A multi-objective rapid iterative optimization design method for IGBT modules of wind turbines

Abstract

The present invention discloses a multi-objective rapid iteration optimization design method for IGBT modules of wind turbines. First, based on the actual processing, manufacturing and use requirements of IGBT modules of large wind turbines, a multi-objective performance improvement optimization design model is established, and the rapid iteration of the performance response model under different design variables is realized by combining parametric modeling; then, based on the efficient proxy modeling method of self-learning adaptive sampling strategy, the performance response of the multi-objective optimization design model is accurately evaluated, and a high-precision and low-cost design proxy model is constructed; then, based on the genetic algorithm of adaptive bandwidth kernel density estimation, an adaptive bandwidth kernel density estimation genetic algorithm assisted by a proxy model is introduced to quickly and accurately iterate and solve the Pareto solution set of the multi-objective optimization design model. The present invention significantly improves the use performance of the IGBT module of the wind turbine, ensures its reliability and economy, and effectively improves the efficiency and accuracy of the multi-objective optimization design.

Description

Technical Field

[0001] The present invention belongs to the technical field of renewable energy power generation, and in particular relates to a multi-objective rapid iterative optimization design method for an IGBT module of a wind turbine generator set. Background Art

[0002] With the booming development of the global wind power market, the unit capacity of wind turbines continues to increase. In large wind turbines, IGBT (Insulate-Gate Bipolar Transistor) modules are the core power components of the unit's power generation, and their performance directly affects the safe and stable operation of the unit. However, the current huge number of wind turbines, coupled with the fact that a considerable number of offshore wind turbines are in harsh service environments with high humidity, high salt, and high corrosion, makes IGBT modules often face complex service conditions such as sudden load changes, overcurrent, overvoltage, and temperature changes, which seriously affects their service life. Therefore, achieving high-performance rapid iterative optimization design of IGBT modules for large wind turbines faces the following challenges:

[0003] (1) Large-capacity wind turbines have put forward more stringent design requirements for the performance of IGBT modules, including high reliability, high power density, low manufacturing cost, long life, etc. However, existing multi-objective performance optimization models for IGBT modules often fail to fully consider the actual design and manufacturing requirements, and have difficulties in balancing multiple complex performance indicators.

[0004] (2) Accurate prediction of the performance of existing large-scale wind turbine IGBT modules often relies on expensive simulations or experiments, combined with proxy model technology to improve design efficiency. However, traditional proxy modeling methods are difficult to achieve an accurate balance between low computational cost and high prediction accuracy, which seriously affects the efficiency and accuracy of optimization design.

[0005] (3) With the rapid iteration of IGBT module design for large wind turbines, the existing multi-objective design methods based on time-consuming performance calculations are difficult to adapt to the needs of rapid iterative design and have poor adaptability. The multi-objective optimization method based on genetic algorithms is difficult to accurately solve the Pareto solution set in high-dimensional engineering design problems. The calculation process is often difficult to converge and the solution set accuracy is low. Summary of the invention

[0006] In order to solve the deficiencies of the prior art and achieve rapid and accurate solution of the multi-objective optimization model, thereby improving the performance and reliability of the IGBT module of the wind turbine, the present invention adopts the following technical solution:

[0007] A multi-objective rapid iterative optimization design method for an IGBT module of a wind turbine generator system comprises the following steps:

[0008] Step S1: Determine the objective function F and the design parameter x that is sensitive to the objective function F through the target to be optimized, determine the constraint function C based on the design requirements, and construct a multi-objective optimization design model M for the IGBT module of the wind turbine generator through the objective function F, the design variable x, and the constraint function C;

[0009] Step S2: Based on the multi-objective optimization design model M of the IGBT module of the wind turbine generator, a parameterized model P between the sensitive design variable x and different objective functions F and constraint functions C is established by using the parametric modeling technology to capture the influence of the design variable x on the key performance indicators, and to perform automated rapid iterative calculations of the performance response model under different design variables x, thereby accelerating the optimization design process;

[0010] Step S3: Based on the multi-objective optimization design model M and the parameterized model P, generate initial design points in the design space, and calculate the values ​​of the constraint function C and the objective function F at different design points respectively, complete the construction of the initial sample set, and establish the initial Kriging proxy model G; use the proposed self-learning adaptive fast iterative sampling strategy to obtain new design sample points in the design space; call the parameterized model P to complete the performance calculation of the new design sample points, and continuously update and iterate the current Kriging proxy model G through the self-learning adaptive fast iterative sampling strategy, and judge whether the proxy model G meets the design accuracy requirements according to the convergence index. If not, repeat the above steps, and repeat until convergence, to complete the construction of a high-precision and low-cost proxy model between the design variables and different objective functions or constraint functions;

[0011] Step S4: setting optimization parameters for solving the multi-objective optimization design model M using a genetic algorithm; solving the optimal Pareto design of the multi-objective optimization design model M using a genetic algorithm based on adaptive bandwidth kernel density estimation; judging whether the optimal Pareto design meets the requirements until the stop condition is met, so as to complete the multi-objective rapid iterative optimization design of the IGBT module of a large wind turbine.

