Cathode open type air-cooled fuel cell multi-objective optimization method based on artificial intelligence

Abstract

The invention discloses a cathode open type air-cooled fuel cell multi-objective optimization method based on artificial intelligence, and relates to the field of batteries, and the method comprises the steps: collecting battery data, and extracting data features through a machine learning technology; establishing a battery model based on machine learning for predicting the performance and health state of the battery, and training the model by using historical data; optimizing design parameters by utilizing an optimization algorithm, combining a plurality of objective functions into a comprehensive objective function, and searching an optimal equilibrium solution through a multi-objective optimization algorithm; the optimized battery design parameters and the operation strategy are applied to an actual battery system, and an optimization algorithm is optimized and adjusted according to an experimental result. A multi-objective optimization algorithm is introduced into fuel cell system performance model development, the engineering suggestion range of the width, the height, the ridge width and the bending angle of a flow channel is further analyzed and given, and the problem that in the background technology, an agent model optimized through a single variable method cannot obtain an optimal equilibrium solution is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of hydrogen fuel cells, and in particular to a multi-objective optimization method for cathode open air-cooled fuel cells based on artificial intelligence. Background Art

[0002] Artificial intelligence (AI) methods have wide applicability, and "AI+" methods have shown great potential in achieving revolutionary progress in technological development. In the field of hydrogen fuel cells, AI methods can be used to predict performance or life, propose intelligent control strategies, and optimize various parameters. Compared with predictions based on experiments or numerical simulations, AI methods can shorten the prediction time from hours or even days to minutes or even seconds without losing accuracy, saving a lot of time and having important application value.

[0003] Since there is an inherent correlation between each parameter of a hydrogen fuel cell, the traditional proxy model that uses a single variable optimization method is not necessarily the optimal proxy model. Therefore, it is necessary to introduce a multi-objective optimization algorithm into the development of a fuel cell system performance proxy model. Summary of the invention

[0004] In view of this, the present invention provides a multi-objective optimization method for proton exchange membrane fuel cells based on artificial intelligence to solve the above technical problems.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A multi-objective optimization method for proton exchange membrane fuel cells based on artificial intelligence, comprising the following steps:

[0007] Collect battery data and use machine learning technology to extract data features;

[0008] Establish a machine learning-based battery model to predict battery performance and health status, and use historical data to train the model;

[0009] Optimize design parameters using optimization algorithms, combine multiple objective functions into a comprehensive objective function, and find the optimal balance solution through multi-objective optimization algorithms;

[0010] Apply the optimized battery design parameters and operating strategies to the actual battery system and conduct field verification, and optimize and adjust the optimization algorithm based on the experimental results.

[0011] Furthermore, the optimization algorithm is used to optimize the design parameters, multiple objective functions are combined into a comprehensive objective function, and the optimal balance solution is found through a multi-objective optimization algorithm, which specifically includes the following steps:

[0012] Determine multiple objective functions to be optimized, select decision variables that affect the value of the objective function, and determine their value range and constraints;

[0013] Obtain sample points, select multiple sample points through experimental design methods within the determined range of design variable values, obtain the outputs corresponding to the sample points, and the constructed sample points include a training set for training the model and a test set for testing the model;

[0014] The model is determined by constructing a model with sample points, training the model with a training set, and testing the accuracy of the trained model with a test set;

[0015] Using the database to train the model through a data-driven approach and test the effectiveness of the proxy model;

[0016] Through iterative calculation and optimization algorithm search space, the approximate value of the optimal solution is found. In each iteration, the value of the decision variable is adjusted according to the value of the objective function and the algorithm rules until the stopping condition is met;

[0017] After obtaining a set of candidate solution sets, the candidate solution sets are screened. Specifically, a non-dominated sorting algorithm is used to divide the solution sets into different optimal levels, and the optimal solution set is selected.

[0018] Furthermore, the method of extracting data features using machine learning technology specifically includes the following steps:

[0019] The data set is randomly divided into training set: test set = 7:3.

