Converter circuit parameter multi-objective gradient optimization method and device

Abstract

The invention provides a converter circuit parameter multi-objective gradient optimization method and device, and the method comprises the steps: determining the performance index of a converter, and carrying out the modeling of the performance index of the converter; obtaining possible design parameters and corresponding performance indexes of the converter, and constructing an input space sample and an output space sample; constructing a performance estimation model, and optimizing a neural architecture of the performance estimation model through a non-dominated sorting genetic algorithm in combination with the input space sample and the output space sample; and determining a limit range of design parameters of the converter, and generating optimal design parameters in the limit range through the performance estimation model. By means of the method, automation of model building can be achieved, the modeling efficiency and the model precision are remarkably improved, and the problem that an optimal parameter combination is difficult to obtain through an existing converter circuit parameter design method is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of converter design, and in particular to a multi-objective gradient optimization method and device for converter circuit parameters. Background Art

[0002] As the core device for power conversion, power converters play a vital role in energy storage, electric vehicles, renewable energy systems and other fields. To meet the increasingly stringent performance requirements in practical applications, power converter design usually requires the simultaneous optimization of multiple conflicting design goals, such as high conversion efficiency, small size, low cost, and low output distortion. However, higher conversion efficiency often requires larger or higher-performance components such as inductors and capacitors, which inevitably leads to an increase in converter size and cost; conversely, in pursuit of a lightweight and compact design, compromises may be made in efficiency or other performance indicators. Therefore, how to achieve the optimal parameter design of the power converter under multi-objective constraints becomes crucial.

[0003] Currently, power converter parameter optimization design methods can be divided into three main categories: traditional human-based design (HD), computer-aided optimization design (CAO-D), and artificial intelligence-based design (AI-D). These optimization design methods generally include two key steps: modeling and search. The modeling step aims to establish a mathematical mapping relationship between design parameters and optimization objectives, while the search step is to find the parameter combination that achieves optimal performance based on the model.

[0004] The HD method relies on engineers' expert experience to perform the entire modeling and parameter search process. Although it played an important role in early power converter design, its drawbacks have become increasingly apparent in the face of increasingly complex power converter topologies and increasingly stringent performance targets. On the one hand, engineers must devote considerable time and effort to tedious formula derivation and parameter tuning, resulting in lengthy design cycles. On the other hand, designs driven by expert experience often struggle to balance multi-objective optimization, easily falling into local optimal solutions and failing to fully tap the converter's performance potential.

[0005] The CAO-D method automates the optimization process through the use of an optimization algorithm, thereby reducing the complexity of power converter parameter optimization design to a certain extent. However, the performance of the CAO-D method still relies heavily on accurate mathematical models. For complex systems such as power converters with strong nonlinearities and multi-physics coupling, establishing a high-precision analytical model is extremely challenging. Insufficient model accuracy directly limits the reliability of the optimization results. Furthermore, the optimization algorithm's search efficiency in high-dimensional parameter spaces can also become a bottleneck.

[0006] In recent years, AI-D methods have demonstrated tremendous potential. They leverage artificial intelligence (AI) to learn the complex relationships between design parameters and performance metrics. They construct data-driven performance models and combine them with optimization algorithms to search for optimal parameter combinations. However, existing AI-D methods still rely on human experience and trial-and-error to select model structures and adjust hyperparameters during the modeling phase. This lack of human experience can also lead to insufficient model generalization. Changes in design requirements or power converter performance may require model readjustment or even reconstruction, making true automation and generalizability difficult. Furthermore, existing methods primarily focus on improving performance metrics such as efficiency, total harmonic distortion (THD), cost, and size, while rarely considering the health of power converter switches.

[0007] There is currently no effective solution to the problem that existing converter circuit parameter design methods are difficult to obtain the optimal parameter combination. Summary of the Invention

[0008] The present invention provides a method and device for multi-objective gradient optimization of converter circuit parameters, which are used to solve the defect that the existing converter circuit parameter design method is difficult to obtain the optimal parameter combination.

[0009] In a first aspect, the present invention provides a multi-objective gradient optimization method for converter circuit parameters, comprising: Determining performance indicators of the converter and modeling the performance indicators of the converter; Obtaining possible design parameters and corresponding performance indicators of the converter, and constructing input space samples and output space samples; Constructing a performance estimation model, combining the input space samples and the output space samples, and optimizing the neural architecture of the performance estimation model through a non-dominated sorting genetic algorithm; The limiting range of the design parameters of the converter is determined, and the optimal design parameters are generated within the limiting range by the performance estimation model.

[0010] According to a multi-objective gradient optimization method for converter circuit parameters provided by the present invention, the performance indicators of the converter include conversion efficiency, total harmonic distortion, bus capacitance ripple, output current ripple, switch tube reliability, volume and cost; The design parameters of the converter include capacitance, inductance, and switching frequency.

[0011] According to a multi-objective gradient optimization method for converter circuit parameters provided by the present invention, the input space samples are different combinations of the design parameters, and the output space samples are performance indicators corresponding to the design parameter combinations.

