A method and system for optimizing the design of a magnetic integrated transformer
By using a graph neural network surrogate model to guide the genetic algorithm, the problems of high computational cost and low efficiency in the design of magnetic integrated transformers are solved. This enables efficient multi-objective optimization under multiple constraints, improving the convergence speed and efficiency of the design.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUILIN UNIV OF ELECTRONIC TECH
- Filing Date
- 2026-03-19
- Publication Date
- 2026-05-29
AI Technical Summary
Existing technologies for designing magnetically integrated transformers suffer from high computational costs, low efficiency, and an inability to effectively utilize historical design experience, making it difficult to achieve efficient multi-objective optimization under multiple constraints.
A graph neural network (GNN) surrogate model is used to guide the genetic algorithm. By constructing a graph neural network model, the implicit mapping relationship between the design parameters of the magnetic integrated transformer and the target value is learned. Combined with the non-dominated sorting genetic algorithm (NSGA-III), the optimization design is carried out without relying on time-consuming electromagnetic simulation.
It achieves rapid multi-objective optimization of magnetic integrated transformers, reduces computational costs, improves convergence speed and optimization efficiency, and can learn design experience from historical data to guide the optimization direction.
Smart Images
Figure CN122113657A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of transformer optimization design, and more specifically, relates to a method and system for optimizing the design of magnetically integrated transformers. Background Technology
[0002] In the field of power electronic converters pursuing high power density and high efficiency, magnetic integration technology, by combining multiple magnetic components into a single magnetic core structure, has become a key direction for reducing system size and losses. However, the design of magnetically integrated transformers involves complex topologies with strong coupling of multiple variables, such as winding arrangement, magnetic circuit coupling, and leakage inductance control. Its parameter space is huge, and there are competing relationships between objectives (such as cost, loss, and size). Traditional design methods have low optimization efficiency and are difficult to find global optimization.
[0003] Existing technologies typically employ multi-objective evolutionary algorithms to solve such optimization problems, with the core being a mechanism to maintain the distribution of the solution set. For example, the Fast Elite Non-Dominated Sorting Genetic Algorithm (NSGA-II) ensures the convergence of solutions through fast non-dominated sorting and uses crowding distance to measure the sparsity of solutions within the target space, thereby selecting individuals with a wider distribution. However, this method experiences a sharp decrease in selection pressure when the number of objectives increases (typically >3), leading to a degradation in optimization performance. The Reference Point Guided Multi-Objective Non-Dominated Sorting Genetic Algorithm (NSGA-III), while inheriting the non-dominated sorting framework, replaces the crowding distance with a reference point-based selection mechanism. It pre-constructs a set of uniformly distributed reference points in the target space and maintains the diversity and uniform distribution of the solution set in the high-dimensional target space by associating population individuals with the nearest reference point and calculating the number of individuals associated with each reference point.
[0004] However, the above method has significant shortcomings in practical applications: in each iteration, high-fidelity electromagnetic field simulations must be called to calculate target values such as losses and costs for a large number of candidate structures, resulting in extremely high computational costs and low efficiency in the overall optimization process. At the same time, the algorithm only processes input parameters and output target values, and cannot directly learn and internalize the implicit mapping relationship between the complex physical fields and topological structures inside the magnetic components from the data. Therefore, each optimization is almost an exploration from scratch, and historical design experience cannot be utilized.
[0005] Therefore, how to achieve multi-objective optimization design of magnetic integrated transformers under multiple constraints without relying on a large amount of time-consuming electromagnetic simulation is an urgent problem to be solved. Summary of the Invention
[0006] To address the shortcomings of existing technologies, the purpose of this application is to provide a method and system for optimizing the design of magnetic integrated transformers, which can achieve multi-objective optimization design of magnetic integrated transformers under multiple constraints without relying on a large amount of time-consuming electromagnetic simulation.
[0007] To achieve the above objectives, in a first aspect, this application provides an optimized design method for a magnetically integrated transformer, comprising the following steps: S10 encodes the design parameters of the magnetic integrated transformer into individuals, generates an initial population, and sets multiple optimization objectives including cost and loss; S20, Construct a graph neural network model, which is used to output parameters to guide genetic operations based on the design parameters of the parent individuals; S30, based on the current population, wherein the current population at the first iteration is the initial population, calculate the target value of each individual in the current population and perform non-dominated sorting, train the graph neural network model using the high-quality individuals obtained by sorting, and perform genetic operations on the current population using the parameters output by the trained graph neural network model to guide genetic operations, thereby generating the offspring population; S40, determine whether the local search condition is met. If it is met, perform a local search on the offspring population and merge the offspring population after the local search with the parent population that generated the offspring population. If it is not met, directly merge the offspring population with the parent population that generated the offspring population. S50: Standardize the target value of the merged population, associate individuals with preset reference points, and select individuals to update the current population based on the association between individuals and reference points. S60: Determine if the termination condition is met. If not, return to step S30 to continue iterating. If the condition is met, output the optimization result.
[0008] As a further preferred embodiment, in step S20, the constructed graph neural network model includes a shared encoder, a feature fusion layer, and three parallel output heads; The shared encoder uses a linear layer, ReLU activation function, and Dropout regularization technique to standardize the feature extraction of the design parameters of the input parent individual; The feature fusion layer is used to integrate the genetic information of different individuals extracted by the shared encoder; The three parallel output heads are a crossover head, a mutation head, and a diversity head. The crossover head outputs a 30-dimensional crossover probability, the mutation head outputs a 1-dimensional mutation probability, and the diversity head outputs a 1-dimensional diversity score.
[0009] As a further preferred embodiment, in step S30, when training the graph neural network model using the high-quality individuals obtained from the ranking, the loss function used simultaneously optimizes the accuracy of crossover probability and mutation probability prediction, and emphasizes the exploration of diverse regions.
[0010] As a further preferred embodiment, in step S30, the non-dominated sort is a fast non-dominated sort.
[0011] As a further preferred embodiment, in step S40, associating individuals with preset reference points specifically includes: pre-constructing a set of uniformly distributed reference points in the target space, associating population individuals with the nearest reference point and calculating the number of individuals associated with that reference point, in order to maintain the diversity and uniform distribution of the solution set in the high-dimensional target space.
[0012] As a further preferred embodiment, in step S10, the design parameters of the magnetic integrated transformer include at least the core size, the number of winding turns, and the wire diameter; the magnetic integrated transformer adopts an EI-shaped iron core, and leakage magnetic laminations for integrating resonant inductance are set on the iron core structure, and the size of the resonant inductance is controlled by adjusting the size parameters of the leakage magnetic laminations; The cross-sectional area of the magnetic circuit formed by the leakage magnetic laminations Determined by the following formula:
[0013] in, This is the magnetic flux density value of the leakage flux shunt, which is less than the saturation magnetic flux density of the iron core.
[0014] As a further preferred embodiment, in step S30, the genetic operation includes selection, crossover, and mutation.
[0015] As a further preferred embodiment, in step S60, the termination condition is reaching a preset maximum number of iterations.
[0016] As a further preferred embodiment, in step S60, the output optimization result is the Pareto optimal solution set obtained after non-dominated sorting.
