A fast calculation method of NiZn high-frequency inductor distributed capacitance based on PSO-BPNN algorithm

Abstract

The application discloses a NiZn high-frequency inductor distributed capacitance fast calculation method based on a PSO-BPNN algorithm, and comprises the following steps: a two-dimensional axisymmetric electric field finite element model of a magnetic core inductance of a NiZn high-frequency inductor is established, and sensitivity analysis is performed to screen out key influence factors; a hybrid sampling strategy of discrete parameter layering and continuous parameter Latin hypercube sampling is adopted to generate multi-parameter samples meeting structure constraints, form a distributed capacitance data set, and train a NiZn high-frequency inductor distributed capacitance fast prediction model based on a particle swarm optimization back propagation neural network; and the key influence factors of the NiZn high-frequency inductor to be calculated are input into the trained prediction model, and corresponding total distributed capacitance, winding capacitance and magnetic core capacitance are output. The application effectively reduces the simulation calculation cost by combining the hybrid sampling strategy with the finite element simulation to construct a high-precision sample data set and by relying on the PSO-BPNN algorithm to establish a parameter intelligent prediction model, and meets the engineering requirements of high-frequency inductor fast design and parameter optimization.

Description

Technical Field

[0001] This invention relates to the field of high-frequency power magnetic technology in power electronics, specifically to a fast calculation method for the distributed capacitance of NiZn high-frequency inductance based on the PSO-BPNN algorithm. Background Technology

[0002] NiZn soft magnetic ferrites are widely used in MHz-level high-frequency magnetic devices due to their high resistivity, low high-frequency loss, and good permeability stability. Compared to MnZn ferrites, NiZn cores have lower dielectric constants and conductivity, and cannot be approximated as ideal conductors under high-frequency electric fields. The electric field can penetrate the core and store energy, making the core capacitance an important component of the high-frequency inductance distributed capacitance. Its high-frequency modeling mechanism differs significantly from that of traditional MnZn cores.

[0003] Existing research has abandoned the ideal conductor assumption of NiZn magnetic cores and confirmed the influence of the electric field energy storage inside the core on the distributed capacitance, but significant technical limitations remain. Traditional analytical models suffer from low accuracy, numerous simplifying assumptions, high technical barriers, and limited practicality. They require solving the Laplace equation, making calculations complex and difficult to apply conveniently for rapid calculations and designs. While finite element simulation offers high accuracy, it suffers from high computational costs and slow iteration speeds, failing to meet the demands of rapid parameter optimization for large-scale engineering projects.

[0004] In addition, existing AI electromagnetic prediction technologies are mostly applied to microwave devices, on-chip inductors and interconnect structures. Although they can efficiently fit the nonlinear relationship of electromagnetic parameters and replace repetitive simulations, they are rarely studied for NiZn high-frequency power inductors. There is a lack of modeling schemes that adapt to their structural constraints, discrete parameter characteristics and joint prediction of multiple capacitance components, and it is impossible to achieve accurate and rapid prediction of winding capacitance, core capacitance and total distributed capacitance.

[0005] In summary, existing technologies suffer from problems such as insufficient modeling specificity, difficulty in balancing accuracy and efficiency, and lack of a parameterized calculation and intelligent prediction system adapted to NiZn high-frequency inductors, making it difficult to meet the engineering requirements for accurate modeling and rapid optimization of high-frequency inductors. Summary of the Invention

[0006] The purpose of this invention is to provide a fast calculation method for the distributed capacitance of NiZn high-frequency inductors based on the PSO-BPNN algorithm. This method aims to solve the technical problems of existing technologies, such as poor modeling specificity for NiZn high-frequency inductors, inability to balance calculation accuracy and simulation efficiency, and difficulty in achieving fast and accurate prediction of multi-component distributed capacitance. By combining a hybrid sampling strategy with finite element simulation to construct a high-precision sample dataset, and relying on algorithms such as Particle Swarm Optimization Backpropagation Neural Network (PSO-BPNN), a parameter intelligent prediction model is established, which effectively reduces the simulation calculation cost and meets the engineering needs of rapid design and parameter optimization of high-frequency inductors.

