A method for predicting cold start time of proton exchange membrane fuel cells

Abstract

This invention discloses a method for predicting the cold start time of a proton exchange membrane fuel cell (PEMFC). This method, which pertains to the field of fuel cell design, involves employing a multilayer perceptron neural network and deep learning methods to establish a nonlinear mapping relationship between initial parameters and cold start time. Finally, cold start time prediction is performed using the trained cold start time prediction model. This method combines high precision, high efficiency, good physical interpretability, and strong generalization capabilities, providing important technical support for low-temperature fuel cell startup control and promising broad application prospects.

Description

Technical Field

[0001] The invention belongs to the field of fuel cell design, and in particular relates to a method for predicting the cold start time of a proton exchange membrane fuel cell. Background Art

[0002] Proton exchange membrane fuel cells (PEMFCs), a clean energy technology fueled by hydrogen, offer significant advantages such as high power generation efficiency, low pollutant emissions, compact size, and low noise, and hold broad application prospects. Due to their unique advantages, PEMFCs have become a preferred energy source to replace traditional rechargeable batteries, particularly for automotive and home power systems, and are considered a strong contender to replace traditional internal combustion engines in the future. Global investment in fuel cell vehicle R&D continues to grow.

[0003] Due to their unique characteristics, proton exchange membrane fuel cells face numerous challenges in comprehensive applications. Icing is a major cause of startup failure during cold starts of PEMFCs in low-temperature environments. During cold-start, water in the electrolyte can freeze, causing the space between the membrane and the catalyst layer to expand, increasing electrical resistance and potentially causing material spalling and damage. Furthermore, ice in the gas diffusion layer can clog its porous structure, altering its hydrophobicity and gas permeability, thus impacting the fuel cell's performance and lifespan.

[0004] Currently, research on cold start performance focuses primarily on enhancing fuel cell stack materials and improving startup methods. However, material upgrades do not guarantee a successful fuel cell startup. Therefore, shifting from troubleshooting to preventive maintenance, predicting and addressing potential issues during PEMFC cold start, has significant practical application value. Currently, methods for predicting fuel cell cold start times and enabling early startup adjustments remain largely unresolved.

[0005] Fuel cell cold-start time directly impacts the system's startup efficiency in low-temperature environments. Traditional approaches suffer from two drawbacks: first, numerical models based on physical mechanisms are computationally complex and difficult to meet real-time control requirements; second, purely data-driven approaches rely on large amounts of experimental data and lack generalizability. Existing solutions have yet to effectively combine the advantages of mechanistic models with those of neural networks. Summary of the Invention

[0006] To solve the above technical problems, the present invention provides a method for predicting the cold start time of a proton exchange membrane fuel cell, comprising:

[0007] Construct a mechanism model for cold start time prediction based on fuel cells;

[0008] Constructing a neural network model, and training the neural network model based on the mechanism model to obtain a prediction model;

[0009] The cold start time of the proton exchange membrane fuel cell is predicted based on the prediction model to generate a prediction result.

[0010] Preferably, the mechanism model includes: a gas transmission model, an output voltage model, a temperature stratification model and a water transmission model.

[0011] Preferably, the gas transmission model includes: a cathode gas transmission model and an anode gas transmission model, and the cathode gas transmission model and the anode gas transmission model are used to reflect the gas diffusion, migration and reaction processes of the fuel cell in a low temperature environment.

[0012] Preferably, the output voltage model is used to obtain the voltage provided to the outside under actual operating conditions of the fuel cell under the influence of ice accumulation factors.

[0013] Preferably, the temperature stratification model is established based on a heat transfer relationship expression at the membrane electrode, and is used to calculate the membrane electrode temperature and ice melting heat.

[0014] Preferably, the water transport model is used to perform water balance calculations and ice accumulation calculations.

[0015] Preferably, the process of obtaining the prediction model includes:

[0016] The startup time and input amount obtained based on the mechanism model are used as a historical data set;

[0017] Normalizing the historical data set to obtain a normalized data set;

[0018] Constructing a neural network model, and training the neural network model based on the normalized data set to obtain a training model;

[0019] The model is evaluated using the test set, and the mean square error and the coefficient of determination are calculated as measures of the model's prediction accuracy to obtain the prediction model.

