Method and device for predicting magnetic core loss of magnetic element

Abstract

The invention relates to a magnetic core loss prediction method and device for a magnetic element, and the method comprises the steps: obtaining historical data sets of different magnetic core materials, thereby considering the influence of the temperature on the magnetic core loss, carrying out the temperature correction of a preset SteinMazz equation according to the historical data sets, obtaining a target SteinMazz equation, and carrying out the prediction of the magnetic core loss of the magnetic element. Therefore, loss prediction can be carried out on the magnetic core material in the historical data set through the target Steinmaz equation, and an updated data set of the prediction value of the temperature factor is obtained. Furthermore, the influence of the magnetic flux density data and the frequency on the magnetic core loss can be considered, the updated data set is merged according to the magnetic core material, the magnetic flux density data and the frequency, a merged data set is obtained, and then the preset prediction model can be trained through the merged data set. The training of the preset prediction model on the magnetic flux density data and frequency is added, so that the predicted magnetic core loss of the target prediction model on the magnetic element better conforms to the actual situation.

Description

Technical Field

[0001] The present invention relates to the technical field of magnetic components, and in particular to a method and device for predicting the core loss of a magnetic component. Background Art

[0002] In recent years, power electronics technology has rapidly developed, with magnetic components widely used. A wide variety of power electronic converters have been developed based on this technology. With the advancement of modern technology, customer requirements for these devices have also increased, demanding not only smaller size but also higher efficiency and power density. According to statistics, the volume and weight of magnetic components generally account for approximately 20% to 30% of the entire device, and losses contribute approximately 30% of the total losses, significantly impacting the power and power density of the entire device.

[0003] The traditional method for calculating core material loss is usually based on test data provided by the manufacturer. This data is often collected from unit loss values at different frequencies and different magnetic flux densities, and most of them are obtained based on a standard sinusoidal excitation voltage. However, in actual situations, the operating conditions of magnetic components are not single, and the loss values calculated using traditional calculation methods are not in line with reality. A more accurate calculation method is undoubtedly necessary.

[0004] Therefore, it is urgent to propose a method and device for predicting the core loss of a magnetic component to solve the technical problem that the prediction of magnetic components in the prior art is based on the test data provided by the manufacturer, resulting in the predicted loss value of the core material loss not being consistent with the actual loss value. Summary of the Invention

[0005] In view of this, it is necessary to provide a method and device for predicting the core loss of a magnetic component to solve the technical problem in the prior art that the loss value of the core material loss prediction does not conform to reality.

[0006] In order to solve the above problems, in a first aspect, the present invention provides a method for predicting core loss of a magnetic component, comprising: Acquire a historical data set of a magnetic component; the historical data set includes magnetic flux density data and frequency of different magnetic core materials; Performing temperature correction on a preset Steinmetz equation according to the historical data set to obtain a target Steinmetz equation, and performing loss prediction on the magnetic core material in the historical data set using the target Steinmetz equation to obtain an updated data set; Merging the updated data set according to the magnetic core material, the magnetic flux density data, and the frequency to obtain a merged data set; A preset prediction model is trained according to the combined data set to obtain a target prediction model, and core loss prediction is performed on the magnetic component according to the target prediction model.

[0007] In a possible implementation, the historical data set further includes an excitation waveform; and before correcting the preset Steinmetz equation according to the historical data set to obtain the target Steinmetz equation, the method further includes: Preprocessing the sample data of the different magnetic core materials in the historical data set to obtain a preprocessed training set; Performing feature extraction on the magnetic flux density data and the excitation waveform in the preprocessing training set to obtain feature data of each sample data; The preprocessed training set is classified and updated according to the feature data to obtain an updated historical data set.

[0008] In a possible implementation, classifying and updating the preprocessed training set according to the feature data to obtain an updated historical data set includes: Training a preset classification model based on the sample data in the preprocessing training set and the feature data to obtain a target classification model; Perform classification prediction on the preprocessed training set according to the target classification model to obtain corresponding classification results; The preprocessing training set is updated according to the classification result to obtain an updated historical data set.

[0009] In a possible implementation, performing temperature correction on a preset Steinmetz equation according to the historical data set to obtain a target Steinmetz equation includes: Performing coefficient fitting on the preset Steinmetz equation according to the historical data set to obtain a plurality of fitting coefficients; The preset Steinmetz equation is temperature-corrected according to the multiple fitting coefficients and a preset temperature function to obtain a target Steinmetz equation.

[0010] In a possible implementation, the historical data set further includes temperature; and merging the updated data set according to the magnetic core material, the magnetic flux density data, and the frequency to obtain a merged data set includes: Performing one-hot encoding on the excitation waveform, the temperature, and the magnetic core material in the updated data set and reindexing to obtain a one-hot encoded column; Determining a corresponding magnetic flux density peak value according to the magnetic flux density data in the updated data set; The one-hot encoded columns, the magnetic flux density peaks, and the frequencies are merged in a column direction to obtain a merged data set.

[0011] In a possible implementation, the excitation waveform includes a sine wave, a cosine wave, and a trapezoidal wave; and preprocessing the sample data of the different magnetic core materials in the historical data set to obtain a preprocessed training set includes: Processing missing values of sample data in the historical data set based on an interpolation method to obtain a complete historical data set; The types of excitation waveforms in the complete historical data set are digitally converted according to the sine wave, the cosine wave, and the trapezoidal wave of each magnetic core material to obtain a preprocessing training set.

[0012] In a possible implementation, performing the target Steinmetz equation on the core material in the historical data set to obtain an updated data set includes: When performing core loss prediction according to the target Steinmetz equation, taking logarithms of both sides of the target Steinmetz equation to obtain a linear target Steinmetz equation; performing a linear target Steinmetz equation on the core material in the historical data set to obtain a new loss prediction; The new loss prediction is added to the historical data set to obtain an updated data set.

[0013] In one possible implementation, the process of determining the preset prediction model includes: According to the combined data set, a linear regression model, a gradient boosting regression model, and an AdaBoost regression model are trained and predicted respectively to obtain corresponding core loss prediction values and residual release data; According to the core loss prediction value and the residual release data, the AdaBoost regression model is determined to be a preset prediction model.

[0014] In a possible implementation, training a preset prediction model according to the merged data set to obtain a target prediction model includes: In the process of training the preset prediction model, the sample data in the combined data set is predicted by a genetic algorithm to obtain the corresponding target core loss; Calculating the peak value of the magnetic flux density and the frequency in the combined data set to obtain the transmitted magnetic energy; Obtaining an objective function value according to the target core loss and the transmitted magnetic energy; When the number of iterations of the genetic algorithm and / or the objective function value meet preset conditions, a target prediction model is obtained.

[0015] In a second aspect, the present invention further provides a device for predicting core loss of a magnetic component, comprising: A data acquisition module is used to acquire a historical data set of the magnetic element; the historical data set includes magnetic flux density data and frequency of different magnetic core materials; an equation correction module, configured to perform temperature correction on a preset Steinmetz equation according to the historical data set to obtain a target Steinmetz equation, and perform loss prediction on the magnetic core material in the historical data set using the target Steinmetz equation to obtain an updated data set; a data set determining module, configured to merge the updated data set according to the magnetic core material, the magnetic flux density data, and the frequency to obtain a merged data set; A model training module is used to train a preset prediction model based on the combined data set to obtain a target prediction model, and predict the core loss of the magnetic component based on the target prediction model.

[0016] The beneficial effects of the present invention are: a historical data set of different magnetic core materials is obtained, so that the influence of temperature on core loss can be considered, and the preset Steinmetz equation is temperature-corrected according to the historical data set to obtain a target Steinmetz equation, so that the loss of the magnetic core material in the historical data set can be predicted by the target Steinmetz equation, and an updated data set with the predicted value of the temperature factor is obtained; further, the influence of magnetic flux density data and frequency on core loss can be considered, and the updated data set is merged according to the magnetic core material, magnetic flux density data and frequency to obtain a merged data set, and then the preset prediction model can be trained by the merged data set, and the training of the preset prediction model on magnetic flux density data and frequency is increased, so that the target prediction model's predicted core loss of the magnetic component is more in line with the actual situation. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic flow chart of an embodiment of a method for predicting core loss of a magnetic component provided by the present invention; Figure 2 A schematic diagram of coordinates of an embodiment of the residual distribution diagrams of the three models provided by the present invention; Figure 3 A schematic diagram of an embodiment of the K-fold cross-validation accuracy curve 2 provided by the present invention Figure 4 For the present invention Figure 1 Another embodiment flow chart before step S102; Figure 5 A schematic diagram of an embodiment of a curve showing the relationship between different temperatures T_i and ln(k_i) provided by the present invention; Figure 6 A schematic diagram of an embodiment of the evaluation index of the preset Steinmetz equation and the target Steinmetz equation at 25°C provided by the present invention; Figure 7 A schematic diagram of an embodiment of the evaluation index of the preset Steinmetz equation and the target Steinmetz equation at 50° C. provided by the present invention; Figure 8 A schematic diagram of an embodiment of the evaluation index of the preset Steinmetz equation and the target Steinmetz equation at 70°C provided by the present invention; Figure 9 A schematic diagram of an embodiment of the evaluation index of the preset Steinmetz equation and the target Steinmetz equation at 90°C provided by the present invention; Figure 10 A coordinate diagram of an embodiment of a relative error comparison analysis diagram of two equations provided by the present invention; Figure 11 A coordinate diagram of an embodiment of a graph of the relationship between the objective function value and the number of iterations provided by the present invention; Figure 12 A coordinate diagram of an embodiment of a core loss variation diagram provided by the present invention; Figure 13 A coordinate diagram of an embodiment of a transmission magnetic energy variation diagram provided by the present invention; Figure 14 This is a schematic structural diagram of an embodiment of a device for predicting core loss of a magnetic element provided by the present invention. DETAILED DESCRIPTION

[0018] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, and are not used to limit the scope of the present invention.

[0019] The excitation waveform refers to the time-varying waveform of the voltage or current used to generate the magnetic field in electromagnetic devices (such as transformers, inductors, and motors). This waveform significantly impacts device performance and losses. Common excitation waveforms include sine, square, and triangular waves.

[0020] The Steinmetz equation (SE) is a classic empirical formula for describing core losses in magnetic materials, such as transformer and inductor cores. Core losses primarily consist of hysteresis losses, eddy current losses, and residual losses. The equation relates these losses to operating frequency and magnetic flux density through a simple mathematical relationship.

[0021] like Figure 1 As shown, a specific embodiment of the present invention discloses a method for predicting core loss of a magnetic component, comprising: S101. Acquire a historical data set of a magnetic component; the historical data set includes magnetic flux density data and frequency of different magnetic core materials.

[0022] The core loss prediction method of a magnetic component provided in an embodiment of the present application can be applied to a core loss prediction system of a magnetic component, wherein the core loss prediction of the magnetic component can be based on a software system running on a terminal device, and the terminal device can be a server, a tablet computer, an augmented reality (AR) / virtual reality (VR) device, a laptop computer, an ultra-mobile personal computer (UMPC), a netbook, a personal digital assistant (PDA), a mobile phone and other terminal devices. The embodiment of the present application does not impose any restrictions on the specific type of the terminal device.

[0023] The historical data set of the magnetic component can be test data provided by the manufacturer, or it can be historical data generated in actual processing, or it can be data detected by experimental means. The specific setting can be based on actual conditions, and the embodiment of the present invention is not limited here. The historical data set can include magnetic flux density data, core loss and frequency of different magnetic core materials. For example, the different magnetic core materials can be material 1, material 2, material 3 and material 4, and the magnetic flux density data, core loss and frequency of each magnetic core material can be different. S102 , performing temperature correction on the preset Steinmetz equation according to the historical data set to obtain a target Steinmetz equation, and performing loss prediction on the magnetic core material in the historical data set using the target Steinmetz equation to obtain an updated data set.

[0024] The preset Steinmetz equation can be the Steinmetz equation (also known as the Steinmetz equation). With technological advancements and increasing application demands, the Steinmetz equation faces increasing challenges and limitations in estimating core losses under a wide variety of core materials and complex, changing operating conditions. First, the Steinmetz equation is only applicable to loss calculations under sinusoidal excitation, affecting accuracy for loss estimation under non-sinusoidal excitation. Second, the Steinmetz equation does not consider the effects of DC bias on core losses, leading to significant errors in certain situations. Third, temperature also affects core losses, and the coefficients in the Steinmetz equation vary with temperature, but the Steinmetz equation does not account for this temperature factor. To account for this temperature effect, the preset Steinmetz equation can be modified by temperature distribution in the historical dataset to obtain a target Steinmetz equation. The target Steinmetz equation can then be used to recalculate and update the core materials in the historical dataset to obtain an updated dataset.

[0025] S103 , merging the updated data set according to the magnetic core material, magnetic flux density data, and frequency to obtain a merged data set.

[0026] In order to make the data set include more factors, the updated data set can be uniquely encoded according to the core material, magnetic flux density data and frequency, and then merged to obtain a merged data set.

[0027] S104 . Training a preset prediction model according to the combined data set to obtain a target prediction model, and predicting the core loss of the magnetic component according to the target prediction model.

[0028] The preset prediction model may be an AdaBoost regression model, so that the AdaBoost regression model may be trained by merging data sets to obtain a target prediction model, and the core loss of the magnetic component may be predicted based on the target prediction model.

[0029] Compared with the prior art, the present embodiment provides a historical data set for obtaining different core materials, so that the preset Steinmetz equation can be corrected according to the influence of temperature on core loss in the historical data set to obtain a target Steinmetz equation, so that the loss of the core material in the historical data set can be predicted by the target Steinmetz equation, and an updated data set with the predicted value of the temperature factor is obtained; further, the magnetic flux density data and frequency also have an impact on the core loss, so the updated data set can be uniquely encoded and merged according to the core material, magnetic flux density data and frequency to obtain a merged data set, and then the preset prediction model can be trained by the merged data set, and the training of the preset prediction model on flux density data and frequency is added, so that the target prediction model's predicted core loss of the magnetic component is more in line with the actual situation.

[0030] In some embodiments of the present invention, the process of determining the preset prediction model includes: According to the combined data set, a linear regression model, a gradient boosting regression model, and an AdaBoost regression model are trained and predicted respectively to obtain corresponding core loss prediction values and residual release data; According to the core loss prediction value and the residual release data, the AdaBoost regression model is determined to be a preset prediction model.

[0031] Three regression models can be selected for training and prediction: linear regression, gradient boosting, and AdaBoost. The gradient boosting model is used for training. A gradient boosting regressor object named model1 is created using GradientBoostingRegressor. The number of weak learners (i.e., decision trees) is set to 20. The weight reduction coefficient for each weak learner is set to 0.2 to control the contribution of the new learner at each iteration. To prevent overfitting, the maximum depth of the decision tree in this embodiment is set to 10. The seed value of the random number generator is set to 42 to ensure consistent results each time the code is run. Finally, training is performed using the fit method. The second model uses LinearRegression to create a linear regressor object named model2, which is then trained using the fit method. The third model defines a decision tree regressor with a maximum depth of 48 as a base learner. This base learner is used to create an AdaBoost regressor containing 50 weak learners (i.e., decision trees). The random seed is set to 42 to ensure reproducibility of the results. The merged data set can then be used to predict the three models separately to obtain the prediction results. The evaluation indicators of the prediction results of the three models are then calculated separately. Each model has two indicators: R² score (coefficient of determination) and mean_absolute_percentage_error. The R² score measures the accuracy of the model prediction and has a value range of 0 to 1. The closer it is to 1, the better the model fit. mean_absolute_percentage_error is a function defined in an embodiment of the present invention, which is used to calculate the mean absolute percentage error (MAPE) between the true value and the predicted value. Using the above two evaluation indicators, the embodiment of the present invention calculates three sets of indicators, corresponding to the prediction results of the three models, as shown in Table 1: Table 1. Performance indicators of three regression models

[0032] In order to intuitively feel the prediction accuracy of the three selected models and to select the most suitable model for modeling in the embodiment of the present invention, the embodiment of the present invention visualizes the predicted results in Table 1. Figure 2As shown, a comparison of the predicted values and true values of the three models, as well as a residual distribution diagram (true value - predicted value), is shown in the figure on the left for the linear regression model, the figure in the middle for the gradient boosting regression model, and the figure on the right for the AdaBoost regression model. The x-axis represents the true value of the core loss, the y-axis represents the predicted value of the core loss by the model, and the dotted line represents the best fit line obtained by fitting the model. It can be observed from the figure that the predicted values of the linear regression model gradually tend to be parallel to the x-axis, indicating that many predicted values do not differ in size. The scatter distribution of the predicted values deviates from the best fit line, indicating that the prediction results of the linear regression model are very inaccurate and perform very poorly in the prediction of core loss. The predicted values of the gradient boosting regression model are evenly distributed around the best fit line, and most of the predicted values are distributed on the best fit line. The predicted values of the AdaBoost regression model are not only evenly distributed around the best fit line, but also converge to the best fit line more than the predicted values of the gradient boosting regression model. By comparing the distribution of the true and predicted values of the three models, it can be concluded that the AdaBoost regression model has a better prediction effect on core loss. The relationship between residuals and frequency can also be determined. The residuals of the linear regression model are mainly distributed between 0 and 3500, indicating that the difference between the predicted and true values is small. However, there is a significant peak, which may be an outlier, indicating that there is a large deviation in the data points in this area. The residuals of this model show a certain degree of symmetry, indicating that the model may not introduce serious bias or noise. The residual values of the gradient boosting regression model are mostly concentrated between -0.5 and 1.5, and the residual values exceeding 1.5 are rare, indicating that the model performs well and the prediction accuracy is high. For the AdaBoost regression model, the residuals are mostly concentrated at 0, indicating high accuracy and the predicted results are close to the true values. At the same time, the residuals of the AdaBoost regression model have small fluctuations, indicating that the model has good stability and generalization ability.