[0012] Furthermore, the step S1 comprises the following steps:

[0013] Step S1.1: Based on the typical usage scenarios, processing and manufacturing costs, and packaging structure damage of key components of IGBT modules of large wind turbines, the busbar, a vulnerable component of the IGBT module, is taken as the research object, and three optimization targets are determined, namely, conductive characteristics, manufacturing costs, and packaging fatigue damage, so as to establish the objective function F for performance improvement of the multi-objective optimization design process of the IGBT module;

[0014] Step S1.2: Screen the key geometric design parameters of the IGBT module structure, observe the impact on the three objective functions by changing the specific design variables of the IGBT module, and use the global sensitivity variance analysis method to identify the three structural design variables x that are highly sensitive to the objective function, including pin height, rib length and busbar thickness;

[0015] Step S1.3: Based on the design standard and safety margin of the IGBT module, the maximum temperature, maximum stress and fatigue damage in the packaging process are determined as the constraint function C for improving the performance of the multi-objective optimization design model;

[0016] Step S1.4: Determine the mathematical model of the multi-objective performance improvement optimization design of the IGBT module of a large wind turbine generator set by combining the objective function F, the design variable x, and the constraint function C; in the step S1.4, the expression of the mathematical model is as follows:

[0017]

[0018] Among them, F(x) represents the objective function, f1(x), f2(x) and f3(x) represent the multiple objective functions of the conductive loss, manufacturing cost and package fatigue damage of the IGBT module components, respectively, and x, x L With x U They represent the selected design variables and the upper and lower boundary conditions of the design variables, respectively. max and [T] represent the maximum temperature and temperature threshold to which the structure is subjected during the packaging process, respectively. and[ ] represent the maximum stress and stress threshold of the structure during the packaging process respectively.

[0019] Furthermore, in step S2, the construction of a parameterized model P between the sensitive design variable x and different objective functions F and constraint functions C includes the following steps:

[0020] Step S2.1: construct a geometric model of the IGBT module, and bind the sensitive design variable x to the design parameters of the actual geometric model, and use the developed parametric geometric modeling module to realize geometric modeling under different design variable sizes;

[0021] Step S2.2: Perform finite element structural simulation on the parameterized geometric model using electrothermal coupling to calculate the conductive loss f1(x) of the objective function under the parameterized design variable x;

[0022] Step S2.3: Measure the volume V of the parameterized geometric model to accurately calculate the manufacturing cost f2(x) of the IGBT module under different design variables x. The calculation expression is as follows:

[0023]

[0024] in, represents the density of IGBT module components, V represents the volume of the geometric model under different design variables, and z represents the manufacturing cost per gram of IGBT raw materials;

[0025] Step S2.4: The different components of the IGBT module are packaged using ultrasonic welding technology. The ultrasonic-thermal-mechanical multi-physics field coupling simulation is used to perform high-fidelity simulation on the parameterized geometric model of the packaging process. The simulation time is set to 0.5 s. The simulation stress results are post-processed to generate a constraint sequence. The accumulated fatigue damage value f3(x) under different constraint sequences is calculated by combining the rain flow counting method and the Miner criterion.

[0026] Step S2.5: Constraint condition of the constraint function during packaging process Maximum temperature T max With maximum stress They are obtained by traversing all nodes and units of the coupled simulation model in step S2.4 and sorting them.

[0027] Furthermore, the self-learning adaptive fast iterative sampling strategy in step 3 specifically includes the following steps:

[0028] Step S3.1: performing Thiessen polygon region division on the generated initial design points so that each region contains only one existing initial design point;

[0029] Step S3.2: Estimating the prediction error of the Kriging proxy model for all existing initial design points;

[0030] Step S3.3: Select the design point with the largest prediction error as the sensitive design point, and screen the new sample points within the Thiessen polygon area where the current sensitive design point is located;

[0031] Step S3.4: Generate a large number of candidate sample points in the entire design space by using the Monte Carlo method;

[0032] Step S3.5: Combine the sensitive samples determined in step S3.3 with the candidate sample points controlled in step S3.4, and select candidate sample points that are relatively far away from the existing sample points, have large prediction errors, and have large Kriging prediction variance as new sample points x new .