[0020] Furthermore, the establishment of a battery proxy model based on machine learning specifically includes the following steps:

[0021] Build and train AI agent models;

[0022] The agent model is used as the fitness function and the genetic algorithm is used to optimize the model to obtain the maximum power density of the AI ​​agent model and the corresponding flow channel structure parameters. The engineering recommended range of the flow channel structure parameters is further analyzed and given.

[0023] Furthermore, the AI ​​agent model includes an SVR model, and the SVR model is expressed as follows:

[0024]

[0025] The first inequality: stf(x i )-y i ≤ε+ξ i

[0026] The second inequality:

[0027]

[0028] Among them, ξ i and is a slack variable, indicating the distance that the points outside the error band exceed the error band; y is the predicted value vector, with a size of 1×N; X is the input matrix, with a size of D in ×N, the component form is [x1,x2,…,x N ]; ω is the coefficient vector, size D in ×1; b is the bias term; C is the regularization constant, ε is the deviation, z ε is the insensitive loss function of ε; is a regularization term used to measure the complexity of the surrogate model to prevent overfitting.

[0029] Furthermore, the dual form of the SVR model is:

[0030]

[0031]

[0032]

[0033] At the same time, the coefficient vector ω is:

[0034]

[0035] Where: α i and The Lagrange multipliers introduced by the first inequality and the second inequality are respectively solved by using quadratic programming.

[0036] Furthermore, the dual form satisfies the KKT condition, which is:

[0037] α i [f(x i )-y i -(ε+ξ i )]=0

[0038]

[0039]

[0040]

[0041] in, Description i and At least one of them is zero, and the solution of SVR is expressed as:

[0042]

[0043] Therefore, the proxy model consists only of The sample points of are completely determined, and these sample points are called support vectors;

[0044] when When This sample point is within the error interval, is not a support vector, and has no effect on the proxy model; when 0<α i <C, or α i = 0, ξ i =0,f(x i )-y i =ε or f(x i )-y i =-ε, that is, the sample point falls exactly on the boundary of the error interval band, and this sample point is the standard support vector; when α i =C, or α i = 0, f(x i )-y i =ε+ξ i ≥ε or That is, the sample point is on the interval band or outside the boundary, the former is the standard support vector, and the latter is the boundary support vector.

[0045] Furthermore, evaluation indicators are introduced to evaluate the prediction performance of the AI ​​agent model, and the evaluation indicators include mean absolute error, root mean square error and goodness of fit.

[0046] Furthermore, the calculation formula of the mean absolute error is as follows:

[0047]

[0048] Among them, MAE is the mean absolute error, P i is the power density value in the data set; P i ′ is the power density value predicted by the proxy model;

[0049] The calculation formula of the root mean square error is as follows:

[0050]

[0051] Among them, RMSE is the root mean square error, P i is the power density value in the data set; P i ′ is the power density value predicted by the proxy model;

[0052] The calculation formula of the goodness of fit is as follows:

[0053]

[0054] Among them, R 2 is the goodness of fit, P i is the power density value in the data set; P i ′ is the power density value predicted by the proxy model, is the average value of the power density in the data set.

[0055] Furthermore, the AI ​​agent model includes a GPR model, which calculates new observation data based on learning samples and assumptions, and outputs its mean as a predicted value. The calculation process is as follows:

[0056] Suppose N sets of learning samples are X = [x1, x2, ..., xN]T and y = [y1, y2, ..., yN]T, and y follows the prior distribution

[0057] y~N(μ y ,k yy )=N((μ y ,κ(X,X))),

[0058] Where: μ is the mean; κ is the kernel function, i.e., the covariance function;

[0059] Given a new prediction point X * , the purpose of GPR is to give the corresponding predicted value y*=f(X*), assuming that y* and y obey the joint Gaussian distribution:

[0060]

[0061] According to the marginal distribution property of the joint Gaussian distribution, we can get that y* obeys the conditional distribution:

[0062]

[0063] That is, the posterior distribution of the new prediction point y* given the known learning sample y is given, and the mean function of this distribution is is the predicted value of the surrogate model, denoted as:

[0064]

[0065] The beneficial effects produced by the technical solution of the present invention are as follows:

[0066] The present invention discloses a multi-objective optimization method for proton exchange membrane fuel cells based on artificial intelligence. The multi-objective optimization algorithm is introduced into the development of a fuel cell system performance model. 240 groups of data obtained by CFD simulation are used as data sets. First, two proxy models, support vector regression and Gaussian process regression, are built and trained to study the feasibility and advantages and disadvantages of the two in predicting battery power density in engineering practice. Then, a genetic algorithm is used to optimize the proxy models to obtain the maximum power density of each model and the corresponding optimal flow channel parameters. Further analysis is performed to provide engineering recommended ranges for the width, height, spine width and bending angle of the flow channel, thereby solving the problem that the proxy model optimized by a single variable method proposed in the background technology cannot obtain the optimal equilibrium solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0068] Figure 1 It is a flowchart of the overall steps of a multi-objective optimization method for proton exchange membrane fuel cells based on artificial intelligence;

[0069] Figure 2 is a specific step flow chart of step S1;

[0070] Figure 3 This is a specific step flow chart of step S3. DETAILED DESCRIPTION

[0071] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0072] Multi-objective optimization is relative to traditional single-objective optimization. How to achieve multi-objective optimization at the same time is a common problem in every field. Multi-objective optimization problems are usually described as vector optimization problems, which consist of multiple mutually influential objective functions. In many cases, the various objective functions conflict with each other. Often, the performance of one sub-objective is improved, while the performance of other objectives is weakened.

[0073] In the embodiment of the present application, the multi-objective optimization design problem can be described by the following formula:

[0074] min x F(x)=[f1(x), f2(x),..., f K (x)]

[0075] stg i (x)≥0, i∈[1,M]; h j (x)=0,j∈[1,L]

[0076] Among them, F(x) represents, f1(x) represents, f2(x) represents, f K (x) indicates that g i (x) indicates, i indicates, M indicates, h indicates j represents, j represents, and L represents.

[0077] Different from single-objective optimization problems, multi-objective optimization problems do not have a unique global optimal solution. It is precisely because it is impossible for every goal to be optimal at the same time, so the main contradiction must be grasped and given a greater weight.

[0078] Specifically, the present application discloses a multi-objective optimization method for proton exchange membrane fuel cells based on artificial intelligence, such as Figure 1 As shown, the following steps are included:

[0079] Step S1, collecting battery data and extracting data features using machine learning technology;

[0080] Specifically, Figure 2 As shown, step S1 also includes the following steps:

[0081] Step S101: Build and train an AI agent model;

[0082] Step S102: taking the proxy model as the fitness function, using a genetic algorithm to find the best solution, obtaining the maximum power density of the AI ​​proxy model and the corresponding flow channel structural parameters, and further analyzing and providing an engineering recommendation range for the flow channel structural parameters.

[0083] Specifically, evaluation indicators are introduced to evaluate the prediction performance of the AI ​​agent model, and the evaluation indicators include mean absolute error, root mean square error and goodness of fit.

[0084] Furthermore, in a preferred embodiment of the present application, the calculation formula of the mean absolute error is as follows:

[0085]

[0086] Among them, MAE is the mean absolute error, P i is the power density value in the data set; P i ′ is the power density value predicted by the proxy model;

[0087] The calculation formula of the root mean square error is as follows:

[0088]

[0089] Among them, RMSE is the root mean square error, P i is the power density value in the data set; P i ′ is the power density value predicted by the proxy model;

[0090] The calculation formula of the goodness of fit is as follows:

[0091]

[0092] Among them, R 2 is the goodness of fit, P i is the power density value in the data set; P i ′ is the power density value predicted by the proxy model, is the average value of the power density in the data set.

[0093] Specifically, in the embodiment of the present application, the training data must be preprocessed before model training to scale each variable to the same range to accelerate the convergence of the neural network. In this article, the data preprocessing method adopts the Z-score method.

[0094] The Z-score method is to standardize the original data through the following formula to make it satisfy the normal distribution:

[0095]

[0096] Where: is the average value of the data in the full range, X std Expressed as the standard deviation of the data over the full range.

[0097] The generalization ability of the model can be improved by adjusting the parameters. Finding the value of an important parameter of a model (the parameter that provides the best generalization performance) is a difficult task, but it is necessary for almost all models and data sets. There are also many studies on parameter tuning, such as non-parametric noise method estimation and genetic algorithm parameter optimization. The most commonly used method is grid search, which mainly refers to trying all possible combinations of the parameters we care about. In the SVR model of this article, we select the Gaussian kernel as the kernel function of the model, and the interval band half-width ε is taken as 0.01. The SVR Gaussian kernel model has two important parameters: the kernel width broadband coefficient β and the regularization parameter C. Among them, C is the penalty parameter, that is, the tolerance to errors. The higher C is, the less tolerant it is to errors and the easier it is to overfit. If C is small, it is easy to underfit. If C is too large or too small, the generalization ability will deteriorate. C and β jointly affect the generalization ability of SVR. The Gaussian kernel parameter changes the complexity of the sample feature subspace distribution by implicitly changing the mapping function, while the penalty parameter affects the generalization ability of SVR by controlling different regression errors. Therefore, we introduce the mean square error as an evaluation indicator:

[0098]

[0099] In order to get a better estimate of generalization performance, grid search is often used in conjunction with cross-validation. The biggest disadvantage of grid search with cross-validation is that it takes too much time to train the model. For the 5×4 parameter set of C∈[1,10,100,1000,10000] and β∈[0.01,0.1,1,10], a 5-fold cross-validation is performed, and each set of hyperparameters is calculated in turn. Finally, the optimal penalty parameter C and kernel width broadband coefficient β are obtained, which are 10000 and 0.1 respectively. They are used as the hyperparameters of the model to train the training set.

[0100] Step S2: Establish a battery model based on machine learning to predict the performance and health status of the battery, and use historical data to train the model;

[0101] Step S3, optimizing the design parameters using an optimization algorithm, combining multiple objective functions into a comprehensive objective function, and finding the optimal equilibrium solution through a multi-objective optimization algorithm;

[0102] Specifically, Figure 3 As shown, step S3 also includes the following steps:

[0103] Step S301, determine multiple objective functions to be optimized, select decision variables that affect the values ​​of the objective functions, and determine their value ranges and constraints;

[0104] Step S302: obtaining sample points, selecting multiple sample points within the determined range of design variable values ​​by an experimental design method, obtaining outputs corresponding to the sample points, and constructing sample points including a training set for the training model and a test set for the test model;

[0105] Step S303: train the model by constructing sample points, train the model by training set, and test the accuracy of the trained model by testing set, so as to determine the model;

[0106] Step S304: Use the database to train the model through a data-driven method, and test the effectiveness of the proxy model;

[0107] Step S305, searching the space through iterative calculation and optimization algorithm to find the approximate value of the optimal solution, and in each iteration, adjusting the value of the decision variable according to the value of the objective function and the algorithm rules until the stop condition is met;

[0108] Step S306: After obtaining a set of candidate solution sets, the candidate solution sets are screened. Specifically, a non-dominated sorting algorithm is used to divide the solution sets into different optimal levels, and the optimal solution set is selected.

[0109] Step S4: Apply the optimized battery design parameters and operation strategies to the actual battery system, conduct field verification, and optimize and adjust the optimization algorithm according to the experimental results.

[0110] In the embodiment of the present application, 240 sets of data obtained from the historical CFD simulation are used as the data set, and the parameters of the cathode flow channel width (0.9mm-1.5mm), depth (1.1mm-1.5mm), bending angle (0°~10°), and spine width are selected for optimization. The data set is randomly divided into training set: test set = 7:3, and the main goal is to improve the performance of the fuel cell, that is, the maximum output power.

