Converter transformer electrostatic ring structure optimization method based on transfer learning and related device

Abstract

The invention discloses a transfer learning-based converter transformer electrostatic ring structure optimization method and a related device, and the method comprises the steps: carrying out the sampling in a value range of a to-be-optimized structure variable parameter of an electrostatic ring, obtaining a plurality of groups of working condition combination data, carrying out the geometric reconstruction and pure electric field calculation of the electrostatic ring, and obtaining two indexes; further obtaining source domain data; selecting a plurality of groups of working condition combination data from the plurality of groups of working condition combination data, performing geometric reconstruction and electrothermal coupling calculation on the electrostatic ring, obtaining two indexes corresponding to each group of working condition combination data, and further obtaining target domain data; the source domain data and the target domain data are calculated through the transfer learning model, and the minimum insulation margin of the oil gap under the alternating current working condition and the maximum electric field intensity in the insulating paperboard under the direct current working condition are obtained through prediction; and then screening is carried out so as to determine the optimal structure variable parameter of the electrostatic ring. The optimization efficiency and performance of the electrostatic ring structure on the valve side of the converter transformer under the electrothermal coupling condition can be improved.

Description

Technical Field

[0001] This invention belongs to the field of insulation structure optimization and intelligent modeling of power equipment, specifically involving a method and related device for optimizing the electrostatic ring structure of converter transformers based on transfer learning. Background Technology

[0002] As a critical connecting device between high-voltage direct current (HVDC) transmission systems and AC systems, converter transformers bear superimposed AC and DC voltages on their valve-side windings for extended periods. This results in a complex internal electric field distribution and a significant temperature rise, posing a severe reliability challenge to the insulation system. Particularly in the end electrostatic ring region, due to its high potential, the electric field concentration is more pronounced, making it a high-risk area for localized insulation breakdown.

[0003] Currently, in the design and optimization of converter transformer insulation structures, most mainstream methods are based on static electric field simulation for modeling and analysis, failing to fully consider the strong coupling effect between the temperature field and the electric field during operation. In actual operation, the dielectric constant and conductivity of the oil-paper insulation material are highly sensitive to temperature and electric field strength, exhibiting significant nonlinear variation characteristics. Ignoring this coupling effect will lead to distortion in electric field prediction, seriously affecting the accuracy and safety margin of insulation design.

[0004] While electrothermal coupling simulation can realistically reflect the physical behavior of materials under actual working conditions, its computational cost is extremely high—the simulation time for each set of design samples can typically reach several hours, significantly limiting the possibility of constructing large-scale samples. Furthermore, to improve optimization efficiency, surrogate models are often used to approximate high-fidelity simulation results, but their prediction accuracy heavily relies on the support of a large number of training samples. Given the high cost of electrothermal coupling simulation, the practical feasibility of this method is limited.

[0005] Therefore, accurately describing the nonlinear behavior of electrothermal coupling while reducing computational overhead and improving optimization efficiency has become a key technical bottleneck that urgently needs to be overcome in the field of converter transformer insulation structure design. An optimization method that balances high efficiency and reliability is urgently needed to provide practical technical support for engineering applications under complex operating conditions. Summary of the Invention