[0033] Furthermore, in step S3.2, a new fast cross-validation error calculation method is designed for error estimation, and the calculation formula is as follows:

[0034]

[0035] Among them, (ii), (i* )and( * i) represent the i-th diagonal element, i-th row and i-th column of the matrix respectively. Vector 1 consists of m components, each of which is equal to 1. R represents the correlation matrix, whose calculation depends on the correlation function R containing the hyperparameter θ of the kriging model. θ (∙), two samples x and x * The correlation function R θ (x, x * , θ), D and H are calculated as follows:

[0036]

[0037]

[0038]

[0039]

[0040] Among them, x s , and θ s Respectively represent x and x * With the sth element of θ, n represents the dimension of x. For m existing samples X and output responses Y, an m×m correlation matrix R will be formed. The optimal hyperparameter Solved by the least squares method, are the regression coefficients of the Kriging proxy model estimated using the least squares method.

[0041] The larger the predicted cross-validation error value, the less information available around the existing samples, and new samples should be added in the corresponding area.

[0042] Furthermore, in step S3.5, the new sample point is calculated as follows:

[0043]

[0044] in, ,i=1,2,…,n c represents the candidate sample point, d represents and k+1 adjacent existing samples x j The spatial distance between It represents the prediction error of the proxy model of the existing sample points adjacent to the current sensitive point, and s represents the prediction variance of the proxy model at any sample point (including existing sample points and candidate sample points).

[0045] Furthermore, the convergence index of the three models is to stop sampling when the error of adding three new sample points in succession is less than the set error threshold of each objective function and constraint function proxy model when constructing the proxy model by self-learning adaptive sampling. The purpose of this setting is to avoid accidental exploration of sampling points and make the proxy model more robust.

[0046] Furthermore, the genetic algorithm based on adaptive bandwidth kernel density estimation in step S4 comprises the following steps:

[0047] Step S4.1: Decode the individual;

[0048] Decoding the individual specifically involves using the proxy model in step S3 to update the constraint function and objective function values ​​of the multi-objective optimization design model M during the optimization iteration solution process, thereby realizing rapid calculation of the performance response during the optimization iteration process.

[0049] Step S4.2: Use the adaptive bandwidth kernel density estimation method to prioritize the fitness of the population, instead of the individual priority comparison based on crowding distance in the original NSGA-II algorithm, to avoid the problem of inaccurate representation of crowding distance of design points at the boundary, and define the priority as the individual kernel density estimate The reciprocal of The calculation of is as follows:

[0050]

[0051] Among them, n p represents the design points on the same non-dominated frontier The number of, h represents the bandwidth, the kernel function K(∙) is calculated as follows:

[0052]

[0053] in, is an intermediate variable, which represents the input of the kernel function; for different design points, The smaller the value, the lower the congestion level near the point, the scarcer the design point, and the higher the priority, which is defined as:

[0054]

[0055]

[0056] in, Indicates the design point by dominate, It has a higher priority and should be retained during the iteration process. and They represent the non-dominated levels where the design points are located;

[0057] Step S4.3: Perform crossover mutation on the population individuals with high priority in step S4.2, and repeat steps S4.1 and S4.2 until the optimal Pareto design is found.

[0058] Furthermore, in step S4.2, when performing kernel density estimation to calculate individual kernel density estimates, an adaptive bandwidth calculation method is used to avoid the problem of inaccurate density estimation caused by the traditional use of fixed bandwidth. Specifically, for different design points at the same non-dominated level, the region partitioning method of step S4.1 is used to divide the objective function space where the design point is located into multiple Thiessen polygons. The design points with common edges of Thiessen polygons are defined as the neighborhood of the design point. The number of neighbors corresponding to the design point is defined as the bandwidth when estimating the density value of the point. The bandwidth of different design points is repeatedly calculated in turn, and their density values ​​are calculated.

[0059] Furthermore, the stopping condition in step S4 is to perform parameterized modeling on the currently calculated optimal Pareto design parameters, and compare the calculated objective function and constraint function with the predicted value of the optimized proxy model. When the error values ​​of all objective functions and constraint functions are less than the error threshold, the design requirements are met and the multi-objective rapid iterative optimization design of the current large wind turbine IGBT module is completed.

