Evaluation method for mutual influence between chips in attenuation process of fuel cell stack

Abstract

The invention relates to a method for evaluating mutual influence between chips in the attenuation process of a fuel cell stack. The method comprises the following steps: S1, acquiring actual voltage data and actual parameter data; s2, inputting the voltage data into a first space-time diagram neural network model to obtain a first voltage predicted value; s3, inputting the parameter data into a second space-time diagram neural network model to obtain a second voltage predicted value; and S4, adopting an SHAP analysis method, obtaining local Shapley values of a plurality of samples based on the first voltage prediction value and the second voltage prediction value, calculating a global Shapley value based on the local Shapley values of the plurality of samples, and obtaining time-space correlation between the voltage of the single battery and the output voltage of each single battery and between each parameter and the output voltage of each single battery. Compared with the prior art, the method has the advantages of comprehensively analyzing the mutual influence between the fuel cell stack attenuation process pieces, reducing the damage to the cell and the like.

Description

Technical Field

[0001] The present invention relates to the technical features of a fuel cell stack, and in particular to a method for evaluating the mutual influence between chips in a fuel cell stack attenuation process. Background Art

[0002] Among hydrogen energy conversion devices, proton exchange membrane fuel cells (PEMFCs) are widely considered to be a promising power technology for mobile, fixed and portable power sources due to their high energy efficiency, low pollution and low noise. Although fuel cell technology has basically met the needs of commercial applications, the limitation of its service life remains one of the main challenges.

[0003] In practical applications, commercial fuel cell stacks are usually composed of multiple single cells connected in series to increase the total output power. However, the interaction between single cells may lead to uneven output voltage, affecting the performance and life of the entire stack. At present, the study of the impact between fuel cell stacks is mainly carried out by artificially applying faults to single cells of models or actual stacks and observing the changes in the output and internal parameters of adjacent cells. These studies provide insights into understanding the interaction between cells, but the analysis method fails to fully capture the impact of internal parameters between single cells and is prone to damage the battery. Summary of the invention

[0004] The purpose of the present invention is to provide a method for evaluating the mutual influence between chips in the attenuation process of a fuel cell stack in order to comprehensively analyze the mutual influence between chips in the attenuation process of a fuel cell stack while reducing damage to the battery.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A method for evaluating the mutual influence between chips in a fuel cell stack attenuation process, the method comprising the following steps:

[0007] S1. Acquire actual voltage data and actual parameter data, wherein the voltage data is the cell voltages of multiple cells at multiple times, and the parameter data is the values ​​of each parameter at the same time;

[0008] S2. Input the voltage data into a first spatiotemporal graph neural network model to obtain a first voltage prediction value;

[0009] S3, inputting the parameter data into a second spatiotemporal graph neural network model to obtain a second voltage prediction value;

[0010] S4. Using the SHAP analysis method, based on the first voltage prediction value and the second voltage prediction value, obtain the local Shapley values ​​of multiple samples corresponding to the two voltage prediction values ​​respectively, and calculate the global Shapley values ​​corresponding to the two voltage prediction values ​​respectively based on the local Shapley values ​​of the multiple samples, and obtain the temporal and spatial correlation between the single battery voltage and various parameters and the output voltage of each single battery.

[0011] Furthermore, the first spatiotemporal graph neural network model and the second spatiotemporal graph neural network model are T-GCN networks.

[0012] Furthermore, the T-GCN network consists of two parts: a graph convolutional network and a gated recurrent unit.

[0013] Furthermore, the training process of the first spatiotemporal graph neural network model is:

[0014] Obtain cell voltages and corresponding graphs of multiple cells at multiple moments for training, input the cell voltages and corresponding graphs of multiple cells at multiple moments into a first space-time graph neural network, the first space-time graph neural network outputs predicted cell voltages, calculates the loss function to train the first space-time graph neural network, and obtains a first space-time graph neural network model.

[0015] Furthermore, the training process of the second spatiotemporal graph neural network model is:

[0016] Obtain each parameter value and corresponding graph at the same time for training, input each parameter value and corresponding graph at the same time into the second space-time graph neural network, the second space-time graph neural network outputs the predicted single battery voltage, calculates the loss function to train the second space-time graph neural network, and obtains the second space-time graph neural network model.

[0017] Furthermore, the corresponding graph is a graph in which each single battery is a node, the connection relationship between the single batteries is composed of edges, and the nodes are connected end to end in sequence.

[0018] Furthermore, the local Shapley value of the sample is:

[0019]

[0020] Where f represents the spatiotemporal graph neural network model, F is the set of all input features, |F| is the size of F; {x ′ 1,…,x ′ n} is the variable to be explained; S refers to the input of the spatiotemporal graph neural network model, which means that F does not contain the subset of feature i; f(S) refers to the voltage prediction value of the model when S is input;

[0021] Among them, when S is the voltage data of a single cell removed, the local Shapley value corresponding to the voltage data is obtained; at this time f(SU{x′ i}) is the first voltage prediction value;

[0022] When S is the parameter data with one parameter removed, the local Shapley value corresponding to the parameter data is obtained; at this time, f(S∪{x′ i}) is the second voltage prediction value.

[0023] Furthermore, the global Shapley value is:

[0024]

[0025] in, is the local Shapley value of feature i in sample j, and N is the number of all samples.

[0026] Furthermore, the actual parameter data is obtained by performing parameter estimation based on actual polarization curve measurement results of each single cell during the durability test and a single cell voltage model.

[0027] Furthermore, the parameter data are current density, open circuit voltage, ohmic resistance, constant terms related to electron transfer coefficient, exchange current density and limiting current density.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] The present invention establishes the relationship between the inter-chip output performance and the inter-chip internal parameters and output performance during the attenuation process through a spatiotemporal graph neural network, and combines explainable artificial intelligence to interpret the relationship established by machine learning, comprehensively analyzes the inter-chip influence during the attenuation process of the fuel cell stack, and then can find the single cell and internal parameters that have a greater impact on the single-chip output performance, which helps to improve the structural design and control strategy of the fuel cell so that it can better operate in a suitable working range, thereby improving the performance and durability of the fuel cell. The parameters used in the present invention are the attenuation data set, which does not affect the normal operation of the fuel cell stack. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is a flow chart of the present invention;

[0031] Figure 2 It is a schematic diagram of the variable load working condition of the present invention;

[0032] Figure 3 Output the affected dynamic SHAP graph for the 8th battery, where Figure 3 (a) is the impact of the historical voltage of battery No. 8 on the current; Figure 3 (b) is the impact of the historical voltage of batteries 6 and 7 on the current; Figure 3 (c) shows the impact of the historical voltage of batteries No. 9 and No. 10 on the current. Figure 3 (d) is an image of the voltage of battery No. 8 changing with time. DETAILED DESCRIPTION

[0033] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0034] The present invention, based on the premise of not damaging the normal operation of the fuel cell stack, combines the space-time graph neural network (ST-GNN) and explainable artificial intelligence (XAI) technology to propose an evaluation method for the mutual influence between chips during the attenuation process of the fuel cell stack, which is used to capture the situation where the output performance of a single cell is affected by the output of adjacent units and their internal parameters. By analyzing the durability experimental data of the fuel cell stack, the main factors affecting the output performance of a single cell and the corresponding internal parameters can be determined. This provides a reliable basis for optimizing the control strategy of the fuel cell stack, improving the structural design and selecting the optimal working conditions. The present invention establishes the relationship between the output performance between chips and the internal parameters between chips and the output performance during the attenuation process through the space-time graph neural network, and combines explainable artificial intelligence to interpret the relationship established by machine learning to find the single cell and internal parameters that have a greater impact on the output performance of the single chip. The flow chart of the method proposed by the present invention is as follows. Figure 1 The method of the present invention comprises the following steps:

[0035] S1. Acquire actual voltage data and actual parameter data, wherein the voltage data is the cell voltages of multiple cells at multiple times, and the parameter data is the values ​​of each parameter at the same time;

[0036] S2. Input the voltage data into a first spatiotemporal graph neural network model to obtain a first voltage prediction value;

[0037] S3, inputting the parameter data into a second spatiotemporal graph neural network model to obtain a second voltage prediction value;

[0038] S4. Using the SHAP analysis method, based on the first voltage prediction value and the second voltage prediction value, obtain the local Shapley values ​​of multiple samples corresponding to the two voltage prediction values ​​respectively, and calculate the global Shapley values ​​corresponding to the two voltage prediction values ​​respectively based on the local Shapley values ​​of the multiple samples, and obtain the temporal and spatial correlation between the single battery voltage and various parameters and the output voltage of each single battery.

[0039] The parameter data of the present invention are current density, open circuit voltage, ohmic resistance, constant term related to electron transfer coefficient, exchange current density and limiting current density.

[0040] The actual parameter data is obtained by parameter estimation based on the actual polarization curve measurement results of each single cell during the durability test and the single cell voltage model. The specific steps of parameter estimation are:

[0041] The model of the change of single cell output voltage with current density i is obtained through theoretical deduction:

[0042]

[0043] Among them, V cell is the cell voltage (V); i is the current density (A / cm2); E OCV is the open circuit voltage (V); R0 is the ohmic resistance (Ω); a, b are constant terms related to the electron transfer coefficient; T is the temperature (K); i0 is the exchange current density (A / cm2); i L is the limiting current density (A / cm2).

[0044] According to the measurement results of polarization curves of each single cell in the durability test of the fuel cell stack, the time-varying parameters in the model are identified based on the single cell voltage model, and the internal parameter estimates of each single cell that change with the attenuation of the battery stack are obtained.

[0045] In S2 and S3 of the present invention, the spatiotemporal graph neural network T-GCN is used to obtain and retain information in the time dimension by using the structure of the gated recurrent unit (GRU) on the graph convolutional network. It can make full use of the historical information and spatiotemporal dependencies of the input variables. Each T-GCN unit consists of a graph convolutional network and a gated recurrent unit. First, data with spatial associations at multiple moments in the past are used as input, and the graph convolutional network is used to extract the association between the features of the nodes in space at each moment. Secondly, the obtained time series with spatial features is input into the gated recurrent unit, and the dynamic changes of the time series are obtained through information transfer between units to capture the time features.

[0046] In the present invention, the training process of the first spatiotemporal graph neural network model is:

[0047] Obtain cell voltages and corresponding graphs of multiple cells at multiple moments for training, input the cell voltages and corresponding graphs of multiple cells at multiple moments into a first space-time graph neural network, the first space-time graph neural network outputs predicted cell voltages, calculates the loss function to train the first space-time graph neural network, and obtains a first space-time graph neural network model.

[0048] The training process of the second spatiotemporal graph neural network model is:

[0049] Obtain each parameter value and corresponding graph at the same time for training, input each parameter value and corresponding graph at the same time into the second space-time graph neural network, the second space-time graph neural network outputs the predicted single battery voltage, calculates the loss function to train the second space-time graph neural network, and obtains the second space-time graph neural network model.

[0050] According to the experimental results of the fuel cell stack under the durability test, the current density of the battery stack, the output voltage of each single cell and the internal parameter estimation of each single cell are obtained. Based on the spatiotemporal graph neural network, the spatiotemporal correlation between the output voltage of each single cell and the current density, the internal parameters of each single cell and the output voltage of each single cell is established respectively.

[0051] The present invention uses the explainable artificial intelligence method SHAP analysis to measure the marginal contribution of each feature in different combinations by calculating the average contribution of each feature to the prediction (i.e., Shapley value). For a model f and a set of features S, the local Shapley value of feature i It focuses on the impact and importance of variables on individual observation samples. The specific calculation formula is:

[0052]

[0053] Where F is the set of all input features; |F| is the size of F; {x ′ 1,…,x ′ n} is the variable that needs to be explained, that is, the input feature; S refers to the subset of F that does not contain feature i; f(S) refers to the predicted value of the model when S is input.

