Forecasting the Sky - AI and the Renewables Grid

EcoLedger Insights — Practitioner note on AI infrastructure for the energy transition.

Forecasting the Sky with AI

AI and the Renewables Grid

At 12:33 on 28 April 2025, Spain and Portugal lost roughly 60% of their electricity generation in seconds. Frequency on the Iberian grid fell from 50 Hz to 47 Hz. Lights went out for the better part of a day across two countries. The post-mortem largely cleared renewables of blame — voltage control, not fuel mix, was the proximate cause — but the incident exposed a harder truth.1 A grid built around predictable, dispatchable generation is now being asked to run on weather. The job is no longer supply the load. It is forecast the sky.

The problem: a grid that runs on weather

Fossil-fuel power was, above all, controllable: a turbine produced what the operator asked of it. Wind and solar are not knobs; they are the sky. Output depends on cloud cover, wind speed, temperature and solar angle. That single shift creates four overlapping problems for system operators.

Figure 1 — As solar penetration rises, midday net load collapses while evening ramping needs grow sharply.

The first is the shape of net load. As solar capacity grows, midday becomes effectively oversupplied while early-evening demand snaps back as the sun sets. California’s duck curve has deepened into a canyon, with ramps that require up to 17 GW of replacement generation in three hours.2 The second is forecast-error reserves: utilities hold standby capacity proportional to how badly they might mis-predict renewable output — currently 6–10% of installed capacity, at an annual cost in the hundreds of millions to over a billion dollars per region.3 The third is curtailment — electricity the grid cannot absorb. The UK discarded ~12% of its wind generation in 2025 (>£1bn). CAISO curtailed 3.4 TWh of solar and wind in 2024, 29% more than 2023. Chile effectively curtailed every megawatt of new solar it added.4 The fourth is voltage and frequency control, which used to come free with rotating fossil mass and now has to be engineered.

How AI changes the calculus

The common thread across all four problems is uncertainty. Reduce the uncertainty and the problem shrinks. AI is now doing this at three layers of the system.


Figure 2 — Reported gains across forecasting, dispatch and atmospheric modelling. The Aurora bar reflects compute speed-up, not accuracy.

At the forecasting layer, machine-learning models trained on satellite imagery, weather reanalysis and turbine telemetry now outperform traditional numerical forecasts by wide margins. National Grid ESO has reported a 33% improvement in solar forecast accuracy from nowcasting models;5 Open Climate Fix and DeepMind have pushed accuracy roughly 40% higher than baseline;6 and Microsoft’s Aurora foundation model matches or beats the European Centre’s flagship forecast on more than 91% of targets at one five-thousandth of the compute.7 At the dispatch layer, reinforcement-learning systems schedule batteries and demand against probabilistic forecasts; DeepMind’s wind operations work raised the market value of 700 MW of wind by ~20% by converting forecast output into 36-hour-ahead firm delivery commitments.8 At the protection layer, anomaly-detection models flag inverter and transformer faults seconds before they cascade.

What the AI-enabled renewables grid looks like

Project the trend out a few years and a different grid comes into focus. Probabilistic forecasts update every minute and feed scheduling directly. Storage assets — utility batteries, EVs, behind-the-meter systems — are dispatched by AI agents bidding into real-time markets. Virtual power plants aggregate millions of distributed devices into firm capacity; the global VPP market is forecast to grow from US$6.3bn in 2025 to US$39.3bn by 2034 (~22.6% CAGR), with demand response already 48% of activity.9 Inverter-based resources provide the voltage and frequency support that synchronous generators used to deliver for free. Curtailment becomes a market signal rather than a waste. Reserves shrink because uncertainty shrinks. The grid stops being something operators balance reactively and becomes something an intelligence layer balances continuously.

The bottom line

A renewables-heavy grid is one of the hardest real-time optimisation problems in the economy. It cannot be solved by transmission alone, or by storage alone, or by markets alone. It needs an intelligence layer that can forecast, dispatch and respond faster than humans can. On the renewables grid, AI is no longer optional infrastructure for the energy transition — it is the substrate.

Sources

  1. ENTSO-E (2025) — Final Report on the Grid Incident in Spain and Portugal on 28 April 2025. Frequency dropped from 50 Hz to 47 Hz; ~60% of generation lost in seconds. ↩
  2. U.S. EIA / CAISO — Solar build-out has deepened the “duck curve” into a canyon; evening ramps now require up to ~17 GW of replacement generation in three hours. Battery capacity grew from ~500 MW (2020) to >13 GW (early 2025). ↩
  3. Lawrence Berkeley National Laboratory — Day-ahead forecast-error reserve requirements grow with renewables share to 6–10% of capacity; reserve sharing is valued at $0.09–$1.24 bn/year for a single multi-utility region. ↩
  4. IEA, Renewables 2025 — UK curtailed ~12% of wind generation in 2025 at >£1bn cost. CAISO curtailed 3.4 TWh of solar and wind in 2024, +29% YoY. Chile’s solar curtailment growth exceeded its solar capacity additions. ↩
  5. National Grid ESO / Open Climate Fix — ESO reports a ~33% improvement in solar forecast accuracy from machine-learning nowcasting on satellite imagery. ↩
  6. Open Climate Fix + Google DeepMind — AI-based solar forecasting reported ~40% accuracy improvement over baseline. ↩
  7. Aurora (Microsoft Research) — 1.3-billion-parameter atmospheric foundation model. Matches or beats ECMWF’s IFS on >91% of targets at roughly 1/5,000th the compute. Nature, 2025. ↩
  8. Google DeepMind — Neural network trained on weather forecasts and turbine data raised the market value of 700 MW of wind capacity by ~20% via 36-hour-ahead delivery commitments. ↩
  9. Precedence Research / Utility Dive (2025–26) — Global VPP market projected from US$6.28 bn (2025) to US$39.31 bn (2034), CAGR 22.6%; demand response represented ~48% of 2025 market share. ↩
Back to blog