[0060] The advantages and beneficial effects of the present invention are:

[0061] (1) The present invention establishes an accurate multi-objective performance improvement optimization design model for IGBT modules of large wind turbine power conversion devices, thereby being able to systematically consider multiple design objectives and constraints, thereby improving the pertinence and effectiveness of the design;

[0062] (2) The present invention combines parametric modeling technology and uses parametric models to automatically and quickly iterate calculations, so that the model can flexibly adjust and optimize design parameters, improve the accuracy of capturing performance responses of design variables, and improve the adaptability and adjustability of the design model, so as to better cope with complex requirements and changes in actual engineering;

[0063] (3) The present invention adopts a new efficient proxy modeling method based on self-learning adaptive sampling strategy. Through fast cross-validation error and dynamic selection of new sample points, a low-cost and high-precision proxy model is constructed, which improves the computational efficiency and optimization accuracy of the proxy model and makes the iteration process faster and more efficient.

[0064] (4) The present invention proposes an advanced genetic algorithm technology based on adaptive bandwidth kernel density estimation, which enables the optimization process to better handle complex design spaces and multi-objective optimization problems, thereby improving the solution speed and optimization quality;

[0065] (5) Under the premise of ensuring design accuracy, the present invention realizes the rapid iterative design of IGBT modules of large wind turbines, which not only improves the conductive performance and reliability, but also effectively reduces the design and manufacturing costs, and improves the overall comprehensive performance and economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 Flow chart of the method in the embodiment of the present invention.

[0067] Figure 2 It is a schematic diagram of the structure of the IGBT module components of a large wind turbine generator system in an embodiment of the present invention.

[0068] Figure 3a This is one of the schematic diagrams of design variables of the IGBT module of a large wind turbine in an embodiment of the present invention.

[0069] Figure 3b This is the second schematic diagram of design variables of the IGBT module of a large wind turbine in an embodiment of the present invention.

[0070] Figure 4 It is a schematic diagram of calculating the conductive performance of the IGBT module of a large wind turbine in an embodiment of the present invention.

[0071] Figure 5 This is a diagram showing a finite element simulation analysis model for fatigue damage, stress and temperature calculations at the packaging stage of an IGBT module of a large wind turbine generator in an embodiment of the present invention.

[0072] Figure 6 It is a historical convergence curve diagram of the iterative sampling process of the fatigue damage proxy model of the IGBT module package of a large wind turbine in an embodiment of the present invention.

[0073] Figure 7 This is a spatial distribution diagram of the Pareto solution set for multi-objective iterative optimization design of IGBT modules of large wind turbines in an embodiment of the present invention. DETAILED DESCRIPTION

[0074] The specific implementation of the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the present invention, and is not used to limit the present invention.

[0075] A multi-objective rapid iterative optimization design method for IGBT modules of large wind turbines. Firstly, according to the actual processing, manufacturing and use requirements of the core power IGBT modules of large wind turbines, a multi-objective performance improvement optimization design model is established based on the global sensitivity variance analysis method, and combined with the parametric modeling technology, the rapid iterative calculation of the performance response model under different design variable conditions is realized; next, an efficient agent modeling method based on self-learning adaptive sampling strategy is proposed to accurately evaluate the performance response of the multi-objective optimization design model and construct a high-precision and low-cost design agent model; then, a genetic algorithm based on adaptive bandwidth kernel density estimation is innovatively proposed, and combined with the agent model-assisted adaptive bandwidth kernel density estimation technology, it is used to quickly and accurately solve the Pareto solution set of the multi-objective optimization design model, and realizes the rapid iterative optimization design of the IGBT module of the wind turbine. Figure 1 As shown, the method specifically comprises the following steps:

[0076] Step S1: According to the actual working requirements, iterative design requirements and processing and manufacturing complexity of the IGBT module of a large wind turbine, key components, vulnerable stages, sensitive design variables and key performance indicators are selected from the numerous components, design parameters, performance indicators and manufacturing stages of the IGBT module, and a comprehensive and accurate IGBT module multi-objective optimization design model M is constructed;

[0077] The construction steps of the multi-objective optimization design model M of the wind turbine IGBT module are as follows:

[0078] Step S1.1: Comprehensively consider the typical usage scenarios, processing and manufacturing costs, and packaging structure damage of key components of large wind turbine IGBT modules, such as Figure 2 As shown, the busbar (1) of the vulnerable component of the IGBT module is taken as the research object, the busbar (1) is arranged on the substrate (4) through the DCB liner (3), and the DCB liner (3) is provided with a probe (2), and three objectives of conductive characteristics, manufacturing cost, and package fatigue damage are determined as the objective function F for improving the performance of the multi-objective optimization design process of the IGBT module;

