A multi-physical field collaborative optimization design method and system for a solid-state transformer based on an image data stream

Abstract

A kind of solid-state transformer multi-physical field collaborative optimization design method based on image data stream, comprising the following steps: obtaining the design parameters of solid-state transformer input by user;Through image generation engine, design parameters are converted into at least including geometric image, waveform image and multi-physical field label image image data;The deep learning agent network is trained using the image data of geometric image, waveform image and multi-physical field label image, and the corresponding physical field distribution image and performance index are predicted according to geometric image and waveform image;In image space, with geometric image as individual, generate Pareto front solution set;Receive the preferred geometric image selected by user from Pareto front solution set, obtain the final design scheme.The seamless fusion of optimization and high-fidelity simulation is realized, and an efficient, accurate and open optimization method is provided for the overall multi-physical field collaborative optimization of solid-state transformer.

Description

Technical Field

[0001] This invention relates to the field of transformer manufacturing technology, and in particular to a multi-physics collaborative optimization design method and system for solid-state transformers based on image data streams. Background Technology

[0002] Solid-state transformers (SSTs), as important devices in the field of power electronics, have received widespread attention in recent years for their advantages such as high power density, strong controllability, and compact size, particularly in applications like smart grids, distributed energy resources, and electric transportation. Current SST research and design primarily rely on parametric modeling and phased simulation verification, with a typical technical process including the following steps: Designers typically select standardized component parameters from existing component libraries, magnetic core libraries, and heat sink libraries based on experience. For example, they might choose a PQ40 magnetic core, a 12:6 winding ratio, and a heat sink with dimensions of 100×80×50 mm to form an initial design scheme. Subsequently, optimization algorithms such as genetic algorithms are used to optimize these parameters with performance indicators such as efficiency and power density as objectives, generating several candidate design parameter sets. For each candidate parameter set, circuit simulation, finite element analysis (FEA), and thermal fluid dynamics (CFD) simulations are performed to verify whether its performance meets the design requirements. When the simulation results do not meet expectations, the design parameters or constraints need to be manually adjusted, and the optimization and simulation process is repeated until the design converges.

[0003] However, existing technologies still have the following shortcomings: First, design space is limited. Parametric modeling can only describe limited geometric topologies, such as standard core shapes and regular winding arrangements, making it difficult to express innovative structures such as asymmetric core air gaps, irregular windings, or biomimetic heat sinks, resulting in many potential high-performance designs remaining unexplored. Second, there is a disconnect between simulation and optimization processes. Optimization algorithms are usually based on simplified models, such as empirical loss formulas or concentrated thermal resistance networks, failing to fully consider the actual physical field distribution; while high-fidelity simulation is only used for verification after optimization, and manual rollback is required if problems are found, resulting in long development cycles and high iteration costs. Third, multi-physics information is lost during parameter transfer. For example, spatial information such as loss density distribution and temperature field distribution generated by electromagnetic simulation are compressed into average values ​​or total power. After being transferred to thermal simulation, the accuracy of thermal analysis decreases, making it difficult to accurately predict failure risks such as local hot spots and thermal stress concentration. Finally, knowledge reuse is difficult. Each new design requires repeated high-cost simulations, and historical design data is difficult to effectively guide new designs, making it difficult to transform R&D experience into reusable knowledge.

[0004] In summary, existing SST design methods have limitations in terms of design space, optimization efficiency, multiphysics coupling, and knowledge reuse, and there is an urgent need to propose new design methods to improve R&D efficiency and design performance. Summary of the Invention

[0005] To address the aforementioned shortcomings, the present invention aims to propose a multi-physics collaborative optimization design method for solid-state transformers based on image data streams, thereby solving the limitations of existing SST design methods in terms of design space, optimization efficiency, multi-physics coupling, and knowledge reuse.

[0006] To achieve this objective, the present invention adopts the following technical solution: a multi-physics collaborative optimization design method for solid-state transformers based on image data streams, comprising the following steps: Step S1: Obtain the solid-state transformer design parameters input by the user; Step S2: Using an image generation engine, convert the design parameters into image data that includes at least geometric images, waveform images, and multiphysics label images; Step S3: Train the deep learning agent network using image data of geometric images, waveform images, and multi-physics field labeled images, and predict the corresponding physical field distribution images and performance indicators based on the geometric images and waveform images; Step S4: In the image space, taking the geometric image as an individual, a multi-objective optimization algorithm is used in combination with the prediction results of the deep learning agent network to perform iterative optimization and generate the Pareto front solution set; Step S5: Receive the preferred geometric image selected by the user from the Pareto front solution set and obtain the final design scheme.

[0007] Preferably, the step of obtaining the geometric image in step S2 is as follows: Based on the design parameters, the grayscale value of the material is selected as the mapping condition, and different pixel values ​​are assigned to each coordinate point in the specified two-dimensional image. The formula for assigning pixel values ​​is as follows: ; in This represents the pixel value at (x, y) coordinates. The indicator function is used when (x, y) is located in the k-th geometric region. At that time, The value is 1, otherwise Take the value 0. This represents the material's grayscale value mapping, where K is the total number of geometric regions. The steps for obtaining the waveform image in step S2 are as follows: The time-domain voltage waveform of a solid-state transformer is obtained by circuit simulation. The time-domain voltage waveform is binarized to generate a first dot plot; The Bresenham algorithm is applied to the first point image to connect adjacent points, where the pixel value of the connection between adjacent points is 1. The steps for acquiring image data for multiphysics label images in step S2 are as follows: The mesh node data output from high-fidelity electromagnetic, thermal, and stress finite element simulations are converted into a physical field image with a regular mesh using barycentric coordinate interpolation or bilinear interpolation, where the barycentric coordinate interpolation satisfies: ; in For the field value corresponding to the corresponding grid vertex, ( () represents the centroid coordinates of the triangular mesh; The multiphysics label image is obtained by interpolating the normalized barycenter coordinates. , These represent the minimum and maximum values ​​for the entire field. Based on the multiphysics label image, a loss density map is obtained through simulation processing. Temperature field diagram and boundary condition images .