[0017] Secondly, this application provides a magnetic integrated transformer optimization design system for implementing the magnetic integrated transformer optimization design method as described in any one of the above statements, comprising: The initialization module is used to encode the design parameters of the magnetic integrated transformer into individuals, generate an initial population, and set multiple optimization objectives including cost and loss. A proxy model building module is used to build a graph neural network model, which is used to output parameters to guide genetic operations based on the design parameters of the parent individual; An evolutionary optimization module is used to calculate the target value of each individual in the current population as a basis, wherein the current population at the first iteration is the initial population, and perform non-dominated sorting. The high-quality individuals obtained by sorting are used to train the graph neural network model, and the parameters output by the trained graph neural network model to guide the genetic operation are used to perform genetic operation on the current population to generate offspring population. The population merging module is used to determine whether the local search conditions are met. If they are met, a local search is performed on the offspring population, and the offspring population after the local search is merged with the parent population that generated the offspring population. If the conditions are not met, the offspring population is directly merged with the parent population that generated the offspring population. The population update module is used to standardize the target value of the merged population, associate individuals with preset reference points, and select individuals to update the current population based on the association between the individual and the reference point. The iteration control module is used to determine whether the termination condition is met. If not, it continues to call the evolution optimization module, population merging module, and population update module for iteration. If the condition is met, it outputs the optimization result.
[0018] The beneficial effects of this application are: (1) The evaluation is greatly accelerated by the proxy model, and the fast forward prediction of the graph neural network is used to replace the time-consuming electromagnetic simulation, reducing the single evaluation time from hours to seconds, breaking through the optimization calculation bottleneck; (2) The algorithm can learn the complex mapping between electromagnetic field and topology from the data, guide the search direction, make up for the lack of domain knowledge in traditional algorithms, and improve the convergence speed of the algorithm. Attached Figure Description
[0019] Figure 1 This is an isometric view of the magnetic integrated transformer provided in the embodiments of this application; Figure 2 This is a cross-sectional view of the magnetic integrated transformer provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of the magnetic integrated transformer provided in the embodiments of this application; Figure 4 This is the equivalent magnetic circuit diagram of the magnetic integrated transformer provided in the embodiments of this application; Figure 5 This is a diagram of the GNN neural network structure provided in the embodiments of this application; Figure 6 This is a flowchart of the GNN-NSGA-III algorithm provided in the embodiments of this application; Figure 7 This is a diagram showing the ZDT1 supervolume test results provided in an embodiment of this application; Figure 8 This is a diagram showing the ZDT2 supervolume test results provided in an embodiment of this application; Figure 9 This is a diagram showing the ZDT3 supervolume test results provided in an embodiment of this application; Figure 10 This is a flowchart of the GNN-NSGA-II algorithm provided in the embodiments of this application. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0021] This embodiment proposes a GNN-NSGA-III optimization algorithm for magnetically integrated transformers, such as... Figure 6 As shown, the specific solution is as follows: 1) Problem Modeling and Initialization Key transformer parameters (such as core size, number of winding turns, wire diameter, etc.) are encoded as individual genes. The problem is set as minimizing total cost and minimizing total losses (including copper and iron losses), constituting a two-objective optimization problem. Within the variable constraints, an improved strategy is used to generate the initial population, ensuring coverage of the design space. The magnetically integrated transformer uses an EI-shaped iron core, and its three-dimensional isometric view is shown below. Figure 1 As shown, the cross-sectional view is as follows Figure 2 As shown.
[0022] Unlike traditional transformers, magnetically integrated transformers incorporate leakage flux laminations into the core structure to integrate a resonant inductor. The magnitude of the resonant inductance is controlled by adjusting the dimensions of these leakage flux laminations. A schematic diagram of a magnetically integrated transformer is shown below. Figure 3 As shown, where, U 1 represents the effective value of the rated voltage of the primary winding. N 1 represents the number of turns in the primary winding. U 2 represents the effective value of the rated voltage of the secondary high-voltage winding. N 2 represents the number of turns in the secondary high-voltage winding. U 3 represents the effective value of the rated voltage of the secondary low-voltage winding. N 3 represents the number of turns in the secondary low-voltage winding. C This is the resonant capacitor. The equivalent magnetic circuit of the transformer is as follows: Figure 4 As shown in the figure, N 1 I 1 represents the magnetomotive force of the primary winding. N 2 I 2 represents the magnetomotive force of the secondary winding. Φ 1 represents the magnetic flux of the primary winding core branch. Φ 2 represents the magnetic flux of the secondary winding core branch. R m1 The magnetic reluctance of the core branch containing the primary winding. R m2 The magnetic reluctance of the core branch containing the secondary winding. R m3 For the core reluctance of the leakage magnetic lamination magnetic shunt, R m1 The air gap reluctance is for the magnetic shunt of the leakage flux stack. The cross-sectional area of the magnetic path formed by the leakage flux stack. Sδ Determined by the following formula:
[0023] In the formula, B δ The flux density value of the leakage flux shunt should be designed to be less than the saturation flux density of the iron core.