[0012] According to a multi-objective gradient optimization method for converter circuit parameters provided by the present invention, a performance estimation model is constructed, and the neural architecture of the performance estimation model is optimized by a non-dominated sorting genetic algorithm based on the input space samples and the output space samples, including: Constructing a training set and a test set based on the input space samples and the output space samples; defining an encoding method of a neural architecture of the performance estimation model, initializing the neural architecture, and determining algorithm parameters of a non-dominated sorting genetic algorithm of the performance estimation model; The neural architecture of the performance estimation model is evaluated and optimized based on the non-dominated sorting genetic algorithm parameters to determine the optimal architecture parameters of the neural architecture of the performance estimation model.

[0013] According to a multi-objective gradient optimization method for converter circuit parameters provided by the present invention, an encoding method of a neural architecture of the performance estimation model is defined, the neural architecture is initialized, and algorithm parameters of a non-dominated sorting genetic algorithm of the neural architecture are determined, including: Define the chromosome structure in a list format, determine the architecture parameters of the neural architecture corresponding to the chromosome gene loci, and set the value range and type; the architecture parameters include the number of layers, the number of neurons per layer, the activation function type, and the batch processing size; Randomly generating neural architecture chromosomes of an initial population to initialize the population of the non-dominated sorting genetic algorithm; Determine the optimization target of the performance estimation model, and determine the algorithm parameters of the non-dominated sorting genetic algorithm for optimizing the performance estimation model in combination with the optimization target; the algorithm parameters include population size, number of iterations, crossover probability, mutation probability and reference point setting.

[0014] According to a multi-objective gradient optimization method for converter circuit parameters provided by the present invention, the neural architecture of the performance estimation model is evaluated and optimized based on the non-dominated sorting genetic algorithm parameters to determine the optimal architecture parameters of the neural architecture of the performance estimation model, including: Decoding each neural architecture chromosome in the population of the non-dominated sorting genetic algorithm, and training the performance estimation model based on the training set and the test set; Evaluating the model performance of the performance estimation model based on the fitness function, and performing iterative optimization based on the non-dominated sorting genetic algorithm until an iterative termination condition is met to generate a final population; A Pareto front solution set in the final population is obtained; each individual in the Pareto front solution set represents a neural architecture of a non-dominated performance evaluation model.

[0015] According to a multi-objective gradient optimization method for converter circuit parameters provided by the present invention, determining a restricted range of the converter design parameters and generating optimal design parameters within the restricted range using the performance estimation model includes: Determining the design parameter limits and performance indicators of the converter; defining an encoding method for design parameters of the converter and a gradient optimization direction for performance indicators of the converter; A design parameter combination is obtained, and an index value of a performance index corresponding to the design parameter combination is predicted by the performance estimation model to determine the optimal design parameters of the converter.

[0016] According to a multi-objective gradient optimization method for converter circuit parameters provided by the present invention, a coding method for defining the converter design parameters is defined, including: The chromosome structure is defined in a list format, with each gene locus corresponding to a set of design parameter combinations; Using the performance index predicted in the performance estimation model as the objective function of the non-dominated sorting genetic algorithm; Randomly generate chromosome structures of design parameter combinations of the initial population.

[0017] According to a multi-objective gradient optimization method for converter circuit parameters provided by the present invention, a design parameter combination is obtained, and the index value of the performance index corresponding to the design parameter combination is predicted by the performance estimation model to determine the optimal design parameters of the converter, including: Decoding the chromosome structure, or lacking the corresponding design parameter combination; Predicting an index value of a performance index corresponding to the design parameter combination based on the performance estimation model; Determining a function value of a fitness function of the indicator value; Selection, crossover and mutation are performed to generate the chromosome structure of the design parameter combination of the next generation population until the iteration termination condition is met and the optimal design parameters are obtained.

[0018] In a second aspect, the present invention further provides a device for multi-objective gradient optimization of converter circuit parameters, comprising: A construction module, configured to determine performance indicators of the converter and model the performance indicators of the converter; An acquisition module, configured to acquire possible design parameters and corresponding performance indicators of the converter, and construct input space samples and output space samples; an optimization module, configured to construct a performance estimation model, and optimize a neural architecture of the performance estimation model by combining the input space samples and the output space samples through a non-dominated sorting genetic algorithm; The determination module is configured to determine a limit range of the design parameters of the converter and generate optimal design parameters within the limit range using the performance estimation model.

[0019] In a third aspect, the present invention also provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the multi-objective gradient optimization method for converter circuit parameters as described in the first aspect above is implemented.

[0020] In a fourth aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the multi-objective gradient optimization method for converter circuit parameters as described in the first aspect above.

[0021] In a fifth aspect, the present invention further provides a computer program product, comprising a computer program, which, when executed by a processor, implements the multi-objective gradient optimization method for converter circuit parameters as described in the first aspect above.

