A PEMFC Degradation State Prediction Method Based on TiDE Network and Semi-Empirical Model

Abstract

The present invention belongs to the technical field of fuel cell health management, and particularly relates to a method for predicting the degradation state of a PEMFC based on a TiDE network and a semi-empirical model, comprising the following steps: collecting multi-dimensional PEMFC operation data through a combination of physical simulation and experiments; screening key state features based on SHAP, and constructing a CF-TiDE prediction model that integrates a convolutional neural network and time-frequency features; combining the physical mechanism interpretability of the semi-empirical model and the high-precision advantage of the data-driven model to achieve multi-scale prediction of the PEMFC degradation state. This method has strong working condition adaptability and can effectively support the life prediction and maintenance decision-making of PEMFCs.

Description

Technical Field

[0001] The present invention belongs to the technical field of fuel cell health management, and particularly relates to a method for predicting the degradation state of PEMFC based on the TiDE network and semi-empirical model. Background Art

[0002] As a clean energy carrier, hydrogen energy has gradually become an important strategic choice to achieve this goal. Proton exchange membrane fuel cell (PEMFC) is one of the core technologies for hydrogen energy utilization, and the degradation problem of PEMFC has become the key bottleneck restricting its large-scale commercial application. Studying the degradation of PEMFC is not only crucial for promoting the development of the hydrogen energy industry, but also of far-reaching significance for achieving China's "dual carbon" goal.

[0003] With the development of artificial intelligence, data-driven methods have been widely used in the field of predicting battery degradation. These methods mainly rely on artificial intelligence machine learning and data analysis technologies. Accurately predicting the degradation state of PEMFC is the key technology affecting its large-scale commercial application. Model-driven methods can be used to explain the operating mechanism of the battery, but the prediction accuracy of these methods is essentially limited by the quality and coverage of the training data. Data-driven methods can improve the efficiency and accuracy of degradation prediction, but these methods have certain limitations in explaining the physical mechanism of the degradation process. At the same time, the lack of training data will also lead to problems such as poor model generalization ability. To sum up, a single fuel cell degradation prediction method will have problems such as inaccurate prediction or unclear mechanism explanation.

[0004] In order to further improve the interpretability and prediction accuracy of PEMFC prediction, the present application proposes a method for predicting the degradation state of PEMFC based on the TiDE network and semi-empirical model, and realizes high-precision prediction of PEMFC degradation by integrating the advantages of data-driven methods and model-driven methods.

[0005] The information disclosed in this background art section is only intended to increase the understanding of the overall background of the present invention, and should not be regarded as an admission or any form of suggestion that this information constitutes the prior art already known to those of ordinary skill in the art. Summary of the Invention

[0006] The object of the present invention is to provide a method for predicting the degradation state of PEMFC based on the TiDE network and semi-empirical model, and realizes high-precision prediction of PEMFC degradation by integrating the advantages of data-driven methods and model-driven methods.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A method for predicting the degradation state of PEMFC based on the TiDE network and semi-empirical model, comprising the following steps:

[0009] S1. Multi-source data collection and fusion:

[0010] S101. Build a high-precision simulation model of PEMFC based on Simulink to simulate the dynamic operation characteristics under different working conditions;

[0011] S102. Conduct degradation experiments, monitor the state characteristics of PEMFC in real time, and generate actual degradation data;

[0012] S103. Under the same working conditions, collect the operation data of the simulation model, and combine the actual degradation data to construct a cross-domain data set;

[0013] S2. Feature engineering and sample construction:

[0014] S201. Perform periodic time encoding on the cross-domain data set;

[0015] S202. Establish the quantitative relationship between the degradation indication voltage and the SOH of PEMFC;

[0016] S203. Calculate the contribution degree of each state feature of the simulation model to the voltage based on SHAP, and screen the top 5 key state features;

[0017] S204. Integrate the time encoding, key state features and voltage data, eliminate noise interference, and divide the training set, validation set and test set according to a certain proportion;

[0018] S3. CF-TiDE model construction and training:

[0019] S301. Construct an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer as the CNN deep feature extraction layer to extract input features;

[0020] Construct a Fourier transform layer as the frequency domain feature extraction layer to extract frequency domain features;

