Proton exchange membrane fuel cell degradation trend prediction method based on frequency enhancement direct prediction

Abstract

The invention discloses a proton exchange membrane fuel cell degradation trend prediction method based on frequency enhancement direct prediction, and belongs to the field of fuel cell life prediction, and the method comprises the steps: obtaining fuel cell voltage time sequence aging data, and carrying out the preprocessing to construct a sample set; a sequence-to-sequence prediction model based on an ITransform framework is constructed; training a model by adopting a frequency enhancement direct prediction training normal form, wherein a loss function fuses time domain and frequency domain loss components; bayesian optimization is combined with cross validation to carry out hyper-parameter optimization so as to determine a final model; and predicting a voltage degradation trend by using the model. According to the method, the multi-step prediction precision is improved through frequency domain loss constraint, the parameter optimization efficiency and stability are improved through an intelligent optimization strategy, and the robustness of the model to actual noise is enhanced through targeted preprocessing, so that the degradation trend of the proton exchange membrane fuel cell is predicted more accurately and efficiently.

Description

Technical Field

[0001] This invention belongs to the field of fuel cell life prediction technology, and particularly relates to a method for predicting the degradation trend of proton exchange membrane fuel cells based on direct prediction of frequency enhancement. Background Technology

[0002] As a key technology in the clean energy field, fuel cells (FCs) play a crucial role in predicting the long-term performance degradation trend for equipment health management, life assessment, and maintenance decisions. Currently, degradation prediction methods for proton exchange membrane fuel cells (PEMFCs) are mainly classified into three categories: model-driven, data-driven, and hybrid-driven. Model-driven methods establish physical models based on electrochemical mechanisms, capable of explaining the intrinsic mechanisms of degradation. Data-driven methods utilize historical operating data and directly learn degradation patterns through machine learning algorithms, exhibiting strong adaptability and flexibility. Hybrid-driven methods combine the above two approaches, aiming to balance mechanistic interpretability and data fitting capabilities. These methods have all been applied in practical engineering and have provided valuable support for PEMFC condition monitoring and life assessment.

[0003] However, existing prediction methods still have several limitations. Model-driven methods rely on accurate mechanistic models, which are difficult to model under complex dynamic conditions and degradation processes coupled with multiple factors, and have high computational complexity. Traditional data-driven methods (such as models based on recurrent neural networks or long short-term memory networks) have limited ability to capture long-range dependencies in long-term series and are easily affected by noise and outliers, affecting the stability and accuracy of predictions. Hybrid-driven methods attempt to combine the advantages of both, but their complex structures and numerous parameters often lead to problems such as high computational resource consumption, low training efficiency, and insufficient generalization performance in practical applications. Therefore, how to construct a method that can effectively capture the long-term nonlinear degradation characteristics of PEMFCs while possessing high prediction efficiency and robustness remains a pressing technical problem to be solved in this field. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a method for predicting the degradation trend of proton exchange membrane fuel cells based on direct frequency enhancement prediction, thereby resolving the issues present in the prior art.

[0005] To achieve the above objectives, the present invention provides a method for predicting the degradation trend of proton exchange membrane fuel cells based on direct prediction of frequency enhancement, comprising: S1. Obtain voltage time-series aging data of proton exchange membrane fuel cells and preprocess it to construct an input sample set; S2. Construct a sequence-to-sequence prediction model based on the ITransformer architecture; S3. The sequence-to-sequence prediction model is trained using the frequency-enhanced direct prediction training paradigm, wherein the total loss function of the frequency-enhanced direct prediction training paradigm is composed of a weighted sum of the time-domain loss component and the frequency-domain loss component. S4. The hyperparameters of the trained sequence-to-sequence prediction model are optimized using the Bayesian optimization method, and the model performance is evaluated using the cross-validation method to determine the final prediction model. S5. The voltage degradation trend of the proton exchange membrane fuel cell is predicted using the final prediction model.

[0006] Preferably, in step S1, the preprocessing includes denoising the voltage time series aging data using variational mode decomposition, and processing the denoised time series using a sliding window method to construct the input sample set, wherein the input time length of each sample is greater than the output prediction time length.

[0007] Preferably, in step S2, the input to the sequence-to-sequence prediction model is a voltage history sequence that has undergone instance normalization.

[0008] Preferably, in step S3, the frequency domain loss component is calculated by transforming the predicted sequence and the real sequence to the frequency domain using Fourier transform, and calculating the L1 norm difference between the two in the frequency domain.

[0009] Preferably, in step S3, the formula for the frequency domain loss component is: ; in, For frequency domain loss components, For the true sequence, This is the predicted sequence.

