PEMFC (proton exchange membrane fuel cell) high-current density performance prediction method, system, equipment and medium

Abstract

The invention relates to a proton exchange membrane fuel cell (PEMFC) high current density performance prediction method, system, equipment and medium. The method comprises the following steps: establishing a multi-physical field coupling model which comprehensively considers complex processes such as electrochemical reaction, proton conduction, gas diffusion and heat transfer, and describing the change of each physical quantity by adopting a partial differential equation based on a basic physical law; performing grid division and numerical discretization on the proton exchange membrane fuel cell model; selecting model parameters, and verifying the model through experimental data of different working conditions; inputting actual working condition parameters into a model to predict performance, and analyzing a simulation result; using a convolutional neural network, a recurrent neural network and an auto-encoder to extract features from different types of data and fuse the features to form a comprehensive feature vector; a deep neural network prediction model is constructed, and a cross entropy loss function and an Adam optimizer are adopted for training; dropout, L1 and L2 regularization, k-fold cross validation and transfer learning are utilized to optimize the model, and the generalization ability is improved; the system, the equipment and the medium realize high current density performance prediction of the proton exchange membrane fuel cell (PEMFC) based on the method; the prediction precision is improved, the experiment cost is reduced, the internal mechanism can be deeply understood, and powerful support is provided for design optimization, operation management and fault diagnosis of the fuel cell.

Description

Technical Field

[0001] The present invention relates to the technical field of proton exchange membrane fuel cells, and particularly relates to a method, a system, a device and a medium for predicting the high current density performance of a proton exchange membrane fuel cell (PEMFC). Background Art

[0002] As a highly efficient and clean energy conversion device, the proton exchange membrane fuel cell (PEMFC) has broad application prospects in the new energy field. With the continuous improvement of its performance requirements, achieving stable and efficient operation at high current density (high current density) has become the key. However, the physical and chemical processes inside the fuel cell at high current density are extremely complex, and various factors are coupled with each other, posing a huge challenge to performance prediction. Traditional prediction methods based on empirical models and semi-empirical models [ARIFM, CHEUNG S C P, ANDREWS J. Different Approaches Used for Modeling and Simulation of Polymer Electrolyte Membrane Fuel Cells: A Review [J]. Energy & Fuels, 2020, 34(10): 11897-11915.], [JIAO K, NI M. Challenges and opportunities in modelling of proton exchange membrane fuel cells (PEMFC) [J]. International Journal of Energy Research, 2017, 41(13): 1793-1797.] are difficult to accurately describe the complex electrochemical reactions, mass and heat transfer, and multiphase flow processes inside the fuel cell at high current density. These methods often ignore the nonlinear interactions between various factors, resulting in a large deviation between the prediction results and the actual situation. Obtaining fuel cell performance data at high current density through experimental tests [ZENG L, DONG D, LU J, et al. Hyperbranched Interpenetrating Hydrogen Bond Network (HIHBN) Proton Exchange Membrane for Fuel Cells Above 220℃ [J]. Advanced Functional Materials, 2025.] not only requires a large amount of time, manpower, and material resources, but also the precise control of experimental conditions is difficult, and the data under different experimental conditions are difficult to directly compare and generalize. There are also many technical problems in measuring the internal state of the fuel cell during the experiment, which limits the in-depth understanding of its internal mechanism. At high current density, the interaction of multiple physical fields such as electricity, heat, mass, and force is involved inside the fuel cell.Most of the existing numerical simulation methods [SONGG-H, MENG H. Numerical modeling and simulation of PEM fuel cells: Progress and perspective [J]. Acta Mechanica Sinica, 2013, 29(3): 318-334.] only consider the influence of partial physical fields, fail to establish a comprehensive and accurate multi-physical field coupling model, and cannot fully reflect the true performance of fuel cells at high current densities. In addition, although the electrical performance data, gas parameter data, temperature data, etc. generated during the operation process are rich, the complex correlations between the data are difficult to be fully explored, and the existing prediction technologies cannot effectively integrate multi-source data, resulting in limited prediction accuracy. Summary of the Invention

[0003] In order to overcome the defects existing in the above-mentioned prior art, the purpose of the present invention is to provide a method, system, device and medium for predicting the performance of a proton exchange membrane fuel cell (PEMFC) at high current densities, a performance prediction technology for a proton exchange membrane fuel cell at high current densities based on multi-modal data fusion and deep learning, to solve the problems of low prediction accuracy, high experimental cost, lack of a comprehensive multi-physical field coupling model, prediction accuracy, data utilization and working condition adaptability in the prior art, and to achieve accurate and efficient prediction of the performance of a proton exchange membrane fuel cell at high current densities.

[0004] In order to achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0005] A method for predicting and optimizing the performance of a proton exchange membrane fuel cell (PEMFC) at high current densities, comprising the following steps:

[0006] Step 1, establish a multi-physical field coupling model of a proton exchange membrane fuel cell (PEMFC);

[0007] Step 2, first perform mesh division on the three-dimensional geometric structure of the fuel cell, then numerically discretize the multi-physical field coupling model established in Step 1 using the finite volume method at the mesh nodes, and solve the discretized equations using an iterative algorithm, and use the SIMPLE algorithm to solve the pressure-velocity coupling equation to obtain a preliminary simulation result of the multi-physical field coupling model;

[0008] Step 3, compare the simulation result of the multi-physical field coupling model obtained in Step 2 with the experimental data, optimize the model parameters in combination with the particle swarm optimization algorithm, and verify the accuracy of the model through multi-condition experiments;

[0009] Step 4: Using the optimized model verified in Step 3, input the actual operating condition parameters to predict the battery performance at high current density, including output voltage, power density, current distribution, temperature distribution, and water content distribution;

[0010] Step 5: Based on the high current density performance prediction results obtained in Step 4, construct a deep learning prediction model to further improve the prediction accuracy;

[0011] Step 6: Optimize and generalize the deep learning model established in Step 5, and output an improved optimization plan to ensure the applicability and reliability of the model under different operating conditions.

