An online estimation method for characteristic frequency impedance of proton exchange membrane fuel cells

Abstract

The present invention relates to an online estimation method for characteristic frequency impedance of a proton exchange membrane fuel cell, comprising: obtaining an original data set for describing the characteristic frequency impedance of a fuel cell through an experiment; processing the original data set using a sliding window algorithm to produce a new data set; dividing the new data set into a training set and a test set, and performing standardization processing, constructing a convolutional residual neural network, and training the model by setting training parameters and using the training set; inputting the test set into the trained model, estimating the characteristic frequency impedance online, and calculating the corresponding accuracy. Compared with the prior art, the present invention is based on a convolutional neural network, takes the entire time series as the network input, and relies on a neural network algorithm to perform online impedance estimation, which can not only effectively shorten the time for online impedance estimation, but also learn the changes in the internal impedance of the battery under different working conditions, and perform high-precision estimation in real time.

Description

Technical Field

[0001] The invention relates to the technical field of fuel cells, and in particular to an online estimation method for characteristic frequency impedance of a proton exchange membrane fuel cell. Background Art

[0002] Fuel cell vehicles are electric vehicles that use a fuel cell system as a single power source, or a fuel cell system and a rechargeable energy storage system as a hybrid power source. Fuel cells are an important field of the new energy revolution and will become a key pillar for building a green, efficient and safe new energy system in the future. It is also an important direction for the transformation and development of the automotive industry, and its development is limited by the service life of the battery. At present, high-frequency and low-frequency impedances are mainly used for fuel cell diagnosis: high-frequency impedance can characterize the membrane water content, mainly used to reflect membrane dry failure; low-frequency impedance characterizes charge transfer loss, mainly to provide a basis for solving gas shortage and flooding. If online fuel cell internal state estimation and feedback can be achieved during the operation of the fuel cell to reasonably control the battery, it can prevent it from experiencing membrane dryness, flooding, gas shortage and other failures, thereby increasing its service life and promoting the large-scale commercial application of fuel cells.

[0003] In existing research, Chinese patent CN202110789795.9 calculates the index parameters of starvation fault diagnosis by combining the real and imaginary parts corresponding to specific frequencies of the electrochemical impedance spectrum and the internal resistance model of the fuel cell with principal component analysis and K-means clustering methods. However, this method cannot realize online calculation and relies on the comparison of cluster center positions for diagnosis, which is not very accurate.

[0004] Chinese patent CN202210685329.0 uses a fast electrochemical impedance spectroscopy method to obtain the impedance spectrum of a proton exchange membrane fuel cell under different working conditions, and uses a classification model based on linear discriminant analysis to select two frequencies of the impedance spectrum and record them as characteristic frequencies, thereby achieving online state monitoring. However, this solution still uses hardware for EIS testing during online monitoring, and its real-time performance is not high.

[0005] In addition, Chinese patent CN202210932056.5 uses K-means clustering for feature extraction and uses neural networks to diagnose various fuel cell fault conditions. However, fuel cell failure is a cumulative effect, and the diagnosis result is not comprehensive if it is only diagnosed through current sensor information.

[0006] Chinese patent CN202210236433.1 analyzes the internal mechanism of proton exchange membrane fuel cells, selects four variables in the equivalent circuit model as characteristic quantities for flooding fault diagnosis, and uses BP neural network to classify faults. However, the equivalent circuit itself has problems with fitting accuracy, the model input is small, the prediction accuracy is not high, and the time accumulation effect is also not considered.

[0007] In summary, there is an urgent need for a fuel cell characteristic frequency impedance estimation method with better accuracy and higher real-time performance to monitor the internal state of the battery. Summary of the invention

[0008] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a method for online estimation of characteristic frequency impedance of a proton exchange membrane fuel cell, which can efficiently, accurately and real-time online estimate the characteristic frequency impedance of a fuel cell.

