A novel framework combining optimized data decomposition and multimodal hybrid neural network Res-CGA for fuel cell performance degradation prognostics

Proton exchange membrane fuel cell (PEMFC) degradation involves complex interactions among multiple aging mechanisms, presenting challenges for accurate voltage prediction due to nonlinear dynamics and stochastic features. This study proposes a framework combining optimized data decomposition and multimodal hybrid neural network Res-CGA for PEMFC voltage degradation prognostics. It develops an optimized decomposition technique integrating the slime mold algorithm with variational mode decomposition to iteratively decompose aging voltage data into time-varying intrinsic mode functions (IMFs). Subsequently, entropy-driven k-means clustering consolidates IMFs into three cooperative IMFs: high-frequency, low-frequency, and trend sequences. The dedicated hybrid neural network Res-CGA (incorporating residual connections, convolutional operations, gating mechanisms, and attention mechanisms) is developed to learn decomposed features. The proposed method demonstrates superior capability in resolving multi-factor superimposed aging features and strong nonstationary. It effectively extracts trend features, long-period patterns, and high time-varying characteristics from aging voltage data, thereby achieving high-precision predictions. Validated across four independent PEMFC aging datasets, the framework achieves satisfactory accuracy with normalized root mean square error (NRMSE) < 0.05 and normalized mean absolute percentage error (NMAPE) < 10 %, outperforming benchmark machine learning and deep learning models by >3.94 % NRMSE and >1.68 % NMAPE reductions. Ablation studies confirm the critical contributions of the optimized decomposition module, the Res-CGA model, and the high-frequency sequence re-decomposition in the proposed method. These modules improve precision of the proposed method, achieving >0.41 % NRMSE and >0.20 % NMAPE reductions. This work establishes a novel technical pathway for PEMFC performance degradation prediction. Future research will focus on integrating this framework into prognostic and health management systems to validate its engineering applicability.