Methodology
How we build forecasts that grid operators can stake reserve decisions on.
Forecast accuracy for grid operations is not the same as general weather model accuracy. RMSE at the synoptic scale does not tell you whether a model correctly characterizes the probability of a 300 MW solar ramp in 45 minutes. Gridvynt publishes its full methodology, accuracy metrics by weather regime and horizon bucket, and validation approach — so your team can assess whether our output is appropriate for your specific reserve commitment workflow.
Methodology sections
What our methodology covers.
NWP Ensemble Construction
We ingest GFS 0.25° (NOAA), ECMWF-IFS 0.1°, and NAM 3-km model runs every 6 hours. For Mountain West deployments, HRRR (High-Resolution Rapid Refresh, 3-km) is ingested hourly as a 4th ensemble member for the 0–18h horizon where its convective-scale resolution provides material improvement over GFS and NAM. Ensemble member weights are recalculated weekly by skill metrics per climate zone and season — not a static average that degrades in shoulder seasons.
Asset-Level Disaggregation
Grid-resolution NWP output (temperature, irradiance, wind speed, wind direction, relative humidity) is disaggregated to your registered asset footprints using GPS coordinates, panel tilt and azimuth, turbine hub height and rotor diameter, and a 30-meter terrain digital elevation model. Solar irradiance is decomposed into direct normal irradiance (DNI), diffuse horizontal irradiance (DHI), and global horizontal irradiance (GHI) components using the Erbs decomposition model.
ML Bias Correction
Gradient-boosted correction models trained on 90+ days of your SCADA actuals correct for systematic NWP errors that pure physics modelling cannot resolve at asset scale: local terrain shading, panel soiling drift (tracked via optical cleanliness proxy), seasonal albedo changes, wind turbine wake sector losses, and inverter-level clipping patterns. Separate correction models are maintained per asset type (utility-scale solar, rooftop aggregation, land-based wind, offshore wind). Weights update every 6 hours on a 90-day rolling training window.
Probabilistic Confidence Bands
P10 and P90 intervals are derived from two components: ensemble member spread (which captures synoptic-scale uncertainty) and historical ML residual distributions by horizon bucket, weather regime, and season (which captures local site uncertainty NWP spread misses). Confidence band calibration is checked monthly against realized actuals using the Brier skill score — if calibration drifts, bands are adjusted automatically at the next 6-hour model cycle. We do not publish symmetric ± intervals.
Accuracy Reporting
We report MAE, RMSE, and skill score versus a persistence baseline and versus the raw NWP ensemble baseline — by horizon bucket (0–12h, 12–36h, 36–72h), by weather regime (clear-sky, variable cloud, ramp event, extreme wind), and by season. You receive this report monthly throughout your subscription, not just at pilot close. Validation methodology follows NREL Solar Radiation Database (NSRDB) and Wind Integration National Dataset (WIND Toolkit) baseline standards where applicable.
Accuracy metrics
The numbers we hold ourselves to.
See the methodology on your asset data.
The pilot includes a back-test against your historical actuals. You see measured accuracy improvement before commit.