[0024] 2) Construction of GNN Proxy Model A diversity-aware GNN model is constructed. Its input consists of design parameters from two parent individuals. An encoder extracts features and outputs three key pieces of information: crossover probability, mutation probability, and diversity score. During algorithm execution, non-dominated solutions (high-quality design schemes) are dynamically collected as training data. During training, the loss function simultaneously optimizes the accuracy of crossover and mutation predictions and emphasizes the exploration of diversity regions, enabling the GNN to learn to generate high-quality and diverse new design schemes.
[0025] The specific improvements are reflected in the following aspects: Uniformity of feature extraction: By introducing a "shared encoder", using linear layers, ReLU activation function and Dropout (0.2) regularization, the standardized extraction of features from different parent inputs is achieved, which effectively reduces the number of model parameters and enhances generalization ability.
[0026] Deep feature fusion: A dedicated “feature fusion” layer can be set up to integrate genetic information from different individuals, providing richer context for subsequent decision-making.
[0027] Decoupling and diversification of output: The most significant improvement is the design of three parallel output heads—"cross head", "mutation head" and "diversity head"—which independently output 30-dimensional crossover probability, 1-dimensional mutation probability and 1-dimensional diversity score, respectively.
[0028] This design allows the neural network to not only output a single metric, but also simultaneously guide multiple core aspects of the evolutionary algorithm (crossover, mutation, and diversity assessment), thus making parameter tuning more adaptive and precise. The GNN neural network structure is as follows: Figure 5 As shown.
[0029] 3) GNN-guided intelligent optimization algorithm We leverage Generative Neural Networks (GNNs) to learn implicit knowledge about magnetically integrated transformer design from historical optimization data and translate it into intelligent guidance for the NSGA-III search process. We calculate the cost and loss of all individuals in the current population. We perform fast non-dominated sorting to categorize solutions into different tiers. We periodically collect data from high-quality solutions to train the GNN, and in subsequent genetic operations, we combine the crossover and mutation probabilities predicted by the GNN to generate new individuals.
[0030] After the program starts, it first initializes the parameters and generates reference points (to guide the population towards a uniform distribution along the Pareto front). Then it enters the main iterative loop, first determining if the maximum number of iterations has been reached. In each iteration, the algorithm first evaluates the fitness of the current population and determines the rank of individuals using non-dominated sorting. Next comes a crucial conditional branch: GNN model intervention: If the training conditions are met, GNN model training is performed, and the neural network learns the features of the solution space to help generate higher-quality offspring; if the conditions are not met, training is skipped and offspring are generated directly.
[0031] Offspring generation: Generate a new batch of offspring population by combining model output or traditional methods.
[0032] After generating offspring, the program introduces a local search mechanism. If the local search conditions are met, the individuals are fine-tuned (local search), and the results are merged with the original population. If the conditions are not met, the program directly proceeds to the subsequent standardization process. To maintain the competitiveness and diversity of the population, the program standardizes the target value and associates individuals with reference points, counting the number of associations at each reference point. Subsequently, through classic genetic algorithm steps—selection, crossover, and mutation—mutated individuals are generated and merged with their parents to update the population. When the number of iterations reaches a preset upper limit, the program exits the loop. A final non-dominated sort is performed, and adaptive weights are calculated to optimize the performance of the final solution. Finally, the results are output, and the program terminates. The overall algorithm flowchart is as follows: Figure 8 As shown.
