Degradation prediction of proton exchange membrane fuel cell based on the multi-inputs Bi-directional long short-term memory

Abstract

The remaining useful life (RUL) is one of the most crucial indicators for proton exchange membrane fuel cell (PEMFC). The existing data-driven prediction algorithms have the problem of low prediction accuracy with finite training data. This paper proposes a multi-input single-output Bi-directional long short-term memory (MISO-BiLSTM) to improve the prediction accuracy. First, a bi-exponential empirical model is developed to reduce the voltage data error caused by stack test start/stop. Second, a Pearson approach is employed to extract RUL-related indicators as input to MISO-BiLSTM to decrease prediction error. Finally, the BiLSTM prediction models are developed for each input separately to improve the prediction performance under limited data. The MISO-BiLSTM prediction method is experimentally validated using experimental data from PEMFC under static and dynamic operating conditions. The results demonstrate that the root mean square error (RMSE) of the MISO-BiLSTM prediction algorithm proposed in this paper are 0.00282 and 0.00386 under 300h train data. The proposed method achieves high accuracy prediction with limited data and has important significance for the health management of PEMFC.