Prediction of PEMFC Life decay based on Machine Learning

Proton exchange membrane fuel cell (PEMFC) has the advantages of fast cold start and zero pollution, however, the safety stability, reliability and durability of mainstream PEMFC restrict its large-scale industrial application and commercialization. In this paper, we propose a method combining wavelet threshold denoising algorithm and machine learning technology to address the problems of low model prediction accuracy due to the high noise of the original PEMFC data set and the unknown influence of PEMFC feature quantity on its lifetime prediction. Firstly, the noise interference in the original data is removed and the PEMFC attenuation information is retained by equal interval sampling and wavelet threshold denoising algorithm, and then the XGBoost-RFECV algorithm is used to calculate the importance of different features of PEMFC and filter out the optimal feature subset, and at the same time, a two-layer bi-directional long and short-term memory (BiLSTM) based neural network is built, and the filtered optimal feature subset is input into the model is trained to achieve life span prediction. The experimental results show that the average absolute error and root mean square error of this method are 0.0013 and 0.0022, respectively, and the coefficient of determination R2 is 0.980, which is more accurate than XGBoost-RNN, XGBoost-LSTM and XGBoost-RFECV-LSTM methods, and can effectively improve the prediction of PEMFC lifetime.