Transformer electromagnetic parameter intelligent prediction and optimization method and system

Abstract

The invention belongs to the technical field of transformers, and particularly relates to a transformer electromagnetic parameter intelligent prediction and optimization method and system. The method aims at solving the problems that due to dynamic changes of process coefficients, theory and actual measurement are not accurate, a single model cannot consider precision and safety at the same time, and efficiency is reduced due to the fact that design iteration depends on manual trial and error. By establishing a historical sample set, special models are respectively trained for two scenes of interpolation and extrapolation, and during actual prediction, a system automatically judges and calls the most suitable model of a to-be-tested transformer based on design parameters; and finally, a prediction result is verified by comparing a prediction value with a measured value, and if the result is not ideal, an alarm and model updating mechanism is triggered, so that the one-time design success rate is improved, and the research and development cost and period are effectively reduced.

Description

Technical Field

[0001] This invention belongs to the field of transformer technology, specifically relating to a method and system for intelligent prediction and optimization of transformer electromagnetic parameters. Background Technology

[0002] Transformers are core equipment in power systems, and the design accuracy of their electromagnetic parameters, such as no-load loss, load loss, no-load current, and load impedance, directly determines the equipment's energy efficiency, manufacturing cost, and operational reliability. Traditional electromagnetic parameter design relies on classical electromagnetic theoretical models such as Maxwell's equations and finite element analysis. It introduces empirical "process coefficients" to compensate for the influence of real-world factors like material properties and manufacturing tolerances on theoretical calculations, and ultimately verifies and corrects the design through prototype manufacturing and actual testing. However, with the continuous advancement of modern materials technology, such as high-performance silicon steel sheets, and the optimization of manufacturing processes, such as high-precision assembly tooling, material properties and process conditions are often dynamically changing. This leads to significant differences between the empirically set fixed "process coefficients" and actual production conditions, increasing the deviation between theoretical and measured values. Consequently, this restricts transformer product performance, cost control, and R&D efficiency.

[0003] Currently, methods to reduce the deviation between theoretical and actual values ​​include manual parameter tuning and static fitting based on historical data. Specifically, manual parameter tuning relies on the experience of senior engineers. By comparing measured data from prototypes with theoretical calculations, various process coefficients are manually and repeatedly adjusted through multiple iterations of "design-prototype-testing" until the product meets standards. This method is highly dependent on personal experience, and the parameter tuning process is prone to human error and inefficient, leading to problems such as difficulty in guaranteeing product performance, lengthy R&D cycles, and high prototyping costs. Static fitting based on historical data, on the other hand, utilizes accumulated experimental data to fit fixed empirical formulas or correction coefficients through simple regression analysis, replacing some of the manual experience. However, this method establishes a "static" mathematical model. Once production conditions change, such as changing material batches, upgrading production lines, or launching new product series, the original model quickly becomes invalid and cannot adapt to the dynamic evolution of process parameters.

[0004] In view of this, the present invention is hereby proposed. Summary of the Invention

[0005] One objective of this invention is to address the problems of inaccuracy between theory and measurement caused by dynamic changes in process coefficients, the inability of a single model to balance accuracy and safety, and the reduced efficiency of design iterations due to reliance on manual trial and error.

[0006] To achieve the above objectives, the present invention provides a method for intelligent prediction and optimization of transformer electromagnetic parameters, comprising:

[0007] The design parameters and corresponding test parameters of multiple historical transformer prototypes were obtained to form a historical sample set.

[0008] Based on the historical sample set, a first type of prediction model suitable for prediction within the sample space and a second type of prediction model suitable for prediction outside the sample space are trained respectively.

[0009] The design parameters of the transformer to be predicted are obtained, and the position of the design parameters in the feature space is determined. If the design parameters are located within the sample space determined by the historical sample set, the first type of prediction model is used to predict the electromagnetic parameters and obtain the predicted value. If the design parameters are located outside the sample space, the second type of prediction model is used to predict the electromagnetic parameters and obtain the predicted value.

[0010] Calculate the deviation between the predicted value and the measured value of the transformer to be predicted, and determine whether the deviation is less than a preset deviation; if yes, output the predicted value of the transformer to be predicted; if no, trigger an alarm, add the current sample to the historical sample set and update the corresponding prediction model.

[0011] Further, the step of training a first-class prediction model suitable for prediction within the sample space and a second-class prediction model suitable for prediction outside the sample space based on the historical sample set includes: preprocessing the historical sample set; wherein, the preprocessing includes encoding the categorical variables in the design parameters and the experimental parameters; dividing the preprocessed historical sample set into a training set and a test set; training and parameter optimization of models in the first-class candidate model pool and the second-class candidate model pool on the training set to obtain multiple first-class candidate models and multiple second-class candidate models; evaluating the performance of the multiple first-class candidate models and the multiple second-class candidate models on the test set; based on the evaluation results, selecting the model with the best performance from the multiple first-class candidate models as the first-class prediction model, selecting the model with the best performance from the multiple second-class candidate models as the second-class prediction model, and serializing and saving the first-class prediction model and the second-class prediction model respectively.

[0012] Furthermore, the step of evaluating the performance of the plurality of first-class candidate models and the plurality of second-class candidate models on the test set includes: inputting the design parameters from any set of the test set into each candidate model for calculation to obtain predicted values ​​of multiple electromagnetic parameters predicted by the plurality of candidate models; scoring all predicted values ​​based on a preset model performance evaluation criterion to obtain a comprehensive score index; and sorting based on all comprehensive score indices; and selecting the model with the highest comprehensive score index as the model with the best performance.

[0013] Furthermore, the step of scoring all predicted values ​​of the candidate model on the test set based on the preset model performance evaluation criteria to obtain a comprehensive scoring index includes: calculating the RMSE root mean square error, MAPE mean absolute percentage error, and R² coefficient of determination for each predicted value and the test parameter corresponding to the design parameter based on the predicted values ​​of multiple electromagnetic parameters; and performing a fusion analysis on the RMSE root mean square error, MAPE mean absolute percentage error, and R² coefficient of determination corresponding to each predicted value to obtain a comprehensive scoring index.

[0014] Furthermore, in addition to the step of outputting the predicted value of the transformer to be predicted, the following parallel steps are also included: calculating at least one error evaluation index for this prediction and storing it in association with the complete information of this prediction; storing the design parameter input value, model prediction output value, and measured value corresponding to this prediction as a complete training sample pair in the historical sample database; wherein, the complete information includes at least the model version identifier that triggered this prediction, the prediction timestamp, and the associated product unique identifier; and strengthening the confidence of the generated prediction model based on the successful prediction result.

[0015] Further, the step of triggering the alarm includes: generating alarm information and pushing the alarm information to design engineers and / or algorithm engineers; wherein, the alarm information includes the results obtained by analyzing a feature set that has a high contribution to the prediction of the target electromagnetic parameters and / or the matching results of historical similar cases; and / or, the step of encoding the categorical variables in the design parameters and the test parameters includes: identifying the categorical variables in the design parameters and test parameters of multiple historical transformer prototypes, and performing numerical encoding conversion on the identified categorical variables to generate numerical features that can be processed by a machine learning model; wherein, the categorical variables include at least one of product identification and specification parameters, operating environment parameters, structural material parameters, and design target parameters.

[0016] Furthermore, the first type of prediction model is set as a tree-based machine learning model, which includes gradient boosting decision tree model and random forest model; wherein, the gradient boosting decision tree model includes XGBoost model or LightGBM model; and / or, the second type of prediction model is set as a regression analysis-based prediction model, which includes linear regression model, multinomial regression model, Gaussian process regression model and quantile regression model.

[0017] In other embodiments, a transformer electromagnetic parameter intelligent prediction and optimization system is provided, capable of executing the transformer electromagnetic parameter intelligent prediction and optimization method described above; the system includes: a data acquisition module configured to acquire design parameters of multiple historical transformer prototypes and their corresponding test parameters, as well as the design parameters of the transformer to be predicted; a training module configured to, based on the historical sample set, train a first type of prediction model suitable for prediction within the sample space and a second type of prediction model suitable for prediction outside the sample space; and add samples whose deviation between the predicted value and the measured value is greater than or equal to a preset deviation to the historical sample set and update the corresponding prediction model; a prediction module configured to determine the position of the design parameter in the feature space; if it is located within the sample space determined by the historical sample set, then the first type of prediction model is used to predict the electromagnetic parameter to obtain a predicted value; if it is located outside the sample space, then the second type of prediction model is used to predict the electromagnetic parameter to obtain a predicted value; a response and self-optimization module configured to calculate the deviation between the predicted value and the measured value of the transformer to be predicted, and determine whether the deviation is less than a preset deviation; if yes, then output the predicted value of the transformer to be predicted; if no, then trigger an alarm.

[0018] In other embodiments, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the intelligent prediction and optimization method for transformer electromagnetic parameters as described above and applies it to the intelligent prediction and optimization system for transformer electromagnetic parameters described above.

[0019] In other embodiments, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, it implements the intelligent prediction and optimization method for transformer electromagnetic parameters as described above and applies it to the intelligent prediction and optimization system for transformer electromagnetic parameters described above.

[0020] Based on the foregoing description, those skilled in the art will understand that this invention establishes a historical sample set and trains dedicated models for interpolation and extrapolation scenarios respectively. During actual prediction, the system automatically determines the most suitable model for the transformer under test based on the design parameters and calls it. Finally, the prediction results are verified by comparing the predicted values ​​with the measured values. If the results are not ideal, an alarm and model update mechanism are triggered. This not only avoids the problem of low efficiency caused by repeated trial and error due to human experience in comparing predicted and measured values, but also solves the problem of systematic deviation between theoretical and measured values ​​caused by changes in process and materials in traditional electromagnetic calculation methods. It also makes up for the insufficient ability of a single prediction model to handle interpolation and extrapolation scenarios, improves the first-time success rate of design, and effectively reduces R&D costs and cycle. Attached Figure Description

[0021] The accompanying drawings, as part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention, but do not constitute an undue limitation of the invention. Obviously, the drawings described below are merely some embodiments, and those skilled in the art can obtain other drawings based on these drawings without creative effort. In the drawings:

[0022] Figure 1 This is a flowchart of a transformer electromagnetic parameter intelligent prediction and optimization system in some embodiments of the present invention;

[0023] Figure 2 The flowchart below shows a transformer electromagnetic parameter intelligent prediction and optimization system in some other embodiments of the present invention.

[0024] Explanation of reference numerals in the attached diagram:

[0025] 100. Intelligent prediction and optimization system for transformer electromagnetic parameters;

[0026] 1. Data acquisition module; 2. Prediction module; 3. Response and self-optimization module; 4. Storage module; 5. Preprocessing module; 6. Training module; 7. Evaluation module. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0028] Those skilled in the art should understand that the embodiments described below are merely a part of the embodiments of the present invention, and not all of the embodiments of the present invention. These partial embodiments are intended to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention. Based on the embodiments provided by the present invention, all other embodiments obtained by those skilled in the art without creative effort should still fall within the scope of protection of the present invention.

[0029] The following reference Figure 1 and Figure 2 This document will provide a detailed description of the intelligent prediction and optimization method and system for transformer electromagnetic parameters in some embodiments of the present invention. Figure 1 This is a flowchart of a transformer electromagnetic parameter intelligent prediction and optimization system in some embodiments of the present invention; Figure 2 The flowchart below shows a transformer electromagnetic parameter intelligent prediction and optimization system in some other embodiments of the present invention.

[0030] In some embodiments of the present invention, an intelligent prediction and optimization method for transformer electromagnetic parameters is provided. This method aims to address the problem of systematic deviations between theoretical and measured values ​​caused by variations in manufacturing processes and materials in traditional electromagnetic calculation methods. Furthermore, it compensates for the limitations of single prediction models in handling interpolation and extrapolation scenarios. Specifically, by establishing a historical sample set, dedicated models are trained for both interpolation and extrapolation scenarios. During actual prediction, the system automatically determines and calls upon the most suitable model for the transformer under test based on design parameters. Finally, the prediction results are verified by comparing the predicted and measured values. If the results are unsatisfactory, an alarm and model update mechanism are triggered, thereby avoiding the inefficiency caused by repeated trial and error based on manual experience in traditional methods of comparing predicted and measured values.

[0031] Specifically, the intelligent prediction and optimization method for transformer electromagnetic parameters includes steps S110 to S140, as follows:

[0032] Step S110: Obtain the design parameters and corresponding test parameters of multiple historical transformer prototypes to form a historical sample set.

[0033] In some specific embodiments, step S110 further includes steps S111 to S114, as follows:

[0034] Step S111: Obtain the theoretical values ​​of design parameters from the design drawings and bill of materials in the product data management system as design parameters; obtain the original measured values ​​of test parameters from the test bench data acquisition system; and obtain the actual process parameters associated with the prototype from the manufacturing execution system as test parameters. The theoretical values ​​of design parameters obtained from the product data management system include at least the magnetic flux density, electrical flux density, theoretical loss, and theoretical impedance parameters derived from the electromagnetic calculation sheet; the actual process parameters obtained from the manufacturing execution system include at least the batch code of the silicon steel sheet material and the winding process identifier.

[0035] Specifically, the design parameters described in this invention include parameters of the following types: product identification and specifications, working environment, structural materials, and design objectives.

[0036] Among them, product identification and specification parameters include product codes, model numbers, and sub-product lines, among other related parameters. These are unique identifiers. Product codes are used to track specific product instances, serving as an index in the dataset or for associating with other data; they are not considered model features. Model numbers are product series identifiers, encompassing design platform standards and technical specifications, and are important category features. Sub-product lines are market / application classifications, such as "distribution," "electricity," and "dry-type," and are key category features that directly impact design rules and performance benchmarks.

[0037] Environmental parameters include capacity, primary voltage, secondary voltage, rated frequency, and altitude. Capacity is the core rated parameter, determining the transformer's size, losses, and cost level; it is the starting point for all design calculations and a key numerical characteristic. Primary voltage reflects the rated electrical conditions, determining insulation design, number of turns, etc., and is an important numerical characteristic. Secondary voltage also reflects rated electrical conditions and is an important numerical characteristic. Rated frequency reflects the system electrical conditions (e.g., 50Hz / 60Hz), affecting core losses and reactance calculations, and is a category or numerical characteristic. Altitude reflects the transformer's operating environment conditions, affecting heat dissipation and insulation, and is a numerical characteristic.

[0038] Structural material parameters include material symbols, heat dissipation structures, core diameter, weight, and weight factor. Material symbols are core material properties (e.g., silicon steel sheet grade), directly determining permeability and iron loss, and are a key category characteristic leading to differences between theoretical and measured values. Heat dissipation structures (e.g., ONAN, OFAF) determine temperature rise and heat dissipation capacity, and are a key category characteristic. Core diameter is a core geometric dimension, directly related to the magnetic circuit cross-sectional area, affecting magnetic flux density, iron weight, and no-load performance, and is a core numerical characteristic. Weight is the core of the overall transformer's physical dimensions, related to material usage and cost, and is a derived numerical characteristic. The weight factor is an indicator of design compactness or material utilization, and is also a derived numerical characteristic.

[0039] Design target parameters include high-voltage rated current, low-voltage rated current, high-voltage current density, basic current density, phase-shifted current density, magnetic flux density, unit iron loss, high-voltage design temperature rise, high-voltage standard temperature rise, low-voltage design temperature rise, low-voltage standard temperature rise, theoretical no-load loss, theoretical no-load current, theoretical load loss, and theoretical load impedance. High-voltage rated current and low-voltage rated current are electrical parameters calculated based on capacity and voltage; they are design input conditions and belong to numerical characteristics. High-voltage current density is the winding current density selected by the operator, affecting winding losses, temperature rise, and cost; it is a core decision characteristic. Basic current density refers to low-voltage or average current density and is a core decision characteristic. Phase-shifted current density is a specific design parameter, such as that applicable to phase-shifting transformers. Magnetic flux density is the core magnetic flux density selected by the operator, affecting iron loss, no-load current, and core size; it is one of the most crucial decision characteristics. Unit iron loss is the theoretical iron loss value obtained by looking up tables or calculating based on material grade and magnetic flux density; it is the basis for no-load loss calculation and an important intermediate calculation characteristic. High-voltage design temperature rise and low-voltage design temperature rise are the winding temperature rise values ​​expected by operators, reflecting the margin of heat dissipation design and representing numerical characteristics. High-voltage standard temperature rise and low-voltage standard temperature rise are the allowable temperature rise limits specified by technical standards, i.e., design constraints, and belong to numerical characteristics. Theoretical no-load loss, theoretical no-load current, theoretical load loss, and theoretical load impedance are expected values ​​based on electromagnetic calculations and are one of the core prediction targets. Subsequent models need to predict the difference between these and actual measurements.

[0040] The test parameters described in this invention include measured no-load loss, measured load loss, measured no-load current, and measured load impedance, among other related parameters.

[0041] Step S112: Based on the product's unique code or a composite key formed by the combination of model, production batch, and serial number, the design parameters, process parameters, and test parameters of the same prototype from different sources are associated and matched to form preliminary sample data pairs.

[0042] Step S113: Perform data quality rule checks on the associated sample data, including at least logical consistency checks and physical range checks; for key features found to be missing during the checks, imputation is performed based on statistical values ​​of the same model of product or domain knowledge formulas, and the imputation marks are recorded to finally form a historical sample set.

[0043] The logical consistency verification step includes: based on basic electrical engineering principles, verifying whether the mathematical relationship between the rated capacity, rated voltage, and rated current values ​​in the sample conforms to a preset physical formula, avoiding basic input errors such as unit confusion (A and kA) and system integration errors such as incorrect field mapping when data is migrated from multiple old systems or integrated into a new platform. The physical range verification step includes: based on the transformer's material properties, insulation class, and cooling method, determining whether the magnetic flux density, electrical flux density, and temperature rise values ​​in the sample are within a preset reasonable physical range, avoiding data that violates physical laws. When a sample fails the logical consistency verification or physical range verification, it is marked as a sample to be verified, and a data correction process is triggered. The data correction process includes repairing based on the correlation of all input data or reminding the operator to perform repairs.

[0044] Step S114: After forming the historical sample set, generate and store versioned metadata for the sample set. The metadata includes at least the time range of data collection, the data source version, the number of samples, and the version of the verification rules applied.

[0045] Step S120: Based on the historical sample set, a first type of prediction model suitable for prediction within the sample space and a second type of prediction model suitable for prediction outside the sample space are trained respectively.

[0046] Step S120 further includes steps S121 to S124, as follows:

[0047] Step S121: Preprocess the historical sample set.

[0048] Step S121, “Preprocessing the historical sample set,” includes a data cleaning step, which includes at least missing value handling and outlier handling.

[0049] In some specific examples, the missing value handling steps include: for a missing continuous design parameter in a sample, selecting other samples from the historical sample set that are identical to that sample in terms of key category characteristics, calculating the statistical value of the parameter in these similar samples, and using the statistical value to fill the missing value. The key category characteristics at least include product line type and voltage level; the specific key category characteristics can be set based on the correlation between actual parameters, and are not specifically limited here.

[0050] In some specific examples, the outlier handling steps include: using statistical methods to identify candidate outliers in the data; comparing the parameter values ​​of the candidate outliers with a pre-set physical rationality rule base; if the value violates the physical rationality rule, it is determined to be erroneous data and is removed or corrected; if the value conforms to the physical rule but belongs to the statistical extreme value, it is marked as a potentially valid sample and retained.

[0051] The physical rationality rule base defines at least the upper and lower limits of reasonable values ​​for parameters such as no-load current, temperature rise, and magnetic flux density. The specific physical rationality rule base can be established based on physics knowledge and the experience of operators in actual transformer production and applications; no specific limitations are specified here.

[0052] In other examples, the data cleaning step also includes: generating a cleaning report that records the missing value imputation, outlier handling actions and reasons for each sample, and storing the report in association with the cleaned dataset.

[0053] Step S121, "Preprocessing the historical sample set," further includes: encoding the categorical variables in the design parameters and the test parameters. The categorical variables include at least one of the following: product identification and specification parameters, operating environment parameters, structural material parameters, and design target parameters.

[0054] In some specific embodiments, the steps of encoding categorical variables in design parameters and experimental parameters include: identifying feature variables in the historical sample set, identifying features with a finite number of discrete values ​​as categorical variables; selecting an encoding method for each categorical variable according to a predefined feature type mapping table; wherein, nominal categorical variables are assigned one-hot encoding, and ordered categorical variables are assigned label encoding; and calling the corresponding encoder to transform each categorical variable to generate a numerical feature matrix that can be processed by the machine learning model.

[0055] The nominal category variables must include at least product line classification, material grade identification, and heat dissipation structure type. The ordinal category variables must include at least insulation class.

[0056] In some specific examples, for a nominal categorical variable, the specific steps of the one-hot encoding transformation are as follows: create k new binary feature columns for the k different values ​​of the variable, such that each row of samples takes the value 1 only in the feature column corresponding to its own value, and takes the value 0 in the other k-1 feature columns.

[0057] In other specific examples, for ordered category variables, the label encoding conversion steps are as follows: based on a predefined mapping relationship that reflects the inherent hierarchical order of the categories, each category value is mapped to an ordered integer.

[0058] Step S122: Divide the preprocessed historical sample set into a training set and a test set, with the training set accounting for 80% and the test set accounting for 20%.

[0059] After step S122 and before step S123, the method further includes: training and parameter optimization of models in the first type of candidate model pool and the second type of candidate model pool on the training set to obtain multiple first type candidate models and multiple second type candidate models.

[0060] Specifically, a feature matrix and target variables are defined for the preprocessed training set. A list of model algorithms suitable for interpolation is selected as the first candidate model pool, such as [Random Forest, XGBoost, LightGBM]. A list of model algorithms suitable for extrapolation is selected as the second candidate model pool, such as [Linear Regression, Multinomial Regression, Gaussian Process Regression]. For each candidate algorithm, within its defined hyperparameter space, the parameter configuration that achieves optimal generalization performance on the given training set is searched, and the final model under that configuration is trained. Ultimately, two sets are formed: an optimized first-class candidate model set and an optimized second-class candidate model set. Each model is an instance of its corresponding algorithm under optimal or near-optimal parameter configurations.

[0061] Step S123: Evaluate the performance of multiple first-class candidate models and multiple second-class candidate models on the test set.

[0062] The steps for evaluating the performance of multiple first-class candidate models and multiple second-class candidate models on the test set include: inputting the design parameters from any set of test sets into each candidate model to calculate the predicted values ​​of multiple electromagnetic parameters predicted by the multiple candidate models; scoring all predicted values ​​based on preset model performance evaluation criteria to obtain a comprehensive score index; ranking all comprehensive score indices; and selecting the model with the highest comprehensive score index as the best performing model.

[0063] The step of scoring all predicted values ​​of candidate models on the test set based on preset model performance evaluation criteria to obtain a comprehensive score index includes: calculating the root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²) of each predicted value and the corresponding experimental parameters based on the predicted values ​​of multiple electromagnetic parameters; and performing a fusion analysis on the RMSE, MAPE, and R² of determination corresponding to each predicted value to obtain a comprehensive score index.

[0064] RMSE (Root Mean Square Error) measures the average difference between the model's predicted value and the actual value, and is calculated using the following formula: ; where the actual value is the value obtained through experimentation from the experimental parameters.

[0065] The Mean Absolute Percentage Error (MAPE) measures the average percentage of prediction error relative to the true value. The formula is as follows:

[0066] ;

[0067] Where n is the total number of samples in the test set; the actual value is the value obtained through experiments in the experimental parameters.

[0068] The R² coefficient of determination represents the proportion of variance captured by the model to the total variance, and is calculated using the following formula:

[0069] ;

[0070] in, This is the actual value. For predicted values, This is the average of the actual values, which are the values ​​obtained through experiments from the test parameters.

[0071] Then, weights are assigned to RMSE (root mean square error), MAPE (mean absolute percentage error), and R² (coefficient of determination), and the results are summed to obtain a comprehensive scoring index. The weights for each of these three parameters can be adjusted based on the operator's actual experimental conditions to suit the corresponding weight; specific weight values ​​are not limited here.

[0072] Step S124: Based on the evaluation results, select the best-performing model from multiple first-class candidate models as the first-class prediction model, select the best-performing model from multiple second-class candidate models as the second-class prediction model, and serialize and save the first-class prediction model and the second-class prediction model respectively.

[0073] By selecting the most accurate model from the first category of candidate models in regions with dense historical data, and selecting the optimal tree model through multiple evaluation metrics (such as RMSE, MAPE, R²), the prediction accuracy for conventional designs can be guaranteed to the greatest extent, directly reducing design iterations and prototype waste.

[0074] In some specific embodiments, the first type of prediction model is set as a tree-based machine learning model, which includes gradient boosting decision tree model and random forest model; wherein, gradient boosting decision tree model includes XGBoost model or LightGBM model.

[0075] In regions with sparse or blank data, security and stability are more important than the limit of accuracy. The second type of candidate model, by evaluating the selected optimal regression model and updating the model, can provide the most reliable risk boundary prediction, enabling the system to give relatively correct results when faced with special designs.

[0076] In some other specific embodiments, the second type of prediction model is set as a prediction model based on regression analysis, and the regression model includes at least a linear regression model, a multinomial regression model, a Gaussian process regression model, and a quantile regression model.

[0077] Step S130: Obtain the design parameters of the transformer to be predicted and determine the position of the design parameters in the feature space; if it is located within the sample space determined by the historical sample set, the first type of prediction model is used to predict the electromagnetic parameters and obtain the predicted value; if it is located outside the sample space, the second type of prediction model is used to predict the electromagnetic parameters and obtain the predicted value.

[0078] Step S140: Calculate the deviation between the predicted value and the measured value of the transformer to be predicted, and determine whether the deviation is less than a preset deviation.

[0079] If so, the predicted value of the transformer to be predicted is output. Simultaneously with the step of outputting the predicted value of the transformer to be predicted, the following parallel steps are also included: calculating at least one error evaluation index for this prediction and storing it in association with the complete information of this prediction; storing the design parameter input value, model prediction output value, and measured value corresponding to this prediction as a complete training sample pair in the historical sample database; wherein, the complete information includes at least the model version identifier that triggered this prediction, the prediction timestamp, and the associated unique product identifier; and, based on the successful prediction result, strengthening the confidence of the generated prediction model. The error indices include RMSE (Root Mean Square Error), MAPE (Mean Absolute Percentage Error), and R² (Coefficient of Determination).

[0080] In some specific examples, the steps to enhance the confidence of the prediction model can be as follows: using the design parameters and measured values ​​of the currently successfully predicted samples, perform online learning updates on the generated prediction model to improve the prediction confidence of the model under similar input patterns or reduce its prediction uncertainty.

[0081] The specific steps for online learning updates are as follows: analyze the individual prediction contribution of each model in the model pool to the current successful samples; and dynamically adjust the weight of each model in the model pool based on the closeness between the prediction results of each model and the measured values, so that the model with more accurate predictions receives higher weights.

[0082] If not, an alarm is triggered, the current sample is added to the historical sample set, and the corresponding prediction model is updated. The steps for triggering an alarm include: generating alarm information and pushing the alarm information to design engineers and / or algorithm engineers. The alarm information includes results obtained from analysis of feature sets that have a high contribution to the prediction of the target electromagnetic parameters and / or matching results from historical similar cases.

[0083] The intelligent prediction and optimization system 100 for transformer electromagnetic parameters will be described below. The intelligent prediction and optimization system 100 for transformer electromagnetic parameters of the present invention can execute any of the intelligent prediction and optimization methods for transformer electromagnetic parameters described above.

[0084] like Figure 1 and Figure 2 As shown, in other embodiments of the present invention, a transformer electromagnetic parameter intelligent prediction and optimization system 100 is provided, which can execute the transformer electromagnetic parameter intelligent prediction and optimization method described in any one of the above descriptions.

[0085] System 100 includes a data acquisition module 1, a training module 6, a prediction module 2, and a response and self-optimization module 3. The data acquisition module 1 is configured to acquire the design parameters of multiple historical transformer prototypes and their corresponding test parameters, as well as the design parameters of the transformer to be predicted. The training module 6 is configured to, based on the historical sample set, train a first-type prediction model suitable for prediction within the sample space and a second-type prediction model suitable for prediction outside the sample space; and add samples whose predicted values ​​deviate from the measured values ​​by a preset deviation to the historical sample set and update the corresponding prediction models. The prediction module 2 is configured to determine the position of the design parameters in the feature space; if they are within the sample space determined by the historical sample set, the first-type prediction model is used to predict the electromagnetic parameters, obtaining the predicted value; if they are outside the sample space, the second-type prediction model is used to predict the electromagnetic parameters, obtaining the predicted value. The response and self-optimization module 3 is configured to calculate the deviation between the predicted value and the measured value of the transformer to be predicted, and determine whether the deviation is less than a preset deviation; if yes, the predicted value of the transformer to be predicted is output; if no, an alarm is triggered.

[0086] In some embodiments, system 100 further includes a preprocessing module 5, configured to preprocess the historical sample set; wherein the preprocessing step includes encoding the categorical variables in the design parameters and the experimental parameters. In this embodiment, training module 6 is further configured to: divide the preprocessed historical sample set into a training set and a test set; train and optimize the models in the first and second candidate model pools on the training set to obtain multiple first-class candidate models and multiple second-class candidate models; evaluate the performance of the multiple first-class candidate models and multiple second-class candidate models on the test set; based on the evaluation results, select the model with the best performance from the multiple first-class candidate models as the first-class prediction model, select the model with the best performance from the multiple second-class candidate models as the second-class prediction model, and serialize and save the first-class prediction model and the second-class prediction model respectively.

