Prediction Of Voltage Degradation Trend For Vehicular Proton Exchange Membrane Fuel Cell

Proton exchange membrane fuel cell（PEMFC）is an energy conversion device that generates electricity from chemical energy.Due to the advantage of high efficiency,zeroemission,low working temperature,and quick launch ability,it has drawn a great deal of attention in new energy vehicle fields.During the process of the PEMFC commercialization,improving durability becomes one of the most important issues.Prognostics and Health Management（PHM）is among the most effective and significant technologies in lifetime extension as well as maintenance cost reductions.The estimation of remaining useful life（RUL）is obtained by the prediction of the degradation trend in this technology,and then the maintenance strategies are formulated.This thesis deals with the prediction issue of the performance degradation trend for PEMFCs equipped in a city bus.In general,the average output voltage of PEMFCs is regraded as the performance index.In the practical operation of the PEMFC city bus,the conditions of the road and environment are actually time-varying,resulting in the variation on the output power and current of PEMFCs.Then a multi-parameter voltage model is constructed to describe the output voltage under different working conditions.Subsequently,aging parameters are selected from the voltage model by resorting to analyzing the degradation mechanism.In addition,the initial values of aging parameters are obtained by the harmony search algorithm（HS）.Thus,the prediction of voltage is essentially the prognosis of aging parameters.Here,two prediction methods are proposed,and the mean absolute error（MAE）of predicted voltage is used as the index of precision for the comparative analysis of the results.As for dataset I,the voltage prediction method based on PF and regression model is proposed.Aging parameters are considered as system states.PF is primarily used for states estimation,and then the Bayesian ridge regression（BRR）model and the Gaussian progression regression（GPR）model are respectively utilized to establish the relationship between the aging parameters and operating time.The aging parameters in predicting stage are obtained by the trained regression models.The BRR model shows better performance than the GPR model while the MAE is more than 8.5 m V.As for dataset II,the prognosis approach is developed by the advantage actor-critic（A2C）algorithm.The actor neural network is utilized to describe the law of parameters aging along with the change of operation time,and the critic neural network is applied to evaluate aging parameters.Then,the environment model is built by voltage estimation error.The training of neural networks is conducted by the interaction between actor-critic neural network and environment model.Accordingly,the parameters are updated until the completion of training.Compared with precision with the former method,the highest MAE is 9.5m V in the same range of working current.The former is slightly superior to the latter in accuracy.Nevertheless,this method is only suitable for the dataset which involved specific loading processes.In order to improve computing speed,the second method simplifies the voltage model.Besides,the restriction on the dataset is less than the former.Thus,the first method is more suitable for theory analysis,and the second one is more potential for online prediction of voltage in particle engineering.