Research On Deep Learning Based Performance Degradation Prediction For Fuel Cell

Proton exchange membrane fuel cell(PEMFC)is a new power generation device based on hydrogen energy,which has outstanding advantages such as high energy conversion efficiency,high specific power,low working temperature,and environmentfriendly,is considered as an energy source solution that realize carbon neutralization.As the key industries during the 14 th Five-Year Plan period,hydrogen energy and fuel cell have received more and more attention and policy support.However,the lower reliability and poor durability of PEMFC still restrict its large-scale application and industrialization.The performance degradation prediction and health management(PHM)technology in the practical application of PEMFC is the main way to solve their reliability and durability problems.At present,there is an urgent need for accurate and efficient PEMFC performance degradation prediction technology to provide information support for system maintenance strategy design,operating condition optimization,and health status assessment.As a complex electrochemical system involving multiple physical quantities,multiple components,and multiple factors,PEMFC has strong nonlinearity,randomness,and uncertainty in its performance degradation process,which significantly increases prediction difficulty.At present,the PEMFC performance degradation prediction method still has problems such as low prediction accuracy,poor generalization,and lack of uncertainty expression.In response to the above problems,this paper focuses on researching and establishing a set of PEMFC performance degradation prediction theories and methods based on deep learning.By making full use of PEMFC degradation data,effectively mining the deep features in the data to improve the accuracy and reliability of PEMFC performance degradation prediction and robustness.The main research content and contributions of the thesis are summarized as follows:(1)The mechanism and characteristics of the fuel cell performance degradation are analyzed in this thesis firstly.The basic principles,implementation framework,specific prediction processes,and mainstream methods of PEMFC performance degradation prediction are introduced subsequently.On this basis,the degradation prediction performance of the shallow prediction methods and deep learning methods is compared,and the prediction performance of the given prediction method trained on different preprocessing data is also compared.The experimental prediction results show that the effective data preprocessing method and deep learning prediction method can improve the performance of PEMFC performance degradation prediction,which also lay a research foundation for the subsequent design of prediction methods.(2)Aiming at the problem that the performance degradation prediction methods based on the shallow machine learning models can not accurately characterize the strong nonlinear degradation characteristics of PEMFC,resulting in low prediction accuracy,a novel performance degradation prediction method based on the deep belief network(DBN)and the extreme learning machine(ELM)is proposed.This method uses the median filtering algorithm to preprocess the degradation data of PEMFC effectively,uses DBN to mine the nonlinear degradation features in the preprocessed data deeply,and then uses ELM to adjust the model parameters and complete the construction of the DBNELM prediction model.Then,the particle swarm optimization algorithm is used to optimize the structural parameters of the DBN-ELM model.Finally,the degradation data of PEMFC under constant load testing and dynamic load testing are used for degradation prediction experiments.The prediction results show that the proposed method has the advantages of the DBN and ELM,and can provide high-precision prediction results and improve the generalization ability of the prediction model.(3)Aiming at the problem that traditional prediction methods cannot effectively quantify the uncertainty of PEMFC performance degradation prediction,which leads to low reliability of prediction results,the performance degradation prediction method of PEMFC based on singular spectrum analysis algorithm and deep Gaussian processes method is proposed.The proposed prediction method first uses the singular spectrum analysis algorithm to reconstruct the degradation features in the raw data to complete the data preprocessing.Then,the deep Gaussian processes method is used to construct the prediction model and the kernel function design of which is optimized according to the degradation characteristics to realize the accurate expression of the degradation law.Finally,two sets of degradation voltage datasets with a test time of up to 1000 h are used to verify the proposed prediction method.The prediction results show that the proposed method not only has high prediction accuracy but also can quantify the uncertainty of the prediction results by constructing high-quality confidence intervals,which improve the reliability of the prediction results.(4)Aiming at the problem of providing both high-precision prediction results and high-quality prediction intervals in practical engineering applications,the performance degradation prediction method of PEMFC based on the non-dominated sorting genetic(NSGA-II)algorithm and the DBN-ELM prediction model is proposed.This method firstly uses the DBN-ELM model and the upper and lower limit estimation method to construct the interval prediction model of PEMFC performance degradation and uses the interval median as the deterministic prediction result of the model.Then a multi-objective optimization framework is built to deal with the simultaneous optimization of prediction accuracy and interval quality.The parameters of the prediction model are optimized using the non-dominated sorting genetic algorithm.Finally,two sets of actual degradation data are used to evaluate the prediction method's performance comprehensively.The results show that this method can provide both high-precision prediction results and high-quality prediction intervals,thereby improving the accuracy and reliability of PEMFC performance degradation prediction.(5)Aiming at the problem that the existing methods do not fully consider the multivariable coupling and abnormal degraded data in the process of PEMFC performance degradation,which leads to low accuracy of prediction results,a multivariable performance degradation prediction method based on the random forest approach and the robust deep Gaussian processes is proposed.The proposed prediction method first uses the random forest approach to select the optimal feature variable set,providing more comprehensive and reliable degradation information for PEMFC performance degradation prediction.Then,the robust deep Gaussian processes model is used to mine the interrelationships of multiple variables to construct the performance degradation prediction model of PEMFC,and Student-t likelihood is used to deal with abnormal data to improve the robustness of the prediction model.Finally,two sets of PEMFC degradation testing data with 24-dimensional variables are used to verify the prediction performance of the proposed method.Experimental results demonstrate that the proposed degradation prediction method has high prediction accuracy and robustness and can provide uncertain information about the prediction results.