A method, system, equipment, and medium for constructing a low-frequency transformer surrogate model based on physical feature guidance and Bayesian active learning.

Abstract

This invention belongs to the field of low-frequency transformer magnetic field optimization technology, and discloses a method, system, device, and medium for constructing a low-frequency transformer surrogate model based on physical feature guidance and Bayesian active learning to solve the problem of prediction distortion in key performance areas. The method includes: constructing a global parameter space; establishing a multi-physics finite element model; defining high-risk performance areas based on low-frequency physical laws; constructing an initial Kriging surrogate model using Latin hypercube sampling; setting penalty weights based on whether unresolved sample points are located in high-risk areas and introducing a point-addition criterion function to solve for the optimal sampling points; calling the finite element model to calculate the true response value and updating the surrogate model; determining whether the prediction accuracy of high-risk areas meets the convergence condition; if it does, outputting the current surrogate model; otherwise, continuing the iteration.

Description

Technical Field

[0001] This invention belongs to the field of low-frequency transformer magnetic field optimization technology, specifically involving a method, system, device, and medium for constructing a low-frequency transformer proxy model based on physical feature guidance and Bayesian active learning. Background Technology

[0002] With the development of low-frequency (e.g., 20Hz) AC power transmission technology, low-frequency transformers, as core hub equipment of the system, are crucial for operational stability. Unlike conventional power-frequency transformers, low-frequency transformers are prone to deep magnetic saturation of the core under low-frequency conditions, and their large size makes heat dissipation through convection of the internal insulating oil difficult. Simultaneously, the eddy current loss density distribution in the winding area exhibits strong non-uniformity, easily leading to localized overheating and insulation aging. Therefore, in the lightweight design of low-frequency transformers, it is necessary to strictly constrain their limiting electromagnetic and thermodynamic performance boundaries.

[0003] In the optimization of the magnetic field and performance prediction of low-frequency transformers, surrogate models (such as the Kriging model) are often used in engineering to replace computationally expensive three-dimensional finite element simulations. The prediction accuracy of the surrogate model largely depends on the quality of the training sample set. Current sampling methods for constructing surrogate models still have shortcomings. On the one hand, static sampling strategies are prone to blindness. Traditional methods often employ Latin hypercube sampling or uniform experimental design, performing a one-time uniform sampling in the global parameter space. However, the electromagnetic response of low-frequency transformers exhibits strong local nonlinear abrupt changes; for example, when the structural dimensions approach a certain critical value, the local magnetic flux density may increase sharply. Uniform sampling methods struggle to actively identify and focus on these sensitive areas crucial to equipment safety, resulting in insufficient prediction accuracy of the surrogate model at critical performance boundaries. On the other hand, existing active learning methods lack guidance from prior physical knowledge. Some studies have introduced Bayesian active learning algorithms, using the prediction variance of the surrogate model for additional sampling. However, these methods treat transformers as pure "black box" models. The algorithms themselves cannot understand the underlying physical mechanisms such as winding eddy current losses and low-frequency magnetic saturation criticality. This leads to sampling points being easily distributed in mathematical variance fluctuation regions that lack engineering physical significance, resulting in a waste of high-cost 3D simulation resources and making it difficult to effectively improve the modeling accuracy of key areas within a limited computational budget. Summary of the Invention

[0004] Based on the aforementioned shortcomings and deficiencies in the existing technology, one of the objectives of this invention is to at least solve one or more of the aforementioned problems in the existing technology. In other words, one of the objectives of this invention is to provide a method, system, device, and medium for constructing a low-frequency transformer proxy model based on physical feature guidance and Bayesian active learning that meets one or more of the aforementioned requirements, so as to improve the prediction accuracy of key performance areas and reduce the computational cost of high-fidelity simulation.

