Machine Learning-Based Multi-Objective Optimization Design Method and Storage Medium for Fuel Cells
By constructing a parameterized flow field model and an adaptive sparsity-weighted penalized boundary objective function, the conflict between net power density, water saturation, and oxygen distribution uniformity in the multi-objective optimization of fuel cell electrode flow field is resolved, achieving efficient and low-cost flow field structure design.
Patent Information
- Application Number
- CN202611585740.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-14
- Publication Date
- 2026-11-10
AI Technical Summary
Existing machine learning optimization techniques struggle to efficiently and robustly balance maximizing net power density, controlling water saturation, and improving oxygen distribution uniformity in the fuel cell plate flow field. Traditional methods are computationally expensive and have lengthy design cycles.
A machine learning-based multi-objective optimization design method is adopted. By constructing a parameterized flow field model, a local search is performed to optimize the flow field structure using a constraint feasibility-driven population reconstruction, a gradient information-guided multi-objective evolution mechanism, and an adaptive sparsity-weighted penalty boundary objective function.
It achieves efficient convergence to the region of optimal net power density while meeting the requirements of water saturation and oxygen distribution uniformity, improving the engineering practicality and optimization efficiency of the design, and reducing computational costs and design cycle.
Smart Images

Figure CN123069931A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of general artificial intelligence technology and swarm intelligence technology. Specifically, it relates to a multi-objective optimization design method and storage medium for fuel cells based on machine learning. Background Technology
[0002] The flow field structure of fuel cell plates is a core component for achieving efficient electrochemical reactions and precise water management, and has become an indispensable key component of vehicle fuel cells and distributed power generation systems. The quality of its flow channel topology, cross-sectional geometry, and three-dimensional micro-disturbance structure directly affects the net power density, water saturation, and oxygen distribution uniformity of the flow field. With fuel cells continuously upgrading towards higher current density, higher power density, and longer lifespan, plate flow field structure design faces significant challenges. It must achieve lightweight and manufacturable plates while ensuring high net power density, controllable water saturation, and excellent oxygen distribution uniformity.
[0003] Traditional plate flow field design relies primarily on engineers' experience and trial-and-error exploration, or on a serial optimization approach based on computational fluid dynamics simulations. It involves iterative verification throughout the entire process using complete numerical simulations, resulting in high computational costs and lengthy design cycles. With technological advancements, machine learning has become a significant breakthrough in engineering optimization. Leveraging data-driven mechanisms, it accurately captures the nonlinear relationships between design parameters and performance indicators, constructing efficient radial basis function models. This significantly reduces reliance on costly simulations and allows for rapid location of optimal solutions in complex design spaces, improving the automation level and overall efficiency of the optimization process.
[0004] However, when dealing with the multi-objective optimization problem of fuel cell plate flow field, existing machine learning optimization techniques need to coordinate multiple conflicting indicators such as maximizing net power density, controlling water saturation, and improving oxygen distribution uniformity. Traditional methods are difficult to efficiently and robustly obtain the globally optimal flow field structure that takes into account electrochemical performance, water management capabilities, and lightweight requirements. Summary of the Invention
[0005] To address the limitations of existing technologies and considering the design requirements of maximizing net power density, optimizing water saturation, and optimizing oxygen distribution uniformity, this invention proposes a multi-objective optimization design method and storage medium for fuel cells based on machine learning.
[0006] To achieve the above objectives, the first aspect of the present invention provides a multi-objective optimization design method for fuel cells based on machine learning, comprising: S1. Based on the flow field structure design requirements and load distribution of fuel cell plates, multiple design parameters are selected to construct a flow field parameterization model and an optimization target simulation model. A three-objective mathematical model of the flow field structure is constructed, with the three objectives including net power density, water saturation, and oxygen distribution uniformity. S2 uses a constraint-feasibility-driven population reconstruction to generate an optimized population, which is then used to build a global database. S3 uses a Gaussian kernel function to construct radial basis functions and derive gradient vectors, constructs a gradient-guided multi-objective evolutionary mechanism to generate offspring sample sets, and updates and optimizes the population; S4. Using the updated current optimal solution as the initial point, construct an adaptive sparsity-weighted penalized boundary objective function to perform local search in order to obtain the current optimal flow field structure. S5. Evaluate the current optimal flow field structure, update the database and optimize the population. If the current optimal flow field structure meets the geometric constraints and the design requirements of the three objectives, output the parameter values of the current optimal flow field structure. Otherwise, return to step S3 until the design requirements are met.
[0007] Optionally, step S1 specifically includes: S1.1 Based on the flow field structure design requirements and load distribution of fuel cell plates, the flow channel width, flow channel spacing, flow channel corner radius, flow channel unit half width, number of transverse flow channel units, number of longitudinal flow channel units and flow channel depth are selected as design parameters, and the design space is constructed according to the upper and lower bounds of the design parameter values. S1.2, Based on the design requirements of the fuel cell flow field structure and the load distribution, construct a parametric model of the flow field in the geometric modeling software; S1.3, Based on the flow field parameterization model, load distribution and boundary conditions, build an optimized target simulation model for net power density, water saturation and oxygen distribution uniformity in the finite element analysis software; S1.4 uses the three optimization objectives of maximizing net power density, optimizing water saturation, and optimizing oxygen distribution uniformity as optimization constraints, and uses flow field geometry constraints as optimization constraints to derive a three-objective mathematical model of the flow field structure.
[0008] Optionally, the specific expression of the three-objective mathematical model of the flow field structure is as follows: , , , In the above formula, This indicates the search for the optimal solution to the design parameters. Indicates the design parameters of the flow field structure. Indicates the width of the flow channel. Indicates the channel spacing. Indicates the radius of the flow channel corner. Indicates the half-width of the flow channel element. Indicates the number of transverse flow channel units. Indicates the number of longitudinal flow channel elements. Indicates the depth of the flow channel. Indicates minimization. The negative number representing net power density. Represents the net power density function. Indicator of water saturation deviation Indicates water saturation. Indicates the optimal water saturation. This indicates uneven distribution of oxygen. The function representing the non-uniformity of oxygen distribution. These are constraints that need to be satisfied. Represents the set of positive integers. Indicates the thickness of the cathode plate. Indicates the length of the cathode plate. This indicates the safe distance between the flow channel and the edge of the cathode plate. Indicates the width of the cathode plate. This indicates taking the minimum value. Indicates the minimum permissible water saturation. This indicates the maximum permissible water saturation level.
[0009] Optionally, step S2 specifically includes: S2.1, using Latin hypercube sampling to generate... An initial population of individuals; based on this, a double-loop optimization process is performed, in which the outer loop executes a total of [number] times. Each time, a new individual is generated through Latin hypercube sampling. The inner loop replaces the new individual with each position in the population and calculates the population diversity contribution value corresponding to each replacement scheme. Finally, the position that maximizes the diversity contribution is selected to perform the replacement operation. The steps for calculating the contribution value of population diversity are as follows: First, calculate the average distance between all individuals in the current population to quantify the overall diversity level of the population. , In the above formula, Indicates the current population set, This represents the average distance between all individuals in the current population. Indicates population size, Indicates the first Individual, Indicates the first Individual, Used for calculation and The Euclidean distance between them; Then, the change in population diversity after replacing an old individual with a new individual is measured to determine whether the replacement operation is beneficial to the improvement of diversity, thereby selecting the replacement scheme that contributes the most to diversity. , In the above formula, Indicates a new candidate individual. This represents the old entity that was replaced. Indicates using replace The change in post-population diversity This represents the population after removing older individuals. Represents a new set of individuals. This indicates the new population after the replacement. This represents the average distance between all individuals in the new population after the replacement. S2.2, using the optimization target simulation model of net power density, water saturation and oxygen distribution uniformity to obtain the net power density value, water saturation value and oxygen distribution uniformity value of the optimized population sample respectively, and using all optimized population samples to construct a global database.
[0010] Optionally, step S3 specifically includes: S3.1, radial basis function models for net power density, water saturation, and oxygen distribution uniformity are constructed using Gaussian kernel functions, with the specific expressions as follows: , , , in, , In the above formula, This represents the input vector to be predicted. Indicates net power density. The radial basis function model predicts the net power density. Indicates water saturation. The radial basis function model predicts the water saturation level. Indicates the uniformity of oxygen distribution. The radial basis function model predicts the uniformity of oxygen distribution. This represents the total number of training samples. Represents the net power density of the first The weights corresponding to each basis function. Indicating the degree of water saturation The weights corresponding to each basis function. The first term in indicating the uniformity of oxygen distribution The weights corresponding to each basis function. Represents the Gaussian kernel function. This represents the Gaussian kernel function corresponding to the net power density. This represents the Gaussian kernel function corresponding to water saturation. The Gaussian kernel function representing the uniformity of oxygen distribution. Indicates the first The input vector of training samples, wherein each training sample corresponds one-to-one with a basis function and serves as its center, and the number of basis functions is equal to the total number of training samples. Used for calculation and The Euclidean distance between them This represents the natural exponential function. This represents the kernel width parameter corresponding to the net power density. This parameter represents the width of the kernel function corresponding to water saturation. The kernel function width parameter represents the uniformity of oxygen distribution. The bias term represents the net power density. The bias term representing water saturation. Bias term indicating the uniformity of oxygen distribution; S3.2, Derive the gradient vector of each radial basis function model, as shown in the following expression: , , , In the above formula, This represents the input vector at the point where the gradient is to be calculated. Represents the radial basis function model of net power density In the input vector The gradient vector at that point, Represents the radial basis function model of water saturation In the input vector The gradient vector at that point, Radial basis function model representing oxygen distribution uniformity In the input vector The gradient vector at that point, Used for calculation and The Euclidean distance between them This represents the first derivative of the Gaussian kernel function with respect to distance; S3.3, Construct a gradient-guided multi-objective evolutionary mechanism to generate a set of offspring samples. The specific steps are as follows: First, calculate the subproblems. The gradient vector of the Chebyshev aggregation function is given by the following formula: , In the above formula, Represents the gradient operator. This represents the Chebyshev aggregation function. Representing subproblems The current solution, Representing subproblems The weight vector, Represents Chebyshev aggregation functions In the current solution The gradient vector with respect to the design parameters, Indicates the dominant target index. Represents the weight vector The Middle One portion, Indicates the first The radial basis function model of the target is in The predicted value at that location, Indicates the first A reference point for each goal. Indicates the first The radial basis function model of the target is in The gradient vector at that point; Then calculate the subproblem. The gradient descent step of the Chebyshev aggregation function is determined using Armijo line search to determine the step size. ,make sure To achieve a sufficient descent, the specific formula is as follows: , , In the above formula, Indicates the first The original candidate solution vector generated after gradient descent of each subproblem Indicates the step size. This represents the descent direction vector. Indicates a probing point The new aggregation function value at that location, This represents the original aggregation function value at the current solution. Indicates the coefficient of full descent; If the above-generated original candidate solutions If the solution exceeds the feasible region boundary, a boundary projection strategy is used to correct it into a feasible solution: , In the above formula, Indicates the first The new candidate solution vector after projection correction for each subproblem This represents the feasible region of the design parameters. Represents the projection operator. Let represent the independent variable that minimizes the objective function. Represents any point within the feasible region. express and The Euclidean distance between them; For all in the optimized population Repeat step S3.3 above for each sub-problem to finally obtain a complete set of offspring samples; S3.4, Selecting the first generation sample from the offspring sample set based on geometric structure constraints. A temporary child, and according to the Chebyshev function from the previous Select the best temporary child with the smallest Chebyshev function value from the temporary children; S3.5, using the best temporary offspring as the basis vector, performs dimensional perturbation mutation to generate... One current sub-library; S3.6, Filter the previous generation from the current child library based on geometric structure constraints. The current child, and according to the Chebyshev function from the previous Select the best current child with the smallest Chebyshev function value from the current children; S3.7, using the optimization target simulation model of net power density, water saturation and oxygen distribution uniformity, respectively, the net power density, water saturation and oxygen distribution uniformity values of the current offspring are obtained, and the optimal current offspring is added to the global database; S3.8 uses the best current offspring to replace the worst sample in the optimized population to achieve an optimized population update.
[0011] Optionally, step S4 specifically includes: S4.1, using the updated current optimal solution as the initial point, constructs an adaptive sparsity-weighted penalized boundary objective function. Based on the PBI penalized boundary cross method, the original multi-objective problem is transformed into a single-objective problem with a penalized term. The PBI penalized boundary cross method defines the search direction through the weight vector, causing the solution to converge to the Pareto front boundary along this direction. The specific formula is as follows: Construct the PBI penalized boundary cross-aggregation function: , , , In the above formula, This represents the design parameter vector, i.e., the candidate solutions in the optimization process; This represents the weight vector, used to define the search direction in the target space and reflect the preference among the targets; This represents a reference point composed of the ideal optimal values of each objective. This represents the cross function value of the PBI aggregation penalty boundary. This represents the projected distance of the candidate solution used to measure convergence along the direction of the weight vector. This represents the vertical distance from the candidate solution used to measure diversity to the weight vector. Indicates the adaptive penalty parameter. Represents the target vector. Represents the weight vector The Euclidean norm, Represents the transpose of a vector; The adaptive penalty parameter is calculated using the following formula: , , , In the above formula, Indicates the first Local density of sub-problems Indicates population size, Indicates the first The neighborhood solution set of each subproblem Indicates the first The target vector of each solution. Indicates the first The target vector of each solution. Indicates the global average density. Indicates the first Adaptive penalty parameters for each sub-problem This represents the global base penalty coefficient, which serves as the benchmark value for the penalty parameters of all subproblems; This represents the smallest positive number used to ensure numerical stability; S4.2, Derive the gradient vector of the final local search objective function. This gradient vector contains information on both the descent direction of the objective and the constraint repair direction, which can guide the search to simultaneously optimize performance and satisfy constraints. The specific formula is as follows: , In the above formula, Indicates the first The final local search objective function for each subproblem Describes the final local search objective function In candidate solutions The gradient vector at that point, Represents the gradient operator. Represents candidate solutions used to measure convergence. Projected distance along the weight vector direction, Represents candidate solutions used to measure diversity. The vertical distance to the weight vector. Indicates the sparsity reward weight. Indicates candidate solutions Corresponding target vector Sparsity at a given location; S4.3, a gradient optimization method is used to perform a local search to generate a set of candidate solutions. The specific formula is as follows: , , , In the above formula, Let represent any target vector used for probability density estimation. Indicating in the target vector The estimated probability density value at that location, Indicates population size, Represents the Gaussian kernel function. Represents the target vector sparsity at that location Indicates the first The final local search objective function for each subproblem; Based on the aforementioned final local search objective function, gradient descent is used to iteratively update the current solution of each subproblem, applying this method to all... Each subproblem is updated sequentially using gradient descent, and the updated solutions constitute the candidate solution set. S4.4. After real evaluation, the current optimal solution is selected from the candidate solution set. Then, geometric structure constraints are used to perform feasible repairs on the current optimal solution to ensure that it fully meets the geometric structure constraints, thus obtaining the current optimal flow field structure.
[0012] In another aspect, the present invention provides a storage medium, which is a computer-readable storage medium and stores a computer program that, when executed by a processor, implements the above-described machine learning-based multi-objective optimization design method for fuel cells.
[0013] In summary, compared with the prior art, the technical solutions conceived in this invention have the following main advantages: This invention addresses the multi-objective optimization problem of fuel cell electrode flow field structure by constructing an adaptive sparsity-weighted penalized boundary objective function for local search. This dynamically reconciles the contradiction between improving objective performance and the feasibility of constraint boundaries during the optimization process, guiding candidate solutions to converge towards the region with optimal net power density while satisfying requirements for water saturation and oxygen distribution uniformity. This improves the algorithm's engineering applicability and optimization efficiency in fuel cell electrode flow field structure design. This invention overcomes the limitations of existing technologies in efficiently coordinating gradient information from radial basis function models with multi-objective evolutionary strategies based on PBI (Penalty-Based Boundary Intersection) decomposition, and the lack of an optimization mechanism that dynamically adjusts penalty parameters based on the sparsity of the objective space, making it difficult to balance convergence speed, population diversity, and constraint handling capabilities.
[0014] This invention combines an adaptive sparsity-weighted penalty boundary objective function for local search, effectively balancing multiple conflicting objectives such as maximizing net power density, controlling water saturation, and improving oxygen distribution uniformity. It obtains the globally optimal current optimal flow field structure parameter values that take into account comprehensive performance. This can solve the limitations of traditional methods in multi-objective optimization of fuel cell electrode flow fields, which rely on time-consuming CFD (Computational Fluid Dynamics) simulation iterations, have high computational costs, and are difficult to accurately handle the coupling relationships between objectives.
[0015] In summary, this invention can efficiently solve multi-objective optimization problems under complex constraints of fuel cell plates, adapt to complex working conditions of multi-objective conflict and constraint coupling in CFD simulation scenarios, improve optimization accuracy and comprehensive flow field performance, and accurately achieve comprehensive optimal control of net power density, water saturation and oxygen distribution uniformity. At the same time, it provides a usable algorithm framework of "RBF (Radial Basis Function) model + PBI decomposition + gradient information guidance" for other multi-objective engineering optimization designs. Attached Figure Description
[0016] Figure 1 A simplified flowchart of the machine learning-based multi-objective optimization design method for fuel cells provided by this invention. Detailed Implementation
[0017] To more clearly illustrate the research objectives, technical solutions, and core advantages of this invention, the following will provide a more detailed description of the invention in conjunction with the accompanying drawings and specific embodiments. It should be noted that the specific embodiments described herein are merely illustrative of the technical concepts of this invention and do not constitute a limitation on the scope of protection of this invention. Furthermore, the technical features involved in the various embodiments of this invention can be arbitrarily combined, as long as there is no logical conflict between them.
[0018] Please see Figure 1 The present invention provides a machine learning-based multi-objective optimization design method for fuel cells, which is applicable to the three-objective optimization design of fuel cell flow field structure. Specifically, the method includes steps S1 to S5.
[0019] S1. Based on the flow field structure design requirements and load distribution of fuel cell plates, multiple design parameters are selected to construct a flow field parameterization model and an optimization target simulation model. A three-objective mathematical model of the flow field structure is constructed, with the three objectives including net power density, water saturation, and oxygen distribution uniformity.
[0020] The specific steps of step S1 are as follows: S1.1 Based on the flow field structure design requirements and load distribution of fuel cell plates, the flow channel width, flow channel spacing, flow channel corner radius, flow channel unit half width, number of transverse flow channel units, number of longitudinal flow channel units and flow channel depth are selected as design parameters, and the design space is constructed according to the upper and lower bounds of the design parameter values. S1.2, Based on the design requirements of the fuel cell flow field structure and the load distribution, construct a parametric model of the flow field in the geometric modeling software; S1.3, Based on the flow field parameterization model, load distribution and boundary conditions, build an optimized target simulation model for net power density, water saturation and oxygen distribution uniformity in the finite element analysis software; S1.4 uses the maximization of net power density, optimization of water saturation, and optimization of oxygen distribution uniformity as the three optimization objectives, and the flow field geometry constraints as the optimization constraints to derive the three-objective mathematical model of the flow field structure. The specific expression is as follows: , , , In the above formula, This indicates the search for the optimal solution to the design parameters. Indicates the design parameters of the flow field structure. Indicates the width of the flow channel. Indicates the channel spacing. Indicates the radius of the flow channel corner. Indicates the half-width of the flow channel element. Indicates the number of transverse flow channel units. Indicates the number of longitudinal flow channel elements. Indicates the depth of the flow channel. Indicates minimization. The negative number representing net power density. Represents the net power density function. Indicator of water saturation deviation Indicates water saturation. Indicates the optimal water saturation. This indicates uneven distribution of oxygen. The function representing the non-uniformity of oxygen distribution. These are constraints that need to be satisfied. Represents the set of positive integers. Indicates the thickness of the cathode plate. Indicates the length of the cathode plate. This indicates the safe distance between the flow channel and the edge of the cathode plate. Indicates the width of the cathode plate. This indicates taking the minimum value. Indicates the minimum permissible water saturation. This indicates the maximum permissible water saturation level.
[0021] S2 uses a constraint-feasibility-driven population reconstruction to generate an optimized population, which is then used to build a global database.
[0022] The specific steps of step S2 are as follows: S2.1, using Latin hypercube sampling to generate... An initial population of individuals; based on this, a double-loop optimization process is performed, in which the outer loop executes a total of [number] times. Each time, a new individual is generated through Latin hypercube sampling. The inner loop replaces the new individual with each position in the population and calculates the population diversity contribution value corresponding to each replacement scheme. Finally, the position that maximizes the diversity contribution is selected to perform the replacement operation. The steps for calculating the contribution value of population diversity are as follows: First, calculate the average distance between all individuals in the current population to quantify the overall diversity level of the population. , In the above formula, Indicates the current population set, This represents the average distance between all individuals in the current population. Indicates population size, Indicates the first Individual, Indicates the first Individual, Used for calculation and The Euclidean distance between them; Then, the change in population diversity after replacing an old individual with a new individual is measured to determine whether the replacement operation is beneficial to the improvement of diversity, thereby selecting the replacement scheme that contributes the most to diversity. , In the above formula, Indicates a new candidate individual. This represents the old entity that was replaced. Indicates using replace The change in post-population diversity This represents the population after removing older individuals. Represents a new set of individuals. This indicates the new population after the replacement. This represents the average distance between all individuals in the new population after the replacement. S2.2, using the optimization target simulation model of net power density, water saturation and oxygen distribution uniformity to obtain the net power density value, water saturation value and oxygen distribution uniformity value of the optimized population sample respectively, and using all optimized population samples to construct a global database.
[0023] S3 uses a Gaussian kernel function to construct radial basis functions and derive gradient vectors, constructs a gradient-guided multi-objective evolutionary mechanism to generate offspring sample sets, and updates and optimizes the population.
[0024] The specific steps of step S3 are as follows: S3.1, radial basis function models for net power density, water saturation, and oxygen distribution uniformity are constructed using Gaussian kernel functions, with the specific expressions as follows: , , , in, , In the above formula, This represents the input vector to be predicted. Indicates net power density. The radial basis function model predicts the net power density. Indicates water saturation. The radial basis function model predicts the water saturation level. Indicates the uniformity of oxygen distribution. The radial basis function model predicts the uniformity of oxygen distribution. This represents the total number of training samples. Represents the net power density of the first The weights corresponding to each basis function. Indicating the degree of water saturation The weights corresponding to each basis function. The first term in indicating the uniformity of oxygen distribution The weights corresponding to each basis function. Represents the Gaussian kernel function. This represents the Gaussian kernel function corresponding to the net power density. This represents the Gaussian kernel function corresponding to water saturation. The Gaussian kernel function representing the uniformity of oxygen distribution. Indicates the first The input vector of training samples, wherein each training sample corresponds one-to-one with a basis function and serves as its center, and the number of basis functions is equal to the total number of training samples. Used for calculation and The Euclidean distance between them This represents the natural exponential function. This represents the kernel width parameter corresponding to the net power density. This parameter represents the width of the kernel function corresponding to water saturation. The kernel function width parameter represents the uniformity of oxygen distribution. The bias term represents the net power density. The bias term representing water saturation. Bias term indicating the uniformity of oxygen distribution; S3.2, Derive the gradient vector of each radial basis function model, as shown in the following expression: , , , In the above formula, This represents the input vector at the point where the gradient is to be calculated. Represents the radial basis function model of net power density In the input vector The gradient vector at that point, Represents the radial basis function model of water saturation In the input vector The gradient vector at that point, Radial basis function model representing oxygen distribution uniformity In the input vector The gradient vector at that point, Used for calculation and The Euclidean distance between them This represents the first derivative of the Gaussian kernel function with respect to distance; S3.3, Construct a gradient-guided multi-objective evolutionary mechanism to generate a set of offspring samples. The specific steps are as follows: First, calculate the subproblems. The gradient vector of the Chebyshev aggregation function is given by the following formula: , In the above formula, Represents the gradient operator. This represents the Chebyshev aggregation function. Representing subproblems The current solution, Representing subproblems The weight vector, Represents Chebyshev aggregation functions In the current solution The gradient vector with respect to the design parameters, Indicates the dominant target index. Represents the weight vector The Middle One portion, Indicates the first The radial basis function model of the target is in The predicted value at that location, Indicates the first A reference point for each goal. Indicates the first The radial basis function model of the target is in The gradient vector at that point; Then calculate the subproblem. The gradient descent step of the Chebyshev aggregation function is determined using Armijo line search to determine the step size. ,make sure To achieve a sufficient descent, the specific formula is as follows: , , In the above formula, Indicates the first The original candidate solution vector generated after gradient descent of each subproblem Indicates the step size. This represents the descent direction vector. Indicates a probing point The new aggregation function value at that location, This represents the original aggregation function value at the current solution. Indicates the coefficient of full descent; If the above-generated original candidate solutions If the solution exceeds the feasible region boundary, a boundary projection strategy is used to correct it into a feasible solution: , In the above formula, Indicates the first The new candidate solution vector after projection correction for each subproblem This represents the feasible region of the design parameters. Represents the projection operator. Let represent the independent variable that minimizes the objective function. Represents any point within the feasible region. express and The Euclidean distance between them; For all in the optimized population Repeat step S3.3 above for each sub-problem to finally obtain a complete set of offspring samples; S3.4, Selecting the first generation sample from the offspring sample set based on geometric structure constraints. A temporary child, and according to the Chebyshev function from the previous Select the best temporary child with the smallest Chebyshev function value from the temporary children; S3.5, using the best temporary offspring as the basis vector, performs dimensional perturbation mutation to generate... One current sub-library; S3.6, Filter the previous generation from the current child library based on geometric structure constraints. The current child, and according to the Chebyshev function from the previous Select the best current child with the smallest Chebyshev function value from the current children; S3.7, using the optimization target simulation model of net power density, water saturation and oxygen distribution uniformity, respectively, the net power density, water saturation and oxygen distribution uniformity values of the current offspring are obtained, and the optimal current offspring is added to the global database; S3.8 uses the best current offspring to replace the worst sample in the optimized population to achieve an optimized population update.
[0025] S4. Using the updated current optimal solution as the initial point, construct an adaptive sparsity-weighted penalty boundary objective function to perform a local search in order to obtain the current optimal flow field structure.
[0026] The specific steps of step S4 are as follows: S4.1, using the updated current optimal solution as the initial point, constructs an adaptive sparsity-weighted penalized boundary objective function. Based on the PBI penalized boundary cross method, the original multi-objective problem is transformed into a single-objective problem with a penalized term. The PBI penalized boundary cross method defines the search direction through the weight vector, causing the solution to converge to the Pareto front boundary along this direction. The specific formula is as follows: Construct the PBI penalized boundary cross-aggregation function: , , , In the above formula, This represents the design parameter vector, i.e., the candidate solutions in the optimization process; This represents the weight vector, used to define the search direction in the target space and reflect the preference among the targets; This represents a reference point composed of the ideal optimal values of each objective. This represents the cross function value of the PBI aggregation penalty boundary. This represents the projected distance of the candidate solution used to measure convergence along the direction of the weight vector. This represents the vertical distance from the candidate solution used to measure diversity to the weight vector. Indicates the adaptive penalty parameter. Represents the target vector. Represents the weight vector The Euclidean norm, Represents the transpose of a vector; The adaptive penalty parameter is calculated using the following formula: , , , In the above formula, Indicates the first Local density of sub-problems Indicates population size, Indicates the first The neighborhood solution set of each subproblem Indicates the first The target vector of each solution. Indicates the first The target vector of each solution. Indicates the global average density. Indicates the first Adaptive penalty parameters for each sub-problem This represents the global base penalty coefficient, which serves as the benchmark value for the penalty parameters of all subproblems; This represents the smallest positive number used to ensure numerical stability; S4.2, Derive the gradient vector of the final local search objective function. This gradient vector contains information on both the descent direction of the objective and the constraint repair direction, which can guide the search to simultaneously optimize performance and satisfy constraints. The specific formula is as follows: , In the above formula, Indicates the first The final local search objective function for each subproblem Describes the final local search objective function In candidate solutions The gradient vector at that point, Represents the gradient operator. Represents candidate solutions used to measure convergence. Projected distance along the direction of the weight vector Represents candidate solutions used to measure diversity. The perpendicular distance to the weight vector. Indicates the sparsity reward weight. Indicates candidate solutions Corresponding target vector Sparsity at a given location; S4.3 uses a gradient optimization method to perform a local search to generate a set of candidate solutions. The specific formula is as follows: , , , In the above formula, Let represent any target vector used for probability density estimation. Indicating in the target vector The estimated probability density value at that location. Indicates population size, Represents the Gaussian kernel function. Represents the target vector sparsity at that location Indicates the first The final local search objective function for each subproblem; Based on the aforementioned final local search objective function, gradient descent is used to iteratively update the current solution of each subproblem, applying this method to all... Each subproblem is updated sequentially using gradient descent, and the updated solutions constitute the candidate solution set. S4.4. After real evaluation, the current optimal solution is selected from the candidate solution set. Then, geometric structure constraints are used to perform feasible repairs on the current optimal solution to ensure that it fully meets the geometric structure constraints, thus obtaining the current optimal flow field structure.
