Establishment of Core Front-End Wind Power Forecasting Technology for Green Hydrogen: Development of Short-Term and Ultra-Short-Term Wind Power Generation Forecasting Models

Green hydrogen is one of the more promising ways to connect renewable power to the parts of decarbonization that electrons alone struggle to reach — heavy industry, long-duration storage, and hard-to-electrify transport. Wind is an attractive power source for the electrolyzers that split water into hydrogen, but wind is also random, seasonal, and volatile, and that volatility propagates straight into hydrogen's cost, reliability, and scalability. In the chain from wind power to green hydrogen, how well you can see the wind coming sets a ceiling on how efficiently the electrolyzer can run — which is why this project, commissioned by the Metal Industries Research & Development Centre (MIRDC) and led by Prof. I-Yun Lisa Hsieh, treats wind forecasting as the critical front-end capability.

Accurate wind-power forecasts do two jobs here. They translate directly into the power available to the electrolyzer, and therefore into how much hydrogen can be made and at what levelized cost (LCOH); and they are the basis on which an energy-management system schedules the electrolyzer's start-ups, shut-downs, and operating points — decisions whose losses and equipment wear depend on knowing the wind input in advance. But forecasting wind well is genuinely hard: wind fields carry strong small-scale turbulence and non-stationarity, terrain (roughness, obstacles, local eddies) makes them spatially uneven, and a turbine's own control — yaw, blade pitch, rotor inertia, power-curve switching — turns wind speed into power output through a highly nonlinear, time-lagged mapping.

The project answers with a two-tier forecasting architecture. A short-term layer (hourly-to-daily) builds a regional meteorological background field from ground weather stations and produces day-ahead wind-power forecasts, correcting turbine power curves for environmental effects to sharpen the wind-speed-to-power step. An ultra-short-term layer (minute-scale, roughly 15–30 minutes ahead) captures fast wind-field behavior — gusts and direction shifts — models the turbine's response delay, and updates continuously from high-frequency SCADA telemetry and live weather data. Throughout, the system carries uncertainty quantification — prediction intervals, probability estimates, and multi-model ensembles — so its outputs come with confidence indicators for risk-aware dispatch, and it converts forecast reserve power into a first estimate of the electrolyzer's real-time hydrogen-production cost under varying load.

Beyond the standalone forecasts, the work is designed to plug into a larger wind–hydrogen–storage picture: it validates against MIRDC's planned pilot site, hands its wind outputs to hydrogen-cost and energy-management modules, and lays groundwork for a coupled wind-to-hydrogen decision tool that could scale across multiple wind farms and sites. For the E3 Center, it extends the group's renewable-forecasting line — building on its earlier work on deep-learning wind-power prediction and the economics of offshore-wind hydrogen — toward the front end of Taiwan's emerging green-hydrogen supply chain.