[0007] To achieve the above objectives, the solution of the present invention is:

[0008] A fast calculation method for the distributed capacitance of NiZn high-frequency inductance based on the PSO-BPNN algorithm includes,

[0009] A two-dimensional axisymmetric finite element model of the core inductance of a NiZn high-frequency inductor was established; the total distributed capacitance, winding capacitance, and core capacitance were obtained; and key influencing factors were screened based on the sensitivity analysis of the core inductor parameters.

[0010] Based on the finite element model, a hybrid sampling strategy of discrete parameter hierarchical and continuous parameter Latin hypercube sampling is adopted to generate multi-parameter samples that satisfy structural constraints, forming a distributed capacitance dataset that includes key influencing factors and total distributed capacitance, winding capacitance and core capacitance.

[0011] A fast prediction model for the distributed capacitance of high-frequency inductors in NiZn is constructed based on a particle swarm optimization backpropagation neural network, and the model is trained based on the distributed capacitance dataset.

[0012] The key influencing factors of the NiZn high-frequency inductor to be calculated are obtained, and the trained NiZn high-frequency inductor distributed capacitance fast prediction model is input to obtain the corresponding total distributed capacitance, winding capacitance and core capacitance.

[0013] This includes obtaining the total distributed capacitance, winding capacitance, and core capacitance, including...

[0014] The total distributed capacitance is obtained by integrating the electric field energy, and the electric field energy density w is obtained first. e for,

[0015]

[0016] Where ε is the dielectric constant of the medium, and E is the electric field strength;

[0017] Thus, the total electric field energy W is obtained. e for,

[0018]

[0019] The total distributed capacitance C ind for,

[0020]

[0021] Where U is the voltage across the winding;

[0022] While keeping the winding geometry, potential boundaries, and external air region settings unchanged, the relative permittivity of the core region is equivalently set to 1. The electric field energy of the winding insulation layer, inter-turn air, and air region near the winding is integrated to obtain the winding electric field energy W. ew Thus, the winding capacitance C is obtained. w for,

[0023]

[0024] The difference between the total distributed capacitance and the winding capacitance is taken as the core capacitance, i.e.

[0025]

[0026] Among them, C ce It is a magnetic core capacitor.

[0027] Among them, based on the sensitivity analysis of the magnetic core inductance parameters, key influencing factors were identified, including:

[0028] Based on the established two-dimensional axisymmetric finite element model of the magnetic core inductance of NiZn high-frequency inductors, several parameters are selected as parameters to be analyzed, and the total distributed capacitance C is taken as the criterion. ind Winding capacitance C w and magnetic core capacitor C ce As the output response quantity of sensitivity analysis;

[0029] Based on a reference inductor structure, a single-parameter scan is performed, where only one parameter to be analyzed is changed each time while other parameters remain constant. Electric field simulations are conducted on the finite element model under different parameter values, and the corresponding C is extracted based on the electric field energy integration method. ind C w C ce ;

[0030] The data obtained from the single-parameter scan were fitted using a log-log regression method. The relationship between parameter x and distributed capacitance C was expressed as lnC = a·lnx + b, where the regression coefficient a serves as the sensitivity coefficient of the corresponding parameter to the distributed capacitance. When the output quantities are C ind C w C ce At that time, the sensitivity coefficients of each parameter to the three types of capacitors were calculated respectively;

[0031] Based on the absolute values ​​of the sensitivity coefficients of each parameter, different parameters are compared with C. ind C w C ce The influence levels were ranked to identify the key influencing factors that significantly affect distributed capacitance.

[0032] Among them, a hybrid sampling strategy employing discrete parameter hierarchical and continuous parameter Latin hypercube sampling includes,

[0033] The number of inductor turns is used as a discrete hierarchical variable to divide the system into multiple parameter subspaces. Within each subspace, key influencing factors other than the number of inductor turns are sampled using Latin hypercube sampling. The sampled samples are then filtered in conjunction with the engineering structural constraints of the inductor to remove invalid samples, resulting in a parameter sample set.

[0034] Among them, a fast prediction model for the distributed capacitance of high-frequency inductance in NiZn is constructed based on a particle swarm optimization backpropagation neural network, including:

[0035] The particle swarm optimization backpropagation neural network consists of an input layer, a shared feature extraction layer, and an output layer. The input layer is the key influencing factor. The shared feature extraction layer consists of at least one fully connected hidden layer. The neurons in each hidden layer perform a weighted summation of the output of the previous layer and output the result after transformation by a nonlinear activation function. The output layer has three regression output nodes, which are used to output the total distributed capacitance, winding capacitance, and core capacitance, respectively.

[0036] The model is trained based on the distributed capacitance dataset, including,

[0037] The distributed capacitance dataset is divided into a training set, a validation set, and a test set.

[0038] Input the input vector into the model to obtain a 3D output result, namely the standardized predicted values ​​of total distributed capacitance, winding capacitance and core capacitance;

[0039] The input-to-hidden-layer connection weights, hidden-to-output-layer connection weights, hidden-layer biases, and output-layer biases in the model are uniformly encoded as particle position vectors. Each particle in the particle swarm corresponds to a complete set of initial weights and bias parameters.

[0040] The initial position and initial velocity of each particle in the particle swarm are randomly generated. The weight and bias corresponding to each particle are assigned to the model. Forward propagation calculation is performed using training set samples to obtain the predicted values ​​of total distributed capacitance, winding capacitance and core capacitance. The mean square error between the predicted value and the finite element label value is used as the particle fitness function.

[0041] Based on the fitness calculation results of each particle, the individual optimal position of each particle is updated, and the global optimal position of the entire particle swarm is further updated.

[0042] The velocity and position of particles are iteratively updated, and the position and velocity of particles that exceed the boundary are corrected.

[0043] After the particle swarm iteration is completed, the position vector of the globally optimal particle is decoded into the initial weights and biases of the model and assigned to each connection layer of the network.

[0044] Using the optimized weights and biases as the initial parameters of the model, forward propagation and backpropagation of the error are performed using the training set. The network weights and biases are updated according to the loss function, and the model training process is monitored using the validation set. The network parameters corresponding to the minimum error in the validation set are selected as the final model parameters, thus obtaining a fast prediction model for the distributed capacitance of NiZn high-frequency inductance.

[0045] The NiZn high-frequency inductors include cylindrical core inductors and improved can core inductors.

[0046] By adopting the above solution, the present invention has the following beneficial effects:

[0047] (1) This invention uses the electric field energy method to theoretically classify the distributed capacitance of NiZn high-frequency inductors and clarify the contribution of the core capacitance. Compared with the traditional model that only considers the winding capacitance, it significantly improves the calculation accuracy of distributed capacitance and provides a more accurate parameter basis for the high-frequency characteristic analysis of high-frequency inductors.

[0048] (2) The present invention adopts a hybrid sampling strategy that combines discrete parameter hierarchical and Latin hypercube sampling to efficiently construct a high-quality dataset covering the entire parameter space, significantly reduce the cost of batch simulation of finite element method, and balance sample representativeness and computational efficiency.

[0049] (3) The present invention constructs a PSO-BPNN multi-output prediction model and introduces physical consistency constraints, which can simultaneously and quickly predict the total distributed capacitance, winding capacitance and core capacitance. Compared with the traditional finite element method, the computational efficiency is significantly improved, and the model has strong generalization ability and high engineering applicability.

[0050] (4) Although the model constructed in this invention is trained on a finite-scale dataset, it has good generalization and extension capabilities and can be extended to the distributed capacitance calculation scenario of magnetic cores of different sizes and specifications. Attached Figure Description

[0051] Figure 1 This is a flowchart of the method of the present invention;

[0052] Figure 2 This is a schematic diagram of the equivalent circuit model of a high-frequency inductor according to an embodiment of the present invention;

[0053] Figure 3 This is a finite element simulation model of the two-dimensional axisymmetric electric field of the improved can-type magnetic core inductor according to an embodiment of the present invention;

[0054] Figure 4 These are electric field distribution diagrams of the cylindrical magnetic core and the improved pot-type magnetic core according to embodiments of the present invention;

