Joint prediction of polarization losses and internal states in fuel cell via time–frequency feature fusion and machine learning

Abstract

The real-time decoupling of polarization losses and internal states is fundamental for extending the lifespan of proton exchange membrane fuel cells (PEMFCs), yet existing methods struggle with the trade-off between measurement speed and information depth. This study proposes a novel synergistic time–frequency fusion framework for the joint prediction of polarization losses and internal state distributions. By leveraging a two-dimensional multi-scale agglomerate model, we construct a high-fidelity dataset that captures the intricate mapping between frequency-domain signatures and microscopic reaction distributions. A comprehensive sensitivity analysis identifies impedance amplitude and phase angle at 79.43 Hz and 10 Hz as optimal features, capturing critical information about reaction interfaces and mass transport that are often neglected in traditional time-domain analysis. These identified features, integrated with macro-level operating conditions, are fed into a Gaussian Process Regression (GPR) model. Results demonstrate superior predictive accuracy with a Mean Absolute Percentage Error (MAPE) below 4% for all key variables. Furthermore, the model exhibits exceptional robustness under 30 dB noise levels and dynamic New European Driving Cycle (NEDC) conditions, successfully tracking transient concentration fluctuations. This work offers a highly efficient and cost-effective approach for online health management by extracting physical insight from less on-board measurement information.