Method for quantifying and decomposing uncertainty of fuel cell aging prediction results

Abstract

The present application relates to the technical field of fuel cell, disclose a kind of uncertainty quantification and decomposition method of fuel cell aging prediction result, comprising the following steps: voltage data acquisition and preprocessing;B-MIXGU aging prediction model modeling and training;Aging prediction and result analysis;Uncertainty quantification and decomposition, the total uncertainty of aging prediction result is quantified using the variance of interval estimation, and the total uncertainty is decomposed into accidental uncertainty and cognitive uncertainty, to explore the source of uncertainty.The present application quantifies and decomposes the uncertainty of the aging prediction result of the fuel cell, decomposes the total uncertainty of the prediction into cognitive uncertainty and accidental uncertainty, and realizes the credibility evaluation of the aging prediction result according to the uncertainty decomposition result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fuel cells, and particularly relates to a method for quantifying and decomposing the uncertainty of a fuel cell aging prediction result. BACKGROUND

[0002] Proton exchange membrane fuel cells (PEMFC) are widely used in the fields of transportation such as hybrid electric vehicles, buses, ships, and airplanes due to their advantages of no pollution, high efficiency, and low noise. However, insufficient life is one of the key problems hindering its large-scale commercialization.

[0003] Under dynamic operating conditions, the interference of external environment and load changes can have a significant impact on the durability of PEMFC. Prognosis and health management (PHM) can estimate the remaining useful life (RUL) by providing degradation information of PEMFC. Therefore, accurately predicting the performance degradation trend of PEMFC is crucial for estimating its RUL.

[0004] Methods for predicting the performance degradation trend of PEMFC mainly include model-based methods, data-driven methods, and hybrid methods combining model and data-driven methods. Related research shows that PEMFC is a nonlinear, multi-physical parameter, and multi-scale characteristic system, and it is currently difficult to establish an accurate physical model to describe its degradation mechanism. Model-based methods rely on complex and high-precision physical models, so it is difficult to achieve ideal prediction results. Data-driven methods learn directly from PEMFC data to predict its performance degradation, which effectively overcomes the difficulty of establishing a high-precision physical model. Hybrid methods attempt to combine the advantages of both, but they rely on high-precision physical models, so it is quite challenging to accurately construct such models.

[0005] In addition, with the rapid development of machine learning technology, especially the strong generalization ability, excellent scalability, and high precision of deep learning technology in handling complex data, data-driven methods based on deep learning have become an effective tool for predicting the performance degradation of PEMFC.

[0006] Currently, there are various data-driven methods based on machine learning, such as support vector machines, particle filters, adaptive unscented Kalman filters, gated recurrent units, long short-term memory recurrent neural networks, etc. Although these methods have achieved many remarkable successes, they can usually only guarantee the effectiveness of the prediction results when the test data distribution is similar to the training data distribution. In practical applications, the distribution of the test data may be domain-shifted from the training data distribution due to external environmental and load changes, and the domain shift has a great impact on the PEMFC performance degradation prediction. When facing unknown data, deep learning models may often give overconfident and unreliable prediction results without explicit warnings. Therefore, in the PEMFC performance degradation prediction, not only the accuracy of the prediction results needs to be focused on, but also the reliability of the prediction results needs to be valued.

[0007] There are mainly two sources of uncertainty in prediction: aleatory uncertainty and epistemic uncertainty. Although uncertainty quantification is crucial for establishing a reliable prediction model, it is often ignored in PEMFC performance degradation prediction. Most deep learning-based methods only provide point estimate predictions and cannot fully convey the reliability of the prediction results.

[0008] Therefore, developing a model that can effectively quantify and convey the prediction uncertainty is of great significance to improve the reliability and practicality of PEMFC aging prediction. SUMMARY

[0009] The purpose of the present application is to overcome the above technical problems, and provide a fuel cell aging prediction result uncertainty quantification and decomposition method, which decomposes the total uncertainty of the prediction into epistemic uncertainty and aleatory uncertainty, and realizes the reliability evaluation of the aging prediction result according to the uncertainty decomposition result.

