Why P90 Forecasts Matter More Than P50 for Reserve Commitment
For most market participants, P50 is the forecast they care about. But for grid operators setting reserve, the P90 tail is where the decision actually lives.
Blog
Technical notes from the Gridvynt team.
We write about renewable energy forecasting, grid balancing, NWP methodology, and what we’re learning in real deployments.
ECMWF vs. GFS: Which NWP Model Performs Better for Solar Forecasting?
A comparison of ECMWF HRES and GFS 0.25° skill scores for irradiance forecasting across different climate regions and forecast horizons.
What SCADA Data Quality Issues Mean for ML Bias Correction
Outlier meter readings and sensor dropout are unavoidable in operational SCADA feeds. Here's how data quality issues propagate into ML correction models and how to handle them.
How Curtailment Attribution Differs from Production Forecasting
Production forecasting asks: what will the asset generate? Curtailment attribution asks: why did it generate less than forecast? These are different technical problems.
Wind Wake Modeling at the Asset Level
Grid-scale NWP models miss the production loss from inter-turbine wake effects. A field note on how sector-based wake correction improves 0-4h wind accuracy.
Integrating with OSIsoft Pi System for Forecast Delivery
A technical walkthrough of the Pi System connector: how we write P10/P50/P90 forecast bands back as PI tags and make them visible in existing PI Vision dashboards.
Resource Adequacy Planning with Probabilistic Solar Forecasts
How utilities can use P10/P90 solar forecast bands in seasonal IRP modeling to account for forecast uncertainty in resource adequacy studies.
The Case for 6-Hour NWP Update Cycles in Operational Forecasting
Why 24-hour forecast update cycles leave reserve commitment decisions stale — and how more frequent model ingestion changes the accuracy math for ramp event prediction.
Gradient Boosting vs. Deep Learning for Solar Site Calibration
We tested both approaches on 18 months of actuals across six solar installations. The results were more nuanced than "deep learning wins in every regime."
Building the Gridvynt REST API: Design Decisions and Tradeoffs
A technical post on how we designed the forecast delivery API — response envelope format, pagination strategy for long-horizon outputs, and webhook push architecture.
How Terrain Effects Degrade Solar Irradiance Forecasts
Mountain shading, elevation gradients, and terrain-channeled cloud motion are all sources of irradiance forecast error that grid-scale NWP models systematically miss.
What We Learned from Our First Six Pilot Deployments
A candid look at what went well and what we had to rebuild after our first six customer pilots — integration surprises, asset data gaps, and forecast calibration lessons.
Measuring Forecast Skill Against NWP Baselines
How we define and report forecast skill score — and why comparing against raw NWP baseline matters more than comparing against climatological persistence models.