[0055] Among them, (a) is the electric field distribution diagram of the cylindrical magnetic core, and (b) is the electric field distribution diagram of the improved pot magnetic core;

[0056] Figure 5 This is a graph showing the comparison of the prediction performance of the PSO-BPNN model in this embodiment of the invention with other models;

[0057] Among them, (a) is a comparison chart of the average MAE, (b) is a comparison chart of the average RMSE, (c) is a comparison chart of the average MAPE, and (d) is a comparison chart of the average RSE. 2 Average value comparison chart;

[0058] Figure 6 This is a scatter plot comparing the actual and predicted values ​​of the distributed capacitance of the model test set in this embodiment of the invention.

[0059] Among them, (a) is a comparison diagram of total distributed capacitance, (b) is a comparison diagram of winding capacitance, and (c) is a comparison diagram of magnetic core capacitance.

[0060] Figure 7 This is a comparison chart of the predicted and measured values ​​of the distributed capacitance of the cylindrical magnetic core and the improved pot magnetic core according to an embodiment of the present invention;

[0061] Among them, (a) is a comparison diagram of the distributed capacitance of cylindrical magnetic cores, and (b) is a comparison diagram of the distributed capacitance of improved pot-type magnetic cores. Detailed Implementation

[0062] The technical solution and beneficial effects of the present invention will be described in detail below with reference to the accompanying drawings.

[0063] The following embodiments are only used to illustrate the technical solutions of the present invention more clearly, and should not be used to limit the scope of protection of the present invention.

[0064] like Figure 1 As shown, this invention provides a fast calculation method for the distributed capacitance of NiZn high-frequency inductance based on the PSO-BPNN algorithm. The specific steps are as follows:

[0065] S1. Based on Ansys Maxwell, a two-dimensional axisymmetric electric field finite element model of NiZn magnetic core inductor was established and finite element simulation was performed to conduct parameter sensitivity analysis.

[0066] According to electromagnetic field theory, the electric field energy density can be expressed as:

[0067]

[0068] Where ε is the dielectric constant of the medium, and E is the electric field strength. Integrating over the entire inductive region, the total electric field energy can be obtained:

[0069]

[0070] According to the definition of energy of a capacitor, the equivalent capacitance satisfies:

[0071]

[0072] Therefore, the total distributed capacitance can be uniformly expressed as:

[0073]

[0074] Based on electric field distribution analysis, the total distributed capacitance of a high-frequency NiZn core inductor can be decomposed into the winding capacitance C. w With magnetic core capacitor C ce Satisfying C ind =C w +C ce .

[0075] While keeping the winding geometry, potential boundaries, and external air region settings unchanged, the relative permittivity of the core region is equivalently set to 1. The electric field energy of the winding insulation layer, inter-turn air, and air region near the winding is integrated to obtain the winding electric field energy W. ew This leads to the winding capacitance:

[0076]

[0077] The difference between the total distributed capacitance and the winding capacitance is taken as the core capacitance, i.e.:

[0078]

[0079] C under different parameters was extracted using finite element simulation. ind C w C ce Sensitivity analysis was performed on parameters such as outer radius, window height, number of turns, and core dielectric constant.

[0080] The parameter sensitivity analysis includes:

[0081] 1) Determine the reference inductor structure and parameters to be analyzed. Based on the established two-dimensional axisymmetric electric field finite element model of the NiZn high-frequency inductor, select the outer radius r t Window height l t End cap height h, center post radius r c Window width w, number of turns N, and relative permittivity of the magnetic core ε r As the parameter to be analyzed, and with the total distributed capacitance C as the parameter to be analyzed. ind Winding capacitance C w and magnetic core capacitor C ce As the output response quantity of sensitivity analysis.

[0082] 2) Perform similarity scaling analysis. Maintain the number of turns N and the relative permittivity of the core ε. r And with the winding topology unchanged, the outer radius r t Window height l t End cap height h, center post radius r c The window width w is synchronously scaled up or down using the same similarity scaling factor α. Calculate C under different scaling factors. ind C w C ce It is used to analyze the influence of changes in overall geometric dimensions on distributed capacitance.

