Forecast solar and wind 72 hours out. Commit the right thermal reserve.
Gridvynt fuses NWP ensemble data with a site-calibrated ML correction layer to produce P10/P50/P90 production forecasts at 15-minute resolution, 72 hours ahead. Grid operators commit thermal reserve against the probability bands — not a single-point estimate that understates ramp risk.
The curtailment problem
The curtailment problem is a forecast problem.
When a single-point NWP forecast assigns low probability to a solar ramp event, the dispatch desk responds rationally: commit extra thermal reserve. That thermal over-commitment physically curtails the renewable generation at the margin. The weather didn’t fail. The forecast — and specifically its failure to communicate ramp probability — did.
The platform
A forecasting platform built for grid operations, not weather apps.
15-Minute, 72-Hour Horizon
288 forecast points per asset per run. Solar irradiance and wind speed disaggregated to your GPS coordinates, panel tilt/azimuth, and turbine rotor diameter — not interpolated from the nearest NWP grid cell.
P-Band to MW Reserve Target
The reserve commitment engine converts P10/P90 production bands into a single spinning and non-spinning reserve MW quantity your day-ahead scheduling team can act on — accounting for your balancing area’s area control error target.
Continuous SCADA Feedback Loop
ML correction model weights retrain every 6 hours against your SCADA actuals feed. Systematic biases from panel soiling, wake effects, and terrain shading are corrected automatically as actuals arrive.
Forecasting capabilities
Two physics engines. One operational decision.
Gridvynt runs a three-member NWP ensemble (GFS at 0.25°, ECMWF-IFS at 0.1°, NAM at 3-km) fused with a gradient-boosted site correction layer. The NWP ensemble handles synoptic-scale dynamics — frontal passages, orographic lifting, mesoscale convection. The ML correction layer handles the systematic biases NWP can’t resolve at asset scale: local terrain shading, panel soiling drift, wind turbine wake losses, and seasonal albedo shifts.
- NWP ensemble ingestion (GFS, ECMWF, NAM)
- Asset-level irradiance and wind speed disaggregation
- Probabilistic confidence bands at P10/P50/P90
- Automated bias correction against rolling actuals
- SCADA and historian feed integration
Operator feedback
From operators who stake reserve decisions on it.
We reduced thermal reserve over-commitment by roughly 9% in the first quarter after switching to Gridvynt’s P10/P90 output bands. Using a single-point deterministic forecast to set spinning reserve for a 1,200 MW solar portfolio was always going to leave us over-committed. The probability bands gave the scheduling desk something it could actually act on rationally.
The 15-minute solar ramp prediction is what we were missing in day-ahead scheduling. Our EMS ingests the forecast via the REST push and we use the P90 band directly for our spinning reserve calculation. The integration took about a week once we had the SCADA actuals feed set up.
Solutions
Built for every node in the grid dispatch chain.
ISO / RTO Operators
P10/P50/P90 forecast inputs at the portfolio and individual-asset level for day-ahead and real-time scheduling. Delivers in OATI-compatible format, Pi System write-back, or REST push — at your scheduling system’s pull cadence.
ISO / RTO detailsUtility Resource Adequacy
Multi-week and seasonal capacity factor distributions for IRP filing and reserve margin stress-testing. Direct export to PROMOD, PLEXOS, and Aurora production cost model formats — no manual data transformation required.
Resource adequacy detailsRenewable Asset Managers
Asset-level P50/P90 production tracking against your modeled budget. Automatic curtailment attribution separates weather variance from grid-ordered curtailment and potential availability events — with confidence scores for each category.
Asset manager detailsGetting started
From raw NWP data to a reserve commitment in four steps.
Connect Your Assets
Point Gridvynt at your asset register — GPS coordinates, panel spec or turbine model, tilt/azimuth. API or CSV upload.
Ingest Ensemble Forecast
Gridvynt ingests GFS, ECMWF, and NAM model runs every 6 hours. No WRF license required.
Run Site Calibration
ML correction layer trains on your actuals data — minimum 90 days of historical SCADA output. Accuracy improves continuously.
Export to Your EMS
Forecasts publish to your energy management system or historian via REST API, CSV schedule, or push webhook — at your EMS's native resolution.
30-day pilot
Ready to cut curtailment from your portfolio?
We run a 30-day pilot on your actual asset data before you sign anything. Talk to a Gridvynt forecasting engineer.