Cross-Renewable Generalization of PEM Electrolyzer Hydrogen-Flow Forecasting: Multi-Horizon Evaluation and Physics Anchoring

Forecasting renewable-powered proton-exchange-membrane (PEM) electrolysis across unfamiliar power trajectories requires evaluation at the renewable-source level rather than within familiar profile families. We evaluated experimentally recorded hydrogen flow from one nominal 1.25 MW PEM-electrolyzer platform under four renewable-source operating domains—Simulated Wind, Simulated Wave, Historical Solar Photovoltaic, and Simulated Tidal—using leave-one-renewable-source-out validation, a 60 s process history, and direct 1, 5, 10, and 30 s horizons. Ridge, HistGradientBoosting, and LightGBM transferred usefully at short horizons but degraded as the horizon increased; Wave had negative R2 for all three models at 30 s. Source balancing and worst-source selection produced only local gains, and model-disagreement error awareness weakened with horizon. For one prespecified five-epoch Conv1d masked-reconstruction design and linear probe, self-supervised embeddings underperformed handcrafted descriptors in source-macro forecasting, without establishing a general limitation of self-supervision. A current-command-based operational gray-box anchor (F1) outperformed its residual-augmented version (G1) at all four horizons; G1 improved on the prespecified Ridge comparator but was less accurate than F1. Equal-weight source-macro errors describe these four held-out operating domains, not a population of systems or sites. The findings do not establish transfer across manufacturers, stacks, capacities, or sites, and no electrochemical parameters were uniquely identified.