Example forecast · wf-north-07 · 120 MW windExample
04080120160MW-12hNow+12h+24h+36h+48h
P10 to P90P50Realized
P50 at +6 h112.4 MW
P10–P90 at +6 h101–124 MW
Next gate closure14:45 CET
Illustrative values, not a real site. Real forecasts carry their run time, model version and calibration window.

Skycast · Forecast

The forecast your site actually sees

Skycast runs an ensemble of AI weather scenarios, downscales each one to the asset with physics-ML, and turns local weather into calibrated P10, P50 and P90 power.

Capabilities

What Skycast does

Ensemble

Scenario fans, not single runs

Many plausible weather futures every run, so uncertainty is measured rather than guessed.

Downscaling

Terrain, wakes and irradiance

Physics-ML surrogates map coarse weather to hub height and panel plane for each asset.

Quantiles

Calibrated P10, P50, P90

Bands are checked against realized output so a P90 really is exceeded about one time in ten.

Horizons

15 minutes to 72 hours

One model chain for intraday and day-ahead, with frequent refresh as new data lands.

Self-calibration

Learns from each site

Realized SCADA output updates each site's error model every day.

Attribution

Error in euros

Forecast error is priced with your settlement data, by site and horizon.

How it works

Inside Skycast

  1. Ensemble

    AI weather models generate many global scenarios several times a day.

  2. Downscale

    Modulus surrogates resolve each scenario at the asset.

  3. Convert

    Per-asset power models turn local weather into output, including curtailment and availability.

  4. Calibrate

    Quantile calibration aligns the bands with realized production.

Specification

Inputs, outputs and hardware

ItemDetail
InputsGlobal weather fields, SCADA (10 min), availability, metering (15 min), site metadata
OutputsP10/P50/P90 per asset and portfolio, 15-minute steps, 15 min to 72 h
RefreshDesign target: intraday data refresh every 60 s per asset; full ensemble runs several times a day
DeliveryREST API, webhooks, CSV and Parquet
Training hardwareDGX or HGX (H100, H200) for weather and downscaling models
ServingTriton and TensorRT; NIM where available

Questions

About Skycast

How is Skycast different from a weather forecast?

A weather forecast gives wind and sun over a grid cell. Skycast gives power at your asset with a calibrated range, after accounting for terrain, wakes, layout and availability.

Do you build your own global weather model?

Not at first. We start from open AI-weather models and put our effort into downscaling, power conversion and calibration. A proprietary global model is a later goal.

What about solar?

The same chain applies: downscaled cloud and irradiance, plane-of-array conversion, inverter limits and soiling, then calibrated quantiles.

Can we compare against our current vendor?

Yes. The backtest and shadow period compare both forecasts on your own sites using the same error and cost metrics.

See Skycast on your own sites

We start with a backtest on your history and show the result side by side with your current approach.