[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for constructing a low-frequency transformer surrogate model based on physical feature guidance and Bayesian active learning, comprising the following steps: S1. Extract the key structural parameters of the low-frequency transformer and set the initial value range of each variable to construct a global parameter space; S2. Establish a multiphysics finite element model for electromagnetic simulation of low-frequency transformers; S3. Based on low-frequency physical laws, the magnetic flux density critical region and / or heat dissipation bottleneck region are defined as high-risk performance areas within the global parameter space. S4. Use Latin hypercube sampling to generate an initial sample set from the global parameter space, call the multiphysics finite element model to calculate the corresponding real response value, and construct the initial Kriging proxy model. S5. Based on whether the unresolved sample points in the global parameter space are located in the high-risk performance region, set a penalty weight, introduce the penalty weight into the active learning point addition criterion function, and solve for the maximum value of the function in the global parameter space as the optimal sampling point. S6. Calculate the true response value of the optimal sampling point using the multiphysics finite element model, add it to the training sample set, and update the Kriging proxy model. S7. Determine whether the prediction accuracy of the updated surrogate model in the high-risk performance region meets the preset convergence condition. If it does, output the current surrogate model; otherwise, return to step S5 to continue iterating.

[0006] As a preferred option: The key structural parameters of the low-frequency transformer include the core column diameter, yoke width, upper window clearance, lower window clearance, phase center distance, and distance from the core end to the boundary.

[0007] As a preferred option, step S3 specifically involves: When the combination of parameters in the key structural parameters of the low-frequency transformer causes the magnetic flux density at the corner inside the core column to be within a preset low-frequency critical magnetic flux range, the spatial domain corresponding to the combination is determined to be the critical region of magnetic flux density. When the combination of parameters in the key structural parameters of the low-frequency transformer causes the insulating oil flow rate at the winding gap to be lower than the preset flow rate threshold and the local heat generation rate to be higher than the preset heat generation rate threshold, the spatial domain corresponding to the combination is determined as the heat dissipation bottleneck area.

[0008] As a preferred option: The actual response values ​​include the core volume, maximum magnetic flux density, and core loss of the low-frequency transformer.

[0009] As a preferred option, step S5 specifically involves: The expected improvement amount and feasibility probability of the unresolved sample points are calculated using the posterior prediction variance of the Kriging surrogate model. Construct a physical feature identification function and set corresponding penalty weights based on whether the unresolved sample points in the global parameter space are located in the high-risk performance region; Multiply the feasibility probability, the expected improvement amount, and the physical feature identification function with the set penalty weight to obtain the active learning addition criterion function; The optimal sampling point is obtained by using a genetic algorithm to find the maximum value of the active learning addition criterion function in the global parameter space.

[0010] As a preferred embodiment, the rules for setting the penalty weights are as follows: When the unresolved sample point is located within the high-risk performance region, the physical feature identification function is set to take an amplified penalty weight; When the pending sample point is located outside the high-risk performance region, the physical feature identification function is set to have a reduced weight.

[0011] As a preferred embodiment, the preset convergence condition in step S7 is: In multiple iterations, the relative error of the surrogate model in predicting the maximum magnetic flux density of newly added sample points in the high-risk performance region and the relative error in predicting the maximum temperature rise of the winding are both less than the preset tolerance.

[0012] Secondly, the present invention provides a low-frequency transformer proxy model construction system based on physical feature guidance and Bayesian active learning, for implementing the low-frequency transformer proxy model construction method as described in the first aspect.

[0013] Thirdly, the present invention provides an electronic device, the electronic device including a memory, a processor and a computer program, wherein when the computer program is executed by the processor, it implements the low-frequency transformer proxy model construction method as described in the first aspect.

[0014] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the low-frequency transformer proxy model construction method as described in the first aspect.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention achieves a deep integration of prior physical knowledge and intelligent sampling algorithms. It breaks through the limitations of traditional mathematical modeling that treats transformers as "data black boxes," innovatively extracting the unique underlying physical characteristics of low-frequency transformers as penalty weights and directly embedding them into the Bayesian active learning addition criterion function. This imbues the intelligent algorithm with "engineering physics common sense," significantly improving the physical orientation of the sampling strategy.

[0016] 2. Significantly improved magnetic field prediction accuracy under extreme operating conditions. With the help of a feature-guided active learning mechanism, the sampling algorithm is no longer limited to the global static distribution, but can spontaneously and dynamically approach the boundary of high prediction uncertainty and located in the physical high-risk critical region. This makes the final Kriging surrogate model have extremely high fidelity in the extreme nonlinear region that determines the safety bottom line of the equipment.

