A machine learning based multi-factor core loss prediction method

Abstract

The present application relates to the technical field of magnetic element modeling and control, and particularly relates to a multi-factor magnetic core loss prediction method based on machine learning. The method first acquires the magnetic flux density data of the magnetic element, extracts its distribution characteristics and shape characteristics, and realizes excitation waveform recognition by constructing a classification model; secondly, a temperature index correction term is introduced based on the Steinmetz equation to establish a temperature-sensitive loss model; further, the entropy weight method and statistical analysis are used to evaluate the importance of multiple factors; an integrated learning model of random forest and LightGBM is used to construct a magnetic core loss prediction model; and a single-objective optimization model is established by combining differential evolution and simulated annealing algorithm with the objectives of minimum loss and maximum magnetic energy transmission. The method improves the generalization ability and prediction accuracy of the model, can adapt to the modeling needs of magnetic core loss under multiple material, multiple waveform and multiple working condition conditions, and is suitable for application scenarios such as high-frequency magnetic element design and power electronic system optimization.

Description

A Multi-Factor Core Loss Prediction Method Based on Machine Learning Technical Field

[0001] This invention relates to the field of magnetic component modeling and control technology, and in particular to a multi-factor magnetic core loss prediction method based on machine learning. Background Technology

[0002] Magnetic components, as core parts of power converters, play a crucial role in power transmission, energy storage, and filtering. Their electrical parameters directly affect the system's size, weight, losses, and cost. With the continuous development of power electronics technology and the new energy industry, core loss modeling and optimization have become key technical paths to improve the reliability and energy efficiency of power converters. Core loss is a core indicator for the design optimization of magnetic components, closely related to equipment efficiency and stability. It is influenced by multiple factors, including operating frequency, magnetic flux density, excitation waveform, operating temperature, and core material, and these factors exhibit complex nonlinear coupling relationships.

[0003] Existing loss separation models and empirical models (such as the Steinmetz equation) suffer from insufficient modeling accuracy, limited applicability, and poor response to complex operating conditions in engineering applications. Specifically, the Steinmetz equation does not consider dynamic temperature effects; it is only applicable to sinusoidal excitation and has poor adaptability to non-standard waveforms such as triangular and trapezoidal waves; furthermore, traditional models lack a synergistic optimization mechanism between core loss and transmitted magnetic energy, making it difficult to meet the design requirements of modern high-performance magnetic components. Although some research has attempted to introduce methods such as multilayer perceptrons (MLP), multi-objective evolutionary algorithms (MOEAD), and long short-term memory networks (LSTM) for modeling and optimization, these methods still suffer from limited model generalization ability, insufficient multi-objective optimization capabilities, and high computational complexity. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention proposes a multi-factor core loss prediction method based on machine learning. By identifying excitation waveforms, modeling temperature correction, and analyzing the weights of influencing factors, a core loss prediction model integrating multi-factor feature inputs is constructed, improving prediction accuracy and model adaptability under complex operating conditions, and solving the problems mentioned in the background technology.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the present invention specifically adopts the following technical solution:

[0008] A multi-factor magnetic core loss prediction method based on machine learning includes the following steps:

[0009] Step 1, Excitation Waveform Feature Extraction and Classification: Obtain magnetic flux density data, extract distribution and shape features, filter out effective feature variables, and construct a classification model to identify the waveform;

[0010] Step 2, Temperature Modeling Correction: Based on the traditional Steinmetz equation, a temperature sensitivity index is introduced to correct the core loss prediction model, and the model is validated.

[0011] Step 3, Analysis of factors affecting core loss: Analysis of variance is used to evaluate the significance and interaction of temperature, material and waveform, and the entropy weight method is used to quantify their influence and determine the dominant factors;

[0012] Step 4, Core Loss Prediction Model Construction: An ensemble learning method is used to construct a core loss model, and the effectiveness of the model is analyzed.

[0013] Step 5, Establishment of the operating condition optimization model: With the goal of minimizing core loss and maximizing transmitted magnetic energy, a single-objective optimization model is constructed by combining differential evolution and simulated annealing algorithms.

[0014] Furthermore, the acquisition of magnetic flux density data in step 1 includes different material types, temperatures, frequencies, excitation waveform types, and corresponding core losses; it also includes dividing the data into training sets and test sets according to a certain ratio.

[0015] The extracted distribution features include peak value, valley value, average value, standard deviation, peak-to-peak value, and number of zero-crossings; the shape features include at least the maximum and minimum values ​​of derivatives, harmonic amplitudes, and the number of inflection points.

[0016] The method for selecting effective feature variables is as follows: first, based on box plot analysis, features that cannot distinguish waveforms are filtered out; then, correlation analysis is performed on the remaining features to filter out duplicate features; finally, the effective feature variables selected include kurtosis, maximum derivative, average peak height, total harmonic distortion, length of flat region, and number of inflection points.

[0017] The classification model is constructed to identify waveforms. It is trained and optimized using a random forest method. The training process uses cross-validation and grid search to adjust parameters, and a confusion matrix is ​​used to evaluate the classification performance.

[0018] Furthermore, the temperature correction model established in step 2 is a negative exponential relationship. Temperature correction model:

[0019]

[0020] In the formula, P vTo correct the core loss in the model, k0 is the baseline loss coefficient, which depends on the material type; e is the base of the natural logarithm; γ is the temperature sensitivity coefficient, and the fitted result is γ = 1.26224e⁻⁰²; T is the operating temperature; f is the frequency; α refers to the frequency-dependent exponential coefficient; B m β represents the peak value of the magnetic flux density, and β is the exponential coefficient related to the magnetic flux density.

[0021] The model validation was performed using the root mean square error (RMSE) and the goodness of fit (R²). 2 Compare the model performance before and after the correction.

[0022] Furthermore, the analysis of factors affecting magnetic core loss in step 3 includes: using one-way ANOVA to evaluate the significance of temperature, material, and waveform on magnetic core loss, and using two-way ANOVA to analyze the interaction between factors;

[0023] The method of combining entropy weight to quantify the degree of influence and determine the dominant factors involves using entropy weight to calculate the weight indices of temperature, excitation waveform and core material, with weights of 0.112, 0.394 and 0.494, respectively.

[0024] Furthermore, in step 4, the core loss prediction model is constructed by considering the influence of characteristics such as temperature, frequency, peak magnetic flux density, excitation waveform and core material, and one-hot encoding is used to eliminate the dimensional influence of the characteristics.

[0025] The ensemble learning method adopted is a combination of random forest and LightGBM model. K-fold cross-validation is used to divide the dataset, and L1 regularization is added during the training process to improve the model's generalization ability.

[0026] The model effectiveness analysis uses metrics such as mean absolute error, median absolute error, root mean square error, and goodness of fit to compare the performance differences between ensemble learning models and single models.

[0027] Furthermore, the operating condition optimization model established in step 5 aims to minimize core loss and maximize transmitted magnetic energy. It transforms the multi-objective problem into a single-objective optimization using a fractional approach, resulting in the constructed single-objective optimization function:

[0028]

[0029] In the formula, minH(P,E) is the minimum energy loss ratio, representing the loss generated per unit of magnetic energy transmitted; P refers to the core loss, which is the loss generated by temperature T, frequency f, and peak magnetic flux density B. m The winding structure W and the core material M are functions; E is the transmitted magnetic energy, determined by f and B. m Decide.

[0030] Because the model has strong nonlinearity, simulated annealing and differential evolution algorithm are used for joint optimization to avoid local optima.

[0031] (III) Beneficial Effects

[0032] Compared with existing technologies, this invention provides a multi-factor core loss prediction method based on machine learning, which has the following beneficial effects:

[0033] By introducing waveform classification and distribution feature extraction mechanisms, the model's ability to perceive complex excitation signals is improved; combined with the temperature index correction model, loss prediction can dynamically respond to temperature changes, enhancing the model's robustness under high-temperature conditions.

[0034] The combination of entropy weighting and statistical analysis effectively identifies and quantifies the importance of each influencing factor, improving the rationality and physical interpretability of the model structure.

[0035] The ensemble learning model built on this basis combines the advantages of multiple algorithms, significantly improving fitting accuracy and generalization ability while keeping model complexity under control.

[0036] By further combining the optimization strategies of differential evolution and simulated annealing algorithms, an optimization model was established with the goal of minimizing core loss and maximizing magnetic energy transfer efficiency, thereby enabling automatic recommendation and scheduling of operating parameters.

[0037] This method can be widely applied to the performance modeling and parameter design of magnetic devices, providing strong support for the design of high-efficiency and energy-saving power electronic systems. Attached Figure Description

[0038] Figure 1 is a flowchart of the multi-factor magnetic core loss prediction method based on machine learning provided by the present invention.

[0039] Figure 2 is a box plot of the distribution characteristic variables provided by the present invention;

[0040] Figure 3 is a feature correlation heatmap provided by the present invention;

[0041] Figure 4 is a confusion matrix diagram of the random forest for different materials provided by the present invention;

[0042] Figure 5 shows the relationship between temperature and core loss provided by the present invention.

[0043] Figure 6 is a comparison of the fitting curves of different mathematical models provided by the present invention as a function of temperature.

[0044] Figure 7 is a comparison of the predicted and actual values ​​of the Steinmetz equation before and after the correction provided by the present invention.

[0045] Figure 8 is a comparison of the error smoothing curves of the Steinmetz equation before and after the correction provided by the present invention.

[0046] Figure 9 is a diagram showing the interaction of three factors: temperature, excitation waveform, and core material, as provided by this invention.

[0047] Figure 10 shows the relationship between peak magnetic flux density and core loss provided by the present invention.

[0048] Figure 11 is a comparison of the predicted and actual values ​​of the random forest and LightGBM ensemble learning model provided by the present invention.

[0049] Figure 12 is a graph showing the iterative curve of the differential optimization algorithm provided by the present invention.

[0050] Figure 13 is an iterative curve of the simulated annealing algorithm provided by the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] Example

[0053] As shown in Figure 1, the multi-factor core loss prediction method based on machine learning proposed in this embodiment of the invention specifically includes the following steps:

[0054] Step 1, Excitation Waveform Feature Extraction and Classification: Obtain magnetic flux density data, extract distribution and shape features, filter out effective feature variables, and construct a classification model to identify the waveform;

[0055] The magnetic flux density data in step 1 includes different material types, temperatures, frequencies, excitation waveform types, and corresponding core losses; it also includes dividing the data into training and testing sets according to a certain ratio.

[0056] The magnetic flux density data consisted of 1024 sampling points, sampled at equal intervals within one cycle, with the unit being Tesla (T). Regarding material types, considering the complexity of the core materials, four typical materials were selected, designated as Material 1, Material 2, Material 3, and Material 4. Four temperature conditions were set for the test: 25℃, 50℃, 70℃, and 90℃. The frequency range for magnetic flux density was 50kHz to 500kHz. The excitation waveform types included sine wave, triangular wave, and trapezoidal wave. Core loss was measured using the AC power method, with the unit being watts per cubic meter (W / m³). 3);

[0057] The distribution characteristics were initially selected as peak value, valley value, mean value, standard deviation, peak-to-peak value, and zero-point crossover number;

[0058] The specific meanings of the selected waveform features are shown in Table 1:

[0059] Table 1 Waveform characteristics of magnetic flux density

[0060]

[0061] The specific method for selecting effective feature variables is as follows: As shown in Figure 2, based on the distribution of the box plot of the distribution feature variables, features that cannot distinguish the waveform are eliminated, and the final selected feature variables are 9: kurtosis, maximum derivative, average peak height, total harmonic distortion, length of flat region, number of inflection points, third harmonic amplitude, fifth harmonic amplitude, and seventh harmonic amplitude.