[0010] Preferably, in step S3, the total loss function is: ; in, For the time-domain loss component, These are the preset weighting coefficients.

[0011] Preferably, in step S4, the Bayesian optimization method is used to optimize at least one hyperparameter among the embedding dimension, number of encoder layers, and number of attention heads of the sequence-to-sequence prediction model.

[0012] Preferably, in step S4, the cross-validation method is K-fold cross-validation, where K is greater than or equal to 5.

[0013] Preferably, in step S4, the average error of multiple verification results obtained by K-fold cross-validation is used as an index to evaluate the performance of the hyperparameter combination.

[0014] Preferably, in step S5, the voltage values ​​of the proton exchange membrane fuel cell at multiple consecutive time points in the future are predicted in multiple steps.

[0015] Compared with the prior art, the present invention has the following advantages and technical effects: This invention significantly improves the accuracy of multi-step prediction and effectively reduces the cumulative error of long-term predictions by employing a "frequency-enhanced direct prediction training paradigm." This paradigm innovatively introduces a frequency-domain loss component based on traditional time-domain loss, mapping the sequence to the frequency domain for difference measurement through Fourier transform. This feature forces the model to focus not only on fitting time-domain points during learning but also on restoring the overall spectral characteristics of the sequence. This overcomes the inherent defect of traditional direct prediction methods that ignore the autocorrelation of the target sequence, making multi-step continuous prediction of the voltage degradation trend of proton exchange membrane fuel cells more accurate and significantly reducing cumulative error.

[0016] This invention improves the efficiency and stability of hyperparameter optimization by employing Bayesian optimization and cross-validation to evaluate model performance. Bayesian optimization, through probabilistic modeling of the hyperparameter space, intelligently guides the search direction, finding better hyperparameter combinations with fewer evaluations, overcoming the high computational cost and low efficiency of traditional grid search or random search. Simultaneously, cross-validation evaluates model performance during the optimization process, eliminating randomness interference from single data partitions and ensuring that the determined optimal hyperparameter combinations have stronger generalization ability and reliability.

[0017] This invention addresses the complex noise and anomalous fluctuations in the voltage signal of proton exchange membrane fuel cells. The preprocessing stage effectively removes high-frequency noise while retaining low-frequency components that reflect the true degradation trend. Simultaneously, the ITransformer architecture is employed to handle sequence-to-sequence prediction tasks, and its multi-head self-attention mechanism better captures long-range dependencies and nonlinear patterns in the voltage degradation process. These features collectively enable the model to more robustly handle disturbance data from real-world operation and more accurately meet the dynamic characteristics prediction requirements of fuel cell voltage degradation. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1This is a schematic diagram comparing 1000 voltage-time series before and after denoising according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the network structure of ITransformer according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the FreDF-ITransformer architecture according to an embodiment of the present invention; Figure 4 This is a schematic diagram showing the comparison results of predicted and experimentally measured voltage degradation curves in an embodiment of the present invention. Detailed Implementation

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0021] Example 1 This embodiment provides a method for predicting the degradation trend of proton exchange membrane fuel cells based on direct prediction of frequency enhancement, including: S1. Obtain voltage time-series aging data of proton exchange membrane fuel cells and preprocess it to construct an input sample set; Further, in step S1, the preprocessing includes using variational mode decomposition to denoise the voltage time series aging data, and using a sliding window method to process the denoised time series to construct the input sample set, wherein the input time length of each sample is greater than the output prediction time length.

[0022] Specifically, step S1 is divided into four parts: (1) The test platform uses Greenlight 20 to test the dynamic durability and polarization characteristics of PEMFC. The FC dynamic load cycle (FC-DLC) is constructed in accordance with the European unified test protocol, and a total of 3076 FC-DLC cycles are completed.

[0023] (2) The test station collects two types of data in real time: voltage / current time series data of dynamic load durability test (1180 seconds per FC-DLC cycle) and performance data of polarization characteristic test. Finally, 1000 hours of voltage-current time series data is selected as the core analysis object.

[0024] (3) Data preprocessing operations employ variational mode decomposition (VMD) for denoising: The filtered time series is input into the variational mode decomposition (VMD) model, which adaptively decomposes the signal into multiple modal components, removes high-frequency noise components, and retains low-frequency components that reflect the true degradation trend of PEMFC. VMD is a non-recursive method that can simultaneously divide the signal into a predefined number of modes, each of which is compactly distributed around a center frequency. The core of VMD lies in constructing and solving the constrained variational problem expressed in equation (1).

[0025] (1) in It is a set of decomposition patterns, output Each represents its corresponding set of center frequencies, while K is the pre-defined number of modes. (Symbol) Represents the convolution operation. This represents the original input signal.

