Accurate short-term forecasts of green-hydrogen output are vital for electrolyzer scheduling, storage management, and grid coordination. We develop a one-hour-ahead benchmark centered on tree-ensemble methods and report accuracy alongside compute energy, carbon emissions, and wall-clock time to enable carbon-aware model selection. Using a multi-year hourly dataset tied to a 1 MW PV-to-electrolyzer conversion chain, we construct meteorology-informed features plus lagged hydrogen output and perform strictly chronological splits to avoid leakage. Across methods, generalization is strong: Random Forest attains MAE 0.67 kg/h, RMSE 1.18 kg/h, and R2 of 0.95 while consuming 1.76 Wh with 1.13 g carbon emissions during training. Gradient-boosting variants achieve comparable accuracy but require more energy and emit more carbon, indicating nontrivial operational trade-offs. These results show that tree ensembles can provide robust one-hour-ahead performance at low computational cost. By pairing predictive metrics with energy, carbon, and time, the benchmark offers a practical basis for selecting hydrogen forecasters under real-world constraints.