Dynamic PEM Hydrogen Prediction Across Operating Domains: Cross-Campaign Transferability and Applicability-Aware Modeling

Reliable time-resolved prediction of hydrogen production is essential for proton exchange membrane (PEM) water electrolysis (PEMWE) under variable operation, yet predictive models are often assessed near the experimental domain used for development. This study introduces a progressive cross-campaign validation framework for dynamic PEM hydrogen prediction across increasingly different operating conditions. A high-resolution experimental dataset from a 100 kW PEM electrolyzer was used to compare static, temporal, physics-structured, and hybrid models under hierarchical sensor representations. Validation combined leave-one-experiment-out testing, transition-resolved analysis, repeated random seeds, an independent PHOEBUS cross-profile stress test, explicit high-current out-of-distribution (OOD) evaluation, and locked transfer to temporally separated 2018 campaigns without model retuning. The proposed physics-structured backbone (PSB) with guarded residual long short-term memory (LSTM), denoted PSB-ResLSTM-G, achieved a mean root mean square error (RMSE) of 0.156 m³ h⁻¹ versus 0.172 m³ h⁻¹ for the Direct LSTM and lower overall RMSE in all 20 matched fold-seed comparisons. Within the PHOEBUS high-current OOD subset, PSB-ResLSTM-G reduced RMSE from 1.559 to 0.237 m³ h⁻¹ (84.8%), with the applicability guard reverting the hybrid prediction to the PSB beyond the development-current support. In the locked 2018 confirmation, PSB achieved the lowest RMSE at 0.547 m³ h⁻¹, followed by PSB-ResLSTM-G at 1.344 m³ h⁻¹ and Direct LSTM at 3.257 m³ h⁻¹, revealing an accuracy–transferability trade-off under the larger cross-era shift. Overall, the proposed applicability-aware hybrid formulation and progressive validation framework jointly characterize dynamic accuracy, OOD behavior, and cross-era transfer within a 100 kW PEM electrolyzer platform.