[0083] 3) Perform single-parameter sweep analysis. Based on the reference inductor structure, only one parameter to be analyzed is changed each time, while other parameters remain unchanged. Electric field simulations are performed on the finite element model under different parameter values, and the corresponding C is extracted based on the electric field energy integration method. ind C w C ce This allows us to obtain the relationship between the various parameters and the distributed capacitance.

[0084] 4) Calculate the normalized sensitivity coefficient. The data obtained from the single-parameter scan is fitted using a log-log regression method. The relationship between parameter x and distributed capacitance C is expressed as lnC = a·lnx + b, where the regression coefficient a serves as the sensitivity coefficient of the corresponding parameter to the distributed capacitance. When the output quantities are C... ind C w C ce At that time, the sensitivity coefficients of each parameter to the three types of capacitors were calculated respectively.

[0085] 5) Screening key influencing factors. Based on the absolute value of the sensitivity coefficients of each parameter, analyze the impact of different parameters on C. ind C w C ce The influence of each factor was ranked, and combined with the inductor structure realizability constraints, key influencing factors that significantly affect distributed capacitance were screened out as input parameters for subsequent sample sampling and PSO-BPNN prediction models.

[0086] S2. A method combining discrete parameter hierarchical sampling and continuous parameter Latin hypercube sampling is used to generate multi-parameter samples that satisfy structural constraints, constructing a sample containing C. ind C w C ce A labeled distributed capacitance dataset; using the number of turns N as a discrete hierarchical variable, ranging from 8 to 13, constructing 6 parameter subspaces; the outer radius r within each subspace. t Window height l t , radius r of the central column c Window width w, relative permittivity of magnetic core εr Latin hypercube sampling was performed using the following method:

[0087] 1) Determine the sample size M;

[0088] 2) Divide the range of each parameter evenly into M intervals;

[0089] 3) Randomly select points within each interval and shuffle them;

[0090] 4) Map to the actual physical parameter range.

[0091] Simultaneously, by combining aspect ratio, window fill rate, and outer magnetic ring thickness constraints, valid samples were screened, and a dataset containing 647 samples was finally constructed.

[0092] S3. Construct a Particle Swarm Optimization Backpropagation Neural Network (PSO-BPNN) and train it to obtain a fast prediction model for the distributed capacitance of NiZn high-frequency inductors. The PSO-BPNN model is a multi-input, multi-output regression prediction model, and its network structure includes an input layer, a shared feature extraction layer, and an output layer. The input layer receives the inductor structure parameters and material parameters after Z-score normalization. The input variables include the number of turns N and the outer radius r. t Window height l t , radius r of the central column c Window width w and relative permittivity ε of the magnetic core r There are a total of 6 input nodes. The shared feature extraction layer consists of at least one fully connected hidden layer. The neurons in each hidden layer perform a weighted summation of the output of the previous layer, and then output the result after transformation by a nonlinear activation function. This summation is used to extract the common features of structural and material parameters on the electric field energy storage distribution. The output layer has 3 regression output nodes, corresponding to the total distributed capacitance C. ind Winding capacitance C w and magnetic core capacitor C ce This enables synchronous prediction of the three capacitors.

[0093] The training process of the PSO-BPNN includes the following steps:

[0094] 1) Preprocess the distributed capacitance dataset. The samples obtained from the finite element simulation are divided into training, validation, and test sets in a ratio of 70%:15%:15%. The input parameters and output capacitance labels are Z-score standardized to reduce the impact of differences in the dimensions and numerical ranges of different parameters on model training.

[0095] 2) Establish a BPNN feedforward network. The standardized 6-dimensional input vector is input into the BPNN input layer. The weighted sum of the fully connected weight matrix from the input layer to the hidden layer and the hidden layer bias is then applied to the hidden layer's non-linear activation function to obtain the hidden layer features. These hidden layer features are further mapped through the fully connected weight matrix from the hidden layer to the output layer and the output layer bias to obtain the 3-dimensional output result, i.e., C. ind C w C ce The standardized predicted value.

[0096] 3) Construct particle encoding. The input-to-hidden-layer connection weights, hidden-to-output-layer connection weights, hidden-layer biases, and output-layer biases in the BPNN are uniformly encoded into particle position vectors. Each particle in the particle swarm corresponds to a complete set of initial BPNN weights and bias parameters.

[0097] 4) Initialize the particle swarm. Randomly generate the initial position and initial velocity of each particle in the particle swarm, where the particle position represents the initial weights and biases of the BPNN, and the particle velocity represents the update direction and update step size of the network parameters in the search space.

[0098] 5) Calculate particle fitness. Assign the weights and biases corresponding to each particle to the BPNN model, and perform forward propagation calculation using the training set samples to obtain C. ind C w C ce The predicted values ​​are used, and the mean square error between the predicted values ​​and the finite element label values ​​is used as the particle fitness function. The smaller the fitness value, the better the initial parameters of the BPNN corresponding to the particle.

[0099] 6) Update the optimal position of particles. Based on the fitness calculation results of each particle, update the individual optimal position of each particle, and further update the global optimal position of the entire particle swarm, so that the particle swarm gradually searches towards the network initial parameter region with smaller prediction error.

[0100] 7) Update particle velocity and position. Based on the velocity and position update formulas of the particle swarm optimization algorithm, and combined with inertia weights, individual optimal positions, and global optimal positions, the velocity and position of particles are iteratively updated, and the positions and velocities of particles that exceed the boundaries are corrected.

[0101] 8) Obtain the optimal initial parameters of the BPNN. After the particle swarm optimization is completed, the position vector of the globally optimal particle is decoded into the initial weights and biases of the BPNN and assigned to each connection layer of the BPNN network, thus completing the global optimization of the BPNN initial parameters by PSO.

[0102] 9) Perform local training of the BPNN. Using the PSO-optimized weights and biases as the initial parameters of the BPNN, perform forward propagation and error backpropagation training using the training set, updating the network weights and biases according to the loss function. During training, the model simultaneously calculates data loss and physical consistency loss, where the data loss is used to constrain C. ind C w C ce The deviation between the predicted value and the finite element simulation label, the physical consistency loss is used to constrain the prediction results to meet C. ind =C w +C ce The relationship between distributed capacitance.

[0103] 10) Determine the final prediction model. Monitor the PSO-BPNN model training process using the validation set, and select the network parameters corresponding to the minimum validation set error as the final model parameters to obtain the fast prediction model of NiZn high-frequency inductance distributed capacitance.

[0104] 11) Perform rapid calculation of distributed capacitance. Obtain the number of turns N and outer radius r of the NiZn high-frequency inductor to be calculated. t Window height l t , radius r of the central column c Window width w and relative permittivity ε of the magnetic core r After being standardized according to the Z-score normalization parameters obtained during the training phase, the input is fed into the trained PSO-BPNN model. The model then propagates forward through the input layer, hidden layer, and output layer, outputting the standardized C. ind C w C ce The predicted results are then denormalized to obtain the actual total distributed capacitance, winding capacitance, and core capacitance.

[0105] Figure 2 This is a schematic diagram of the equivalent circuit model of a high-frequency inductor, including the DC resistance of the winding, inductance, iron loss equivalent resistance, and distributed capacitance. The distributed capacitance directly affects the high-frequency impedance and self-resonance characteristics.

[0106] Figure 3 This is a finite element simulation model of the two-dimensional axisymmetric electric field of an improved can-core inductor. Figure 4 The electric field distribution diagrams for cylindrical and improved pot-type magnetic cores are shown. The model includes a central column, outer magnetic ring, end caps, and a single-layer winding. The distributed capacitance parameters are extracted using the electric field energy integration method.

[0107] Figure 5 The chart compares the prediction performance of the PSO-BPNN model with other models. The results show that the PSO-BPNN model performs better in terms of mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). 2It is the best in all metrics, and its prediction accuracy is significantly better than traditional BPNN, SVR and RF models.

[0108] Figure 6 This is a comparison chart of the actual and predicted values ​​of the distributed capacitance on the test set of the PSO-BPNN model. The scatter points closely match the diagonal, indicating that the model has strong nonlinear fitting ability and good physical consistency.