[0026] A time series comparison chart of denoised and undenoised sequences, such as... Figure 1 As shown, the denoised time series contains 1000 operating points, corresponding to 1000 hours of PEMFC aging experiment. The results show that the VMD method exhibits excellent performance, capable of simultaneously eliminating outliers and high-frequency noise.

[0027] (4) The sliding window method is used to generate input-output samples to construct the sample set: the window length is set to 21 hours (input dimension), the output length is set to 5 hours (prediction dimension), and the window step size is set to 1 hour. That is, the window length is 21 hours, the data of the first 21 hours in the window is used as input, and the voltage data of the next 5 consecutive hours is used as output. The window slides over the entire time series with a step size of 1 hour, and a new input-output sample is generated at each step.

[0028] S2. Construct a sequence-to-sequence prediction model based on the ITransformer architecture; Furthermore, in step S2, the input to the sequence-to-sequence prediction model is a voltage history sequence that has undergone instance normalization.

[0029] Specifically, the ITransformer model retains the Transformer model structure while removing irrelevant auxiliary modules such as classification and reconstruction. The encoder block has L layers, the specific number of which is determined through subsequent hyperparameter optimization. Each layer contains a self-attention layer and a feedforward network layer, both employing residual connections and layer normalization structures. A schematic diagram of the ITransformer model is shown below. Figure 2 As shown.

[0030] The process of transforming an input time series into a prediction using this architecture involves the following key steps: (1) For multivariate time series with a backtracking window as input, the input data will be converted into: (2) Where T represents the time series length; N represents the number of variables (directly corresponding to the types of battery parameters that need to be detected); X represents the transposed input matrix (matrix elements are historical time series data of the battery's own parameters; in this proposal, it is the battery voltage-time series for the first 21 hours). Each row of the matrix corresponds to one type of battery parameter, and each column corresponds to the sampled value at one time point. The time series is used as the main axis so that subsequent processing and encoding can be performed according to variables.

[0031] (2) Next is the multilayer perceptron (MLP) embedding. An MLP algorithm is applied to the time series of each variable, mapping it to a D-dimensional feature space, thus giving each variable a corresponding "label" vector, such as: (3) in Let represent the initial embedding matrix, D represent the dimension of the feature space, and MLP represent a multilayer perceptron.

[0032] (3) Stacking of ITransformer Blocks: Variable tokens are processed sequentially through L layers of ITransformer blocks to enhance representation capabilities. Each layer performs two key operations: Self-attention layer: (4) Feedforward network layer: (5) Among them, SelfAttn is a self-attention mechanism, and FeedForward is a feedforward network.

[0033] (4) Prediction mapping: After passing through the L-layer ITransformer block Projecting to predict values ​​at future time steps yields future sequence predictions for each variable, with the following results: (6) The output matrix is ​​a prediction matrix, and the matrix elements are the predicted future values ​​of the battery's own parameters (predicted for the next 5 hours). Each row corresponds to one future time step, and each column corresponds to the predicted value of one type of battery parameter. S is the future prediction duration of the battery parameter.

[0034] (5) The transposed output is: (7) It conforms to the general time series format (with time step as the main axis), which is convenient for subsequent use.

[0035] S3. The sequence-to-sequence prediction model is trained using the frequency-enhanced direct prediction training paradigm, wherein the total loss function of the frequency-enhanced direct prediction training paradigm is composed of a weighted sum of the time-domain loss component and the frequency-domain loss component. Further, in step S3, the frequency domain loss component is calculated as follows: the predicted sequence and the real sequence are respectively transformed to the frequency domain by Fourier transform, and the L1 norm difference between the two in the frequency domain is calculated.

[0036] Specifically, Frequency Enhancement Direct Prediction (Fre-DF) transforms both the predicted and true sequences to the frequency domain using Fourier transform. The prediction error is then calculated in this domain, taking advantage of the fact that most frequency components are uncorrelated. A simplified schematic diagram of the FreDF-ITransformer is shown below. Figure 3 As shown, the execution process of Fre-DF is completed through the following key stages: (1) Output transpose: Consistent with the standard DF model paradigm, the prediction model used in this study as the basis (Specifically implemented as ITransformer) processes the historical input sequence T to generate a multi-step prediction result Y.

[0037] (2) Calculation of double loss: Temporal loss L tmp The L2 norm is used to quantify the time-domain difference between the predicted and actual values.

[0038] (8) Frequency domain loss L freq The sequence is transformed to the frequency domain by using Fast Fourier Transform (FFT) and the difference is calculated using the L1 norm, which solves the problem of traditional prediction ignoring the autocorrelation of the target.