[0017] 3. Significantly reduces computational redundancy in magnetic field simulation. Compared to the traditional approach of exponentially increasing the global sample size to forcibly improve accuracy, this invention effectively avoids ineffective and high-cost 3D finite element simulations in safe and smooth regions and mathematically insignificant fluctuation areas while ensuring accuracy in key regions, thus bringing the construction cost of the transformer proxy model close to the theoretical lower limit.

[0018] Further or more detailed beneficial effects will be described in conjunction with specific embodiments in the detailed implementation. Attached Figure Description

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

[0020] Figure 1 This is a flowchart illustrating the method for constructing a low-frequency transformer proxy model according to Embodiment 1 of the present invention.

[0021] Figure 2 This is a schematic diagram illustrating the physical high-risk area delineation under no-load conditions of a low-frequency transformer as described in Embodiment 1 of the present invention.

[0022] Figure 3 This is a schematic diagram illustrating the physical high-risk area delineation under the load conditions of a low-frequency transformer as described in Embodiment 1 of the present invention.

[0023] Figure 4 This is a schematic diagram of the adaptive peak-finding sampling process in which the penalty weight is introduced into the active learning addition criterion function as described in Embodiment 1 of the present invention.

[0024] Figure 5 This is a comparison diagram of the sample point distribution in the parameter space between conventional uniform sampling and Bayesian active adaptive sampling as described in Embodiment 1 of the present invention.

[0025] Figure 6 This is a convergence curve of the prediction error of the maximum magnetic flux density and the maximum temperature rise under continuous iteration of the Kriging surrogate model described in Embodiment 1 of the present invention.

[0026] Figure 7 This is a structural diagram of the electronic device described in Embodiment 3 of the present invention.

[0027] Icon labels: 700. Electronic equipment; 701. Processor; 702. Communication bus; 703. User interface; 704. Network interface; 705. Memory. Detailed Implementation

[0028] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0029] In the following description, several embodiments of the present invention are provided. Different embodiments can be substituted or combined. Therefore, the present invention can also be considered to include all possible combinations of the same and / or different embodiments described. Thus, if one embodiment includes features A, B, and C, and another embodiment includes features B and D, then the present invention should also be considered to include embodiments containing one or more other possible combinations of A, B, C, and D, even if such embodiments are not explicitly described in the following text.

[0030] The following description provides examples and does not limit the scope, applicability, or examples set forth in the claims. Changes may be made to the function and arrangement of the described elements without departing from the scope of the invention. Various processes or components may be appropriately omitted, substituted, or added to the various examples. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Furthermore, features described with respect to some examples may be combined into other examples.

[0031] To facilitate a better understanding of the embodiments of the present invention, its application scenarios will be explained before providing a detailed explanation of the specific implementation methods.

[0032] The low-frequency transformer surrogate model construction method described in this specification is applied to lightweight design and electromagnetic performance optimization of low-frequency transformers, particularly in low-frequency (e.g., 20Hz) AC transmission systems where precise evaluation of the transformer's magnetic saturation characteristics, winding eddy current loss distribution, and thermal performance boundaries is required. In these scenarios, the application of this low-frequency transformer surrogate model construction method aims to construct an adaptive surrogate model with high predictive accuracy in key performance regions with a limited number of costly 3D finite element simulation calls. This replaces a large number of repetitive physical simulation calculations, providing an efficient and reliable predictive tool for optimizing structural parameters, lightweight design, and safety boundary verification of low-frequency transformers.

[0033] The following is a brief explanation of the low-frequency transformer, surrogate model, physical feature guidance, Bayesian active learning, key structural parameters, multiphysics finite element model, magnetic flux density critical region, heat dissipation bottleneck region, Latin hypercube sampling, unresolved sample points, active learning addition criterion function, and genetic algorithm involved in several embodiments of this specification: Low-frequency transformers refer to power transformers that operate at frequencies lower than the power frequency (50Hz / 60Hz), and are typically used in special applications such as 20Hz AC transmission systems. Compared to conventional power frequency transformers, low-frequency transformers are more prone to deep magnetic saturation of the core under low-frequency conditions, and the eddy current loss density distribution in the windings exhibits strong non-uniformity.

[0034] A surrogate model, also known as a meta-model or response surface model, is an approximate model with a computational cost far lower than high-fidelity physical simulation. This embodiment employs the Kriging surrogate model, which provides a quantitative assessment of prediction uncertainty (i.e., posterior prediction variance) along with the predicted value, making it suitable for adaptive sampling strategies based on active learning.

[0035] Physical feature guidance refers to embedding the unique underlying physical mechanisms of low-frequency transformers (such as low-frequency magnetic saturation threshold, winding eddy current loss accumulation, and insulating oil flow characteristics) as prior knowledge into the decision-making process of intelligent sampling algorithms, so that the sampling strategy has common sense of engineering physics and avoids treating the transformer as a pure "black box" model.

[0036] Bayesian active learning refers to an active learning method based on the Bayesian statistical inference framework. It uses the posterior prediction variance output by the surrogate model to quantify the prediction uncertainty of unsampled points. By constructing a point addition criterion function, it automatically selects the next sampling point with the most information, thereby obtaining the highest modeling accuracy with the least amount of samples.

[0037] Key structural parameters refer to the core geometric dimensions that affect the electromagnetic and thermal performance of low-frequency transformers, including the core column diameter, yoke width, upper window clearance, lower window clearance, phase-to-phase center distance, and the distance from the core end to the boundary. These parameters together constitute the global parameter space to be optimized.

[0038] A multiphysics finite element model refers to a three-dimensional finite element simulation model that can simultaneously solve for the electromagnetic field and thermal flow field of a low-frequency transformer. In this embodiment, the COMSOL simulation software is used to establish this model to calculate the magnetic flux density distribution under no-load conditions and the winding eddy current loss density distribution under rated load conditions.

[0039] The critical region of magnetic flux density refers to the parameter range in the global parameter space where the combination of key structural parameters causes the magnetic flux density at the corners inside the transformer core to approach the low-frequency deep saturation threshold. This region is a high-risk area that must be strictly avoided in the lightweight design of transformers.

[0040] The heat dissipation bottleneck zone refers to the parameter range in the global parameter space where a combination of key structural parameters simultaneously causes the insulating oil flow rate at the winding gap to fall below a preset threshold, while the local heat generation rate exceeds a preset threshold. This region is a high-risk area in transformer design that is prone to causing localized overheating and insulation aging.

[0041] Latin hypercube sampling is a stratified sampling experimental design method that can achieve relatively uniform spatial filling in a multidimensional parameter space with fewer sample points. In this embodiment, it is used to generate an initial sparse sample set as the basic data for constructing the initial Kriging surrogate model.

[0042] Undecided sample points refer to candidate sample points that have not yet been used in the finite element simulation model to calculate the true response value during the active learning iteration process. The active learning point addition criterion function evaluates all undecided sample points and selects the point with the largest function value as the next optimal sampling point.

[0043] The active learning criterion function, also known as the acquisition function, is the core function used to guide sampling decisions in Bayesian active learning. In this embodiment, the function is composed of the product of the feasibility probability of the undecided sample point, the expected improvement amount, and the physical feature identification function. The maximum value of this function corresponds to the optimal sampling point for the next high-cost simulation calculation.

[0044] Genetic algorithms are global optimization algorithms that simulate natural selection and genetic mechanisms. They search for the global optimum in the parameter space through operations such as selection, crossover, and mutation. In this embodiment, a genetic algorithm is used to find the maximum point of the active learning addition criterion function in the global parameter space. It has the advantages of not relying on gradient information and being suitable for high-dimensional nonlinear optimization problems.

[0045] Example 1: like Figure 1 As shown, this embodiment provides a method for constructing a low-frequency transformer proxy model based on physical feature guidance and Bayesian active learning, including the following steps S1 to S7.

[0046] Step S1: Construct the global parameter space: Key structural parameters of the low-frequency transformer are extracted as a set of optimization variables, and initial value ranges for each variable are set to construct a global parameter space. In this embodiment, the key structural parameters include the core column diameter. D Magnetic yoke width b Upper window clearance h 1. Lower window clearance h 2. Center distance between phases M and the distance from the end of the core to the boundary E Based on insulation and assembly requirements, the allowable range of values ​​for each parameter is set to form an initial six-dimensional global parameter space.