[0009] The object of the present invention can be achieved by the following technical solution: A method for online estimation of characteristic frequency impedance of a proton exchange membrane fuel cell comprises the following steps:

[0010] S1. The original data set used to describe the characteristic frequency impedance of the fuel cell is obtained through experiments;

[0011] S2, using the sliding window algorithm to process the original data set to produce a new data set;

[0012] S3, divide the new data set into a training set and a test set, perform standardization, build a convolutional residual neural network, and train the model by setting training parameters and using the training set;

[0013] S4. Input the test set into the trained model, estimate the characteristic frequency impedance online, and calculate the corresponding accuracy.

[0014] Furthermore, the step S1 specifically involves acquiring the original data set through multiple bench experiments.

[0015] Furthermore, the original data set in step S1 includes model input data and model output data, wherein the model input data is specifically the output current, output voltage, battery temperature, anode dew point temperature, cathode dew point temperature, cathode air stoichiometric ratio, anode gas pressure, and cathode gas pressure of the vehicle-mounted fuel cell;

[0016] The model output data is specifically the real part of the high-frequency impedance at 2500 Hz and the real part of the low-frequency impedance at 10 Hz.

[0017] Furthermore, the step S2 specifically utilizes a sliding window algorithm to clean the original data set to remove boundary data that is not conducive to subsequent training.

[0018] Furthermore, the step S2 specifically includes the following steps:

[0019] S21, set the window width;

[0020] S22, taking all signals in the window as a whole as network input, and taking the characteristic frequency impedance at the last moment of the window as output, thereby forming a complete data;

[0021] S23, move the window backward by one unit, and repeat step S22;

[0022] S24. Combine all data into a new data set.

[0023] Furthermore, the specific process of the standardization processing in step S3 is: firstly standardize the training set, and then standardize the test set using the standardization parameters.

[0024] Furthermore, the calculation formula for the normalization process in step S3 is specifically:

[0025]

[0026]

[0027]

[0028] Among them, x i is the original data sequence, i is the i-th sampling point, m is the number of all sampling points, μ B is the mean, σ B is the variance, is the normalized sequence.

[0029] Furthermore, the convolutional residual neural network constructed in step S3 includes a convolutional layer, a pooling layer, an activation layer, a shortcut, a batch normalization layer, a dropout layer and a fully connected layer;

[0030] The set training parameters include but are not limited to learning rate, number of training times, batch size, and gradient threshold.

[0031] Furthermore, step S4 specifically inputs the test set into the trained model, performs online estimation of the characteristic frequency impedance, obtains sequences of high-frequency impedance and low-frequency impedance respectively, and then calculates the accuracy of the online estimation result according to three evaluation indicators: relative error root mean square error, average relative error and relative error.

[0032] Furthermore, the calculation formula of the relative error root mean square error is:

[0033]

[0034] The calculation formula of the average relative error is:

[0035]

[0036] The calculation formula of the relative error is:

[0037]

[0038] Among them, y i is the actual characteristic frequency impedance series, is the estimated characteristic frequency impedance series.

[0039] Compared with the prior art, the present invention proposes a method for estimating characteristic frequency impedance of a proton exchange membrane fuel cell based on a convolutional residual neural network. First, the original data set used to describe the characteristic frequency impedance of the fuel cell is obtained through experiments; secondly, the original data set is processed using a sliding window algorithm to create a new data set; then, the training set and the test set are divided, standardized, a convolutional residual neural network is constructed, training parameters are set, and the model is trained using the training set; finally, the test set is input into the trained model, the characteristic frequency impedance is estimated online, and its accuracy is calculated. The present invention does not require the establishment of a complex or empirical fuel cell impedance mechanism model, nor does it require complex hardware as an aid. It completely relies on a neural network algorithm to achieve online impedance estimation. The online impedance estimation takes a short time, can learn the changes in the internal impedance of the battery under different working conditions, and can perform high-precision online and real-time estimation.

[0040] The present invention takes into account that the change of the internal state is related to the cumulative effect of time, and therefore uses a convolutional neural network to take the entire time series as the network input, thereby ensuring a higher accuracy of the state estimation.