[0012] The multi-physics coupling model described in Step 1 includes an electrochemical reaction sub-model, a proton conduction sub-model, and a gas diffusion sub-model; among them,

[0013] The electrochemical reaction sub-model uses the Butler-Volmer equation to describe the reaction rate on the electrode surface of the proton exchange membrane fuel cell, including the kinetic expressions of the anodic reaction rate and the cathodic reaction rate;

[0014] The proton conduction sub-model uses a conductivity model based on Nafion membrane, and the relationship between its conductivity and temperature and humidity is characterized by an exponential function:

[0015]

[0016] Among them, σ0 is the reference conductivity, E a is the activation energy, R is the gas constant, and f(λ) is the humidity-dependent function;

[0017] The gas diffusion sub-model uses a modified Fick-Knudsen diffusion model, which combines Fick molecular diffusion and Knudsen diffusion to describe the gas transport process in the catalyst layer and the microporous layer. The calculation of its effective diffusion coefficient is as follows:

[0018]

[0019] Among them, the molecular diffusion coefficient D molecular is obtained from Fick's law, and the Knudsen diffusion coefficient D Knudsen is determined by the pore size and the mean free path of gas molecules.

[0020] The specific method of Step 2 is as follows:

[0021] Step 2.1: Mesh the three-dimensional geometric structure of the fuel cell;

[0022] Adopt non-uniform grid division, locally refine the grid in key areas including the catalyst layer and the proton exchange membrane, and control the grid size within 10 - 20 μm; in non-critical areas including the flow channels, use a coarser grid of 50 - 100 μm; perform grid independence verification, and gradually refine the grid until the change rates of key variables including current density and temperature distribution are less than 1%.

[0023] Step 2.2 Numerically discretize the multi-physics coupling model established in Step 1 using the finite volume method at the grid nodes, and solve the discretized equations using an iterative algorithm; use the SIMPLE algorithm to solve the pressure-velocity coupling equation, where: the convection term adopts the second-order upwind scheme (QUICK scheme); the diffusion term adopts the central difference scheme; set the convergence criterion: the residuals of each equation are less than 1×10 -6 ; Obtain the preliminary simulation results of the multi-physics coupling model.

[0024] The specific method of Step 3 is as follows:

[0025] Step 3.1 Measure and obtain the basic performance data including voltage, temperature, and power density through polarization curve experiments, and determine the initial parameter range. The parameters include: material physical property parameters, reaction kinetic parameters, and transport characteristic parameters.

[0026] Step 3.2 Compare the simulation results of the multi-physics coupling model obtained in Step 2 with the polarization curve experimental data in Step 3.1, and use the particle swarm optimization algorithm to adjust the parameters. Specifically, customize the population size and the number of iterations, calculate the root mean square error of the voltage data in the simulation results obtained in Step 2 and the measured voltage of the polarization curve through the fitness function F(x) = RMSE(x), and stop the optimization when the error is less than 5%.

[0027] Step 3.3 Select multiple operating conditions different from the polarization curve experiment, including different temperatures, different humidities, different pressures, and different current density ranges, conduct experiments, and perform simulations under the same multiple operating conditions as the experiments to obtain the errors between the predictions and the experimental data, so as to verify the accuracy of the model through multi-condition experiments.

[0028] The specific method of Step 4 is as follows:

[0029] Input the following actual operating condition parameters into the multi-physics coupling model: current density 0.5 - 2.0 A / cm 2 , inlet humidity 50 - 100%, coolant flow rate 0.1 - 0.5 L / min; Output the prediction results, including output voltage, power density, current distribution, temperature distribution, gas flow rate, pressure, and water content distribution data.

[0030] The specific method of Step 5 is as follows:

[0031] Based on the prediction results obtained in Step 4, preprocess the data, including normalizing the electrical performance data (including voltage and current) to [0, 1]; performing time alignment on the gas parameters (including flow rate and pressure); filling in the missing values of the temperature data through bicubic interpolation; using a convolutional neural network (CNN) to extract the features related to polarization and internal resistance from the electrical performance data during the operation of the PEMFC, capturing the temporal dependence relationship and gas dynamic change characteristics from the gas state data (including pressure, flow rate, and concentration) through the LSTM unit in the recurrent neural network (RNN), and using an autoencoder to extract the temperature distribution and change trend features;

[0032] Obtain the weights between modalities through the multi-head self-attention mechanism of the Transformer, capture the complex relationships between multi-modal data for fusion, and form a comprehensive feature vector reflecting the operating state of the PEMFC;

[0033] Use the fused feature vector as the input layer and the predicted power density as the output layer to construct a deep neural network prediction model; train the model using several groups of multi-modal data covering simulation and experimental data under different temperature, humidity, and pressure conditions, use the cross-entropy loss function to quantify the difference between the power density predicted by the model and the simulation data in Step 2, and use the Adam optimizer to dynamically adjust the update direction and step size of the weights between modalities; use the backpropagation algorithm to iteratively adjust the model parameters using the gradient descent strategy, enabling the model to learn the mapping relationship between multi-modal data and the performance at high current densities, and the prediction results are used to guide the adjustment of the control parameters of the fuel cell system.

[0034] The specific method of Step 6 is as follows:

[0035] Step 6.1 Pre-training stage:

[0036] On the historical data set of typical operating conditions of the fuel cell, including rated conditions, variable load conditions, and low-temperature start-up conditions, train the deep neural network prediction model constructed in Step 5 to obtain the trained pre-training model and parameter values, including the convolutional layer and its parameter values, the bidirectional LSTM layer and its parameter values, and the encoder and its parameter values; the parameter values of the convolutional layer include the weights and bias terms of the convolutional kernel; the parameter values of the bidirectional LSTM layer include the weights of the input gate and forget gate; the encoder and its parameter values minimize the prediction error loss function L through the Adam optimizer:

[0037] L = α1L_voltage + α2L_resistance + α3L_temperature

[0038] where α1, α2, and α3 correspond to the prediction error weights of voltage, internal resistance, and temperature respectively;

[0039] Step 6.2 Perform parameter transfer on the pre-trained model obtained in Step 6.1:

[0040] 6.2.1, CNN weight transfer: Transfer the weights and bias terms of the convolutional kernels of the convolutional layers used for voltage distribution feature extraction in the pre-trained model trained on the historical dataset in Step 6.1 to the deep neural network prediction model constructed in Step 5 through parameter transfer to obtain a new model, and maintain its spatial feature extraction ability;

