Fusion prediction method for proton exchange membrane fuel cell under dynamic working condition

Abstract

The invention relates to the technical field of fuel cells, and discloses a proton exchange membrane fuel cell fusion prediction method under a dynamic working condition, and the method comprises the steps: carrying out the dynamic load test cycle of a plurality of groups of vehicle-mounted fuel cells under the dynamic working condition, obtaining voltage data, and building a physical aging model in a time domain based on the voltage data; then, parameter identification is carried out by using an algorithm to obtain dynamic DI; and finally, short-term and long-term aging prediction is carried out on the DI by using a CPO-CNN-BiLSTM neural network. According to the fusion prediction method for the proton exchange membrane fuel cell under the dynamic working condition, short-term prediction has higher prediction precision, long-term prediction shows better prediction performance, and the overall aging trend can be shown more accurately.

Description

Technical Field

[0001] The present invention relates to the technical field of fuel cells, and in particular to a method for predicting the fusion of proton exchange membrane fuel cells under dynamic conditions. Background Art

[0002] Traditional energy sources based on fossil materials produce a large amount of greenhouse gases during the energy conversion process, which has many adverse effects on the environment. Proton exchange membrane fuel cells (PEMFCs) can directly convert the energy in fuel into electrical energy. They have various excellent properties such as clean, efficient and continuous energy conversion capabilities. They are widely regarded as the "ultimate energy" of the 21st century and are used in transportation, military equipment, aerospace and other fields. However, the high cost and short life of PEMFCs have become the main reasons restricting their large-scale commercial development. Improving the power and power density of PEMFCs stacks, cold start temperature, durability and efficiency of stacks has become a key research direction in the field of fuel cell science research. Among them, the durability of fuel cells has received increasing attention in meeting the requirements of transportation applications. In order to slow down the performance degradation of PEMFCs, in addition to achieving material breakthroughs and improving the design of the stack, health management systems are usually used to maintain and extend the life. Accurate prognosis of PEMFCs is the key to health management, which helps to monitor the health status of PEMFCs and estimate their remaining useful life (RUL), thereby achieving timely management and repair of PEMFCs systems. Prognosis is mainly divided into short-term prognosis and long-term prognosis. Short-term prediction can capture local changes with higher accuracy, while long-term prediction can capture aging trends.

[0003] The aging trend prediction and RUL prediction of PEMFCs can be divided into two types: model-based prediction methods and data-based prediction methods. Model-based methods attempt to establish mechanism, empirical or semi-empirical aging models for aging prediction. The accuracy of the model will have a decisive effect on the accuracy of the aging prediction results. Data-based methods do not need to consider the complex physical field information inside PEMFCs. They only need to analyze historical data and learn the connection between history and the future to make aging predictions. Artificial intelligence has been widely used in various fields due to its powerful learning ability. Therefore, the use of artificial intelligence methods for aging prediction has received more and more attention. At present, many prediction methods based on machine learning have been applied to the prediction of PEMFCs, such as adaptive neural fuzzy inference system (ANFIS), relevance vector machine (RVM), echo state network (ESN) and deep neural network (DNN).

[0004] Nowadays, the aging prediction of PEMFCs under constant load conditions (static and quasi-dynamic) has been widely studied, but the research on vehicle-mounted PEMFCs under dynamic conditions is relatively small. When PEMFCs work under constant load conditions, the stack voltage of PEMFCs will show a significant downward trend with FC degradation, so the stack voltage can be directly selected 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. Directly using the stack voltage as DI is no longer effective in this case. Summary of the invention

[0005] The purpose of the present invention is to address the deficiencies of the above-mentioned technology and provide a proton exchange membrane fuel cell fusion prediction method under dynamic conditions. The short-term prediction has higher prediction accuracy, the long-term prediction shows better prediction performance, and can more accurately show the overall aging trend.

[0006] To achieve the above-mentioned purpose, the present invention relates to a proton exchange membrane fuel cell fusion prediction method under dynamic conditions. Under dynamic conditions, several groups of vehicle-mounted fuel cells are subjected to dynamic load test cycles to obtain voltage data, and a physical aging model is established in the time domain based on this. Then, an algorithm is used to perform parameter identification to obtain dynamic DI, and finally, the CPO-CNN-BiLSTM neural network is used to perform short-term and long-term aging predictions on DI.

[0007] Preferably, the method comprises 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 the proton exchange membrane fuel cell, and analyze the voltage, current and hydrogen inlet and outlet pressures collected in the dynamic durability test;

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

[0010] S3) Extraction and preprocessing of dynamic DI: Improve the Gray Wolf optimization algorithm, perform parameter identification on each dynamic load test cycle for the physical aging model proposed in step S2), analyze the trend of each parameter change, select the appropriate dynamic DI, and filter the DI;

[0011] S4) Establishment of neural network and prediction strategy for aging prediction: Combining the spatial extraction ability of convolutional neural network and the time series prediction ability of BiLSTM, a CNN-BiLSTM model was constructed, and the crown porcupine optimization algorithm was used to optimize the hyperparameters of the CNN-BiLSTM model 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 perform short-term and long-term aging predictions on DI respectively.

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

[0014]

[0015] Where i is the stack current, i loss is the leakage current, i ex is the exchange current, i L is the limiting current, E cell is the single cell voltage, E ocv is the open circuit voltage, T is the operating temperature, R is the molar gas constant, F is the Faraday constant, α is a parameter related to the electrode charge transfer coefficient, and R eq 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 Levy flight and dynamic weight strategy are introduced to improve the Grey Wolf Optimization Algorithm for parameter identification, which includes the following steps:

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

[0018] S302) using chaotic mapping to initialize population positions and randomly generate various parameter values;

[0019] S303) Substituting the parameter value into the physical aging model to calculate the theoretical voltage;

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

[0021] S305) updating the position and speed of the population;

[0022] S306) Repeat step S303) to step S305) until the stop condition is met, and output the global optimal solution, that is, 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 eq As a dynamic DI to represent the overall aging.

[0024] Preferably, in step S3), when dynamic DI preprocessing is performed, 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 transform the filtered DI parameters α and R eq As the input of the neural network, the training set and test set are divided for the training and prediction comparison of the CPO-CNN-BiLSTM neural network respectively. The data first passes through the convolution layer, activation layer and pooling layer in the CNN, and then is input into the forward and reverse LSTM networks in the BiLSTM. Finally, the fully connected layer is used to connect the entire network. The hyperparameters of the CNN-BiLSTM model are optimized using CPO. The optimization ranges for the number of CNN convolution kernels, the number of BiLSTM units, the initial learning rate and the L2 regularization coefficient are [20, 60], [30, 90], [0.001, 0.01] and [0.001, 0.01] respectively. The network parameters are updated, and the network is trained with the new parameters, and then short-term and long-term aging predictions are performed respectively.

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

[0028] 1. Short-term prediction has higher prediction accuracy, and long-term prediction shows better prediction performance, which can more accurately show the overall aging trend;

[0029] 2. In view of the fact that the stack voltage of on-board PEMFCs cannot be directly used as an aging indicator to represent the aging of the battery under dynamic conditions, a method for extracting dynamic aging indicators is proposed, that is, a physical aging model is established in the time domain, and the parameters in the aging model are identified using the improved Gray Wolf optimization algorithm. The equivalent resistance R of the two aging indicators is successfully extracted. eq and the electrode charge transfer coefficient α;

[0030] 3. The CNN-BiLSTM neural network optimized by the crested porcupine algorithm (CPO) uses the powerful global search capability of CPO to optimize the hyperparameters in the CNN-BiLSTM network, greatly improving the prediction ability of the network;

[0031] 4. In the 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 is kept at a low level. In the long-term prediction, CPO-CNN-BiLSTM also shows better prediction performance and can more accurately show the overall aging trend. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a graph showing the changes in the overall voltage and anode pressure of PEMFCs under dynamic conditions in the proton exchange membrane fuel cell fusion prediction method under dynamic conditions of the present invention;

