Physics informed deep learning for real time optimization of PEM fuel cells in electric vehicles

Proton Exchange Membrane (PEM) fuel cells are central to zero-emission transportation, but their power output is governed by coupled electrochemical and thermal dynamics difficult to predict for real-time control. Existing data-driven models either ignore the underlying electrochemistry or embed physics without validating against independently measured stack power, leaving open whether reported accuracy reflects genuine skill or recovery of a known equation. We address this gap with a physics-informed framework supplying first-principles electrochemical quantities (Nernst voltage, activation, ohmic, and concentration overpotentials) as input features to a Time-Series Mixer (TS-Mixer) network, trained and evaluated directly against the IEEE PHM 2014 (FC1 Ageing) dataset’s measured stack voltage and current. Operating conditions vary little over the test, so cumulative operating time, a proxy for degradation, is the dominant driver of real power ( $$R^2=0.91$$ alone) and is included as an input. The model achieves $$R^2=0.969$$ , RMSE $$=0.295$$ W, and MAPE $$=0.097\%$$ , with 5-fold cross-validation giving $$R^2=0.967\pm 0.001$$ and no train-test gap. To confirm this is genuine rather than leakage-driven, the TS-Mixer is benchmarked against LSTM, GRU, TCN, Transformer, SVR, and linear regression on identical data, achieving the best accuracy, while an ablation study quantifies the physics-derived features’ contribution beyond degradation. Model footprint and inference latency are reported for embedded feasibility, and SHAP, LIME, and permutation importance provide feature attributions. The underlying equations are also validated as a sizing tool, reproducing $$\sim$$ 60 kW for an illustrative 370-cell stack, establishing a framework validated against real measurement for fuel cell electric vehicles.