Degradation prediction of proton exchange membrane fuel cell based on Bi-LSTM-GRU and ESN fusion prognostic framework. (12th September 2022)

• Record Type: Journal Article
• Authors: Li, Songyang; Luan, Weiling; Wang, Chang; Chen, Ying; Zhuang, Zixian
• Is Part Of: International journal of hydrogen energy. Volume 47:Number 78(2022)
• Page Start: 33466
• Page End: 33478
• Publication Date: 2022-09-12
• Subjects: PEMFC -- Prognostic -- Remaining useful life -- Bi-LSTM-GRU -- Deep learning; Hydrogen as fuel -- Periodicals; Hydrogène (Combustible) -- Périodiques; Hydrogen as fuel; Periodicals; 665.81
• DOI: 10.1016/j.ijhydene.2022.07.230
• Languages: English
• ISSNs: 0360-3199

Abstract

Abstract: The durability of proton exchange membrane fuel cell (PEMFC) is one of the technical challenges restricting its commercial applications. To enhance the reliability and durability of PEMFC, a fusion prognostic framework is proposed based on bi-direction long short-term memory (Bi-LSTM), bi-direction gated recurrent unit (Bi-GRU) and echo state network (ESN), which can achieve short-term degradation prediction and remaining useful life (RUL) estimation of PEMFC with fewer training datasets. For short-term prediction, using the first 200 h of voltage degradation data for training can achieve an acceptable and accurate prediction, with the root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (R 2 ) of 0.0235, 0.0195 and 0.9822, respectively. Compared with traditional machine learning methods, the proposed fusion prognostic framework shows a better predictive performance. In addition, a 100-step-sliding-windows method based on the fusion prognostic framework was implemented for RUL estimation. The results show that the percentage error ( E r ) is only 1.22% with the first 200 h of training data. The proposed method has great significance for online testing and health management of PEMFC. Highlights: Bi-LSTM and Bi-GRU are deployed to extract degradation characteristics of PEMFC. The introduction of ESN reduces the risk of overfitting. Both the short-term degradation prediction and RUL estimation can be achieved. Only the first fifth of the dataset is used as the training dataset. The prediction accuracy is improved compared with the traditional model.