[0033] In order to demonstrate the generalization ability of the model, the embodiment of the present invention also performed k-fold cross-validation on the AdaBoost regression model. In the k-fold cross-validation experiment conducted in the embodiment of the present invention, the merged dataset was used as the dataset and target variable for the k-fold cross-validation. The k value is taken as 5. That is to say, the original dataset is randomly divided into 5 subsets (also called "folds"), one of which is used as the test set each time, and the remaining 4 subsets are merged into the training set. This process is repeated 5 times, and each subset is used as a test set in turn. Finally, the average performance indicators of the 5 iterations, such as accuracy, recall rate, etc. (the embodiment of the present invention uses accuracy and standard deviation), are calculated as the final performance evaluation of the model, as shown in Table 2: Table 2. Accuracy and standard deviation of three regression models

[0034] Table 2 shows the accuracy and standard deviation of the k-fold cross validation experiment of different models on the same data set. It is obvious that the accuracy of the AdaBoost regression model and the gradient boosting regression model in the experiment reached more than 90%, and the standard deviation was 0.06 and 0.07. This shows that the two models can maintain good accuracy under relatively stable conditions. It further shows that the AdaBoost regression model has good generalization ability. The embodiment of the present invention also performs a visualization of the K-fold cross validation experiment, as shown in the figure. Figure 3 As shown, Figure 3 This is the K-fold cross validation accuracy curve. In K-fold cross validation, the dataset is divided into K non-overlapping subsets, each of which is called a "Fold". The "Fold" on the horizontal axis represents the fold number in each iteration, from 1 to K. For example, if K=5, then the "Fold" on the horizontal axis may be displayed as 1, 2, 3, 4, and 5, corresponding to the first subset selected as the test set in the first iteration, the second subset selected as the test set in the second iteration, and so on. The vertical axis represents the accuracy. Figure 3 It can be seen that the average accuracy of the AdaBoost regression model in the K-fold cross-validation experiment reached more than 93%, indicating that the model has good generalization ability and accuracy.

[0035] The AdaBoost regression model achieved a high accuracy rate in this prediction and can significantly improve the overall prediction accuracy. It can capture complex patterns and nonlinear relationships in the data, thereby better fitting real-world data. Although the AdaBoost regression model is a black box model, the decision-making process of the model can be understood by looking at the weights and contributions of each weak learner. This helps to understand how the model makes predictions and provides a basis for business decisions. At the same time, the embodiment of the present invention uses K-fold cross-validation to prove the generalization ability of the model, which can help the industry select a better-performing and more stable model. It also proves that the AdaBoost regression model can be applied to various regression tasks, and its flexibility enables it to adapt to different data types and problem scenarios. It is then determined that the preset prediction model is the AdaBoost regression model, so that the AdaBoost regression model can be trained again by merging the data sets, and the target prediction model can be obtained, and the core loss of the magnetic component can be predicted based on the target prediction model.

[0036] In some embodiments of the present invention, the historical data set also includes an excitation waveform; before step S102, as Figure 4 As shown, it also includes: S401 , preprocessing sample data of different magnetic core materials in a historical data set to obtain a preprocessed training set.

[0037] S402 : performing feature extraction on the magnetic flux density data and the excitation waveform in the pre-processed training set to obtain feature data of each sample data.

[0038] S403: Classify and update the pre-processed training set according to the feature data to obtain an updated historical data set.

[0039] In a specific embodiment of the present invention, sample data from different magnetic core materials in a historical dataset can be preprocessed to generate a preprocessed training set. The specific preprocessing process can be customized based on actual conditions. The historical dataset can also include excitation waveforms, with excitation waveforms categorized as sinusoidal, triangular, and trapezoidal. By analyzing the temporal changes in magnetic flux density data and extracting key features, characteristic data for each sample data can be obtained. The preprocessed training set can then be classified and updated based on the characteristic data, resulting in an updated historical dataset.

[0040] In order to improve data quality, in some embodiments of the present invention, the excitation waveform includes a sine wave, a cosine wave, and a trapezoidal wave; step S401 includes: The missing values of sample data in the historical data set are processed based on the interpolation method to obtain a complete historical data set; The types of excitation waveforms in the complete historical data set are digitally converted according to the sine wave, cosine wave and trapezoidal wave of each magnetic core material to obtain a preprocessed training set.

[0041] In a specific embodiment of the present invention, the find function can be used to check whether there are missing values in the historical data set. The processing of missing values generally adopts interpolation or direct deletion to obtain a complete historical data set, thereby ensuring the integrity of the data and the validity and feasibility of the prediction results. Among them, the excitation waveform is a text data that cannot be directly recognized by the computer. The excitation waveform type needs to be converted into a digital label, that is, the sine wave, cosine wave and trapezoidal wave of each magnetic core material are digitally converted to the type of excitation waveform in the complete historical data set. The excitation waveform type in the complete historical data set is converted into a digital label to facilitate the use of the subsequent classification model. Among them, 1 can represent a sine wave, 2 can represent a triangle wave, and 3 can represent a trapezoidal wave for conversion, thereby obtaining a pre-processed training set that has completed the conversion.

[0042] Furthermore, basic statistical information about the waveform data can be calculated, including maximum, minimum, mean, and standard deviation; the number of peaks and valleys in the waveform data; and the rise and fall times of the waveform. For example, assuming the time-varying relationship of magnetic flux density is described as B(t), the key features extracted include: maximum value (max(Bi), minimum value (min(Bi), mean(Bi), standard deviation; and peak position, which indicates the sampling points where the magnetic flux density reaches its maximum and minimum values. These statistical features can help convert complex magnetic flux density data into finite feature vectors, facilitating subsequent classification and modeling.

[0043] In some embodiments of the present invention, step S403 includes: The preset classification model is trained based on the sample data and feature data in the preprocessing training set to obtain the target classification model.

[0044] The preset classification model can be a random forest classification model. Random forest is an ensemble learning machine learning algorithm that performs classification or regression tasks by building multiple decision trees and combining prediction results. During model training, each decision tree randomly extracts samples and features from the dataset for training. Bagging is used to train multiple base learners in parallel to reduce variance and improve model stability and accuracy.

[0045] The random forest classification model uses bootstrap sampling to extract multiple sub-datasets from the original dataset. These datasets include repeated samples to ensure consistency with the size of the original dataset. At each node, a feature subset is computationally selected from which the optimal partitioning feature is selected. The number of features is usually selected to be the square root or logarithm of the total number of features. For each decision tree, the random forest does not perform pruning, allowing it to grow to the maximum extent to ensure the diversity of the model. The advantages and disadvantages of random forest are also obvious. It has high accuracy, good stability, strong noise resistance, and is suitable for high-dimensional data and can be easily parallelized. However, its parameters are relatively complex and need to be constantly tuned, and the training and prediction speeds are relatively slow.

[0046] To evaluate the performance of the random forest classification model, a pre-set classification model was trained based on the sample data and feature data from the preprocessed training set. A classification report was output, including data such as accuracy, precision, recall, f1-score, and support. A confusion matrix was also output to demonstrate the model's classification performance on the validation set. Analysis of the confusion matrix and classification report results showed that the accuracy of the random forest classification model used in this embodiment of the present invention reached 1.0 for Materials 1, 2, and 4, and was the lowest in Material 3, but still reached 0.99. The accuracy, recall, and F-1 score of the training results were all close to 1, indicating that the model performed well overall, resulting in the target classification model.

[0047] The preprocessed training set is classified and predicted according to the target classification model to obtain the corresponding classification results.

[0048] Among them, the classification model is used to predict the excitation waveform of the sample data and feature data in the preprocessing training set, which can achieve good expected results, so that the number of predicted excitation waveforms in each material can be output, that is, the classification result, as shown in Table 3: Table 3. Statistical number of predicted waveforms for each material

[0049] The preprocessed training set is updated according to the classification results to obtain an updated historical data set.

[0050] After the classification result is obtained, the classification result can be input into the preprocessing training set, and the sample data in the preprocessing training set can be summarized and updated, so as to obtain an updated historical data set.

[0051] In some embodiments of the present invention, step S102 includes: Perform coefficient fitting on the preset Steinmetz equation according to the historical data set to obtain multiple fitting coefficients; The preset Steinmetz equation is modified according to a plurality of fitting coefficients and a preset temperature function to obtain a target Steinmetz equation.

[0052] In a specific embodiment of the present invention, the coefficients of the historical data set can be fitted according to the Steinmetz equation to obtain multiple fitting coefficients, which are: f 、 B 、 a 、 b Based on the Steinmetz equation, considering the influence of temperature, the coefficient k Set as a function of temperature k ( T), to construct the correction equation, the temperature function of the embodiment of the present invention It is expressed as shown in formula (1): (1) The complete modified target Steinmetz equation formula is shown in formula (2): (2) in, represents the coefficient at standard temperature, c represents the temperature coefficient, T Indicates temperature (℃), P 、 f 、 B 、 a 、 b It has the same meaning as the Steinmetz equation and will not be repeated here.

[0053] Furthermore, in order to make the target Steinmetz equation applicable to the prediction of core loss under different temperature changes, the embodiment of the present invention first verifies that "the coefficients of the Steinmetz equation will change with temperature changes." In Material 1 of the historical data set, the four columns of information, namely frequency, peak value of magnetic flux density, waveform type, and core loss, are screened and retained. Only all data with a waveform type of sine wave are taken to ensure that the data used is applicable to the Steinmetz equation. These data are used as input to fit the Steinmetz equation to obtain the coefficients. a 、 b 、 k The fitting results of the embodiment of the present invention are shown in Table 4: Table 4. Fitting coefficients of the Steinmetz equation

[0054] Add temperature analysis, group the data by temperature, and divide the temperature groups into four categories: 25℃, 50℃, 70℃, and 90℃. Input the data with the same temperature value into the Steinmetz equation for fitting, and observe the influence of different temperatures. k The fitting results of the embodiment of the present invention are shown in Table 5: Table 5. Steinmetz equation at different temperatures k Fitted values

[0055] Table 5 shows that the coefficients of the Steinmetz equation do change with temperature. Therefore, even with the same core material, sine wave, and input data, temperature variations can lead to errors in the Steinmetz equation's calculation of core loss, further demonstrating the necessity of including a temperature parameter in the Steinmetz equation.

[0056] In order to ensure that the modified target Steinmetz equation and the Steinmetz equation only differ in temperature, the embodiment of the present invention uses the target Steinmetz equation to predict the core loss. a and b The values of remain the same as in Section 6.3.1, which are 1.4296 and 2.4713 respectively.

[0057] In order to explore the effect of temperature T on core loss P The impact of Function, taking the logarithm of this function, we get formula (3): (3) The temperature-corrected and c the value of the parameter, , c =-0.0116. Then we plotted the relationship curve between different temperatures T_i and ln(k_i), as shown in the figure below: Figure 5 As shown, Figure 5 For different temperatures T_i and ln ( k_i ) relationship curve, given by Figure 5 It can be seen that T_i and ln ( k_i ) shows a negative correlation. That is, the higher the temperature T, The smaller the core loss P The smaller it is, the further it is verified that temperature has an impact on the core loss.

[0058] In order to further confirm the accuracy of the prediction of the modified target Steinmetz equation, the prediction error analysis of the preset Steinmetz equation and the target Steinmetz equation can be performed. Common error analysis methods include systematic error, random error, absolute error, and relative error. The embodiment of the present invention chooses to analyze the relative error. The relative error calculation is shown in formula (4): (5) in, is the absolute error, that is, the difference between the measured value and the true value; Represents the true value.

[0059] The preset Steinmetz equation and the modified target Steinmetz equation are used to predict the core loss of the data in Material 1 in the updated data set, and the relative errors are calculated. Finally, the prediction result samples are output. Due to the large number of data items, Figure 6-Figure 9 Only 5 data points at temperatures of 25℃, 50℃, 75℃ and 90℃ are displayed. Figure 6is the evaluation index of the preset Steinmetz equation and the target Steinmetz equation at 25°C. Figure 7 is the evaluation index of the preset Steinmetz equation and the target Steinmetz equation at 50°C. Figure 8 is the evaluation index of the preset Steinmetz equation and the target Steinmetz equation at 70°C. Figure 9 It is the evaluation index of the preset Steinmetz equation and the target Steinmetz equation at 90℃. In the figure, P_pred_SE represents the core loss predicted by the preset Steinmetz equation, P_pred_Mod represents the core loss predicted by the modified target Steinmetz equation, Error_SE represents the relative error of the preset Steinmetz equation, and Error_Mod represents the relative error of the modified target Steinmetz equation. Figure 6-Figure 9 A horizontal comparison was made between the actual value of the core loss and the preset SE predicted value and the target SE predicted value. It was found that the target SE predicted value was closer to the actual loss value. From the perspective of relative error, the relative error of the target Steinmetz equation was smaller than the relative error of the preset Steinmetz equation. Both indicate that the target Steinmetz equation with temperature correction has better core loss prediction effect.