[0041] 6.2.2, RNN weight transfer: Transfer the recurrent unit parameters of the bidirectional LSTM layer in the pre-trained model obtained in Step 6.1, including the weights of the input gate and forget gate, to the new model;

[0042] 6.2.3, AE weight transfer: Transfer the overall parameters of the encoder part of the pre-trained model obtained in Step 6.1 to the new model;

[0043] Step 6.3 Fine-tune the new model after parameter transfer:

[0044] 6.3.1, Hierarchical learning rate setting:

[0045] Customize and set the learning rate of the frozen layers (the first 3 convolutional layers of the CNN and the input gate of the RNN);

[0046] Customize and set the learning rate of the fine-tuning layers (the last 2 convolutional layers of the CNN and the output gate of the RNN);

[0047] Customize and set the learning rate of the newly added layer (Transformer attention layer);

[0048] 6.3.2, Adopt gradient clipping and set the upper limit value of the gradient norm to prevent gradient explosion during the fine-tuning process;

[0049] Step 6.4 Adaptive optimization:

[0050] 6.4.1, Set the initial learning rate;

[0051] 6.4.2, Adjust the learning rate using the cosine annealing strategy:

[0052] η_t = η_min + 0.5(η_max - η_min)(1 + cos(tπ / T))

[0053] where η_max is the initial learning rate, η_min is the minimum learning rate, and T is the cycle length.

[0054] Step 6.5 Generalization enhancement:

[0055] 6.5.1, Data augmentation strategy:

[0056] Operating condition disturbance simulation: Add ±5% random noise to the input data;

[0057] Time series data truncation: Randomly delete 10%-20% of the time steps;

[0058] 6.5.2. During the test phase, Monte Carlo Dropout is adopted, the dropout rate is set for uncertainty estimation, and the variance threshold is set to measure whether the prediction variance uncertainty of the new model using Monte Carlo Dropout during the test phase is too high. When the prediction variance > the variance threshold, model recalibration is triggered.

[0059] A proton exchange membrane fuel cell (PEMFC) high current density performance prediction and optimization system, including the following modules:

[0060] Multi-physics field coupling modeling module, including a three-dimensional geometric modeling unit for constructing the battery geometric structure including flow channels, gas diffusion layers, catalyst layers, and proton exchange membranes in steps 1 and 2; a multi-field coupling solver that simultaneously solves the governing equations by the finite volume method for numerical simulation to obtain data such as voltage and temperature fields;

[0061] Experimental data acquisition and parameter optimization module for calibrating the multi-physics field coupling model parameters in step 3, including a multi-channel data acquisition unit for real-time acquisition of voltage, current, temperature, gas flow rate, and pressure data; a parameter calibration unit that optimizes the model parameters using an improved particle swarm algorithm, with the optimization objective being to make the root mean square error between the simulated voltage and experimental data < 5%;

[0062] Deep learning prediction module for multi-modal data fusion and performance prediction in step 5, where a convolutional neural network (CNN) processes voltage distribution images and infrared thermal imaging data, and a recurrent neural network (LSTM) processes gas parameter time series; dynamically weights the features of each modality through a multi-head attention mechanism; outputs the prediction of the performance decay rate based on the Transformer architecture;

[0063] Optimization decision-making module for outputting an improved optimization plan in step 6 based on the deep learning model after optimization and generalization processing in step 5 to ensure the applicability and reliability of the model under different operating conditions.

[0064] A proton exchange membrane fuel cell (PEMFC) high current density performance prediction and optimization device, including:

[0065] A memory for storing computer programs;

[0066] A processor for executing the computer program to perform the proton exchange membrane fuel cell (PEMFC) high current density performance prediction method according to any one of steps 1 to 6.

[0067] A computer-readable storage medium stores a computer program, which, when executed by a processor, can realize the prediction of the high current density performance of a proton exchange membrane fuel cell (PEMFC) based on the high current density performance prediction and optimization method of the proton exchange membrane fuel cell (PEMFC).

[0068] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0069] 1. The three-dimensional two-phase non-isothermal model proposed in step 1-2 of the present invention integrates multiple physical fields such as electrochemical reactions, proton / electron conduction, gas diffusion, heat transfer, and water transfer. In the electrochemical reaction model, the Butler-Volmer equation is introduced to describe the electrochemical reaction rate on the electrode surface; in the proton conduction model, the relationship between the proton conductivity of the proton exchange membrane and temperature and humidity is considered; in the gas diffusion model, the Knudsen diffusion is combined to correct the gas diffusion process in the catalyst layer and microporous layer, significantly improving the model's ability to describe the complex interaction mechanism inside the fuel cell. The proton conduction model considers the influence of temperature and humidity, and the gas diffusion model introduces microporous layer correction, which is closer to the actual working conditions, especially suitable for the simulation of non-linear effects at high current densities, and corrects the limitations of traditional models. It overcomes the problems of high experimental difficulty and high experimental cost at high current densities, and combines non-uniform grid technology (such as boundary layer encryption) to reasonably divide the three-dimensional geometric structure of the fuel cell, and uses the finite volume method to discretize the established multi-physical field coupling model; linearize the discretized equations and use iterative algorithms to solve them to ensure the stability and convergence of numerical calculations, reduce the calculation amount while ensuring accuracy, and reduce the consumption of computing resources;

[0070] 2. In step 5 of the present invention, deep learning algorithms are used for feature extraction for different types of data. For electrical performance data, convolutional neural networks are used to extract features related to polarization and internal resistance; for gas parameter data, recurrent neural networks are used to capture the dynamic change features of gases; for temperature data, autoencoders are used to extract the temperature distribution and change trend features. The extracted features are fused according to specific rules to form a comprehensive feature vector that comprehensively reflects the operating state of the fuel cell.