[0006] To address the problems existing in the prior art, the present invention aims to provide a method and related apparatus for optimizing the electrostatic ring structure of a converter transformer based on transfer learning. The present invention can improve the optimization efficiency and performance of the valve-side electrostatic ring structure of the converter transformer under electrothermal coupling conditions.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A method for optimizing the electrostatic ring structure of a converter transformer based on transfer learning includes the following steps: Sampling was performed within the range of values ​​for the structural variable parameters to be optimized in the electrostatic ring to obtain multiple sets of working condition combination data; The electrostatic ring is geometrically reconstructed and a simple electric field is calculated using the multiple sets of operating condition combination data to obtain two indices corresponding to each set of operating condition combination data. The multiple sets of operating condition combination data and the two indices corresponding to each set of operating condition combination data constitute source domain data. The two indices are the minimum insulation margin of the oil gap of the electrostatic ring under AC operating conditions and the maximum electric field strength of the electrostatic ring in the insulating paperboard under DC operating conditions. Several sets of working condition combination data are selected from the multiple sets of working condition combination data, and the electrostatic ring is geometrically reconstructed and electrothermal coupling calculation is performed using the several sets of working condition combination data to obtain the two indicators corresponding to each set of working condition combination data. The several sets of working condition combination data and the two indicators corresponding to each set of working condition combination data constitute the target domain data. The minimum insulation margin of the oil gap under AC operating conditions is predicted by calculating the source domain data using the pre-trained first transfer learning model. The maximum electric field strength in the insulating paperboard under DC conditions is predicted by calculating the target domain data using a pre-trained second transfer learning model. The predicted minimum insulation margin of the oil gap under AC conditions and the maximum electric field strength in the insulating paperboard under DC conditions were screened, and the optimal structural variable parameters of the electrostatic ring were determined based on the screening results.

[0008] Preferably, the Latin hypercube sampling method is used to sample within the range of values ​​of the structural variable parameters to be optimized in the electrostatic ring, thereby obtaining multiple sets of working condition combination data.

[0009] Preferably, the structural variable parameters to be optimized for the electrostatic ring include the radius of the outer corner of the upper side of the outer insulating layer, the radius of the outer corner of the upper side of the electrostatic ring, the radius of the inner corner of the upper side of the outer insulating layer, the radius of the inner corner of the upper side of the electrostatic ring, the radius of the outer corner of the lower side of the outer insulating layer, the radius of the outer corner of the lower side of the electrostatic ring, the thickness of the outer insulating layer, the distance from the inner side of the outer insulating layer to the center line, and the distance between the lower side of the outer insulating layer and the first disc winding.

[0010] Preferably, the SOBOL sensitivity analysis method is used to screen the structural variable parameters to be optimized for each electrostatic ring, and the structural variable parameters to be optimized for the electrostatic ring include the radius of the outer corner of the upper side of the outer insulating layer, the radius of the outer corner of the upper side of the electrostatic ring, the radius of the outer corner of the lower side of the outer insulating layer, the thickness of the outer insulating layer, the distance from the inner side of the outer insulating layer to the center line, and the distance between the lower side of the outer insulating layer and the first disc winding.

[0011] Preferably, the network structure and hyperparameter settings of the first and second transfer learning models are identical, and the training process includes: Pre-training is performed using source domain data, and the parameters of the frozen layer are fixed after pre-training is completed. Then, the target domain data is used for training. During training, the weights of the trainable layers are fine-tuned. After training, the final first or second transfer learning model network is obtained.

[0012] Preferably, the activation function of the first transfer learning model and the second transfer learning model is the ReLU activation function, the loss function is the root mean square error, and the optimizer is Adam.

[0013] Preferably, the process of screening the predicted minimum insulation margin of the oil gap under AC conditions and the maximum electric field strength in the insulating paperboard under DC conditions, and determining the optimal structural variable parameters of the electrostatic ring based on the screening results, includes the following steps: With the objectives of maximizing the oil gap insulation margin and minimizing the electric field strength in the insulating paperboard, a non-dominated sorting genetic algorithm was used to evaluate the predicted minimum oil gap insulation margin under AC conditions and the maximum electric field strength in the insulating paperboard under DC conditions. The set of operating condition combinations with the highest evaluation scores was obtained and used as the optimal structural variable parameters of the electrostatic ring.

