A Weather Foundation Model for the Power Grid

TL;DR

This work demonstrates that Weather Foundation Models (WFMs) can be effectively adapted to power-grid operations by post-training a pretrained Generative Forecasting Transformer (GFT) on utility asset data. The resulting GFT-HQ delivers hyper-local forecasts for temperature, precipitation, hub-height wind, wind-farm icing, and rime-ice risk on transmission lines, outperforming strong NWP baselines across multiple lead times and metrics, including an unprecedented average precision of $AP\approx 0.72$ for day-ahead rime-ice detection. The approach achieves significant reductions in traditional forecast errors (e.g., $MAE$ for temperature by ~15%, precipitation by ~35%, wind by ~15%) and enables new operational capabilities such as proactive de-icing and improved dynamic line ratings. Practically, this work shows that minimal post-training data can unlock asset-specific, multivariate forecasts with cross-variable coherence, providing actionable insights for grid resilience and reliability under changing climate and asset portfolios.

Abstract

Weather foundation models (WFMs) have recently set new benchmarks in global forecast skill, yet their concrete value for the weather-sensitive infrastructure that powers modern society remains largely unexplored. In this study, we fine-tune Silurian AI's 1.5B-parameter WFM, Generative Forecasting Transformer (GFT), on a rich archive of Hydro-Québec asset observations--including transmission-line weather stations, wind-farm met-mast streams, and icing sensors--to deliver hyper-local, asset-level forecasts for five grid-critical variables: surface temperature, precipitation, hub-height wind speed, wind-turbine icing risk, and rime-ice accretion on overhead conductors. Across 6-72 h lead times, the tailored model surpasses state-of-the-art NWP benchmarks, trimming temperature mean absolute error (MAE) by 15%, total-precipitation MAE by 35%, and lowering wind speed MAE by 15%. Most importantly, it attains an average precision score of 0.72 for day-ahead rime-ice detection, a capability absent from existing operational systems, which affords several hours of actionable warning for potentially catastrophic outage events. These results show that WFMs, when post-trained with small amounts of high-fidelity, can serve as a practical foundation for next-generation grid-resilience intelligence.

Figures (22)

• Figure 1: Weather foundation model for the power grid.
• Figure 2: Hydro-Québec's major facilities and transmission line infrastructure: Current (Left), and planned (Right) expansion; compiled from public Hydro-Québec materials hq_action_plan_2035_2023hq_axes_2025hq_supply_plan_2023hq_strategic_plan_2024.
• Figure 3: (a) Precision–recall curve for rime-ice forecast over the next 24 h forecasts. (b) Cumulative Distribution Function (CDF) of F1 scores of rime-ice over the next 24 h forecast over all stations. Note that Makkonen$^\dagger$ is derived from ERA5 reanalysis and cannot be used operationally.
• Figure 4: Romaine rime-ice event (Nov 2024) at two nearby Sygivre stations 11 km apart. Top: MONTAG_C; bottom: ERIC_C. Heatmap shows GFT-HQ 1-h rime-ice probabilities from successive 6-hourly initializations (y-axis) verifying at each valid time (x-axis; UTC) over 2024-11-06 to 2024-11-24. The black strip labelled "Icing" is the observed binary occurrence. At ERIC_C, a stable high-probability signal appears by 16 Nov and peaks on 18--19 Nov, aligning with the longest observed episode; several later bursts are also forecast. At MONTAG_C, the signal is weaker and more episodic, with only modest probabilities near the brief observed bursts, illustrating strong micro-scale differences despite close proximity.
• Figure 5: Wind-speed forecast improvements from finetuning. (a) Mean absolute errors of forecasts across all stations (b) Fraction of stations with lower forecasting errors than ECMWF-IFS (c) Mean absolute error across all the stations in the evening during peak load