This is the authors' abstract. We don't add key points for this paper.
Green hydrogen has become a pivotal clean energy carrier essential for achieving a net-zero economy. It holds immense potential to decarbonize industries and transportation by harnessing renewable solar energy, yet its production from solar faces significant challenges due to inaccurate irradiance forecasting under dynamic weather conditions and inefficient linkage to electrolysis processes. In this study, we propose the HyUNet model, a novel deep learning framework that utilizes ground-based sky image datasets from Stanford (USA) to enhance ultra-short-term photovoltaic (PV) output forecasting, overcoming the limitations of existing methods like physical, statistical, and hybrid models that neglect real-time visual data and multi-seasonal variability. By capturing detailed cloud positions and movements, HyUNet improves hydrogen production estimation by coupling accurate PV forecasts with a deterministic photovoltaic-to‑hydrogen conversion model of the PEM electrolyzer within photovoltaic-electrochemical (PV-EC) systems, achieving RMSE values of 1.058 at 5-min intervals with Skill Scores of 0.638, 0.660, and 0.444 for 5-min, 10-min, and 15-min horizons, respectively. The model shows improvement in RMSE by 63.8% at 5-min, 49.0% at 10-min, and 44.4% at 15-min compared to benchmark models. In terms of hydrogen production, the model reduces the standard deviation of hourly hydrogen yield deviations during peak hours (10:00–16:00) by 36.8–67.9% relative to benchmark models at the 5-min horizon, corresponding to at least 20% lower day-to-day variability in peak-hour hydrogen yield, with consistent hydrogen production levels across the multi-seasonal evaluation period. Weibull probability analysis and Williams plot assessments confirm the model's reliability, with production peaks aligning with solar irradiance and demonstrating improved operational efficiency. While practical PEM electrolyzers are subject to minimum load and ramp rate limitations, this study assumes sufficient rated capacity to absorb available PV power, focusing on the influence of forecasting accuracy on hydrogen production estimation. While the validation is based on a single-site dataset spanning multiple seasons and weather conditions, these findings position HyUNet as a promising framework for optimizing green hydrogen generation, whose generalizability across climates and geographies remains to be confirmed through multi-site validation.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Jonathan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: