Degradation Prediction of Proton-Exchange Membrane Fuel Cell Tested Under Static and Quasi-Dynamic Operating Conditions

Abstract

The viability and extended lifetime of proton-exchange membrane (PEM) fuel cells are essential to enable hydrogen electric aviation. Their reliability, that is, the ability to maintain stable performance over long operational periods under varying environmental and load conditions, is a critical factor for their adoption in aviation systems. Prognostics and Health Management (PHM) is a popular strategy used extensively in aero engines. This data-driven approach is used to estimate the best maintenance times and to prevent utilisation of equipment beyond its end-of-life. Despite being considerably well established in classical aeronautics, PHM is not yet mature for PEM fuel cells. Currently PHM for PEM fuel cells face three technical problems: (1) reliance on multi-sensor datasets that limit applicability when only voltage and current measurements are available, (2) short-term prediction windows of typically below 100 h, and (3) limited linkage between physics models and machine learning models for RUL predictions. This study presents a hybrid long short-term memory (LSTM) and echo state network (ESN) framework to address gaps (1) and (2) under static and quasi-dynamic conditions. Using seasonal and trend decomposition using LOESS (STL), voltage and current are separated into components, enabling medium-term prediction up to 125 h. Existing membrane degradation models are utilised to establish a foundation for future linkage with a machine learning model to improve remaining useful life (RUL) predictions. Our results suggest that the hybrid LSTM–ESN framework enables 125-h degradation predictions using only voltage and current measurements.