[0021] S302. Construct a TiDE time series encoding-decoding architecture, and use the residual block ResidualBlock to construct a series-connected encoder Encoder and decoder Decoder;

[0022] S303. Integrate the time encoding in S204 and the features extracted in S301, input them into the TiDE time series encoding-decoding architecture to generate a historical state encoding vector, and output the future voltage prediction sequence of PEMFC through the time decoder;

[0023] S304. Use the optimizer and loss function to iteratively optimize the parameters of the CF-TiDE model;

[0024] S305. Calculate the error of the iterative result. After passing the test, a degraded dynamic prediction model is obtained;

[0025] S4. Online prediction of the degraded state:

[0026] S401. Real-time collect the operation data of the target PEMFC and process the data according to the steps of S2;

[0027] S402. Input the processed data into the trained CF-TiDE model to output the future voltage prediction sequence;

[0028] S403. Generate a degraded state evaluation report through inverse normalization and the SOH mapping function.

[0029] Preferably, in S302, a TiDE time series encoding-decoding architecture is constructed, and a residual block ResidualBlock is used to construct a cascaded encoder Encoder and decoder Decoder, specifically including:

[0030] Construct a multi-layer perceptron layer as the basic residual block ResidualBlock of the CF-TiDE model;

[0031] Based on the residual block, construct a cascaded encoder Encoder (number of layers n e ) and decoder Decoder (number of layers n d );

[0032] Use a residual block to perform dimensionality reduction processing on each time point in the past and future :

[0033] ;

[0034] The obtained after dimensionality reduction and and are connected together and sent to an encoder Encoder composed of multiple residual blocks (number n e ) to be mapped into a coding vector :

[0035] ;

[0036] The vector obtained after encoding is sent to a decoder Decoder composed of multiple residual blocks (number n d ) to be mapped into an intermediate vector and then reshaped to form a decoded vector :

[0037] ;

[0038] Among them, i represents the i-th time series, L represents the length of past time, H represents the length of future time, and the historical data is expressed as ; the dynamic covariates are expressed as ; the static attributes are expressed as .

[0039] Preferably, in S303, the time encoding of S204 and the features extracted in S301 are fused, and the input is fed into the TiDE time series encoding-decoding architecture to generate the historical state encoding vector, and the future voltage prediction sequence of the PEMFC is output through the time decoder, specifically including:

[0040] The CNN deep feature extraction layer extracts the integrated features of the input information ; the Fourier transform layer extracts the integrated features of the frequency domain information ; the time encoding of S204 is extracted as a feature ; the input is fed into the TiDE time series encoding-decoding architecture constructed in S302 and is obtained;

[0041] For each time point's decoding vector and the dynamic covariates of the future time point are stacked and sent to the time decoder of a residual block to obtain the final decoding result;

[0042] Using the mean value of the historical data as the baseline, a global residual connection ResidualConnection is added and superimposed on the decoding result to improve the long-term prediction stability, forming the future prediction :

[0043] .

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] (1) The PEMFC degradation state prediction method based on the TiDE network and the semi-empirical model of the present invention combines the simulation data and experimental data of the semi-empirical physical model to predict the battery degradation, uses the SHAP value to measure the importance of influencing factors, and also combines the convolutional neural network and the Fourier transform to propose a new model CF-TiDE, which improves the interpretability and accuracy of the data-driven model for predicting the battery SOH.

[0046] (2) The PEMFC degradation state prediction method based on the TiDE network and the semi-empirical model of the present invention realizes the comprehensive improvement of the PEMFC degradation prediction method, solves the limitations such as low accuracy, high cost and low efficiency of a single degradation prediction method, and thus meets the development requirements of the PEMFC technology industry. Description of the Drawings

[0047] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0048] Figure 2 It is a schematic diagram of the physical simulation model adopted by the present invention;

[0049] Figure 3 It is a schematic diagram of the network structure of the convolutional depth feature extraction layer of the present invention;

[0050] Figure 4 It is a schematic diagram of the TiDE network structure of the present invention;

[0051] Figure 5 It is a schematic diagram of the online prediction process of the PEMFC degradation state of the present invention;