[0060] Will Figure 6-Figure 9 A longitudinal comparison focuses on the relative errors of the two equations' predictions at different temperatures. At 25°C, Error_Mod is approximately 20% smaller than Error_SE; at 50°C, Error_Mod is approximately 1% smaller than Error_SE; at 70°C, Error_Mod is approximately 30% smaller than Error_SE; and at 90°C, Error_Mod is approximately 70% smaller than Error_SE. Furthermore, the relative error of the default Steinmetz equation varies significantly with temperature, even exceeding 100% at 90°C. The relative error gap of the modified target Steinmetz equation is significantly smaller, demonstrating its applicability to core loss predictions under varying temperatures.

[0061] In order to more intuitively show the comparison between the predicted values and the true values of the two equations, box plots of the preset Steinmetz equation and the target Steinmetz equation are drawn to show the relative error comparison analysis of the two equations, such as Figure 10 As shown in the figure, different colors are used to distinguish the relative errors between the preset Steinmetz equation and the target Steinmetz equation. The horizontal axis represents four different temperatures, each temperature value corresponds to a data point, and the vertical axis represents the percentage of relative error. Figure 10The data points in the figure are divided into four sections. The boxes represent the interquartile range; the upper and lower edges represent the maximum and minimum relative errors, respectively; and the median represents the value at the middle position. As can be seen from the figure, the target SE error is small at 25°C; the SE error is large at 90°C; as the temperature increases, the target SE error gradually decreases, while the preset SE error gradually increases; overall, the target SE error is smaller than the preset SE error at all temperatures. This indicates that the modified target Steinmetz equation performs better in terms of prediction error.

[0062] In some embodiments of the present invention, step S102 includes: When predicting the core loss based on the target Steinmetz equation, the logarithm of both sides of the target Steinmetz equation is taken to obtain the linear target Steinmetz equation; The loss prediction of the core material in the historical data set is performed based on the linear target Steinmetz equation to obtain a new loss prediction; The new loss prediction is added to the historical dataset to obtain an updated dataset.

[0063] In a specific embodiment of the present invention, when predicting core loss, the Steinmetz equation is generally used to describe the relationship between loss and parameters such as frequency and magnetic flux density. The Steinmetz equation is generally nonlinear and involves complex exponential relationships, which makes the calculation and fitting process time-consuming and complicated. In order to simplify the calculation, the logarithm of both sides of the target Steinmetz equation can be taken to convert the nonlinear equation into a linear form. Linear relationships are easier to handle, reducing calculation time and complexity. The linearized equation is more convenient to use fitting methods such as linear regression, so that the loss of the core material in the historical data set can be predicted based on the linear target Steinmetz equation to obtain a new loss prediction for temperature. The new loss prediction can then be added to the historical data set to obtain an updated data set, thereby improving the efficiency and accuracy of data fitting.

[0064] In some embodiments of the present invention, the historical data set further includes temperature; step S103 includes: Perform one-hot encoding on the excitation waveform, temperature, and core material in the updated dataset and reset the index to obtain a one-hot encoded column. According to the magnetic flux density data in the updated data set, the corresponding magnetic flux density peak value is determined; Merge the one-hot encoded columns, magnetic flux density peaks, and frequencies in the column direction to obtain a merged dataset.

[0065] In a specific embodiment of the present invention, the three columns "Core Material," "Temperature, °C," and "Excitation Waveform" in the updated dataset are one-hot encoded and reindexed to ensure that the newly generated one-hot encoded column has a continuous integer index. This generates a one-hot encoded column. The column of magnetic flux density data in the updated dataset is extracted and its maximum value is calculated to obtain the peak magnetic flux density value and form a new column. The one-hot encoded column is then merged with the "Frequency" and "Peak Magnetic Flux Density" columns in the updated dataset along the column direction, ultimately generating the merged dataset X_predict.

[0066] In some embodiments of the present invention, step S104 includes: In the process of training the preset prediction model, the sample data in the combined data set is predicted by the genetic algorithm to obtain the corresponding target core loss; The peak value and frequency of the magnetic flux density in the combined data set are calculated to obtain the transmitted magnetic energy; According to the target core loss and transmitted magnetic energy, the objective function value is obtained; When the number of iterations of the genetic algorithm and / or the objective function value meet the preset conditions, the target prediction model is obtained.

[0067] In a specific embodiment of the present invention, a decision tree regressor with a maximum depth of 48 is defined as a base learner. This base learner is used to create an AdaBoost regressor, which contains 50 weak learners (i.e., decision trees), and the random seed is set to 42 to ensure the repeatability of the results. The merged data set is used to fit the model, and the column "frequency" in the merged data set is multiplied by the column "magnetic flux density peak" in the merged data set to obtain the transmitted magnetic energy and assign it to the "transmitted magnetic energy", and finally the optimized merged data set is obtained. The genetic (GA) algorithm is a search optimization method that simulates natural selection and genetic mechanisms and is widely used to solve complex mathematical and engineering problems. The GA algorithm can be applied to the optimization of core loss and transmitted magnetic energy. The purpose of the GA algorithm is to find the optimal solution to minimize core loss and maximize transmitted magnetic energy. The specific application process is: Initialize the population: First, define an individual structure, each of which contains five attributes: temperature index, material index, waveform index, frequency, and magnetic flux density. Then use toolbox.population(n=100) to create an initial population of 100 randomly generated individuals.

[0068] Evaluate Fitness: For each individual in the population, use the evaluate function to calculate its fitness. This function converts the individual's attributes into actual input data and then uses a pre-trained model to predict nuclear loss. The objective function is the ratio of nuclear loss to transferred energy, and this embodiment of the present invention aims to minimize this ratio.

[0069] Selection: Use the toolbox.select method to select a subset of individuals from the current population as parents for the next generation. Tournament selection is used here, where several individuals are randomly selected from the population for comparison and the individuals with the highest fitness are selected.

[0070] Crossover: To generate new offspring, a custom crossover operation, custom_crossover, is used. In this operation, integer variables (temperature, material, and waveform index) are swapped, and continuous variables (frequency and flux density) are subjected to mixed crossover. Mixed crossover is a method between two-point crossover and uniform crossover, combining the advantages of both.

[0071] Mutation Operation: The custom_mutate operation is used to introduce new features or change the value of existing features. The mutation probability is controlled by the indpb parameter. In this example, the embodiment of the present invention mutates temperature, material, waveform index, frequency, and magnetic flux density.

