Study On Methods Of Health And Remaining Service Life Of Proton Exchange Membrane Fuel Cell

Proton exchange membrane fuel cell(PEMFC)is considered to have broad application prospects because of its high energy density,environmental protection and low noise.With the advancement of commercial application,the demand for PEMFC health management is also increasing.Accurate prediction of PEMFC health state(SOH)and remaining useful life(RUL)has been proved to be the research hotspot and difficulty to guide equipment fault maintenance and optimize management strategy.Therefore,based on the three degradation indicators of voltage,power and internal resistance,this study explores the impact of single PEMFC on the prediction ability of the model under various prediction types,the improvement strategy of the relevant PEMFC group under long-term prediction,and the generalization ability of the prediction model under simulated missing data.The full text is summarized as follows:Firstly,in order to monitor the continuous change of PEMFC state and predict the state value in the next step,SOH single-step prediction is carried out based on voltage,power and internal resistance.The results show that using voltage and power as degradation indicators,the prediction effect of kernel ridge regression(KRR)model is the best under different sampling interval,and the RMSE of its prediction result is at least 2.6% higher than that of other models;Convolutional neural network(CNN)has the best prediction effect at different time points,and the RMSE of its prediction result is at least 17.7% more than that of other models.The internal resistance of the calculation method is greatly affected by the prediction result of the internal resistance of the calculation method,which is greatly affected by the internal resistance of the calculation method.Secondly,in order to study the feasibility of multi-step prediction of PEMFC under single indicator,multiple models are used to predict under different data sampling interval.The results show that in the medium and short-term prediction,the better the prediction effect of CNN model with the increase of prediction step size;The prediction effect of all models in long-term prediction is poor.Based on the data of multiple parameters(temperature,pressure,humidity,etc.)measured by PEMFC in the degradation process,the regression model between multiple parameters and voltage is established to explore the correlation between multiple parameters and voltage.The results show that the multi parameter regression prediction is obviously affected by fluctuating conditions,and there is no strong functional correlation between multi parameter and voltage.In order to solve the problem of poor accuracy of single indicator long-term prediction,a two index(voltage,internal resistance)and three indicators(voltage,power,internal resistance)model for long-term prediction is proposed by combining multiple indicators.The results show that the SOH prediction accuracy of multi indicator long-term prediction is higher than that of single indicator longterm prediction,and the RUL prediction convergence effect of three indicators model is better than that of two indicators model.Then,aiming at the problem of poor stability of single PEMFC in long-term prediction,a combined vertical and horizontal prediction framework is proposed,and the long-term prediction of two related PEMFC groups is explored.The results show that the fitting degree between the long-term prediction results and the measured data under the new framework is higher,and the long-term prediction results are better than those of single indicator and multi indicator.Finally,in order to verify the generalization of the model,the medium and short-term prediction effects of multiple models under four PEMFC degradation data with huge differences,unknown and missing information are explored.In order to improve the prediction accuracy,the attention mechanism is used in the seq2 seq model to strengthen the correlation between the input data.The results show that the prediction effect of seq2seq＿attention model is better in two of the data set,and the prediction effect of the other two data set is not outstanding,but as a neural network model,it has more optimization space.