[0010] To achieve the above purpose, the fuel cell aging prediction result uncertainty quantification and decomposition method provided by the present application comprises the following steps:

[0011] S1) Voltage data acquisition and preprocessing: collect voltage data output under the dynamic load cycle working condition of the PEMFC, perform preprocessing operation to obtain a voltage data set, and divide the voltage data set into training set data and test set data;

[0012] S2) Modeling and training of the B-MIXGU aging prediction model: the model parameters of the MIXGU model are replaced by probability distributions to obtain a Bayesian MIXGU model B-MIXGU, the posterior distribution of the B-MIXGU model parameters is obtained based on Bayes' theorem, and the output of B-MIXGU is obtained according to the posterior distribution, thereby establishing the B-MIXGU aging prediction model, the method of variational inference is used to approximately solve the posterior distribution of the model parameters of the B-MIXGU aging prediction model, by solving the minimum KL divergence between the variational distribution and the posterior distribution, it is further transformed into optimizing the lower bound of the evidence, and the trained B-MIXGU aging prediction model is obtained by combining the training set data;

[0013] S3) Aging prediction and result analysis: input the test set data into the trained B-MIXGU aging prediction model to obtain the point estimation and interval estimation of the aging prediction results of PEMFC, and analyze the aging prediction results through the error evaluation index;

[0014] S4) Uncertainty quantification and decomposition: the cognitive uncertainty measured by the randomness of the model parameters and the accidental uncertainty caused by the data noise jointly act on the total uncertainty, the variance of the interval estimation is used to quantify the total uncertainty of the aging prediction results, and the total uncertainty is decomposed into accidental uncertainty and cognitive uncertainty to explore the source of uncertainty.

[0015] Preferably, in step B), MIXGU includes a gated recurrent unit GRU, a minimum gated unit MGU and an adaptive weight adjustment module AWAM, and the expression is as follows:

[0016] GRU:

[0017] z k = σ(W z [h k-1 ,x k ]+b z )

[0018] r k = σ(W r [h k-1 ,x k ]+b r )

[0019]

[0020] In the formula, x k is the input information at the current time k, h k-1 is the hidden state at the last time, h k1 is the hidden state of GRU transmitted to AWAM, is the candidate hidden state of GRU, and z kis the update gate of GRU, r k is the reset gate of GRU, σ is the sigmoid function, tanh is the tanh function, W z Update the weight matrix of the GRU gate, b z The bias matrix of the GRU update gate, W r Reset the gate weight matrix for GRU, b r Reset the bias matrix of the GRU gate, W h1 is the weight matrix of the GRU candidate hidden state, b h1 is the bias matrix of the GRU candidate hidden state;

[0021] MGU:

[0022] f k =σ(W f [h k-1 ,x k ]+b f )

[0023]

[0024] Where, f k is the update gate of MGU, h k2 is the hidden state passed from MGU to AWAM, is the candidate hidden state of MGU, W f Update the weight matrix of the MGU gate, b f Update the bias matrix of the MGU gate, W h2 is the weight matrix of the MGU candidate hidden state, b h2 is the bias matrix of the MGU candidate hidden state;

[0025] AWAM:

[0026] MIXGU k =σ(W s [h k1 ,h k2 ]+b s )

[0027] h k =(1-MIXGU k )*h k1 +MIXGU k *h k2

[0028] Where, MIXGU k is the weight adjustment module of MIXGU at the current moment k, W s for MIXGU k The corresponding weight matrix of the weight adjustment module, b sMIXGU k corresponding bias matrix of the weight adjustment module, h k MIXGU k MIXGU k MIXGU k MIXGU k1 MIXGU k MIXGU k MIXGU k2 MIXGU

[0029] Preferably, the step B) comprises the following steps:

[0030] S201) replace the model parameters of MIXGU with a probability distribution to obtain B-MIXGU, and obtain the posterior distribution of the model parameters ω of B-MIXGU based on Bayes theorem:

[0031]