[0020] Preferably, the prediction model has an input layer with 2 nodes, a hidden layer with 60 neurons, and an output layer with 1 node.

[0021] On the other hand, the present invention further provides an electronic device, comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the method is implemented when the processor executes the computing program.

[0022] On the other hand, the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program implements the method when executed by a processor.

[0023] Compared with the prior art, the present invention has the following advantages and technical effects:

[0024] This paper proposes a hybrid modeling-based method for predicting the cold start time of proton exchange membrane fuel cells. This method utilizes a mechanistic model to generate a large amount of high-quality simulation training data. Based on this data, the method employs a multilayer perceptron neural network (MLP) and deep learning methods to establish a nonlinear mapping relationship between initial parameters and cold start time. Ultimately, the trained cold start time prediction model is used to predict cold start time. This method combines high accuracy, high efficiency, good physical interpretability, and strong generalization capabilities, providing important technical support for low-temperature start control of fuel cells and promising broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0026] Figure 1 A system flow chart of an embodiment of the present invention;

[0027] Figure 2 Schematic diagram of the feedback effect of ice formation and heat generation inside a fuel cell during low-temperature startup according to an embodiment of the present invention;

[0028] Figure 3 A diagram showing the temperature stratification and energy transfer model of a fuel cell unit according to an embodiment of the present invention;

[0029] Figure 4 This is a schematic diagram of a MATLAB / Simulink model based on the fuel cell mechanism of an embodiment of the present invention;

[0030] Figure 5 Schematic diagram of the neural network topology structure of an embodiment of the present invention. DETAILED DESCRIPTION

[0031] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0032] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0033] Example 1

[0034] like Figure 1As shown, this embodiment provides a method for predicting the cold start time of a proton exchange membrane fuel cell, including:

[0035] Step 1: First, build a mechanism model for cold start time prediction.

[0036] Step 1-1, taking the fuel cell monomer as the research object, divide the low-temperature startup process into stages and analyze the influence relationship between its internal icing and heat generation.

[0037] Step 1-2: Establish a PEMFC mechanism numerical model to simulate the low-temperature startup process, and determine the electrochemical reactions, water balance, material transfer and heat transfer processes inside the fuel cell monomer from a mechanism level.

[0038] The mechanism model takes initial external conditions such as current and ambient temperature as input and outputs the voltage and membrane electrode temperature during model operation. Specifically, the mechanism model provided in this embodiment includes four parts: a gas transmission model, an output voltage model, a temperature stratification model, and a water transmission model.

[0039] Establish a gas transmission model: including cathode gas transmission model and anode gas transmission model.

[0040] The cathode and anode gas transport models describe the gas diffusion, migration, and reaction processes in fuel cells under low-temperature conditions. These transport and reaction processes are affected by factors such as temperature, pressure, gas concentration, and current density. Therefore, the hydrogen and oxygen flow rates consumed by the fuel cell to achieve a set current vary at different reaction temperatures. By calculating these gas consumptions, the fuel cell inlet and outlet gas flow rates and water vapor partial pressure can be further calculated, allowing the assessment of membrane water content and water management, which are key factors affecting the cold-start performance of fuel cells.

[0041] An output voltage model is established to obtain the voltage provided to the outside under the actual operating conditions of the fuel cell under the influence of ice accumulation factors. By calculating the actual operating voltage of the fuel cell under different ice accumulation levels, the degree of ice accumulation can be reflected and the cold start situation of the fuel cell can be predicted.

[0042] Based on a comprehensive consideration of factors such as the high thermal conductivity of fuel cell materials, the rapid heat transfer characteristics of thin-layer structures, and the practicality of engineering calculations, this model assumes that the membrane electrode structure composed of the proton exchange membrane and catalyst layer and other layers of the monomer structure are a whole with uniform temperature distribution. This allows for simplified calculations while ensuring prediction accuracy, thereby improving the real-time and applicability of cold start time prediction.

[0043] Based on the following heat transfer relationship expression at the membrane electrode, a temperature stratification model is established. The membrane electrode temperature and ice melting heat are calculated using the known ambient temperature, current, output voltage, and inlet flow calculated in the above process:

[0044] ;

[0045] in, is the membrane electrode specific heat capacity, is the membrane electrode mass, is the membrane electrode temperature, 、 They are the heat generated in the membrane electrode and the heat transferred to the gas diffusion layer.

[0046] Establish water transport models and perform water balance and ice accumulation calculations;

[0047] The water balance calculations include the ice accumulation rate, the water generation rate in the cathode catalyst layer, the water absorption rate by the membrane, the water discharge rate through the gas diffusion layer, and the water back diffusion rate.

[0048] Based on porous media theory and numerical modeling requirements, this paper defines the ice volume fraction as the ratio of the ice volume in the pores of the catalyst layer to the total voids.

[0049] Step 1-3, according to the low-temperature startup process divided in this paper, calculate the ice accumulation in different stages and judge the cold start situation.

[0050] By judging the ice volume fraction, it is possible to determine whether the cold start is successful. If the cold start is successful, the specific time of the cold start can be predicted. If the ice volume fraction continues to accumulate to the maximum value and the ice completely blocks the pores of the catalyst layer, it will cause the low-temperature start to fail.

[0051] If the stack membrane electrode temperature rises to the freezing point before the ice volume fraction rises to the critical value, the accumulated ice will begin to melt in the third stage.

[0052] Based on the energy balance, the ice volume fraction and the water transfer flow rate between the two poles under the current condition are calculated using the available parameters such as current and membrane electrode temperature. This output is the influencing factor in the output voltage model.

[0053] Step 1-4, here the cold start time is defined as: the duration required for the PEMFC to start at a temperature below freezing and after the membrane electrode temperature rises to zero degrees Celsius. Among them, the boundary temperature of the fuel cell cold start time is when the membrane electrode temperature reaches the freezing point. After that, the fuel cell can operate smoothly without startup failure caused by ice.

[0054] In steps 1-5, the model monitors the membrane electrode temperature at different currents and ambient temperatures. When the membrane electrode temperature reaches the defined boundary, the model runtime is recorded. This time is the cold start time. A sufficient number of data sets, each containing the startup current, ambient temperature, and cold start time at a specific moment, are recorded as sample data for the following model training and verification.

[0055] Step 2. Secondly, build a neural network and use the mechanism model data to obtain an accurate and fast prediction model.

[0056] Step 2-1: To improve the speed of cold start time prediction, a multi-layer perceptron (MLP) neural network prediction method is used to improve the speed of prediction. By training on historical data, the relationship between input features (ambient temperature and starting current) and target output (start time) under different environments is learned.

[0057] Step 2-2, when the membrane electrode temperature reaches the boundary condition, the startup time and input quantities (ambient temperature and startup current) obtained by the mechanism model are used as the historical data set.

[0058] In Step 2-3, data normalization method is used as the preprocessing method of the original data to improve the prediction accuracy of the model, eliminate the influence of different input feature dimensions, and improve the training stability and convergence speed of the model.

[0059] In step 2-4, in order to evaluate the learning effect and generalization ability of the model, this embodiment splits the dataset into a training set and a test set, using a random partitioning strategy where 70% is used for training and the remaining 30% is used for validation.

[0060] Step 2-5, design a neural network structure to optimize the cold start time prediction model. This embodiment adopts a feedforward neural network with two nodes in the input layer, 60 neurons in one hidden layer, and one node in the output layer.

[0061] Step 2-6, using the ReLU activation function in the hidden layer, can converge faster during training and effectively handle complex nonlinear relationships. The output layer uses a linear activation function, which can directly output the prediction results and is suitable for regression tasks.

[0062] In Step 2-7, the mean square error (MSE) is used as the loss function to measure the error between the predicted value and the true value.

[0063] Step 2-8: In order to improve the convergence speed of model training, the Adam optimizer is used to improve training efficiency.

[0064] In Step 2-9, the model is trained using the training set. 300 training rounds are set, with a batch size of 64. Early stopping is used to prevent overfitting.

[0065] In Step 2-10, the He normal initialization method is used to initialize the weights during the model optimization process. Dropout regularization is introduced and the Dropout ratio is set to 0.2.

[0066] Step 2-11: After training, the model is evaluated using the test set, and the mean square error (MSE) and coefficient of determination R² are calculated as measures of the model's prediction accuracy.