[0008] Preferably, the training process of the deep learning agent network includes: Step S31: Receive the geometric image using the first network, and obtain the loss density map prediction value from the pre-trained loss network NEM using the geometric image. ; Predicted values ​​from the loss density map and loss density map Obtain the first loss; The formula for obtaining the first loss is as follows: ; Where H and W are the length and width of the geometric image, respectively. The similarity coefficient ranges from 0.01 to 0.05. To learn perceptual image patch similarity; Step S32: Use the second network receive loss density map and boundary condition images Through the loss density map and boundary condition images Obtain temperature field map predictions from a pre-trained temperature network (Nthermal). ; Predicted values ​​from the temperature field map and temperature field diagram Obtain a second loss; The formula for obtaining the second loss is as follows: ; Where H and W are the length and width of the geometric image, respectively. The similarity coefficient ranges from 0.1 to 0.2. The gradient of the predicted values ​​in the temperature field map. This represents the gradient of the temperature field diagram; Step S33: Receive waveform images using a third network, and use the waveform images to perform online performance regression on predicted EMI values ​​in the Nperf network. By sharing the bottleneck features of U-Net, the outputs of the first network and the second network are connected to obtain the efficiency prediction value and the power density prediction value. The third loss is obtained by using the predicted EMI value, efficiency prediction value, and power density prediction value; The formula for obtaining the third loss is as follows: ; in These represent the task weights, with values ​​ranging from 1 to 2, 0.1 to 0.3, and 0.5 to 0.8, respectively. These are the efficiency prediction value and the efficiency input by the user, respectively. These are the predicted power density and the power density input by the user, respectively. Step S34: Obtain the total loss by summing the first loss, the second loss, and the third loss. Apply the total loss to the Adam optimizer and set the learning rate to [value missing]. Train 50-200 times.

[0009] Preferably, the step S4, which uses a multi-objective optimization algorithm combined with the prediction results of the deep learning agent network for iterative optimization, is as follows: Step S41: Randomly select a first number of geometric images from the knowledge base; Randomly generate a second number of geometric images; The solid-state transformer design parameters input by the user are randomly perturbed, and a third number of geometric images are generated; Step S42: Using the geometric image generated in step S41 as the iterative individual, and performing the following constraints to select the geometric image for iteration; The constraint is that the predicted value of the largest temperature field in the geometric image cannot be greater than the first temperature value; Step S43: Perform population update on the iterative geometric image; The population update operation is as follows: Two parent geometric images are randomly selected, and their pixel values ​​are swapped within a random rectangular region to generate a child image. Add Gaussian noise N(0,0.05) to each pixel with a probability of 0.1, or perform morphological dilation or erosion operations on the geometric image; And retain the top 10% of nondominated solutions that should be optimal for the next iteration; The expression for the appropriate value is: ; ; ; ; ; in Represents the efficiency prediction value of the k-th geometric image. Represents the predicted power density value of the k-th geometric image. Let be the EMI value of the k-th geometric image. This represents the area in the k-th geometric image where the predicted temperature field value is greater than the second temperature threshold, where the second temperature threshold is less than the first temperature threshold. These represent the standard deviation and mean of the temperature field, respectively. Step S44: Iteratively execute steps S42~43 until the number of iterations equals the iteration threshold, or the Pareto front is stable.

[0010] Preferably, the steps for obtaining the final design scheme in step S5 are as follows: The ANSYS Maxwell and Fluent are used to perform finite element simulations of electromagnetic and thermal fields on the 3D model corresponding to the user-selected preferred geometry image. The accurate loss distribution, temperature field and EMI are obtained and compared with the prediction results of the deep learning agent network to generate a verification report. The system performs edge detection and contour extraction on the selected geometric images, identifies the core boundaries and calculates their dimensions, identifies the center coordinates and area of ​​the winding region, calculates the number of turns and wire diameter, and outputs the identified geometric features as CAD models and parameter tables. The verification report, CAD model, and parameter table are sent to the user for final selection, resulting in the final design scheme.

[0011] A multi-physics collaborative optimization design system for solid-state transformers based on image data streams employs a multi-physics collaborative optimization design method for solid-state transformers based on image data streams, including: Data acquisition module: used to acquire solid-state transformer design parameters input by the user; The image generation module is used to convert design parameters into image data, including at least geometric images, waveform images, and multiphysics label images, through the image generation engine. The model prediction module is used to predict the corresponding physical field distribution image and performance indicators based on the geometric image and waveform image using a pre-trained deep learning agent network. The iterative module uses geometric images as individuals in the image space and employs a multi-objective optimization algorithm combined with the prediction results of the deep learning agent network to iteratively optimize and generate a Pareto front solution set. The selection module is used to receive the preferred geometric image selected by the user from the Pareto front solution set and obtain the final design scheme.

[0012] Preferably, the image generation module includes a first image submodule, a second image submodule, and a third image submodule; The first image submodule is used to select the gray value of the material as the mapping condition based on the design parameters, and assign different pixel values ​​to each coordinate point in the specified two-dimensional image. The second image submodule is used to perform circuit simulation on the solid-state transformer to obtain its time-domain voltage waveform. The time-domain voltage waveform is binarized to generate a first dot plot; The Bresenham algorithm is applied to the first point image to connect adjacent points, where the pixel value of the connection between adjacent points is 1. The third image submodule is used to convert the mesh node data output from high-fidelity electromagnetic, thermal, and stress finite element simulations into a physical field image with a regular mesh through barycentric coordinate interpolation or bilinear interpolation, wherein the barycentric coordinate interpolation satisfies: ; in For the field value corresponding to the corresponding grid vertex, ( () represents the centroid coordinates of the triangular mesh; The multiphysics label image is obtained by interpolating the normalized barycenter coordinates. , These represent the minimum and maximum values ​​for the entire field. Based on the multiphysics label image, a loss density map is obtained through simulation processing. Temperature field diagram and boundary condition images .

[0013] Preferably, the model prediction module includes a first training submodule, a second training submodule, a third training submodule, and a fourth training submodule; The first training submodule is used to receive the geometric image and obtain the loss density map prediction value in the pre-trained loss network NEM using the geometric image. ; Predicted values ​​from the loss density map and loss density map Obtain the first loss; The second training submodule is used to receive the loss density map. and boundary condition images Through the loss density map and boundary condition images Obtain temperature field map predictions from a pre-trained temperature network (Nthermal). ; The third training submodule is used to receive waveform images and predict EMI values ​​in the online performance regression network Nperf based on the waveform images. By sharing the bottleneck features of U-Net, the outputs of the first network and the second network are connected to obtain the efficiency prediction value and the power density prediction value. The third loss is obtained by using the predicted EMI value, efficiency prediction value, and power density prediction value; The fourth training submodule is used to sum the first loss, the second loss, and the third loss to obtain the total loss. The Adam optimizer is then applied using this total loss, and the learning rate is set to... Train 50-200 times.

[0014] Preferably, the iteration module includes a first iteration submodule, a second iteration submodule, a third iteration submodule, and a fourth iteration submodule: The first iteration submodule is used to randomly select a first number of geometric images from the knowledge base; Randomly generate a second number of geometric images; The solid-state transformer design parameters input by the user are randomly perturbed, and a third number of geometric images are generated; The first iteration submodule is used to select the geometric image to be iterated based on the geometric image generated in the first iteration submodule and to perform the following constraints to select the geometric image to be iterated. The constraint is that the predicted value of the largest temperature field in the geometric image cannot be greater than the first temperature value; The third iteration submodule is used to perform population updates on the iterative geometric image; The population update operation is as follows: Two parent geometric images are randomly selected, and their pixel values ​​are swapped within a random rectangular region to generate a child image. Add Gaussian noise N(0,0.05) to each pixel with a probability of 0.1, or perform morphological dilation or erosion operations on the geometric image; And retain the top 10% of nondominated solutions that should be optimal for the next iteration; The fourth iteration submodule is used to re-call the second and third iteration submodules until the number of iterations equals the iteration threshold, or the Pareto front is stable.

[0015] Preferably, the selection module performs the following steps: The ANSYS Maxwell and Fluent are used to perform finite element simulations of electromagnetic and thermal fields on the 3D model corresponding to the user-selected preferred geometry image. The accurate loss distribution, temperature field and EMI spectrum are obtained and compared with the prediction results of the deep learning agent network to generate a verification report. The system performs edge detection and contour extraction on the selected geometric images, identifies the core boundaries and calculates their dimensions, identifies the center coordinates and area of ​​the winding region, calculates the number of turns and wire diameter, and outputs the identified geometric features as CAD models and parameter tables. The verification report, CAD model, and parameter table are sent to the user for final selection, resulting in the final design scheme.

[0016] One of the above technical solutions has the following advantages or beneficial effects: the present invention significantly expands the design freedom, enabling the exploration of non-standardized, biomimetic, or irregular structures; it achieves seamless integration of optimization and high-fidelity simulation, and provides an efficient, accurate, and open optimization method for multi-physics collaborative optimization of solid-state transformers. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the structure of one embodiment of the method of the present invention.

[0018] Figure 2 This is a flowchart of one embodiment of the system of the present invention. Detailed Implementation

[0019] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0020] In the description of embodiments of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0021] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0022] like Figures 1-2 As shown, a multi-physics collaborative optimization design method for solid-state transformers based on image data streams includes the following steps: Step S1: Obtain the solid-state transformer design parameters input by the user; Step S2: Using an image generation engine, convert the design parameters into image data that includes at least geometric images, waveform images, and multiphysics label images; Step S3: Train the deep learning agent network using image data of geometric images, waveform images, and multi-physics field labeled images, and predict the corresponding physical field distribution images and performance indicators based on the geometric images and waveform images; Step S4: In the image space, taking the geometric image as an individual, a multi-objective optimization algorithm is used in combination with the prediction results of the deep learning agent network to perform iterative optimization and generate the Pareto front solution set; Step S5: Receive the preferred geometric image selected by the user from the Pareto front solution set and obtain the final design scheme.

[0023] Traditional parametric modeling can only describe limited geometric topologies such as standard core shapes and regular winding arrangements, limiting spatial design and preventing the exploration of many potential high-performance designs to obtain excellent optimization parameters. To address this design space limitation, step S2 of this invention first uses an image generation engine to generate high-fidelity geometric images. These geometric images are then used as design variables, freely expressing arbitrarily complex shapes such as asymmetric core air gaps, irregular winding arrangements, and biomimetic heatsink structures, greatly expanding the design space and enabling the exploration of numerous potential high-performance innovative structures. The image generation engine first generates a structure that most closely approximates the user-input solid-state transformer design parameters. However, this structure may not necessarily optimize the solid-state transformer after application. Therefore, a pre-trained deep learning agent network is needed to acquire data related to the generated geometric image. This data includes physical field distribution images and performance indicators, such as power density, efficiency, or EMI values. These data indicate the operating parameters of the solid-state transformer, and their quality can be judged based on the design parameters.

[0024] Therefore, in step S3, the image data of the geometric image, waveform image, and multiphysics label image generated in step S2 are used as training data for the deep learning agent network. After training, it can quickly predict the distribution image and accurate performance indicators close to the actual physical field during the optimization iteration process, and directly embed high-fidelity evaluation into the optimization loop without the need for manual adjustment of constraints and rerunning of simulations, thereby significantly shortening the R&D cycle and reducing iteration costs.

[0025] Although the deep learning agent network can determine whether the structure generated in step S2 is usable, the generated structure may not be the optimal solution. Therefore, in step S4, several random geometric images will be generated in the image space based on the design parameters. At the same time, several geometric images will be randomly selected from the knowledge base. These obtained geometric images will be input into the trained deep learning agent network for optimization iteration to select the Pareto front solution set. Although the Pareto front solution set is the optimal solution, it may not be suitable for the user. Therefore, it is necessary to send it to the user for confirmation and make the final selection.

[0026] This invention significantly expands design freedom, enabling the exploration of non-standardized, biomimetic, or irregular structures; it achieves seamless integration of optimization and high-fidelity simulation, and provides an efficient, accurate, and open optimization method for multi-physics collaborative optimization of solid-state transformers.

[0027] Preferably, the step of obtaining the geometric image in step S2 is as follows: Based on the design parameters, the grayscale value of the material is selected as the mapping condition, and different pixel values ​​are assigned to each coordinate point in the specified two-dimensional image. The formula for assigning pixel values ​​is as follows: ; in This represents the pixel value at (x, y) coordinates. The indicator function is used when (x, y) is located in the k-th geometric region. At that time, The value is 1, otherwise Take the value 0. This represents the material's grayscale value mapping, where K is the total number of geometric regions. For acquiring geometric images, this invention selects the grayscale value of the material as a mapping condition based on the design parameters. In the two-dimensional image, each coordinate point is assigned a pixel value. Specifically, an indicator function determines whether the point is located within the k-th geometric region and superimposes the corresponding material grayscale mapping value, thereby generating a grayscale image capable of representing arbitrarily complex geometric shapes (such as asymmetric air gaps, irregular windings, biomimetic heat sinks, etc.). This avoids the limitations of traditional parametric modeling, which can only use standard magnetic core shapes and regular arrangements, greatly expanding the design space.

[0028] The steps for obtaining the waveform image in step S2 are as follows: The time-domain voltage waveform of a solid-state transformer is obtained by circuit simulation. The time-domain voltage waveform is binarized to generate a first dot plot; The Bresenham algorithm is applied to the first point image to connect adjacent points, where the pixel value of the connection between adjacent points is 1. For waveform image acquisition, this method first performs circuit simulation on a solid-state transformer to obtain the time-domain voltage waveform, then binarizes the waveform to generate the first dot plot, and then uses the Bresenham algorithm to connect adjacent points and set the pixel value of the connection line to 1, thereby generating a clear and continuous binary waveform image. This processing can convert dynamic electrical signals in circuit simulation into image format.

[0029] The steps for acquiring image data for multiphysics label images in step S2 are as follows: The mesh node data output from high-fidelity electromagnetic, thermal, and stress finite element simulations are converted into a physical field image with a regular mesh using barycentric coordinate interpolation or bilinear interpolation, where the barycentric coordinate interpolation satisfies: ; in For the field value corresponding to the corresponding grid vertex, ( () represents the centroid coordinates of the triangular mesh; The multiphysics label image is obtained by interpolating the normalized barycenter coordinates. , These represent the minimum and maximum values ​​for the entire field. Based on the multiphysics label image, a loss density map is obtained through simulation processing. Temperature field diagram and boundary condition images .

[0030] The physical field image is converted into a regular grid through barycentric coordinate interpolation (for triangular meshes) or bilinear interpolation. The barycentric coordinate interpolation formula uses the field values ​​at the vertices of the triangles and the barycentric coordinates to obtain the field value at any point, and then normalizes it through the minimum and maximum values ​​of the entire field to obtain a standard image. The resulting labeled image fully preserves the spatial distribution information in the loss density map, temperature field map, and boundary condition image, such as the location of local hot spots and areas of thermal stress concentration, and not just the total power or average temperature.

[0031] Preferably, the training process of the deep learning agent network includes: Step S31: Receive the geometric image using the first network, and obtain the loss density map prediction value from the pre-trained loss network NEM using the geometric image. ; Predicted values ​​from the loss density map and loss density map Obtain the first loss; The formula for obtaining the first loss is as follows: ; Where H and W are the length and width of the geometric image, respectively. The similarity coefficient ranges from 0.01 to 0.05. To learn perceptual image patch similarity; Step S32: Use the second network receive loss density map and boundary condition images Through the loss density map and boundary condition images Obtain temperature field map predictions from a pre-trained temperature network (Nthermal). ; Predicted values ​​from the temperature field map and temperature field diagram Obtain a second loss; The formula for obtaining the second loss is as follows: ; Where H and W are the length and width of the geometric image, respectively. The similarity coefficient ranges from 0.1 to 0.2. The gradient of the predicted values ​​in the temperature field map. This represents the gradient of the temperature field diagram; Step S33: Receive waveform images using a third network, and use the waveform images to perform online performance regression on predicted EMI values ​​in the Nperf network. By sharing the bottleneck features of U-Net, the outputs of the first network and the second network are connected to obtain the efficiency prediction value and the power density prediction value. The third loss is obtained by using the predicted EMI value, efficiency prediction value, and power density prediction value; The formula for obtaining the third loss is as follows: ; in These represent the task weights, with values ​​ranging from 1 to 2, 0.1 to 0.3, and 0.5 to 0.8, respectively. These are the efficiency prediction value and the efficiency input by the user, respectively. These are the predicted power density and the power density input by the user, respectively. Step S34: Obtain the total loss by summing the first loss, the second loss, and the third loss. Apply the total loss to the Adam optimizer and set the learning rate to [value missing]. Train 50-200 times.