[0032] In the formula, p(ω|D) is the posterior distribution of the parameter ω, D is the training set data, p(ω) is the prior distribution of the parameter ω, p(D|ω) is the likelihood function of the parameter ω, and p(D) is the prior distribution of the training set data. The parameter ω includes W z , W r , b z , b r , W h1 , b h1 , W f , b f , W h2 , b h2 , W s , b s ;

[0033] S202) obtain the output of B-MIXGU according to the posterior distribution, and establish a B-MIXGU aging prediction model:

[0034] p(y|x,D)=∫p(y|x,ω)p(ω,D)dω

[0035] In the formula, x is input data, y is model output, p(y|x,D) is an output distribution under given training data D and input data x, and p(y|x,ω) is an output distribution under given parameter ω and input data x.

[0036] S203) approximate the posterior distribution of the model parameters of the B-MIXGU aging prediction model by using a variational inference method, and create a variational distribution q θ(ω) to approximate the true posterior distribution p(ω|D);

[0037] S204) Further transformed into optimization of maximizing the evidence lower bound ELBO by solving the minimum KL divergence between the variational distribution and the posterior distribution:

[0038] argminKL[q θ (ω)||p(ω|D)]=argmaxELBO;

[0039] S205) Obtain the trained B-MIXGU aging prediction model combined with the training set data.

[0040] Preferably, the step S4) comprises the following steps:

[0041] S401) The cognitive uncertainty measured by the randomness of the model parameters and the accidental uncertainty caused by the data noise jointly act on the total uncertainty;

[0042] p(y|x,D)=∫p(y|x,ω)q θ (ω)dω

[0043] Where the randomness of the model parameters is represented by the variational distribution q θ (ω), which quantifies the cognitive uncertainty, and the data noise is represented by the probability distribution p(y|x,ω), which quantifies the accidental uncertainty;

[0044] S402) Adopt the variance of interval estimation to quantify the total uncertainty of the aging prediction result, decompose the total uncertainty into accidental uncertainty and cognitive uncertainty, and explore the source of uncertainty;

[0045] D(y)=E(y 2 )-E(y) 2

[0046] =E[E(y 2 |x)]-E[E(y|x)] 2

[0047] =E[D(y|x)+E(y|x) 2 ]-E[E(y|x)] 2

[0048] =E[E(y|x)] 2 -E[E(y|x) 2 ]+E[D(y|x)]

[0049] =D[E(y|x)]+E[D(y|x)]

[0050] Wherein, D(y) represents the total prediction uncertainty, D[E(y|x)] represents the fluctuation of the expected E(y|x) caused by the variability of the model parameters, measures the cognitive uncertainty, E[D(y|x)] represents the average value of D(y|x), eliminates the influence caused by the variability of the model parameters, and extracts the accidental uncertainty caused by random noise.

[0051] Compared with the prior art, the present application has the following advantages:

[0052] 1. The B-MIXGU model has very excellent PEMFC aging prediction accuracy, and the point estimation prediction accuracy of B-MIXGU is much better than that of other advanced recurrent neural network models through dynamic load cycle working condition durability experimental data set verification, and B-MIXGU can accurately predict PEMFC aging according to very little training data, and is more suitable for practical application.

[0053] 2. An uncertainty perception method for PEMFC aging prediction in a probabilistic Bayesian framework is proposed, the variance of the interval prediction result under the probability distribution of the model parameters can be used to measure the total uncertainty, the uncertainty information can be used to distinguish the invisible domain, the reliability evaluation of the aging prediction result is realized, and the transparency of the aging prediction model is improved.

[0054] 3. An uncertainty decomposition method is proposed to decompose the total prediction uncertainty into cognitive uncertainty and accidental uncertainty, the uncertainty decomposition and source tracing method proposed in terms of uncertainty caused by invisible load current working conditions and invisible noise condition changes is effective, and can help experts analyze whether the uncertainty is caused by insufficient cognition of measurement data or noise interference. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 A flowchart of the uncertainty quantification and decomposition method for the fuel cell aging prediction result of the present application;

[0056] Figure 2 A flowchart of the uncertainty quantification and decomposition method for the fuel cell aging prediction result of the present application. DETAILED DESCRIPTION

[0057] The present application will be further described in detail below in combination with the drawings and specific embodiments.

[0058] As shown in Figure 1 and Figure 2 , an uncertainty quantification and decomposition method for a fuel cell aging prediction result includes the following steps:

[0059] S1) Collection and preprocessing of voltage data: collect the voltage data output by the PEMFC under dynamic load cycle conditions, perform preprocessing operations to obtain a voltage data set, and divide the voltage data set into training set data and test set data;

[0060] S2) Modeling and training of B-MIXGU aging prediction model: replace the model parameters of the MIXGU model with a probability distribution to obtain the Bayesian mixture gate unit model B-MIXGU, obtain the posterior distribution of the B-MIXGU model parameters based on Bayes' theorem, and obtain the output of B-MIXGU according to the posterior distribution, thereby establishing the B-MIXGU aging prediction model. The method of variational inference is used to approximate the posterior distribution of the model parameters of the B-MIXGU aging prediction model, and by solving the minimum KL (Kullback-Leibler) divergence between the variational distribution and the posterior distribution, it is further transformed into an optimization of the lower bound of the evidence. Combined with the training set data, a trained B-MIXGU aging prediction model is obtained;

[0061] S3) Aging prediction and result analysis: input the test set data into the trained B-MIXGU aging prediction model to obtain the point estimate and interval estimate of the aging prediction results of the PEMFC, and analyze the aging prediction results through the error evaluation index;

[0062] S4) Uncertainty quantification and decomposition: the cognitive uncertainty measured by the randomness of the model parameters and the accidental uncertainty caused by the data noise jointly act on the total uncertainty. The variance of the interval estimate is used to quantify the total uncertainty of the aging prediction results, and the total uncertainty is decomposed into accidental uncertainty and cognitive uncertainty to explore the source of uncertainty.

[0063] In step B), MIXGU includes: a gated recurrent unit GRU, a minimum gated unit MGU, and an adaptive weight adjustment module AWAM, expressed as follows:

[0064] GRU:

[0065] z k =σ(W z [h k-1 ,x k ]+b z )

[0066] r k =σ(W r [h k-1 ,x k ]+b r )

[0067]

[0068] In the formula, xk is the input information for the current time instance k, h k-1 is the hidden state of the previous time instance, h k1 is the hidden state of the GRU passed to the AWAM, is the candidate hidden state of the GRU, z k is the update gate of the GRU, r k is the reset gate of the GRU, σ is the sigmoid function, tanh is the tanh function, W z is the weight matrix of the GRU update gate, b z is the bias matrix of the GRU update gate, W r is the weight matrix of the GRU reset gate, b r is the bias matrix of the GRU reset gate, W h1 is the weight matrix of the GRU candidate hidden state, b h1 is the bias matrix of the GRU candidate hidden state;

[0069] MGU:

[0070] f k = σ(W f [h k-1 ,x k ]+b f )

[0071]

[0072] where f k is the update gate of the MGU, h k2 is the hidden state of the MGU passed to the AWAM, is the candidate hidden state of the MGU, W f is the weight matrix of the MGU update gate, b f is the bias matrix of the MGU update gate, W h2 is the weight matrix of the MGU candidate hidden state, b h2 is the bias matrix of the MGU candidate hidden state;

[0073] AWAM:

[0074] MIXGU k = σ(W s [h k1 ,h k2 ]+b s )

[0075] h k = (1 - MIXGU k )* h k1 + MIXGU k * h k2

[0076] Where, MIXGU k is the weight adjustment module of MIXGU at the current moment k, W s for MIXGU k The corresponding weight matrix of the weight adjustment module, b s for MIXGU k The corresponding bias matrix of the weight adjustment module, h k is the final hidden state output, MIXGU k Decision, when MIXGU k When it approaches 0, h k tends to h k1 , indicating that GRU is dominant, when MIXGU k When it approaches 1, h k tends to h k2 , indicating that MGU is dominant.