[0072] Iterative evolution: The above steps are repeated multiple times, and each iteration produces a new population. At the end of each generation, the fitness value of the best individual is recorded and used for plotting.

[0073] Output results: Finally, the best individuals after multiple generations of evolution are extracted. These individuals represent the best solutions found under given constraints.

[0074] GA parameter settings. Population size: The population size refers to the number of individuals in each generation. A larger population can increase exploration capabilities, but also increases computational complexity. It is generally set between 50 and 200.

[0075] Crossover probability: The crossover probability determines the frequency of individual gene information exchange and directly affects the algorithm's search ability. It is usually set between 0.6 and 1.0.

[0076] Mutation probability: Mutation probability affects the diversity of individual genetic information. An appropriate mutation rate helps avoid falling into a local optimal solution. It is generally between 0.01 and 0.1.

[0077] To achieve minimum core loss and maximum transmitted magnetic energy under certain conditions (temperature, frequency, waveform, peak magnetic flux density, and core material), it is necessary to optimize the core loss and transmitted magnetic energy. A genetic algorithm is used for this optimization.

[0078] 1. Define constants: temperature (can only be 25, 50, 70, 90), material (can only be material 1, material 2, material 3, material 4), excitation waveform (can only be triangle wave, trapezoidal wave, sine wave), frequency range and flux density range.

[0079] 2. Set up the basic elements of the genetic algorithm: Create a fitness class called FitnessMin to minimize core loss. Then create an individual class called Individual with a fitness attribute. Finally, create a toolbox object to store the various operations of the genetic algorithm.

[0080] 3. Define property generators: These generators are used to generate individual property values, including temperature, material, waveform, and magnetic flux density.

[0081] 4. Define the structure initializers: This defines how to initialize individuals and populations. The initCycle method loops through the given attribute generator to create an individual, while the initRepeat method repeats this process to create a population.

[0082] 5. Define the evaluation function: This function takes an individual as input and returns a tuple containing the objective function value. It first converts the individual's attributes into integer indices, then prepares the input data and passes it to the pre-trained model for prediction. Finally, it calculates the ratio of core loss to transferred magnetic energy as the objective function value.

[0083] 6. Register custom crossover and mutation operations: Crossover and mutation operations will be called during the genetic algorithm process.

[0084] 7. Run the genetic algorithm: Set the number of iterations to 50 and the initial population size to 100 as the initial values to start the genetic algorithm, clone the population, apply crossover and mutation, evaluate the fitness of individuals with invalid fitness, select the next generation population, and record the best individual.

[0085] Use visualization to clearly show the change of fitness as the number of iterations increases. Draw an iteration graph with the number of iterations as the horizontal axis and the fitness as the vertical axis, such as Figure 11 As shown, Figure 11 It is a graph showing the relationship between the objective function value and the number of iterations. It can be seen that the best fitness, that is, the optimal value of the objective function, can be found within 5 iterations. The optimal values are as follows: optimal material: Material 2 in the merged data set, optimal temperature: 78.00 degrees Celsius, optimal frequency: 2524689.99Hz, optimal waveform type: trapezoidal wave, optimal magnetic flux density peak: 0.0243T, optimal transmission magnetic energy: 1849.9077, minimum core loss: 1758.1820W / m3. In addition, the embodiment of the present invention also pays attention to the changes in core loss and transmission magnetic energy. Figure 12 and Figure 13 As shown, Figure 12 is the core loss variation diagram, Figure 13To transmit the magnetic energy change diagram, Figure 12 It can be seen that the core loss has many peaks and many valleys, and reaches higher peaks when the sample numbers are 2500, 5000, and 9000. It can be seen from the merged data set that at the peak of 2500, the core material is material one, the temperature is 25 degrees Celsius, and the excitation waveform is a trapezoidal wave. At the peak of 5000, the core material is material one, the temperature is 25 degrees Celsius, and the excitation waveform is a trapezoidal wave. At the peak of 5000, the core material is material two, the temperature is 70 degrees Celsius, and the excitation waveform is a triangular wave. At the peak of 9000, the core material is material 3, the temperature is 50 degrees Celsius, and the excitation waveform is a trapezoidal wave. When the sample number is 10000, the core loss reaches a valley, the core material is material 4, the temperature is 50 degrees Celsius, and the excitation waveform is a sine wave. Figure 12 , we can clearly see the approximate values of information loss under different working conditions, and from the overall perspective, the core loss fluctuates greatly, with the maximum difference between the peak and valley values being 3.5e6W / m3. Figure 13 The values of the transmitted magnetic energy (the product of frequency and peak magnetic flux density) for different samples are shown. Similar to the core loss variation graph, the transmitted magnetic energy still fluctuates significantly, with a maximum value of approximately 78,000 and a minimum value close to 0. In actual operating conditions, core loss and transmitted magnetic energy fluctuate significantly under different conditions, demonstrating that achieving both minimum core loss and maximum transmitted magnetic energy during the optimization process is a very challenging task. The trained and optimized target prediction model has a low average relative error and high accuracy, enabling the prediction of core loss in magnetic components.

[0086] The embodiment of the present invention fully considers the distribution characteristics and shape characteristics of the magnetic flux density through the random forest model; the modified target Steinmetz equation considers the coefficient k as a function of temperature change, which can be adapted to the prediction of core loss at different temperatures, and the prediction result is better than the preset Steinmetz equation. The independent and synergistic effects of the three influencing factors on the core loss are fully considered, and the constructed multivariate regression model can accurately give the combination of the three factors to minimize the core loss. Multiple models are used to train the training set separately, and by comparing the evaluation indicators of different models, it is found that the average relative error of the AdaBoostRegressor model is small and the accuracy is high. The constructed model can minimize the core loss value and maximize the magnetic energy transmission with a very small number of iterations.