[0041] When acquiring the original data set, the network input selected by the present invention is data that can be collected by common sensors of vehicle-mounted fuel cells at this stage, which has the advantage of low measurement cost and is easy to implement in a vehicle-mounted fuel cell system. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 is a flow chart of the method of the present invention;

[0043] Figure 2a to 2d Examples of two different current modes used in the impedance characteristic experiment in the embodiment;

[0044] Figure 3 This is a flow chart of the sliding window algorithm of the present invention;

[0045] Figure 4 This is a characteristic frequency impedance estimation framework diagram of the present invention;

[0046] Figure 5 This is a structural diagram of the convolutional residual network of the present invention;

[0047] Figure 6 Schematic diagram of the convolution operation of the present invention;

[0048] Figure 7 It is a schematic diagram of the shortcut structure of the present invention;

[0049] Figure 8 Schematic diagram of the fully connected layer of the present invention;

[0050] Figure 9a and 9b This is a diagram of high-frequency impedance estimation results in the embodiment;

[0051] Fig.10a and 10b Graph showing the low-frequency impedance estimation results in the embodiment. DETAILED DESCRIPTION

[0052] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0053] Example

[0054] like Figure 1 As shown, a method for online estimation of characteristic frequency impedance of a proton exchange membrane fuel cell comprises the following steps:

[0055] S1. The original data set used to describe the characteristic frequency impedance of the fuel cell is obtained through experiments;

[0056] S2, using the sliding window algorithm to process the original data set and create a new data set;

[0057] S3, divide the training set and test set, and perform standardization; construct a convolutional residual neural network, set training parameters, and use the training set to train the model;

[0058] S4. Input the test set into the trained model, perform online estimation of the characteristic frequency impedance, and calculate its accuracy.

[0059] This embodiment applies the above technical solution, mainly including:

[0060] 1. Obtain a data set for describing the characteristic frequency impedance of the fuel cell through experiments.

[0061] During this process, a large amount of raw data was obtained through a large number of bench tests, all sensor data were analyzed, and the most commonly used and low-cost measurement signals for vehicle fuel cells were selected as model inputs. At the same time, the characteristic frequencies were determined to be 2500Hz and 10Hz, that is, the real part of the high-frequency impedance of 2500Hz and the real part of the low-frequency impedance of 10Hz were used as the output of the model, that is, as an indicator for internal state identification.

[0062] In this embodiment, the bench test of commercial membrane electrode produced by Yangtze Energy Technologies is taken as an example. Four groups of current inputs are used, two of which are used for high-frequency impedance (2500Hz) estimation and two for low-frequency impedance (10Hz) estimation. Each impedance estimation is divided into two groups of current types: short sequence step-growth current and long sequence current. Figure 2a to 2d As shown. The characteristics of short sequence step growth current are that the amount of data is large and the current changes faster over time; while the long sequence current has only one set of such data, the current changes relatively slowly and the time is longer. In addition, this embodiment also selects 8 signals that are most commonly used and have low measurement costs for on-board fuel cells as the input of the convolutional residual network, namely output current, output voltage, battery temperature, anode dew point temperature, cathode dew point temperature, cathode air stoichiometric ratio, anode gas pressure, and cathode gas pressure, to form the original data set.

[0063] Second, use the sliding window algorithm to process the original data set (remove some boundary data that is not conducive to training) and create a new data set.

[0064] First, the original data is processed by sliding window, that is, the idea of ​​time accumulation effect is added to the network estimation, and all the signals in a period of time before a certain moment to be estimated are taken as a whole, so that the relationship of the original single signal becomes a collection of signals about time series, so that the network can estimate the results based on historical information, making the results more accurate. The specific operation of the algorithm is as follows Figure 3 As shown, the specific steps are as follows:

[0065] 1) Set a window with width w.

[0066] 2) All signals in the window are taken as a whole as the network input, and the characteristic frequency impedance of the last moment of the window is taken as the output, thus forming a complete data.

[0067] 3) Move the window back one unit and repeat the steps in step 2).

[0068] 4) Combine all the data into a brand new dataset.

[0069] 3. Divide the training set and test set, standardize them, build a convolutional residual neural network, set training parameters, and use the training set to train the model. The specific operations are as follows: Figure 4 As shown, the specific steps are as follows:

[0070] 1) The new data set is randomly divided into a training set and a test set. The training set is used to train the network, and the test set is used to test and evaluate the network.