[0109] Figure 7 The chart compares the predicted and measured inductance values ​​for cylindrical and improved pot-type magnetic cores. The predicted error for the cylindrical core is 2.9%, while the predicted error for the improved pot-type core is 17.1%, both meeting engineering accuracy requirements. However, the improved pot-type core exhibits a larger error due to its more complex structure.

[0110] In this embodiment, an Ansys Maxwell 2023 finite element model was established, and a dataset was generated through batch simulation. A PSO-BPNN model was built using Python to complete training and testing. NiZn magnetic core inductor samples were used in the experiment, and a Keysight E4990A impedance analyzer was used to measure the high-frequency impedance. The total distributed capacitance was calculated using the resonant frequency.

[0111] This invention classifies distributed capacitance using the electric field energy method, accurately quantifying the capacitance contribution of NiZn magnetic cores; it employs a hybrid sampling strategy to construct a high-quality dataset, significantly reducing simulation costs; and it builds a PSO-BPNN multi-output prediction model to achieve rapid and accurate prediction of distributed capacitance. Compared with traditional analytical methods and finite element methods, this invention balances accuracy and efficiency, requires no additional hardware investment, is easy to implement, highly practical, and can be directly used for rapid design and parameter optimization of high-frequency inductors.

[0112] This invention also proposes a NiZn high-frequency inductance distributed capacitance prediction system, including a finite element simulation module, a dataset construction module, and a PSO-BPNN prediction module. It should be noted that each module in the above system corresponds to a specific step in the method of this invention, possessing corresponding functions and beneficial effects; details not described in detail can be found in the method embodiments.

[0113] This invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-described method steps and has corresponding functions and beneficial effects.

[0114] This invention also proposes a computer-readable storage medium storing a computer program, which, when run by a processor, implements the above-described method steps and has corresponding functions and beneficial effects.

[0115] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A fast calculation method for the distributed capacitance of NiZn high-frequency inductance based on the PSO-BPNN algorithm, characterized in that: include, A two-dimensional axisymmetric finite element model of the magnetic core inductor of NiZn high-frequency inductor was established; the total distributed capacitance, winding capacitance and magnetic core capacitance were obtained, and key influencing factors were screened based on the sensitivity analysis of the magnetic core inductor parameters. Based on the aforementioned finite element model, a hybrid sampling strategy of discrete parameter hierarchical and continuous parameter Latin hypercube sampling is adopted to generate multi-parameter samples that satisfy structural constraints, forming a distributed capacitance dataset that includes key influencing factors and total distributed capacitance, winding capacitance, and core capacitance. A fast prediction model for the distributed capacitance of high-frequency inductors in NiZn is constructed based on a particle swarm optimization backpropagation neural network, and the model is trained based on the distributed capacitance dataset. The key influencing factors of the NiZn high-frequency inductor to be calculated are obtained, and the trained NiZn high-frequency inductor distributed capacitance fast prediction model is input to obtain the corresponding total distributed capacitance, winding capacitance and core capacitance.

2. The method as described in claim 1, characterized in that: Obtain the total distributed capacitance, winding capacitance, and core capacitance, including: The total distributed capacitance is obtained by integrating the electric field energy, and the electric field energy density w is obtained first. e for, ， Where ε is the dielectric constant of the medium, and E is the electric field strength; Thus, the total electric field energy W is obtained. e for, ， The total distributed capacitance C ind for, ， Where U is the voltage across the winding; While keeping the winding geometry, potential boundaries, and external air region settings unchanged, the relative permittivity of the core region is equivalently set to 1. The electric field energy of the winding insulation layer, inter-turn air, and air region near the winding is integrated to obtain the winding electric field energy W. ew Thus, the winding capacitance C is obtained. cw for, ， The difference between the total distributed capacitance and the winding capacitance is taken as the core capacitance, i.e. ， Among them, C ce It is a magnetic core capacitor.