[0111] Further, in a preferred embodiment of the present application, the AI ​​agent model includes an SVR model, and the SVR model is expressed as follows:

[0112]

[0113] The first inequality: stf(x i )-y i ≤ε+ξ i

[0114] The second inequality:

[0115]

[0116] Among them, ξ i and is a slack variable, indicating the distance that the points outside the error band exceed the error band; y is the predicted value vector, with a size of 1×N; X is the input matrix, with a size of D in ×N, the component form is [x1,x2,…,x N ]; ω is the coefficient vector, size D in ×1; b is the bias term; C is the regularization constant, ε is the deviation, z ε is the insensitive loss function of ε; is a regularization term used to measure the complexity of the surrogate model to prevent overfitting.

[0117] In the embodiment of the present application, two AI agent models of support vector regression and Gaussian process regression were first built and trained to study the feasibility and advantages and disadvantages of the two in predicting battery power density in engineering practice;

[0118] The surrogate model is used as the fitness function, and the genetic algorithm is used to find the optimal solution to obtain the maximum power density of the two surrogate models and the corresponding flow channel structure parameters.

[0119] The analysis gives the engineering recommended range of flow channel structural parameters.

[0120] Further, in a preferred embodiment of the present application, the dual form of the SVR model is:

[0121]

[0122]

[0123]

[0124] At the same time, the coefficient vector ω is:

[0125]

[0126] Where: α i and The Lagrange multipliers introduced by the first inequality and the second inequality are respectively solved by using quadratic programming.

[0127] Further, in a preferred embodiment of the present application, the dual form satisfies the KKT condition, and the KKT condition is:

[0128] α i [f(x i )-y i -(ε+ξ i )]=0

[0129]

[0130]

[0131]

[0132] in, Description i and At least one of them is zero, and the solution of SVR is expressed as:

[0133]

[0134] Therefore, the proxy model consists only of The sample points of are completely determined, and these sample points are called support vectors;

[0135] when When This sample point is within the error interval, is not a support vector, and has no effect on the proxy model; when 0<α i <C, or α i = 0, ξ i =0,f(x i )-y i =ε or f(x i )-y i =-ε, that is, the sample point falls exactly on the boundary of the error interval band, and this sample point is the standard support vector; when α i =C, or α i = 0, f(x i )-y i =ε+ξ i ≥ε or That is, the sample point is on the interval band or outside the boundary, the former is the standard support vector, and the latter is the boundary support vector.

[0136] The kernel function actually maps the sample from the original space to a higher-dimensional space, so that the linear regression effect of the sample in the high-dimensional space is better. The dot product is actually the simplest kernel function, that is, the linear kernel Other commonly used kernel functions in SVR include: polynomial kernel Gaussian kernel κ(x i ,x j ) = exp(-β||x i -x j ||2) and Sigmoid kernel wait.

[0137] SVR has many advantages, such as being very effective for high-dimensional data, and since the model only contains a subset of the training set, it has high memory resource utilization. In addition, the introduction of the kernel function makes the model more versatile. The disadvantages of SVR include not being able to directly provide the uncertainty of the prediction, usually requiring cross-validation, and having high computational complexity when the sample size is large.

[0138] In the embodiment of the present application, an example is given in which the output voltage and power density have a maximum power density in one polarization curve. Generally, the polarization curve and the maximum power density are the main factors for measuring the output performance of PEMFC. In the embodiment of the present application, the maximum power density is selected as the target to be optimized, which specifically includes the following steps:

[0139] Determine the research objectives and design variables, that is, determine the input and output of the proxy model. After the research objectives are selected, the factors that affect the research objectives are determined immediately, and then the factors that have the main impact on the research objectives are determined from multiple influencing factors as design variables. When the design variables are determined, the value range of the design variables needs to be determined, that is, the design domain is determined. The size of the design domain must meet the engineering design requirements;

