Multi-scale degradation forecasting of PEMFCs under non-stationary operating conditions: A novel spatio-temporal deep learning framework

Predicting the degradation is crucial for effective fuel cell health management, which directly impacts the reliability and cost-efficiency of proton exchange membrane fuel cells (PEMFC). While deep learning have achieved success in PEMFC degradation prediction, existing methods face challenges in modeling non-stationary, spatial correlations and long-term dependencies of multi-sensor data. To address these challenges, we propose SANformer, a novel Transformer-based spatio-temporal modeling framework. Specifically, SANformer consist of a Transformer based on spatial attention and a seasonal-trend decomposition module, which work in parallel to jointly capture complex spatial and temporal dependencies in degradation data. Furthermore, we introduce an adaptive non-stationary reconstruction attention (ANRA) mechanism that can dynamically reconstruct attention maps by identifying non-stationary factors from multi-sensor time series data. Experimental results on both static and quasi-dynamic datasets demonstrate that SANformer significantly outperforms existing state-of-the-art methods in prediction accuracy and maintains robust performance even under limited training data conditions.