[0032] To further address the problems in the background technology, such as limited design space, disconnect between simulation and optimization, loss of multiphysics information, and difficulties in knowledge reuse, this training method first uses a first network to receive geometric images, directly predicts the loss density map, and introduces a first loss function that fuses L1 absolute error and learned perceptual image patch similarity. ,in It can measure the perceptual structural similarity between the predicted loss distribution and the actual loss distribution, preserving the image structure. This enables the first network to accurately reconstruct a spatially detailed loss density image from arbitrarily complex geometries, avoiding the simplification of loss to total power or average value in traditional parametric modeling. This preserves local hotspot information and solves the problem of multiphysics information loss.

[0033] The second network also introduces a difference term between the predicted temperature field gradient and the actual temperature field gradient to avoid physically unreasonable local abrupt changes, making the thermal simulation results more consistent with the principles of heat transfer, thereby improving the accuracy of predicting thermal stress concentration and failure risk.

[0034] The third network receives waveform images, and the online performance regression network predicts EMI values. Simultaneously, by sharing the bottleneck features of the U-Net, it connects the outputs of the first and second networks to obtain the predicted values ​​of efficiency and power density, and calculates the third loss. This loss adds the prediction errors of EMI, efficiency, and power density according to the task weights. The multi-task joint learning method enables the third network to simultaneously consider the coupling relationship between electrical performance, electromagnetic interference, and thermal field. Furthermore, the shared bottleneck features promote information interaction between different physical fields, avoiding the inconsistencies caused by independent optimization in traditional segmented simulation.

[0035] Finally, the total loss is the sum of the first, second, and third losses. The Adam optimizer is used, the learning rate is 1e-4, and the training is conducted for 50 to 200 epochs to ensure that the network converges to a high-precision state.

[0036] Preferably, the step S4, which uses a multi-objective optimization algorithm combined with the prediction results of the deep learning agent network for iterative optimization, is as follows: Step S41: Randomly select a first number of geometric images from the knowledge base; Randomly generate a second number of geometric images; The solid-state transformer design parameters input by the user are randomly perturbed, and a third number of geometric images are generated; Step S42: Using the geometric image generated in step S41 as the iterative individual, and performing the following constraints to select the geometric image for iteration; The constraint is that the predicted value of the largest temperature field in the geometric image cannot be greater than the first temperature value; Step S43: Perform population update on the iterative geometric image; The population update operation is as follows: Two parent geometric images are randomly selected, and their pixel values ​​are swapped within a random rectangular region to generate a child image. Add Gaussian noise N(0,0.05) to each pixel with a probability of 0.1, or perform morphological dilation or erosion operations on the geometric image; And retain the top 10% of nondominated solutions that should be optimal for the next iteration; The expression for the appropriate value is: ; ; ; ; ; in Represents the efficiency prediction value of the k-th geometric image. Represents the predicted power density value of the k-th geometric image. Let be the EMI value of the k-th geometric image. This represents the area in the k-th geometric image where the predicted temperature field value is greater than the second temperature threshold, where the second temperature threshold is less than the first temperature threshold. These represent the standard deviation and mean of the temperature field, respectively. Step S44: Iteratively execute steps S42~43 until the number of iterations equals the iteration threshold, or the Pareto front is stable.

[0037] First, step S41 is executed: the population is initialized using three complementary geometric image generation methods. These include randomly selecting a first number of existing geometric images from the knowledge base, generating a second number of geometric images completely randomly, and generating a third number of geometric images by randomly perturbing the user-input solid-state transformer design parameters. This initialization strategy directly addresses the difficulty of knowledge reuse, as the knowledge base stores excellent geometric images from historical designs, allowing each new design to inherit prior experience instead of starting from scratch. Random generation and parameter perturbation ensure population diversity. Then, according to step S42, the temperature field can be predicted using a pre-trained deep learning proxy model. From a usage perspective, if the temperature field value is greater than a first temperature value, the solid-state transformer corresponding to that geometric shape is unusable and needs to be discarded. Iterative updates are then performed, swapping pixel values ​​within random rectangular regions to generate offspring images. Gaussian noise or morphological dilation / erosion operations are added to each pixel with a probability of 0.1, and the top 10% of non-dominated solutions based on fitness values ​​are directly adopted into the next generation. Among them, rectangular region pixel swapping can simulate the local feature recombination in biomimetic structures, while Gaussian noise and morphological operations allow the generation of innovative geometric topologies that cannot be described by traditional parametric modeling, such as asymmetric air gaps, irregular winding arrangements, and biomimetic heat sinks, thus greatly expanding the design space. Finally, the iterations are waited for the number of iterations to equal the iteration threshold, or the Pareto front is stably completed for iterative updates.

[0038] The steps to generate the Pareto front solution set can be to select earlier solutions from the set of solutions in the last iteration based on the appropriate value, and combine them to form the Pareto front solution set.

