When rime ice builds up on a transmission line, it’s already too late. The escalating weight and drag can lead to line clashing, flashovers, and catastrophic outages. For decades, utility companies have been forced to react to these events, often scrambling crews, managing costly repairs, and dealing with grid instability.
What if you could move from reaction to prevention?
The Gap: Why No One Offers This Operationally
Reliable, asset-level rime-ice forecasts have been missing from the operator's toolkit for two simple reasons:
- NWP Models lack sharpness: Traditional numerical weather prediction (NWP) models forecast at a multi-kilometer scale. This "smeared out" view misses the micro-scale physics that create ice on a specific transmission line. The result is vague guidance for a region, not actionable intelligence for an asset.
- You Can't "Tack On" Precision: Many localized AI/ML solutions merely post-process the imprecise NWP forecasts. This approach is unable to generate a dependable rime-ice forecast that was absent from the core model. You’re correcting a signal that was never there.
A New Capability: GFT-HQ Learning Rime-Ice Natively
Our collaboration with Hydro-Québec validates this new approach. By post-training Silurian’s GFT weather foundation model with data from Hydro-Québec’s own network of icing sensors, we created GFT-HQ.
This model learns to natively forecast rime-ice, rather than just post-processing a NWP forecast. The result is a new, operational capability: GFT-HQ achieves an Average Precision of 0.73 for day-ahead rime-ice detection. This provides operators with a reliable, full-day warning that aligns directly with their existing planning windows.
AP = 0.37
AP = 0.72
Beyond aggregate skill, we also analyzed the model's ability to capture high-impact episodes that posed particular challenges for Hydro-Québec’s operations.


What This Unlocks: From Reactive to Proactive
- 01Prevent false dispatches
Targeted De-icing
Shift from reacting to surprise accretion events to scheduling proactive crew runs on priority spans with quantified false-dispatch risk.
- 02Derate with context
Hazard-Aware DLR
Fuse probabilistic icing trajectories with your DLR engine so you can preemptively derate corridors before load threatens system stability.
- 03Day-ahead foresight
Smarter Market Decisions
Fold weather-driven outage probabilities into hedging, unit commitment, and congestion planning to improve revenue capture.
- 04Unified forecasting
Single, Coherent Model
Leverage the same GFT backbone to enhance wind, temperature, and precipitation guidance wherever sensors report.
As atmospheric volatility increases, the cost of a failed forecast is amplifying. To move beyond simple precision to true decision optimization, a new approach is required.
The GFT-HQ model proves this new capability is unlocked by post-training weather foundation models with high-value, operational sensor data.
Don't just monitor. Predict. See how your sensor network can power a new class of high-precision forecasts. Read more in the preprint here.