[0140] Determine the sample points of the proxy model, that is, determine the training set and test set for constructing the proxy model. First, you need to obtain the sample points. Select enough sample points through experimental design methods within the determined range of design variable values. The distribution of sample points should cover the entire design domain as evenly as possible. Secondly, obtain the output corresponding to the sample points. The output of the sample points can be obtained through experiments or numerical proxy model methods. The constructed sample points include the training set for training the proxy model and the test set for testing the proxy model;

[0141] Determine the proxy model, that is, train the proxy model through the constructed sample points. Train the proxy model through the training set, and test the accuracy of the trained model through the test set. If the accuracy of the proxy model does not meet the engineering design requirements, on the one hand, the accuracy of the proxy model can be improved by increasing the sample points through experimental design, and on the other hand, the accuracy of different proxy models generated by the training set can be compared to select the best proxy model;

[0142] Use the database to train the surrogate model through a data-driven approach and test the effectiveness of the developed surrogate model. Well-validated surrogate models can replace physical models and effectively predict power density under new operating conditions;

[0143] The surrogate model and stochastic optimization algorithm are used to find the optimal parameter composition that maximizes the maximum power density.

[0144] For many practical problems, a single simulation experiment may take several minutes, hours or even days to complete. Therefore, for tasks such as design optimization, design space search, sensitivity analysis and what-if analysis, which require thousands or even millions of simulations, it will be impossible to solve the original model directly, while the model can quickly predict the experimental results. The model mainly deals with some engineering models that require millions of iterations. The model (also called approximate model, response surface model, etc.) is fast in calculation and can maintain the same calculation accuracy as the experimental data or simulation model, but the amount of calculation is very small. The model is established using a data-driven approach. The intermediate physical process is not concerned or even unknown, but the input-output behavior criteria of the model are very important. Usually a finite number of inputs and a finite number of outputs are selected to establish the model, and then the model is used to predict the variables of this physical model.

[0145] To solve this problem, the present application proposes to select sampling through sequential design, optimal experimental design (OED) or active learning to establish a proxy model and optimize model parameters, wherein the model parameters are weighed by deviation and variance to evaluate the accuracy of the model.

[0146] In the embodiments of the present application, the accuracy of the proxy models is also evaluated, and the proxy models include polynomial response surface method, kriging method, gradient enhanced kriging method (GEK), support vector machine, spatial mapping, and artificial neural network.

[0147] Furthermore, in a preferred embodiment of the present application, the AI ​​agent model includes a GPR model, which calculates new observation data based on learning samples and assumptions, and outputs its mean as a predicted value. The calculation process is as follows:

[0148] Suppose N sets of learning samples are X = [x1, x2, ..., xN]T and y = [y1, y2, ..., yN]T, and y follows the prior distribution

[0149] y~N(μ y ,k yy )=N((μ y ,κ(X,X))),

[0150] Where: μ is the mean; κ is the kernel function, i.e., the covariance function;

[0151] Given a new prediction point X * , the purpose of GPR is to give the corresponding predicted value y*=f(X*), assuming that y* and y obey the joint Gaussian distribution:

[0152]

[0153] According to the marginal distribution property of the joint Gaussian distribution, we can get that y* obeys the conditional distribution:

[0154]

[0155] That is, the posterior distribution of the new prediction point y* given the known learning sample y is given, and the mean function of this distribution is is the predicted value of the surrogate model, denoted as:

[0156]

[0157] In terms of model scalability, both GPR and SVR have more optional kernel functions, so their scalability is higher. However, compared with SVR, GPR has more advantages in this respect because it has more commonly used kernel functions. For example, its sine square kernel is very suitable for learning periodic data, and the white kernel has the function of estimating the noise level of data, etc., which are not available in SVR.

[0158] SVR cannot directly give the confidence level of the prediction, while GPR itself as a probability model can do this. In terms of model structure and form, the biggest advantage of SVR compared to GPR is the saving of computing resources. Although it consumes the most time in hyperparameter optimization, for the final model, it only uses a part of the data set, namely the support vector, while GPR needs to perform a large number of matrix calculations on the entire training set, so SVR has a clear advantage in this regard.

[0159] Taking into account the prediction performance, speed performance and intrinsic characteristics of the model, the following conclusions are drawn about the selection of surrogate models in fuel cell power density prediction: for small sample and low-dimensional data sets, GPR is the best model choice, while for large sample and high-dimensional data sets, the advantages of SVR are more obvious.

[0160] Specifically, in the embodiments of the present application, SVR and GPR both have good effects when used as proxy models to predict power density. The R2 on both the training set and the test set is greater than 0.995, and the prediction time is in seconds or minutes, indicating that they can be used as alternative models for predicting fuel cell power density in engineering practice.

[0161] When selecting a proxy model for fuel cell power density prediction, it is necessary to comprehensively consider the model's predictive performance, speed performance and intrinsic characteristics. Analysis of the two models shows that for small sample, low-dimensional data sets, GPR is the best model choice, while for large sample, high-dimensional data sets, the advantages of SVR are more obvious.

[0162] By combining the proxy model with the genetic algorithm optimization method, the embodiment of the present application narrowed the recommended ranges of the width, height, ridge width and bending angle of the optimal flow channel obtained by CFD simulation to w = 1.05mm ~ 1.2mm, d = 1.3mm ~ 1.5mm and θ tending to 0°, respectively. Within this range, the output power density of a single cell will not be less than 0.489W / cm2. This also shows that this method has good feasibility.

[0163] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0164] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0165] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0166] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0167] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, read-only memory), random access memory (RAM, random access memory), disk or optical disk and other media that can store program code.

[0168] In the description of the specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0169] The above contents are merely examples and explanations of the structure of the present invention. The technicians in this technical field may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the structure of the invention or exceed the scope defined by the claims, they should all fall within the protection scope of the present invention.

Claims

1. A multi-objective optimization method for cathode open air-cooled fuel cells based on artificial intelligence, characterized in that: The following steps are involved: Collect battery data and use machine learning technology to extract data features; Establish a machine learning-based battery model to predict battery performance and health status, and use historical data to train the model; Optimize design parameters using optimization algorithms, combine multiple objective functions into a comprehensive objective function, and find the optimal balance solution through multi-objective optimization algorithms; Apply the optimized battery design parameters and operating strategies to the actual battery system and conduct field verification, and optimize and adjust the optimization algorithm based on the experimental results.

2. The method for multi-objective optimization of cathode open air-cooled fuel cells based on artificial intelligence according to claim 1, characterized in that: The method of optimizing design parameters by using an optimization algorithm, combining multiple objective functions into a comprehensive objective function, and finding the optimal equilibrium solution by using a multi-objective optimization algorithm specifically includes the following steps: Determine multiple objective functions to be optimized, select decision variables that affect the value of the objective function, and determine their value range and constraints; Obtain sample points, select multiple sample points through experimental design methods within the determined range of design variable values, obtain the outputs corresponding to the sample points, and the constructed sample points include a training set for training the model and a test set for testing the model; The model is determined by constructing a model with sample points, training the model with a training set, and testing the accuracy of the trained model with a test set; Using the database to train the model through a data-driven approach and test the effectiveness of the proxy model; Through iterative calculation and optimization algorithm search space, the approximate value of the optimal solution is found. In each iteration, the value of the decision variable is adjusted according to the value of the objective function and the algorithm rules until the stopping condition is met; After obtaining a set of candidate solution sets, the candidate solution sets are screened. Specifically, a non-dominated sorting algorithm is used to divide the solution sets into different optimal levels, and the optimal solution set is selected.

3. The method for multi-objective optimization of cathode open air-cooled fuel cells based on artificial intelligence according to claim 2, characterized in that: The method of extracting data features by using machine learning technology specifically includes the following steps: The data set is randomly divided into training set: test set = 7:

3.

4. The method for multi-objective optimization of cathode open air-cooled fuel cells based on artificial intelligence according to claim 1, characterized in that: The establishment of a battery model based on machine learning specifically includes the following steps: Build and train AI agent models; The agent model is used as the fitness function and the genetic algorithm is used to optimize the model to obtain the maximum power density of the AI ​​agent model and the corresponding flow channel structure parameters. The engineering recommended range of the flow channel structure parameters is further analyzed and given.

