High-Resolution Rapid Refresh (HRRR)

The HRRR executes a complete data assimilation and forecast cycle every hour, producing a new 18-hour to 48-hour forecast. This latency of approximately 90 minutes from observation time to product delivery ensures that the latest radar, satellite, and surface observations are ingested to correct for rapidly evolving mesoscale phenomena like thunderstorm initiation, which directly impacts irradiance ramp rate predictions.

The model operates on a 3-kilometer grid, resolving atmospheric convection explicitly without the need for deep convective parameterization schemes. This convection-allowing resolution is critical for accurately simulating cloud motion vectors and cellular cloud structures that cause sharp spatial gradients in Global Horizontal Irradiance (GHI) across a solar farm, a detail lost in coarser global models.

HRRR is distinguished by its direct assimilation of WSR-88D radar reflectivity and radial velocity data via the Gridpoint Statistical Interpolation (GSI) system. This process initializes hydrometeor fields (rain, snow, graupel) within the model, drastically improving the short-term prediction of cloud cover and precipitation. For energy traders, this translates to superior 0-6 hour solar forecasts compared to models relying solely on satellite radiances.

The operational HRRR-Smoke configuration integrates a biomass burning emission model that simulates the transport and radiative effects of smoke plumes. This is vital for Direct Normal Irradiance (DNI) prediction during wildfire season, as elevated aerosol optical depth can attenuate solar radiation by 10-30% in regions far from the fire source, an effect not captured by standard clear-sky models.

The model employs the Mellor-Yamada-Nakanishi-Niino (MYNN) planetary boundary layer scheme, which parameterizes turbulent mixing in the lower atmosphere. Accurate boundary layer height and stability forecasts are essential for predicting hub-height wind speeds at wind farms, as the vertical mixing of momentum dictates the wind profile between the surface and turbine rotor layers.

While the primary HRRR is a single deterministic run, it is nested within the High-Resolution Ensemble Forecast (HREF) system. The HREF combines multiple time-lagged HRRR runs with other convection-allowing models to generate probabilistic forecast products. These ensemble-derived probabilities of cloud cover or wind ramps allow grid operators to quantify uncertainty and set dynamic operating reserves.

A statistical post-processing technique that corrects systematic biases in raw NWP output like HRRR. By establishing a regression relationship between historical model forecasts and local observations, MOS removes persistent errors caused by terrain smoothing or parameterization flaws. This step is critical for converting HRRR's gridded irradiance into accurate, site-specific Global Horizontal Irradiance (GHI) predictions for individual solar farms.

Unlike HRRR's single deterministic run, ensemble systems like the High-Resolution Ensemble Forecast (HREF) generate multiple future atmospheric states by perturbing initial conditions or model physics. This produces a distribution of outcomes, explicitly quantifying forecast uncertainty. For grid operators, ensemble spread is a direct proxy for reserve requirements, enabling risk-based decisions that a single HRRR run cannot support.

A recursive Bayesian algorithm that optimally estimates a dynamic system's state from noisy sensor measurements. In renewable forecasting, a Kalman filter adaptively corrects HRRR's systematic bias in real-time as new SCADA or pyranometer observations arrive. Unlike static MOS, the Kalman gain dynamically adjusts the correction weight based on recent model performance, making it ideal for non-stationary error patterns.

A computationally efficient alternative to dynamical ensembles. AnEn searches a historical archive for past atmospheric states similar to the current HRRR forecast. The corresponding historical observations form the predictive distribution. This method implicitly captures local effects and model biases without running multiple NWP simulations, providing a probabilistic forecast from a single deterministic HRRR run.

The final deliverable for grid operators, expressing future wind or solar generation as a probability distribution or set of quantiles (e.g., P10, P50, P90). HRRR provides the high-resolution deterministic core, but post-processing via Quantile Regression or ensemble dressing converts this into a full uncertainty profile. This enables operators to hold dynamic operating reserves based on the quantified risk of a ramp event.

The standard metric for evaluating probabilistic forecasts derived from HRRR post-processing. CRPS measures the integrated squared difference between the forecast's cumulative distribution function and the empirical observation. Unlike point metrics like RMSE, CRPS evaluates both calibration (statistical consistency) and sharpness (concentration), punishing forecasts that are overconfident or poorly resolved.