<p>Accurately predicting long-term degradation patterns in proton exchange membrane fuel cell (PEMFC) stacks under automotive operating conditions remains challenging. Prediction methods are largely constrained by laboratory-scale experiments and limited stack sizes, resulting in insufficient accuracy and generalization capability. To address these limitations, in this paper we propose a multi-scale bidirectional fusion network (MBFNet) tailored for an industrial 215-channel PEMFC stack, enabling accurate degradation prediction under accelerated real-world dynamic conditions using gas-heat-electricity (GHE) co-simulation data. A channel-joint adaptive noise correlation threshold (NCT) algorithm is introduced to account for variable correlations across sensors and operating conditions without relying on prior physical modeling. A multi-scale decomposition module captures degradation dynamics at different temporal scales, while a bidirectional fusion module integrates global trends and local details into the final prediction. Experimental results show that MBFNet achieves 18.6% lower prediction error and 36.8% fewer parameters than the long short-term memory (LSTM)-attention benchmark under real operating scenarios. In multi-step prediction tasks, MBFNet reduces root mean square error by an average of 24.5% relative to LSTM-attention and 55.2% relative to a one-dimensional convolutional neural network (1D-CNN) across four prediction horizons, better satisfying automotive application requirements. Moreover, MBFNet exhibits strong physical interpretability, making it efficient to implement and promising for practical deployment.</p>