[0014] This invention also provides a converter transformer electrostatic ring structure optimization system based on transfer learning, used to implement the converter transformer electrostatic ring structure optimization method based on transfer learning described above, comprising: Sampling unit: Used to sample within the range of values ​​of the structural variable parameters to be optimized in the electrostatic loop, and obtain multiple sets of working condition combination data; The first calculation unit is used to perform geometric reconstruction of the electrostatic ring and perform simple electric field calculation using the multiple sets of operating condition combination data to obtain two indicators corresponding to each set of operating condition combination data. The multiple sets of operating condition combination data and the two indicators corresponding to each set of operating condition combination data constitute source domain data. Among them, the two indicators are the minimum insulation margin of the oil gap of the electrostatic ring under AC operating conditions and the maximum electric field strength of the electrostatic ring in the insulating paperboard under DC operating conditions. The second calculation unit is used to select several sets of working condition combination data from the multiple sets of working condition combination data, and use the several sets of working condition combination data to perform geometric reconstruction of the electrostatic ring and perform electrothermal coupling calculation to obtain the two indicators corresponding to each set of working condition combination data, and to construct the target domain data by combining the several sets of working condition combination data and the two indicators corresponding to each set of working condition combination data. First prediction unit: used to calculate the source domain data using the pre-trained first transfer learning model and predict the minimum insulation margin of the oil gap under AC operating conditions; The second prediction unit is used to calculate the target domain data using a pre-trained second transfer learning model and predict the maximum electric field strength in the insulating paperboard under DC conditions. Optimization unit: Used to screen the predicted minimum insulation margin of the oil gap under AC conditions and the maximum electric field strength in the insulating paperboard under DC conditions, and determine the optimal structural variable parameters of the electrostatic ring based on the screening results.

[0015] The present invention also provides an electronic device, comprising: One or more processors; A storage device on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the transfer learning-based electrostatic ring structure optimization method for converter transformers as described above.

[0016] The present invention also provides a storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the transfer learning-based electrostatic ring structure optimization method for converter transformers as described above.

[0017] The present invention has the following beneficial effects: This invention presents a transfer learning-based optimization method for the electrostatic ring structure of converter transformers. This method effectively overcomes the technical bottleneck of balancing accuracy and computational efficiency in the design of converter transformer insulation structures, demonstrating significant practicality and advancement. Specifically, addressing the problem of prediction distortion caused by neglecting electrothermal coupling in existing static electric field simulations, this invention introduces thermoelectric coupling calculations into the target domain data construction. It fully considers the nonlinear characteristics of the dielectric constant and conductivity of the oil-paper insulation material as a function of temperature and electric field strength, accurately capturing the strong coupling effect between the temperature field and the electric field in actual operation. This improves the reliability of insulation performance index predictions from a data-driven perspective and resolves the hidden danger of inaccurate safety margin assessments in traditional designs. Meanwhile, addressing the pain points of high cost and difficulty in constructing large-scale samples for electrothermal coupling simulation, this invention innovatively adopts a collaborative application of source and target domain data partitioning and dual transfer learning models: the source domain constructs large-scale samples based on simple electric field calculations, providing a sufficient training foundation for the first transfer learning model to achieve efficient prediction of AC operating conditions; the target domain only needs to select a small number of samples for high-cost electrothermal coupling calculations to support the second transfer learning model in accurately predicting key DC operating conditions, without relying on a large number of electrothermal coupling simulation samples, significantly reducing computational overhead and sample construction costs, and solving the problem of limited accuracy caused by insufficient training samples in traditional surrogate models. Finally, by screening the prediction results of the dual models, the optimal parameters are determined, ensuring comprehensive coverage and accurate evaluation of key insulation indicators under AC and DC superimposed operating conditions, while significantly improving optimization efficiency, achieving an organic unity of high efficiency and reliability, and providing practical technical support for the optimization of the insulation structure of the electrostatic ring region of converter transformers. Attached Figure Description

[0018] Figure 1 This is a flowchart of the converter transformer electrostatic ring structure optimization method based on transfer learning in an embodiment of the present invention.

[0019] Figure 2 The schematic diagram of the structure to be optimized and the schematic diagram of each design parameter in the embodiment of the present invention.

[0020] Figure 3 This is a flowchart of the electrothermal coupling calculation in an embodiment of the present invention.

