Forecasting engine
NWP ensemble plus asset-level ML correction. Not one or the other.
Most commercial forecast products pick one approach. Gridvynt fuses NWP physics with a continuously retrained ML bias-correction layer — because neither alone is accurate enough for reserve commitment decisions.
Two layers
What each layer handles — and why both matter.
Layer 1: NWP ensemble
The NWP layer handles synoptic-scale weather patterns — frontal passages, orographic lifting, mesoscale convection. These are the effects where physical atmospheric modelling has a genuine edge over pure data-driven approaches. We ingest GFS 0.25°, ECMWF 0.1°, and NAM 3-km runs, then weight the ensemble by recent skill metrics per region and season.
- GFS 0.25° global model (NOAA)
- ECMWF 0.1° high-resolution (ERA5 calibrated)
- NAM 3-km continental US mesoscale
- Dynamic ensemble weighting by region × season
Layer 2: ML site correction
The ML layer targets the systematic bias that NWP can’t resolve: local terrain shading, panel soiling drift, wake loss in wind farms, seasonal albedo changes, and inverter-level clipping patterns. It trains on your rolling actuals feed and updates weights every 6 hours — so forecast accuracy compounds over a deployment period.
- Gradient boosting ensemble with rolling retrain
- Minimum 90-day actuals warm-up period
- Separate correction models per asset type
- 6-hour weight update cycle against SCADA actuals
Output specification
What the forecast API returns — and what each band is for.
P10
10th percentile
Production floor: 90% probability that actual output will exceed this level. Use for conservative spinning reserve commitment when the grid is operating near its minimum reserve margin. During high-VRE dispatch periods, P10 is the defensible floor for your BAL-001 compliance calculation.
P50
Best estimate / median
The central forecast — equal probability of over and under. Use for day-ahead energy quantity bids, hourly scheduling, and production accounting. The P50 is calibrated monthly against realized actuals. We report out-of-sample P50 MAE and RMSE by horizon bucket so you can see accuracy directly.
P90
90th percentile
Production ceiling: 90% probability that actual output will fall below this level. Use for curtailment risk assessment — if a solar plant is near its interconnection limit and P90 approaches that limit, automated curtailment pre-notice is warranted. Also relevant for capacity factor crediting and ancillary services upper-bound calculations.
| Parameter | Value | Notes |
|---|---|---|
| Forecast horizon | 72 hours |
T+1h through T+72h from model run time |
| Temporal resolution | 15 minutes |
288 data points per asset per model run |
| Update cadence | Every 6 hours |
Triggered by NWP model run availability (00Z, 06Z, 12Z, 18Z) |
| Output bands | P10, P50, P90 |
Derived from ensemble spread + historical residual distributions |
| Solar units | MW (AC) |
Irradiance → power conversion using registered inverter spec |
| Wind units | MW (AC) |
Wind speed → power via turbine power curve with wake correction |
| Spatial scope | Asset-level + portfolio roll-up |
Individual site forecasts plus aggregated portfolio P-bands |
See forecast accuracy on your assets.
We run a 30-day back-test against your historical actuals before the pilot goes live. You see the accuracy improvement before you commit.