Physics Informed Deep Learning for Real Time Optimization of PEM Fuel Cells in Electric Vehicles

Proton Exchange Membrane (PEM) fuel cells are pivotal for zero-emission trans portation, yet their dynamic power output is challenging to predict due to complex elec trochemical and thermal interactions. To address this, we develop a novel physics in formed deep learning framework that integrates a first-principles electrochemical model with a Time-Series Mixer (TS-Mixer) network to forecast real time PEM power stack performance. Using engineered inputs temperature, pressure, and current density to capture key dynamics, the model is trained on the IEEE PHM 2014 dataset (steady state data from a prognostic challenge) and augmented with synthetic transient load profiles. It achieves exceptional accuracy: R ² of 0.998, MAE of 0.347 W, and RMSE of 0.571 W. Interpretability is ensured via SHAP and LIME analyses, while overfit ting is mitigated through residual bias checks, early stopping, dropout, and adaptive learning-rate scheduling critical for reliable fuel cell modeling. This scalable framework supports digital-twin integration and advanced control, enhancing the efficiency and sustainability of fuel cell electric vehicles.