[0021] Figure 4 This is a schematic diagram of the transfer learning agent model structure in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0023] This invention presents a transfer learning-based optimization method for the electrostatic ring structure of converter transformers. This method improves the optimization efficiency and performance of the valve-side electrostatic ring structure under electrothermal coupling conditions. It has two optimization objectives: first, to increase the insulation margin in the insulating oil gap near the valve-side winding electrostatic ring under AC conditions; and second, to reduce the electric field strength in the insulating paperboard under DC conditions. By introducing a transfer learning strategy, this method significantly reduces the amount of target domain data required for training a high-precision surrogate model, reducing model construction time by approximately 67.76% while maintaining optimization accuracy.

[0024] In this embodiment, the valve-side electrostatic ring structure of a 500kV converter transformer is optimized. The specific implementation steps are as follows, and the overall optimization design process is shown in the appendix. Figure 1 : Step 1: Parameter Filtering The initial structure has a total of 9 variable design parameters, each with a reasonable range of variation set, as detailed in the appendix. Figure 2 Taking into account the sensitivity of each design variable to the optimization variable, the variable with higher sensitivity to the optimization design parameter is selected as the key structural variable as the optimization object (where the outer insulation layer covers the surface of the electrostatic ring): outer corner radius of the upper side of the outer insulation layer X1, outer corner radius of the upper side of the electrostatic ring X2, inner corner radius of the upper side of the outer insulation layer X3, inner corner radius of the upper side of the electrostatic ring X4, outer corner radius of the lower side of the outer insulation layer X5, outer corner radius of the lower side of the electrostatic ring X6, thickness of the outer insulation layer X7, distance from the inner side of the outer insulation layer to the center line X8, distance from the lower side of the outer insulation layer to the first disc winding X9.

[0025] The design ranges for each parameter are shown in Table 1: Table 1

[0026] Given the relatively uniform temperature distribution in the vicinity of the electrostatic ring, the pure electric field calculation results and the electrothermal coupling calculation results show a high degree of consistency in their response trends to structural parameters during the geometric parameter changes. Considering that the computational cost of pure electric field simulation is significantly lower than that of electrothermal coupling simulation, this embodiment further selects the pure electric field model for sample space construction in the global sensitivity analysis stage to improve computational efficiency and ensure analysis accuracy.

[0027] This embodiment uses the SOBOL sensitivity analysis method to screen each initial design parameter, selecting those more sensitive to the optimization variables for further study. Specifically, the SOBOL sensitivity analysis method requires a large sample size. To balance accuracy and efficiency, this embodiment first trains a backpropagation neural network (BPNN) surrogate model using 300 sets of finite element analysis (FEA) results. The model input consists of the aforementioned nine parameters (i.e., the outer corner radius of the upper side of the outer insulating layer X1, the outer corner radius of the upper side of the electrostatic ring X2, the inner corner radius of the upper side of the outer insulating layer X3, the inner corner radius of the upper side of the electrostatic ring X4, the outer corner radius of the lower side of the outer insulating layer X5, and the inner corner radius of the lower side of the electrostatic ring X6). The design parameters (X6 for the outer edge fillet radius, X7 for the outer insulation layer thickness, X8 for the distance from the inner side of the outer insulation layer to the center line, and X9 for the distance from the lower side of the outer insulation layer to the first disc winding) are output as two optimization target values ​​(i.e., the insulation margin in the insulating oil gap near the electrostatic ring of the valve-side winding under AC conditions and the electric field strength in the insulating paperboard under DC conditions). Subsequently, using the Saltelli sampling strategy, 2000 sets of samples are generated on the trained BPNN surrogate model, and the overall effect sensitivity index is based on this. S T For the two indicators (i.e., the optimization objectives mentioned above), the corresponding S... T An equal-weighted summation (0.5 / 0.5) was used. When the weighted overall sensitivity was greater than 0.10, the variable was considered a key parameter requiring further optimization. In this embodiment, six parameters were ultimately selected for optimization: the outer corner radius of the upper side of the outer insulation layer (X1), the outer corner radius of the upper side of the electrostatic ring (X2), the outer corner radius of the lower side of the outer insulation layer (X5), the thickness of the outer insulation layer (X7), the distance from the inner side of the outer insulation layer to the center line (X8), and the distance from the lower side of the outer insulation layer to the first winding (X9) for further research.

