AI in renewable energy is already in practical use. Its clearest job today is forecasting: predicting how much electricity solar panels and wind turbines will produce, so that grid operators can balance supply and demand more cheaply. In Great Britain, a machine-learning solar forecast has been running in the national control room since November 2025.
There is a second side to the story. The data centres that train and run AI are a fast-growing source of electricity demand. A fair picture has to cover both what AI does for the grid and what it asks of it.
How it works
Wind and solar power vary with the weather, and the grid operator has to match supply and demand second by second. If forecasts are wrong, it must pay other generators to turn up or down at short notice, which costs money.
Older forecasting methods were simple. For solar, one approach was to multiply expected sunlight (irradiance) by the installed capacity. Machine-learning models instead learn from years of past weather data and actual output, and can combine many inputs at once. They do not change how a panel or turbine works; they improve the predictions that the grid is run on.
AI is also used on the network side, for example to spot faults faster or to estimate how much power existing lines can safely carry in real conditions.
Real-world examples
NESO and the Alan Turing Institute (2019)
In 2019 National Grid ESO, now the National Energy System Operator (NESO), worked with the Alan Turing Institute on a machine-learning solar forecast. It used a random forest model with about 80 inputs to replace the simple irradiance-times-capacity method. NESO reported the new forecast as 33% more accurate, though it did not define the measure or baseline. The work was funded through Ofgem’s Network Innovation Allowance.
NESO has also worked with the University of Sheffield to map rooftop solar that the grid cannot see directly. These “invisible” panels lower the demand the grid sees on sunny days, so knowing where they are matters for forecasting.
Open Climate Fix’s Quartz Solar (2025)
In November 2025 Quartz Solar, an AI tool from Open Climate Fix, went live in NESO’s control room. NESO’s Head of AI, Digitalisation and Innovation, Carolina Tortora, said it “can accurately predict solar generation up to 36 hours in advance”.
The other figures come from Open Climate Fix itself. The company says the tool is 2.8 times more accurate than NESO’s previous solar forecasts, could avoid at least £30 million a year in imbalance costs, and could save about 300,000 tonnes of CO2 a year. These are the company’s estimates, not NESO’s own audited results. NESO went on to publish an Energy Forecasting Strategy in March 2026.
DeepMind and Google wind farms (2019)
In 2019 DeepMind applied machine learning to 700 MW of Google’s wind capacity in the central United States, predicting output 36 hours ahead. This let Google commit power to the grid in advance. DeepMind said the approach “boosted the value of our wind energy by roughly 20 percent” compared with making no advance commitments, and described the results as early.
The key word is value. The turbines did not produce more electricity; the same output was sold on better terms.
Data centre cooling (2016)
In 2016 DeepMind reported that its AI system cut the energy used to cool Google’s data centres by 40%, equal to a 15% reduction in overall power usage effectiveness (PUE) overhead. This is not a renewable energy result, and it is nearly a decade old, but it shows AI being used to trim the energy demand of computing itself.
The IEA’s view on grids (2025)
The International Energy Agency’s 2025 report Energy and AI says AI-based fault detection can cut the length of outages by 30–50%, and that sensors combined with AI management could unlock up to 175 GW of transmission capacity without building new lines. In one scenario, it projects that wider use of existing AI applications could cut energy-related emissions by around 5% in 2035. That is a projection, not a measured result.
The data centre question
According to the IEA, data centres used about 415 TWh of electricity in 2024, roughly 1.5% of the world total. The US accounted for 45% of that, China 25% and Europe 15%.
The IEA’s Base Case projects data centre demand rising to about 945 TWh by 2030, more than double the 2024 level, and about 1,200 TWh by 2035. It calls AI “the most important driver” of that growth. The IEA reports data centres as a whole; AI is a growing part of their load, not the whole of it.