[0052] Figure 6 It is a prediction effect diagram of the TiDE model provided by the embodiment of the present invention on the experimental data set;

[0053] Figure 7 It is a prediction effect diagram of the TiDE model provided by the embodiment of the present invention on the simulation data set;

[0054] Figure 8 It is a prediction effect diagram of the CF-TiDE model provided by the embodiment of the present invention on the experimental data set;

[0055] Figure 9 It is a prediction effect diagram of the CF-TiDE model provided by the embodiment of the present invention on the simulation data set. Detailed implementation manners

[0056] The technical solutions of the present invention patent are described clearly and completely below. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0057] Referring to the attached Figure 1 , a method for predicting the degradation state of PEMFC based on TiDE network and semi-empirical model, the steps are as follows:

[0058] S1. Multi-source data collection and fusion, specifically including:

[0059] S101. Build a high-precision simulation model of PEMFC including a battery stack, a hydrogen / oxygen supply system, a cooling system and a load based on Simulink to simulate the dynamic operation characteristics under different working conditions;

[0060] S102. Implement an accelerated degradation experiment on the test bench, monitor the state characteristics such as PEMFC voltage, temperature, pressure, humidity, etc. in real time, and generate actual degradation data;

[0061] S103. Under the same working conditions, collect the operation data of the simulation model, and combine it with the actual degradation data to construct a cross-domain dataset;

[0062] S2. Feature engineering and sample construction, specifically including:

[0063] S201. Perform periodic time encoding on the cross-domain dataset, and map the time features (year, month, day, hour, minute) to the interval;

[0064] S202. Select voltage as the degradation indication parameter, and establish the quantitative relationship between the degradation indication voltage and the SOH of PEMFC;

[0065] S203. Based on SHAP, calculate the contribution degree of each state feature of the simulation model to the voltage, and screen the top 5 key state features (such as current density, hydrogen inlet pressure, etc.) to improve the interpretability of degradation prediction;

[0066] S204. Integrate the time encoding, key state features and voltage data, eliminate noise interference through the Savitzky-Golay filter, and divide the training set, validation set and test set proportionally;

[0067] S3. CF-TiDE model construction and training, specifically including:

[0068] S301. Construct a convolutional Fourier double-branch feature extraction layer, refer to Appendix Figure 3 , specifically including:

[0069] Construct an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer as the CNN deep feature extraction layer to extract input features;

[0070] Construct a Fourier transform layer as the frequency domain feature extraction layer to extract frequency domain features;

[0071] S302. Construct a TiDE time series encoding-decoding architecture, and use the residual block ResidualBlock to construct a series-connected encoder Encoder and decoder Decoder to achieve long-term and short-term dependence modeling;

[0072] S303. Fuse the time encoding in S204 and the features extracted in S301, input them into the TiDE time series encoding-decoding architecture to generate a historical state encoding vector, and output the PEMFC future voltage prediction sequence through the time decoder TemporalDecoder;

[0073] S304. Use an optimizer and a loss function to prevent overfitting by the early stopping method, and iteratively optimize the parameters of the CF-TiDE model;

[0074] S305. Calculate the error of the iterative result, and obtain a degradation dynamic prediction model after passing the test;

[0075] S4. Online prediction of the degradation state, specifically including:

[0076] S401. Real-time collect the operation data of the target PEMFC, and process the data according to the steps of S2;

[0077] S402. Input the processed data into the trained CF-TiDE model, and output the future voltage prediction sequence;

[0078] S403. Generate a degradation state evaluation report through denormalization and the SOH mapping function to guide the formulation of the maintenance strategy.

[0079] In this embodiment, the physical simulation model diagram in the process of S101 is shown, referring to the appendix Figure 2 .