[0062] As shown in Figure 3, Spearman correlation analysis is then used to filter duplicate features. The calculation formula is as follows:

[0063]

[0064] In the formula, ρ s This is the Spearman correlation coefficient, where n is the sample size and d is the number of samples. i It is the grade difference of the i-th pair of data points;

[0065] As can be seen from the heat map in Figure 3, there is a strong correlation between the amplitudes of the third, fifth, and seventh harmonics, so they are removed.

[0066] The specific method for constructing the classification model is as follows: For the six selected feature variables—kurtosis, maximum derivative, average peak height, total harmonic distortion, length of flat regions, and number of inflection points—a random forest is constructed. 80% of the data is used as the training set, and the remaining 20% ​​is used as the test set. Grid search is used to optimize the model's hyperparameters. The optimal parameters obtained after training are shown in Table 2.

[0067] Table 2 Optimal Parameters for Random Forest

[0068]

[0069] Simultaneously, accuracy, recall, and F1 score are used to evaluate the model, and the calculation formula is as follows:

[0070]

[0071] In the formula, TP indicates that the positive class is correctly predicted, TN indicates that the negative class is correctly predicted, FP indicates that the negative class is misclassified as positive, FN indicates that the positive class is misclassified as negative, Accuracy represents the proportion of samples correctly predicted by the model out of the total samples, which measures the overall accuracy of the prediction, Recall represents the proportion of samples that are actually positive that were correctly predicted by the model, which reflects the model's coverage of the positive class, Precision represents the proportion of samples that are predicted as positive by the model that are actually positive, which reflects the precision of the prediction, and F1 is the harmonic mean of precision and recall, which is used to balance the two.

[0072] The evaluation results of the random forest model on four different materials are shown in Table 3:

[0073] Table 3 Evaluation results of the random forest model on 4 different materials.

[0074]

[0075] Figure 4 shows the confusion matrix results of the random forest model. The values ​​on the diagonal of the confusion matrix represent the number of correctly classified samples, while those outside the diagonal represent the number of misclassified samples. On the coordinate axis, category 0 represents sine waves, category 1 represents triangular waves, and category 2 represents trapezoidal waves. As can be seen from the figure, the random forest model has no errors in identifying and classifying sine waves, but makes a few errors in distinguishing between triangular and trapezoidal waves. Furthermore, the waveform recognition accuracy of material 4 is 1.00, indicating that the waveform classification effect of the model meets the requirements of effectiveness and rationality.

[0076] Step 2, Temperature Modeling Correction: Based on the traditional Steinmetz equation, a temperature sensitivity index is introduced to correct the core loss prediction model, and the model is validated.

[0077] The traditional Steinmetz equation in step 2 refers to the core loss under sinusoidal excitation, and the calculation method is as follows:

[0078]

[0079] In the formula, P is the core loss; f is the frequency; B m α is the peak value of the magnetic flux density; k1, α1, and β1 are coefficients fitted based on experimental data. The formula shows that the core loss per unit volume depends on a power function of the frequency and the peak value of the magnetic flux density. This equation only applies to sinusoidal excitation, and the coefficient parameters may be affected by differences in the working environment, temperature, frequency, excitation waveform, etc. of the magnetic material.

[0080] As shown in Figure 5, the core loss gradually decreases with increasing temperature, and the relationship between the two is not a simple linear one. At the same time, this invention compares four mathematical relationships: linear function, quadratic function, exponential function, and logarithmic function.

[0081] As shown in Figure 6, a comparison chart of the fitting curves of the above four mathematical relationships and the actual data was plotted, and the goodness of fit was calculated. The results show that the goodness of fit of the exponential model is better, reaching 0.997; therefore, the negative exponential relationship is chosen.