[0028] Step 2: Obtaining the Transfer Learning Sample Set Using the Latin hypercube, 200 sets of operating condition combinations were pre-calculated within the value range of each selected parameter to be optimized. From these 200 sets, any 50 combinations were selected to reconstruct the geometric model of the electrostatic loop and perform electrothermal coupling calculations, yielding the minimum insulation margin under AC operating conditions. Margin AC,min Maximum electric field strength under DC operating conditions E DC,max The data obtained from the electrothermal coupling calculation is used as the target domain data. See the appendix for the calculation process. Figure 3 First, a simplified calculation process was employed for thermal calculations. Using the transformer loss under rated load provided by the factory test as a benchmark, the influence of temperature on conductor resistivity was considered during the calculation process, ultimately obtaining the steady-state thermal flow field distribution of the winding. Then, the obtained transformer winding temperature field distribution data was imported into the electric field calculation module. Considering the influence of temperature and electric field strength on the dielectric constant and conductivity of the insulating material, the steady-state AC and DC electric field distribution of the winding was calculated. Further calculations were performed on the maximum DC electric field strength of the electrostatic ring at the valve-side winding end and the minimum AC oil gap insulation margin.

[0029] In addition, all 200 sets of data were geometrically reconstructed and simple electric field calculations were performed (the nonlinear effects of temperature and electric field on the oil-paper insulation material were not considered when performing simple electric field calculations, and the dielectric performance parameters were calculated as constants), and the data from the simple electric field calculations were used as source domain data.

[0030] Step 3: Training the transfer learning agent model Based on the training parameter sample set constructed in step 2, a transfer learning neural network model suitable for the research problem of this invention was trained.

[0031] This invention introduces a transfer learning strategy to address the problem of insufficient prediction accuracy of surrogate models under conditions of limited sample size for electrothermal coupling data. The transfer learning model used in this embodiment is shown in the attached figure. Figure 4As shown, the transfer learning method adopted is based on the classic framework of "pre-training-fine-tuning". The main training process includes the following three stages: (1) Data preprocessing stage: The sample data of the source domain and the target domain are mixed and processed, and the input variables are uniformly scaled by the max-min normalization method to improve the generalization ability and convergence speed of the model; (2) Pre-training stage: The model is trained using sufficient source domain data. Among them, the first two layers of the network are frozen layers (FL), which are responsible for extracting the general features of the source domain samples; the remaining layers are trainable layers (TL) to implement specific prediction tasks. (3) Fine-tuning stage: In this stage, the model does not need to rebuild the network structure, but directly transfers the network architecture and its parameters from the pre-training stage to the target domain. The parameters of the frozen layers remain unchanged, and only the weights of the trainable layers are fine-tuned to adapt to the feature distribution of the target domain data, thereby completing the target domain proxy model. Specifically, the training data used in this embodiment includes 200 sets of source domain data samples and 50 sets of target domain data samples from step 2, with the training set and test set divided in a ratio of 0.8:0.2. Two independent surrogate models were constructed for the two optimization objectives: one for predicting the minimum insulation margin of the oil gap under AC operating conditions (denoted as the first transfer learning model), and the other for predicting the maximum electric field strength in the insulating paperboard under DC operating conditions (denoted as the second transfer learning model). The network structure and hyperparameter settings of these two models are identical; detailed parameter configurations are shown in Table 2.

[0032] Table 2

[0033] For the prediction results of the maximum electric field strength of the insulating paperboard under DC conditions, the average relative percentage error of the corresponding model on the test set is 0.17%; for the prediction results of the minimum insulation margin of the oil gap under AC conditions, the average relative percentage error of the corresponding model on the test set is 1.34%.