[0027] S5. Evaluate the current optimal flow field structure, update the database and optimize the population. If the current optimal flow field structure meets the geometric constraints and the design requirements of the three objectives, output the parameter values of the current optimal flow field structure. Otherwise, return to step S3 until the design requirements are met.
[0028] Example 1 illustrates the optimization performance of the proposed machine learning-based multi-objective optimization design method for fuel cells using a benchmark function. The benchmark function includes three optimization objectives and three optimization constraints. The expression is as follows: , , , , , , , , , In the above formula, To simultaneously minimize the three objectives of the benchmark function, , , These are the first, second, and third objective values of the benchmark function, used to verify the convergence performance and distribution uniformity of the algorithm. It is the input vector of the benchmark function. It is the dimension of the benchmark function. Represents the input vector of the benchmark function The first in One input variable, Represents the distance function. , Represents the input vector of the benchmark function The first and second components are used to control the position of the target value on the Pareto front. Represents the input vector of the benchmark function The remaining portion, This represents the remaining input variables, used to control the distance of the population to the Pareto front; 、 、 These are the first, second, and third constraint functions of the benchmark function, used to verify the algorithm's ability to handle constrained optimization problems; This represents the constraint radius parameter; in this embodiment, we take... ; This indicates the range of values that the input variables of the benchmark function can take.
[0029] The above benchmark test functions are processed through steps S1 to S5 of the machine learning-based fuel cell multi-objective optimization design method (hereinafter referred to as the method of this embodiment) provided in this embodiment to obtain experimental results.
[0030] To further illustrate this embodiment, the method described herein is compared with another classic differential evolution algorithm. The maximum number of simulation evaluations for this embodiment is set to 1000, and the experimental results are shown in Table 1. With the same number of simulation samples, IGD (Inverted Generational Distance) is used as a comprehensive evaluation index for population convergence and distribution. A smaller value indicates better overall performance. The IGD value of the method in this embodiment is significantly smaller than that of the differential evolution algorithm, indicating that the solution set obtained by the method in this embodiment has significant advantages in terms of Pareto front approximation and distribution uniformity. It can be considered that the method in this embodiment performs well in the three-objective optimization design problem of maximizing net power density, optimizing water saturation, optimizing oxygen distribution uniformity, and constraining channel width, channel spacing, and channel turning radius.
[0031] Table 1. Comparison of optimization results of different methods
[0032] This invention provides a multi-objective optimization design method for fuel cells based on machine learning. It designs an adaptive sparsity-weighted penalized boundary cross method to construct a local search objective function (i.e., constructing an adaptive sparsity-weighted penalized boundary objective function for local search). This method can efficiently solve multi-objective optimization problems of electrode flow fields under complex CFD simulation scenarios, improve optimization accuracy and overall performance balance, and provide a systematic solution for multi-objective optimization design with constraints on maximizing net power density and water saturation and oxygen distribution uniformity.
[0033] Secondly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the machine learning-based multi-objective optimization design method for fuel cells described in the foregoing embodiments.
[0034] Computer storage media can include various media that can store program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0035] Specifically, computer program products include: data signals, and data signals embodied in a carrier wave.
[0036] It should be noted that the technical solutions described in this invention can be combined arbitrarily without conflict.
[0037] Those skilled in the art will readily understand that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A multi-objective optimization design method for fuel cells based on machine learning, characterized in that, include: S1. Based on the flow field structure design requirements and load distribution of fuel cell plates, multiple design parameters are selected to construct a flow field parameterization model and an optimization target simulation model. A three-objective mathematical model of the flow field structure is constructed, with the three objectives including net power density, water saturation, and oxygen distribution uniformity. S2 uses a constraint-feasibility-driven population reconstruction to generate an optimized population, which is then used to build a global database. S3 uses a Gaussian kernel function to construct radial basis functions and derive gradient vectors, constructs a gradient-guided multi-objective evolutionary mechanism to generate offspring sample sets, and updates and optimizes the population; S4. Using the updated current optimal solution as the initial point, construct an adaptive sparsity-weighted penalized boundary objective function to perform local search in order to obtain the current optimal flow field structure. S5. Evaluate the current optimal flow field structure, update the database and optimize the population. If the current optimal flow field structure meets the geometric constraints and the design requirements of the three objectives, output the parameter values of the current optimal flow field structure. Otherwise, return to step S3 until the design requirements are met.