[0047] Step S2: Establish a multiphysics finite element model: A multiphysics finite element model for electromagnetic simulation of low-frequency transformers was established. In this embodiment, a magnetic field simulation model of a 220kV low-frequency transformer was established using COMSOL simulation software. This model can solve for the three-dimensional magnetic flux density distribution and winding eddy current loss density distribution under low-frequency operating conditions.

[0048] Step S3: Define high-risk performance areas: Based on low-frequency physical laws, the magnetic flux density critical region and / or heat dissipation bottleneck region are defined as high-risk performance areas within the global parameter space.

[0049] Specifically, when the combination of key structural parameters of a low-frequency transformer causes the magnetic flux density at the corner inside the core column to fall within a preset low-frequency critical flux range, the spatial domain corresponding to this combination is determined as the critical flux density region. Analysis of no-load conditions, such as... Figure 2 As shown, transformer cores are prone to deep saturation at low frequencies. Simulations revealed that magnetic flux density is concentrated at the core column and internal corners, with a maximum value of 1.85T. Setting the low-frequency safe operation threshold to 1.75T, the preset low-frequency critical magnetic flux range is [1.75, 1.85]. The structural parameter combination range that causes the magnetic flux density at the internal corners of the core column to approach this threshold is defined as the critical flux density region.

[0050] When the combination of key structural parameters of a low-frequency transformer causes the insulating oil flow velocity at the winding gap to be lower than the preset flow velocity threshold of 0.15 m / s, and the local heat generation rate to be higher than the preset heat generation rate threshold of 18000 W / m, 3At that time, the spatial domain corresponding to this combination is identified as the heat dissipation bottleneck area. Analyze the rated load conditions, such as... Figure 3 As shown, leakage magnetic field and losses are mainly concentrated in the gap between the high and low voltage windings and at the ends. When the window clearance (upper window clearance) is... h 1. Clear distance between the lower window and the window h When the distance (2) is too small, the flow velocity of the insulating oil at the winding gap decreases significantly, leading to a high accumulation of eddy current loss heat. The parameter range that causes oil flow stagnation and the local heat generation rate to exceed the warning value is designated as the heat dissipation bottleneck area.

[0051] Step S4: Construct the initial Kriging proxy model: An initial sample set is generated from the global parameter space using Latin hypercube sampling (LHS). The corresponding true response values ​​are calculated using the multiphysics finite element model, and an initial Kriging surrogate model is constructed. In this embodiment, 50 initial sparse sample points are generated in the six-dimensional global parameter space using Latin hypercube sampling. The true response values ​​of each sample point are calculated using the finite element model. These true response values ​​include the core volume, maximum magnetic flux density, and core loss of the low-frequency transformer. A model containing predicted mean is constructed. With posterior prediction variance The initial Kriging proxy model.

[0052] Step S5: Adaptive sampling guided by physical features: Based on whether the unresolved sample points in the global parameter space are located in the high-risk performance region, a penalty weight is set, and the penalty weight is introduced into the active learning point addition criterion function. The maximum value of this function in the global parameter space is then used as the optimal sampling point.

[0053] In this embodiment, step S5 is specifically implemented as follows: First, the expected improvement (EI) and probability of feasibility (PoF) of the undecided sample point are calculated using the posterior prediction variance of the Kriging surrogate model. The expected improvement is used to assess the potential of the sample point in achieving the lightweight objective, while the probability of feasibility quantifies the probability that the sample point will not exceed the magnetic saturation and overheating safety limits.

[0054] Secondly, construct the physical feature identification function. W ( x The penalty weight is set based on whether the unresolved sample points in the global parameter space are located in the high-risk performance region. For example... Figure 2 and Figure 3 As shown, when unresolved sample points xWhen the performance is located within the high-risk performance area defined in step S3, the physical feature identification function is set to amplify the penalty weight. W ( x ) = 1.5; when unresolved sample points x When located outside the high-risk performance region, the physical feature identification function is set to reduce the weight. W ( x =0.5.