[0022] Compared with the prior art, the present invention has the following beneficial effects: The multi-objective gradient optimization method for converter circuit parameters provided by the present invention overcomes the defect that converter intelligent modeling relies on expert experience by constructing a performance estimation model mapping converter performance indicators and design parameters, and optimizing it through the neural architecture search technology of the non-dominated sorting genetic algorithm. Moreover, the process does not require frequent debugging of the network architecture and hyperparameters, and can realize the automation of model construction without any human intervention, significantly improving modeling efficiency and model accuracy, having good versatility, and solving the problem that the existing converter circuit parameter design method is difficult to obtain the optimal parameter combination. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0024] Figure 1 This is a flow chart of the multi-objective gradient optimization method for converter circuit parameters provided by the present invention; Figure 2 4L-ANPC inverter topology according to an embodiment of the present invention; Figure 3 is a schematic diagram of a fully automatic modeling process of a power converter performance estimation model according to an embodiment of the present invention; Figure 4 Schematic diagram of the performance indicator gradient optimization rule in an embodiment of the present invention; Figure 5 This is a schematic diagram of the NSGA-III search network architecture process in an embodiment of the present invention; Figure 6 This is a structural block diagram of the converter circuit parameter multi-objective gradient optimization device provided by the present invention; Figure 7 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0025] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0026] The present invention provides a multi-objective gradient optimization method for converter circuit parameters. Figure 1 Flowchart of the multi-objective gradient optimization method for converter circuit parameters provided by the present invention. Figure 1 As shown, the method includes the following steps: Step S101, determining performance indicators of the converter and modeling the performance indicators of the converter; Step S102, obtaining possible design parameters of the converter and corresponding performance indicators, and constructing input space samples and output space samples; Step S103: constructing a performance estimation model, combining the input space samples and the output space samples, and optimizing the neural architecture of the performance estimation model through a non-dominated sorting genetic algorithm; Step S104 : determining the restricted range of the design parameters of the converter, and generating the optimal design parameters within the restricted range by using the performance estimation model.

[0027] In this method, the performance indicators of the converter to be designed are first determined, modeled, and the representation of these indicators is determined. Next, possible converter design parameters and their corresponding performance indicators are obtained. These design parameters and performance indicators are used as input space samples and output space samples, respectively, for subsequent model training. A performance estimation model is then constructed, using the input space samples as input and the output space samples as labels. The neural architecture of the performance estimation model is optimized using a non-dominated sorting genetic algorithm (NSGA-III). The non-dominated sorting genetic algorithm uses the non-dominated sorting genetic algorithm III (NSGA-III). Finally, based on the constraints on the converter design parameters, the performance estimation model is used to search for the optimal design parameters within this constrained range. In the above process, a performance estimation model mapping the converter performance indicators and design parameters is constructed, and optimized through the neural architecture search technology of the non-dominated sorting genetic algorithm, which overcomes the defect that the converter intelligent modeling relies on expert experience. Moreover, this process does not require frequent debugging of the network architecture and hyperparameters, and can realize the automation of model construction without any human intervention, significantly improving the modeling efficiency and model accuracy, with good versatility, and solving the problem that the existing converter circuit parameter design method is difficult to obtain the optimal parameter combination.

[0028] The following is an example of the 4L-ANPC inverter topology. Figure 2 Schematic diagram of the topology of the 4L-ANPC inverter in an embodiment of the present invention. Figure 2 As shown in the figure, this topology has many design parameters and the multi-objective optimization design is difficult, which can fully verify the effectiveness and superiority of the method proposed in this paper in dealing with complex converter design problems. The design parameters include capacitance (C bus1 , C bus2 , C bus3 , C a , C b and C c ), inductance (L a , L b and L c ), switching frequency (f s ). Performance indicators include conversion efficiency (E f ), total harmonic distortion (Thd), volume (V), cost (Cost), bus capacitor ripple (Ri_C bus ), output current ripple (Ri_I output) and the health of the switch tube (H_S). Detailed design specifications and goals are shown in Tables 1 and 2, respectively.

[0029] Table 14 L-ANPC inverter design parameters and specifications

[0030] Table 24L-ANPC inverter design objectives

[0031] In some embodiments, the performance indicators of the converter include conversion efficiency, total harmonic distortion, bus capacitance ripple, output current ripple, switch tube reliability, volume and cost; the design parameters of the converter include capacitance, inductance and switching frequency.

[0032] For example, the conversion efficiency of the converter is defined as: E f = (1-P loss ) / P total Among them, E f Indicates conversion efficiency, P total Indicates total power, P loss Indicates power loss, P loss The simplified calculation is: P loss = 18P loss_S +n P loss_C + 3P loss_L in, n Indicates the number of capacitors used, P loss_S Indicates the switching tube loss, P loss_L Represents the magnetic component loss, P loss_C Represents capacitor loss.

[0033] Switching tube loss P loss_S Considering the conduction loss P loss_S_on , switching loss P loss_S_sw And the conduction loss P of the anti-parallel diode loss_SD_on and reverse recovery loss P loss_SD_rec , the specific formula is as follows: P loss_S = P loss_S_on + P loss_S_sw + P loss_SD_on + P loss_SD_rec

[0034]

[0035]

[0036]

[0037] Where D is the duty cycle, V f is the conduction voltage drop, E on_nom With E off_nom They are the turn-on and turn-off energies of the switch tube under standard test, I d_nom With V ds_nom They are the on-state current and off-state voltage of the switch tube under standard test, E rec_nom , I f_nom With V f _nom They are diode recovery loss, forward current and reverse voltage under standard test. represents the phase angle, represents the on-resistance, Indicates the angle The instantaneous current of the device in position, Indicates the dead time, represents the switching frequency, Indicates that the device is at an angle Position is the function of opening and closing, Indicates the drain-source voltage drop of the device. The parameters under standard test are obtained from the switching tube data sheet.