[0071] 2) Standardize the training set to improve the convergence speed and training accuracy of the network. Since the test set is unknown a priori, it also needs to be standardized, so it is standardized using the standardized parameters of the training set. The standardization method includes the following formulas:

[0072]

[0073]

[0074]

[0075] Among them, x i is the original data sequence, i is the i-th sampling point, m is the number of all sampling points, μ B is the mean, σ B is the variance, is the normalized sequence.

[0076] 3) Build a residual neural network, which mainly consists of convolutional layers, pooling layers, activation layers, shortcuts, batch normalization layers, dropout layers and fully connected layers. The structure of the convolutional residual network is as follows: Figure 5 As shown, and set the parameters of each layer.

[0077] 4) Set the training parameters of the network, including learning rate, number of trainings, batch size, gradient threshold, etc. Use mean square error as the loss function, select the Adam algorithm to train the network, continuously optimize and iterate, and save the optimal network.

[0078] At this point, the model training process is completed. The convolutional residual network used is as follows Figure 5 As shown, each layer of the network is introduced below:

[0079] 1) The main operations of the convolution layer are as follows Figure 6 As shown. The convolution layer is the core layer of the convolutional neural network. Its function is to perform feature extraction on the input data containing multiple convolution kernels. For one-dimensional convolution, if the input of the convolution layer can be expressed as:

[0080] x 1:n =[x1,x2,…,x n ]′

[0081] The output of the convolutional layer is:

[0082] c i =wx i:i+h-1 +b

[0083] Among them, w and b are the weight and bias of the convolution respectively.

[0084] 2) The essence of the pooling layer is subsampling. It is sandwiched between consecutive convolutional layers to compress the amount of data and parameters and reduce overfitting.

[0085] 3) The function of the activation layer is to nonlinearly map the output of the convolutional layer to approximate any nonlinear model. The activation function used in convolutional neural networks is usually the rectified linear unit (ReLU) in the following formula, which is characterized by fast convergence and simple gradient calculation.

[0086]

[0087] Here, f(x) is the output and x is the input.

[0088] 4) Shortcut is the core of the residual network. Its structure is as follows Figure 7 Compared with ordinary convolutional neural networks, the residual network adds a shortcut, whose function is to add the previous layer directly to the next layer to avoid the loss of data in the previous layer during transmission and improve the estimation accuracy of the network. The formula is as follows:

[0089] x l+1 =x l +F(x l ,W l )

[0090] 5) The batch normalization layer is used to normalize each batch of data between convolutional layers. It not only speeds up the convergence of the model, but also alleviates the "gradient dispersion" problem in the deep network to a certain extent, making the training of the deep network model easier and more stable. The formula is as follows:

[0091]

[0092]

[0093]

[0094]

[0095] Among them, μ B and are the mean and variance respectively, ε is an infinitesimal value to prevent division by zero, xi and i The input and output, γ and β are the parameters learned by this layer to recover the features learned in some previous layer, and m is the number of samples in the current batch.

[0096] 6) The function of the dropout layer is to randomly discard a part of neurons in each training to improve the generalization ability of the network.

[0097] 7) The fully connected layer is where all neurons between two layers have weights and bias connections, usually located at the end of a convolutional neural network, such as Figure 8 shown.

[0098] Fourth, input the test set into the trained model, perform online estimation of the characteristic frequency impedance, and evaluate its accuracy.

[0099] like Figure 4 As shown in the figure, the test set is input into the network, and the characteristic frequency impedance is estimated respectively, and the sequences of high-frequency impedance and low-frequency impedance are obtained respectively. The accuracy is evaluated, and the main evaluation indicators are relative error root mean square error (RMSE), mean relative error (MAPE) and relative error (error).

[0100]

[0101]

[0102]

[0103] Among them, y i is the actual characteristic frequency impedance series, is the estimated characteristic frequency impedance series.