[0055] Then, the feasibility probability, the expected improvement amount, and the physical feature identification function with the set penalty weights are multiplied to obtain the active learning addition criterion function, i.e.: , In the formula: AF ( x ) represents the number of unresolved sample points in the global parameter space. x The larger the comprehensive expected value of the addition (i.e. the value of the active learning addition criterion function), the higher the value of calling high-fidelity multiphysics simulation for real sampling at this point; x It represents the unresolved sample points in the global parameter space, i.e., the multidimensional input vector of key structural parameters; PoF ( x ) indicates pending sample points x The feasibility probability is the probability that the design scheme corresponding to the point satisfies all the above external constraints, based on the mean and variance prediction of the current proxy model (such as the Kriging model). EI ( x ) indicates pending sample points x The expected improvement, i.e., the calculation of the unknown point relative to the currently known optimal feasible solution, with respect to the optimization objective. x The expected improvement in the objective function that may result; W ( x ) represents the physical feature identification function, which serves as an adaptive penalty and incentive weight factor in the spatial domain, used to embed prior physical field knowledge into a purely mathematical point-addition strategy.

[0056] Finally, the maximum value of the active learning addition criterion function in the global parameter space is used as the optimal sampling point. For example... Figure 4 As shown, this embodiment uses a genetic algorithm to solve for the maximum coordinates of the criterion function in a continuous space. This mechanism forces the sampling points to spontaneously cluster towards the boundary where the physical risk is high and the model is extremely uncertain. Figure 4The figure illustrates the response surface morphology of the point-addition criterion function under two-dimensional parameter space projection. As can be seen from the figure, compared to the traditional PoF-EI criterion, the criterion function in this embodiment exhibits a distinct "bimodal" characteristic at the boundary of the high-risk performance region, with higher function values ​​near the magnetic flux density critical region and the heat dissipation bottleneck region. The genetic algorithm gradually converges to the global maximum point through selection, crossover, and mutation operations within the parameter space. Figure 4 The contour plot clearly illustrates the optimization path: the initial population is evenly distributed within the parameter space, and as the number of iterations increases, the individuals gradually cluster towards the boundary of the high-risk performance region, eventually converging to the optimal sampling point around the 50th generation. This process demonstrates that the method described in this invention can adaptively tilt sampling resources towards regions with high physical risk and high model prediction uncertainty, avoiding wasting computational resources in safe, flat regions or purely mathematically volatile regions.

[0057] like Figure 5 As shown, comparing the sample point distribution of conventional Latin hypercube uniform sampling with that of the method in this embodiment, it can be seen that the traditional uniform sampling method ( Figure 5 In (a) of the model, the sampling points are evenly distributed throughout the parameter space. Although the total number of samples is large, the sample points are sparsely distributed in high-risk performance areas such as the magnetic flux density critical region and the heat dissipation bottleneck region, resulting in insufficient prediction accuracy of the surrogate model in these key areas. The sampling method based on physical feature guidance and Bayesian active learning described in this embodiment (…) Figure 5 In (b) of the sample, the initial sample is uniformly distributed using Latin hypercube sampling. Then, guided by the active learning of the point criterion function, the subsequent sampling points spontaneously cluster towards the boundary of the high-risk performance region. Figure 5 It is evident that the sampling point density is significantly higher in the vicinity of the magnetic flux density critical region (highlighted area in the figure) and the heat dissipation bottleneck region (shaded area in the figure) than in other regions. Statistical analysis shows that, to achieve the same prediction accuracy in high-risk areas, the simulation calls required by the method in this embodiment are approximately 40% of those required by the traditional uniform sampling method, thus reducing computational costs by about 60%. This demonstrates that the method described in this invention can effectively focus on critical performance boundaries and avoid ineffective, high-cost three-dimensional finite element simulations in safe and flat regions.

[0058] Step S6: Dynamically update the agent model: The multiphysics finite element model is used to calculate the true response value of the optimal sampling point, which is then added to the training sample set to update the Kriging surrogate model. In this embodiment, the optimal sampling point determined in step S5 is input into the finite element simulation model. After high-fidelity solving, the true response values ​​such as the core volume, maximum magnetic flux density, and highest hot spot temperature rise of the winding under these parameters are obtained. This set of high-fidelity data is then fed back into the training sample set to dynamically update the hyperparameters of the Kriging surrogate model.

[0059] Step S7, Convergence Determination and Model Output: Determine whether the prediction accuracy of the updated surrogate model in the high-risk performance region meets the preset convergence condition. If it does, output the current surrogate model; otherwise, return to step S5 to continue iterating.

