EU AI Act transparency duties apply now; high-risk duties from December 2027. Check your exposure
Insights About us Careers
Contact us
Industries / 8 verticals

AI for Energy and Utilities

Long asset lives, unforgiving physics, and a regulator who will ask what your model did during the incident. Eight verticals, written for the person accountable for keeping the system standing.

8 verticals
Individual pages
2 Dec 2027
Annex III critical infrastructure
Safety component?
The question to settle
Why this sector is different

Forecasting is only as good as your measurement.

Energy is the sector where AI is most often sold as a modelling improvement and most often limited by an instrumentation gap. Renewable forecasting research is consistent on this: forecasts built on ground measurement at the site substantially outperform those built on numerical weather prediction alone, and both outperform remote sensing. The gap between those data sources is larger than the gap between a good model and a mediocre one, which means the highest-return investment is frequently a met mast rather than a data scientist.

The second constraint is that most of your operational history describes a system that was working. Grids, plants and networks are run to avoid the events you most want to predict, so the failures, excursions and instabilities are rare by design. That is not a reason to avoid the work; it is a reason to build condition monitoring against known-good behaviour rather than failure prediction against eleven examples, and to be sceptical of anyone selling the second.

Evidence, not enthusiasm

Where AI actually earns its place in energy

Ranked by evidence rather than by conference frequency. The maturity column is our own read; the catch column is what the vendor case study leaves out.

WhereWhat it doesMaturityThe catch
Renewable generation forecastingPredicts wind and solar output hours to days ahead for dispatch and trading.ProvenGenuinely effective, and accuracy is driven more by your measurement data than by model choice.
Demand and load forecastingPredicts consumption by network area and time horizon.ProvenMature and reliable. Weather and behavioural change are the error sources, and both are modellable.
Asset condition monitoringDetects degradation in transformers, turbines, pumps and rotating plant.ProvenWorks well against known-good signatures. Failure prediction is a different and much weaker claim.
Vegetation and network inspectionAnalyses imagery from drones, helicopters and satellites for encroachment and defects.StrongMature and measurable. The constraint is the crew capacity to act on findings, not the detection.
Energy trading and price forecastingForecasts prices and imbalance, supports position and dispatch decisions.StrongReal for short-horizon and imbalance work. Regime changes and market coupling break models that ignore them.
Grid state estimation and operations supportImproves visibility of network state and supports operator decisions.MixedValuable and close to the safety boundary. Advisory works; anything actuating is a different regulatory programme.
Predictive maintenance on major assetsPredicts failure of transformers, turbines and large plant.MixedThe most over-sold application here. Major assets fail rarely by design, so the failure examples do not exist.
Autonomous grid controlAI making switching or protection decisions without a person.Not readyA safety component under any reasonable reading, with the conformity assessment and functional safety work that implies.

The number that should change where you spend first

Reviews of renewable forecasting consistently find that the input data source dominates model choice. Forecasts built on in-situ ground measurement reach a median absolute percentage error in the region of eight per cent; numerical weather prediction alone lands nearer eleven; remote sensing alone is materially worse again. The best ensemble and hybrid methods reach five to six per cent — but only where the measurement data supports them. If your sites are not instrumented, the highest-return investment is instrumentation, and any supplier proposing a model without asking what you measure at the site is answering a question you did not ask. See time series forecasting.

What the rules require

Critical infrastructure, and the line inside it

Energy sits under network security regulation, sector cybersecurity codes and, for some systems, the AI Act's critical infrastructure category. The classification question is finer than it first appears. This is our reading as at September 2026 and we work alongside your regulatory, safety and legal functions rather than in place of them.

EU AI Act

Annex III, from 2 December 2027

AI intended to be used as a safety component in the management and operation of critical infrastructure, including the supply of electricity, gas and heating, is high-risk, with obligations from 2 December 2027 following the Digital Omnibus deferral. A safety component is one whose failure could cause physical damage or harm to people or property — which is a narrower test than 'anything touching the grid' and a broader one than 'anything that trips a breaker'.