[0033] In this embodiment, the Genetic Neural Network (GNN) learns the complex mapping between design parameters and performance from historical data, becoming a high-performance internal agent. This allows for rapid evaluation of candidate solutions and intelligent guidance of genetic operations towards more promising directions. This improves the efficiency and effectiveness of multi-objective optimization, enabling the rapid acquisition of transformer designs that balance cost and losses.
[0034] The algorithm was run for 500 generations using the standard test functions ZDT1, ZDT2, and ZDT3, and its performance was evaluated using two metrics: hypervolume and inverse generation distance (IGD).
[0035] The ZDT1 function is a biobjective optimization function with a global optimum and a local optimum. The Pareto front is convex and continuous. The function is defined as follows:
[0036] The ZDT2 function is also a biobjective optimization function, possessing a global optimum and a local optimum. The Pareto front is non-convex and continuous. The function is defined as follows:
[0037] The ZDT3 function is a biobjective optimization function with a global optimum and a local optimum. The Pareto front consists of multiple discontinuous convex segments. The function is defined as follows:
[0038] Hypervolume is a core comprehensive metric for evaluating the performance of multi-objective optimization algorithms; it measures the volume of the solution set in the objective space. Specifically, it is defined by the size of the space enclosed by the reference point and the non-dominated solutions in the Pareto solution set. For these three test functions, there exists an ideal value of 0.24 for the hypervolume metric. The backward generation distance (IGD) measures how close the algorithm's solution set is to the true Pareto front. A smaller IGD value indicates that the solution set is closer to the true Pareto front. Hypervolume results are shown below. Figure 7 , Figure 8 and Figure 9 As shown, the algorithm combined with GNN has a significant improvement in convergence speed compared to the standard non-dominated genetic algorithm.
[0039] The data comparisons for the three indicators are shown in Tables 1, 2, and 3: Table 1 ZDT1 Test Results
[0040] Table 2 ZDT2 Test Results
[0041] Table 3 ZDT3 Test Results
[0042] The results show that, for the three test functions, the IGD index of the algorithm combined with GNN is improved compared with the standard non-dominated genetic algorithm, but the overall results are not significantly different. As for the hypervolume index, the algorithm combined with GNN has a significant improvement in convergence speed. The combined algorithm reaches the ideal value of hypervolume within 100 generations, while the standard non-dominated genetic algorithm only reaches the ideal value after 300 generations. The ZDT3 test results even failed to reach the ideal value after running for 500 generations.
[0043] Alternative solutions exist to the above technical approach: the idea of combining GNN with NSGA-III can be applied to related algorithms such as NSGA-II. GNN can guide NSGA-II in optimizing the magnetic integrated transformer. The combined program flow is as follows: Figure 10 As shown.
[0044] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A method for optimizing the design of a magnetically integrated transformer, characterized in that, Includes the following steps: S10 encodes the design parameters of the magnetic integrated transformer into individuals, generates an initial population, and sets multiple optimization objectives including cost and loss; S20, Construct a graph neural network model, which is used to output parameters to guide genetic operations based on the design parameters of the parent individuals; S30, based on the current population, wherein the current population at the first iteration is the initial population, calculate the target value of each individual in the current population and perform non-dominated sorting, train the graph neural network model using the high-quality individuals obtained by sorting, and perform genetic operations on the current population using the parameters output by the trained graph neural network model to guide genetic operations, thereby generating the offspring population; S40, determine whether the local search condition is met. If it is met, perform a local search on the offspring population and merge the offspring population after the local search with the parent population that generated the offspring population. If it is not met, directly merge the offspring population with the parent population that generated the offspring population. S50: Standardize the target value of the merged population, associate individuals with preset reference points, and select individuals to update the current population based on the association between individuals and reference points. S60: Determine if the termination condition is met. If not, return to step S30 to continue iterating. If the condition is met, output the optimization result.