[0087] In order to better implement the core loss prediction method of the magnetic element in the embodiment of the present invention, based on the core loss prediction method of the magnetic element, the embodiment of the present invention also provides a core loss prediction device for the magnetic element, such as Figure 14 As shown, the core loss prediction device 1400 of a magnetic component includes: The data acquisition module 1401 is used to acquire a historical data set of a magnetic element; the historical data set includes magnetic flux density data and frequency of different magnetic core materials; An equation correction module 1402 is configured to perform temperature correction on a preset Steinmetz equation based on a historical data set to obtain a target Steinmetz equation, and to use the target Steinmetz equation to predict the loss of the magnetic core material in the historical data set to obtain an updated data set; A data set determination module 1403 is configured to merge the updated data sets according to the magnetic core material, magnetic flux density data, and frequency to obtain a merged data set; The model training module 1404 is used to train a preset prediction model according to the combined data set to obtain a target prediction model, and predict the core loss of the magnetic component according to the target prediction model.

[0088] The core loss prediction device 1400 for a magnetic element provided in the above embodiment can implement the technical solution described in the above embodiment of the core loss prediction method for a magnetic element. The specific implementation principles of the above modules or units can be found in the corresponding contents in the above embodiment of the core loss prediction method for a magnetic element, which will not be repeated here.

[0089] The above is a detailed introduction to the core loss prediction method and device for the magnetic element provided by the present invention. Specific examples are used in the embodiments of the present invention to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas. At the same time, for those skilled in the art, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for predicting core loss of a magnetic component, characterized in that: include: Acquire a historical data set of a magnetic component; the historical data set includes magnetic flux density data and frequency of different magnetic core materials; Performing temperature correction on a preset Steinmetz equation according to the historical data set to obtain a target Steinmetz equation, and performing loss prediction on the magnetic core material in the historical data set using the target Steinmetz equation to obtain an updated data set; Merging the updated data set according to the magnetic core material, the magnetic flux density data, and the frequency to obtain a merged data set; A preset prediction model is trained according to the combined data set to obtain a target prediction model, and core loss prediction is performed on the magnetic component according to the target prediction model.

2. The method for predicting core loss of a magnetic component according to claim 1, wherein: The historical data set also includes an excitation waveform; and before correcting the preset Steinmetz equation according to the historical data set to obtain the target Steinmetz equation, the method further includes: Preprocessing the sample data of the different magnetic core materials in the historical data set to obtain a preprocessed training set; Performing feature extraction on the magnetic flux density data and the excitation waveform in the preprocessing training set to obtain feature data of each sample data; The preprocessed training set is classified and updated according to the feature data to obtain an updated historical data set.

3. The method for predicting core loss of a magnetic component according to claim 2, wherein: The classifying and updating the preprocessed training set according to the feature data to obtain an updated historical data set includes: Training a preset classification model based on the sample data in the preprocessing training set and the feature data to obtain a target classification model; Perform classification prediction on the preprocessed training set according to the target classification model to obtain corresponding classification results; The preprocessing training set is updated according to the classification result to obtain an updated historical data set.

4. The method for predicting core loss of a magnetic component according to claim 1, wherein: The step of performing temperature correction on the preset Steinmetz equation according to the historical data set to obtain the target Steinmetz equation includes: Performing coefficient fitting on the preset Steinmetz equation according to the historical data set to obtain a plurality of fitting coefficients; The preset Steinmetz equation is temperature-corrected according to the multiple fitting coefficients and a preset temperature function to obtain a target Steinmetz equation.

5. The method for predicting core loss of a magnetic component according to claim 2, wherein: The historical data set further includes temperature; and the updated data set is merged according to the magnetic core material, the magnetic flux density data, and the frequency to obtain a merged data set, including: Performing one-hot encoding on the excitation waveform, the temperature, and the magnetic core material in the updated data set and reindexing to obtain a one-hot encoded column; Determining a corresponding magnetic flux density peak value according to the magnetic flux density data in the updated data set; The one-hot encoded columns, the magnetic flux density peaks, and the frequencies are merged in a column direction to obtain a merged data set.

6. The method for predicting core loss of a magnetic component according to claim 2, wherein: The excitation waveform includes a sine wave, a cosine wave, and a trapezoidal wave; and the preprocessing of the sample data of the different magnetic core materials in the historical data set to obtain a preprocessed training set includes: Processing missing values of sample data in the historical data set based on an interpolation method to obtain a complete historical data set; The types of excitation waveforms in the complete historical data set are digitally converted according to the sine wave, the cosine wave, and the trapezoidal wave of each magnetic core material to obtain a preprocessing training set.

7. The method for predicting core loss of a magnetic component according to claim 1, wherein: The step of performing loss prediction on the magnetic core material in the historical data set by using the target Steinmetz equation to obtain an updated data set includes: When performing core loss prediction according to the target Steinmetz equation, taking logarithms of both sides of the target Steinmetz equation to obtain a linear target Steinmetz equation; Performing a loss prediction on the primary magnetic core material in the historical data set according to the linear target Steinmetz equation to obtain a new loss prediction; The new loss prediction is added to the historical data set to obtain an updated data set.

8. The method for predicting core loss of a magnetic component according to claim 5, wherein: The process of determining the preset prediction model includes: According to the combined data set, a linear regression model, a gradient boosting regression model, and an AdaBoost regression model are trained and predicted respectively to obtain corresponding core loss prediction values and residual release data; According to the core loss prediction value and the residual release data, the AdaBoost regression model is determined to be a preset prediction model.

9. The method for predicting core loss of a magnetic component according to claim 5, wherein: The step of training a preset prediction model according to the combined data set to obtain a target prediction model includes: In the process of training the preset prediction model, the sample data in the combined data set is predicted by a genetic algorithm to obtain the corresponding target core loss; Calculating the peak value of the magnetic flux density and the frequency in the combined data set to obtain the transmitted magnetic energy; Obtaining an objective function value according to the target core loss and the transmitted magnetic energy; When the number of iterations of the genetic algorithm and / or the objective function value meet preset conditions, a target prediction model is obtained.

10. A device for predicting core loss of a magnetic component, characterized in that: include: A data acquisition module is used to acquire a historical data set of the magnetic element; the historical data set includes magnetic flux density data and frequency of different magnetic core materials; an equation correction module, configured to perform temperature correction on a preset Steinmetz equation according to the historical data set to obtain a target Steinmetz equation, and perform loss prediction on the magnetic core material in the historical data set using the target Steinmetz equation to obtain an updated data set; a data set determining module, configured to merge the updated data set according to the magnetic core material, the magnetic flux density data, and the frequency to obtain a merged data set; A model training module is used to train a preset prediction model based on the combined data set to obtain a target prediction model, and predict the core loss of the magnetic component based on the target prediction model.

Images