Most icing forecasts that fail in production don't fail because the underlying weather model is weak, they fail because the forecast was never calibrated to the specific turbine it's supposed to protect. Renewcast's approach starts from the site, not the model: an AI-native digital twin per turbine, dynamic weekly recalibration, and real-time retraining layered on top of an ensemble that dynamically selects the best-performing NWP models for each forecast. Validated on Renewcast's own operational data, that combination reaches 80% accuracy and 91% recall on icing events, with a false-alarm rate of just 5–6%.
Why generic icing forecasts fail
Icing onset depends on a very local combination of temperature, humidity, liquid water content and turbine-specific microclimate. A forecast tuned to "cold and humid, statistically" will miss the site that ices at –2°C and the one that doesn't ice until –6°C. Generic accuracy on a generic benchmark says nothing about whether a forecast actually knows a given site.
How local calibration changes the outcome
Two forecasts can start from the same base weather model and still disagree completely on when a specific turbine ices. The difference is what happens after the base forecast: whether it is corrected using that site's own historical icing events through an AI-native digital twin recalibrated weekly, or left as a generic regional output refreshed on a fixed schedule. Local calibration and real-time retraining are the difference between a forecast that's directionally right and one a trading desk can actually act on.
Why this matters now
- Cold snaps that drive up regional imbalance costs are exactly when a poorly calibrated icing forecast is most expensive to trust.
- Portfolios with cold-climate exposure carry this risk on every icing-prone site, whether or not it's being actively managed.
- Reviewing forecasting calibration ahead of the next cold season costs far less than discovering the gap during it.
How Renewcast works
- The AI-native digital twin per turbine and weekly recalibration we introduced in our latest blog, the foundation that makes site-specific calibration possible.
- Dynamic selection of the top-performing NWP models feeding each forecast, rather than committing to a single source.
- A robust RNN framework built to stay reliable even when input data quality drops, with real-time retraining triggered by new meteo forecast runs.
Weather data integration: models, frequency, resolution
Renewcast fuses multiple meteorological sources and models to reduce single-model bias; the platform ingests global NWP outputs and local observations, applies point-level correction, and updates forecasts hourly to reflect the latest model runs and on-site telemetry, the same multi-source approach applied specifically to icing-relevant variables.
Technical architecture: accuracy, speed, reliability
- Cloud-native SaaS with API access and client UI for parametric reports and downloads, refreshed hourly.
- Digital-twin layer mapping machine-level behaviour to forecast outputs, improving per-asset correction and reducing site-specific bias.
- Performance metrics and SLAs focused on NMAE/RMSE/skill scores rather than vanity metrics, so clients can quantify forecast value on icing-exposed sites specifically.
Support for trading and operations
- Locally calibrated day-ahead and intraday forecasts reduce imbalance exposure and enable tighter positions before settlement.
- API access and frequent updates allow automated trading systems to ingest corrected positions as icing risk evolves through the day.
Comparison with traditional solutions and other forecasting platforms
- Traditional single-model or purely statistical forecast services correct for site bias, if at all, on a fixed schedule, by the time a correction lands, the cold snap that needed it may already be over.
- Renewcast's weekly recalibration plus real-time retraining keeps the site-specific correction current through the event itself, which is what turns a directionally-right forecast into one a trading desk can act on intraday.
References
Wind turbine icing characteristics and icing-induced power losses to utility-scale wind turbines. PMC.
An Ensemble Network for High-Accuracy and Long-Term Forecasting of Icing on Wind Turbines. PMC.
Predicting Wind Farm Production and Icing Losses with AFRY DAPS. AFRY.
Cold Climate Wind: Challenges, Technological Solutions and Policy. WIREs Energy and Environment.
Record high electricity prices in Northern Norway. High North News.