[0080] In S302, construct the TiDE time series encoding-decoding architecture, and use the residual block ResidualBlock to construct the cascaded encoder Encoder and decoder Decoder to achieve long-term and short-term dependence modeling, specifically including:

[0081] Construct a multi-layer perceptron layer as the basic residual block ResidualBlock of the CF-TiDE model;

[0082] Based on the residual block, construct the cascaded encoder Encoder (number of layers n e ) and decoder Decoder (number of layers n d ); i represents the i-th time series, L represents the past time length, H represents the future time length, the historical data is expressed as , the dynamic covariate is expressed as , and the static attribute is expressed as ;

[0083] Use a residual block to perform dimensionality reduction processing on at each time point in the past and future:

[0084] ;

[0085] The obtained after dimensionality reduction, and are connected together and sent to an encoder Encoder composed of multiple residual blocks (number n e ) to be mapped into an encoded vector :

[0086] ;

[0087] The encoded vector is sent to a decoder Decoder containing multiple residual blocks (number n d ), and is mapped into an intermediate vector , and after reshaping, a decoded vector is formed:

[0088] .

[0089] Referring to the appendix Figure 4 , in S303, the time encoding of S204 is fused with the features extracted in S301, and is input into the TiDE network to generate a historical state encoding vector, and a future voltage prediction sequence is output through the temporal decoder TemporalDecoder, specifically including:

[0090] The CNN deep feature extraction layer extracts the integrated features of the input information , the Fourier transform layer extracts the integrated features of the frequency domain information , the time encoding of S204 is extracted as a feature , and is input into the TiDE time series encoding-decoding architecture constructed in S302 and obtains ;

[0091] For each decoded vector at each time point in and the dynamic covariates at future time points , after stacking, they are sent to the time decoder of a residual block to obtain the final decoding result;

[0092] Taking the mean of the historical data as the baseline, a global residual connection ResidualConnection is added and stacked on the decoding result to improve the long-term prediction stability, and a future prediction is formed:

[0093] .

[0094] In addition, in S305, the error of the iterative result is calculated, and after passing the test, a degraded dynamic prediction model is obtained, specifically including:

[0095] The relative root mean square error (RRMSE), mean absolute percentage error (MAPE), and mean absolute error (MAE) are used as evaluation indicators;

[0096] If RRMSE, MAPE, and MAE all meet the expected requirements, the test passes, and thus a remaining life dynamic prediction model that passes the test is obtained;

[0097] Otherwise, continue training to reach the target accuracy.

[0098] The calculation formulas of RRMSE, MAPE and MAE are as follows:

[0099]

[0100]

[0101]

[0102] Wherein, is the predicted value, y i is the true value, and n is the number of samples. Among them, mean(y i ) represents the average value of the true values. The smaller the RRMSE value, MAPE value and MAE value, the higher the prediction accuracy.

[0103] In this embodiment, the online prediction flowchart of the PEMFC degradation state for S4 is shown, referring to the appendix Figure 5 .

[0104] This embodiment also gives a specific example of the PEMFC degradation state prediction method based on the TiDE network and the semi-empirical model. Taking the PEMFC produced by Yuchai Xinlan Technology Co., Ltd. as the object, a corresponding simulation model is built according to step S101, 533h of dynamic operation experimental data is collected, and the simulation data of a certain driving condition is introduced on this basis, so as to obtain the corresponding experimental and simulation data sample sets for training and testing the remaining life dynamic prediction model. Table 1 below shows the test error evaluation results of the TiDE and CF-TiDE models.

[0105] Table 1 Test error evaluation results of TiDE and CF-TiDE models

[0106]

[0107] In order to verify the effectiveness of the improved model in this embodiment, after the model training is completed, the test sample set is respectively brought into different models for testing to conduct a comparative experiment. Referring to the appendix Figures 6 - 9 , the comparison between the test values of the TiDE and CF-TiDE models on the sample set and the true values of the test sample set is shown respectively. Combining the data analysis in Table 1, it can be seen that the test results of the CF-TiDE model are closest to the true values.

[0108] Those skilled in the art can recognize that the technical solutions described in the above embodiments can be driven by a computer program to execute all or part of the steps on corresponding hardware devices, and the relevant programs can be stored in a non-transitory computer-readable storage medium, including but not limited to carriers such as magnetic disks, optical storage media, read-only memories or random access memories.