2. The multi-objective optimization design method for fuel cells based on machine learning as described in claim 1, characterized in that, Step S1 specifically includes: S1.1 Based on the flow field structure design requirements and load distribution of fuel cell plates, the flow channel width, flow channel spacing, flow channel corner radius, flow channel unit half width, number of transverse flow channel units, number of longitudinal flow channel units and flow channel depth are selected as design parameters, and the design space is constructed according to the upper and lower bounds of the design parameter values. S1.2, Based on the design requirements of the fuel cell flow field structure and the load distribution, construct a parametric model of the flow field in the geometric modeling software; S1.3, Based on the flow field parameterization model, load distribution and boundary conditions, build an optimized target simulation model for net power density, water saturation and oxygen distribution uniformity in finite element analysis software; S1.4 uses the three optimization objectives of maximizing net power density, optimizing water saturation, and optimizing oxygen distribution uniformity as optimization constraints, and uses flow field geometry constraints as optimization constraints to derive a three-objective mathematical model of the flow field structure.
3. The multi-objective optimization design method for fuel cells based on machine learning as described in claim 2, characterized in that, The specific expression of the three-objective mathematical model of the flow field structure is as follows: , , , In the above formula, This indicates the search for the optimal solution to the design parameters. Indicates the design parameters of the flow field structure. Indicates the width of the flow channel. Indicates the channel spacing. Indicates the radius of the flow channel corner. Indicates the half-width of the flow channel element. Indicates the number of transverse flow channel units. Indicates the number of longitudinal flow channel elements. Indicates the depth of the flow channel. Indicates minimization. The negative number representing net power density. Represents the net power density function. Indicator of water saturation deviation Indicates water saturation. Indicates the optimal water saturation. This indicates uneven distribution of oxygen. The function representing the non-uniformity of oxygen distribution. These are constraints that need to be satisfied. Represents the set of positive integers. Indicates the thickness of the cathode plate. Indicates the length of the cathode plate. This indicates the safe distance between the flow channel and the edge of the cathode plate. Indicates the width of the cathode plate. This indicates taking the minimum value. Indicates the minimum permissible water saturation. This indicates the maximum permissible water saturation.
4. The multi-objective optimization design method for fuel cells based on machine learning as described in claim 1, characterized in that, Step S2 specifically includes: S2.1, using Latin hypercube sampling to generate... An initial population of individuals; based on this, a double-loop optimization process is performed, in which the outer loop executes a total of [number] times. Each time, a new individual is generated through Latin hypercube sampling. The inner loop replaces the new individual with each position in the population and calculates the population diversity contribution value corresponding to each replacement scheme. Finally, the position that maximizes the diversity contribution is selected to perform the replacement operation. The steps for calculating the contribution value of population diversity are as follows: First, calculate the average distance between all individuals in the current population to quantify the overall diversity level of the population. , In the above formula, Indicates the current population set, This represents the average distance between all individuals in the current population. Indicates population size, Indicates the first Individual, Indicates the first Individual, Used for calculation and The Euclidean distance between them; Then, the change in population diversity after replacing an old individual with a new individual is measured to determine whether the replacement operation is beneficial to the improvement of diversity, thereby selecting the replacement scheme that contributes the most to diversity. , In the above formula, Indicates a new candidate individual. This represents the old entity that was replaced. Indicates using replace The change in post-population diversity This represents the population after removing older individuals. Represents a new set of individuals. This indicates the new population after the replacement. This represents the average distance between all individuals in the new population after the replacement. S2.2, using the optimization target simulation model of net power density, water saturation and oxygen distribution uniformity to obtain the net power density value, water saturation value and oxygen distribution uniformity value of the optimized population sample respectively, and using all optimized population samples to construct a global database.
5. The multi-objective optimization design method for fuel cells based on machine learning according to claim 4, characterized in that, Step S3 specifically includes: S3.1, radial basis function models for net power density, water saturation, and oxygen distribution uniformity are constructed using Gaussian kernel functions, with the specific expressions as follows: , , , in, , In the above formula, This represents the input vector to be predicted. Indicates net power density. The radial basis function model predicts the net power density. Indicates water saturation. The radial basis function model predicts the water saturation level. Indicates the uniformity of oxygen distribution. The radial basis function model predicts the uniformity of oxygen distribution. This represents the total number of training samples. Represents the net power density of the first The weights corresponding to each basis function. Indicating the degree of water saturation The weights corresponding to each basis function. The first term in indicating the uniformity of oxygen distribution The weights corresponding to each basis function. Represents the Gaussian kernel function. This represents the Gaussian kernel function corresponding to the net power density. This represents the Gaussian kernel function corresponding to water saturation. The Gaussian kernel function representing the uniformity of oxygen distribution. Indicates the first The input vector of training samples, wherein each training sample corresponds one-to-one with a basis function and serves as its center, and the number of basis functions is equal to the total number of training samples. Used for calculation and The Euclidean distance between them This represents the natural exponential function. This represents the kernel width parameter corresponding to the net power density. This parameter represents the width of the kernel function corresponding to water saturation. The kernel function width parameter represents the uniformity of oxygen distribution. The bias term represents the net power density. The bias term representing water saturation. Bias term indicating the uniformity of oxygen distribution; S3.2, Derive the gradient vector of each radial basis function model, as shown in the following expression: , , , In the above formula, This represents the input vector at the point where the gradient is to be calculated. Represents the radial basis function model of net power density In the input vector The gradient vector at that point, Represents the radial basis function model of water saturation In the input vector The gradient vector at that point, Radial basis function model representing oxygen distribution uniformity In the input vector The gradient vector at that point, Used for calculation and The Euclidean distance between them This represents the first derivative of the Gaussian kernel function with respect to