3. The method as described in claim 1, characterized in that: Based on the sensitivity analysis of the magnetic core inductance parameters, key influencing factors were identified, including: Based on the established two-dimensional axisymmetric finite element model of the magnetic core inductance of NiZn high-frequency inductors, several parameters are selected as parameters to be analyzed, and the total distributed capacitance C is taken as the criterion. ind Winding capacitance C w and magnetic core capacitor C ce As the output response quantity of sensitivity analysis; Based on a reference inductor structure, a single-parameter scan is performed, where only one parameter to be analyzed is changed each time while other parameters remain constant. Electric field simulations are conducted on the finite element model under different parameter values, and the corresponding C is extracted based on the electric field energy integration method. ind C w C ce ; The data obtained from the single-parameter scan were fitted using a log-log regression method. The relationship between parameter x and distributed capacitance C was expressed as lnC = a·lnx + b, where the regression coefficient a serves as the sensitivity coefficient of the corresponding parameter to the distributed capacitance. When the output quantities are C ind C w C ce At that time, the sensitivity coefficients of each parameter to the three types of capacitors were calculated respectively; Based on the absolute values ​​of the sensitivity coefficients of each parameter, different parameters are compared with C. ind C w C ce The influence levels were ranked to identify the key influencing factors that significantly affect distributed capacitance.

4. The method as described in claim 1, characterized in that: A hybrid sampling strategy employing discrete parameter hierarchical and continuous parameter Latin hypercube sampling is adopted, including: The number of inductor turns is used as a discrete hierarchical variable to divide multiple parameter subspaces. In each subspace, key influencing factors other than the number of inductor turns are sampled by Latin hypercube sampling. By combining the constraints of the inductor engineering structure, the sampled samples are screened, invalid samples are removed, and a parameter sample set is obtained.

5. The method as described in claim 1, characterized in that: A fast prediction model for the distributed capacitance of high-frequency inductance in NiZn is constructed based on a particle swarm optimization backpropagation neural network, including: The particle swarm optimization backpropagation neural network consists of an input layer, a shared feature extraction layer, and an output layer. The input layer is the key influencing factor. The shared feature extraction layer consists of at least one fully connected hidden layer. The neurons in each hidden layer perform a weighted summation of the output of the previous layer and then output the result after transformation by a nonlinear activation function. The output layer is set with three return output nodes, which are used to output the total distributed capacitance, winding capacitance and core capacitance, respectively.

6. The method as described in claim 5, characterized in that: The model is trained based on the aforementioned distributed capacitance dataset, including: The distributed capacitance dataset is divided into a training set, a validation set, and a test set. Input the input vector into the model to obtain a 3D output result, namely the standardized predicted values ​​of total distributed capacitance, winding capacitance and core capacitance; The input-to-hidden-layer connection weights, hidden-to-output-layer connection weights, hidden-layer biases, and output-layer biases in the model are uniformly encoded as particle position vectors. Each particle in the particle swarm corresponds to a complete set of initial weights and bias parameters. The initial position and initial velocity of each particle in the particle swarm are randomly generated. The weight and bias corresponding to each particle are assigned to the model. Forward propagation calculation is performed using training set samples to obtain the predicted values ​​of total distributed capacitance, winding capacitance and core capacitance. The mean square error between the predicted value and the finite element label value is used as the particle fitness function. Based on the fitness calculation results of each particle, the individual optimal position of each particle is updated, and the global optimal position of the entire particle swarm is further updated. The velocity and position of particles are iteratively updated, and the position and velocity of particles that exceed the boundary are corrected. After the particle swarm iteration is completed, the position vector of the globally optimal particle is decoded into the initial weights and biases of the model and assigned to each connection layer of the network. Using the optimized weights and biases as the initial parameters of the model, forward propagation and backpropagation of the error are performed using the training set. The network weights and biases are updated according to the loss function, and the model training process is monitored using the validation set. The network parameters corresponding to the minimum error in the validation set are selected as the final model parameters, thus obtaining a fast prediction model for the distributed capacitance of NiZn high-frequency inductance.

7. The method as described in claim 1, characterized in that: The NiZn high-frequency inductors include cylindrical core inductors and improved can core inductors.

Images