[0039] Preferably, the steps for obtaining the final design scheme in step S5 are as follows: The ANSYS Maxwell and Fluent are used to perform finite element simulations of electromagnetic and thermal fields on the 3D model corresponding to the user-selected preferred geometry image. The accurate loss distribution, temperature field and EMI are obtained and compared with the prediction results of the deep learning agent network to generate a verification report. The system performs edge detection and contour extraction on the selected geometric images, identifies the core boundaries and calculates their dimensions, identifies the center coordinates and area of ​​the winding region, calculates the number of turns and wire diameter, and outputs the identified geometric features as CAD models and parameter tables. The verification report, CAD model, and parameter table are sent to the user for final selection, resulting in the final design scheme.

[0040] A multi-physics collaborative optimization design system for solid-state transformers based on image data streams employs a multi-physics collaborative optimization design method for solid-state transformers based on image data streams, including: Data acquisition module: used to acquire solid-state transformer design parameters input by the user; The image generation module is used to convert design parameters into image data, including at least geometric images, waveform images, and multiphysics label images, through the image generation engine. The model prediction module is used to predict the corresponding physical field distribution image and performance indicators based on the geometric image and waveform image using a pre-trained deep learning agent network. The iterative module uses geometric images as individuals in the image space and employs a multi-objective optimization algorithm combined with the prediction results of the deep learning agent network to iteratively optimize and generate a Pareto front solution set. The selection module is used to receive the preferred geometric image selected by the user from the Pareto front solution set and obtain the final design scheme.

[0041] Preferably, the image generation module includes a first image submodule, a second image submodule, and a third image submodule; The first image submodule is used to select the gray value of the material as the mapping condition based on the design parameters, and assign different pixel values ​​to each coordinate point in the specified two-dimensional image. The second image submodule is used to perform circuit simulation on the solid-state transformer to obtain its time-domain voltage waveform. The time-domain voltage waveform is binarized to generate a first dot plot; The Bresenham algorithm is applied to the first point image to connect adjacent points, where the pixel value of the connection between adjacent points is 1. The third image submodule is used to convert the mesh node data output from high-fidelity electromagnetic, thermal, and stress finite element simulations into a physical field image with a regular mesh through barycentric coordinate interpolation or bilinear interpolation, wherein the barycentric coordinate interpolation satisfies: ; in For the field value corresponding to the corresponding grid vertex, ( () represents the centroid coordinates of the triangular mesh; The multiphysics label image is obtained by interpolating the normalized barycenter coordinates. , These represent the minimum and maximum values ​​for the entire field. Based on the multiphysics label image, a loss density map is obtained through simulation processing. Temperature field diagram and boundary condition images .

[0042] Preferably, the model prediction module includes a first training submodule, a second training submodule, a third training submodule, and a fourth training submodule; The first training submodule is used to receive the geometric image and obtain the loss density map prediction value in the pre-trained loss network NEM using the geometric image. ; Predicted values ​​from the loss density map and loss density map Obtain the first loss; The second training submodule is used to receive the loss density map. and boundary condition images Through the loss density map and boundary condition images Obtain temperature field map predictions from a pre-trained temperature network (Nthermal). ; The third training submodule is used to receive waveform images and predict EMI values ​​in the online performance regression network Nperf based on the waveform images. By sharing the bottleneck features of U-Net, the outputs of the first network and the second network are connected to obtain the efficiency prediction value and the power density prediction value. The third loss is obtained by using the predicted EMI value, efficiency prediction value, and power density prediction value; The fourth training submodule is used to sum the first loss, the second loss, and the third loss to obtain the total loss. The Adam optimizer is then applied using this total loss, and the learning rate is set to... Train 50-200 times.

[0043] Preferably, the iteration module includes a first iteration submodule, a second iteration submodule, a third iteration submodule, and a fourth iteration submodule: The first iteration submodule is used to randomly select a first number of geometric images from the knowledge base; Randomly generate a second number of geometric images; The solid-state transformer design parameters input by the user are randomly perturbed, and a third number of geometric images are generated; The first iteration submodule is used to select the geometric image to be iterated based on the geometric image generated in the first iteration submodule and to perform the following constraints to select the geometric image to be iterated. The constraint is that the predicted value of the largest temperature field in the geometric image cannot be greater than the first temperature value; The third iteration submodule is used to perform population updates on the iterative geometric image; The population update operation is as follows: Two parent geometric images are randomly selected, and their pixel values ​​are swapped within a random rectangular region to generate a child image. Add Gaussian noise N(0,0.05) to each pixel with a probability of 0.1, or perform morphological dilation or erosion operations on the geometric image; And retain the top 10% of nondominated solutions that should be optimal for the next iteration; The fourth iteration submodule is used to re-call the second and third iteration submodules until the number of iterations equals the iteration threshold, or the Pareto front is stable.

[0044] Preferably, the selection module performs the following steps: The ANSYS Maxwell and Fluent are used to perform finite element simulations of electromagnetic and thermal fields on the 3D model corresponding to the user-selected preferred geometry image. The accurate loss distribution, temperature field and EMI spectrum are obtained and compared with the prediction results of the deep learning agent network to generate a verification report. The system performs edge detection and contour extraction on the selected geometric images, identifies the core boundaries and calculates their dimensions, identifies the center coordinates and area of ​​the winding region, calculates the number of turns and wire diameter, and outputs the identified geometric features as CAD models and parameter tables. The verification report, CAD model, and parameter table are sent to the user for final selection, resulting in the final design scheme.

[0045] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0046] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A multi-physics collaborative optimization design method for solid-state transformers based on image data streams, characterized in that, Includes the following steps: Step S1: Obtain the solid-state transformer design parameters input by the user; Step S2: Using an image generation engine, convert the design parameters into image data that includes at least geometric images, waveform images, and multiphysics label images; Step S3: Train the deep learning agent network using image data of geometric images, waveform images, and multi-physics field labeled images, and predict the corresponding physical field distribution images and performance indicators based on the geometric images and waveform images; Step S4: In the image space, taking the geometric image as an individual, a multi-objective optimization algorithm is used in combination with the prediction results of the deep learning agent network to perform iterative optimization and generate the Pareto front solution set; Step S5: Receive the preferred geometric image selected by the user from the Pareto front solution set and obtain the final design scheme.