distance; S3.3, Construct a gradient-guided multi-objective evolutionary mechanism to generate a set of offspring samples. The specific steps are as follows: First, calculate the subproblems. The gradient vector of the Chebyshev aggregation function is given by the following formula: , In the above formula, Represents the gradient operator, This represents the Chebyshev aggregation function. Representing subproblems The current solution, Representing subproblems The weight vector, Represents Chebyshev aggregation functions In the current solution The gradient vector with respect to the design parameters, Indicates the dominant target index. Represents the weight vector The Middle One portion, Indicates the first The radial basis function model of the target is in The predicted value at that location, Indicates the first A reference point for each goal. Indicates the first The radial basis function model of the target is in The gradient vector at that point; Then calculate the subproblem. The gradient descent step of the Chebyshev aggregation function is determined using Armijo line search to determine the step size. ,make sure To achieve a sufficient descent, the specific formula is as follows: , , In the above formula, Indicates the first The original candidate solution vector generated after gradient descent of each subproblem Indicates the step size. This represents the descent direction vector. Indicates a probing point The new aggregation function value at that location, This represents the original aggregation function value at the current solution. Indicates the coefficient of full descent; If the above-generated original candidate solutions If the solution exceeds the feasible region boundary, a boundary projection strategy is used to correct it into a feasible solution: , In the above formula, Indicates the first The new candidate solution vector after projection correction for each subproblem This represents the feasible region of the design parameters. Represents the projection operator. Denotes the independent variable that minimizes the objective function. Represents any point within the feasible region. express and The Euclidean distance between them; For all in the optimized population Repeat step S3.3 above for each sub-problem to finally obtain a complete set of offspring samples; S3.4, Selecting the first generation sample from the offspring sample set based on geometric structure constraints. A temporary child, and according to the Chebyshev function from the previous Select the best temporary child with the smallest Chebyshev function value from the temporary children; S3.5, using the best temporary offspring as the basis vector, performs dimensional perturbation mutation to generate... One current sub-library; S3.6, Filter the previous generation from the current child library based on geometric structure constraints. The current child, and according to the Chebyshev function from the previous Select the best current child with the smallest Chebyshev function value from the current children; S3.7, using the optimization target simulation model of net power density, water saturation and oxygen distribution uniformity, respectively, the net power density, water saturation and oxygen distribution uniformity values of the current offspring are obtained, and the optimal current offspring is added to the global database; S3.8 uses the best current offspring to replace the worst sample in the optimized population to achieve an optimized population update.
6. The multi-objective optimization design method for fuel cells based on machine learning as described in claim 1, characterized in that, Step S4 specifically includes: S4.1, using the updated current optimal solution as the initial point, constructs an adaptive sparsity-weighted penalized boundary objective function. Based on the PBI penalized boundary cross method, the original multi-objective problem is transformed into a single-objective problem with a penalized term. The PBI penalized boundary cross method defines the search direction through the weight vector, causing the solution to converge to the Pareto front boundary along this direction. The specific formula is as follows: Construct the PBI penalized boundary cross-aggregation function: , , , In the above formula, This represents the design parameter vector, i.e., the candidate solutions in the optimization process; This represents the weight vector, used to define the search direction in the target space and reflect the preference among the targets; This represents a reference point composed of the ideal optimal values of each objective. This represents the cross function value of the PBI aggregation penalty boundary. This represents the projected distance of the candidate solution used to measure convergence along the direction of the weight vector. This represents the vertical distance from the candidate solution used to measure diversity to the weight vector. Indicates the adaptive penalty parameter. Represents the target vector. Represents the weight vector The Euclidean norm, Represents the transpose of a vector; The adaptive penalty parameter is calculated using the following formula: , , , In the above formula, Indicates the first Local density of sub-problems Indicates population size, Indicates the first The neighborhood solution set of each subproblem Indicates the first The target vector of each solution. Indicates the first The target vector of each solution. Indicates the global average density. Indicates the first Adaptive penalty parameters for each sub-problem This represents the global base penalty coefficient, which serves as the benchmark value for the penalty parameters of all subproblems; This represents the smallest positive number used to ensure numerical stability; S4.2, Derive the gradient vector of the final local search objective function. This gradient vector contains information on both the descent direction of the objective and the constraint repair direction, which can guide the search to simultaneously optimize performance and satisfy constraints. The specific formula is as follows: , In the above formula, Indicates the first The final local search objective function for each subproblem Describes the final local search objective function In candidate solutions The gradient vector at that point, Represents the gradient operator, Represents candidate solutions used to measure convergence. Projected distance along the weight vector direction, Represents candidate solutions used to measure diversity. The vertical distance to the weight vector. Indicates the sparsity reward weight. Indicates candidate solutions Corresponding target vector Sparsity at a given location; S4.3 uses a gradient optimization method to perform a local search to generate a set of candidate solutions. The specific formula is as follows: , , , In the above formula, Let represent any target vector used for probability density estimation. Indicating in the target vector The estimated probability density value at that location, Indicates population size, Represents the Gaussian kernel function. Represents the target vector sparsity at that location Indicates the first The final local search objective function for each subproblem; Based on the aforementioned final local search objective function, gradient descent is used to iteratively update the current solution of each subproblem, applying this method to all... Each subproblem is updated sequentially using gradient descent, and the updated solutions constitute the candidate solution set. S4.
4. After real evaluation, the current optimal solution is selected from the candidate solution set. Then, geometric structure constraints are used to perform feasible repairs on the current optimal solution to ensure that it fully meets the geometric structure constraints, thus obtaining the current optimal flow field structure.
7. A storage medium, characterized in that, The storage medium is a computer-readable storage medium and stores a computer program, which, when executed by a processor, implements the multi-objective optimization design method for fuel cells based on machine learning as described in any one of claims 1-6.