[0033] Figure 2 A graph showing changes in dynamic aging indicators in the proton exchange membrane fuel cell fusion prediction method under dynamic conditions of the present invention;

[0034] Figure 3 It is a short-term prediction result diagram of two DIs when the training set accounts for 50% in the proton exchange membrane fuel cell fusion prediction method under dynamic conditions of the present invention;

[0035] Figure 4 It is a short-term prediction result diagram of voltage converted from different training set proportions in the proton exchange membrane fuel cell fusion prediction method under dynamic conditions of the present invention;

[0036] Figure 5 This is a long-term prediction result diagram of two DIs when the training set length is 450h in the proton exchange membrane fuel cell fusion prediction method under dynamic conditions of the present invention. DETAILED DESCRIPTION

[0037] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0038] A proton exchange membrane fuel cell fusion prediction method under dynamic conditions is proposed. Under dynamic conditions, dynamic load test cycles are performed on several groups of vehicle-mounted fuel cells to obtain voltage data. Based on this, a physical aging model is established in the time domain. Then, an algorithm is used to perform parameter identification to obtain dynamic DI. Finally, the CPO-CNN-BiLSTM neural network is used to perform short-term and long-term aging predictions on DI.

[0039] Specifically, the steps include:

[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 the proton exchange membrane fuel cell, and analyze the voltage, current and hydrogen inlet and outlet pressures collected in the dynamic durability test;

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

[0042] S3) Extraction and preprocessing of dynamic DI: Improve the Gray Wolf optimization algorithm, perform parameter identification on each dynamic load test cycle for the physical aging model proposed in step S2), analyze the trend of each parameter change, select the appropriate dynamic DI, and filter the DI;

[0043] S4) Establishment of neural network and prediction strategy for aging prediction: Combining the spatial extraction ability of convolutional neural network and the time series prediction ability of BiLSTM, a CNN-BiLSTM model was constructed, and the crown porcupine optimization algorithm was used to optimize the hyperparameters of the CNN-BiLSTM model 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 perform short-term and long-term aging predictions on DI respectively.

[0045] This embodiment is based on the measured voltage data of 3076 groups of fuel cell dynamic load test cycles (FC-DLC) of the vehicle-mounted fuel cell stack under dynamic conditions, establishes a physical aging model in the time domain, and then uses the improved Grey Wolf optimization algorithm to perform parameter identification to obtain the dynamic DI. Finally, the CPO-CNN-BiLSTM neural network is used to perform short-term and long-term aging predictions on DI.

[0046] Specifically, the dynamic durability test data collection includes the following: the entire durability experiment has a total of 3076 sets of fuel cell dynamic load test cycles, about 1008 hours, the sampling frequency of the PEMFCs durability test is 1Hz, each FC-DLC cycle lasts 1180s, contains 35 test steps, and each test cycle consists of 152 FC-DLC cycles (about 50 hours), followed by a characterization test. After completing each 304 FC-DLC cycles, there is a 12-hour rest period to simulate the shutdown behavior of fuel cell vehicles.

[0047] When analyzing the data, since each FC-DLC cycle includes 35 test steps, the current and voltage will change with the different test steps. Therefore, the full-time voltage cannot be directly used as dynamic DI. The optimal value of the hydrogen inlet and outlet pressure is 110Kpa. Figure 1As shown in the hydrogen inlet and outlet pressure diagram, the hydrogen inlet pressure oscillates violently around 150 hours, causing a shutdown voltage drop at this time. After 650 hours, the entire PEMFCs system experiences abnormal gas pressure / flow operation, which inevitably changes the operating state and increases the difficulty of prediction.

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

[0049]

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

[0051] In the step S3), when solving complex optimization problems, the Gray Wolf Algorithm has problems such as insufficient global search capability and easy to fall into local optimal values. Therefore, Levy flight and dynamic weight strategy are introduced to improve the Gray Wolf optimization algorithm for parameter identification, including the following steps:

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

[0053] S302) using chaotic mapping to initialize population positions and randomly generate various parameter values;

[0054] S303) Substituting the parameter value into the physical aging model to calculate the theoretical voltage;

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

[0056] S305) updating the position and speed of the population;

[0057] S306) Repeat step S303) to step S305) until the stop condition is met, and output the global optimal solution, that is, 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 As time goes by, there is an obvious trend feature until 640 hours ago. Figure 1 It can be seen that abnormal gas pressure / flow operation occurred after 640 hours. Figure 2 As shown, this leads to α and R eq After 640 hours, a phenomenon that was obviously opposite to the overall trend appeared. Therefore, the parameters α and R eq As a dynamic DI to represent the overall aging, where R eq is called DI1 and α is called DI2.

[0059] In step S3), when dynamic DI preprocessing is performed, the original parameter data volume is huge and contains spikes and a large amount of noise, which will lead to large errors in the calculation results and long calculation time. In order to improve the prediction accuracy and reduce the calculation burden, Gaussian filtering is used to process the original data, where the size of the Gaussian kernel is 20. This method has a strong suppression effect on noise and can effectively retain 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 meta-heuristic algorithm that simulates four different protection mechanisms of the crested porcupine (CP). From the least aggressive to the most aggressive, the crested porcupine uses four distinct protection mechanisms: vision, sound, smell, and physical attack. The first two protection mechanisms (vision and sound) reflect the exploratory behavior of the CPO, while the second two protection mechanisms (smell and physical attack) reflect the exploitative behavior of the CPO. The specific steps of the CPO algorithm are as follows:

[0062] In practice, not all CPs will activate the defense mechanism. Only the threatened CPs will activate the defense mechanism. The cyclic population reduction technology can simulate the above situation very well. It can not only speed up the convergence speed, but also maintain the diversity of the population. Its mathematical model is as follows:

[0063]

[0064] Where T is the variable that determines the number of cycles, t is the current function evaluation, and Tmax is the maximum value of the function evaluation, % is the remainder or modulus operator, N min is the minimum number of individuals in the newly generated population. The following are four specific defense strategies:

[0065] 1) First defense strategy

[0066] When the CP is aware of the predator, it will spread out its sharp spines to give the predator a visual shock. When the predator chooses to approach the CP, the distance between the predator and the CP decreases, encouraging exploration of the area between the predator and the CP to speed up convergence. When the predator chooses to leave, the distance between the predator and the CP is maximized, encouraging exploration of farther areas to identify unvisited areas. It can be expressed by the following mathematical formula:

[0067]

[0068] in is the optimal solution for the function to estimate t, is a vector generated between the current CP and a randomly selected CP from the population, representing the position of the predator at iteration t, τ1 is a random number based on a normal distribution, and τ2 is a random value in the interval [0,1], generating The mathematical formula is as follows:

[0069]

[0070] Where r is a random number between [1, N];

[0071] 2) Second defense strategy

[0072] CP uses sound to make noise and threaten predators, and the sound becomes louder as the predator approaches it. In order to mathematically model this behavior, the following formula was proposed:

[0073]

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

[0075] 3) The third defense strategy

[0076] In this strategy, CP secretes a foul smell that spreads around it to prevent predators from approaching it. The mathematical model is as follows:

[0077]

[0078] Where r3 is a random number between [1,N], τ3 is a random value in the interval [0,1], is the position of the ith individual at iteration t, is a parameter used to control the search direction, γ t As a defense factor, is the odor diffusion factor, and the three expressions are as follows:

[0079]

[0080]

[0081] in represents the objective function value of the ith individual at iteration t, ε is a small value that avoids division by zero, is a vector containing randomly generated values ​​between 0 and 1, rand is a variable containing randomly generated numbers between 0 and 1, N is the population size, t is the number of current iterations, and t max is the maximum number of iterations;

[0082] 4) The fourth defense strategy

[0083] When the predator is very close, the CP will use its own spines to physically attack the predator, which is manifested as a one-dimensional inelastic collision. The physical attack behavior can be expressed by the following mathematical formula:

[0084]

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

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

[0087] The aging prediction of proton exchange membrane fuel cells is mainly divided into short-term prediction and long-term prediction. Short-term prediction can capture local changes with high accuracy, and the general prediction time is several hours or days. Long-term prediction can capture the aging trend, and the prediction range is usually hundreds of hours or months. Due to the dynamic load conditions, long-term prediction faces greater challenges. The evaluation indicators of performance prediction usually use the root mean square error (RMSE), mean absolute error (MAE) and determination coefficient (R 2 ), the calculation formulas of the three evaluation indicators are as follows:

[0088]

[0089] The method of the present invention is compared with CNN-BiLSTM and BiLSTM without parameter optimization when the training set accounts for 40%, 50%, 60% and 70% respectively, to verify the superiority of the method proposed in this paper in predicting DI.

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

[0091] Since the amount of data in the dynamic data set is huge, few people directly use the original voltage data for short-term prediction. Therefore, the present invention converts the parameters into voltage through the aging model, where DI1 and DI2 use the predicted values ​​and other parameters take the average value of the training set. Figure 4 The short-term voltage prediction results are converted to different training set ratios. The black box in the range of about 750h to 1000h has a large error. This is because the long-term operation of the battery stack and the drastic change in the inlet pressure cause the aging model to have a poor fitting effect under full load. Therefore, as the training set ratio increases, the prediction error will also increase due to the poor fitting effect of the aging model under full load. However, since the aging model can still fit well under other load conditions, the overall prediction error remains 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] For the long-term aging prediction results, when evaluating the performance of the proposed fusion method in long-term prediction, the degradation trend of DI also changes accordingly due to the change in operating conditions after 640 h. Therefore, the termination time of the 640 h test, that is, 650 h, is selected as the end of life time, and the data after 650 h are set as failure values. Figure 5 The long-term prediction results of CPO-CNN-BiLSTM, CNN-BiLSTM and BiLSTM networks for the next 200 hours of data when the first 450 hours are used as training sets, where Figure 5 (a) is the equivalent resistance, Figure 5 (b) is the charge transfer coefficient. The blue data represents available data, and the gray data represents failed data. In the test, each shutdown will cause the internal parameters of the battery stack to change dramatically and then the parameters will recover, which makes long-term prediction more difficult. Figure 5 It can be seen that compared with a single BiLSTM network, the original CNN-BiLSTM and the CPO-optimized CNN-BiLSTM network can roughly predict the future trend of DI. The CPO-optimized CNN-BiLSTM network can more accurately predict the overall trend of DI1 and DI2 than the unoptimized network, and accurately predict the value at the end of life, greatly improving the overall performance of long-term prediction.

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

[0094] Table 1 Long-term prediction error results of three algorithms on 450 hours training set

[0095]

[0096] The fusion prediction method of proton exchange membrane fuel cells under dynamic conditions of the present invention has higher prediction accuracy in short-term prediction and better prediction performance in long-term prediction, and can more accurately show the overall aging trend. In view of the fact that the stack voltage of vehicle-mounted PEMFCs under dynamic conditions cannot be directly used as an aging indicator to represent the aging of the battery, a method for extracting dynamic aging indicators is proposed, that is, a physical aging model is established in the time domain, and the parameters in the aging model are identified using the improved gray wolf optimization algorithm, and the equivalent resistance R of the two aging indicators is successfully extracted.eq and electrode charge transfer coefficient α; the CNN-BiLSTM neural network optimized by the crested porcupine algorithm (CPO) uses 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; in the 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 the aging indicators into voltage is kept at a low level. In the long-term prediction, CPO-CNN-BiLSTM also shows better prediction performance and can more accurately show the overall aging trend.

