Answers are grounded in this paper and cite their sources.
Introduction Accurate proton exchange membrane fuel cell (PEMFC) prognosis requires evaluation beyond low errors on familiar operating records. We assess the reliability, robustness, and transferability of a previously explored random‐forest recursive feature elimination–temporal convolutional network (RF‐RFE‐TCN) pipeline for voltage degradation and remaining useful life (RUL) prediction. Methods Particle swarm optimization tunes the network architecture. Feature selection and standardization are fitted to training data only, with a fixed 24‐step input history and chronological evaluation. Voltage forecasts are mapped to threshold‐based RUL, while Monte Carlo dropout provides prediction intervals. Synthetic perturbation and cross‐regime transfer tests characterize sensitivity to changing conditions. Result The test mean absolute errors (MAEs) are 0.0452 V (FC1) and 0.0478 V (FC2). Under the reported synthetic perturbations, R 2 ranges from 0.91 to 0.95 and MAE from 0.05 to 0.07 V. Relative to the 0.045 V FC1 reference error, direct high‐to‐low‐load and FC1‐to‐FC2 transfer increase MAE to 0.068 and 0.082 V, respectively. Discussion These findings support leakage-controlled and uncertainty‐aware evaluation while highlighting the limits of direct transfer. The predictor uses physically motivated inputs but lacks explicit electrochemical constraints; independently varied thermal, humidity, stoichiometric, and load‐cycling regimes require further validation.
No takes yet. Share an insight, caveat, or question.
Ma et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: