Fuel cell runner-electrode coupling design and attenuation optimization method based on machine learning

Abstract

The invention discloses a fuel cell flow channel-electrode coupling design and attenuation optimization method based on machine learning, and the method employs a machine learning algorithm to construct a data driving model based on a fuel cell physical model, and achieves the rapid prediction and feature quantitative analysis of the performances of a cell under different flow field and electrode coupling structures. A gradient descent algorithm is further combined, multi-target and multi-parameter flow field structure optimization is carried out, key parameters in the reaction process are accurately regulated and controlled, and the stability and long-term performance of the fuel cell are improved. According to the method, coupling comprehensive optimization can be accurately and comprehensively carried out on the runner and electrode parameters, and a multi-target comprehensive optimization result which is difficult to realize by a traditional optimization method is obtained. By locally regulating and controlling flow field parameters and optimizing reactant concentration uniformity, substance exchange efficiency and temperature and humidity control, attenuation of the fuel cell is effectively improved, and the overall performance of the fuel cell is improved.

Description

Technical Field

[0001] This invention belongs to the field of fuel cell technology, and specifically relates to a method for fuel cell flow channel-electrode coupling design and attenuation optimization based on machine learning. Background Technology

[0002] Proton exchange membrane fuel cells (PEMFCs), as efficient and clean energy conversion devices, have attracted widespread attention due to their excellent energy density, compact structure, and zero emissions. PEMFCs convert hydrogen and oxygen into electrical energy through electrochemical reactions, and their internal working process involves multiple complex physical processes such as gas-liquid two-phase flow, heat and mass transfer, and electrochemical reactions. The flow-field coupled electrode, as one of the core components of the fuel cell, is the site of the electrochemical reaction, and its design directly affects the cell's performance, stability, and lifespan. Therefore, how to improve the efficiency and durability of fuel cells by optimizing the flow-field coupled electrode structure has become a key issue in fuel cell technology research.

[0003] In the design of flow-coupled electrodes for fuel cells, optimizing the structural parameters of these electrodes is a complex process. The flow-coupled electrodes involve coupling between multiple scales and multiple physical fields, with strong nonlinear relationships between parameters, making it difficult for traditional optimization methods to achieve ideal results. Typically, the optimization of flow-coupled electrode parameters employs a single-variable method, which fixes some parameters and studies their impact on battery performance by changing each parameter one by one. While this method can yield some results in simple cases, it performs poorly when dealing with complex nonlinear problems and can only obtain local optima, making it difficult to effectively analyze the global solution space. Therefore, the single-variable method faces significant limitations when optimizing the complex structure of flow-coupled electrodes.

[0004] Currently, although numerous studies based on optimization algorithms have been applied to the optimization of flow-field coupled electrode structures, most focus on optimizing the peak power density of the fuel cell, often neglecting the multi-dimensional and multi-objective optimization of fuel cell performance. For example, in addition to power density, factors such as reactant concentration distribution, velocity field within the electrode, and temperature and humidity field also play a crucial role in the performance and lifespan of the fuel cell. Single-objective optimization methods are insufficient to comprehensively improve the overall performance of the fuel cell. Furthermore, many studies have not performed feature screening of optimization parameters, causing the optimization algorithm to be interfered with by fluctuations in low-sensitivity parameters during processing, resulting in slower convergence speed, low optimization efficiency, and thus limiting its practical application.

[0005] To overcome the aforementioned problems, researchers have begun to introduce machine learning techniques to more accurately optimize the flow-field coupled electrode structure of fuel cells using data-driven models. This approach utilizes a large amount of simulation data generated through high-fidelity physical models to train a neural network model, thereby predicting the impact of different flow-field coupled electrode structure parameters (such as fin height and fin width) on battery performance. Through model optimization, it is possible to quickly predict battery performance under complex multiphysics coupling conditions with different parameters and effectively optimize these parameters to achieve the optimal design for overall battery performance.

[0006] However, despite the great potential of machine learning methods, current research faces several challenges. First, training machine learning models requires a large amount of high-quality training data, which necessitates high-performance simulation models and sufficient computational resources. Second, the "black box" nature of machine learning models results in relatively poor interpretability, potentially posing risks in practical engineering applications. Finally, how to combine optimization algorithms with data-driven models for global optimization, improve optimization efficiency, and balance different metrics in multi-objective optimization remains a pressing issue.

