Spatiotemporal sensitivity analysis of fuel cell performance under dynamic cycling using a sliding-window-integrated hybrid Sobol-Shapley method

The spatiotemporal evolution of parameter sensitivity in proton exchange membrane fuel cells (PEMFCs) under dynamic loads exhibits marked non-stationary dynamic characteristics, which traditional static methods fail to resolve due to their inability to capture transient coupling and load-dependent interactions, thereby constraining real-time control adaptability. This study proposes a hybrid Sobol-Shapley (HSS) framework integrating sliding windows (50-s window, 10-s step) and conflict coefficient weighting, validated through 1000-h experiments (25 cm2 seven-layer stack, 16 parameters), to systematically quantify time-varying contributions of mechanical, thermodynamic, and fluid parameters to voltage degradation. The hybrid random forest (RF)-linear regression ensemble achieves 0.94 R2 prediction accuracy with 32.6% computational efficiency improvement over pure RF, while grid search-optimized XGBoost regressors enhance SHAP interpretability. Key findings include: (1) Current density dominates voltage response (HSS = 0.52) with a 46% sensitivity surge during load transients, necessitating hierarchical voltage regulation for overshoot suppression; (2) Spatial homogeneity in anode/cathode pressure and flow sensitivities contrasts with significant temperature sensitivity divergence between electrodes; (3) The system-level response delay for thermal parameters is 42.2 s, and for fluid parameters below 15 s, inducing temporal decoupling. Anode HSS sensitivity exceeds cathode values, while hydrogen purging prolongs anode parameter delays; (4) Cathode flow sensitivity increases by 105% post-900 h aging, with a 23% interaction intensification with humidity/temperature parameters, demanding adaptive water-thermal co-optimization over rigid supply strategies. This framework provides methodological innovation for PEMFC transient dynamics prediction and durability-oriented control.