On supply, the IEA expects renewables to meet about half of the growth in data centre demand to 2035, with gas adding about 175 TWh and nuclear a similar amount. Emissions from data centre electricity rise from about 180 Mt today to 300 Mt by 2035 in the Base Case, and up to 500 Mt in its “Lift-Off” case, still under 1.5% of energy-sector emissions.
In Great Britain, NESO’s 2025 Future Energy Scenarios put current data centre demand at about 7.6 TWh, from 2.4 GW of connected sites. By 2050 the range runs from about 30 TWh to 71 TWh depending on the scenario, a spread that reflects how uncertain the outlook is.
What the numbers say
- Solar forecasting: 33% more accurate (NESO, 2019); 2.8 times more accurate than NESO’s previous tools (Open Climate Fix’s claim, 2025).
- Wind: about 20% more value, not more electricity, from 700 MW of Google wind capacity (DeepMind, 2019).
- Grids: outage durations cut by 30–50% with AI fault detection; up to 175 GW of extra transmission capacity possible (IEA, 2025).
- Global demand: data centres used about 415 TWh in 2024, projected at about 945 TWh by 2030 (IEA, 2025).
- Great Britain: about 7.6 TWh now, 30–71 TWh by 2050 depending on the scenario (NESO, 2025).
To see how much of Britain’s electricity comes from wind and solar at any moment, see our live GB electricity mix.
Limitations and honest assessment
- Many headline figures are company claims. The £30 million saving and 300,000 tonnes of CO2 come from Open Climate Fix, and DeepMind’s wind result was described as early by DeepMind itself.
- Accuracy claims are hard to compare. NESO’s 2019 “33%” was not tied to a published metric, so it cannot be set against the 2025 “2.8 times” figure.
- Forecasts do not change the weather. Better predictions help the grid prepare for a calm or cloudy day, but they do not make wind or solar output any steadier. A plant’s capacity factor stays the same however good the forecast is.
- Projections are not outcomes. The IEA’s 5% emissions cut and its data centre figures for 2030 and 2035 are scenarios.
- AI adds load. Any savings from AI forecasting have to be weighed against the electricity data centres use, which the IEA expects to more than double this decade.
The UK angle
Britain is a useful test case. Its grid operator has used machine learning for solar forecasting since 2019 and adopted Open Climate Fix’s tool in its control room in 2025. At the same time, data centre demand is set to grow, though NESO’s own range for 2050 is wide.
The government’s AI Growth Zones policy aims to place AI data centres where grid capacity and clean power are available. Regulation is also developing. Ofgem published guidance on “Ethical AI use in the energy sector” on 20 May 2025, built on safety, security, fairness and sustainability. A second version followed on 13 May 2026, noting that AI had moved from pilots to operational use in the sector.
Frequently asked questions
How is AI used in renewable energy?
Mainly for forecasting wind and solar output, so the grid can be balanced more cheaply, and increasingly for grid tasks such as spotting faults. In Britain, NESO uses a machine-learning solar forecast in its control room.
How does machine learning improve solar forecasting?
It learns from past weather data and actual output rather than using a fixed formula. NESO’s 2019 model with the Alan Turing Institute used about 80 inputs, and Open Climate Fix’s Quartz Solar predicts output up to 36 hours ahead.
Did DeepMind make wind turbines produce more power?
No. DeepMind’s 2019 work with Google raised the value of the wind energy by roughly 20% by forecasting output 36 hours ahead. The amount of electricity generated did not change.
How much electricity do data centres use?
About 415 TWh worldwide in 2024, or 1.5% of global electricity, according to the IEA, which projects about 945 TWh by 2030. In Great Britain, NESO puts current demand at about 7.6 TWh.
Can AI reduce power cuts?
The IEA says AI-based fault detection can cut the length of outages by 30–50%. It shortens outages rather than preventing every fault.
Is AI in the energy sector regulated in the UK?
Ofgem, the energy regulator, has issued guidance on ethical AI use, first in May 2025 and in an updated version in May 2026.