[0007] Therefore, the deep integration of physical modeling and machine learning, especially neural network-based models, can provide more accurate optimization results even under conditions of strong nonlinear relationships and mutual coupling between parameters. Applying this method to the design of flow field coupled electrodes for fuel cells will help overcome the limitations of traditional optimization methods, promote the further development of fuel cell technology, and provide new ideas and solutions for achieving more efficient, stable, and durable fuel cells. Summary of the Invention

[0008] To overcome the shortcomings of the prior art, the present invention aims to provide a machine learning-based method for fuel cell flow channel-electrode coupling design and degradation optimization. This method addresses issues such as fuel product mixing and uneven fuel concentration distribution during the flow reaction of proton exchange membrane fuel cells, which lead to decreased battery performance and reduced battery life. The method constructs a neural network model, combines it with a physical model and data-driven methods, and utilizes optimization algorithms to optimize key structural parameters of the catalyst layer. This improves fuel cell performance and durability, shortens the design cycle, and reduces computational costs.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] A machine learning-based method for fuel cell flow channel-electrode coupling design and degradation optimization includes the following steps:

[0011] S1: Build a physical model of the fuel cell flow field coupled with the electrode, obtain high-dimensional physical field data through physical model simulation, and form a sample dataset together with the fuel cell performance indicators and flow channel design parameters. The physical field data includes the pressure drop of the flow field, reactant concentration, velocity field inside the electrode, and temperature and humidity field.

[0012] S2: Perform data preprocessing;

[0013] S3: A battery performance prediction model is built based on a neural network. The flow channel design parameters of the fuel cell are used as inputs, the performance index of the fuel cell is used as outputs, the target prediction accuracy is set, and the model is trained using a preprocessed sample dataset.

[0014] S4: Set the numerical range and minimum allowable change of the optimization design parameters, and iteratively optimize through the gradient descent optimization algorithm to obtain the optimal combination of design parameters in order to optimize the fuel cell performance.

[0015] In one embodiment, the design parameters are the rib height and rib width of the fuel cell flow channel, and the performance indicators include concentration distribution uniformity and velocity and temperature / humidity distribution within the electrodes.

[0016] In one embodiment, the preprocessing includes normalization, noise reduction, and feature selection; the feature selection refers to screening out key features closely related to fuel cell performance from the original high-dimensional data, including the following steps:

[0017] Feature selection: By calculating the correlation between each input feature and the performance index, design parameters that have an impact on battery performance greater than a set threshold are selected;

[0018] Dimensionality reduction: By removing redundant or noisy features, the dimensionality of the model input is reduced.

[0019] In one embodiment, S3 involves constructing a battery performance prediction model using a U-Net neural network. The U-Net neural network includes an input layer, an output layer, and multiple hidden layers. The input layer takes in the channel design parameters, and the output layer outputs the predicted performance index values. The hyperparameters of the neural network are then tuned, and the model is trained using a preprocessed sample dataset.

[0020] In one embodiment, step S3 measures the model's prediction accuracy by calculating the coefficient of determination between the predicted and actual values. The formula for calculating the coefficient of determination is:

[0021]

[0022] Among them, R 2 As the coefficient of determination, y i This is the actual value. For predicted values, R is the mean of the actual values, n is the sample size, and R0 is the mean of the actual values. 2 The value ranges from 0 to 1, where the closer the value is to 1, the stronger the predictive ability of the model, and the closer the value is to 0, the weaker the predictive ability of the model.

[0023] In one embodiment, adjusting the training data includes removing outliers, increasing the amount of data, and data augmentation.

[0024] In one embodiment, the optimization method for step S4 is as follows:

[0025] The numerical range and the minimum allowable change are used as the input boundary conditions for the optimization algorithm; the gradient descent optimization algorithm is initialized, and the flow channel design parameters are optimized step by step. Within the set maximum number of iterations, the optimization algorithm iterates until the optimal parameter value is obtained, thereby realizing the optimal design of the flow field coupled electrode. The optimal design is the electrode porosity obtained under the flow channel design parameters.

[0026] In one embodiment, step S4 incorporates a coupling constraint between compression and porosity during the optimization process, ensuring that the mapping ε(h) between the rib height h and the catalyst layer porosity ε remains within [ε]. min ,ε max The interval is defined, and the minimum feature size is constrained to meet the requirements, where ε min , ε max These represent the minimum and maximum values ​​of the catalytic layer porosity ε, respectively.