2. The multi-physics collaborative optimization design method for solid-state transformers based on image data streams according to claim 1, characterized in that, The steps for obtaining the geometric image in step S2 are as follows: Based on the design parameters, the grayscale value of the material is selected as the mapping condition, and different pixel values ​​are assigned to each coordinate point in the specified two-dimensional image. The formula for assigning pixel values ​​is as follows: ； in This represents the pixel value at (x, y) coordinates. The indicator function is used when (x, y) is located in the k-th geometric region. At that time, The value is 1, otherwise Take the value 0. This represents the material's grayscale value mapping, where K is the total number of geometric regions. The steps for obtaining the waveform image in step S2 are as follows: The time-domain voltage waveform of a solid-state transformer is obtained by circuit simulation. The time-domain voltage waveform is binarized to generate a first dot plot; The Bresenham algorithm is applied to the first point image to connect adjacent points, where the pixel value of the connection between adjacent points is 1. The steps for acquiring image data for multiphysics label images in step S2 are as follows: The mesh node data output from high-fidelity electromagnetic, thermal, and stress finite element simulations are converted into a physical field image with a regular mesh using barycentric coordinate interpolation or bilinear interpolation, where the barycentric coordinate interpolation satisfies: ； in For the field value corresponding to the corresponding grid vertex, ( () represents the centroid coordinates of the triangular mesh; The multiphysics label image is obtained by interpolating the normalized barycenter coordinates. , These represent the minimum and maximum values ​​for the entire field. Based on the multiphysics label image, a loss density map is obtained through simulation processing. Temperature field diagram and boundary condition images .

3. The multi-physics collaborative optimization design method for solid-state transformers based on image data streams according to claim 1, characterized in that, The training process for the deep learning agent network includes: Step S31: Receive the geometric image using the first network, and obtain the loss density map prediction value from the pre-trained loss network NEM using the geometric image. ; Predicted values ​​from the loss density map and loss density map Obtain the first loss; The formula for obtaining the first loss is as follows: ； Where H and W are the length and width of the geometric image, respectively. The similarity coefficient ranges from 0.01 to 0.

05. To learn perceptual image patch similarity; Step S32: Use the second network receive loss density map and boundary condition images Through the loss density map and boundary condition images Obtain temperature field map predictions from a pre-trained temperature network (Nthermal). ; Predicted values ​​from the temperature field map and temperature field diagram Obtain a second loss; The formula for obtaining the second loss is as follows: ； Where H and W are the length and width of the geometric image, respectively. The similarity coefficient ranges from 0.1 to 0.

2. The gradient of the predicted values ​​in the temperature field map. This represents the gradient of the temperature field diagram; Step S33: Receive waveform images using a third network, and use the waveform images to perform online performance regression on predicted EMI values ​​in the Nperf network. By sharing the bottleneck features of U-Net, the outputs of the first network and the second network are connected to obtain the efficiency prediction value and the power density prediction value. The third loss is obtained by using the predicted EMI value, efficiency prediction value, and power density prediction value; The formula for obtaining the third loss is as follows: ； in These represent the task weights, with values ​​ranging from 1 to 2, 0.1 to 0.3, and 0.5 to 0.8, respectively. These are the efficiency prediction value and the efficiency input by the user, respectively. These are the predicted power density and the power density input by the user, respectively. Step S34: Obtain the total loss by summing the first loss, the second loss, and the third loss. Apply the total loss to the Adam optimizer and set the learning rate to [value missing]. Train 50-200 times.

4. The multi-physics collaborative optimization design method for solid-state transformers based on image data streams according to claim 1, characterized in that, The steps in step S4, which use a multi-objective optimization algorithm combined with the prediction results of the deep learning agent network for iterative optimization, are as follows: Step S41: Randomly select a first number of geometric images from the knowledge base; Randomly generate a second number of geometric images; The solid-state transformer design parameters input by the user are randomly perturbed, and a third number of geometric images are generated; Step S42: Using the geometric image generated in step S41 as the iterative individual, and performing the following constraints to select the geometric image for iteration; The constraint is that the predicted value of the largest temperature field in the geometric image cannot be greater than the first temperature value; Step S43: Perform population update on the iterative geometric image; The population update operation is as follows: Two parent geometric images are randomly selected, and their pixel values ​​are swapped within a random rectangular region to generate a child image. Add Gaussian noise N(0,0.05) to each pixel with a probability of 0.1, or perform morphological dilation or erosion operations on the geometric image; And retain the top 10% of nondominated solutions that should be optimal for the next iteration; The expression for the appropriate value is: ； ； ； ； ； in Represents the efficiency prediction value of the k-th geometric image. Represents the predicted power density value of the k-th geometric image. Let be the EMI value of the k-th geometric image. This represents the area in the k-th geometric image where the predicted temperature field value is greater than the second temperature threshold, where the second temperature threshold is less than the first temperature threshold. These represent the standard deviation and mean of the temperature field, respectively. Step S44: Iteratively execute steps S42~43 until the number of iterations equals the iteration threshold, or the Pareto front is stable.