[0087] In some embodiments, the prediction module 2 is further configured to input design parameters from any set of the test sets into each candidate model for calculation, thereby obtaining predicted values ​​of multiple electromagnetic parameters predicted by the multiple candidate models. In this embodiment, the system 100 further includes an evaluation module 7, which is configured to score all predicted values ​​based on preset model performance evaluation criteria to obtain a comprehensive score index; sort all comprehensive score indices; and select the model with the highest comprehensive score index as the best-performing model.

[0088] In some embodiments, the evaluation module 7 is further configured to: calculate the root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²) of each predicted value and the test parameter corresponding to the design parameter based on the predicted values ​​of the multiple electromagnetic parameters; and perform a fusion analysis on the RMSE, MAPE, and R² corresponding to each predicted value to obtain a comprehensive scoring index.

[0089] In some embodiments, the evaluation module 7 is further configured to, while outputting the predicted value of the transformer to be predicted, also include the following parallel step: calculating at least one error evaluation index for this prediction. In this embodiment, the system 100 further includes a storage module 4, which is configured to: store the complete information associated with this prediction; store the design parameter input value, model prediction output value, and measured value corresponding to this prediction as a complete training sample pair in the historical sample database; wherein, the complete information includes at least the model version identifier that triggered this prediction, the prediction timestamp, and the associated product unique identifier; and strengthen the confidence of the generated prediction model based on the successful prediction result.

[0090] In some embodiments, the response and self-optimization module 3 is configured to: generate alarm information and push the alarm information to design engineers and / or algorithm engineers; wherein the alarm information includes the results obtained by analyzing a feature set that has a high contribution to the prediction of the target electromagnetic parameters and / or the matching results of historical similar cases.

[0091] In some embodiments, the preprocessing module 5 is further configured to identify categorical variables in the design parameters and test parameters of multiple historical transformer prototypes, and to perform numerical encoding conversion on the identified categorical variables to generate numerical features that can be processed by the machine learning model; wherein, the categorical variables include at least one of product identification and specification parameters, operating environment parameters, structural material parameters, and design target parameters.

[0092] The first type of prediction model is set as a tree-based machine learning model, which includes gradient boosting decision tree model and random forest model; wherein the gradient boosting decision tree model includes XGBoost model or LightGBM model; and / or, the second type of prediction model is set as a regression analysis-based prediction model, which includes linear regression model, multinomial regression model, Gaussian process regression model and quantile regression model.

[0093] In other embodiments of the present invention, a computer device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the computer program, it implements the relevant steps of the intelligent prediction and optimization method for transformer electromagnetic parameters as described above and applies them to the intelligent prediction and optimization system 100 for transformer electromagnetic parameters described above.

[0094] In other embodiments of the present invention, a computer-readable storage medium is also provided, which stores a computer program, characterized in that the computer program, when executed by a processor, implements the intelligent prediction and optimization method for transformer electromagnetic parameters as described above and is applied to the intelligent prediction and optimization system 100 for transformer electromagnetic parameters described above.

[0095] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, prediction models, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0096] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0097] Those skilled in the art will understand that this invention establishes a historical sample set and trains dedicated models for interpolation and extrapolation scenarios respectively. During actual prediction, the system automatically determines the most suitable model for the transformer under test based on the design parameters and calls it. Finally, the prediction results are verified by comparing the predicted values ​​with the measured values. If the results are not ideal, an alarm and model update mechanism are triggered. This not only avoids the problem of low efficiency caused by repeated trial and error due to human experience in comparing predicted and measured values, but also solves the problem of systematic deviation between theoretical and measured values ​​caused by changes in process and materials in traditional electromagnetic calculation methods. It also makes up for the insufficient ability of a single prediction model to handle interpolation and extrapolation scenarios, improves the first-time success rate of design, and effectively reduces R&D costs and cycle.

[0098] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination can be used to combine all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or device so disclosed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose.

[0099] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are intended to be within the scope of the invention and form different embodiments. For example, in the claims, any of the claimed embodiments can be used in any combination.

[0100] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-described technical content to create equivalent embodiments without departing from the scope of the present invention. The implementation schemes in the above embodiments can be further combined or replaced. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for intelligent prediction and optimization of transformer electromagnetic parameters, characterized in that, include: The design parameters and corresponding test parameters of multiple historical transformer prototypes were obtained to form a historical sample set. Based on the historical sample set, a first type of prediction model suitable for prediction within the sample space and a second type of prediction model suitable for prediction outside the sample space are trained respectively. The design parameters of the transformer to be predicted are obtained, and the position of the design parameters in the feature space is determined. If the design parameters are located within the sample space determined by the historical sample set, the first type of prediction model is used to predict the electromagnetic parameters and obtain the predicted value. If the design parameters are located outside the sample space, the second type of prediction model is used to predict the electromagnetic parameters and obtain the predicted value. Calculate the deviation between the predicted value and the measured value of the transformer to be predicted, and determine whether the deviation is less than a preset deviation; if yes, output the predicted value of the transformer to be predicted; if no, trigger an alarm, add the current sample to the historical sample set and update the corresponding prediction model.