[0054] The local Shapley values ​​calculated above are local Shapley values ​​calculated based on the first voltage prediction value and the second voltage prediction value. i}), the model input S is the voltage data of a cell removed, the feature i is the voltage data of the removed cell, and F is the voltage data of all cells. The local Shapley value calculated at this time is the impact of the removed cell.

[0055] For the second voltage prediction value (f(S∪{x′ i})), the input S of the model is the parameter data after removing one parameter, that is, the parameter data after removing any parameter of current density, open circuit voltage, ohmic resistance, constant term related to electron transfer coefficient, exchange current density and limiting current density. Among the parameter data, current density does not rely on parameter estimation, but is the data obtained by direct measurement. At this time, the feature i is the parameter data of the parameter removed, and F is all the parameter data at the same time. The local Shapley value calculated at this time is the influence of the removed parameter.

[0056] The global Shapley value focuses on using the entire data set to reveal the relationship between input features and output. At this time, it is necessary to consider the input and output of the model in each observation sample:

[0057]

[0058] in, is the Shapley value of feature i in sample j; N is the number of all samples. Samples represent data at different times (periods). For the cell voltages of multiple cells at multiple times (t1 to tn), the time period from t1 to tn is one sample, and t2 to tn+1 is another sample. For each parameter value at the same time, each time corresponds to a sample.

[0059] The spatiotemporal correlation between the output voltage and current density of each single cell, the internal parameters of each single cell and the output voltage of each single cell is established based on the spatiotemporal graph neural network. Based on explainable artificial intelligence, the global contribution of the entire life test process and the dynamic local contribution that changes over time are obtained. Then, the inter-chip influence is analyzed through the Shapley contribution value, taking the influencing parameters and the affected output as the objects.

[0060] At present, most studies need to introduce artificial faults on a normally operating fuel cell stack, and use a direct comparison method to observe the affected parameters in order to analyze the impact between each piece. The present invention uses a spatiotemporal graph neural network and an explainable artificial intelligence method to analyze the situation in which the output performance of a single cell is affected by the output and internal parameters of adjacent cells during the attenuation of the fuel cell stack without affecting the normal operation of the fuel cell stack. By analyzing the durability test data of the fuel cell stack, we can determine the cells that have a greater impact on the output performance of a single cell and the corresponding internal parameters. This method helps to improve the structural design and control strategy of the fuel cell, so that it can better operate in a suitable working range, thereby improving the performance and durability of the fuel cell. Therefore, the present invention is of great significance to the development of fuel cell technology.

[0061] The following is a detailed description of the inter-chip influence analysis method for the attenuation process of a fuel cell stack proposed by the present invention through a specific embodiment. The mutual influence of inter-chip output during the attenuation process of a fuel cell stack under variable load conditions and the influence of the internal parameters of adjacent cells on the output of a single cell are evaluated.

[0062] First, a durability test is performed to obtain the original attenuation data set. This data set should at least include the voltage of each battery in the stack and the steady-state polarization curve data. The variable load conditions of the durability test are as follows: Figure 2As shown, the single cycle time is about 90 minutes. Subsequently, the data is resampled, smoothed, and normalized before model training. Furthermore, according to different evaluation requirements, the battery voltage, current density, and internal parameter estimates are used to set node features for the spatiotemporal graph model. In addition, the static graph structure should be set according to the actual connection status of each single battery in the battery stack.

[0063] After setting the number of hidden layer neurons, loss function, and optimizer, the model's hyperparameters are adjusted using the training set and test set to ensure that the model can accurately reflect the relationship between input and output.

[0064] After the model training is completed, the SHAP method is further used to explain the contribution of the input features to the relationship established by the model. Relying on the entire life test data, the calculated SHAP contribution value is used to evaluate the impact of each input feature on the output of a single battery from both a global and dynamic perspective. Figure 3 As shown in the figure, the voltage change at around 0.75V will strongly change the impact of adjacent batteries on the output of the observed battery. The SHAP value contribution of the battery with the first sequence number increases first and then decreases, while the opposite is true for the battery with the last sequence number. By adding up the SHAP values ​​that change over time, the global contribution evaluation of the entire durability test can be obtained.

