Research On Hydrogen Fuel Cell Remaining Useful Life Prediction

With the intensification of energy shortages and environmental pollution,hydrogen fuel cells are widely used in transportation,distributed generation and portable power due to the advantages of high efficiency and environmental protection.However,the durability of fuel cell restricts its large-scale commercial development.Due to the complex degradation mechanism of the performance of fuel cell,it is difficult to predict the remaining life of the battery throu gh the mechanism model.Although the current short-term prediction based on experimental data has a high fittness for the prediction results of the relative measured values,it cannot realize the judgment of the future long-term state of the battery.In order to solve this problem,this paper uses deep learning and particle filtering to construct the long-term prediction frameworks for the remaining useful life of hydrogen fuel cells based on experimental data.The former corresponds to the data model,and the latter to the empirical and semi-empirical models,the long-term prediction frameworks respectively verify their effectiveness of long-term prediction,to realize the research of remaining useful life prediction of hydrogen fuel cell based on data.Since the data model only relies on the data itself,the selected experimental data is analyzed and processed to strengthen the model’s capture of the voltage decay trend.Based on the short-term prediction framework,a long short-term memory network model is used to train the data.Change the input method of data in the prediction phase to test the effect of long-term prediction.The long short-term memory network model has been debugged to improve the accuracy of long-term prediction result,but it is unable to perform long-term prediction of the voltage attenuation trend affected by random fluctuations.In order to realize the long-term prediction of battery degradation,the empirical voltage attenuation model is determined according to the experimental data,and the voltage attenuation state is tracked by particle filter to estimate the model parameters.Adding the voltage recovery effect caused by the start and stop of the stack,the empirical model is used to extrapolate the state of the future voltage to realize the long-term prediction of the remaining battery life.For the deviations in the prediction process,the iterative training method and the genetic algorithm are used to optimize the parameters of the empirical model.Through the comparison between the models and the test of multiple prediction starting points,it is verified that the particle filter model based on genetic algorithm optimization can achieve fast and accurate long-term prediction.In view of the limitation of the empirical model based on the voltage decay trend,there is a lack of inherent insight into battery degradation.In this regard,a semi-empirical model derived from the polarization equation is used,and the time-varying parameters of the model are inversely deduced according to the changes in the polarization curves measured in the experiment.The time-varying parameters and their derivatives are used to restore the attenuation trend of the voltage.The tracking of parameters by particle filter can achieve a more rapid and accurate estimation of the remaining useful life of the fuel cell.Based on the experimental data of hydrogen fuel cells,this paper tests the influence of data model and empirical and semi-empirical models on the remaining useful life prediction of the battery.Through the prediction frameworks of long short-term memory network and particle filtering,the research of long-term prediction methods is completed.The rapid and accurate prediction of the remaining useful life of fuel cell has practical engineering application significance and provides a new idea for fuel cell life prediction.