[0079] Step S1.2: Screen the key geometric design parameters of the IGBT module structure, observe the impact on the three objective functions by changing specific design variables, and use the global sensitivity variance analysis method to identify three structural design variables x that are highly sensitive to the objective function, including pin height, rib length and busbar thickness, such as Figure 3a , Figure 3b As shown;

[0080] Step S1.3: Based on the design standard and safety margin of the IGBT module, the maximum temperature, maximum stress and fatigue damage in the packaging process are determined as the constraint function C of the multi-objective performance improvement optimization design model;

[0081] Step S1.4: Combining steps S1.1, S1.2 and S1.3, the mathematical model expression for the multi-objective performance improvement optimization design of the IGBT module of a large wind turbine is determined as follows:

[0082]

[0083] Where F(x) represents the objective function, f1(x), f2(x) and f3(x) represent the three objective functions of the IGBT module busbar (1) conduction loss, manufacturing cost and package fatigue damage, respectively, and x, x L With x U They represent the selected design variables and the upper and lower boundary conditions of the design variables, respectively. max [T] represents the maximum temperature and temperature threshold to which the structure is subjected during the packaging process. and[ ] represents the maximum stress and stress threshold to which the structure is subjected during the packaging process.

[0084] Step S2: Based on the multi-objective optimization design model M of the IGBT module in step S1, a parametric model P between sensitive design variables and different objective functions F and constraint functions C is established by using parametric modeling technology, so as to effectively capture the influence of design variables on key performance indicators, and realize automatic and rapid iterative calculation of performance response models under different design variables, thereby accelerating the optimization design process;

[0085] like Figures 4 to 6 As shown, U0 and U1 represent voltage values, △U represents conduction loss, and the construction steps of the parameterized model P between the sensitive design variable x and different objective functions F and constraint functions C are as follows:

[0086] Step S2.1: construct a geometric model of the IGBT module, and bind the sensitive design variable x to the design size of the actual geometric model, and use the developed parametric geometric modeling module to realize geometric modeling under different design variable sizes;

[0087] Step S2.2: Perform finite element structural simulation on the parameterized geometric model using electrothermal coupling technology to calculate the conductive loss f1(x) of the objective function under the parameterized design variable x;

[0088] Step S2.3: Measure the volume V of the parameterized geometric model to accurately calculate the manufacturing cost f2(x) of the IGBT module under different design variables x. The calculation expression is as follows:

[0089]

[0090] in, represents the density of the IGBT module busbar (1), V represents the volume of the geometric model under different design variables, and z represents the manufacturing cost per gram of IGBT raw material;

[0091] Step S2.4: The different components of the IGBT module are packaged using ultrasonic welding technology, and the parametric geometric model of this process is simulated with high fidelity using ultrasonic-thermal-mechanical multi-physics field coupling simulation technology. The simulation time is set to 0.5 s, and the simulation stress results are post-processed to generate a stress sequence. The accumulated fatigue damage value f3(x) under different stress sequences is calculated by combining the rain flow counting method and the Miner criterion;

[0092] Step S2.5: Maximum temperature T of constraint function C during packaging max With maximum stress They are all obtained through post-processing calculation of the data of the coupled simulation model in step S2.4, specifically, traversing all nodes and units of the finite element simulation model and sorting them to obtain the maximum temperature and stress.

[0093] Step S3: Based on the multi-objective optimization design model M in step S1 and the parameterized model P in step S2, generate initial design points in the design space, and calculate the values ​​of the constraint function C and the objective function F at different design points respectively, complete the construction of the initial sample set, and establish the initial Kriging proxy model G; use the proposed self-learning adaptive fast iterative sampling strategy to obtain new design sample points in the design space; call the parameterized model P to complete the performance calculation of the new design sample points, and continuously update and iterate the current Kriging proxy model G through the self-learning adaptive fast iterative sampling strategy; judge whether the proxy model G meets the design accuracy requirements according to the convergence index, if not, repeat the above steps, and repeat until the construction of a high-precision and low-cost proxy model between the design variables and different objective functions or constraint functions is completed.

[0094] The self-learning adaptive fast iterative sampling strategy includes the following steps:

[0095] Step S3.1: performing Thiessen polygon region division on the generated initial design points so that each region contains only one existing initial design point;

[0096] Step S3.2: The prediction error of the Kriging proxy model is estimated for all existing initial design points. A new fast cross-validation error calculation method is proposed. The calculation formula is as follows:

[0097]

[0098] Among them, (ii), (i * )and( * i) represents the i-th diagonal element, i-th row, and i-th column of the matrix, respectively. Vector 1 consists of m components, each of which is equal to 1. R is the correlation matrix, whose calculation depends on the correlation function R containing the hyperparameters θ of the kriging model θ (∙). In addition, two samples x and x * The correlation function R θ (x, x * , θ), D and H are calculated as follows:

[0099]

[0100]

[0101]

[0102]

[0103] Among them, x s , and θ s They are x and x respectively. * With the sth element of θ, n is the dimension of x. For m existing samples X and output responses Y, an m×m correlation matrix R will be constructed. In addition, the optimal hyperparameters The solution is obtained by the least squares method. are the regression coefficients of the Kriging proxy model estimated using the least squares method.

[0104] The larger the predicted cross-validation error value, the less information available around the existing samples, and new samples should be added in the corresponding area.

[0105] Step S3.3: Select the design point with the largest prediction error as the sensitive design point, and screen the new sample points within the Thiessen polygon area where the current sensitive design point is located;

[0106] Step S3.4: Generate a large number of candidate sample points in the entire design space by using the Monte Carlo method ,i=1,2,…,n c ;

[0107] Step S3.5: Combine the sensitive samples determined in step S3.3 with the candidate sample points controlled in step S3.4, and select candidate sample points that are relatively far away from the existing sample points, have larger prediction errors, and have larger Kriging prediction variance as new sample points x new , calculated as follows:

[0108]

[0109] where d is and k+1 adjacent existing samples x j The spatial distance between.

[0110] The convergence index of the proxy model is to stop sampling when three new sample points are added continuously during the self-learning adaptive sampling to construct the proxy model, and all of them are less than the set error threshold of the proxy model of each objective function and constraint function. The purpose of this setting is to avoid accidental exploration of sampling points and make the proxy model more robust.

[0111] Step S4: Set the optimization parameters for solving the multi-objective optimization design model M using the genetic algorithm; the optimization parameters specifically include: determining the individuals and populations of the optimization process, set to 100, the number of iterations to 10000, the objective function dimension to 3, and the design variable dimension to 3. The proposed genetic algorithm based on adaptive bandwidth kernel density estimation is used to solve the optimal Pareto design of the multi-objective optimization design model M, such as Figure 7 As shown in the figure. By setting the stop condition, it is judged whether the obtained Pareto design meets the requirements. If not, the optimal Pareto design is repeatedly solved until the stop condition is met, completing the multi-objective rapid iterative optimization design of the IGBT module of the large wind turbine.

[0112] The genetic algorithm based on adaptive bandwidth kernel density estimation has the following specific implementation steps:

[0113] Step S4.1: Decode the individual;

[0114] Decoding the individual specifically involves using the proxy model in step S6 to update the constraint function and objective function values ​​of the multi-objective optimization design model M during the optimization iteration solution process, thereby achieving rapid calculation of the performance response during the optimization iteration process.

[0115] Step S4.2: Use the proposed adaptive bandwidth kernel density estimation method to prioritize the population by fitness;

[0116] The proposed adaptive bandwidth kernel density estimation method is used to prioritize the fitness of the population, replacing the individual priority comparison based on crowding distance in the original NSGA-II algorithm, avoiding the problem of inaccurate representation of crowding distance of design points at the boundary. The priority is defined as the individual kernel density estimate The reciprocal of The calculation of is as follows:

[0117]

[0118] Among them, n p is the number of design points in the same non-dominated frontier, h represents the bandwidth, and K(∙) is the kernel function. The calculation formula of the kernel function K(∙) is as follows:

[0119]

[0120] For different design points, The smaller the value, the lower the congestion level near the point, the scarcer the design point, and the higher the priority, which is defined as:

[0121]

[0122]

[0123] Indicates the design point by dominate, The higher the priority, the more it should be retained during the iteration process. and They respectively represent the non-dominated levels of the design points.

[0124] When performing kernel density estimation to calculate individual kernel density estimates, an adaptive bandwidth calculation method is used to avoid the problem of inaccurate density estimation caused by the traditional fixed bandwidth. The details are as follows:

[0125] For different design points at the same non-dominated level, the region partitioning method of step S4.1 is used to divide the objective function space where the design points are located into multiple Thiessen polygons. The design points with common edges of Thiessen polygons are defined as the neighborhood of the design point. The number of neighbors corresponding to the design point is defined as the bandwidth when estimating the density value of the point. The bandwidth of different design points is repeatedly calculated in turn, and their density values ​​are calculated.

