediction model, DC bias prediction is performed on the transformer under test. The target training data is determined based on the DC bias magnetization simulation results of the target transformer, including: Construct a target geometric model based on the target transformer's geometry and material properties; The DC bias magnetization simulation results of the target transformer were obtained by performing DC bias magnetization simulation on the target geometric model using the finite element analysis method. The DC bias simulation results are preprocessed to obtain target training data; The DC bias prediction model is obtained by supervising the training of a pre-built neural network model using the target training data, including: A multi-stage training method is used to supervise the training of the neural network model based on the target training data and the physical constraints. The weights of the data loss and the physical loss are configured differently in different training stages. The training stages include a first stage, a second stage, and a third stage. The sum of the physical loss weight and the data loss weight is the same in each training stage. The physical loss weight is greater than the data loss weight in the first and second stages. The difference between the physical loss weight and the data loss weight in the first stage is greater than the difference between the physical loss weight and the data loss weight in the second stage. The physical loss weight is equal to the data loss weight in the third stage. The neural network model that first meets the preset accuracy requirements is identified as the DC bias prediction model.

2. The method according to claim 1, characterized in that, The electromagnetic field constraints corresponding to the transformer material properties include zero transformer current density in the air region and core region, and a known constant transformer reluctance in the air region and winding region.

3. The method according to claim 1, characterized in that, The process of constructing the physical loss includes: The first loss is constructed based on the total current law equation in Maxwell's equations; The second loss is constructed based on Gauss's theorem equation for the magnetic field in Maxwell's equations; A third loss is constructed based on the field-path coupling equation; A fourth loss is constructed based on the aforementioned nonlinear magnetization characteristics; The physical loss is determined by a weighted sum of the first loss, the second loss, the third loss, and the fourth loss.

4. A transformer DC bias prediction device, characterized in that, The device includes: The training data determination module is used to determine the target training data based on the DC bias simulation results of the target transformer; wherein, the DC bias simulation results include multiple sets of electromagnetic field quantity simulation data under different DC bias excitation conditions; The model training module is used to supervise the training of a pre-constructed neural network model using the target training data to obtain a DC bias prediction model. The neural network model embeds physical constraints, including a first constraint and a second constraint. The first constraint is described based on electromagnetic field limitations corresponding to transformer material properties, and the second constraint is described based on physical losses in a loss function. The loss function is constructed based on a weighted sum of data losses and physical losses. The data losses are constructed based on the DC bias simulation results and the prediction results of the neural network model, while the physical losses are constructed based on Maxwell's equations, field-circuit coupling equations, and nonlinear magnetization characteristics. The model prediction module is used to predict the DC bias of the transformer under test based on the DC bias prediction model. The training data determination module is specifically used for: Construct a target geometric model based on the target transformer's geometry and material properties; The DC bias magnetization simulation results of the target transformer were obtained by performing DC bias magnetization simulation on the target geometric model using the finite element analysis method. The DC bias simulation results are preprocessed to obtain target training data; Specifically, the model training module is used for: A multi-stage training method is used to supervise the training of the neural network model based on the target training data and the physical constraints. The weights of the data loss and the physical loss are configured differently in different training stages. The training stages include a first stage, a second stage, and a third stage. The sum of the physical loss weight and the data loss weight is the same in each training stage. The physical loss weight is greater than the data loss weight in the first and second stages. The difference between the physical loss weight and the data loss weight in the first stage is greater than the difference between the physical loss weight and the data loss weight in the second stage. The physical loss weight is equal to the data loss weight in the third stage. The neural network model that first meets the preset accuracy requirements is identified as the DC bias prediction model.

5. An electronic device, characterized in that, The electronic device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the transformer DC bias prediction method according to any one of claims 1-3.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a processor, implement the transformer DC bias prediction method according to any one of claims 1-3.

Images