Remaining Useful Life Prediction Based on LSTM with Peephole for PEMFC

Nowadays, proton exchange membrane fuel cells (PEMFC) have attracted more and more attention. However, its large-scale commercial development is limited by its short service life. With Prognostics and Health Management (PHM), the operating state of the fuel cell can be tested and the future state can be predicted to improve the service life of the fuel cell. As an important part of PHM, more and more attention is paid to the prediction of the remaining useful life (RUL) of PEMFC. RNN and LSTM networks are the most common method to predict RUL. In this paper, the LSTM model with peephole is proposed to predict the remaining service life of PEMFC. After being smoothed with LOWESS, the test data of 1154-hour steady operation are used to compare the proposed model with some existing model. The results show that the absolute average error (MAE) and root mean square error (RMSE) predicted by this method are 0.007 and 0.1558, respectively, which are better than the RNN and the common LSTM network. At the same time, the decay behavior of the fuel cell in the next 20 hours is predicted by the Recursive Multi-step Forecast, and the results show that the LSTM with the peephole is more consistent. Therefore, the LSTM with peephole is more accurate in predicting RUL.