[0027] In one embodiment, S4, during the optimization process, utilizes the pressure drop difference Δp between adjacent flow channels. adj The intensity of convection under the ribs is regulated and introduced as a bonus to reduce the risk of accelerated decay caused by concentration polarization. At the same time, an upper limit is set on the total pressure drop Δp to avoid excessive pump power and structural fatigue.

[0028] In one embodiment, in step S4, during the optimization process, the carbon corrosion risk index CCRI and the membrane chemical degradation risk index MCDI are introduced. CCRI is measured by the spatial integral of the temperature-oxygen concentration-velocity coupled field in the sub-rib region, and MCDI is measured by the region with high temperature-high oxygen and low water content. Both are added as penalty terms to the multi-objective optimization function.

[0029] By introducing constraint or guidance terms during the S4 optimization process, the optimization results can not only satisfy the improvement of performance, but also the effects of low voltage drop and longer life.

[0030] Compared with existing technologies, the advantages of this invention lie in that, by combining machine learning and physical models, it not only improves the efficiency of fuel cell flow channel-electrode coupling design but also provides a more scientific optimization basis for improving fuel cell performance, overcoming the shortcomings of traditional optimization methods such as large computational load, slow convergence, and lack of global optimization capabilities. Furthermore, the model of this invention has good scalability and can be applied to the flow channel-electrode coupling design of different types of fuel cells, providing strong technical support for the future industrial application of fuel cells. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the process of the present invention.

[0032] Figure 2 This is a schematic diagram of the fuel cell flow channel-electrode coupling design of the present invention.

[0033] In the figure: 1-Anode flow field plate, 2-Anode current collector plate, 3-Anode rib, 4-Anode electrode, 5-Anode catalyst layer, 6-Exchange membrane, 7-Cathode catalyst layer, 8-Cathode electrode, 9-Cathode rib, 10-Cathode current collector plate, 11-Cathode flow field plate.

[0034] Figure 3 This is a comparison chart of polarization curves before and after machine learning optimization in this invention.

[0035] Figure 4 This is a voltage decay curve before and after machine learning optimization according to the present invention.

[0036] Figure 5 This is a distribution diagram of reactant concentrations before and after machine learning optimization according to the present invention. Detailed Implementation

[0037] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings and examples.

[0038] Proton exchange membrane fuel cells (PEMFCs) are prone to problems such as fuel product mixing and uneven concentration distribution during fuel flow and reaction, which may lead to performance degradation and battery attenuation. To address this, this invention proposes a machine learning-based method for fuel cell flow channel-electrode coupling design and attenuation optimization. This method generates a large dataset by establishing a high-fidelity fuel cell physical model and trains the dataset using a neural network. This enables a nonlinear mapping between the flow channel-electrode coupling structure and battery performance, as well as rapid prediction and quantitative analysis of battery performance under different flow fields and electrode coupling structures. Specifically, this invention first generates multi-physical field data, including reactant concentration, electrode velocity field, and temperature and humidity field, under different fin heights and widths, using a physical model and experimental data. Through normalization and data preprocessing of the input samples, characteristic parameters closely related to fuel cell performance are extracted. These parameters include fin height, fin width, and other design parameters affecting the flow channel-electrode coupling efficiency. Based on this, a U-Net neural network model is used to train the input samples, establishing a mapping relationship between the output fields, such as reactant concentration and electrode velocity field, and the catalyst layer structure. Furthermore, by combining gradient descent algorithm, multi-objective and multi-parameter flow field structure optimization is performed to precisely control key parameters in the reaction process, thereby improving the stability and long-term performance of the fuel cell. This method can accurately and comprehensively optimize the coupled flow channel and electrode parameters, achieving multi-objective comprehensive optimization results that are difficult to achieve with traditional optimization methods. By locally controlling flow field parameters, the uniformity of reactant concentration, mass exchange efficiency, and temperature and humidity control are optimized, thereby effectively improving fuel cell degradation and enhancing the overall performance of the fuel cell. Ultimately, this invention yields a high-performance fuel cell flow field-electrode coupling structure with uniform fuel concentration distribution and long lifespan, providing a strong guarantee for the long-term operation of the fuel cell.