[0071] 3. In step 6 of the present invention, the features of different data sources are weighted and fused through an attention mechanism, fusing electrical performance, gas parameters, and temperature data. The characterization ability of the comprehensive feature vector for the battery state is improved by 40%. It comprehensively reflects the operating state of the PEMFC, significantly improves the accuracy of PEMFC performance prediction under high current density, combines the particle swarm optimization algorithm to optimize the model parameters, and is verified by multi-condition experimental data to ensure that the model prediction error is less than 5%, providing a reliable basis for subsequent performance analysis; Based on the fused feature vector, a deep neural network prediction model is constructed. The model selects the Transformer architecture, trains the model with a large amount of data, uses the cross-entropy loss function and the Adam optimizer, and continuously adjusts the model parameters to enable the model to learn the complex mapping relationship between multi-modal data and performance under high current density; Dropout, regularization, and cross-validation techniques are adopted to enhance the generalization ability of the model;

[0072] 4. Real-time prediction: The model can be deployed in an actual PEMFC system to achieve real-time performance prediction and status monitoring. In the design stage, the influence of parameters such as flow field plate and catalyst layer thickness on performance can be quickly evaluated; in the operation stage, the risk of hydrothermal imbalance can be diagnosed in real time and the control strategy can be optimized; in the maintenance stage, the aging of the membrane electrode or the blockage of the gas diffusion layer can be pre-warned based on the prediction results, thus supporting the full life cycle management of the fuel cell.

[0073] In summary, through the technical chain of "multi-physical field modeling - parameter optimization - deep learning fusion", the present invention realizes high-fidelity simulation and intelligent prediction under high current density conditions of fuel cells, provides a systematic solution for performance improvement, life extension, and fault prevention, significantly reduces the R & D cost and experimental cycle, and has important scientific value and engineering application prospects. Brief Description of the Drawings

[0074] Figure 1 is the schematic flow diagram of the present invention.

[0075] Figure 2 is the schematic diagram of local encryption of the three-dimensional structure grid of the fuel cell of the present invention.

[0076] Figure 3 is the schematic diagram of the multi-physical field coupling model of the present invention.

[0077] Figure 4 is the comparison chart of the simulation results and experimental data of the present invention.

[0078] Figure 5 is the architecture diagram of the deep learning prediction model of the present invention. Detailed Embodiment

[0079] The specific embodiments disclosed in the present invention will be described in detail below in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only for the purpose of illustration and explanation of the present disclosure, and are not intended to limit the present disclosure.

[0080] The specific technical solutions are as follows:

[0081] A method for predicting and optimizing the high current density performance of a proton exchange membrane fuel cell (PEMFC), comprising the following steps:

[0082] Step 1: Establish a multi-physics coupling model of a proton exchange membrane fuel cell (PEMFC) to provide a theoretical basis for subsequent numerical simulations; the basic control equations involved in the model include: mass conservation equation, momentum conservation equation, component transport equation, energy conservation equation, liquid water transport equation, membrane water transport equation, hydraulic conservation, and charge conservation equation:

[0083] 1) Mass conservation equation, used to describe the overall mass balance of the gas mixture:

[0084]

[0085] In the formula, ρ g is the density of the gas mixture / kg·m -3 , is the velocity of the gas mixture / m·s -1 , S m is the mass source term / kg·m -3 ·s -1 .

[0086] 2) Momentum conservation equation, used to calculate the flow velocity distribution of the gas in the porous medium:

[0087]

[0088] In the formula, ε is the porosity (the volume fraction of pores in the porous medium), s lq is the liquid water saturation (the volume fraction of liquid water in the pores), p g is the gas pressure / Pa, μ g is the viscosity of the gas mixture / kg·m -1 ·s -1 , S u is the momentum source term / kg·m -2 ·s -1 .

[0089] 3) Component transport equation, used to predict the concentration distribution of each component of the gas:

[0090]

[0091] In the formula, i represents the gas components to be solved (including: hydrogen, oxygen, and water vapor), and Y i is the mass fraction of each gas component, is the effective diffusion coefficient of each gas component / m 2 ·s -1 , S i is the mass source term / kg·m -3 ·s -1 .

[0092] 4) Hydraulic conservation (GDL / MPL / CL), used to describe the infiltration behavior of liquid water in porous media:

[0093]

[0094] In the formula, p lq is the hydraulic pressure / Pa, ρ lq is the density of liquid water / kg·m -3 , K is the permeability of the porous media / m 2 , k lq is the relative permeability of liquid water, μ lq is the viscosity of liquid water / kg·m -1 ·s -1 , S lq is the mass source term of liquid water / kg·m -3 ·s -1 .

[0095] Among them, the liquid water saturation is obtained through the Leverett-J equation:

[0096]

[0097] In the formula, p c is the capillary pressure / Pa, σ is the gas-liquid surface tension / N·m -1 , θ is the contact angle / °.

[0098] 5) Liquid water saturation conservation (GC), used to calculate the migration velocity of liquid water in the gas channel:

[0099]

[0100] In the formula, is the liquid water velocity, according to the model assumption

[0101] 6) Membrane water conservation (CL / PEM), used to simulate the migration of water molecules in the proton exchange membrane:

[0102]

[0103] In the formula, n dis the electrohydrodynamic drag coefficient, and F is the Faraday constant / C·mol -1 , is the effective proton conductivity / S·m -1 , φ ion is the proton potential / V, ρ mem is the membrane density / kg·m -3 , and EW is the equivalent mass / kg·mol -1 , is the effective diffusion coefficient of membrane-bound water, λ mw is the membrane-bound water content (moles of water molecules bound per mole of sulfonic acid groups), S mw is the membrane-bound water source term / mol·m -3 ·s -1 .

[0104] 7) Conservation of electron potential (BP / GDL / MPL / CL), used to calculate the electron conduction process in the solid phase and generate the current density distribution:

[0105]

[0106] In the formula, is the effective electron conductivity / S·m -1 , φ ele is the electron potential / V, S ele is the potential source term / A·m -3 .

[0107] 8) Conservation of proton potential (CL / PEM), used to describe the proton conduction behavior of the proton exchange membrane and the catalyst layer:

[0108]

[0109] In the formula, S ion is the potential source term / A·m -3 .

[0110] 9) Energy conservation equation (global domain), used to calculate the global temperature distribution:

[0111]

[0112] In the formula, C p is the specific heat capacity of the mixed gas / J·mol -1 ·K -1 , k eff is the effective thermal conductivity / W·m -1 ·K -1 , T is the temperature / K, S E is the energy source term / W·m -3 .

[0113] The multi-physics coupling model satisfies:

[0114] 1) The PEMFC operates in a steady state, i.e., the distribution of each physical quantity does not change with time;

[0115] 2) The flow is laminar;

[0116] 3) The gas is an ideal gas;

[0117] 4) The water produced by the oxygen reduction reaction at the cathode is assumed to be liquid water initially, and then undergoes a phase change to form water vapor;

[0118] 5) The contact between components is good, without contact thermal resistance;

[0119] 6) Each gas component and liquid water cannot penetrate through the proton exchange membrane;

[0120] 7) The drainage rate of liquid water in the channel is very fast, so the velocity of liquid water in the channel is equal to that of the gas.

[0121] The multi-physical field coupling model described in step 1 includes an electrochemical reaction sub-model, a proton conduction sub-model, and a gas diffusion sub-model; among them,

[0122] The electrochemical reaction sub-model uses the Butler-Volmer equation to describe the electrochemical reaction rate on the electrode surface, and its current source term is used as the coupled input of the proton conduction model and the electron conduction model, affecting the proton concentration distribution and the electric potential field. The reaction rate of this model is affected by the local temperature and gas concentration, forming a two-way coupling;

[0123] The proton conduction sub-model uses a conductivity model based on the Nafion membrane, and the relationship between its conductivity and temperature and humidity is characterized by an exponential function:

[0124]

[0125] where, σ0 is the reference conductivity, E a is the activation energy, R is the gas constant, and f(λ) is the humidity-dependent function;

[0126] The gas diffusion sub-model uses a modified Fick-Knudsen diffusion model, which combines Fick molecular diffusion and Knudsen diffusion, and is used to describe the gas transport process in the catalyst layer and the microporous layer. Its effective diffusion coefficient is calculated as follows:

[0127]

[0128] where, the molecular diffusion coefficient D molecular is obtained from Fick's law, and the Knudsen diffusion coefficient D Knudsen is determined by the pore size and the mean free path of gas molecules.

[0129] The gas concentration distribution of this model directly affects the reactant supply of the electrochemical reaction model. At the same time, its pore water saturation is coupled with the phase change transport of water. The model implementation uses the numerical simulation software ANSYS Fluent to implement the multi-physics coupling model, and defines the geometric structure, material properties, boundary conditions and initial conditions of the proton exchange membrane fuel cell in the software. A user-defined function (UDF) written in C language is used to calculate and update the transport coefficients and source terms during the iterative solution process.

[0130] Step 2: First, mesh the three-dimensional geometric structure of the fuel cell. Subsequently, use the finite volume method to numerically discretize the multi-physics coupling model established in Step 1 at the mesh nodes, and use an iterative algorithm to solve the discretized equations. Use the SIMPLE algorithm to solve the pressure-velocity coupling equation to obtain the preliminary simulation results of the multi-physics coupling model.

[0131] The specific method of the said Step 2 is as follows:

[0132] Step 2.1 Mesh the three-dimensional geometric structure of the fuel cell.

[0133] Use non-uniform mesh division, and locally refine the key areas including the catalyst layer and the proton exchange membrane, with the mesh size controlled within 10 - 20 μm to ensure high-precision calculations of electrochemical reactions and proton transport; in the non-critical areas including the flow channels, use a coarser mesh of 50 - 100 μm to reduce the computational amount; use mesh independence verification, and gradually refine the mesh until the change rates of the key variables including the current density and temperature distribution are less than 1% to ensure the calculation accuracy.

[0134] Step 2.2 Use the finite volume method to numerically discretize the multi-physics coupling model established in Step 1 at the mesh nodes, and use an iterative algorithm to solve the discretized equations; use the SIMPLE algorithm to solve the pressure-velocity coupling equation, where: the convection term uses the second-order upwind scheme (QUICK scheme); the diffusion term uses the central difference scheme; set the convergence criterion: the residuals of each equation are less than 1×10 -6 ; Obtain the preliminary simulation results of the multi-physics coupling model.

[0135] Step 3: Compare the simulation results of the multi-physics coupling model obtained in Step 2 with the experimental data, optimize the model parameters in combination with the particle swarm optimization algorithm, and verify the model accuracy through multi-condition experiments.

[0136] The specific method of the said Step 3 is as follows:

[0137] Step 3.1 Obtain the basic performance data including voltage, temperature, and power density through polarization curve experiments, and determine the initial parameter range in combination with the literature [ZHANGR, CHEN L, MINT, et al. Multiscale study of reactive transport and multiphase heat transfer processes in catalyst layers of proton exchange membrane fuel cells[J]. Carbon Neutrality, 2024, 3(1).]. The parameters include: material physical property parameters, reaction kinetic parameters, and transport characteristic parameters;

[0138] Step 3.2 Compare the simulation results of the multi - physical - field coupling model obtained in Step 2 with the polarization curve experimental data in Step 3.1, and use the particle swarm optimization algorithm to adjust the parameters. Specifically, custom - set the population size to 50 - 100 and the number of iterations to 100 - 200. Calculate the root - mean - square error of the voltage data in the simulation results obtained in Step 2 and the measured voltage of the polarization curve through the fitness function F(x) = RMSE(x). Stop the optimization when the error is less than 5%;

[0139] Step 3.3 Select multiple working conditions different from the polarization curve experiment, including different temperatures, different humidities, different pressures, and different current density ranges, conduct experiments, and perform simulations under the same multi - working - condition conditions as the experiments to obtain the error between the prediction and the experimental data, so as to verify the model accuracy through multi - working - condition experiments.

[0140] This embodiment uses the experimental data of 3 different working conditions (low / medium / high current density) for verification. Compare:

[0141] The error of the polarization curve (voltage - current density) is less than 3%; the temperature distribution error is less than 2°C; the correlation coefficient between the local current density distribution and the experimental measurement (such as segmented cell testing) is greater than 0.95.

[0142] If the error exceeds the limit, return to adjust the mesh division strategy or the optimization algorithm parameters.

[0143] Step 4, use the optimized model verified in Step 3, input the actual operating condition parameters, and predict the battery performance at high current density, including output voltage, power density, current distribution, temperature distribution, and water content distribution;

[0144] The specific method of Step 4 is as follows:

[0145] Input the following working condition parameters into the multi - physical - field coupling model: current density 0.5 - 2.0 A / cm 2, intake air humidity 50-100%, coolant flow 0.1-0.5L / min; output prediction results, including output voltage, power density, current distribution (identification of local hot spots) (highest temperature position and gradient), temperature distribution, gas flow, pressure and water content distribution (liquid water saturation, indicating flooding risk areas) data.

[0146] Analysis of multi-physics coupling interaction mechanisms:

[0147] (1) Temperature-electrochemical coupling effect:

[0148] When the local current density is greater than 1.5A / cm 2 When the temperature rises to above 70°C due to Joule heat, the water content λ of the membrane electrode drops below 4, causing the proton conductivity to decrease by more than 30% and eventually causing the output voltage to decay faster (>5% / h);

[0149] (2) Two-phase flow-reaction kinetics coupling:

[0150] The accumulation of liquid water in the microporous layer (saturation > 40%) causes the effective diffusion coefficient of oxygen to drop to 50% of that in the dry state, which in turn causes a local current density difference at the cathode of > 15%, thereby inducing the risk of reverse polarity battery (potential < -0.1 V).

[0151] Step 5: Based on the high current density performance prediction results obtained in step 4, a deep learning prediction model is constructed to further improve the prediction accuracy;

[0152] The specific method of step 5 is:

[0153] Based on the prediction results obtained in step 4, the data is preprocessed, including normalizing the electrical performance data (including voltage and current) to [0,1]; time alignment of gas parameters (including flow rate and pressure); missing value filling of temperature data through bicubic interpolation; using convolutional neural network (CNN) to extract the characteristics of electrical performance data related to polarization and internal resistance during PEMFC operation, capturing the time series dependency and gas dynamic change characteristics from gas state data (including pressure, flow rate, concentration) through LSTM units in recurrent neural network (RNN), and extracting temperature distribution and change trend characteristics using autoencoder;

[0154] The inter-modal weights are obtained through the Transformer's multi-head self-attention mechanism (head number 8), which captures the complex relationship between multi-modal data and fuses them to form a comprehensive feature vector that fully reflects the operating status of PEMFC.

[0155] Based on the fused feature vector, use the fused feature vector as the input layer and the predicted power density as the output layer to construct a deep neural network prediction model; use 10,000 groups of multi-modal data covering different temperature, humidity, and pressure conditions in simulation and experimental data to train the model. Adopt the cross-entropy loss function to quantify the difference between the power density predicted by the model (output layer) and the simulation data in step 2. Use the Adam optimizer to dynamically adjust the update direction and step size of the weights between modalities; utilize the gradient descent strategy through the backpropagation algorithm to iteratively adjust the model parameters, enabling the model to learn the mapping relationship between multi-modal data and performance under high current density. The prediction results are used to guide the adjustment of the control parameters of the fuel cell system.

[0156] Step 6. Optimize and generalize the deep learning model established in step 5, and output an improved optimization plan to ensure the applicability and reliability of the model under different working conditions.

[0157] The specific method of step 6 is as follows:

[0158] Step 6.1 Pre-training stage:

[0159] On the historical data set of typical operating conditions of fuel cells, including rated conditions, variable load conditions, and low-temperature startup conditions, train the deep neural network prediction model constructed in step 5 to obtain a trained pre-trained model and parameter values, including the convolutional layer and its parameter values, the bidirectional LSTM layer and its parameter values, and the encoder and its parameter values; the parameter values of the convolutional layer include the weights and bias terms of the convolutional kernel; the parameter values of the bidirectional LSTM layer include the weights of the input gate and forget gate; the encoder and its parameter values minimize the prediction error loss function L through the Adam optimizer:

[0160] L = α1L_voltage + α2L_resistance + α3L_temperature

[0161] Where α1, α2, and α3 respectively correspond to the prediction error weights of voltage, internal resistance, and temperature;

[0162] Step 6.2 Perform parameter transfer on the pre-trained model obtained in step 6.1:

[0163] 6.2.1, CNN weight transfer: Transfer the weights and bias terms of the convolutional kernel of the convolutional layer used for voltage distribution feature extraction in the pre-trained model trained on the historical data set in step 6.1 to the deep neural network prediction model constructed in step 5 through parameter transfer to obtain a new model, and maintain its spatial feature extraction ability;

[0164] 6.2.2, RNN Weight Transfer: Transfer the recurrent unit parameters of the bidirectional LSTM layer in the pre-trained model obtained in step 6.1, including the weights of the input gate and forget gate, to the new model;

[0165] 6.2.3, AE Weight Transfer: Transfer the overall parameters of the encoder part of the pre-trained model obtained in step 6.1 to the new model;

[0166] Step 6.3 Fine-tune the new model after parameter transfer:

[0167] 6.3.1, Hierarchical Learning Rate Setting:

[0168] Set the learning rate of the frozen layers (the first 3 convolutional layers of the CNN, the input gate of the RNN) = 0;

[0169] Set the learning rate of the fine-tuning layers (the last 2 convolutional layers of the CNN, the output gate of the RNN) = 1e-4;

[0170] Set the learning rate of the newly added layer (Transformer attention layer) = 5e-4;

[0171] 6.3.2, Apply gradient clipping and set the gradient norm upper limit value = 1.0 to prevent gradient explosion during the fine-tuning process;

[0172] Step 6.4 Adaptive Optimization:

[0173] 6.4.1, Set the initial learning rate;

[0174] 6.4.2, Adjust the learning rate using the cosine annealing strategy:

[0175] η_t = η_min + 0.5(η_max - η_min)(1 + cos(tπ / T))

[0176] Where the initial learning rate η_max = 3e-4, the minimum learning rate η_min = 1e-5, and the cycle length T = 100 epochs.

[0177] Step 6.5 Generalization Enhancement:

[0178] 6.5.1, Data Augmentation Strategy:

[0179] Operating Condition Disturbance Simulation: Add ±5% random noise to the input data;

[0180] Time Series Data Truncation: Randomly delete 10% - 20% of the time steps;

[0181] 6.5.2. During the testing phase, Monte Carlo Dropout is adopted, and the dropout rate (dropout rate = 0.2) is set for uncertainty estimation. A variance threshold is set to measure whether the prediction variance uncertainty of the new model using Monte Carlo Dropout during the testing phase is too high. When the prediction variance > variance threshold, model recalibration is triggered.

