摘要
针对质子交换膜燃料电池(PEMFC)的剩余使用寿命(RUL)预测问题,提出一种基于老化特征的混合预测方法。PEMFC整个生命周期内的老化特征可以分解为整体老化特征、电压波动特征和电压恢复特征。针对整体老化特征,首先利用平方根无迹卡尔曼滤波(SRUKF)进行预测;针对恢复电压特征,使用二阶RQ-RLC等效电路模型拟合电化学阻抗谱数据,获得恢复电压和时间的间接联系;针对电压波动特征,使用Xception-长短期记忆(LSTM)神经网络进行训练和预测。最后将三种老化特征的预测结果相叠加并计算PEMFC的RUL预测结果。基于两组工况数据验证,结果表明,相比SRUKF和Xception-LSTM独立预测,混合预测方法的精度提高了约20%。
For the life prediction issue of Proton Exchange Membrane Fuel Cells(PEMFC),a hybrid prediction method based on aging characteristics was proposed.The aging characteristics of PEMFC throughout its lifecycle could be decomposed into overall aging characteristics,voltage fluctuation characteristics,and voltage recovery characteristics.For the overall aging characteristics,the Square Root Unscented Kalman Filter(SRUKF)was utilized for prediction;for the voltage recovery characteristics,a second-order RQ-RLC equivalent circuit model was used to fit the electrochemical impedance spectroscopy data,obtaining an indirect relationship between recovery voltage and time;for the voltage fluctuation characteristics,the Xception-Long Short Term Memory(LSTM)network was employed for training and prediction.Finally,the prediction results of the three aging characteristics were superimposed to calculate the RUL prediction result for PEMFC.Based on the verification of two sets of operating condition data,the results show that compared to independent predictions by SRUKF and Xception-LSTM,the accuracy of the hybrid prediction method is improved by about 20%.Certainly,here is the translation with the sequence numbers replaced by ordinal terms:Firstly,the collection and preprocessing of voltage aging data.Preprocess the collected PEMFC data,which includes removing monitoring noise,outlier processing,and timestamp reconstruction;Secondly,voltage aging feature prediction.Utilize the PEMFC empirical model to represent the degradation trend and apply SRUKF to identify the hidden state of the voltage.The output of this module is the prediction of future voltage aging trends under the overall trend;Thirdly,voltage fluctuation feature prediction.This module is designed to analyze the periodic fluctuations of PEMFC voltage during long-term durability operation.The Xception-LSTM model is extensively trained with voltage fluctuation feature datasets for long-term prediction of these characteristics.Fourthly,EIS test data collection.Gather EIS experimental data of the stack at specific time points during long-term durability experiments at regular intervals.Fifthly,recovery voltage feature prediction.Based on EIS data,identify the parameters of the second-order RQ-RLC model and further establish the relationship between resistance parameters and recovery voltage.Lastly,overlay the results of the three feature predictions to obtain the long-term prediction of the aging voltage characteristic values,and calculate the RUL prediction value according to the aging voltage threshold.This paper proposes a hybrid prediction method based on aging characteristics,which decomposes the aging characteristics into overall aging features,voltage fluctuation features,and voltage recovery features,and carries out long-term predictions for the three types of features,finally overlaying the prediction results.The method was validated using data from two sets of operating conditions,FC1 and FC2,and the results indicate:(1)The short-term prediction results show that SRUKF has a smaller prediction error compared to UKF,and after incorporating the Xception network for feature extraction,the Xception-LSTM model exhibits better prediction performance;(2)Under both operating conditions,the hybrid prediction method has a high RUL prediction accuracy.For FC1,most prediction starting points can achieve an accuracy of over 95%,while for FC2,the accuracy can reach over 90%;(3)A comparison among the SRUKF,Xception-LSTM,and hybrid prediction methods shows that,compared to predicting a single aging feature,the RUL prediction accuracy is higher when combining multiple aging features.
作者
徐冰
张敏
杨文涛
张领先
Xu Bing;Zhang Min;Yang Wentao;Zhang Lingxian(Beijing Sifang Automation Company Limited,Beijing,100085,China)
出处
《电工技术学报》
北大核心
2025年第17期5703-5714,共12页
Transactions of China Electrotechnical Society
基金
国家重点研发计划资助项目(2021YFB2400700)。