[0039] like Figure 1 As shown, the present invention mainly includes the following steps:

[0040] Step 1: Build a physical model of the fuel cell flow field and electrode coupling to obtain the data driving source.

[0041] This invention first constructs a two-dimensional mathematical and physical model of flow-channel electrode coupling based on the multi-physical coupling process of the internal flow field and electrodes of a fuel cell. This model involves the coupled calculation of multiple physical fields, including gas-liquid two-phase flow, heat transfer, mass transfer, and electrochemical reactions. The model describes the flow field and electrode structure within the fuel cell using physical laws such as the mass conservation equation, momentum conservation equation, energy conservation equation, and proton conservation equation. Through physical model simulation and discretization using the finite element method, the battery output polarization curve and related high-dimensional physical field data (mainly including pressure drop in the flow field, reactant concentration distribution, velocity field within the electrodes, and temperature and humidity field) were obtained. The model results were verified by experimental data to ensure their accuracy, providing a reliable physical basis for the subsequent development of data-driven models.

[0042] Figure 2 The detailed structure of the fuel cell is shown, along with the optimized cell flow parameters (fin height and fin width) of this invention. The influence of fin height and fin width on electrode parameters (porosity) is also revealed. The physical model of this invention is based on this structure. In the fuel cell, the anode flow field plate 1 and the cathode flow field plate 11 are symmetrically arranged, with an exchange membrane 6 in the center. On either side of the exchange membrane 6 are the anode catalyst layer 5 and the cathode catalyst layer 7, respectively. Anode current collectors 2 are arranged at intervals on the side of the anode flow field plate 1 closest to the anode catalyst layer 5. Anode ribs 3 are located between adjacent anode current collectors 2, and anode electrodes 4 are arranged between the anode ribs 3 and the anode catalyst layer 5. The cathode flow field plate 11 correspondingly consists of a cathode electrode 8, cathode ribs 9, and a cathode current collector 10.

[0043] Step 2: Select the flow channel structure parameters and battery performance targets to generate a training dataset.

[0044] Based on actual production and design needs, this embodiment selects key design parameters affecting fuel cell performance as input features for the neural network model. Specifically, the rib height and rib width of the flow channel are selected as input parameters to describe the geometric structure of the flow channel. By simulating the physical model under different rib height and rib width parameters, the relationship between the flow field structure and battery performance indicators (such as concentration distribution uniformity and velocity, temperature, and humidity distribution within the electrodes) is obtained, which, together with the aforementioned physical field data, constitutes a sample dataset. In this dataset, the input features include the rib height and rib width of the flow channel, while the output targets are parameters such as the reactant concentration distribution, velocity field, and temperature and humidity field of the battery.

[0045] Step 3: Data preprocessing.

[0046] After the dataset is generated, standard preprocessing methods are used to normalize it, mapping all input feature and output target values ​​to the range of 0-1, eliminating differences between different orders of magnitude. Subsequently, noise reduction and feature selection are performed. Feature selection refers to filtering key features closely related to fuel cell performance from the original high-dimensional data to improve model training efficiency and prediction accuracy. Specifically, feature selection includes the following steps:

[0047] Feature selection: By calculating the correlation between each input feature and the performance index, design parameters that have an impact on battery performance greater than a set threshold are selected;

[0048] Dimensionality reduction: By removing redundant or noisy features, the dimensionality of the model input is reduced, thereby improving model training efficiency and avoiding overfitting.

[0049] To further improve optimization efficiency and avoid ineffective computation, this invention employs a feature selection method based on physical consistency and maximum mutual information coefficient. The correlation between input parameters and performance indicators is evaluated using the Pearson correlation coefficient and the maximum mutual information coefficient. This method accurately identifies key parameters that significantly impact battery flow channel-electrode coupling performance, reducing redundant variables and thus optimizing computational efficiency. Combined with a data-driven model, this invention enables precise optimization under multi-dimensional and multi-objective conditions, overcoming the limitations of traditional single-objective optimization methods.

[0050] The preprocessed dataset was divided into training and testing sets in a ratio of 0.75:0.25, which were used for training and validation of the neural network model, respectively.

[0051] Step 4: Construct a battery performance prediction model based on a neural network, set a target prediction accuracy, train the model using a preprocessed sample dataset, and perform battery performance prediction and sensitivity analysis based on the trained model.