[0126] Step S4.3: Perform crossover mutation on the population individuals with high priority in step S4.2, and repeat steps S4.1 and S4.2 until the optimal Pareto design is found.

[0127] The stopping condition is to parameterize the Pareto design parameters currently calculated, calculate the corresponding objective function and constraint function, and compare them with the predicted value of the optimized proxy model. When the error value of all objective functions and constraint functions is less than 5%, the design requirements are met and the multi-objective rapid iterative optimization design of the IGBT module of the current large wind turbine is completed.

[0128] 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 replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-objective rapid iterative optimization design method for IGBT modules of wind turbines, characterized in that The steps include: Step S1: Determine the objective function and the design parameters sensitive to the objective function through the target to be optimized, determine the constraint function based on the design requirements, and construct a multi-objective optimization design model for the IGBT module of the wind turbine generator through the objective function, design variables and constraint function; Step S2: Based on the multi-objective optimization design model of the IGBT module of the wind turbine generator, a parameterized model between the design variables and different objective functions and constraint functions is established to capture the impact of the design variables on the key performance indicators, and to perform automatic rapid iterative calculation of the performance response model under different design variables; the construction of the parameterized model between the design variables and different objective functions and constraint functions includes the following steps: Step S2.1: construct a geometric model of the IGBT module, and bind the design variable x to the design parameters of the actual geometric model, and use the developed parametric geometric modeling module to realize geometric modeling under different design variable sizes; Step S2.2: Perform finite element structural simulation on the parameterized geometric model using electrothermal coupling to calculate the conductive loss f1(x) of the objective function under the parameterized design variable x; Step S2.3: Measure the volume V of the parameterized geometric model to calculate the manufacturing cost f2(x) of the IGBT module under different design variables x. The calculation expression is as follows: f2(x)=ρ×V×z Where ρ represents the density of IGBT module components, V represents the volume of the geometric model under different design variables, and z represents the manufacturing cost per gram of IGBT raw materials; Step S2.4: Encapsulate different components of the IGBT module, perform high-fidelity simulation on the parameterized geometric model of the encapsulation process, post-process the simulation results to generate a constraint sequence, and calculate the accumulated fatigue damage value f3(x) under different constraint sequences; Step S2.5: Constraint condition of the constraint function during packaging process Maximum temperature T max and the maximum stress σ max All of them are obtained by traversing all nodes and units of the coupled simulation model in step S2.4 and sorting them; Step S3: Based on the multi-objective optimization design model and the parameterized model, an initial design point is generated in the design space, and the values ​​of the constraint function and the objective function under different design points are calculated respectively, so as to complete the construction of the initial sample set and establish the initial Kriging proxy model; the proposed self-learning adaptive fast iterative sampling strategy is used to obtain new design sample points in the design space; the parameterized model is called to complete the performance calculation of the new design sample points, and the current Kriging proxy model is continuously updated and iterated through the self-learning adaptive fast iterative sampling strategy until convergence; the self-learning adaptive fast iterative sampling strategy specifically includes the following steps: Step S3.1: performing Thiessen polygon region division on the generated initial design points so that each region contains only one existing initial design point; Step S3.2: Estimating the prediction error of the Kriging proxy model for all existing initial design points; Step S3.3: The design point with the largest prediction error is selected as the sensitive design point, and the new sample point is screened within the Thiessen polygon area where the current sensitive design point is located; Step S3.4: Generate a large number of candidate sample points in the entire design space by using the Monte Carlo method; Step S3.5: Combining the sensitive samples determined in step S3.3 with the candidate sample points controlled in step S3.4, by selecting candidate sample points that are relatively far away from the existing sample points, have large prediction errors, and have large Kriging prediction variance as new sample points; Step S4: setting optimization parameters for solving the multi-objective optimization design model process using a genetic algorithm; solving the optimal design of the multi-objective optimization design model using a genetic algorithm based on adaptive bandwidth kernel density estimation; judging whether the optimal design meets the requirements until the stop condition is met, so as to complete the multi-objective rapid iterative optimization design of the IGBT module of the wind turbine generator; the genetic algorithm based on adaptive bandwidth kernel density estimation solves the optimal design of the multi-objective optimization design model, including the following steps: Step S4.1: Decode the individual; Step S4.2: Use the adaptive bandwidth kernel density estimation method to prioritize the fitness of the population, and define the priority as the individual kernel density estimate The reciprocal of The calculation of is as follows: Among them, n p represents the design point x in the same non-dominated frontier i The number of h represents bandwidth, and the kernel function K(·) is calculated as follows: Among them, u is an intermediate variable, which represents the input of the kernel function; for different design points, the The smaller the value, the lower the congestion level near the point, the scarcer the design point, and the higher the priority, which is defined as: in, represents the design point p j Subject to p i Domination, p i It has a higher priority and should be retained during the iteration process. and They represent the non-dominated levels where the design points are located; Step S4.3: Perform crossover mutation on the population individuals with high priority in step S4.2, and repeat steps S4.1 and S4.2 until the optimal design is found.