5. The method for multi-objective optimization of cathode open air-cooled fuel cells based on artificial intelligence according to claim 4, characterized in that: The AI ​​agent model includes an SVR model, and the SVR model is expressed as follows: The first inequality: stf(x i )-y i ≤ε+ξ i The second inequality: Among them, ξ i and is a slack variable, indicating the distance that the points outside the error band exceed the error band; y is the predicted value vector, with a size of 1×N; X is the input matrix, with a size of D in ×N, the component form is [x1,x2,…,x N ]; ω is the coefficient vector, size D in ×1; b is the bias term; C is the regularization constant, ε is the deviation, z ε is the insensitive loss function of ε; is a regularization term used to measure the complexity of the surrogate model to prevent overfitting.

6. The method for multi-objective optimization of cathode open air-cooled fuel cells based on artificial intelligence according to claim 5, characterized in that: The dual form of the SVR model is: At the same time, the coefficient vector ω is: Where: α i and The Lagrange multipliers introduced by the first inequality and the second inequality are respectively solved by using quadratic programming.

7. The method for multi-objective optimization of cathode open air-cooled fuel cells based on artificial intelligence according to claim 6, characterized in that: The dual form satisfies the KKT condition, which is: a i [f(x i )-y i -(e+x) i )]=0 in, Description i and At least one of them is zero, and the solution of SVR is expressed as: Therefore, the proxy model consists only of The sample points of are completely determined, and these sample points are called support vectors; when When This sample point is within the error interval, is not a support vector, and has no effect on the proxy model; when 0<α i <C, or α i = 0, ξ i =0,f(x i )-y i =ε or f(x i )-y i =-ε, that is, the sample point falls exactly on the boundary of the error interval band, and this sample point is the standard support vector; when α i =C, or α i = 0, f(x i )-y i =ε+ξ i ≥ε or That is, the sample point is on the interval band or outside the boundary, the former is the standard support vector, and the latter is the boundary support vector.

8. The multi-objective optimization method for proton exchange membrane fuel cells based on artificial intelligence according to claim 4 is characterized in that: Evaluation indicators are introduced to evaluate the prediction performance of the AI ​​agent model, and the evaluation indicators include mean absolute error, root mean square error and goodness of fit.

9. The method for multi-objective optimization of cathode open air-cooled fuel cells based on artificial intelligence according to claim 8, characterized in that: The calculation formula of the mean absolute error is as follows: Among them, MAE is the mean absolute error, P i is the power density value in the data set; P i ′ is the power density value predicted by the proxy model; The calculation formula of the root mean square error is as follows: Among them, RMSE is the root mean square error, P i is the power density value in the data set; P i ′ is the power density value predicted by the proxy model; The calculation formula of the goodness of fit is as follows: Among them, R 2 is the goodness of fit, P i is the power density value in the data set; P i ′ is the power density value predicted by the proxy model, is the average value of the power density in the data set.

10. The method for multi-objective optimization of cathode open air-cooled fuel cells based on artificial intelligence according to claim 4, characterized in that: The AI ​​agent model includes a GPR model, which calculates new observation data based on learning samples and assumptions, and outputs its mean as a predicted value. The calculation process is as follows: Suppose N sets of learning samples are X = [x1, x2, ..., xN]T and y = [y1, y2, ..., yN]T, and y follows the prior distribution y~N(μ y ,k yy )=N((μ y ,κ(X,X))), Where: μ is the mean; κ is the kernel function, i.e., the covariance function; Given a new prediction point X * , the purpose of GPR is to give the corresponding predicted value y*=f(X*), assuming that y* and y obey the joint Gaussian distribution: According to the marginal distribution property of the joint Gaussian distribution, we can get that y* obeys the conditional distribution: That is, the posterior distribution of the new prediction point y* given the known learning sample y is given, and the mean function of this distribution is is the predicted value of the surrogate model, denoted as:

Images