An improved iTransformer framework for health prediction of proton exchange membrane fuel cells

Accurate prediction of the state of health (SOH) and remaining useful life (RUL) in proton exchange membrane fuel cells (PEMFCs) is critical for optimizing maintenance strategies. However, conventional data-driven methods often struggle with nonlinear degradation dynamics, non-stationary behaviors, and limited interpretability. To overcome these challenges, this study proposes an improved iTransformer framework that integrates with a lightweight temporal convolutional network. It incorporates neural ordinary differential equations embedding to model continuous-time degradation, Kolmogorov-Arnold attention for feature interaction, and quantum-based extended long short-term memory for long-term sequence modelling. Interpretability is achieved via deep learning-based feature importance Shapley additive explanations, while conformal prediction provides robust uncertainty quantification. Validated on the IEEE 2014 PHM Challenge datasets (FC1 and FC2), the model demonstrated superior SOH prediction performance, achieving R 2 values up to 0.9956 (FC1) and 0.9930 (FC2). For RUL estimation, the model achieved exceptional accuracy with an absolute error of 0.48 and a relative error of 0.23% for FC2, while maintaining well-calibrated prediction intervals aligned with the nominal 90% confidence level. Comparative analysis confirms that the proposed framework outperforms state-of-the-art methods, demonstrating its effectiveness, robustness, and practical applicability for PEMFC prognostics and maintenance decision-making.