[0038] Magnetic component P loss_L The loss takes into account the hysteresis loss P of the toroidal inductor. loss_L_Fe and eddy current loss P loss_L_Cu , the specific formula is as follows: P loss_L_Cu = I rms 2 × R Cu P loss_L_Fe ≈ a × B b × f s + c × B 2 × f s 2 B =μ×H H =

[0039] l =

[0040] in, I rms represents the effective value of the inductor current, R Cu represents the DC resistance, a 、 b and c is the core material constant, B represents the inductor flux density, f s represents the switching frequency, μ represents the magnetic permeability, H represents the magnetizing force, Indicates the number of coil turns, represents the current flowing through the inductor, Indicates the effective magnetic circuit length, Indicates the outer diameter of the core, Indicates the inner diameter of the core. Inductance-related parameters are obtained from the core data sheet. After selecting the core, the maximum number of turns and inductance that can be selected are calculated as:

[0041]

[0042] in, represents the window utilization coefficient, Indicates the cross-sectional area of the conductor. Indicates the maximum number of turns corresponding to the core; Indicates the maximum inductance value, Indicates that the toroidal core is high, Represents the inductance factor.

[0043] Capacitor loss P loss_C Considering the equivalent resistance ESR and equivalent inductance ESL of the capacitor under high-frequency operation, the calculation is:

[0044]

[0045] in, represents the loss angle tangent value, represents capacitance, Indicates the equivalent inductance factor, which is obtained from the capacitor data sheet.

[0046] The total harmonic distortion is considered as the average value of the three-phase current distortion, and the calculation formula is as follows: Thd = (Thd A + Thd B + Thd C ) / 3 Among them, Thd A Indicates the total harmonic distortion of phase A current, Thd B Indicates the total harmonic distortion of phase B current, Thd C Indicates the total harmonic distortion of phase C current.

[0047] The calculation formula of bus capacitor ripple is as follows:

[0048] in, represents the bus capacitor ripple, Indicates rated power, f m Indicates the power frequency, f m =50Hz, U bus Indicates the bus voltage.

[0049] The output current ripple is calculated as follows: Ri_ I output = U bus / (12× f s ×L) Among them, Ri_ I output Indicates the output current ripple, U bus represents the bus voltage, f s represents the switching frequency, and L represents the inductance value.

[0050] In this embodiment, the reliability of the switch tube is defined as the quotient of the junction temperature of the switch tube and the rated junction temperature, and the calculation formula is as follows: H_S = ( T j / T vj )×100% Among them, H_S represents the reliability of the switch tube, T j Indicates the junction temperature of the switching tube, T vj Indicates the rated junction temperature, which is obtained from the switch tube data sheet. The third-order RC model used is used to represent the transient thermal characteristics of the switch tube. In order to simplify the calculation, the switch tube junction temperature T is defined as j Calculated as: T j = T C + (P loss_S (t)×Z th (t))

[0051]

[0052] Among them, T CIndicates the switch case temperature, P loss_S (t) represents the power loss of the switch tube, t represents time, Z th Indicates the switching tube impedance, R thi Indicates the thermal resistance of the switching tube, Indicates the thermal capacity of the switching tube, Indicates the i The time constant of the RC branch.

[0053] In this embodiment, the calculation of volume mainly considers the volume of the designed device (the volume of the capacitor V C The volume of the inductor V L ), the volume of the inductor V L Consider the core volume V L_core and winding volume V L_wire , the calculation formula is as follows: V = n ×V C + 3×V L V C = × r 2 ×G C V L = V L_core + V L_wire

[0054] V L_wire = N × A W ×K u Where V represents the volume of the converter, n Indicates the number of capacitors used, V C Represents the volume of the capacitor, V L represents the volume of the inductor, r Represents the capacitor radius, G C Represents the capacitance height, V L_core Indicates the core volume, V L_wire represents the winding volume, Indicates the outer diameter of the core, Indicates the inner diameter of the core, Indicates that the toroidal core is high, Indicates the number of coil turns, represents the window utilization coefficient, Indicates the cross-sectional area of the conductor.

[0055] In this embodiment, the cost calculation mainly considers the cost of designing components (the cost of capacitors C Cost of inductor L), Cost C Obtained from the data table, Cost L Consider the core cost L_core Cost of winding L_wire , the calculation formula is as follows: Cost = n ×Cost C + 3×Cost L Cost L = 1×V L_core + 2×2N (G L +OD-ID) Among them, Cost represents the converter cost, n Indicates the number of capacitors used, Cost C Indicates the capacitor cost, Cost L represents the inductor cost, 1 represents the estimated cost of the core per unit volume, V L_core represents the core voltage, 2 represents the estimated cost per unit length of winding, Indicates the outer diameter of the core, Indicates the inner diameter of the core, Indicates that the toroidal core is high, Indicates the number of coil turns.