[0067] Step 3: Finally, the trained and verified accurate model is solidified into a prediction engine.

[0068] After the above training and verification, it was confirmed that a high-precision and high-efficiency fuel cell cold start time prediction model was successfully constructed. Finally, the predicted values ​​were restored to actual units through inverse normalization.

[0069] Therefore, the cold start time of the fuel cell under this condition can be quickly predicted by the ambient temperature and starting current.

[0070] Example 2

[0071] This embodiment provides a method for predicting the cold start time of a proton exchange membrane fuel cell, comprising:

[0072] The first part is the construction of mechanism model.

[0073] The following settings are made based on the fuel cell manufacturing process referenced in this embodiment:

[0074] First, the membrane electrode of the proton exchange membrane fuel cell is a three-layer structure consisting of one proton exchange membrane and two catalyst layers, with uniform temperature distribution.

[0075] Secondly, the specific process of low-temperature startup is divided into four stages according to the state of water in the catalyst layer;

[0076] Phase 1: The fuel cell starts working, water is continuously generated in the cathode catalyst layer, and the water accumulates in the ionomer and gas phase.

[0077] Stage 2: Once the catalytic layer is saturated with water vapor, the generated water will accumulate in the catalytic layer in the form of ice.

[0078] Stage 3: If the fuel cell temperature reaches 0°C before the catalyst layer surface is completely frozen, the ice accumulated inside begins to melt.

[0079] Stage 4: When the accumulated ice completely melts, the fuel cell temperature rises further and low-temperature startup is successful.

[0080] like Figure 2 As shown, this embodiment provides a schematic diagram of the feedback effect of ice formation and heat generation inside the fuel cell during low-temperature startup, revealing the results of the mutual influence and changes between internal electrochemical reactions and material transport in several stages of low-temperature startup.

[0081] According to the working reaction mechanism of PEMFC, a simulation model of the fuel cell system is built in Simulink.

[0082] First, a gas transmission model is built based on the principle of conservation of gas mass, the laws of thermodynamics and the ideal gas state equation.

[0083] Establish a cathode gas transport model to calculate the mass flow of oxygen at the cathode inlet of the fuel cell;

[0084] ;

[0085] in, is the volume fraction of oxygen at the cathode inlet; is the air mass flow rate at the cathode inlet; , are the molar masses of oxygen and nitrogen, respectively; is the specific humidity of the gas at the cathode inlet.

[0086] In addition, the mass flow rate of water vapor discharged from the cathode gas diffusion layer is calculated using the water balance model described later.

[0087] Similarly, an anode gas transmission model is established to calculate the hydrogen mass flow rate at the anode inlet of the fuel cell;

[0088] ;

[0089] ;

[0090] in, is the gas mass flow rate at the anode inlet; is the molar mass of hydrogen; is the gas specific humidity at the anode inlet; is the relative humidity at the anode inlet; is the gas temperature at the anode inlet.

[0091] When the generated water is transferred from the anode to the cathode, its mass flow rate is regulated by both electroosmosis and back diffusion, and its value can be calculated based on the electroosmosis coefficient.

[0092] The output voltage of a fuel cell is determined by a variety of factors, and its specific calculation formula is expressed as:

[0093] ;

[0094] in, represents the Nernst voltage; represents activation loss; Corresponding concentration loss; is the ohmic loss.

[0095] The activation loss is affected by two factors: reactant concentration and ice accumulation, and its solution process is based on the Tafel equation.

[0096] like Figure 3 As shown, this embodiment constructs a temperature stratification and heat transfer model of a fuel cell monomer. The model focuses on the generation and transfer of heat, and does not involve the details of the flow field structure and the influence of water distribution.

[0097] This enables the calculation of the temperature distribution inside the battery cell.

[0098] In the membrane electrode, the heat generated It mainly consists of the following three parts: First, the reversible heat generated by the electrochemical reaction , followed by irreversible heat caused by polarization phenomenon and water phase change , and finally the latent heat , expressed as follows:

[0099] ;

[0100] The reversible heat is controlled by both the membrane electrode temperature and the current density, wherein the membrane electrode temperature is obtained by heat transfer.

[0101] The irreversible heat is affected by the electrochemical reaction and is calculated using the output voltage mentioned above.

[0102] Use the following expression to calculate the membrane electrode temperature.

[0103] ;

[0104] in, is the membrane electrode specific heat capacity, is the membrane electrode mass, is the membrane electrode temperature, 、 They are the heat generated in the membrane electrode and the heat transferred to the gas diffusion layer.

[0105] Taking the cathode catalyst layer as the main research object, a water balance and ice accumulation model was established;

[0106] The water balance formula is as follows:

[0107] ;

[0108] in, Expressed as the ice accumulation rate, is the water generation rate in the cathode catalyst layer, represents the rate at which the membrane absorbs water, represents the rate of water discharge through the gas diffusion layer, represents the rate at which water undergoes reverse diffusion;

[0109] The ice volume fraction is defined as the ratio of the ice volume in the pores of the catalyst layer to the total voids to better express the icing phenomenon.

[0110] According to the low-temperature startup process divided in this embodiment, in the first stage, a water storage balance relationship can be obtained.

[0111] At this time, the ice volume status inside the fuel cell is judged.

[0112] If the water storage obtained by integrating the ice accumulation rate in the first stage exceeds the maximum water storage capacity of the ionomer in the catalytic layer, it means that the ionomer has reached saturation and the water accumulation rate has become the ice accumulation rate.

[0113] On the contrary, if the water storage obtained by integrating the ice accumulation rate over the first stage does not exceed the maximum water storage of the ionomer in the catalytic layer, it indicates that the ionomer has not reached saturation and there is no ice accumulation. At this time, the fuel cell can start normally.

[0114] Furthermore, when the ionomer reaches saturation, it enters the second stage and ice is formed evenly in the pores.

[0115] In the second stage, the ice accumulation is the integral of the water generation rate in the cathode catalyst layer minus the water absorption rate of the membrane and the water discharge rate through the gas diffusion layer over the second stage time.

[0116] Therefore, the ice volume fraction in the second stage is calculated as the initial ice volume fraction plus the generated ice accumulation.

[0117] During this process, the ice volume fraction accumulates continuously to a maximum value. If the ice completely blocks the pores in the catalytic layer, it can be determined that the battery shutdown has caused the low-temperature start-up failure and the start-up time cannot be calculated.

[0118] If the stack temperature can rise to the freezing point before the ice volume fraction reaches the critical value, the ice accumulated in the previous stage will begin to melt in the third stage and the stack can still be started at a low temperature.

[0119] The ice melting rate in the third stage is calculated using the following energy balance.

[0120] Then the ice volume fraction is calculated as follows;

[0121] ;

[0122] Where, is the initial ice volume fraction, is the molar mass of ice.

[0123] During this process, the volume fraction of ice gradually increases until it reaches a maximum value. If the ice completely fills the pores of the catalyst layer, it may cause the battery to shut down, resulting in low-temperature start-up failure.

[0124] When the ice volume fraction has not yet reached the critical value, if the stack temperature rises to the freezing point, it enters the third stage and begins to melt the accumulated ice. The melting rate of the ice can be calculated through energy balance.

[0125] ;

[0126] As the ice melts, it absorbs all the heat generated by the battery stack, while the battery temperature remains stable at freezing point.

[0127] like Figure 4 As shown in the figure, combined with the modeling content of the fuel cell monomer in the above steps, a low-temperature startup lumped parameter model of the fuel cell monomer was built based on the matlab / simulink platform.

[0128] Add an oscilloscope module to the model and run it to record the changes in membrane electrode temperature under different ambient temperatures and starting current inputs.

[0129] The time it takes for the model to run after the membrane electrode temperature reaches the freezing point under different inputs is counted, and a data set D containing N groups of samples is obtained. , , )}, i=1,...,n historical dataset.

[0130] The second part is the construction of neural network.

[0131] The Min-Max normalization method is used to normalize the input features [ , ] and output They are normalized to the interval [0, 1] so that all features have the same numerical range, thus preventing some features from dominating the training process.

[0132] The normalization formula is as follows:

[0133] ;

[0134] in: is the original eigenvalue; and are the minimum and maximum values ​​of the feature respectively; is the normalized eigenvalue. After normalization, the numerical distribution of the input data is more uniform, which helps speed up the training of the neural network and improve the prediction accuracy of the model.

[0135] To evaluate the model's learning effectiveness and generalization capabilities, this example splits the dataset into a training set and a test set. 70% of the data is used for training, optimizing model parameters and learning data distribution characteristics. 30% of the data is used for testing, evaluating model performance on unseen data to prevent overfitting. A random partitioning strategy is used during data splitting to ensure the dataset covers a variety of possible operating conditions (e.g., varying temperatures and gas flow rates), enhancing the model's adaptability and robustness.

[0136] like Figure 5 As shown, this embodiment uses a feedforward neural network to optimize the cold start time prediction model and constructs a deep learning model for cold start time prediction. The specific structure is as follows:

[0137] The input layer contains two nodes, corresponding to two key input parameters related to cold start, ambient temperature and starting current;

[0138] , ;

[0139] The hidden layer contains 1 hidden layer and 60 neurons, which are used to capture the nonlinear relationship between input data and cold start time. The hidden layer uses a fully connected layer to implement nonlinear transformation. The ReLU activation function has the advantages of sparse activation, simple calculation, and less gradient disappearance problem, which enables the network to converge faster during training and effectively handle complex nonlinear relationships. According to the following formula:

[0140] ;

[0141] The output layer contains one node, which outputs the predicted cold start time of the fuel cell using linear regression. Since the cold start time is a continuous value, the linear activation function can directly output the prediction result, which is suitable for regression tasks. The specific calculation of the predicted cold start time is as follows:

[0142] ;

[0143] During model training, the mean square error (MSE) is used as the loss function to measure the error between the predicted value and the true value. The formula is as follows:

[0144] ;

[0145] in: is the actual cold start time; is the cold start time of neural network prediction; n is the total number of samples.

[0146] The mean squared error can more strongly penalize larger prediction errors, making the model focus more on accurate prediction of cold start time.

[0147] To improve the convergence speed of model training, this embodiment uses the Adam optimizer, which combines the advantages of momentum and adaptive learning rate. It can automatically adjust the learning rate of each parameter to adapt to different gradient changes, thereby improving training efficiency. The Adam formula is as follows:

[0148] ;

[0149] ;

[0150] ;

[0151] in: is the gradient; are the first-order and second-order moment estimates, respectively; is a hyperparameter, take ; For the learning rate, set the initial learning rate to 0.001.

[0152] The model is trained using the training set, with 300 training rounds set, that is, the network performs 300 complete iterations on the data to ensure that the data features are fully learned; the batch size is set to 64, that is, 64 samples are used for gradient calculation each time the weight is updated to balance computational efficiency and training stability.

[0153] Set up an early stopping mechanism to avoid errors in network training. Using the early stopping mechanism, if the verification error no longer decreases after 10 rounds, the training is stopped to prevent overfitting.

[0154] During the model optimization process, the weight initialization method has a significant impact on the model's convergence speed and final prediction performance. If the initial weight is too large, it may cause gradient explosion; if it is too small, it may cause gradient vanishing, making it difficult for the network to learn effective features. This embodiment uses He normal initialization, which is designed specifically for the ReLU function and can effectively alleviate the problem of gradient vanishing or gradient explosion. The mathematical expression of He normal initialization is as follows:

[0155] ;

[0156] in: is the neural network weight; It means that the mean is 0 and the standard deviation is Normal distribution; is the number of neurons in the previous layer.

[0157] The number of nodes (number of neurons) in a neural network's hidden layer directly affects the model's learning ability and computational complexity. In this embodiment, a grid search method was used to determine the optimal number of hidden layer nodes. First, several possible numbers of hidden layer neurons (32, 60, 128, and 256) were preset as candidate parameters. Next, models with different hidden layer sizes were trained on the training data, and their prediction accuracy was evaluated using cross-validation. The final configuration with the lowest test error was selected. Ultimately, this embodiment determined the optimal number of hidden layer nodes to be 60. This configuration allows the neural network to learn sufficient feature information while maintaining computational efficiency, achieving optimal prediction results.

[0158] To improve the generalization ability of the model, this embodiment introduces Dropout regularization in the hidden layer and sets the Dropout ratio to 0.2. The mathematical representation of the Dropout effect is as follows:

[0159] , if i is not dropped;

[0160] , if i is dropped;

[0161] in: is the neuron value after Dropout processing; is the activation value of the original neuron; is the Dropout ratio (set to 0.2 in this example).

[0162] After the network is trained with the training set, the test set is used to evaluate the model and verify the loss to determine the correctness of the model. The verification indicators are the root mean square error RMSE and the coefficient of determination. ,both verify that a lower value indicates that the model has a higher prediction accuracy for the cold start time of the fuel cell.

[0163] The root mean square error is calculated according to the following formula:

[0164] ;

[0165] The coefficient of determination is calculated according to the following formula:

[0166] ;

[0167] After training the neural network with the above parameters, the prediction results are highly consistent with the test set, and the model is accurate.

[0168] The trained model is then solidified into a prediction engine for use in predicting fuel cell cold start times. It receives real-time data from ambient temperature and current sensors to obtain the current and ambient temperature at that moment.

[0169] The neural network model is based on the collected and Perform forward propagation calculation and output the cold start time prediction value The calculation process is as follows:

[0170] ;

[0171] in: Represents a trained multilayer perceptron neural network; and As the input variable, after the nonlinear mapping of the neural network, the output is the predicted cold start time .

[0172] Finally, the predicted value is converted to its actual physical meaning through inverse normalization to obtain the predicted value of the fuel cell cold start time to be predicted. This can be used for control system decision-making, avoiding the numerical error caused by normalization and ensuring that the prediction result is consistent with the actual operation of the fuel cell.

[0173] In this embodiment, the input data is processed using Min-Max normalization, so its inverse normalization calculation is:

[0174] ;

[0175] in: and are the minimum and maximum cold start times in the training data; inverse normalization will Restore to the actual unit of seconds to obtain the actual fuel cell cold start time, which is used for fuel cell system decision-making.

[0176] On the other hand, this embodiment further provides an electronic device, comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the method is implemented when the processor executes the computing program.

[0177] On the other hand, this embodiment further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program implements the method when executed by a processor.

[0178] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for predicting the cold start time of a proton exchange membrane fuel cell, characterized in that: include: Construct a mechanism model for cold start time prediction based on fuel cells; The mechanism model includes: gas transmission model, output voltage model, temperature stratification model and water transmission model; The gas transmission model includes: a cathode gas transmission model and an anode gas transmission model, wherein the cathode gas transmission model and the anode gas transmission model are used to reflect the gas diffusion, migration and reaction process of the fuel cell in a low temperature environment; The output voltage model is used to obtain the voltage provided to the outside under actual operating conditions of the fuel cell under the influence of ice accumulation factors; The temperature stratification model is established based on the heat transfer relationship expression at the membrane electrode, and is used to calculate the membrane electrode temperature and ice melting heat; The water transport model is used to perform water balance calculations and ice accumulation calculations; Constructing a neural network model, and training the neural network model based on the mechanism model to obtain a prediction model, wherein the neural network model is a multilayer perceptron neural network; Predicting the cold start time of the proton exchange membrane fuel cell based on the prediction model and generating a prediction result; The process of obtaining the prediction model includes: The startup time and input quantity obtained based on the mechanism model are used as a historical data set, wherein the input quantity is the startup current and the ambient temperature; Normalizing the historical data set to obtain a normalized data set; Constructing a neural network model, and training the neural network model based on the normalized data set to obtain a training model; Evaluate the training model using a test set, calculate the mean square error and the coefficient of determination as indicators of model prediction accuracy, and obtain the prediction model; Establish a cathode gas transport model to calculate the mass flow of oxygen at the cathode inlet of the fuel cell; ； in, is the volume fraction of oxygen at the cathode inlet; is the air mass flow rate at the cathode inlet; , are the molar masses of oxygen and nitrogen, respectively; is the specific humidity of the gas at the cathode inlet; Establish an anode gas transmission model to calculate the hydrogen mass flow rate at the anode inlet of the fuel cell; ； in, is the gas mass flow rate at the anode inlet; is the specific humidity of the gas at the anode inlet.

2. The method according to claim 1, characterized in that The prediction model has an input layer with 2 nodes, a hidden layer with 60 neurons, and an output layer with 1 node.

3. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein: When the processor executes the computing program, the method described in any one of claims 1 to 2 is implemented.

4. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 2 is implemented.

Images