How AI can transform the UK’s electricity grid

Lucy Yu, the UK Government’s AI Champion for Clean Energy, says that artificial intelligence (AI) could provide the transformative capability required to operate the energy system more efficiently and reliably, but that new governance is required in order to safeguard and allow AI use to be optimised.

Lucy Yu was appointed to lead an independent review of the opportunities and barriers associated with the use of AI in the UK’s electricity grid. Over seven months, her team engaged with more than 200 people across network companies, industry, technology, academia, research, and consumer-facing organisations. The resulting review contains 34 recommendations.

The case for applying AI begins with the changing nature of the electricity system itself. Yu points to the UK’s move away from fossil fuel generation and towards weather-dependent renewables.

“In 1990, coal provided about 80 per cent of Britain’s electricity, but just under two years ago, on 30 September 2024, the UK closed its last coal-fired power plant, and that ended 142 years of coal-generated electricity. A big chunk of that coal-generated electricity has been replaced by wind and solar power.”

At the same time, changes are taking place on the demand side. Households and businesses are increasingly adopting technologies which alter how they interact with the electricity network.

“Behind-the-meter technologies are becoming more significant, and they change import and export behaviours from households and businesses with the main grid, but they are often not visible to the system operator, and although their adoption is still early overall, the trend is upwards.”

For the system operator, this makes visibility and forecasting increasingly important. Yu says that operating the grid involves combining “dozens of different tasks and models and forecasts, including continuously updated assessments of supply and demand”.

“We need the best possible observability of the entire system. We need to know what assets exist and where, and what their states are. We cannot make a good forecast with an incomplete picture of the system.”

The second requirement is stronger forecasting of power generation, including generation which varies according to the weather. However, the transition itself creates a challenge for the models used to forecast the system.

“Historical real-world data can help us to train and build better forecasting models, but when we move into a new regime, by its very nature, that means we naturally have less historical data available for that purpose, so we have a challenge to work on improving these forecasts over time.”

She argues that this is an area where AI can make a practical contribution. “High-quality forecasts are important, but we do not just forecast supply and demand; we also need to understand different potential failure modes in the system and how likely they are, so that we can make dynamic assessments about the security of the system.”

Yu says that, in practical terms, this means being able to answer a question: “If something does go wrong, will the grid stay stable?”

Scenario planning

The increasing complexity of the electricity system means that there are more possible scenarios to consider. This creates a computational challenge, particularly when decisions have to be made quickly.

“We need to be able to run a large number of scenarios or simulations cost-effectively, simultaneously, and certainly extremely quickly,” Yu says. “If they take too long to compute, then by the time we have the results, we cannot use them to inform real-time operational decisions.”

One potential solution is the use of surrogate models, which can replicate the behaviour of more complex models while requiring substantially less computational power. Advances in GPUs also provide the capacity to run large numbers of scenarios much faster.

Yu points to research demonstrating that the best-performing GPU tested could provide a speed-up of almost 375 times on a network with 118 nodes, while a model representing 1,354 buses of the European transmission network recorded a speed-up of more than 70 times.

She argues that the combination of these technologies could create a significant new capability for grid operators.

“We can start to think about combining different things like these smaller surrogate models and more efficient GPUs with better AI-driven forecasts themselves. Collectively, this becomes a powerful new capability for the system operator to run a renewables-heavy grid efficiently and reliably.”

Yu also sees scope for improving the way energy sector models themselves are developed and assessed, and she argues that there needs to be a better understanding of which models perform best for particular applications.

“We need to be able to understand and compare the performance of different models against relevant metrics and if we can find a way to do that, then we can also exploit that to rapidly improve the quality of the whole discipline overall.”

Yu highlights ImageNet and the Critical Assessment of Protein Structure Prediction (CASP) as examples where shared datasets and testing frameworks helped drive rapid advances.

“High-quality verified and structured data sets for training, for shared testing, and clear performance benchmarks are all things that we can also do for energy systems-related forecasting, with the ultimate aim of rapidly driving model improvements,” she says.

AI could also increasingly affect consumers directly. Low-carbon technologies such as electric vehicles, heat pumps, and batteries can be optimised alongside time-of-use tariffs, allowing consumption to shift towards cheaper or greener periods.

As consumers become more accustomed to AI in other areas of their lives, Yu expects “greater willingness to engage with it in energy or to sign up to products and services which use AI”.

Unintended consequences

However, although AI optimisation can act rapidly and at scale, poorly designed incentives or signals could produce unintended consequences.

“The key point here is that if you give an optimiser any signal to optimise against, it has the possibility to aggressively exploit that, and some forms of AI could turbocharge that,” Yu says.

Her recommendations therefore include ensuring that AI optimisers receive effective signals about grid congestion on operational timescales, potentially through mechanisms such as regional flexibility markets, locational wholesale pricing, and dynamic network pricing.

The longer-term question is whether AI could contribute to an autonomous grid. Yu is careful to distinguish autonomy from automation.

“Automation is already fairly widely used in the grid to perform actions on timeframes that are not possible for humans to deliver,” she says. However, an autonomous system is different as it has “delegated authority to both pursue objectives and take actions within a range of operating conditions without needing ongoing human direction”.

That distinction has implications for governance, testing, safety, accountability, and liability. “All of this needs to begin with a shared language and definitions for how we understand these systems and their risks,” Yu says.

“The rules that govern how we plan and operate the grid today, were put in place for a world that we no longer live in,” she adds.

The grid now has decades of asset failure data, increasing quantities of real-time information, more sophisticated forecasting and optimisation techniques, faster simulation capabilities, and the ability to rapidly control network assets.

“None of these things existed when the rules were set decades ago. They exist today, but we are still using legacy rules that prevent us from exploiting all the benefits they could bring.”

Yu argues that this could allow the system to move from deterministic approaches, based on predefined worst-case scenarios, towards probabilistic and risk-based management.

“We are already doing this in small pockets of the grid,” she says, pointing to dynamic line rating as one example.

“On windier days, overhead lines are cooled more, and so the safe limit of the line can be higher. This is useful because we can transmit more wind instead of curtailing it.”

Scaling that approach across the electricity system, however, is considerably more challenging. Yu concludes: “Every process that feeds into the planning and operation of the grid has a system transformation problem. It needs high levels of observability and interoperability of data between physical and institutional boundaries.”