[0109] The foregoing description of specific exemplary embodiments of the present invention has been presented for purposes of illustration and example. These descriptions are not intended to limit the invention to the precise forms disclosed, and obviously, many modifications and variations are possible in light of the above teaching. The purpose of selecting and describing the exemplary embodiments is to explain the specific principles of the invention and its practical applications, so that those skilled in the art can implement and utilize the various different exemplary embodiments of the invention, as well as various different selections and modifications. The scope of the invention is intended to be defined by the claims and their equivalents.

Claims

1. A method for predicting the degradation state of PEMFC based on the TiDE network and semi-empirical model, characterized in that, It includes the following steps: S1. Multi-source data collection and fusion: S101. Build a high-precision PEMFC simulation model based on Simulink; S102. Conduct degradation experiments, monitor the state characteristics of PEMFC in real time, and generate actual degradation data; S103. Under the same working conditions, collect the operation data of the simulation model, and combine the actual degradation data to construct a cross-domain dataset; S2. Feature engineering and sample construction: S201. Perform periodic time encoding on the cross-domain dataset; S202. Establish the quantitative relationship between the degradation indication voltage and the SOH of PEMFC; S203. Calculate the contribution degree of each state characteristic of the simulation model to the voltage based on SHAP, and select the top 5 key state characteristics; S204. Integrate the time encoding, key state characteristics and voltage data, eliminate noise interference, and divide them into training set, validation set and test set; S3. CF-TiDE model construction and training: S301. Construct an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer as the CNN deep feature extraction layer to extract input features; Construct a Fourier transform layer as the frequency domain feature extraction layer to extract frequency domain features; S302. Construct a TiDE time series encoding-decoding architecture, and use basic residual blocks to construct a cascaded encoder and decoder; S303. Integrate the time encoding in S204 and the features extracted in S301, input them into the TiDE time series encoding-decoding architecture to generate a historical state encoding vector, and output the future voltage prediction sequence of PEMFC through the time decoder; S304. Use an optimizer and a loss function to iteratively optimize the parameters of the CF-TiDE model; S305. Calculate the error of the iterative result, and obtain the degradation dynamic prediction model after passing the test; S4. Online prediction of degradation state: S401. Collect the operation data of the target PEMFC in real time, and process the data according to the steps in S2; S402. Input the processed data into the trained CF-TiDE model to output the future voltage prediction sequence; S403. Generate a degradation state evaluation report through inverse normalization and SOH mapping function.

2. The method for predicting the degradation state of PEMFC based on the TiDE network and semi-empirical model according to claim 1, wherein In S302, the TiDE time series encoding-decoding architecture is constructed, and basic residual blocks are used to construct a cascaded encoder and decoder, which specifically includes: Construct a multi-layer perceptron layer as the basic residual block of the CF-TiDE model; Construct a cascaded encoder and decoder based on the residual block; Use a residual block to perform dimensionality reduction on each time point in the past and future as follows: ； Obtained by dimensionality reduction is connected with and and then sent to an encoder composed of multiple residual blocks to be mapped into a coded vector : ； The encoded vector is sent to a decoder containing multiple residual blocks and mapped into an intermediate vector After that, it is reshaped to form a decoded vector : ； where \(i\) represents the \(i\)-th time series, \(L\) represents the length of past time, \(H\) represents the length of future time, historical data is expressed as , dynamic covariates are expressed as , and static attributes are expressed as .

3. The method for predicting the degradation state of PEMFC based on the TiDE network and semi-empirical model according to claim 1, characterized in that, In S303, the time encoding in S204 and the features extracted in S301 are integrated, input into the TiDE time series encoding-decoding architecture to generate a historical state encoding vector, and output the future voltage prediction sequence of PEMFC through the time decoder, which specifically includes: The CNN deep feature extraction layer extracts the integrated features of the input information , and the Fourier transform layer extracts the integrated features of the frequency domain information , extracts the time encoding of S204 as features , fuses the above features, inputs the TiDE time series encoding-decoding architecture constructed by S302 and obtains ; For the decoding vectors at each time point in and the dynamic covariates at future time points After stacking, they are sent to a time decoder of a residual block to obtain the final decoding result; Using the mean of historical data as the baseline, adding a global residual connection ResidualConnection and superimposing it on the decoding result to improve the long-term prediction stability, thus forming a future prediction : 。

Images