2. A multi-objective rapid iterative optimization design method for IGBT modules of wind turbine generators according to claim 1, characterized in that: The step S1 comprises the following steps: Step S1.1: Based on the IGBT module of the wind turbine generator set, determine the target to be optimized to establish the objective function F for improving the performance of the multi-objective optimization design process of the IGBT module; Step S1.2: By changing the specific design variables of the IGBT module, using the global sensitivity variance analysis method, identify the design variables x that are highly sensitive to the objective function; Step S1.3: Based on the design standard and safety margin of the IGBT module, determine the constraint function C for improving the performance of the multi-objective optimization design model; Step S1.4: Determine the mathematical model of the multi-objective performance improvement optimization design of the wind turbine IGBT module by combining the objective function F, the design variable x, and the constraint function C; in the step S1.4, the expression of the mathematical model is as follows: Min F(x)=(f1(x),f2(x),f3(x)) Wherein, F(x) represents the objective function, f1(x), f2(x) and f3(x) represent multiple objective functions of the IGBT module components, x, x L With x U They represent the selected design variables and the upper and lower boundary conditions of the design variables, respectively. max and [T] represent the maximum temperature and temperature threshold of the structure during the packaging process, respectively, max and [σ] represent the maximum stress and stress threshold of the structure during the packaging process, respectively.

3. A multi-objective rapid iterative optimization design method for IGBT modules of wind turbine generators according to claim 1, characterized in that: In step S3.2, a new fast cross-validation error calculation method is designed for error estimation, and the calculation formula is as follows: Among them, (ii), (i * )and( * i) represent the i-th diagonal element, i-th row and i-th column of the matrix respectively. Vector 1 consists of m components, each of which is equal to 1. R represents the correlation matrix, whose calculation depends on the correlation function R containing the hyperparameter θ of the kriging model. θ (·), two samples x and x * The correlation function R θ (x,x * ,θ), D and H are calculated as follows: H＝(1 T ×1) -1 ×1×1 T Among them, x s , and θ s Respectively represent x and x * With the sth element of θ, n represents the dimension of x. For m existing samples X and output responses Y, an m×m correlation matrix R will be formed. The optimal hyperparameter Solved by the least squares method, are the regression coefficients of the Kriging proxy model estimated using the least squares method.

4. A multi-objective rapid iterative optimization design method for IGBT modules of wind turbine generators according to claim 1, characterized in that: In step S3.5, the new sample point is calculated as follows: in, represents the candidate sample point, d represents and k+1 adjacent existing samples x j The spatial distance between represents the prediction error of the proxy model for the existing sample points adjacent to the current sensitive point, and s represents the prediction variance of the proxy model at any sample point.

5. A multi-objective rapid iterative optimization design method for IGBT modules of wind turbine generators according to claim 1, characterized in that: The convergence index of the model in step S3 is to stop sampling when the error of continuously adding new sample points is less than the set error threshold of each objective function and constraint function when constructing the proxy model for self-learning adaptive sampling.

6. A multi-objective rapid iterative optimization design method for IGBT modules of wind turbine generators according to claim 1, characterized in that: In step S4.2, for different design points at the same non-dominated level, the region partitioning method used is used to divide the objective function space where the design points are located into multiple Thiessen polygons, and the design points with common edges of the Thiessen polygons are defined as the neighborhood of the design point. The number of neighbors corresponding to the design point is defined as the bandwidth when estimating the density value of the point. The bandwidth of different design points is repeatedly calculated in sequence, and their density values ​​are calculated.

7. A multi-objective rapid iterative optimization design method for IGBT modules of wind turbine generators according to claim 1, characterized in that: The stopping condition in step S4 is to perform parameterized modeling on the currently calculated optimal design parameters, and compare the calculated objective function and constraint function with the predicted value of the optimized proxy model. When the error values ​​of all objective functions and constraint functions are less than the error threshold, the design requirements are met and the multi-objective rapid iterative optimization design of the IGBT module of the current large wind turbine is completed.

Images