[0182] Verification by simulation experiment:

[0183] A three-dimensional multi-physics coupling model of PEMFC is established. The particle swarm optimization algorithm (PSO) is used to calibrate the model parameters to ensure that the simulation results are consistent with the experimental data. The deep learning model is trained using the simulation data to further improve the prediction accuracy and calculation efficiency. The robustness of the model under variable working conditions is ensured through parameter transfer. Comparison with the existing technology: The prediction authenticity is improved through multi-physics coupling; the deep learning model can predict the transient response in real time, while the traditional CFD cannot meet this requirement; by combining PSO parameter optimization and deep learning, the two major problems of "model calibration" and "fast prediction" are solved simultaneously. Through quantitative analysis verified by simulation and experiment, it can be obtained that the voltage prediction error < 2% (experiment vs simulation), a 60%-80% reduction compared with the existing technology (typical value 5%-10%); the detection accuracy of the hot spot position > 90% (simulation vs experimental infrared thermal imaging), about 10% improvement compared with the traditional CFD model; in addition, the inference time of the deep learning model is extremely short, accelerating by more than a hundred times compared with a single simulation of the traditional CFD. The model can predict the flooding risk 10-15 minutes in advance to guide the regulation strategy.

Claims

1. A method for predicting and optimizing the high current density performance of a proton exchange membrane fuel cell (PEMFC), characterized in that, It includes the following steps: Step 1, establish a multi-physics coupling model of a proton exchange membrane fuel cell (PEMFC); Step 2, first perform mesh division on the three-dimensional geometric structure of the fuel cell, then numerically discretize the multi-physics coupling model established in Step 1 using the finite volume method at the mesh nodes, solve the discretized equations using an iterative algorithm, and use the SIMPLE algorithm to solve the pressure-velocity coupling equation to obtain the preliminary simulation results of the multi-physics coupling model; Step 3, compare the simulation results of the multi-physics coupling model obtained in Step 2 with the experimental data, optimize the model parameters in combination with the particle swarm optimization algorithm, and verify the model accuracy through multi-condition experiments; Step 4, use the optimized model verified in Step 3, input the actual operating condition parameters, and predict the battery performance at high current density, including output voltage, power density, current distribution, temperature distribution, and water content distribution; Step 5, based on the high current density performance prediction results obtained in Step 4, construct a deep learning prediction model to further improve the prediction accuracy; Step 6, optimize and generalize the deep learning model established in Step 5, and output an improved optimization plan to ensure the applicability and reliability of the model under different working conditions.

2. The method according to claim 1, wherein The multi-physics coupling model described in Step 1 includes an electrochemical reaction sub-model, a proton conduction sub-model, and a gas diffusion sub-model; among them, The electrochemical reaction sub-model uses the Butler-Volmer equation to describe the reaction rate on the surface of the proton exchange membrane fuel cell electrode, including the kinetic expressions of the anode reaction rate and the cathode reaction rate; The proton conduction sub-model uses a conductivity model based on Nafion membrane, and its relationship between conductivity and temperature and humidity is characterized by an exponential function: where σ0 is the reference conductivity, E a is the activation energy, R is the gas constant, and f(λ) is the humidity-dependent function; The gas diffusion sub-model uses a modified Fick-Knudsen diffusion model, which combines Fick molecular diffusion and Knudsen diffusion to describe the gas transport process in the catalyst layer and the microporous layer, and its effective diffusion coefficient is calculated as follows: Among them, the molecular diffusion coefficient D molecular is obtained from Fick's law, and the Knudsen diffusion coefficient D Knudsen is determined by the pore size and the mean free path of gas molecules.

3. The method according to claim 1, characterized in that, The specific method of Step 2 is as follows: Step 2.1 Perform mesh division on the three-dimensional geometric structure of the fuel cell; Use non-uniform mesh division, locally encrypt in the key areas including the catalyst layer and the proton exchange membrane, and control the mesh size within 10 - 20μm; in the non-key areas including the flow channels, use a coarser mesh of 50 - 100μm; use mesh independence verification, gradually encrypt the mesh until the change rate of the key variables including current density and temperature distribution is less than 1%; Step 2.2 Numerically discretize the multi-physics coupling model established in Step 1 on the grid nodes by using the finite volume method, and solve the discretized equations by using an iterative algorithm; use the SIMPLE algorithm to solve the pressure-velocity coupling equation, where: the convective term adopts the second-order upwind scheme (QUICK scheme); the diffusion term adopts the central difference scheme; set the convergence criterion: the residuals of each equation are less than 1×10 -6 ; obtain the preliminary simulation results of the multi-physics coupling model.

4. The method according to claim 1, characterized in that, The specific method of Step 3 is as follows: Step 3.1 Measure and obtain the basic performance data including voltage, temperature, and power density through polarization curve experiments to determine the initial parameter range, and the parameters include: material physical property parameters, reaction kinetic parameters, and transport characteristic parameters; Step 3.2 Compare the simulation results of the multi-physics coupling model obtained in Step 2 with the polarization curve experimental data in Step 3.1, and use the particle swarm optimization algorithm to adjust the parameters, specifically including: customizing the population size and the number of iterations, calculating the root mean square error of the voltage data in the simulation results obtained in Step 2 and the experimentally measured voltage of the polarization curve through the fitness function F(x) = RMSE(x), and stopping the optimization when the error is less than 5%; Step 3.3 Select multiple operating conditions different from the polarization curve experiment, including different temperatures, different humidities, different pressures, and different current density ranges, conduct experiments, and perform simulations under the same multiple operating conditions as the experiment to obtain the error between the prediction and the experimental data, so as to verify the accuracy of the model through multiple operating condition experiments.

5. The method according to claim 1, characterized in that The specific method of Step 4 is as follows: Input the following actual operating condition parameters into the multi-physical field coupling model: current density of 0.5 - 2.0 A / cm 2 , inlet humidity of 50 - 100%, coolant flow rate of 0.1 - 0.5 L / min; output the prediction results, including output voltage, power density, current distribution, temperature distribution, gas flow rate, pressure, and water content distribution data.