Claims

1. A method for predicting the fusion of a proton exchange membrane fuel cell under dynamic conditions, characterized in that: Under dynamic working conditions, dynamic load test cycles are carried out on several groups of vehicle fuel cells to obtain voltage data. Based on this, a physical aging model is established in the time domain. Then, the algorithm is used to perform parameter identification to obtain the dynamic DI. 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 the fusion of a proton exchange membrane fuel cell under dynamic conditions as claimed in claim 1, characterized in that: The steps include: 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 the proton exchange membrane fuel cell, and analyze the voltage, current and hydrogen inlet and outlet pressures collected in the dynamic durability test; S2) Establish a physical aging model of PEMFCs for parameter identification: Establish a physical aging model in the time domain to explore the changing trends of various parameters in each dynamic load test cycle; S3) Extraction and preprocessing of dynamic DI: Improve the Gray Wolf optimization algorithm, perform parameter identification on each dynamic load test cycle for the physical aging model proposed in step S2), analyze the trend of each parameter change, select the appropriate dynamic DI, and filter the DI; S4) Establishment of neural network and prediction strategy for aging prediction: Combining the spatial extraction ability of convolutional neural network and the time series prediction ability of BiLSTM, a CNN-BiLSTM model was constructed, and the crown porcupine optimization algorithm was used to optimize the hyperparameters of the CNN-BiLSTM model 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 perform short-term and long-term aging predictions on DI respectively.

3. The method for predicting the fusion of a proton exchange membrane fuel cell under dynamic conditions as claimed in claim 2, characterized in that: In the step S2), the PEMFCs physical aging model is: Where i is the stack current, i loss is the leakage current, i ex is the exchange current, i L is the limiting current, E cell is the single cell voltage, E ocv is the open circuit voltage, T is the operating temperature, R is the molar gas constant, F is the Faraday constant, α is a parameter related to the electrode charge transfer coefficient, and R eq is the equivalent resistance, B c It is a parameter related to the number of electrons transferred in the overall reaction.

4. The method for predicting the fusion of a proton exchange membrane fuel cell under dynamic conditions as claimed in claim 3, characterized in that: In the step S3), the Levy flight and dynamic weight strategy are introduced to improve the Grey Wolf Optimization Algorithm for parameter identification, including the following steps: S301) Initializing parameters, including population size, maximum iteration step length, and convergence factor extreme value; S302) using chaotic mapping to initialize population positions and randomly generate various parameter values; S303) Substituting the parameter value into the physical aging model to calculate the theoretical voltage; S304) Calculate the population fitness value, and update the global optimal solution and the individual optimal solution according to the fitness value; S305) updating the position and speed of the population; S306) Repeat step S303) to step S305) until the stop condition is met, and output the global optimal solution, that is, the optimal parameter value.

5. The method for predicting the fusion of a proton exchange membrane fuel cell under dynamic conditions as claimed in claim 4, characterized in that: 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 eq As a dynamic DI to represent the overall aging.

6. The method for predicting the fusion of a proton exchange membrane fuel cell under dynamic conditions as claimed in claim 5, characterized in that: In the 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.

7. The method for predicting the fusion of a proton exchange membrane fuel cell under dynamic conditions as claimed in claim 6, characterized in that: In the 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.

8. The method for predicting the fusion of a proton exchange membrane fuel cell under dynamic conditions as claimed in claim 7, characterized in that: 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 the input of the neural network, the training set and test set are divided for the training and prediction comparison of the CPO-CNN-BiLSTM neural network respectively. The data first passes through the convolution layer, activation layer and pooling layer in the CNN, and then is input into the forward and reverse LSTM networks in the BiLSTM. Finally, the fully connected layer is used to connect the entire network. The hyperparameters of the CNN-BiLSTM model are optimized using CPO. The optimization ranges for the number of CNN convolution kernels, the number of BiLSTM units, the initial learning rate and the L2 regularization coefficient are [20, 60], [30, 90], [0.001, 0.01] and [0.001, 0.01] respectively. The network parameters are updated, and the network is trained with the new parameters, and then short-term and long-term aging predictions are performed respectively.

Images