2. The method according to claim 1, characterized in that, The steps of training a first-class prediction model suitable for prediction within the sample space and a second-class prediction model suitable for prediction outside the sample space based on the historical sample set include: The historical sample set is preprocessed; wherein, the preprocessing includes encoding the categorical variables in the design parameters and the experimental parameters; The preprocessed historical sample set is divided into a training set and a test set; On the training set, models in the first type of candidate model pool and the second type of candidate model pool are trained and their parameters are optimized to obtain multiple first type candidate models and multiple second type candidate models; The performance of multiple first-type candidate models and multiple second-type candidate models is evaluated on the test set; Based on the evaluation results, the model with the best performance is selected from multiple first-class candidate models as the first-class prediction model, and the model with the best performance is selected from multiple second-class candidate models as the second-class prediction model. The first-class prediction model and the second-class prediction model are then serialized and saved respectively.

3. The method according to claim 2, characterized in that, The step of evaluating the performance of the plurality of first-class candidate models and the plurality of second-class candidate models on the test set includes: The design parameters of any set of test parameters are input into each candidate model for calculation, and the predicted values ​​of multiple electromagnetic parameters predicted by multiple candidate models are obtained. Based on the preset model performance evaluation criteria, all predicted values ​​are scored to obtain a comprehensive score index; and all comprehensive score indices are then ranked. The model with the highest comprehensive score index is selected as the best performing model.

4. The method according to claim 3, characterized in that, The step of scoring all predicted values ​​of the candidate model on the test set based on preset model performance evaluation criteria to obtain a comprehensive score index includes: Based on the predicted values ​​of multiple electromagnetic parameters, the root mean square error of RMSE, the mean absolute percentage error of MAPE, and the coefficient of determination of R² are calculated for each predicted value and the test parameter corresponding to the design parameter. The root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²) corresponding to each predicted value are integrated and analyzed to obtain a comprehensive scoring index.

5. The method according to claim 1, characterized in that, Simultaneously with the step of outputting the predicted value of the transformer to be predicted, the following parallel steps are also included: Calculate at least one error evaluation index for this prediction and store it in association with the complete information of this prediction; store the design parameter input value, model prediction output value and measured value corresponding to this prediction as a complete training sample pair in the historical sample database; wherein, the complete information includes at least the model version identifier that triggered this prediction, the prediction timestamp and the associated product unique identifier; Based on the successful prediction results, the confidence level of the generated prediction model was enhanced.

6. The method according to claim 2, characterized in that, The step of triggering the alarm includes: generating alarm information and pushing the alarm information to design engineers and / or algorithm engineers; wherein, the alarm information includes results obtained from analysis based on feature sets that have a high contribution to the prediction of the target electromagnetic parameters and / or matching results of historical similar cases; and / or, The step of encoding the categorical variables in the design parameters and test parameters includes: identifying the categorical variables in the design parameters and test parameters of multiple historical transformer prototypes, and performing numerical encoding conversion on the identified categorical variables to generate numerical features that can be processed by the machine learning model; wherein, the categorical variables include at least one of product identification and specification parameters, operating environment parameters, structural material parameters, and design target parameters.

7. The method according to claim 1, characterized in that, The first type of prediction model is set as a tree-based machine learning model, which includes gradient boosting decision tree models and random forest models; wherein, the gradient boosting decision tree model includes XGBoost or LightGBM models; and / or, The second type of prediction model is set as a prediction model based on regression analysis, and the regression model includes linear regression model, multinomial regression model, Gaussian process regression model and quantile regression model.

8. A smart prediction and optimization system for transformer electromagnetic parameters, characterized in that, The system is capable of executing the intelligent prediction and optimization method for transformer electromagnetic parameters as described in any one of claims 1 to 7; the system comprises: The data acquisition module is configured to acquire the design parameters of multiple historical transformer prototypes and their corresponding test parameters, as well as the design parameters of the transformer to be predicted. The training module is configured to train a first-class prediction model suitable for prediction within the sample space and a second-class prediction model suitable for prediction outside the sample space based on the historical sample set; and to add samples whose deviation between the predicted value and the measured value is greater than or equal to a preset deviation to the historical sample set and update the corresponding prediction model. The prediction module is configured to determine the position of the design parameters in the feature space; if the parameters are located within the sample space determined by the historical sample set, the first type of prediction model is used to predict the electromagnetic parameters and obtain the predicted value; if the parameters are located outside the sample space, the second type of prediction model is used to predict the electromagnetic parameters and obtain the predicted value. The response and self-optimization module is configured to calculate the deviation between the predicted value and the measured value of the transformer to be predicted, and determine whether the deviation is less than a preset deviation; if so, the predicted value of the transformer to be predicted is output; if not, an alarm is triggered.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the intelligent prediction and optimization method for transformer electromagnetic parameters as described in any one of claims 1 to 7 and applies it to the intelligent prediction and optimization system for transformer electromagnetic parameters as described in claim 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the intelligent prediction and optimization method for transformer electromagnetic parameters as described in any one of claims 1 to 7 and applies it to the intelligent prediction and optimization system for transformer electromagnetic parameters as described in claim 8.

Images