Where the line falls

Optimisation is not automatically a safety component

Draft guidance indicates that grid optimisation tools where the core safety functions are handled separately are not caught, and that AI used purely for cybersecurity is excluded. But the boundary is genuinely fine, regulators are expected to read safety component broadly in critical infrastructure, and under-classifying carries more risk than over-classifying. Document the reasoning whichever conclusion you reach — the reasoning is what a supervisor will ask for.

Network Code on Cybersecurity

Regulation (EU) 2024/1366, rolling out now

In force since May 2024 and directly applicable. National competent authorities designated from December 2024; harmonised risk assessment frameworks from March 2025. Designated high-impact and critical-impact entities must operate a cybersecurity management system, submit a risk report within twelve months of designation and every three years after, implement controls scaled to impact classification, manage supply chain risk, and demonstrate compliance by audit within twenty-four months. It reaches TSOs, DSOs, significant generation assets and critical ICT service providers.

NIS2

The layer underneath

Energy entities are essential entities under NIS2, with risk management, incident reporting and management accountability obligations. Transposition was due in October 2024 and has been uneven across member states, which means the practical position varies by jurisdiction more than the directive intends. An AI system in an operational technology environment inherits these obligations rather than sitting beside them.

Market integrity

REMIT, revised in 2024

The revised REMIT took effect on 7 May 2024, extending scope to energy storage and certain financial instruments, giving ACER cross-border investigatory powers, and increasing what must be disclosed about algorithmic trading. For anyone running automated trading or dispatch optimisation against market signals, market manipulation exposure is a design constraint rather than a compliance afterthought.

Functional safety

The regime that was already there

IEC 61508 and the sector standards built on it remain operative. Reconciling a machine learning component with a framework built around deterministic behaviour and quantified failure rates is genuinely hard, and it is the reason credible deployments in this sector keep AI advisory — informing an operator — rather than actuating.

What this means for a build

The decisive question is whether the system performs a safety function or informs a person who does. An advisory model that recommends a dispatch, flags an asset or ranks an inspection route sits outside almost all of the above. The same model wired to switch, trip or curtail automatically is a safety component with conformity assessment, AI Act high-risk duties and functional safety reconciliation attached. Decide it in week one and document the reasoning, because the two architectures diverge from the first sprint. See EU AI Act compliance readiness and AI risk assessment.

Questions

FAQ

Marked up with FAQPage schema so these answers can surface in search results and inside AI assistant responses.

Do you have energy sector experience?

Yes in forecasting, asset condition analytics, inspection imagery and operational data engineering. Less in market trading strategy, and none in protection scheme design or control system engineering — on those we work alongside your engineers rather than in place of them. Each of the eight vertical pages states our depth in that area rather than implying uniform expertise across the sector.

Is our AI system a critical infrastructure safety component?

It depends on whether its failure could cause physical damage or harm, not on whether it touches the grid. An optimisation tool where core safety functions are handled separately is generally outside the category; a system whose malfunction could damage plant or endanger people is inside it. The line is fine, regulators are expected to read it broadly, and under-classifying is riskier than over-classifying — so settle it in week one and document the reasoning either way.

Why do our predictive maintenance projects stall?

Because major energy assets fail rarely, which is the point of maintaining them. A transformer fleet might give you a handful of relevant failures across decades, which will not support a model however it is framed. Condition monitoring against a known-good signature works, needs no failure examples, and catches degradation trends — it is a weaker claim than prediction and it is the one the data actually supports.

How much can AI improve our renewable forecasts?

It depends far more on what you measure than on the model. Published reviews consistently find that forecasts built on in-situ ground measurement outperform those built on numerical weather prediction alone, which in turn outperform remote sensing — and the gaps between those data sources exceed the gap between a good model and an ordinary one. If your sites are not instrumented, start there. Any supplier who quotes an accuracy figure without asking what you measure on site is quoting someone else's data.

What is the most common way energy AI projects fail?

Operational data that was never collected for analysis: SCADA historians with tags nobody can identify, maintenance records as free text, asset registers that disagree with what is in the ground, and time stamps recorded at insertion rather than at measurement. Teams scope a modelling project and find a data engineering one in month three. We front-load that assessment, and it has ended engagements before anyone spent money on a model.

Start with the problem, not the technology.

Thirty minutes, no charge, no deck. Tell us what is going wrong in your organisation and we will tell you whether AI is the right instrument — including when it plainly is not.