2. The optimized design method for magnetic integrated transformers as described in claim 1, characterized in that, In step S20, the constructed graph neural network model includes a shared encoder, a feature fusion layer, and three parallel output heads; The shared encoder uses a linear layer, ReLU activation function, and Dropout regularization technique to standardize the feature extraction of the design parameters of the input parent individual; The feature fusion layer is used to integrate the genetic information of different individuals extracted by the shared encoder; The three parallel output heads are a crossover head, a mutation head, and a diversity head. The crossover head outputs a 30-dimensional crossover probability, the mutation head outputs a 1-dimensional mutation probability, and the diversity head outputs a 1-dimensional diversity score.
3. The optimized design method for magnetic integrated transformers as described in claim 1, characterized in that, In step S30, when training the graph neural network model using the high-quality individuals obtained from the ranking, the loss function used simultaneously optimizes the accuracy of crossover probability and mutation probability prediction, and emphasizes the exploration of diverse regions.
4. The optimized design method for magnetic integrated transformers as described in claim 1, characterized in that, In step S30, the non-dominated sort is a fast non-dominated sort.
5. The optimized design method for magnetic integrated transformers as described in claim 1, characterized in that, In step S40, associating individuals with preset reference points specifically includes: pre-constructing a set of uniformly distributed reference points in the target space, associating population individuals with the nearest reference point and calculating the number of individuals associated with that reference point, in order to maintain the diversity and uniform distribution of the solution set in the high-dimensional target space.
6. The optimized design method for a magnetically integrated transformer as described in claim 1, characterized in that, In step S10, the design parameters of the magnetic integrated transformer include at least the core size, the number of winding turns, and the wire diameter; the magnetic integrated transformer adopts an EI-shaped iron core, and leakage magnetic laminations for integrating resonant inductance are set on the iron core structure, and the size of the resonant inductance is controlled by adjusting the size parameters of the leakage magnetic laminations; The cross-sectional area of the magnetic circuit formed by the leakage magnetic laminations Determined by the following formula: in, This is the magnetic flux density value of the leakage flux shunt, which is less than the saturation magnetic flux density of the iron core.
7. The optimized design method for magnetic integrated transformers as described in claim 1, characterized in that, In step S30, the genetic operations include selection, crossover, and mutation.
8. The optimized design method for a magnetically integrated transformer as described in claim 1, characterized in that, In step S60, the termination condition is reaching the preset maximum number of iterations.
9. The optimized design method for a magnetically integrated transformer as described in claim 1, characterized in that, In step S60, the output optimization result is the Pareto optimal solution set obtained after non-dominated sorting.
10. A magnetic integrated transformer optimization design system, characterized in that, The method for optimizing the design of a magnetically integrated transformer as described in any one of claims 1 to 9 includes: The initialization module is used to encode the design parameters of the magnetic integrated transformer into individuals, generate an initial population, and set multiple optimization objectives including cost and loss. A proxy model building module is used to build a graph neural network model, which is used to output parameters to guide genetic operations based on the design parameters of the parent individual; An evolutionary optimization module is used to calculate the target value of each individual in the current population as a basis, wherein the current population at the first iteration is the initial population, and perform non-dominated sorting. The high-quality individuals obtained by sorting are used to train the graph neural network model, and the parameters output by the trained graph neural network model to guide the genetic operation are used to perform genetic operation on the current population to generate offspring population. The population merging module is used to determine whether the local search conditions are met. If they are met, a local search is performed on the offspring population, and the offspring population after the local search is merged with the parent population that generated the offspring population. If the conditions are not met, the offspring population is directly merged with the parent population that generated the offspring population. The population update module is used to standardize the target value of the merged population, associate individuals with preset reference points, and select individuals to update the current population based on the association between the individual and the reference point. The iteration control module is used to determine whether the termination condition is met. If not, it continues to call the evolution optimization module, population merging module, and population update module for iteration. If the condition is met, it outputs the optimization result.