[0052] This invention employs a U-Net neural network to construct a battery performance prediction model. The U-Net neural network is suitable for processing spatially correlated input data, such as flow field data and electrode structure data. It includes an input layer, an output layer, and multiple hidden layers. The input layer takes in the flow channel design parameters, and the output layer outputs predicted performance indicators. The hyperparameters of the neural network are optimized, and it is trained using a pre-processed sample dataset.

[0053] During training, the network structure (such as the number of hidden layers, number of nodes, activation function, etc.) and hyperparameters (such as learning rate, regularization coefficient, etc.) are adjusted to ensure the model has high prediction accuracy. Cross-validation is used during training to further improve the model's robustness.

[0054] After training, the neural network model is validated using test set data, and the prediction accuracy is calculated. The model's performance is evaluated by comparing the error between the predicted and actual values. Specifically, in this step, the coefficient of determination between the predicted and actual values ​​is calculated to measure the model's prediction accuracy. The formula for calculating the coefficient of determination is:

[0055]

[0056] Among them, R 2 As the coefficient of determination, y i This is the actual value. For predicted values, R is the mean of the actual values, n is the sample size, and R0 is the mean of the actual values. 2 The value ranges from 0 to 1, where the closer the value is to 1, the stronger the predictive ability of the model, and the closer the value is to 0, the weaker the predictive ability of the model.

[0057] In this embodiment, the target prediction accuracy is set to R. 2 The threshold is 0.98, when the calculated R... 2 When the value is greater than or equal to this threshold, the model prediction accuracy is considered to meet the requirements, and the resulting battery performance prediction model is the final prediction model, ensuring the reliability of the model in practical applications. When R 2 If the value is less than this threshold, the prediction accuracy is considered insufficient, and it is necessary to continue to optimize the model or adjust the training data (i.e., remove outliers, increase the amount of data, and perform data augmentation).

[0058] The specific methods for adjusting training data are as follows:

[0059] Outlier removal: This involves analyzing abnormal data points in the training dataset and removing outliers that do not contribute to model training or may cause model instability. The presence of outliers can negatively impact model training performance, thereby reducing prediction accuracy.

[0060] Increase the amount of data: By adding more training samples, especially those covering a broad design space, the model can learn a more comprehensive relationship between features and performance. More training data can improve the model's generalization ability and reduce overfitting.

[0061] Data augmentation: Data augmentation techniques (such as making small perturbations to existing data or generating new samples) increase the diversity of data. This can help neural network models avoid local optima and improve their adaptability to various input conditions.

[0062] Meanwhile, the trained model of this invention can be used for sensitivity analysis of input features to quantify the impact of various design parameters on battery performance. For example, the influence of fin height and fin width on reactant concentration, the influence of fin height on the velocity field within the electrode, and the influence of fin width on temperature and humidity control. Through sensitivity analysis, parameters with a significant impact on battery performance can be further identified, providing a basis for subsequent optimization.

[0063] Step 4: Perform gradient descent optimization based on the optimization objective.

[0064] The numerical range and minimum allowable changes of the optimization design parameters are defined, and the gradient descent optimization algorithm is used to iteratively optimize the fin height and fin width of the flow channel. The optimization goal is to achieve a more uniform distribution of reactant concentration, a larger mean velocity field, and more reasonable control of the temperature and humidity field, thereby improving the battery performance and extending its service life. Based on the set optimization goals, the gradient descent algorithm iteratively calculates the optimal combination of fin height and fin width parameters, ultimately achieving the globally optimal design for fuel cell performance.

[0065] Specifically, in this step, the numerical range of the optimized design parameters and the minimum allowable change are used as the input boundary conditions of the optimization algorithm; the gradient descent optimization algorithm is initialized, and the flow channel design parameters are optimized step by step. Within the set maximum number of iterations, the optimization algorithm iterates until the optimal parameter value is obtained, thereby achieving the optimal design of the flow field coupled electrode, that is, the optimal electrode porosity obtained under the flow channel design parameters.

[0066] Furthermore, a coupling constraint between compression and porosity is incorporated into the optimization process to maintain the mapping ε(h) between the rib height h and the catalyst layer porosity ε within [ε]. min ,ε max The interval is defined, and the minimum feature size is constrained to meet the requirements, where ε min , ε max These represent the minimum and maximum values ​​of the catalytic layer porosity ε, respectively.

[0067] Here, the minimum characteristic dimension refers to the minimum dimension of the ribs in the fuel cell flow channel, specifically the lower limit of geometric parameters such as rib height or rib width. As a component of the flow channel-electrode coupling structure, the size of the ribs affects the performance of the fuel cell, including fluid flow, mass transfer, and reaction efficiency. Since there is a mapping relationship between rib height and catalyst layer porosity, limiting the minimum dimension of the rib height can ensure the effectiveness and stability of the structure.

[0068] Furthermore, to further improve fuel cell performance and reduce degradation, the pressure drop difference Δp between adjacent flow channels is used during the optimization process. adjAdjusting the under-fin convection intensity and introducing it as a bonus item to further optimize battery design parameters (fin height, fin width) to reduce the risk of accelerated degradation caused by concentration polarization. At the same time, setting an upper limit on the total pressure drop Δp to avoid excessive pump power and structural fatigue.

[0069] Furthermore, the carbon corrosion risk index CCRI and the membrane chemical degradation risk index MCDI are introduced into the optimization. CCRI is measured by the spatial integral of the temperature-oxygen concentration-velocity coupled field in the under-rib region, while MCDI is measured by the region with high temperature-high oxygen and low water content. Both are added as penalty terms to the multi-objective optimization function.

[0070] In the multi-objective optimization function, CCRI and MCDI are added as penalty terms to guide the optimization process and avoid excessive corrosion and film degradation risks. The optimization objective can be expressed in the following form:

[0071] Objective Function = α·P performance -β·CCRI-γ·MCDI

[0072] P performance These are the performance indicators of the fuel cell (such as power density, concentration uniformity, etc.); α, β, and γ are weighting coefficients that control the relative importance of performance, carbon corrosion, and membrane degradation risk, respectively.

[0073] Constraints:

[0074] CCRI≤CCRI max

[0075] MCDI≤MCDI max

[0076] Δp≤Δp max

[0077] h min ≤h≤h max

[0078] Among them, CCRI max This is the maximum tolerable risk of carbon corrosion. MCDI max This is the maximum tolerable risk of membrane degradation. Δp max It is the set maximum tolerable voltage drop h min and h max These are the minimum and maximum size limits for rib height, respectively.

[0079] This invention achieves globally optimal performance by iteratively optimizing key parameters using gradient descent or other optimization algorithms during the optimization process. Optimization objectives include improving the uniformity of reactant concentration, enhancing mass exchange efficiency within the electrodes, and rationally controlling the temperature and humidity field to extend battery life. By optimizing design parameters such as fin height and fin width, the pressure drop and velocity distribution in the flow field are improved, ultimately resulting in enhanced fuel cell performance and extended lifespan.

[0080] While ensuring battery stability, this invention comprehensively considers various performance indicators to avoid performance imbalances that may result from optimizing a single objective. By precisely controlling the design parameter range and combining it with an appropriate number of iterations, the optimal structure of the fuel cell flow channel-electrode coupling is finally obtained.

[0081] Step 5: Optimize Result Validation and Practical Application

[0082] The optimized rib height and rib width parameters were incorporated into the physical model for verification, ensuring that the optimized design could provide better performance in practical applications. By applying the optimized structural parameters to the design of the fuel cell, significant performance improvements were achieved, especially under high current density conditions, where concentration polarization losses were significantly reduced, and the discharge capacity and thermal uniformity of the cell were improved.

[0083] Figure 3 The polarization and power curves of the fuel cell before and after optimization using the method of this invention are shown. It can be seen that the optimized battery performance is significantly improved compared to the unoptimized version.

[0084] Figure 4 The degradation curves of the fuel cell before and after optimization using the method of this invention are shown. The current density is set at 2 A / cm². 2 By observing the voltage change over time, it can be seen that the decay rate after optimization is slower than that before optimization.

[0085] Figure 5 The battery parameters before and after optimization are set in the same fuel cell physical model for comparative analysis. The two-dimensional current density distribution obtained by the physical model simulation shows that the optimized current density distribution is more uniform and the concentration polarization is reduced.

[0086] Therefore, the method of this invention can effectively achieve multi-objective optimization of the fuel cell flow channel-electrode structure, overcome the limitations of traditional optimization methods, and provide new ideas and technical support for the design and optimization of fuel cells.

Claims

1. A machine learning based fuel cell flow channel-electrode coupling design and decay optimization method, characterized in that, The method comprises the following steps: S1: a physical model of coupling of a fuel cell flow field and an electrode is built, high-dimensional physical field data are obtained through simulation of the physical model, and the physical field data, performance indicators of the fuel cell and flow channel design parameters are collectively used to form a sample data set, wherein the physical field data include pressure drop, reactant concentration, velocity field and temperature and humidity field in the electrode of the flow field; S2: data preprocessing is performed; S3: a battery performance prediction model is constructed based on a neural network, the flow channel design parameters of the fuel cell are used as input, the performance indicators of the fuel cell are used as output, a target prediction accuracy is set, and the preprocessed sample data set is used for training; S4: a numerical range and a minimum change amount of an optimized design parameter are set, a gradient descent optimization algorithm is iteratively optimized, and an optimal design parameter combination is obtained to optimize the performance of the fuel cell.

2. The method of claim 1, wherein, The design parameters are rib height and rib width of the flow channel of the fuel cell, and the performance indicators include concentration distribution uniformity, velocity in the electrode and temperature and humidity distribution. 3.The method of claim 1 or 2, wherein, The preprocessing includes normalization processing, noise reduction and feature selection; The feature selection refers to screening key features closely related to the performance of the fuel cell from original high-dimensional data, and comprises the following steps: Feature screening: the correlation between each input feature and the performance indicator is calculated, and the design parameter whose influence on the performance of the battery is greater than a set threshold is selected; Dimension reduction: the dimension of the model input is reduced by removing redundant features or noise features.

4. The method of claim 1, wherein, In S3, a U-Net neural network is used to construct the battery performance prediction model, the U-Net neural network comprises an input layer, an output layer and a plurality of hidden layers, wherein the input layer inputs the flow channel design parameters, and the output layer outputs the predicted value of the performance indicator; the hyperparameters of the neural network are optimized, and the preprocessed sample data set is used for training.

5. The method of claim 1, wherein, In S3, the prediction accuracy of the model is measured by calculating the determination coefficient between the predicted value and the actual value, and the calculation formula of the determination coefficient is: where R 2 is the coefficient of determination, y i is the actual value, is the predicted value, is the mean of the actual values, n is the number of samples, and R 2 is the value of the coefficient of determination. The value of R 2 ranges from 0 to 1, where a value closer to 1 indicates a stronger predictive ability of the model, and a value closer to 0 indicates a weaker predictive ability of the model.

6. The method of claim 5, wherein the method further comprises: The training data is adjusted, including removing outliers, increasing data volume and data enhancement.

7. The method of claim 1, wherein the method further comprises: determining a plurality of flow channel-electrode coupling designs; and determining a plurality of flow channel-electrode coupling designs that are optimized for a plurality of fuel cell stack performance metrics. In S4, the optimization method is as follows: The numerical range and the minimum change amount are used as input boundary conditions of the optimization algorithm; the gradient descent optimization algorithm is initialized, the flow channel design parameters are optimized step by step, and the optimization algorithm is iteratively calculated within a set maximum number of iterations until the optimal parameter value is obtained, so that the optimal design of the flow field coupled with the electrode is realized, and the optimal design is the optimal electrode porosity obtained under the flow channel design parameters. 8.The method of claim 1 or 7, wherein, The S4, in the optimization process, adds the coupling constraint of compression and porosity, makes the mapping of the rib height h and the catalyst layer porosity ε, ε(h), remain in the interval [ε min ,ε max ], and limits the minimum feature size to meet the requirements, where ε min and ε max are the minimum value and the maximum value of the catalyst layer porosity ε, respectively. 9.The method of claim 1 or 7, wherein, S4, in the optimization process, through the adjacent flow channel pressure drop difference Δp adj The sub-rib convection intensity is regulated, and the sub-rib convection intensity is introduced as a reward term to reduce the risk of accelerated decay caused by concentration polarization, and an upper limit of the total pressure drop Δp is set to avoid excessive pump power and structural fatigue. 10.The method of claim 1 or 7, wherein, In S4, during the optimization process, a carbon corrosion risk index CCRI and a membrane chemical degradation risk index MCDI are introduced, wherein the CCRI is measured by the spatial integral of the temperature-oxygen concentration-velocity coupling field in the rib under area, and the MCDI is measured by the area with high temperature, high oxygen and low water content; Both of them are added as penalty terms to the multi-objective optimization function.

Images