[0056] In traditional converter parameter setting methods, the establishment of an artificial neural network (ANN) model requires the data set, number of network layers, number of neurons per layer, activation function (AF), and batch size (BS). In addition to the data set, other information is often set through expert experience. In order to achieve fully automated converter modeling based on artificial neural networks, this embodiment uses the neural architecture search technology based on NSGA-III to build a fully automated modeling process for the power converter performance estimation model (PEM), such as Figure 3 As shown, Figure 3 Schematic diagram of the fully automatic modeling process of the power converter performance estimation model according to an embodiment of the present invention. Detailed description is as follows: First, the search space is defined. The search space consists of two parts: input space samples and output space samples. The input space samples are different design parameter combinations (DPC) (C bus1 、C bus2、C bus3 、C a 、C b 、C c 、L a 、L b 、L c and f s ); Specifically, based on the limited range of design parameters, a finite set of parameter combinations is generated for busbar capacitance, filter capacitance, inductance, and switching frequency. Based on performance index modeling and simulation analysis, the output space sample is the performance index (E f, Thd ,V,Cost,R i_Cbus 、R i_Ioutput and H_S).

[0057] On this basis, step S103 constructs a performance estimation model, combines the input space samples and the output space samples, and optimizes the neural architecture of the performance estimation model through a non-dominated sorting genetic algorithm, including: constructing a training set and a test set based on the input space samples and the output space samples; defining the encoding method of the neural architecture of the performance estimation model, initializing the neural architecture, and determining the algorithm parameters of the non-dominated sorting genetic algorithm of the performance estimation model; evaluating and optimizing the neural architecture of the performance estimation model based on the non-dominated sorting genetic algorithm parameters, and determining the optimal architecture parameters of the neural architecture of the performance estimation model.

[0058] In this embodiment, an encoding method of a neural architecture of a performance estimation model is defined, and the neural architecture is initialized, and algorithm parameters of a non-dominated sorting genetic algorithm of the neural architecture are determined, including: defining a chromosome structure in a list form, determining the architecture parameters of the neural architecture corresponding to the chromosome gene sites and setting the value range and type; the architecture parameters include the number of layers, the number of neurons in each layer, the activation function type and the batch processing size; randomly generating an initial population of neural architecture chromosomes, and initializing the population of a non-dominated sorting genetic algorithm; determining an optimization target of the performance estimation model, and determining the algorithm parameters of a non-dominated sorting genetic algorithm for optimizing the performance estimation model in combination with the optimization target; the algorithm parameters include population size, number of iterations, crossover probability, mutation probability and reference point setting.

[0059] For example, the input control samples and output space samples are divided into training sets and test sets in a ratio of 7:3. Then, the encoding method of the neural architecture is defined, the chromosome structure is defined in a list form, the architectural parameters of the neural architecture corresponding to the chromosome gene site are determined, and the value range and type are set. The architectural parameters of the neural architecture include the number of layers, the number of neurons in each layer, the activation function type, and the batch processing size. Then, the neural architecture chromosomes of the initial population are randomly generated. Taking the accuracy and complexity of the performance estimation model as the optimization goal, the maximum compromise between the accuracy and complexity of the performance estimation model is achieved. The specific formula is as follows:

[0060]

[0061]

[0062] in, Indicates the accuracy of the model, is the fitness index, Z is the model complexity, NL is the number of neural network layers, k nl is the number of neurons in this layer, P is the number of predicted performance indicators, M is the number of output space samples, y m,i is the true output of the i-th performance indicator sample, is the predicted output of the i-th performance indicator, is the true output mean of the mth performance indicator, Then, set the algorithm parameters of the non-dominated sorting genetic algorithm, such as population size, number of iterations, crossover probability, mutation probability, reference point setting, etc.

[0063] Based on this, the PEM fully automated modeling problem can be expressed as:

[0064] subject to:Maximum

[0065] Minimal Z Among them, NL is the number of neural network layers, k nl is the number of neurons in this layer, is the activation function, represents the batch size, represents the accuracy of the model, Z represents the complexity of the model, is a normalization factor for the maximum model complexity in the search space.

[0066] The neural architecture of the performance estimation model is evaluated and optimized based on the parameters of the non-dominated sorting genetic algorithm to determine the optimal architecture parameters of the neural architecture of the performance estimation model, including: decoding each neural architecture chromosome in the population of the non-dominated sorting genetic algorithm, and training the performance estimation model based on the training set and the test set; evaluating the model performance of the performance estimation model based on the fitness function, and iteratively optimizing based on the non-dominated sorting genetic algorithm until the iteration termination condition is met to generate the final population; obtaining the Pareto front solution set in the final population; each individual in the Pareto front solution set represents a non-dominated neural architecture of the performance evaluation model.

[0067] For example, each neural architecture chromosome in the population is decoded, and a performance estimation model is trained using the training set data. The model performance is evaluated based on the fitness function. The specific formula is as follows: F1 = R total - / Z max Among them, F1 represents the fitness value, represents the accuracy of the model, Z represents the complexity of the model, is a normalizing factor for the maximum model complexity in the search space. Furthermore, based on the fitness value of the current population, the non-dominated sorting genetic algorithm performs selection, crossover, and mutation operations to generate the next generation population and determine whether the maximum number of iterations has been reached. After the non-dominated sorting genetic algorithm iterations are completed, the Pareto front solution set in the final population is obtained. Each individual in the Pareto front solution set represents a non-dominated neural architecture, and an equilibrium point on the Pareto front is selected based on the output of the fitness function.