[0039] (9) in, For frequency domain loss components, For the true sequence, This is the predicted sequence.

[0040] Weighted total loss: This integrates the two loss components into a final objective function, which is a weighted sum of the time-domain loss and the frequency-domain loss. Hyperparameters Used to balance the influence weights of the optimization objective in the time and frequency domains.

[0041] (10) in, For the time-domain loss component, These are the preset weighting coefficients.

[0042] In this embodiment, the input data preprocessing is as follows: each input 21×1 voltage sequence is independently standardized, and the sequence is processed to have zero mean and zero unit variance through normalization; the normalized sequence values ​​and time index are input into the embedding layer, and mapped to the D-dimensional feature space through a multilayer perceptron (MLP); the state of the encoder at the last time step is extracted, and the input to the linear layer is mapped to the voltage prediction values ​​for 5 consecutive hours, and finally restored to the actual voltage unit through inverse normalization.

[0043] S4. The hyperparameters of the trained sequence-to-sequence prediction model are optimized using the Bayesian optimization method, and the model performance is evaluated using the cross-validation method to determine the final prediction model. Further, in step S4, the Bayesian optimization method is used to optimize at least one hyperparameter among the embedding dimension, number of encoder layers, and number of attention heads of the sequence-to-sequence prediction model.

[0044] Furthermore, in step S4, the cross-validation method is K-fold cross-validation, where K is greater than or equal to 5.

[0045] Furthermore, in step S4, the average error of multiple verification results obtained by K-fold cross-validation is used as an indicator to evaluate the performance of the hyperparameter combination.

[0046] Specifically, step S4 mainly consists of two parts: (1) Key hyperparameters of Bayesian optimization of ITransformer include embedding dimension D, number of encoder layers L, number of multi-head attention heads, dimension of feedforward hidden layers, and learning rate.

[0047] (2) Five-fold cross-validation experiment: First, 60% of the training set (the top 60% of the total samples) is randomly divided into 5 equal-sized subsets, denoted as F1, F2, F3, F4, and F5. Then, 100 Bayesian hyperparameter searches are performed. After each search, five-fold cross-validation is immediately performed, that is, the entire training set is divided into 5 folds. 4 folds are used as the sub-training set, and 1 fold is used as the sub-validation set. A new ITransformer is created using the hyperparameters found in this search, and it is trained on the sub-training set. Then, predictions are made on the sub-validation set and the root mean square error (RMSE) is calculated. This process is repeated 5 times, and the average of the 5 obtained RMSE values ​​is used as the RMSE index of this hyperparameter search. Among the average root mean square error (RMSE_avg) of 100 searches, the hyperparameter combination corresponding to the minimum value is selected as the optimal hyperparameters for subsequent full model training.

[0048] S5. The voltage degradation trend of the proton exchange membrane fuel cell is predicted using the final prediction model.

[0049] Furthermore, in step S5, the voltage values ​​of the proton exchange membrane fuel cell at multiple consecutive time points in the future are predicted in multiple steps.

[0050] Specifically, step S5 includes: (1) Dataset partitioning and model training: The total samples were divided into a training set (the first 60% of samples, corresponding to the early to mid-stage degradation of PEMFC) and a validation set (the last 40% of samples, corresponding to the mid to late-stage degradation) in a 6:4 ratio. Each sample contained 26 consecutive hours of voltage data (21 hours of input + 5 hours of output) to ensure that the validation set covered the more complex late-stage degradation scenario. The ITransformer model was trained using the complete training set and the optimal hyperparameters, and mini-batch gradient descent was used until the loss function converged.

[0051] (2) Validation Set Prediction and Performance Evaluation: Input validation samples and output 5-hour voltage prediction values ​​through the trained model; average the prediction values ​​for overlapping periods to ensure a continuous prediction sequence in the output. Performance evaluation uses quantitative indicators to calculate the MAE, RMSE, and R² of the validation set prediction values ​​and measured values; specific predicted voltage time series curves and measured voltage time series curves are shown in the figure. Figure 4 As shown.

[0052] This embodiment includes a comparative experiment. Since Long Short-Term Memory (LSTM) is a classic model in the field of time series prediction and is widely used in PEMFC degradation prediction research, it was chosen as the comparative model. The training process, dataset partitioning, hyperparameter optimization method (same as Bayesian optimization + five-fold cross-validation), and evaluation metrics (same as MAE / RMSE / R²) of LSTM are completely consistent with FreDF-ITransformer.

