Steady-State Model Enabled Dynamic PEMFC Performance Degradation Prediction via Recurrent Neural Network

Proton exchange membrane fuel cells (PEMFC), distinguished by rapid refueling capability and zero tailpipe emissions, have emerged as a transformative energy conversion technology for automotive applications. Nevertheless, their widespread commercialization remains constrained by technical limitations mainly in operational longevity. Precise prognostics of performance degradation could enable real-time optimization of operation, thereby extending service life. This investigation proposes a hybrid prognostic framework integrating steady-state modeling with dynamic condition. First, a refined semi-empirical steady-state model was developed. Model parameters’ identification was achieved using grey wolf optimizer. Subsequently, dynamic durability testing data underwent systematic preprocessing through a correlation-based screening protocol. The processed dataset, comprising model-calculated reference outputs under dynamic conditions synchronized with filtered operational parameters, served as inputs for a recurrent neural network (RNN). Comparative analysis of multiple RNN variants revealed that the hybrid methodology achieved superior prediction fidelity, demonstrating a root mean square error of 0.6228%. Notably, the integration of steady-state physics could reduce the RNN structural complexity while maintaining equivalent prediction accuracy. This model-informed data fusion approach establishes a novel paradigm for PEMFC lifetime assessment. The proposed methodology provides automakers with a computationally efficient framework for durability prediction and control optimization in vehicular fuel cell systems.