[0068] In some embodiments, step S104 determines the limit range of the design parameters of the converter, and generates the optimal design parameters within the limit range through a performance estimation model, including: determining the limit range and performance indicators of the design parameters of the converter, the input space samples become all possible values within the limit range of the design parameters of the converter, and the output space samples are all possible performance indicators of the converter; defining the encoding method of the design parameters of the converter, and the gradient optimization direction of the performance indicators of the converter; obtaining a design parameter combination, and predicting the index value of the performance indicator corresponding to the design parameter combination through a performance estimation model to determine the optimal design parameters of the converter.

[0069] In this embodiment, the encoding method for defining the design parameters of the converter includes: defining the chromosome structure in a list form, where each gene point corresponds to a set of design parameter combinations; using the performance indicators predicted in the performance estimation model as the objective function of the non-dominated sorting genetic algorithm; and randomly generating the chromosome structure and related parameters of the design parameter combinations of the initial population.

[0070] On this basis, the definition Figure 4 The performance index gradient optimization rule shown, Figure 4 Schematic diagram of the performance index gradient optimization rule in the embodiment of the present invention. Based on the hierarchical weighted theory, the fitness function F2 of the optimization problem is defined as:

[0071]

[0072]

[0073]

[0074] in, Represents the fitness function value, α, β, γ, ∈[0,1] is the weight distribution function, and α1>α2>α3>α4, 、 、 Represents the performance index weight constant, E f ', THD', V', Cost', Ri_C bus ',Ri_I output 'And H_S' are normalized performance indicators, and the normalization method is as follows:

[0075] in, x Represents any performance indicator sequence in the set {·}. x max and x min They represent the maximum and minimum values of the performance indicators respectively, Therefore, the optimization problem of converter parameter combination can be expressed as the following formula:

[0076] subject to:H_S: Best Health; (first priority) E f >99 %; THD <1 %; (Second priority) Ri_ C bus <5 %; Ri_ I output <5%; (third priority) V: Appropriate volume; Cost: Appropriate cost; (Fourth priority) Obtain a design parameter combination, and predict the index value of the performance index corresponding to the design parameter combination through a performance estimation model to determine the optimal design parameters of the converter, including: decoding the chromosome structure, or the corresponding design parameter combination; predicting the index value of the performance index corresponding to the design parameter combination based on the performance estimation model; determining the function value of the fitness function of the index value; performing selection, crossover and mutation processing to generate the chromosome structure of the design parameter combination of the next generation population until the iteration termination condition is met to obtain the optimal design parameters.

[0077] For example, the chromosome is first decoded to obtain a specific design parameter combination. Next, the performance estimation model predicts the corresponding performance indicator value. Finally, the fitness function value is calculated, and selection, crossover, and mutation operations are performed to generate the design parameter chromosome for the next generation population. Iterations are then determined to determine whether the termination criteria are met until the optimal design parameter combination is output, which is the optimal design parameter combination.

[0078] In order to verify the effectiveness of the above method, the automatic modeling effect of the converter was verified as follows: Based on the design parameter restrictions in Table 1, busbar capacitance, filter capacitance, inductance and switching frequency are divided into 12, 3, 10 and 20 parts respectively. =7200 sets of design parameter combinations DPCs.

[0079] Based on the performance index modeling and simulation analysis, the converter performance index (E f ,Thd, V, Cost, Ri_C bus , Ri_I output and H_S). Generate an optimized design data set. Further, the data set is divided into a training set and a test set in a ratio of 7:3. Specifically expressed as: Original dataset:

[0080] Training set:

[0081] Test set:

[0082] in, = [C bus1 C bus2 C bus3 C a C b C c L a L b L c f s ] is the modeling sample, and the other parameters are the modeling targets.

[0083] Based on the training set, NSGA-III was used to automatically build a performance estimation model. The NSGA-III initialization parameters are shown in Table 3, and the neural architecture parameter search range of the performance estimation model is shown in Table 4.

[0084] Table 3 NSGA-III initialization parameters

[0085] Table 4 Parameter search range of neural architecture

[0086] Furthermore, based on the test set, the neural architecture of the established performance estimation model is verified. The NSGA-III search network architecture process is as follows Figure 5 As shown, Figure 5 This is a schematic diagram of the NSGA-III network architecture search process in an embodiment of the present invention. Each colored dot represents the optimal solution obtained after a single iterative operation. After multiple iterations, the NSGA-III population converged on an optimal solution. The model achieved a mean absolute error of only 0.32% and a model complexity of 196. This indicates that the NSGA-III method has found the optimal neural architecture. The architectural parameters of the neural architecture are shown in Table 5.

[0087] Table 5 Architecture parameters of the optimal neural architecture

[0088] Define the design parameter constraints in Table 1 as the search space, and apply NSGA-III to search for the optimal converter design parameters DPC. The NSGA-III algorithm is initialized with the settings in Table 3, and the search results are shown in Table 6.