[0065] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for evaluating the mutual influence between chips in a fuel cell stack attenuation process, characterized in that: The method comprises the following steps: S1. Acquire actual voltage data and actual parameter data, wherein the voltage data is the cell voltages of multiple cells at multiple times, and the parameter data is the values ​​of each parameter at the same time; S2. Input the voltage data into a first spatiotemporal graph neural network model to obtain a first voltage prediction value; S3, inputting the parameter data into a second spatiotemporal graph neural network model to obtain a second voltage prediction value; S4. Using the SHAP analysis method, based on the first voltage prediction value and the second voltage prediction value, obtain the local Shapley values ​​of multiple samples corresponding to the two voltage prediction values ​​respectively, and calculate the global Shapley values ​​corresponding to the two voltage prediction values ​​respectively based on the local Shapley values ​​of the multiple samples, and obtain the temporal and spatial correlation between the single battery voltage and various parameters and the output voltage of each single battery.

2. A method for evaluating the mutual influence between chips in a fuel cell stack attenuation process according to claim 1, characterized in that: The first spatiotemporal graph neural network model and the second spatiotemporal graph neural network model are T-GCN networks.

3. The method for evaluating the mutual influence between chips in the attenuation process of a fuel cell stack according to claim 2, characterized in that: The T-GCN network consists of two parts: a graph convolutional network and a gated recurrent unit.

4. A method for evaluating the mutual influence between chips in a fuel cell stack attenuation process according to claim 3, characterized in that: The training process of the first spatiotemporal graph neural network model is: Obtain cell voltages and corresponding graphs of multiple cells at multiple moments for training, input the cell voltages and corresponding graphs of multiple cells at multiple moments into a first space-time graph neural network, the first space-time graph neural network outputs predicted cell voltages, calculates the loss function to train the first space-time graph neural network, and obtains a first space-time graph neural network model.

5. The method for evaluating the mutual influence between chips in the attenuation process of a fuel cell stack according to claim 3, characterized in that: The training process of the second spatiotemporal graph neural network model is: Obtain each parameter value and corresponding graph at the same time for training, input each parameter value and corresponding graph at the same time into the second space-time graph neural network, the second space-time graph neural network outputs the predicted single battery voltage, calculates the loss function to train the second space-time graph neural network, and obtains the second space-time graph neural network model.

6. A method for evaluating the mutual influence between chips during fuel cell stack attenuation process according to claim 4 or 5, characterized in that: The corresponding graph is a graph in which each single battery is a node, the connection relationship between the single batteries is a graph composed of edges, and the nodes are connected end to end in sequence.

7. The method for evaluating the mutual influence between chips in the attenuation process of a fuel cell stack according to claim 1, characterized in that: The local Shapley value of a sample is: Where f represents the spatiotemporal graph neural network model, F is the set of all input features, |F| is the size of F; {x′1,...,x′ n } is the variable to be explained; S refers to the input of the spatiotemporal graph neural network model, which means that F does not contain the subset of feature i; f(S) refers to the voltage prediction value of the model when S is input; Among them, when S is the voltage data of a single cell removed, the local Shapley value corresponding to the voltage data is obtained; at this time f(S∪{x′ i }) is the first voltage prediction value; When S is the parameter data with one parameter removed, the local Shapley value corresponding to the parameter data is obtained; at this time, f(S∪{x′ i }) is the second voltage prediction value.

8. The method for evaluating the mutual influence between chips during the attenuation process of a fuel cell stack according to claim 7, characterized in that: The global Shapley value is: in, is the local Shapley value of feature i in sample j, and N is the number of all samples.

9. The method for evaluating the mutual influence between chips in the attenuation process of a fuel cell stack according to claim 1, characterized in that: The actual parameter data is obtained by performing parameter estimation based on actual polarization curve measurement results of each single cell during the durability test and a single cell voltage model.

10. The method for evaluating the mutual influence between chips in the attenuation process of a fuel cell stack according to claim 9, characterized in that: The parameter data are current density, open circuit voltage, ohmic resistance, constant terms related to electron transfer coefficient, exchange current density and limiting current density.

Images