[0053] The comparative experimental data of FreDF-ITransformer and LSTM are shown in Table 1. Table 1 presents the differences in metrics between the two models on the training and validation sets. FreDF-ITransformer has a lower RMSE, lower MAE, and higher R² than LSTM, which quantitatively demonstrates its advantages in generalization ability and resistance to outliers.

[0054] Table 1 The beneficial effects of this embodiment: In this embodiment, FreDF introduces a dual-loss calculation: the time-domain loss ensures the basic prediction accuracy, while the frequency-domain loss maps the sequence to the frequency domain through Fourier transform and uses the L1 norm to balance the sensitivity of frequency components with different amplitudes. This compensates for the deficiency of traditional direct prediction (DF) in ignoring target autocorrelation and significantly reduces the cumulative error of multi-step prediction.

[0055] This embodiment introduces the FreDF-ITransformer algorithm into PEMFC degradation prediction for the first time, adding "instance normalization + time embedding" preprocessing. By extracting the aggregated information of the last time step, the prediction results are mapped to adapt to the nonlinear degradation characteristics under PEMFC dynamic load.

[0056] This embodiment uses Bayesian optimization instead of traditional grid search, predicting the performance of hyperparameter combinations through probabilistic modeling to guide the search direction. In terms of effectiveness, the optimal parameters can be found in only 100 searches, saving significant computation time compared to grid search (which requires hundreds of iterations), while ensuring that the model's mean absolute error (MAE) on the training set is close to the theoretical optimum.

[0057] This embodiment adopts a combination of "Bayesian optimization + five-fold cross-validation". Bayesian optimization intelligently searches for hyperparameters through probabilistic modeling, while five-fold cross-validation eliminates the randomness of a single data partition. The combination of the two saves computational resources and ensures the stability of model performance evaluation.

[0058] This embodiment verifies the superiority of FreDF-ITransformer in the three core metrics of MAE, RMSE, and R² through comparative experiments with the classic LSTM model. It is particularly outstanding in capturing long-term dependencies and resisting outlier interference, providing a more reliable technical option for PEMFC degradation prediction.

[0059] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting the degradation trend of proton exchange membrane fuel cells based on direct prediction of frequency enhancement, characterized in that, Includes the following steps: S1. Obtain voltage time-series aging data of proton exchange membrane fuel cells and preprocess it to construct an input sample set; S2. Construct a sequence-to-sequence prediction model based on the ITransformer architecture; S3. The sequence-to-sequence prediction model is trained using the frequency-enhanced direct prediction training paradigm, wherein the total loss function of the frequency-enhanced direct prediction training paradigm is composed of a weighted sum of the time-domain loss component and the frequency-domain loss component. S4. The hyperparameters of the trained sequence-to-sequence prediction model are optimized using the Bayesian optimization method, and the model performance is evaluated using the cross-validation method to determine the final prediction model. S5. The voltage degradation trend of the proton exchange membrane fuel cell is predicted using the final prediction model.

2. The prediction method according to claim 1, characterized in that, In step S1, the preprocessing includes using variational mode decomposition to denoise the voltage time series aging data, and using a sliding window method to process the denoised time series to construct the input sample set, wherein the input time length of each sample is greater than the output prediction time length.

3. The prediction method according to claim 1, characterized in that, In step S2, the input to the sequence-to-sequence prediction model is a voltage history sequence that has undergone instance normalization.

4. The prediction method according to claim 1, characterized in that, In step S3, the frequency domain loss component is calculated as follows: the predicted sequence and the real sequence are transformed to the frequency domain by Fourier transform, and the L1 norm difference between the two in the frequency domain is calculated.

5. The prediction method according to claim 4, characterized in that, In step S3, the formula for the frequency domain loss component is: ； in, For frequency domain loss components, For the true sequence, This is the predicted sequence.

6. The prediction method according to claim 4, characterized in that, In step S3, the total loss function is: ； in, For the time-domain loss component, These are the preset weighting coefficients.

7. The prediction method according to claim 1, characterized in that, In step S4, the Bayesian optimization method is used to optimize at least one hyperparameter of the sequence-to-sequence prediction model, including the embedding dimension, the number of encoder layers, and the number of attention heads.

8. The prediction method according to claim 7, characterized in that, In step S4, the cross-validation method is K-fold cross-validation, where K is greater than or equal to 5.

9. The prediction method according to claim 8, characterized in that, In step S4, the average error of multiple verification results obtained by K-fold cross-validation is used as an indicator to evaluate the performance of the hyperparameter combination.

10. The prediction method according to claim 1, characterized in that, In step S5, the voltage values ​​of the proton exchange membrane fuel cell at multiple consecutive time points in the future are predicted in multiple steps.

Images