[0089] Table 6 Optimal design parameters and performance indicators of the converter

[0090] In summary, this paper proposes a multi-objective gradient optimization method for converter circuit parameters. First, the neural architecture search technology of NSGA-III is proposed to fully automatically construct a performance estimation model that maps converter performance indicators to design parameters. This method eliminates the need for frequent debugging of network architecture and hyperparameters and can automate model construction. Second, a gradient optimization method for converter performance indicators is proposed. Based on the established performance indicator priority, this method quickly and systematically searches for converter parameters. Furthermore, an innovative switch health optimization indicator based on switch case temperature calculation is proposed, which can effectively assess the health level of switch tubes. Finally, the correctness of the proposed method is verified based on a constructed 4L-ANPC inverter prototype. This paper provides a new example reference for the fully automated design of power converters, with good scalability. In theory, it can effectively address complex converter design scenarios with numerous design variables and design objectives and high design complexity.

[0091] The present invention also provides a multi-objective gradient optimization device for converter circuit parameters. The multi-objective gradient optimization device for converter circuit parameters provided by the present invention is described below. The multi-objective gradient optimization device for converter circuit parameters described below and the multi-objective gradient optimization method for converter circuit parameters described above can be referenced to each other. Figure 6 This is a structural block diagram of the multi-objective gradient optimization device for converter circuit parameters provided by the present invention, such as Figure 6 As shown, the device includes: A construction module 601 is used to determine the performance index of the converter and model the performance index of the converter; An acquisition module 602 is configured to acquire possible design parameters of the converter and corresponding performance indicators, and construct input space samples and output space samples; An optimization module 603 is used to construct a performance estimation model, combine input space samples and output space samples, and optimize the neural architecture of the performance estimation model through a non-dominated sorting genetic algorithm; The determination module 604 is configured to determine a limit range of the design parameters of the converter and generate optimal design parameters within the limit range using a performance estimation model.

[0092] When using this device, the construction module 601 first determines the performance indicators of the converter to be designed, models the converter's performance indicators, and determines the representation form of the performance indicators. Then, the acquisition module 602 obtains the converter's possible design parameters and corresponding performance indicators, using the possible design parameters and performance indicators as input space samples and output space samples, respectively, for subsequent model training. The optimization module 603 then constructs a performance estimation model, using the input space samples as input and the output space samples as labels, and optimizes the neural architecture of the performance estimation model using a non-dominated sorting genetic algorithm. The non-dominated sorting genetic algorithm uses the Non-dominated Sorting Genetic Algorithms-III (NSGA-III). Finally, the determination module 604 searches for the optimal design parameters within the restricted range of the converter's design parameters using the performance estimation model. In the above process, a performance estimation model mapping the converter performance indicators and design parameters is constructed, and optimized through the neural architecture search technology of the non-dominated sorting genetic algorithm, which overcomes the defect that the converter intelligent modeling relies on expert experience. Moreover, this process does not require frequent debugging of the network architecture and hyperparameters, and can realize the automation of model construction without any human intervention, significantly improving the modeling efficiency and model accuracy, with good versatility, and solving the problem that the existing converter circuit parameter design method is difficult to obtain the optimal parameter combination.

[0093] Figure 7An example of a physical structure diagram of an electronic device is shown below. Figure 7 As shown, the electronic device may include: a processor 701, a communication interface 702, a memory 703, and a communication bus 704, wherein the processor 701, the communication interface 702, and the memory 703 communicate with each other via the communication bus 704. The processor 701 may call the logic instructions in the memory 703 to execute a multi-objective gradient optimization method for converter circuit parameters, the method including: Determine the performance indicators of the converter and model the performance indicators of the converter; Obtain possible design parameters and corresponding performance indicators of the converter, and construct input space samples and output space samples; Construct a performance estimation model, combine input space samples and output space samples, and optimize the neural architecture of the performance estimation model through a non-dominated sorting genetic algorithm; The limiting range of the converter design parameters is determined, and the optimal design parameters are generated within the limiting range through a performance estimation model.

[0094] In addition, the logical instructions in the aforementioned memory 703 can be implemented in the form of a software functional unit and, when sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or the portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the various embodiments of the method of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0095] In another aspect, the present invention further provides a computer program product, which includes a computer program. The computer program may be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the multi-objective gradient optimization method for converter circuit parameters provided by the above methods. The method includes: Determine the performance indicators of the converter and model the performance indicators of the converter; Obtain possible design parameters and corresponding performance indicators of the converter, and construct input space samples and output space samples; Construct a performance estimation model, combine input space samples and output space samples, and optimize the neural architecture of the performance estimation model through a non-dominated sorting genetic algorithm; The limiting range of the converter design parameters is determined, and the optimal design parameters are generated within the limiting range through a performance estimation model.

[0096] In another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for multi-objective gradient optimization of converter circuit parameters provided by the above methods is implemented. The method includes: Determine the performance indicators of the converter and model the performance indicators of the converter; Obtain possible design parameters and corresponding performance indicators of the converter, and construct input space samples and output space samples; Construct a performance estimation model, combine input space samples and output space samples, and optimize the neural architecture of the performance estimation model through a non-dominated sorting genetic algorithm; The limiting range of the converter design parameters is determined, and the optimal design parameters are generated within the limiting range through a performance estimation model.