6. The method according to claim 1, wherein The specific method of constructing the deep learning prediction model in Step 5 is as follows: Based on the prediction results obtained in Step 4, preprocess the data, including normalizing the electrical performance data (including voltage and current) to [0, 1]; aligning the time of the gas parameters (including flow rate and pressure); filling the missing values of the temperature data through bicubic interpolation; using a convolutional neural network (CNN) to extract the features related to polarization and internal resistance from the electrical performance data during the operation of the PEMFC, capturing the temporal dependence relationship and the dynamic change characteristics of the gas from the gas state data (including pressure, flow rate, and concentration) through the LSTM unit in the recurrent neural network (RNN), and using an autoencoder to extract the temperature distribution and change trend features; Obtain the weights between modalities through the multi-head self-attention mechanism of the Transformer, capture the complex relationships between multi-modal data for fusion, and form a comprehensive feature vector reflecting the operating state of the PEMFC; Use the fused feature vector as the input layer and the predicted power density as the output layer to construct a deep neural network prediction model; train the model with multi-modal data of several groups of simulation and experimental data covering different temperature, humidity, and pressure conditions, use the cross-entropy loss function to quantify the difference between the power density predicted by the model and the simulation data in Step 2, and use the Adam optimizer to dynamically adjust the update direction and step size of the weights between modalities; use the backpropagation algorithm to iteratively adjust the model parameters using the gradient descent strategy, so that the model learns the mapping relationship between multi-modal data and the performance at high current densities, and the prediction results are used to guide the adjustment of the control parameters of the fuel cell system.

7. The method according to claim 6, wherein The specific method of Step 6 is as follows: Step 6.1 Pre-training stage: On the historical data set of typical operating conditions of the fuel cell, including rated conditions, variable load conditions, and low-temperature start-up conditions, train the deep neural network prediction model constructed in Step 5 to obtain the trained pre-trained model and parameter values, including the convolutional layer and its parameter values, the bidirectional LSTM layer and its parameter values, and the encoder and its parameter values; the parameter values of the convolutional layer include the weights and bias terms of the convolutional kernel; the parameter values of the bidirectional LSTM layer include the weights of the input gate and the forget gate; the encoder and its parameter values minimize the prediction error loss function L through the Adam optimizer: L = α1L_voltage + α2L_resistance + α3L_temperature where α1, α2, and α3 correspond to the prediction error weights of voltage, internal resistance, and temperature, respectively; Step 6.2 performs parameter transfer on the pre-trained model obtained in Step 6.1: 6.2.1, CNN weight transfer: Transfer the weights and bias terms of the convolutional kernels of the convolutional layers used for voltage distribution feature extraction in the pre-trained model trained on the historical dataset in Step 6.1 to the deep neural network prediction model constructed in Step 5 through parameter transfer to obtain a new model, maintaining its spatial feature extraction ability; 6.2.2, RNN weight transfer: Transfer the recurrent unit parameters of the bidirectional LSTM layer in the pre-trained model obtained in Step 6.1, including the input gate and forget gate weights, to the new model; 6.2.3, AE weight transfer: Transfer the overall parameters of the encoder part of the pre-trained model obtained in Step 6.1 to the new model; Step 6.3 fine-tunes the new model after parameter transfer: 6.3.1, Hierarchical learning rate setting: Customize the learning rate of the frozen layers (the first 3 convolutional layers of the CNN and the input gate of the RNN); Customize the learning rate of the fine-tuning layers (the last 2 convolutional layers of the CNN and the output gate of the RNN); Customize the learning rate of the newly added layer (the Transformer attention layer); 6.3.2, Use gradient clipping and set the upper limit of the gradient norm to prevent gradient explosion during the fine-tuning process; Step 6.4 Adaptive optimization: 6.4.1, Set the initial learning rate; 6.4.2, Use the cosine annealing strategy to adjust the learning rate: η_t = η_min + 0.5(η_max - η_min)(1 + cos(tπ / T)) where η_max is the initial learning rate, η_min is the minimum learning rate, and T is the cycle length. Step 6.5 Generalization enhancement: 6.5.1, Data augmentation strategy: Operating condition perturbation simulation: Add ±5% random noise to the input data; Time series data truncation: Randomly delete 10% - 20% of the time steps; 6.5.2, Use Monte Carlo Dropout in the test phase, set the dropout rate for uncertainty estimation, and set the variance threshold to measure whether the prediction variance uncertainty of the new model using Monte Carlo Dropout in the test phase is too high. When the prediction variance > variance threshold, trigger model recalibration.

8. A high current density performance prediction and optimization system for a proton exchange membrane fuel cell (PEMFC), characterized in that, It includes the following interconnected modules: Multi-physics field coupling modeling module, including a three-dimensional geometry modeling unit for constructing the battery geometry structure containing flow channels, gas diffusion layers, catalytic layers, and proton exchange membranes in Steps 1 and 2; a multi-field coupling solver that simultaneously solves the governing equations by the finite volume method for numerical simulation to obtain data such as voltage and temperature fields; Experimental data acquisition and parameter optimization module for calibrating the multi-physics field coupling model parameters in Step 3, including a multi-channel data acquisition unit for real-time acquisition of voltage, current, temperature, gas flow, and pressure data; a parameter calibration unit that uses an improved particle swarm algorithm to optimize the model parameters, with the optimization goal of making the root mean square error between the simulated voltage and experimental data < 5%; A deep learning prediction module for multi-modal data fusion and performance prediction in Step 5, where a Convolutional Neural Network (CNN) processes voltage distribution images and infrared thermal imaging data, and a Long Short-Term Memory (LSTM) network processes gas parameter time series; dynamically weights the features of each modality through a multi-head attention mechanism; and outputs a prediction of the performance decay rate based on the Transformer architecture. An optimization decision module for Step 6, which outputs an improved optimization plan based on the deep learning model after optimization and generalization in Step 5 to ensure the applicability and reliability of the model under different working conditions.

9. A device for predicting and optimizing the high current density performance of a proton exchange membrane fuel cell (PEMFC), characterized in that, It includes: A memory for storing computer programs; A processor for performing the proton exchange membrane fuel cell (PEMFC) high current density performance prediction method according to any one of Steps 1 to 6 when executing the computer program.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it can realize the high current density performance prediction of the proton exchange membrane fuel cell (PEMFC) based on the proton exchange membrane fuel cell (PEMFC) high current density performance prediction and optimization method.

Images