[0077] Specifically, step B) includes the following steps:

[0078] S201) The model parameters of MIXGU are replaced with probability distribution to obtain B-MIXGU, and the posterior distribution of the B-MIXGU model parameter ω is obtained based on Bayes' theorem:

[0079]

[0080] Where p(ω|D) is the posterior distribution of parameter ω, D is the training set data, p(ω) is the prior distribution of parameter ω, p(D|ω) is the likelihood function of parameter ω, p(D) is the prior distribution of training set data, and parameter ω includes W in the GRU expression z , W r , b z , b r , W h1 , b h1 , W in the MGU expression f , b f , W h2 , b h2 , W in the AWAM expression s , b s ;

[0081] S202) Obtaining the output of B-MIXGU according to the posterior distribution and establishing a B-MIXGU aging prediction model:

[0082] p(y|x,D)=∫p(y|x,ω)p(ω,D)dω

[0083] In the formula, x is input data, y is model output, p(y|x, D) is an output distribution under the given training data D and input data x, and p(y|x, ω) is an output distribution under the given parameters ω and input data x.

[0084] S203) The posterior distribution of the model parameters of the B-MIXGU aging prediction model is approximately solved by using the method of variational inference, and a variational distribution q θ (ω) is used to approximate the real posterior distribution p(ω|D);

[0085] S204) By solving the minimum KL divergence between the variational distribution and the posterior distribution, it is further transformed into optimizing the maximum evidence lower bound ELBO:

[0086] argminKL[q θ (ω)||p(ω|D)]=argmaxELBO;

[0087] S205) The trained B-MIXGU aging prediction model is obtained in combination with the training set data.

[0088] In addition, step S4) includes the following steps:

[0089] S401) The cognitive uncertainty measured by the randomness of the model parameters and the accidental uncertainty caused by the data noise jointly act on the total uncertainty;

[0090] p(y|x,D)=∫p(y|x,ω)q θ (ω)dω

[0091] Wherein, the randomness of the model parameters is represented by the variational distribution q θ (ω), which quantifies the cognitive uncertainty, and the data noise is represented by the probability distribution p(y|x, ω), which quantifies the accidental uncertainty;

[0092] S402) The variance of interval estimation is used to quantify the total uncertainty of the aging prediction result, and the total uncertainty is decomposed into accidental uncertainty and cognitive uncertainty to explore the source of uncertainty;

[0093] D(y)=E(y 2 )-E(y) 2

[0094] =E[E(y 2 |x)]-E[E(y|x)] 2

[0095] =E[D(y|x)+E(y|x) 2 ]-E[E(y|x)] 2

[0096] = E [E (y | x)] 2 - E [E (y | x) 2 + E [D (y | x)]

[0097] = D [E (y | x)] + E [D (y | x)]

[0098] Where D (y) represents the total prediction uncertainty, D [E (y | x)] represents the fluctuation of E (y | x) due to the variability of model parameters, which measures the epistemic uncertainty, E [D (y | x)] represents the average value of D (y | x), which eliminates the influence caused by the variability of model parameters, so as to extract the accidental uncertainty caused by random noise.

[0099] In the PEMFC aging prediction process, when facing unknown data, the deep learning model based on machine learning data-driven method may give overconfident and unreliable prediction results without explicit warning. Therefore, although uncertainty quantification is crucial for establishing a reliable prediction model, it is often ignored in PEMFC performance degradation prediction. In order to fill the research gap of quantifying and decomposing the uncertainty in the PEMFC aging prediction process, the present application proposes a novel Bayesian mixed gate unit model (B-MIXGU) under the framework of probabilistic Bayesian deep learning, which is used for PEMFC performance degradation prediction and uncertainty quantification and decomposition, and provides the reliability of the aging prediction result. The aging prediction model based on B-MIXGU combines probabilistic Bayesian theory with MIXGU, adopts random probability distribution to replace MIXGU model parameters, uses variational inference method to derive the optimal distribution of model parameters, greatly improves the prediction accuracy, especially when the proportion of training data set is very low. The uncertainty perception method of PEMFC aging prediction under the framework of probabilistic Bayesian framework uses the variance of interval estimation result to quantify the total uncertainty, uses the posterior distribution of model parameters to perceive epistemic uncertainty, and uses the probability distribution of model output to perceive accidental uncertainty, realizes the reliability evaluation of aging prediction result, and improves the transparency of aging prediction model. And through the change of uncertainty component under the condition of invisible working condition and invisible noise, the source of aging prediction result uncertainty is traced back, which helps experts to analyze whether the uncertainty is caused by insufficient cognition of measurement data or noise interference.

