Proton exchange membrane fuel cell fusion prediction method under dynamic operating conditions

Abstract

This invention relates to the field of fuel cell technology and discloses a method for predicting proton exchange membrane fuel cell fusion under dynamic operating conditions. Under dynamic operating conditions, several groups of on-board fuel cells undergo dynamic load test cycles to obtain voltage data. Based on this data, a physical aging model is established in the time domain. Then, an algorithm is used to identify parameters and obtain dynamic aging indices (DI). Finally, a CPO-CNN-BiLSTM neural network is used to predict the short-term and long-term aging of the DI. This invention's method for predicting proton exchange membrane fuel cell fusion under dynamic operating conditions exhibits higher prediction accuracy in the short term and better prediction performance in the long term, more accurately reflecting the overall aging trend.

Description

Technical Field

[0001] This invention relates to the field of fuel cell technology, and specifically to a method for predicting the fusion of proton exchange membrane fuel cells under dynamic operating conditions. Background Technology

[0002] Traditional energy sources, primarily based on fossil fuels, generate significant amounts of greenhouse gases during energy conversion, causing numerous adverse environmental impacts. Proton exchange membrane fuel cells (PEMFCs) directly convert the energy in fuel into electrical energy, possessing superior performance characteristics such as cleanliness, high efficiency, and continuous energy conversion capabilities. Widely considered the "ultimate energy source" of the 21st century, they have found applications in transportation, military equipment, and aerospace. However, the high cost and short lifespan of PEMFCs are major factors hindering their large-scale commercialization. Improving the power and power density, cold start temperature, stack durability, and efficiency of PEMFC stacks has become a key research focus in fuel cell science, with fuel cell durability receiving increasing attention to meet the requirements of transportation applications. To mitigate the performance degradation of PEMFCs, in addition to material breakthroughs and improved stack design, health management systems are typically employed to maintain and extend their lifespan. Accurate prognostication of PEMFCs is crucial for health management, helping to monitor their health status and estimate their remaining useful life (RUL), thereby enabling timely management and repair of PEMFC systems. Prognosis is mainly divided into short-term prognosis and long-term prognosis. Short-term prediction can capture local changes with high accuracy, while long-term prediction can capture aging trends.

[0003] Aging trend prediction and RUL prediction for PEMFCs can be categorized into two types: model-based prediction methods and data-based prediction methods. Model-based methods attempt to establish mechanistic, empirical, or semi-empirical aging models for aging prediction, and the accuracy of the model plays a decisive role in the accuracy of the prediction results. Data-based methods do not need to consider the complex physical field information within PEMFCs; they only need to analyze historical data to learn the relationship between the past and the future, thereby making aging predictions. Artificial intelligence, due to its powerful learning capabilities, has been widely applied in various fields, and therefore, the use of artificial intelligence methods for aging prediction is receiving increasing attention. Currently, many machine learning-based prediction methods have been applied to PEMFC prediction, such as Adaptive Neural Fuzzy Inference System (ANFIS), Relevance Vector Machine (RVM), Echo State Network (ESN), and Deep Neural Network (DNN).

[0004] Currently, aging prediction of PEMFCs operating under constant load conditions (static and quasi-dynamic) has been extensively studied; however, research on automotive PEMFCs under dynamic conditions is relatively limited. When PEMFCs operate under constant load conditions, the stack voltage of PEMFCs exhibits a significant decreasing trend with FC degradation, thus the stack voltage can be directly used as the aging index (DI) for aging prediction. However, under dynamic conditions, the stack voltage is not only affected by the degree of FC degradation but also depends on time-varying operating conditions and highly dynamic loads, making direct use of the stack voltage as the DI no longer effective. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the above-mentioned technologies by providing a method for predicting the fusion of proton exchange membrane fuel cells under dynamic operating conditions. This method has higher prediction accuracy in the short term and better prediction performance in the long term, and can more accurately reflect the overall aging trend.

[0006] To achieve the above objectives, the proton exchange membrane fuel cell fusion prediction method under dynamic operating conditions involved in this invention performs dynamic load test cycles on several groups of on-board fuel cells under dynamic operating conditions to obtain voltage data. Based on this data, a physical aging model is established in the time domain. Then, an algorithm is used to identify parameters and obtain dynamic DI. Finally, a CPO-CNN-BiLSTM neural network is used to predict the short-term and long-term aging of DI.

[0007] Preferably, the method includes the following steps:

[0008] S1) Collection and analysis of voltage aging data: Determine the durability test standard, conduct dynamic durability and polarization characterization tests on the test station, collect various state data of proton exchange membrane fuel cells, and analyze the voltage, current and hydrogen inlet and outlet pressures collected in the dynamic durability test.

[0009] S2) Establish a physical aging model for PEMFCs to identify parameters: Establish a physical aging model in the time domain to explore the changing trends of each parameter in each dynamic load test cycle.

[0010] S3) Extraction and preprocessing of dynamic DI: Improve the gray wolf optimization algorithm, identify parameters for each dynamic load test cycle of the physical aging model proposed in step S2), analyze the changing trend of each parameter, select a suitable dynamic DI, and filter the DI.

[0011] S4) Neural network construction and prediction strategy for aging prediction: Combining the spatial extraction capability of convolutional neural networks and the time series prediction capability of BiLSTM, a CNN-BiLSTM model is constructed. The hyperparameters of the CNN-BiLSTM model are optimized using the Crowned Porcupine optimization algorithm to obtain the CPO-CNN-BiLSTM neural network.

[0012] S5) Short-term and long-term aging prediction results: The CPO-CNN-BiLSTM neural network in step S4) is used to predict the short-term and long-term aging of DI respectively.

[0013] Preferably, in step S2), the physical aging model of PEMFCs is:

[0014]

[0015] Where i is the stack current, i loss It is leakage current, i ex It is the alternating current, i L It limits the current, E cell It is the voltage of a single cell, E ocv Here, R is the open-circuit voltage, T is the operating temperature, R is the molar gas constant, F is the Faraday constant, and α is a parameter related to the electrode charge transfer coefficient. eq It is the equivalent resistance, B c It is a parameter related to the number of electrons transferred in the overall reaction.

[0016] Preferably, in step S3), the improved gray wolf optimization algorithm based on Lévy flight and dynamic weight strategy is introduced for parameter identification, including the following steps:

[0017] S301) Initialize parameters, including population size, maximum iteration step size, and convergence factor extreme value;

[0018] S302) The population position is initialized using chaotic mapping, and the values ​​of each parameter are randomly generated;

[0019] S303) Substitute the parameter values ​​into the physical aging model to calculate the theoretical voltage;

[0020] S304) Calculate the population fitness value, and update the global optimal solution and individual optimal solution based on the fitness value;

[0021] S305) Update the population's position and velocity;

[0022] S306) Repeat steps S303) to S305) until the stopping condition is met, and output the global optimal solution, i.e. the optimal parameter value.

[0023] Preferably, in step S3), based on the physical aging model, parameter identification is performed on the data in each dynamic load test cycle, and the parameters α and R are... eq The overall aging is represented by dynamic DI.

[0024] Preferably, in step S3), when performing dynamic DI preprocessing, Gaussian filtering is used to process the original data, wherein the size of the Gaussian kernel is 20.

[0025] Preferably, in step S4), CNN-BiLSTM is used as a sequence prediction model, and CPO is used to optimize the number of CNN convolution kernels, the number of BiLSTM units, the initial learning rate, and the L2 regularization coefficient in the model.

[0026] Preferably, in step S5), the prediction strategy of the CPO-CNN-BiLSTM neural network is to use the filtered DI parameters α and R... eq As input to the neural network, training and test sets were divided for training and prediction comparison of the CPO-CNN-BiLSTM neural network, respectively. The data first passed through the convolutional, activation, and pooling layers in the CNN, and then was input into the forward and backward LSTM networks in the BiLSTM. Finally, the entire network was connected using fully connected layers. The hyperparameters of the CNN-BiLSTM model were optimized using CPO. The optimization ranges for the number of CNN convolutional kernels, the number of BiLSTM units, the initial learning rate, and the L2 regularization coefficient were [20, 60], [30, 90], [0.001, 0.01], and [0.001, 0.01], respectively. The network parameters were updated, and the network was trained with the new parameters. Short-term and long-term aging predictions were then performed.

[0027] Compared with the prior art, the present invention has the following advantages:

[0028] 1. Short-term forecasts have higher forecast accuracy, while long-term forecasts show better forecast performance and can more accurately reflect the overall aging trend.

[0029] 2. Addressing the issue that the stack voltage of automotive PEMFCs cannot be directly used as an aging indicator to represent battery aging under dynamic operating conditions, a method for extracting dynamic aging indicators is proposed. This method involves establishing a physical aging model in the time domain and using an improved Grey Wolf optimization algorithm to identify the parameters in the aging model, successfully extracting two aging indicators: the equivalent resistance R. eq and electrode charge transfer coefficient α;

[0030] 3. The CNN-BiLSTM neural network optimized by the Crowned Porcupine (CPO) algorithm utilizes the powerful global search capability of CPO to optimize the hyperparameters in the CNN-BiLSTM network, which greatly improves the prediction ability of the network.

[0031] 4. In short-term prediction of aging indicators, the prediction results of the optimized CNN-BiLSTM network are compared with those of a single network. The optimized network has higher prediction accuracy. The overall prediction error of the prediction results after converting the aging indicators into voltage remains at a low level. In long-term prediction, CNN-BiLSTM also shows better prediction performance and can more accurately represent the overall aging trend. Attached Figure Description

[0032] Figure 1 This is a graph showing the overall voltage and anode pressure changes of PEMFCs under dynamic operating conditions in the dynamic operating condition prediction method of the present invention.

[0033] Figure 2 This is a graph showing the changes in dynamic aging indicators in the proton exchange membrane fuel cell fusion prediction method under dynamic operating conditions of the present invention.

[0034] Figure 3 This is a graph showing the short-term prediction results of two DI methods when the training set accounts for 50% in the proton exchange membrane fuel cell fusion prediction method under dynamic operating conditions of the present invention.

[0035] Figure 4 This is a graph showing the short-term prediction results of voltage converted from different training set proportions in the proton exchange membrane fuel cell fusion prediction method under dynamic operating conditions of the present invention.

[0036] Figure 5 This is a graph showing the long-term prediction results of two DI methods in the proton exchange membrane fuel cell fusion prediction method under dynamic operating conditions of the present invention when the training set length is 450h. Detailed Implementation

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

[0038] A dynamic operating condition proton exchange membrane fuel cell fusion prediction method is proposed. Under dynamic operating conditions, several groups of on-board fuel cells are subjected to dynamic load test cycles to obtain voltage data. Based on this data, a physical aging model is established in the time domain. Then, the algorithm is used to identify parameters and obtain dynamic DI. Finally, the CPO-CNN-BiLSTM neural network is used to predict the short-term and long-term aging of DI.

[0039] Specifically, it includes the following steps:

[0040] S1) Collection and analysis of voltage aging data: Determine the durability test standard, conduct dynamic durability and polarization characterization tests on the test station, collect various state data of proton exchange membrane fuel cells, and analyze the voltage, current and hydrogen inlet and outlet pressures collected in the dynamic durability test.

[0041] S2) Establish a physical aging model for PEMFCs to identify parameters: Establish a physical aging model in the time domain to explore the changing trends of each parameter in each dynamic load test cycle.

[0042] S3) Extraction and preprocessing of dynamic DI: Improve the gray wolf optimization algorithm, identify parameters for each dynamic load test cycle of the physical aging model proposed in step S2), analyze the changing trend of each parameter, select a suitable dynamic DI, and filter the DI.

[0043] S4) Neural network construction and prediction strategy for aging prediction: Combining the spatial extraction capability of convolutional neural networks and the time series prediction capability of BiLSTM, a CNN-BiLSTM model is constructed. The hyperparameters of the CNN-BiLSTM model are optimized using the Crowned Porcupine optimization algorithm to obtain the CPO-CNN-BiLSTM neural network.

[0044] S5) Short-term and long-term aging prediction results: The CPO-CNN-BiLSTM neural network in step S4) is used to predict the short-term and long-term aging of DI respectively.

[0045] This embodiment is based on the measured voltage data of 3076 fuel cell dynamic load test cycles (FC-DLC) of on-board fuel cell stacks under dynamic operating conditions. A physical aging model is established in the time domain. Then, the improved gray wolf optimization algorithm is used to identify parameters and obtain dynamic DI. Finally, the CPO-CNN-BiLSTM neural network is used to predict the short-term and long-term aging of DI.

[0046] Specifically, the dynamic durability test data acquisition includes the following: The entire durability experiment consists of 3076 sets of fuel cell dynamic load test cycles, approximately 1008 hours. The sampling frequency for the PEMFCs durability test is 1Hz. Each FC-DLC cycle lasts 1180s and includes 35 test steps. Each test cycle consists of 152 FC-DLC cycles (approximately 50 hours). Characterization tests are then performed. After completing every 304 FC-DLC cycles, there is a 12-hour rest period to simulate the shutdown behavior of the fuel cell vehicle.

[0047] During data analysis, because each FC-DLC cycle contains 35 test steps, the current and voltage will change depending on the test step. Therefore, the voltage over the entire time cannot be directly used as the dynamic DI. The optimal value for the hydrogen inlet and outlet pressure is 110 kPa. Figure 1As shown in the hydrogen inlet and outlet pressure diagram, a violent oscillation of hydrogen inlet pressure occurred around 150 hours later, causing a voltage drop during shutdown. After 650 hours, the entire PEMFCs system experienced abnormal gas pressure / flow operation, which inevitably changed the operating status and increased the difficulty of prediction.

[0048] In step S2), since the output voltage of a fuel cell is one of the most important physical quantities for measuring fuel cell performance, but the aging trend of PEMFCs cannot be found from the overall voltage change trend graph, voltage cannot be directly used as DI under dynamic operating conditions. The entire durability test includes 3076 FC-DLCs, each with the same current change. An accurate physical aging model is established for each FC-DLC, and the dynamic DI is found by exploring the changing trends of parameters in the aging model. The physical aging model of PEMFCs is as follows:

[0049]

[0050] Where i is the stack current, i loss It is leakage current, i ex It is the alternating current, i L It limits the current, E cell It is the voltage of a single cell, E ocv Here, R is the open-circuit voltage, T is the operating temperature, R is the molar gas constant, F is the Faraday constant, and α is a parameter related to the electrode charge transfer coefficient. eq It is the equivalent resistance, B c It is a parameter related to the number of electrons transferred in the overall reaction.

[0051] In step S3), to address the issues of insufficient global search capability and susceptibility to local optima in solving complex optimization problems, the Grey Wolf algorithm is improved by introducing Lévy flight and dynamic weighting strategies for parameter identification, including the following steps:

[0052] S301) Initialize parameters, including population size, maximum iteration step size, and convergence factor extreme value;

[0053] S302) The population position is initialized using chaotic mapping, and the values ​​of each parameter are randomly generated;

[0054] S303) Substitute the parameter values ​​into the physical aging model to calculate the theoretical voltage;

[0055] S304) Calculate the population fitness value, and update the global optimal solution and individual optimal solution based on the fitness value;

[0056] S305) Update the population's position and velocity;

[0057] S306) Repeat steps S303) to S305) until the stopping condition is met, and output the global optimal solution, i.e. the optimal parameter value.