[0034] From the perspective of computational time cost, the computing platform configuration used in this embodiment is as follows: an Intel Xeon W9-3495X processor with a main frequency of 1.90GHz, 128GB of memory, and an NVIDIA RTX A4000 graphics processor. Under these hardware conditions, the average computation time for a single pure electric field simulation is approximately 5 minutes, while the average computation time for a single electrothermal coupling simulation is as high as 70 minutes.

[0035] In the method proposed in this invention, the training samples required for transfer learning include 200 sets of source domain (simple electric field calculation) samples and 50 sets of target domain (electrothermal coupling calculation) samples, with a total sample construction time of 75 hours. Furthermore, the training phase of the transfer learning surrogate model is accelerated using GPUs, with the training time for a single model being approximately 8 minutes. Therefore, the total time for the entire surrogate model construction phase using the method of this invention is approximately 75.27 hours.

[0036] In contrast, if method 2 (i.e., electrothermal coupling full data training) is used to obtain model accuracy comparable to the method of this invention, no less than 200 sets of electrothermal coupling samples need to be constructed, and the sample preparation stage takes up to 233.33 hours; in addition, the training time for a single surrogate model is about 5 minutes, and the total time for the overall construction process is about 233.5 hours.

[0037] In summary, compared with Method 2, the transfer learning modeling method proposed in this invention can reduce the total time cost of model construction by about 67.76% while ensuring model accuracy, significantly improving the computational efficiency in engineering applications.

[0038] Step 4: Multi-objective optimization process The optimization objective has been clearly stated above: to increase the minimum insulation margin in the insulating oil gap. Reduce the maximum DC electric field strength in the insulating paper Simultaneously, it ensures that the maximum DC field strength in the insulating oil does not exceed 12 kV / mm. To efficiently achieve multi-objective optimization, this invention employs a non-dominated sorting genetic algorithm (NSGA-II) to globally optimize the design parameters. Specifically, the NSGA-II algorithm calls two pre-trained transfer learning surrogate models to achieve rapid prediction of optimization indicators. The population size is set to 100, and the maximum number of iterations is set to 200.

[0039] This invention introduces the TOPSIS method to score each solution in the Pareto front solution set and selects the solution with the highest score as the final optimal structure. In the TOPSIS evaluation process, the weights of both optimization objectives are 0.5. According to the TOPSIS evaluation results, the highest score is 0.6664, and the corresponding parameter combination is the optimal solution obtained in this optimization. The optimal parameters after the relevant model optimization are shown in Table 3, with a corresponding predicted value of 1.846 and a predicted value of 23.05 kV / mm.

[0040] Table 3

[0041] Verification showed that the minimum insulation margin of the oil gap was 1.84, an improvement of 8.87% compared to the original value. The maximum DC electric field was 26.1 kV / mm, a decrease of 12.71% compared to the original value. The maximum DC electric field in the oil region was 8.62 kV / mm, also a decrease compared to the original value. In summary, the optimization results meet the basic design requirements, verifying the effectiveness of the multi-objective optimization method.

[0042] Furthermore, embodiments of the present invention also provide a converter transformer electrostatic ring structure optimization system based on transfer learning, used to implement the converter transformer electrostatic ring structure optimization method based on transfer learning described above. The system includes: Sampling unit: Used to sample within the range of values ​​of the structural variable parameters to be optimized in the electrostatic loop, and obtain multiple sets of working condition combination data; The first calculation unit is used to perform geometric reconstruction of the electrostatic ring and perform simple electric field calculation using the multiple sets of operating condition combination data to obtain two indicators corresponding to each set of operating condition combination data. The multiple sets of operating condition combination data and the two indicators corresponding to each set of operating condition combination data constitute source domain data. Among them, the two indicators are the minimum insulation margin of the oil gap of the electrostatic ring under AC operating conditions and the maximum electric field strength of the electrostatic ring in the insulating paperboard under DC operating conditions. The second calculation unit is used to select several sets of working condition combination data from the multiple sets of working condition combination data, and use the several sets of working condition combination data to perform geometric reconstruction of the electrostatic ring and perform thermal coupling calculation to obtain the two indicators corresponding to each set of working condition combination data, and to construct the target domain data by combining the several sets of working condition combination data and the two indicators corresponding to each set of working condition combination data. First prediction unit: used to calculate the source domain data using the pre-trained first transfer learning model and predict the minimum insulation margin of the oil gap under AC operating conditions; The second prediction unit is used to calculate the target domain data using a pre-trained second transfer learning model and predict the maximum electric field strength in the insulating paperboard under DC conditions. Optimization unit: Used to screen the predicted minimum insulation margin of the oil gap under AC conditions and the maximum electric field strength in the insulating paperboard under DC conditions, and determine the optimal structural variable parameters of the electrostatic ring based on the screening results.

[0043] The embodiments of the present invention also provide corresponding electronic devices and computer-readable storage media for implementing the solutions provided in the embodiments of the present invention.

[0044] The electronic device includes a storage device and one or more processors. The storage device stores instructions or code, and the processors execute the instructions or code to enable the device to perform the transfer learning-based electrostatic ring structure optimization method for converter transformers as described in any embodiment of this application.

[0045] The storage medium stores a computer program, which, when executed by a processor, implements the converter transformer electrostatic ring structure optimization method based on transfer learning as described in any embodiment of this application.

[0046] Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for optimizing the structure of a converter transformer electrostatic ring based on transfer learning, characterized in that, The method comprises the following steps: Sampling in the value range of the structure variable parameters of the electrostatic ring to be optimized to obtain a plurality of groups of working condition combination data; Geometric reconstruction of the electrostatic ring is performed using the plurality of groups of working condition combination data, and simple electric field calculation is performed to obtain two indexes corresponding to each group of working condition combination data, and the plurality of groups of working condition combination data and the two indexes corresponding to each group of working condition combination data constitute source domain data; wherein the two indexes are minimum insulation margin of oil gap of the electrostatic ring under alternating current working condition and maximum electric field strength in the insulation paperboard of the electrostatic ring under direct current working condition; Selecting a plurality of groups of working condition combination data from the plurality of groups of working condition combination data, and performing geometric reconstruction of the electrostatic ring using the plurality of groups of working condition combination data and performing electric-thermal coupling calculation to obtain the two indexes corresponding to each group of working condition combination data, and the plurality of groups of working condition combination data and the two indexes corresponding to each group of working condition combination data constitute target domain data; The first pre-trained transfer learning model is used to calculate the source domain data to predict the minimum insulation margin of the oil gap under the alternating current working condition; The second pre-trained transfer learning model is used to calculate the target domain data to predict the maximum electric field strength in the insulation paperboard under the direct current working condition; The predicted minimum insulation margin of the oil gap under the alternating current working condition and the maximum electric field strength in the insulation paperboard under the direct current working condition are screened, and the optimal structure variable parameters of the electrostatic ring are determined according to the screening result.

2. The method of claim 1, wherein the method is based on transfer learning. The Latin hypercube sampling method is used to sample in the value range of the structure variable parameters of the electrostatic ring to be optimized to obtain a plurality of groups of working condition combination data.

3. The method of claim 1, wherein the method is based on transfer learning. The structure variable parameters of the electrostatic ring to be optimized include the upper side outer edge fillet radius of the outer insulation layer, the upper side outer edge fillet radius of the electrostatic ring, the upper side inner edge fillet radius of the outer insulation layer, the upper side inner edge fillet radius of the electrostatic ring, the lower side outer edge fillet radius of the outer insulation layer, the lower side outer edge fillet radius of the electrostatic ring, the thickness of the outer insulation layer, the inner side distance of the outer insulation layer to the center line, and the distance between the lower side of the outer insulation layer and the first pancake winding.

4. The converter transformer electrostatic ring structure optimization method based on transfer learning according to claim 3, characterized in that, The SOBOL sensitivity analysis method is used to screen each structure variable parameter of the electrostatic ring to be optimized to obtain the structure variable parameters of the electrostatic ring to be optimized, including the upper side outer edge fillet radius of the outer insulation layer, the upper side outer edge fillet radius of the electrostatic ring, the lower side outer edge fillet radius of the outer insulation layer, the thickness of the outer insulation layer, the inner side distance of the outer insulation layer to the center line, and the distance between the lower side of the outer insulation layer and the first pancake winding.

5. The method of claim 1, wherein the method is based on transfer learning. The network structure and hyperparameter setting of the first transfer learning model and the second transfer learning model are consistent, and the training process comprises: Pre-training is performed using the source domain data, and after the pre-training is completed, the parameters of the frozen layer are fixed; Then, the target domain data is used for training, the weights of the trainable layers are fine-tuned during the training, and finally the first transfer learning model or the second transfer learning model network is obtained after the training is completed.

6. The method of claim 1, wherein the method is based on transfer learning. The activation function of the first transfer learning model and the second transfer learning model adopts the Relu activation function, the loss function adopts the root mean square error, and the optimizer adopts Adam.

7. The method of claim 1, wherein the method is based on transfer learning. The predicted oil gap minimum insulation margin under alternating current working condition and the maximum electric field strength in the insulation paperboard under direct current working condition are screened, and the optimal structure variable parameter of the electrostatic ring is determined according to the screening result. With the maximum oil gap insulation margin and the minimum electric field strength in the insulation paperboard as the target, the non-dominated sorting genetic algorithm is used to evaluate the predicted oil gap minimum insulation margin under alternating current working condition and the maximum electric field strength in the insulation paperboard under direct current working condition, and a group of working condition combination data with the highest evaluation result score is obtained, which is used as the optimal structure variable parameter of the electrostatic ring.

8. A converter transformer electrostatic ring structure optimization system based on transfer learning, used to implement the converter transformer electrostatic ring structure optimization method based on transfer learning in any one of claims 1-7, characterized in that, Comprise: a sampling unit for sampling in the value range of the structure variable parameter to be optimized of the electrostatic ring to obtain a plurality of groups of working condition combination data; a first calculation unit for using the plurality of groups of working condition combination data to perform geometric reconstruction on the electrostatic ring and perform simple electric field calculation to obtain two indexes corresponding to each group of working condition combination data, and the plurality of groups of working condition combination data and the two indexes corresponding to each group of working condition combination data constitute source domain data; wherein the two indexes are respectively the minimum oil gap insulation margin of the electrostatic ring under alternating current working condition and the maximum electric field strength in the insulation paperboard of the electrostatic ring under direct current working condition; a second calculation unit for selecting a plurality of groups of working condition combination data from the plurality of groups of working condition combination data, and using the plurality of groups of working condition combination data to perform geometric reconstruction on the electrostatic ring and perform electric-thermal coupling calculation to obtain the two indexes corresponding to each group of working condition combination data, and the plurality of groups of working condition combination data and the two indexes corresponding to each group of working condition combination data constitute target domain data; a first prediction unit for calculating the source domain data through a pre-trained first transfer learning model to predict the minimum oil gap insulation margin under alternating current working condition; a second prediction unit for calculating the target domain data through a pre-trained second transfer learning model to predict the maximum electric field strength in the insulation paperboard under direct current working condition; an optimization unit for screening the predicted minimum oil gap insulation margin under alternating current working condition and the maximum electric field strength in the insulation paperboard under direct current working condition, and determining the optimal structure variable parameter of the electrostatic ring according to the screening result.

9. An electronic device, comprising: Comprise: one or more processors; a storage device having one or more programs stored thereon; when the one or more programs are executed by the one or more processors, the one or more processors implement the transfer learning-based converter transformer electrostatic ring structure optimization method of any one of claims 1 to 7.

10. A storage medium, characterized by a computer program is stored thereon, wherein the computer program is executed by a processor to implement the transfer learning-based converter transformer electrostatic ring structure optimization method of any one of claims 1 to 7.

Images