Prediction of fuel cell performance degradation using a combined approach of machine learning and impedance spectroscopy

Accurate prediction of performance degradation in complex systems such as solid oxide fuel cells is crucial for expediting technological advancements. However, significant challenges still persist due to limited comprehension of degradation mechanisms and difficulties in acquiring in-situ features. In this study, we propose an effective approach that integrates long short-term memory(LSTM) neural network and dynamic electrochemical impedance spectroscopy(DEIS). This integrated approach enables precise prediction of future evolutions in both current-voltage and EIS features using historical testing data,without prior knowledge of degradation mechanisms. For short-term predictions spanning hundreds of hours, our approach achieves a prediction accuracy exceeding 0.99, showcasing promising prospects for diagnostic applications. Additionally, for long-term predictions spanning thousands of hours, we quantitatively determine the significance of each degradation mechanism, which is crucial for enhancing cell durability. Moreover, our proposed approach demonstrates satisfactory predictive ability in both time and frequency domains, offering the potential to reduce EIS testing time by more than half.