[0082] Temperature-corrected core loss prediction model:

[0083]

[0084] In the formula, P v To correct the core loss in the model, k0 is the baseline loss coefficient, which depends on the material type; e is the base of the natural logarithm; γ is the temperature sensitivity coefficient, and the fitted result is γ = 1.26224e⁻⁰²; T is the operating temperature; f is the frequency; α refers to the frequency-dependent exponential coefficient; B m β represents the peak value of the magnetic flux density, and β is the exponential coefficient related to the magnetic flux density.

[0085] The model validation in this embodiment analyzed the root mean square error (RMSE) and goodness of fit (R²) of the Steinmetz equation before and after correction. 2 :

[0086]

[0087] In the formula, y i It is the i-th true value. This is the i-th model prediction, where n is the number of observations. The average of the true values; This represents the sum of squared errors in the model's predictions. R0 represents the sum of squared deviations between the true value and its mean. RMSE measures the magnitude of the error between the predicted and true values; a smaller value indicates a more accurate model prediction. 2 This represents the proportion of variation explained by the model, and its value ranges from [0,1]. The closer it is to 1, the better the model fits.

[0088] The results are shown in Table 4. After correction, the accuracy and stability of the model have been greatly improved.

[0089] Table 4 Comparison of model indicators before and after correction

[0090]

[0091] As shown in Figure 7, the comparison between the predicted and actual values ​​of the Steinmetz equation before and after the correction shows that the corrected model fits the actual values ​​better.

[0092] As shown in Figure 8, the error smoothing curves of the Steinmetz equation before and after correction show that the error is significantly reduced after correction.

[0093] Step 3, Analysis of factors affecting core loss: Analysis of variance is used to evaluate the significance and interaction of temperature, material and waveform, and the entropy weight method is used to quantify their influence and determine the dominant factors;

[0094] In step 3, one-way ANOVA was used to analyze temperature, material and waveform respectively. The results are shown in Table 5. In Table 5, sum_sq is the sum of squares, df is the degrees of freedom, F value is used to measure the significance of the factor, and P value represents the probability that the factor has no effect when the current F value is observed. The smaller the value, the more significant the effect. Usually, <0.05 is used as the significance standard.

[0095] Table 5 Results of One-Way ANOVA

[0096]

[0097] As shown in Table 5, different temperatures, different types of magnetic core materials, and different types of excitation waveforms all have a significant impact on magnetic core loss.

[0098] For the synergistic effect analysis, a two-way ANOVA was used, as shown in Figure 9. The waveform codes are: 2 for sine wave, 1 for trapezoidal wave, and 0 for triangular wave.

[0099] The interaction between temperature and excitation waveform shows that the core loss of a sinusoidal wave is smaller than that of a triangular wave and a trapezoidal wave; different combinations of excitation waveforms at different temperatures will have a significant impact on the core loss.

[0100] Through the interaction between temperature and core material, it can be seen that the core loss of material 2 is greater than that of material 3, which is greater than that of material 1, which is greater than that of material 4. The synergistic effect of temperature and core material on core loss does not show a significant interdependence.

[0101] The interaction between the excitation waveform and the core material shows that different combinations of core materials and excitation waveforms will have a significant impact on core loss.

[0102] For the influence degree analysis, the weights of the influencing factors obtained using the entropy weight method are shown in Table 6:

[0103] Table 6. Weighting Results of Influencing Factors Using the Entropy Weight Method

[0104]

[0105] The results show that the core material has the most significant impact on core loss, the excitation waveform also has a relatively important impact on core loss, while temperature has a relatively small impact on core loss.

[0106] Step 4, Core Loss Prediction Model Construction: An ensemble learning method is used to construct a core loss model, and the effectiveness of the model is analyzed.

[0107] In step 4, as shown in step 3, temperature, core material, and excitation waveform all affect core loss. As shown in step 1, frequency and peak magnetic flux density also affect core loss. As shown in Figure 10, core loss increases with the increase of peak magnetic flux density. Therefore, when constructing the core loss prediction model, the effects of temperature, frequency, peak magnetic flux density, excitation waveform, and core material should be considered. One-hot encoding should be used to eliminate dimensions before establishing the model.

[0108] The ensemble learning method refers to building a combined model of random forest and LightGBM for the processed dataset. The optimal parameter tuning results are shown in Table 7. Here, objective is set to "regression" to indicate that the target task type of the LightGBM model is set to regression, max_depth is the maximum depth of the decision tree, and n_estimators refers to the number of trees in the model.

[0109] Table 7 Optimal parameter values ​​for training Random Forest and LightGBM

[0110]

[0111] The evaluation metrics are Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Median Absolute Error (MedAE), and R² (goodness of fit). The calculation methods for MAE and MedAE are as follows:

[0112]

[0113] In the formula, n is the number of samples, y i It is the actual value. The median is the predicted value; median represents the median function.

[0114] Random Forest, LightGBM, and the combination of Random Forest and LightGBM were compared on the test set in terms of MAE, RMSE, MedAE, and R. 2 As shown in Table 8:

[0115] Table 8 Evaluation metrics of the three models on the test set

[0116]

[0117] Referring to Figure 11, the ensemble learning model combining Random Forest and LightGBM showed the best performance across all four evaluation metrics on the test set, with a significant fit between predicted and actual values. Compared to a single Random Forest model, the combined model reduced the mean squared error by 21596.69, significantly improving both prediction accuracy and model robustness. Furthermore, the introduction of L1 regularization and cross-validation strategies during training further enhanced its generalization ability.

[0118] Step 5, Establishment of the operating condition optimization model: With the goal of minimizing core loss and maximizing transmitted magnetic energy, a single-objective optimization model is constructed by combining differential evolution and simulated annealing algorithms.

[0119] In step 5, before establishing the operating condition optimization model, the core loss prediction model established in step 4 is denoted as:

[0120] P(T,f,Bm,W,M)

[0121] The transfer magnetic energy function is denoted as:

[0122] E(f,B m )=f×B m

[0123] To minimize core losses and maximize magnetic energy transfer, the objective function is:

[0124]

[0125] The constraints are:

[0126]

[0127] In step 5, the operating condition optimization model is established with the objectives of minimizing core loss and maximizing transmitted magnetic energy. The multi-objective problem is transformed into a single-objective optimization using a fractional form, resulting in the constructed single-objective optimization function:

[0128]

[0129] In the formula, minH(P,E) is the minimum loss energy ratio, which represents the loss generated per unit of magnetic energy transmitted; P refers to the core loss, which is a function of temperature T, frequency f, peak magnetic flux density Bm, winding structure W, and core material M; E is the transmitted magnetic energy, which is determined by f and Bm.

[0130] Based on this optimization model, this invention employs a hybrid optimization strategy combining differential evolution algorithm and simulated annealing algorithm. A population iterative update mechanism is used to find the global optimum, and the convergence characteristics of the optimization process are shown in Figures 12 and 13. Experimental verification shows that this optimization method can effectively obtain the optimal combination of multiple key parameters such as temperature, operating frequency, core material, and excitation waveform. As shown in Table 9, under this optimal parameter combination, the system achieves the minimization of core loss and the maximization of transmitted magnetic energy E, reaching the expected optimization objective.