5. The multi-physics collaborative optimization design method for solid-state transformers based on image data streams according to claim 1, characterized in that, The steps to obtain the final design scheme in step S5 are as follows: The ANSYS Maxwell and Fluent are used to perform finite element simulations of electromagnetic and thermal fields on the 3D model corresponding to the user-selected preferred geometry image. The accurate loss distribution, temperature field and EMI are obtained and compared with the prediction results of the deep learning agent network to generate a verification report. The system performs edge detection and contour extraction on the selected geometric images, identifies the core boundaries and calculates their dimensions, identifies the center coordinates and area of ​​the winding region, calculates the number of turns and wire diameter, and outputs the identified geometric features as CAD models and parameter tables. The verification report, CAD model, and parameter table are sent to the user for final selection, resulting in the final design scheme.

6. A multi-physics collaborative optimization design system for solid-state transformers based on image data streams, characterized in that, The method for multi-physics collaborative optimization design of solid-state transformers based on image data streams, as described in any one of claims 1 to 5, includes: Data acquisition module: used to acquire solid-state transformer design parameters input by the user; The image generation module is used to convert design parameters into image data, including at least geometric images, waveform images, and multiphysics label images, through the image generation engine. The model prediction module is used to predict the corresponding physical field distribution image and performance indicators based on the geometric image and waveform image using a pre-trained deep learning agent network. The iterative module uses geometric images as individuals in the image space and employs a multi-objective optimization algorithm combined with the prediction results of the deep learning agent network to iteratively optimize and generate a Pareto front solution set. The selection module is used to receive the preferred geometric image selected by the user from the Pareto front solution set and obtain the final design scheme.

7. The solid-state transformer multi-physics collaborative optimization design system based on image data stream according to claim 6, characterized in that, The image generation module includes a first image submodule, a second image submodule, and a third image submodule; The first image submodule is used to select the gray value of the material as the mapping condition based on the design parameters, and assign different pixel values ​​to each coordinate point in the specified two-dimensional image. The second image submodule is used to perform circuit simulation on the solid-state transformer to obtain its time-domain voltage waveform. The time-domain voltage waveform is binarized to generate a first dot plot; The Bresenham algorithm is applied to the first point image to connect adjacent points, where the pixel value of the connection between adjacent points is 1. The third image submodule is used to convert the mesh node data output from high-fidelity electromagnetic, thermal, and stress finite element simulations into a physical field image with a regular mesh through barycentric coordinate interpolation or bilinear interpolation, wherein the barycentric coordinate interpolation satisfies: ； in For the field value corresponding to the corresponding grid vertex, ( () represents the centroid coordinates of the triangular mesh; The multiphysics label image is obtained by interpolating the normalized barycenter coordinates. , These represent the minimum and maximum values ​​for the entire field. Based on the multiphysics label image, a loss density map is obtained through simulation processing. Temperature field diagram and boundary condition images .

8. The solid-state transformer multi-physics collaborative optimization design system based on image data stream according to claim 6, characterized in that, The model prediction module includes a first training submodule, a second training submodule, a third training submodule, and a fourth training submodule. The first training submodule is used to receive the geometric image and obtain the loss density map prediction value in the pre-trained loss network NEM using the geometric image. ; Predicted values ​​from the loss density map and loss density map Obtain the first loss; The second training submodule is used to receive the loss density map. and boundary condition images Through the loss density map and boundary condition images Obtain temperature field map predictions from a pre-trained temperature network (Nthermal). ; The third training submodule is used to receive waveform images and predict EMI values ​​in the online performance regression network Nperf based on the waveform images. By sharing the bottleneck features of U-Net, the outputs of the first network and the second network are connected to obtain the efficiency prediction value and the power density prediction value. The third loss is obtained by using the predicted EMI value, efficiency prediction value, and power density prediction value; The fourth training submodule is used to sum the first loss, the second loss, and the third loss to obtain the total loss. The Adam optimizer is then applied using this total loss, and the learning rate is set to... Train 50-200 times.

9. The multi-physics collaborative optimization design system for solid-state transformers based on image data streams according to claim 6, characterized in that, The iteration module includes a first iteration submodule, a second iteration submodule, a third iteration submodule, and a fourth iteration submodule: The first iteration submodule is used to randomly select a first number of geometric images from the knowledge base; Randomly generate a second number of geometric images; The solid-state transformer design parameters input by the user are randomly perturbed, and a third number of geometric images are generated; The first iteration submodule is used to select the geometric image to be iterated based on the geometric image generated in the first iteration submodule and to perform the following constraints to select the geometric image to be iterated. The constraint is that the predicted value of the largest temperature field in the geometric image cannot be greater than the first temperature value; The third iteration submodule is used to perform population updates on the iterative geometric image; The population update operation is as follows: Two parent geometric images are randomly selected, and their pixel values ​​are swapped within a random rectangular region to generate a child image. Add Gaussian noise N(0,0.05) to each pixel with a probability of 0.1, or perform morphological dilation or erosion operations on the geometric image; And retain the top 10% of nondominated solutions that should be optimal for the next iteration; The fourth iteration submodule is used to re-call the second and third iteration submodules until the number of iterations equals the iteration threshold, or the Pareto front is stable.

10. A multi-physics collaborative optimization design system for solid-state transformers based on image data streams according to claim 6, characterized in that, The selected module will perform the following steps: The ANSYS Maxwell and Fluent are used to perform finite element simulations of electromagnetic and thermal fields on the 3D model corresponding to the user-selected preferred geometry image. The accurate loss distribution, temperature field and EMI spectrum are obtained and compared with the prediction results of the deep learning agent network to generate a verification report. The system performs edge detection and contour extraction on the selected geometric images, identifies the core boundaries and calculates their dimensions, identifies the center coordinates and area of ​​the winding region, calculates the number of turns and wire diameter, and outputs the identified geometric features as CAD models and parameter tables. The verification report, CAD model, and parameter table are sent to the user for final selection, resulting in the final design scheme.

Images