[0060] In this embodiment, the preset convergence condition is: in multiple consecutive iterations, the relative error of the surrogate model's prediction of the maximum magnetic flux density of newly added sample points in the high-risk performance region and the relative error of its prediction of the maximum temperature rise of the winding are both less than a preset tolerance. For example... Figure 6 As shown, Figure 6 This figure illustrates the convergence curve of the surrogate model's prediction error for sample points within the high-risk performance region during continuous iterative sampling in this embodiment. The horizontal axis represents the number of Bayesian active learning iterations, and the vertical axis represents the relative prediction error (expressed as a percentage). The solid curve represents the trend of the relative prediction error for the maximum magnetic flux density, while the dashed curve represents the trend of the relative prediction error for the maximum temperature rise of the winding. Figure 6 It can be seen that in the first 10 iterations, due to the small number of initial samples, both prediction errors are relatively high. As the number of iterations increases, the sampling points gradually gather towards the boundary of the high-risk area, and the prediction error shows a rapid downward trend. After 50 iterations, both prediction errors stabilize below 1.5%, and there is no error rebound in 10 consecutive iterations. In this embodiment, the preset convergence condition is: in 10 consecutive iterations, the relative error of the surrogate model in predicting the maximum magnetic flux density of newly added sample points in the high-risk performance area and the relative error in predicting the maximum temperature rise of the winding are both less than 1.5%. When this convergence condition is met, sampling stops, and the current surrogate model is output as the final high-fidelity adaptive surrogate model; if it is not met, the process returns to step S5 to continue iterating. Figure 6 The method described in this invention has been fully verified to have good convergence characteristics and can achieve high-precision modeling of high-risk regions within a limited computational budget.

[0061] Example 2: This embodiment provides a low-frequency transformer proxy model construction system based on physical feature guidance and Bayesian active learning, which is used to implement the low-frequency transformer proxy model construction method as described in Embodiment 1.

[0062] Example 3: like Figure 7 As shown, this embodiment provides an electronic device, which may include: at least one processor, at least one network interface, a user interface, a memory, and at least one communication bus.

[0063] The communication bus can be used to enable communication between the various components mentioned above.

[0064] The user interface may include buttons, and optional user interfaces may also include standard wired interfaces and wireless interfaces.

[0065] The network interface may include, but is not limited to, Bluetooth modules, NFC modules, Wi-Fi modules, etc.

[0066] The processor may include one or more processing cores. It connects various parts of the electronic device via various interfaces and lines, executing instructions, programs, code sets, or instruction sets stored in memory, and accessing data stored in memory to perform various functions and process data. Optionally, the processor can be implemented using at least one hardware form of DSP, FPGA, or PLA. The processor may integrate one or more of the following: CPU, GPU, and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also be implemented as a separate chip without being integrated into the processor.

[0067] The memory may include RAM or ROM. Optionally, the memory may include a non-transitory computer-readable medium. The memory can be used to store instructions, programs, code, code sets, or instruction sets. The memory may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor. The memory, as a computer storage medium, may include an operating system, a network communication module, a user interface module, and application programs. The processor can be used to call the application programs stored in the memory and execute the steps of the low-frequency transformer proxy model construction method mentioned in the foregoing embodiments.

[0068] Example 4: This embodiment provides a computer-readable storage medium storing instructions that, when executed on a computer or processor, cause the computer or processor to perform the above-described instructions. Figure 1 One or more steps in the illustrated embodiment. If the constituent modules of the above-described electronic device are implemented as software functional units and sold or used as independent products, they can be stored in the computer-readable storage medium.

[0069] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this specification are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., Digital Versatile Discs (DVDs)), or semiconductor media (e.g., Solid State Disks (SSDs)).

[0070] Those skilled in the art will understand that all or part of the processes in the method of Embodiment 1 described above can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks. Unless otherwise specified, the technical features of this embodiment and the implementation scheme can be combined arbitrarily.

[0071] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0072] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0073] The above description is merely an exemplary embodiment of the present invention and should not be construed as limiting the scope of the invention. Any equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of embodiments of the invention upon considering the specification and practicing the disclosure herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not described herein. The specification and embodiments are to be considered exemplary only, and the scope and spirit of the invention are defined by the claims.