[0131] Table 9 Optimization Results

[0132]

[0133] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-factor core loss prediction method based on machine learning, characterized in that, Includes the following steps: Step 1, Excitation Waveform Feature Extraction and Classification: Obtain magnetic flux density data, extract distribution and shape features, filter out effective feature variables, and construct a classification model to identify the waveform; Step 2, Temperature Modeling Correction: Based on the traditional Steinmetz equation, introduce a temperature sensitivity index to correct the core loss prediction model, and verify the model; The corrected core loss prediction model in Step 2 has a negative exponential relationship, specifically: In the formula, To correct the core loss of the model, This is the baseline loss factor, which is related to the type of material. is the base of the natural logarithm. Let be the temperature sensitivity coefficient, and the fitting result is: = 1.26224e-02， Operating temperature For frequency, The exponential coefficient related to frequency. The peak value of the magnetic flux density. The exponential coefficients are related to the magnetic flux density; the model validation uses the root mean square error (RMSE) and the goodness of fit (R²). 2 Step 3: Analysis of factors affecting core loss: Analysis of variance is used to evaluate the significance and interaction of temperature, material, and waveform; entropy weight method is combined to quantify their influence and determine the dominant factors. Step 4: Construction of core loss prediction model: An ensemble learning method is used to construct a core loss model, and model effectiveness analysis is performed. Step 5: Establishment of operating condition optimization model: A single-objective optimization model is constructed using differential evolution and simulated annealing algorithms, with the objectives of minimizing core loss and maximizing transmitted magnetic energy. The operating condition optimization model in Step 5, with the objectives of minimizing core loss and maximizing transmitted magnetic energy, transforms the multi-objective problem into a single-objective optimization using a fractional form. The constructed single-objective optimization function is as follows: In the formula, It is the minimum energy loss ratio, representing the loss generated per unit of magnetic energy transmitted; Core loss refers to temperature. ,frequency Peak magnetic flux density Winding structure and magnetic core materials The function; To transmit magnetic energy, by and The decision was made to use a combination of simulated annealing and differential evolution algorithm to optimize the model, given its strong nonlinearity, in order to avoid local optima.

2. The multi-factor core loss prediction method based on machine learning according to claim 1, characterized in that, The acquisition of magnetic flux density data in step 1 includes different material types, temperatures, frequencies, excitation waveform types, and corresponding core losses; it also includes dividing the data into training and testing sets according to a certain ratio; the extraction of distribution features includes peak value, valley value, average value, standard deviation, peak-to-peak value, and zero-crossing number; shape features include at least the maximum and minimum derivative values, harmonic amplitudes, and the number of inflection points; the method for selecting effective feature variables is as follows: first, based on box plot analysis, features that cannot distinguish waveforms are filtered out; then, correlation analysis is performed on the remaining features to filter out duplicate features; the finally selected effective feature variables include kurtosis, maximum derivative value, average peak height, total harmonic distortion, length of flat regions, and the number of inflection points; the construction of a classification model to identify waveforms uses a random forest method for training and optimization, with cross-validation and grid search used to adjust parameters during training, and a confusion matrix used to evaluate classification performance.

3. The multi-factor core loss prediction method based on machine learning according to claim 1, characterized in that, The analysis of factors affecting core loss in step 3 includes: using one-way ANOVA to assess the significance of temperature, material, and waveform on core loss; using two-way ANOVA to analyze the interaction between factors; and combining entropy weighting to quantify the degree of influence and determine the dominant factors, specifically by using entropy weighting to calculate the weight indices of temperature, excitation waveform, and core material, with weights of 0.112, 0.394, and 0.494, respectively.

4. The multi-factor core loss prediction method based on machine learning according to claim 1, characterized in that, In step 4, the core loss prediction model is constructed, taking into account the influence of temperature, frequency, peak magnetic flux density, excitation waveform, and core material characteristics. One-hot encoding is used to eliminate the influence of the dimensions of the features. The ensemble learning method adopted is a combination of random forest and LightGBM model. K-fold cross-validation is used to divide the dataset, and L1 regularization is added during training to improve the model's generalization ability. The model effectiveness analysis uses mean absolute error, median absolute error, root mean square error, and goodness of fit to compare the performance differences between the ensemble learning model and the single model.

Images