Accuracy improvement of fuel cell prognostics based on voltage prediction

Proton exchange membrane fuel cell (PEMFC) is a promising hydrogen technique with various application prospects. However, all the PEMFCs are subject to degradation resulting from mechanical and chemical aging. To tackle this challenge, accurately predicting fuel cell degradation is essential for its durability optimization. In this study, an enhanced data-driven prognostic framework is developed to accurately predict short-term and medium-term degradation using only fuel cell voltage as the input feature. Firstly, a local outlier factor (LOF) algorithm is adopted for automatic detection of outliers in raw data collected from actual sensing environments. Then, an advanced deep learning model, residual–CNN–LSTM-random attention, is proposed to optimize voltage prediction to better indicate future PEMFC degradation trend. The proposed work is validated by the IEEE PHM 2014 Data Challenge. Compared to state-of-the-art methods, the proposed framework provides superior prediction accuracies with high stability. For instance, the framework improves short-term prediction, achieving a root mean square error (RMSE) of 0.0021 and a mean absolute percentage error (MAPE) of 0.0323 at steady state when training stops at 600 h. For medium-term prediction, our method also attains better results with an RMSE of 0.0085 and a MAPE of 0.4237 under same working conditions. Additionally, the comparative analyses demonstrate a lower computational burden and higher suitability of proposed work for practical applications.