[0097] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0098] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A multi-objective gradient optimization method for converter circuit parameters, characterized in that: include: Determining performance indicators of the converter and modeling the performance indicators of the converter; Obtaining possible design parameters and corresponding performance indicators of the converter, and constructing input space samples and output space samples; Constructing a performance estimation model, combining the input space samples and the output space samples, and optimizing the neural architecture of the performance estimation model through a non-dominated sorting genetic algorithm; The limiting range of the design parameters of the converter is determined, and the optimal design parameters are generated within the limiting range by the performance estimation model.

2. The multi-objective gradient optimization method for converter circuit parameters according to claim 1, characterized in that: The performance indicators of the converter include conversion efficiency, total harmonic distortion, bus capacitance ripple, output current ripple, switch tube reliability, volume and cost; The design parameters of the converter include capacitance, inductance, and switching frequency.

3. The multi-objective gradient optimization method for converter circuit parameters according to claim 1, characterized in that: The input space samples are different combinations of the design parameters, and the output space samples are performance indicators corresponding to the design parameter combinations.

4. The multi-objective gradient optimization method for converter circuit parameters according to claim 1, characterized in that: Constructing a performance estimation model, combining the input space samples and the output space samples, and optimizing the neural architecture of the performance estimation model through a non-dominated sorting genetic algorithm, including: Constructing a training set and a test set based on the input space samples and the output space samples; defining an encoding method of a neural architecture of the performance estimation model, initializing the neural architecture, and determining algorithm parameters of a non-dominated sorting genetic algorithm of the performance estimation model; The neural architecture of the performance estimation model is evaluated and optimized based on the non-dominated sorting genetic algorithm parameters to determine the optimal architecture parameters of the neural architecture of the performance estimation model.

5. The multi-objective gradient optimization method for converter circuit parameters according to claim 4, characterized in that: Defining an encoding method of a neural architecture of the performance estimation model, initializing the neural architecture, and determining algorithm parameters of a non-dominated sorting genetic algorithm of the neural architecture, including: Define the chromosome structure in a list format, determine the architecture parameters of the neural architecture corresponding to the chromosome gene loci, and set the value range and type; the architecture parameters include the number of layers, the number of neurons per layer, the activation function type, and the batch processing size; Randomly generating neural architecture chromosomes of an initial population to initialize the population of the non-dominated sorting genetic algorithm; Determine the optimization target of the performance estimation model, and determine the algorithm parameters of the non-dominated sorting genetic algorithm for optimizing the performance estimation model in combination with the optimization target; the algorithm parameters include population size, number of iterations, crossover probability, mutation probability and reference point setting.

6. The multi-objective gradient optimization method for converter circuit parameters according to claim 4, characterized in that: Evaluating and optimizing the neural architecture of the performance estimation model based on the non-dominated sorting genetic algorithm parameters to determine optimal architecture parameters of the neural architecture of the performance estimation model includes: Decoding each neural architecture chromosome in the population of the non-dominated sorting genetic algorithm, and training the performance estimation model based on the training set and the test set; Evaluating the model performance of the performance estimation model based on the fitness function, and performing iterative optimization based on the non-dominated sorting genetic algorithm until an iterative termination condition is met to generate a final population; A Pareto front solution set in the final population is obtained; each individual in the Pareto front solution set represents a neural architecture of a non-dominated performance evaluation model.

7. The multi-objective gradient optimization method for converter circuit parameters according to claim 1, characterized in that: Determining a limit range of design parameters of the converter, and generating optimal design parameters within the limit range using the performance estimation model, comprises: Determining the design parameter limits and performance indicators of the converter; defining an encoding method for design parameters of the converter and a gradient optimization direction for performance indicators of the converter; A design parameter combination is obtained, and an index value of a performance index corresponding to the design parameter combination is predicted by the performance estimation model to determine the optimal design parameters of the converter.

8. The multi-objective gradient optimization method for converter circuit parameters according to claim 7, characterized in that: The encoding method for defining the design parameters of the converter includes: The chromosome structure is defined in a list format, with each gene locus corresponding to a set of design parameter combinations; Using the performance index predicted in the performance estimation model as the objective function of the non-dominated sorting genetic algorithm; Randomly generate chromosome structures of design parameter combinations of the initial population.

9. The multi-objective gradient optimization method for converter circuit parameters according to claim 7, characterized in that: Obtaining a design parameter combination, and predicting the index value of the performance index corresponding to the design parameter combination by the performance estimation model to determine the optimal design parameters of the converter, including: Decoding the chromosome structure, or lacking the corresponding design parameter combination; Predicting an index value of a performance index corresponding to the design parameter combination based on the performance estimation model; Determining a function value of a fitness function of the indicator value; Selection, crossover and mutation are performed to generate the chromosome structure of the design parameter combination of the next generation population until the iteration termination condition is met and the optimal design parameters are obtained.

10. A multi-objective gradient optimization device for converter circuit parameters, characterized in that: include: A construction module, configured to determine performance indicators of the converter and model the performance indicators of the converter; An acquisition module, configured to acquire possible design parameters and corresponding performance indicators of the converter, and construct input space samples and output space samples; an optimization module, configured to construct a performance estimation model, and optimize a neural architecture of the performance estimation model by combining the input space samples and the output space samples through a non-dominated sorting genetic algorithm; The determination module is configured to determine a limit range of the design parameters of the converter and generate optimal design parameters within the limit range using the performance estimation model.

Images