[0100] In the Bayesian framework, the aleatoric and epistemic uncertainties are integrated into the probability distribution of the prediction results, and then the variance of the interval estimation is used to represent the total uncertainty, and the contributions of the aleatoric and epistemic uncertainties to the prediction variance are analyzed. This method can effectively reflect the credibility of the prediction results in the PEMFC performance degradation prediction. High total uncertainty indicates low credibility of the prediction results, and low total uncertainty indicates high credibility of the prediction results. Higher epistemic uncertainty indicates greater distribution difference between the test set and the training set. Higher aleatoric uncertainty indicates more unknown quantities in the test set data. In summary, by quantifying the uncertainty, the prediction results of the model can be better understood, and the prediction credibility analysis results can be combined with human expert judgment to accurately estimate the RUL of the PEMFC.

Claims

1. A method for quantifying and decomposing uncertainty in fuel cell aging prediction results, characterized by: The steps include: S1) voltage data collection and preprocessing: collecting the voltage data output under the PEMFC dynamic load cycle condition, performing preprocessing operations to obtain a voltage data set, and dividing the voltage data set into training set data and test set data; S2) Modeling and training of the B-MIXGU aging prediction model: The model parameters of the mixed gated unit model MIXGU are replaced with probability distributions to obtain the Bayesian mixed gated unit model B-MIXGU. The posterior distribution of the B-MIXGU model parameters is obtained based on Bayes' theorem. The output of B-MIXGU is obtained based on the posterior distribution, thereby establishing the B-MIXGU aging prediction model. The posterior distribution of the model parameters of the B-MIXGU aging prediction model is approximated using variational inference. By solving the minimum KL divergence between the variational distribution and the posterior distribution, it is further converted into optimizing the maximum evidence lower bound. Combined with the training set data, the trained B-MIXGU aging prediction model is obtained. S3) Aging prediction and result analysis: The test set data is input into the trained B-MIXGU aging prediction model to obtain the point estimate and interval estimate aging prediction results of PEMFC. The aging prediction results are analyzed using error evaluation indicators. S4) Uncertainty quantification and decomposition: Epistemic uncertainty measured by the randomness of model parameters and aleatory uncertainty caused by data noise contribute jointly to the total uncertainty. The variance of interval estimates is used to quantify the total uncertainty of aging prediction results. The total uncertainty is then decomposed into aleatory uncertainty and epistemic uncertainty to explore the sources of uncertainty.

2. The method for quantifying and decomposing uncertainty in fuel cell aging prediction results according to claim 1, wherein: In step S2), MIXGU includes: a gated recurrent unit GRU, a minimum gated unit MGU and an adaptive weight adjustment module AWAM, and the expression is as follows: GRU: z k =σ(W z [h k-1 ,x k ]+b z ) r k =σ(W r [h k-1 ,x k ]+b r ) Where x k Input information for the current moment k, h k-1 is the hidden state at the previous moment, h k1 is the hidden state passed from GRU to AWAM, is the candidate hidden state of GRU, z k is the update gate of GRU, r k is the reset gate of GRU, σ is the sigmoid function, tanh is the tanh function, W z Update the weight matrix of the GRU gate, b z The bias matrix of the GRU update gate, W r Reset the gate weight matrix for GRU, b r Reset the bias matrix of the GRU gate, W h1 is the weight matrix of the GRU candidate hidden state, b h1 is the bias matrix of the GRU candidate hidden state; MGU: f k =σ(W f [h k-1 ,x k ]+b f ) Where, f k is the update gate of MGU, h k2 is the hidden state passed from MGU to AWAM, is the candidate hidden state of MGU, W f Update the weight matrix of the MGU gate, b f Update the bias matrix of the MGU gate, W h2 is the weight matrix of the MGU candidate hidden state, b h2 is the bias matrix of the MGU candidate hidden state; AWAM: MIXGU k =σ(W s [h k1 ,h k2 ]+b s ) h k =(1-MIXGU k )*h k1 +MIXGU k *h k2 Where, MIXGU k is the weight adjustment module of MIXGU at the current moment k, W s for MIXGU k The corresponding weight matrix of the weight adjustment module, b s for MIXGU k The corresponding bias matrix of the weight adjustment module, h k is the final hidden state output, MIXGU k Decision, when MIXGU k When it approaches 0, h k tends to h k1 , indicating that GRU is dominant, when MIXGU k When it approaches 1, h k tends to h k2 , indicating that MGU is dominant.