[0058] In step S3), based on the physical aging model, parameter identification is performed on 1180 sets of data in each dynamic load test cycle. From the results of all parameter identifications, it can be seen that α and R... eq It exhibits a clear trend over time, continuing until 640 hours ago, from Figure 1 It can be seen that abnormal gas pressure / flow rate operation occurred after 640 hours, combined with Figure 2 As shown, this leads to α and R eq After 640 hours, a phenomenon emerged that was significantly contrary to the overall trend. Therefore, parameters α and R were used. eq The overall aging is represented by dynamic DI, where R eq It is called DI1, and α is called DI2.

[0059] In step S3), during dynamic DI preprocessing, the large amount of data in the original parameters, which contains spikes and a lot of noise, can lead to large errors in the calculation results and excessively long calculation time. In order to improve the prediction accuracy and reduce the computational burden, Gaussian filtering is used to process the original data, with a Gaussian kernel size of 20. This method has a strong noise suppression effect and can effectively preserve the useful information of the original signal.

[0060] In step S4), CNN-BiLSTM is used as a sequence prediction model, and CPO is used to optimize the number of CNN convolution kernels, the number of BiLSTM units, the initial learning rate, and the L2 regularization coefficient in the model.

[0061] The Crested Porcupine Optimization (CPO) algorithm is a metaheuristic algorithm that simulates four different protection mechanisms of the crested porcupine (CP). From least aggressive to most aggressive, the CP employs four distinct protection mechanisms: visual, auditory, olfactory, and physical attack. The first two mechanisms (visual and auditory) reflect the exploratory behavior of the CPO, while the latter two (olfactory and physical attack) reflect its exploitative behavior. The specific steps of the CPO algorithm are as follows:

[0062] In reality, not all CPs activate their defense mechanisms; only those CPs under threat do. The cyclic population reduction technique effectively simulates this situation, not only accelerating convergence but also maintaining population diversity. Its mathematical model is as follows:

[0063]

[0064] In the formula, T is the variable that determines the number of iterations, and t is the current function evaluation.max The maximum value of the function evaluation is %, % represents the remainder or modulo operator, and N is the maximum value of the function evaluation. min This is the smallest number of individuals in the newly formed population. The following describes four specific defense strategies:

[0065] 1) First defense strategy

[0066] When a predator (CP) becomes aware of a predator, it extends its sharp spines to create a visual impact. When the predator chooses to approach the CP, the distance between them decreases, encouraging exploration of the area between them to accelerate convergence. When the predator chooses to leave, the distance between them is maximized, encouraging exploration of more distant areas to identify unexplored territories. This can be represented by the following mathematical formula:

[0067]

[0068] in It is the optimal solution for estimating the function t. This is a vector generated between the current CP and a CP randomly selected from the population, representing the predator's position at iteration t. τ1 is a random number based on a normal distribution, and τ2 is a random value in the interval [0,1]. The mathematical formula is as follows:

[0069]

[0070] In the formula, r is a random number between [1, N];

[0071] 2) Second defense strategy

[0072] The CP uses sound to create noise and threaten predators, increasing the volume as the predator approaches. To mathematically model this behavior, the following formula is proposed:

[0073]

[0074] In the formula, r1 and r2 are two random integers between [1, N], and τ3 is a random value generated between 0 and 1;

[0075] 3) Third defense strategy

[0076] In this strategy, the CP secretes a foul odor that diffuses in its surrounding area to prevent predators from approaching it. The mathematical model is as follows:

[0077]

[0078] Where r3 is a random number between [1, N], and τ3 is a random value within the interval [0, 1]. Let be the position of the i-th individual at iteration t. It is a parameter used to control the search direction, γ t As a defensive factor, The three factors, namely odor diffusion factor, are expressed as follows:

[0079]

[0080]

[0081] in This represents the objective function value of the i-th individual at iteration t, where ε is a small value to avoid being divisible by zero. `rand` is a vector containing randomly generated numbers between 0 and 1, `rand` is a variable containing randomly generated numbers between 0 and 1, `N` is the population size, and `t` is the current iteration number. max It is the maximum number of iterations;

[0082] 4) Fourth Defense Strategy

[0083] When predators get very close, CP will use its spikes to launch a physical attack on the predator, which manifests as a one-dimensional inelastic collision. This physical attack behavior can be represented by the following mathematical formula:

[0084]

[0085] in It is the optimal solution, representing CP. is the position of the i-th individual at iteration t, representing the predator at that position; α is the convergence rate factor discussed later in the parameter setting section; and τ4 is a random value within the interval [0,1]. It is the average force exerted by CP on the i-th predator.

[0086] In step S5), the prediction strategy of the CPO-CNN-BiLSTM neural network is to use the filtered DI parameters α and R... eq As input to the neural network, training and test sets were divided for training and prediction comparison of the CPO-CNN-BiLSTM neural network, respectively. The data first passed through the convolutional, activation, and pooling layers in the CNN, and then was input into the forward and backward LSTM networks in the BiLSTM. Finally, the entire network was connected using fully connected layers. The hyperparameters of the CNN-BiLSTM model were optimized using CPO. The optimization ranges for the number of CNN convolutional kernels, the number of BiLSTM units, the initial learning rate, and the L2 regularization coefficient were [20, 60], [30, 90], [0.001, 0.01], and [0.001, 0.01], respectively. The network parameters were updated, and the network was trained with the new parameters. Short-term and long-term aging predictions were then performed.

[0087] Aging prediction for proton exchange membrane fuel cells is mainly divided into short-term and long-term prediction. Short-term prediction can capture local changes with high accuracy, typically with a prediction period of several hours or days. Long-term prediction can capture aging trends, usually with a prediction range of several hundred hours or months. Due to the dynamically changing load conditions, long-term prediction faces greater challenges. Performance evaluation metrics commonly used in regression problems include root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 The calculation formulas for the three evaluation indicators are as follows:

[0088]

[0089] The method of this invention is compared with CNN-BiLSTM and BiLSTM without parameter optimization at training set percentages of 40%, 50%, 60%, and 70%, respectively, to verify the superiority of the proposed method in prediction in DI.

[0090] Figure 3 The results of three methods for short-term prediction of two DIs are shown when the training set accounts for 50% of the total training set. Figure 3 (a) is the equivalent resistance. Figure 3 (b) represents the parameters related to the electrode charge transfer coefficient. After 640 hours, the abnormal gas pressure caused drastic changes in the aging parameters, making prediction more difficult. Compared with the CNN-BiLSTM and BiLSTM algorithms, the method proposed in this invention can predict the changing trend very well.

[0091] Because the dynamic dataset is very large, few people use the raw voltage data directly for short-term prediction. Therefore, this invention converts the parameters into voltage through an aging model, where DI1 and DI2 are the predicted values, and the other parameters are the average values ​​of the training set. Figure 4 The voltage prediction results were converted to different training set proportions. The error was larger in the black box between approximately 750h and 1000h, due to the prolonged operation of the fuel cell stack and drastic changes in inlet pressure, leading to a deterioration in the fitting performance of the aging model under full load conditions. Therefore, as the training set proportion increases, the prediction error also increases due to the worse fitting performance of the aging model under full load conditions. However, because the aging model still fits well under other load conditions, the overall prediction error remained at a low level, with R² values ​​greater than 0.99, and RMSE and MAE values ​​less than 0.01 and 0.007, respectively.

[0092] Regarding the long-term aging prediction results, when evaluating the performance of the proposed fusion method in long-term prediction, since the degradation trend of DI also changes due to the change in operating conditions after 640h, the termination time of this test at 640h, i.e., 650h, is selected as the end time of lifetime, and the data after 650h is set as the failure value. Figure 5 The results show the long-term predictions of CPO-CNN-BiLSTM, CNN-BiLSTM, and BiLSTM networks on the following 200 hours of data, using the first 450 hours of training data. Figure 5 (a) is the equivalent resistance, Figure 5 (b) represents the charge transfer coefficient; blue data represents usable data, and gray data represents failed data. During testing, each shutdown causes drastic changes in the internal parameters of the fuel cell stack, leading to parameter recovery, which increases the difficulty of long-term prediction. Figure 5 As can be seen, compared to a single BiLSTM network, both the original CNN-BiLSTM and the CPO-optimized CNN-BiLSTM can roughly predict the future trend of DI. The CPO-optimized CNN-BiLSTM predicts the overall trends of DI1 and DI2 more accurately than the unoptimized network, and accurately predicts the value at the end of the lifetime, resulting in a significant improvement in long-term prediction performance.

[0093] This invention uses two error metrics, RMSE and MAE, to evaluate the long-term prediction performance of the network. The long-term prediction errors of the three algorithms on a 450-hour training set are shown in Table 1. As can be seen from Table 1, for DI1, the CPO-CNN-BiLSTM network reduced the long-term prediction RMSE and MAE by 15.53% and 14.81%, respectively, compared to the CNN-BiLSTM network. For DI2, the CPO-CNN-BiLSTM network reduced the long-term prediction RMSE and MAE by 12.5% ​​and 14.29%, respectively, compared to the CNN-BiLSTM network.

[0094] Table 1 shows the long-term prediction error results of the three algorithms on a 450-hour training set.

[0095]

[0096] This invention presents a dynamic prediction method for proton exchange membrane fuel cells (PEMFCs). It exhibits higher prediction accuracy in the short term and better prediction performance in the long term, more accurately reflecting the overall aging trend. Addressing the issue that stack voltage cannot be directly used as an aging indicator for vehicle-mounted PEMFCs under dynamic conditions, a method for extracting dynamic aging indicators is proposed. This involves establishing a physical aging model in the time domain and using an improved gray wolf optimization algorithm to identify parameters in the aging model, successfully extracting two aging indicators: the equivalent resistance R.eq The electrode charge transfer coefficient α; the CNN-BiLSTM neural network optimized by the Crowned Porcupine Algorithm (CPO) utilizes the powerful global search capability of CPO to optimize the hyperparameters in the CNN-BiLSTM network, greatly improving the network's predictive ability; in short-term prediction of aging indicators, the prediction results of the optimized CNN-BiLSTM network are compared with those of a single network. The optimized network has higher prediction accuracy, and the overall prediction error of the prediction results after converting aging indicators into voltage remains at a low level. In long-term prediction, CPO-CNN-BiLSTM also shows better prediction performance and can more accurately represent the overall aging trend.

Claims

1. A method for predicting the fusion of proton exchange membrane fuel cells under dynamic operating conditions, characterized in that: Under dynamic operating conditions, several sets of on-board fuel cells were subjected to dynamic load test cycles to obtain voltage data. Based on this data, a physical aging model was established in the time domain. The physical aging model is as follows: ,in, It is the fuel cell current. It is leakage current. It is an alternating current. It limits the current. It is the voltage of a single battery. It is open-circuit voltage. T It is the operating temperature. R It is the molar gas constant. F It is Faraday's constant. These are parameters related to the electrode charge transfer coefficient. It is the equivalent resistance. These parameters are related to the number of electrons transferred in the overall reaction. Then, based on a physical aging model, the Grey Wolf optimization algorithm is improved using the Lévy flight and dynamic weighting strategy to identify parameters in the data from each dynamic load test cycle, and select the appropriate parameters. and The overall aging is represented by dynamic DI, and finally, the CPO-CNN-BiLSTM neural network is used to predict the short-term and long-term aging of DI.

2. The method for predicting proton exchange membrane fuel cell fusion under dynamic operating conditions as described in claim 1, characterized in that: Includes the following steps: S1) Collection and analysis of voltage aging data: Determine the durability test standard, conduct dynamic durability and polarization characterization tests on the test station, collect various state data of proton exchange membrane fuel cells, and analyze the voltage, current and hydrogen inlet and outlet pressures collected in the dynamic durability test. S2) Establish a physical aging model for PEMFCs to identify parameters: Establish a physical aging model in the time domain to discover the changing trends of each parameter in each dynamic load test cycle; S3) Extraction and preprocessing of dynamic DI: Improve the gray wolf optimization algorithm, identify parameters for each dynamic load test cycle of the physical aging model proposed in step S2), analyze the changing trend of each parameter, select a suitable dynamic DI, and filter the DI. S4) Neural network construction and prediction strategy for aging prediction: Combining the spatial extraction capability of convolutional neural networks and the time series prediction capability of BiLSTM, a CNN-BiLSTM model is constructed. The hyperparameters of the CNN-BiLSTM model are optimized using the Crowned Porcupine optimization algorithm to obtain the CPO-CNN-BiLSTM neural network. S5) Short-term and long-term aging prediction results: The CPO-CNN-BiLSTM neural network in step S4) is used to predict the short-term and long-term aging of DI respectively.

3. The method for predicting proton exchange membrane fuel cell fusion under dynamic operating conditions as described in claim 2, characterized in that: In step S3), the Lévy flight and dynamic weight strategy are used to improve the gray wolf optimization algorithm for parameter identification, including the following steps: S301) Initialize parameters, including population size, maximum iteration step size, and convergence factor extreme value; S302) The population position is initialized using chaotic mapping, and the values ​​of each parameter are randomly generated; S303) Substitute the parameter values ​​into the physical aging model to calculate the theoretical voltage; S304) Calculate the population fitness value, and update the global optimal solution and individual optimal solution based on the fitness value; S305) Update the population's position and velocity; S306) Repeat steps S303) to S305) until the stopping condition is met, and output the global optimal solution, i.e. the optimal parameter value.

4. The method for predicting proton exchange membrane fuel cell fusion under dynamic operating conditions as described in claim 2, characterized in that: In step S3), when performing dynamic DI preprocessing, Gaussian filtering is used to process the original data, where the size of the Gaussian kernel is 20.

5. The method for predicting proton exchange membrane fuel cell fusion under dynamic operating conditions as described in claim 4, characterized in that: In step S4), CNN-BiLSTM is used as a sequence prediction model, and CPO is used to optimize the number of CNN convolution kernels, the number of BiLSTM units, the initial learning rate, and the L2 regularization coefficient in the model.

6. The method for predicting proton exchange membrane fuel cell fusion under dynamic operating conditions as described in claim 5, characterized in that: In step S5), the prediction strategy of the CPO-CNN-BiLSTM neural network is to filter the DI parameters. and As input to the neural network, training and test sets were divided for training and prediction comparison of the CPO-CNN-BiLSTM neural network, respectively. The data first passed through the convolutional, activation, and pooling layers in the CNN, and then was input into the forward and backward LSTM networks in the BiLSTM. Finally, the entire network was connected using fully connected layers. The hyperparameters of the CNN-BiLSTM model were optimized using CPO. The optimization ranges for the number of CNN convolutional kernels, the number of BiLSTM units, the initial learning rate, and the L2 regularization coefficient were [20, 60], [30, 90], [0.001, 0.01], and [0.001, 0.01], respectively. The network parameters were updated, and the network was trained with the new parameters. Short-term and long-term aging predictions were then performed.

Images