[0104] To verify the effectiveness of the method of the present invention, this embodiment also uses different algorithms to estimate the high-frequency and low-frequency impedances. The comparison results are as follows: Figure 9a , 9b, 10a, 10b and Table 1, where BP is an artificial neural network, SVR is a support vector machine regression algorithm, 1-CNN is a one-dimensional convolutional neural network, 2-CNN is a two-dimensional convolutional neural network, LSTM represents a long short-term memory network, and ResNet is a convolutional residual network used in the present invention. It can be seen that the estimation accuracy of the convolutional residual network (ResNet) used in the present invention is much better than that of the traditional machine learning methods SVR and BP, and it is also better than 1-CNN and 2-CNN, and is slightly higher than the LSTM network specifically used to solve timing problems in deep learning. Moreover, compared to LSTM, ResNet is much better than LSTM in both offline training time and online estimation time, because the network structure is relatively simple, which shows that the method proposed in the present invention is very suitable for online applications and is suitable for a large number of commercial vehicle data samples in the future.

[0105] Table 1 Comparison of the estimation results of characteristic frequency impedance by different algorithms

[0106]

Claims

1. A method for online estimation of characteristic frequency impedance of a proton exchange membrane fuel cell, characterized in that: The following steps are involved: S1. The original data set used to describe the characteristic frequency impedance of the fuel cell is obtained through experiments; S2, using the sliding window algorithm to process the original data set to produce a new data set; S3, divide the new data set into a training set and a test set, perform standardization, build a convolutional residual neural network, and train the model by setting training parameters and using the training set; S4, input the test set into the trained model, estimate the characteristic frequency impedance online, and calculate the corresponding accuracy; The original data set in step S1 includes model input data and model output data, wherein the model input data specifically includes the output current, output voltage, battery temperature, anode dew point temperature, cathode dew point temperature, cathode air stoichiometric ratio, anode gas pressure, and cathode gas pressure of the vehicle-mounted fuel cell; The model output data is specifically the real part of the high-frequency impedance of 2500 Hz and the real part of the low-frequency impedance of 10 Hz; The step S2 specifically uses a sliding window algorithm to clean the original data set to remove boundary data that is not conducive to subsequent training. The step S2 specifically includes the following steps: S21, set the window width; S22, taking all signals in the window as a whole as network input, and taking the characteristic frequency impedance at the last moment of the window as output, thereby forming a complete data; S23, move the window backward by one unit, and repeat step S22; S24. Combine all data into a new data set.

2. The method for online estimation of characteristic frequency impedance of a proton exchange membrane fuel cell according to claim 1, characterized in that: The step S1 specifically involves obtaining the original data set through multiple bench experiments.

3. The method for online estimation of characteristic frequency impedance of a proton exchange membrane fuel cell according to claim 1, characterized in that: The specific process of the standardization processing in step S3 is: first standardize the training set, and then standardize the test set using the standardization parameters.

4. A method for online estimation of characteristic frequency impedance of a proton exchange membrane fuel cell according to claim 3, characterized in that: The calculation formula for the normalization process in step S3 is specifically: Among them, x i is the original data sequence, i is the i-th sampling point, m is the number of all sampling points, μ B is the mean, σ B is the variance, is the normalized sequence.

5. The method for online estimation of characteristic frequency impedance of a proton exchange membrane fuel cell according to claim 1, characterized in that: The convolutional residual neural network constructed in step S3 includes a convolutional layer, a pooling layer, an activation layer, a shortcut, a batch normalization layer, a dropout layer and a fully connected layer; The training parameters set include learning rate, number of training times, batch size, and gradient threshold.

6. The method for online estimation of characteristic frequency impedance of a proton exchange membrane fuel cell according to claim 1, characterized in that: The step S4 specifically inputs the test set into the trained model, estimates the characteristic frequency impedance online, obtains the sequences of high-frequency impedance and low-frequency impedance respectively, and then calculates the accuracy of the online estimation result according to the three evaluation indicators of relative error root mean square error, average relative error and relative error.

7. A method for online estimation of characteristic frequency impedance of a proton exchange membrane fuel cell according to claim 6, characterized in that: The calculation formula of the relative error root mean square error is: The calculation formula of the average relative error is: The calculation formula of the relative error is: Among them, y i is the actual characteristic frequency impedance series, is the estimated characteristic frequency impedance series.

Images