3. The method for quantifying and decomposing uncertainty in fuel cell aging prediction results according to claim 2, wherein: The step S2) comprises the following steps: S201) The model parameters of MIXGU are replaced with probability distribution to obtain B-MIXGU, and the posterior distribution of the B-MIXGU model parameter ω is obtained based on Bayes' theorem: Where p(ω|D) is the posterior distribution of parameter ω, D is the training set data, p(ω) is the prior distribution of parameter ω, p(D|ω) is the likelihood function of parameter ω, p(D) is the prior distribution of training set data, and parameter ω includes W in the GRU expression z , W r , b z , b r , W h1 , b h1 , W in the MGU expression f , b f , W h2 , b h2 , W in the AWAM expression s , b s ; S202) Obtaining the output of B-MIXGU according to the posterior distribution and establishing a B-MIXGU aging prediction model: p(y|x,D)=∫p(y|x,ω)p(ω|D)dω Where x is the input data, y is the model output, p(y|x,D) is the output distribution given the training data D and the input data x, and p(y|x,ω) is the output distribution given the parameter ω and the input data x. S203) Using the variational inference method to approximate the posterior distribution of the model parameters of the B-MIXGU aging prediction model, a variational distribution q with a parameter θ is created. θ (ω) to approximate the true posterior distribution p(ω|D); S204) By solving the minimum KL divergence between the variational distribution and the posterior distribution, it is further transformed into optimizing the maximization of the evidence lower bound ELBO: argminKL[q θ (ω)||p(ω|D)]6argmaxELBO. S205) Combining the training set data to obtain a trained B-MIXGU aging prediction model.

4. The method for quantifying and decomposing uncertainty in fuel cell aging prediction results according to claim 3, wherein: The step S4) comprises the following steps: S401) Epistemic uncertainty measured by the randomness of model parameters and aleatoric uncertainty caused by data noise contribute jointly to the total uncertainty; p(y|x,D)=∫p(y|x,ω)q θ (ω)dω Among them, the randomness of the model parameters is given by the variational distribution q θ (ω) represents,quantifying epistemic uncertainty, and data noise is represented by the probability distribution p(y|x,ω),quantifying aleatoric uncertainty; S402) using the variance of interval estimates to quantify the total uncertainty of aging prediction results, decomposing the total uncertainty into aleatory uncertainty and epistemic uncertainty, and exploring the sources of uncertainty; D(y)=E(y 2 )-Hey) 2 =E[E(y 2 |x)]-E[E(y|x)] 2 =E[D(y|x)+E(y|x) 2 ]-E[E(y|x)] 2 =E[E(y|x)] 2 -E[E(y|x) 2 ]+E[D(y|x)] =D[E(y|x)]+E[D(y|x)] Where D(y) represents the total prediction uncertainty, D[E(y|x)] represents the fluctuation of expected E(y|x) due to the variability of model parameters, which measures epistemic uncertainty, and E[D(y|x)] represents the calculation of the average value of D(y|x), eliminating the effects caused by the variability of model parameters, thereby extracting the accidental uncertainty caused by random noise.

Images