Claims

1. A method for constructing a low-frequency transformer surrogate model based on physical feature guidance and Bayesian active learning, characterized in that, Including the following steps: S1. Extract the key structural parameters of the low-frequency transformer and set the initial value range of each variable to construct a global parameter space; S2. Establish a multiphysics finite element model for electromagnetic simulation of low-frequency transformers; S3. Based on low-frequency physical laws, the magnetic flux density critical region and / or heat dissipation bottleneck region are defined as high-risk performance areas within the global parameter space. S4. Use Latin hypercube sampling to generate an initial sample set from the global parameter space, call the multiphysics finite element model to calculate the corresponding real response value, and construct the initial Kriging proxy model. S5. Based on whether the unresolved sample points in the global parameter space are located in the high-risk performance region, set a penalty weight, introduce the penalty weight into the active learning point addition criterion function, and solve for the maximum value of the function in the global parameter space as the optimal sampling point. S6. Calculate the true response value of the optimal sampling point using the multiphysics finite element model, add it to the training sample set, and update the Kriging proxy model. S7. Determine whether the prediction accuracy of the updated surrogate model in the high-risk performance region meets the preset convergence condition. If it does, output the current surrogate model; otherwise, return to step S5 to continue iterating.

2. The method for constructing a low-frequency transformer proxy model according to claim 1, characterized in that: The key structural parameters of the low-frequency transformer include the core column diameter, yoke width, upper window clearance, lower window clearance, phase center distance, and distance from the core end to the boundary.

3. The method for constructing a low-frequency transformer proxy model according to claim 2, characterized in that, Step S3 is as follows: When the combination of parameters in the key structural parameters of the low-frequency transformer causes the magnetic flux density at the corner inside the core column to be within a preset low-frequency critical magnetic flux range, the spatial domain corresponding to the combination is determined to be the critical region of magnetic flux density. When the combination of parameters in the key structural parameters of the low-frequency transformer causes the insulating oil flow rate at the winding gap to be lower than the preset flow rate threshold and the local heat generation rate to be higher than the preset heat generation rate threshold, the spatial domain corresponding to the combination is determined as the heat dissipation bottleneck area.

4. The method for constructing a low-frequency transformer proxy model according to claim 1, characterized in that: The actual response values ​​include the core volume, maximum magnetic flux density, and core loss of the low-frequency transformer.

5. The method for constructing a low-frequency transformer proxy model according to claim 1, characterized in that, Step S5 is as follows: The expected improvement amount and feasibility probability of the unresolved sample points are calculated using the posterior prediction variance of the Kriging surrogate model. Construct a physical feature identification function and set corresponding penalty weights based on whether the unresolved sample points in the global parameter space are located in the high-risk performance region; Multiply the feasibility probability, the expected improvement amount, and the physical feature identification function with the set penalty weight to obtain the active learning addition criterion function; The optimal sampling point is obtained by using a genetic algorithm to find the maximum value of the active learning addition criterion function in the global parameter space.

6. The method for constructing a low-frequency transformer proxy model according to claim 5, characterized in that, The rules for setting the penalty weights are as follows: When the unresolved sample point is located within the high-risk performance region, the physical feature identification function is set to take an amplified penalty weight; When the pending sample point is located outside the high-risk performance region, the physical feature identification function is set to have a reduced weight.

7. The method for constructing a low-frequency transformer proxy model according to claim 1, characterized in that, The preset convergence condition in step S7 is: In multiple iterations, the relative error of the surrogate model in predicting the maximum magnetic flux density of newly added sample points in the high-risk performance region and the relative error in predicting the maximum temperature rise of the winding are both less than the preset tolerance.

8. A system for constructing a low-frequency transformer surrogate model based on physical feature guidance and Bayesian active learning, characterized in that, This method is used to implement the low-frequency transformer proxy model construction method as described in any one of claims 1 to 7.

9. An electronic device, the electronic device comprising a memory, a processor, and a computer program, characterized in that, When the computer program is executed by the processor, it implements the low-frequency transformer proxy model construction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the low-frequency transformer proxy model construction method as described in any one of claims 1 to 7.

Images