UK Power Networks directly monitors only about 14% of its more than 120,000 secondary substation transformers, yet the British electricity distributor says machine learning, smart-meter information and asset data allow it to estimate utilisation across the entire low-voltage network. The company reports 96.7% forecast accuracy for what it calls secondary-network visibility, while locations predicted to be heavily utilised are passed to network planners for expert assessment. That figure is important, but it should not be confused with 96.7% accuracy in predicting equipment failures.
That distinction captures both the opportunity and the limits of artificial intelligence in electricity networks. Utilities do not need an algorithm capable of predicting the exact hour when a transformer will fail to create commercial value. Software that reliably identifies which assets deserve closer attention can already change where engineers inspect, where capital is deployed and which equipment is replaced first.
The model is spreading through other parts of the industry. E.ON says around 180 drones are already operating across its German grid companies and generating high-resolution imagery for artificial-intelligence applications. The European utility also says AI-assisted predictive maintenance and image analysis are already in use rather than existing solely as research programmes. Enel similarly says artificial intelligence and computer vision are being used in distribution-grid predictive maintenance to identify components and anomalies.
The commercial question is therefore becoming less about whether utilities will use AI and more about which maintenance decisions can safely be moved from fixed schedules toward continuous assessment of asset condition and failure probability.
How is artificial intelligence being used in power-grid maintenance?
Grid AI is not a single technology. Several distinct approaches are being applied to different maintenance problems.
Computer-vision systems analyse photographs, thermal images and video collected by drones, helicopters, fixed cameras or field crews. Algorithms can be trained to identify corrosion, damaged components, vegetation encroachment and other visible abnormalities before an engineer manually reviews the flagged image.
Machine-learning systems working with operational data take a different approach. They analyse patterns in measurements such as voltage, current, temperature, loading or smart-meter information and compare them with normal behaviour or historical equipment problems.
Digital twins add another layer by creating software representations of physical assets or systems. GE Vernova’s SmartSignal predictive-maintenance platform, for example, combines digital-twin models with near-real-time sensor information to monitor equipment behaviour and identify developing abnormalities. GE Vernova describes the technology as capable of detecting, diagnosing and forecasting equipment problems rather than merely displaying measurements on a dashboard.
The practical objective is not perfect foresight. It is to create a ranked list of where maintenance resources are likely to create the greatest value.
Where are utilities already using AI in day-to-day grid operations?
UK Power Networks provides one of the clearest examples because machine-learning outputs are already influencing low-voltage network planning and operational workflows.
Smart-meter information is used to support automatic fault reporting and to check remotely whether repairs have restored electricity supply. The company also combines data from monitored substations, smart meters and asset records to estimate utilisation at secondary substations without dedicated monitoring equipment. It says the resulting model provides effective visibility across 100% of the low-voltage network even though only 14% of secondary substations have direct monitoring.
E.ON operates on a substantially larger geographic scale. The company says its electricity and gas networks extend roughly 1.6 million kilometres across Europe and serve approximately 47 million customers. Around 180 drones are currently in use across its German grid businesses, generating imagery that can be analysed through artificial-intelligence applications. E.ON also states that AI-based predictive maintenance is already used to detect anomalies and help prevent incidents.
Enel says machine learning is being used for predictive analysis of electricity-distribution infrastructure, while computer vision supports identification of components and anomalies. The company describes the objective as concentrating inspection activity on equipment facing greater failure risk rather than applying the same level of inspection to every asset.
These examples are important because they move the discussion beyond laboratory algorithms. AI is already influencing where grid companies look, which assets they investigate and how maintenance teams allocate time.
Can AI predict transformer failures before customers lose power?
Pacific Gas and Electric Company provides one of the most instructive examples because its predictive-maintenance programme exposes both the value and the limitations of the technology.
PG&E’s EPIC 3.20 programme developed machine-learning models using information already available to the utility, including smart-meter voltage data, asset location and weather. During development, engineers reviewed more than 270 model predictions between April 2021 and February 2022. PG&E reported that 64% corresponded to relevant transformer anomalies requiring further investigation, while another 26% identified other distribution-system issues.
The programme subsequently moved beyond the demonstration stage. PG&E’s 2023 EPIC annual report states that EPIC 3.20 moved to production during 2023 and estimated potential customer savings of approximately $3.2 million annually through reduced outage duration and proactive replacement of failing transformers.
The later regulatory record adds an important qualification. PG&E subsequently told California regulators that the original model was strong at identifying voltage-related anomalies but had difficulty precisely predicting when a transformer would fail. The utility operationalised part of the technology as a power-quality management tool while continuing development of more precise transformer-failure prediction outside the EPIC programme.
That is perhaps a more useful illustration of grid AI than a perfect prediction would be. The technology did not need to tell PG&E exactly when every transformer would break to create operational value. It could still identify unusual behaviour earlier and narrow the number of assets requiring investigation.
How are AI and autonomous drones changing power-line inspection?
Transmission and distribution networks contain enormous numbers of geographically dispersed components, making visual inspection a natural target for automation.
National Grid has previously tested drone-based pylon inspection with Keen AI and autonomous-flight specialist sees.ai, using imagery to identify corrosion and other structural conditions. The objective was to replace part of the repetitive process of collecting and manually analysing pylon photographs with more consistent automated inspection.
The concept has now expanded considerably. In June 2026, National Grid and Keen AI secured £355,985 of Alpha-stage funding through Ofgem’s Strategic Innovation Fund for Foundational Shared Model Operations, or FoSMo. The programme brings together National Grid, SP Energy Networks, SSEN Transmission, UK Power Networks and Electricity North West to develop a common artificial-intelligence foundation for analysing visual electricity-network asset data.
The project remains an Alpha-stage innovation programme, not a fully deployed nationwide AI maintenance system. Its commercial logic is nevertheless notable. Grid operators often inspect similar pylons, insulators, cables and fittings, but each utility historically possesses only its own relatively limited collection of examples of rare defects.
FoSMo aims to pool anonymised information so participating operators can improve the underlying model without separately rebuilding the same capability. National Grid estimates that full adoption could save the industry around £22.6 million over five years from 2027 by reducing duplicated AI development and improving asset-condition assessment. It also projects roughly 85,000 fewer customer interruptions and 5.2 million fewer minutes of lost supply annually if the technology achieves the anticipated performance at scale. Those figures are forecasts, not achieved savings.
How could predictive maintenance change the economics of electricity utilities?
Electricity networks traditionally combine preventive and reactive maintenance.
Preventive maintenance replaces or inspects equipment according to defined intervals, engineering standards or asset age. Reactive maintenance begins when equipment has already failed or produced an obvious operational problem.
Both approaches can be expensive in different ways. Replacing healthy equipment too early wastes remaining asset life and capital, while waiting until equipment fails can create emergency repair costs, customer outages and potentially significant safety consequences.
Predictive maintenance attempts to operate between those extremes. If data indicates that one transformer is behaving abnormally while another similar transformer remains healthy, engineers can prioritise inspection of the first unit without automatically replacing both.
That becomes increasingly valuable as utility capital programmes grow. National Grid invested a record £11.6 billion in its 2025/26 financial year and now expects cumulative capital investment of at least £70 billion between 2026/27 and 2030/31. E.ON plans approximately €48 billion of investment across Europe between 2026 and 2030, with network infrastructure taking a major share.
Every new substation, transformer, cable and overhead line eventually joins the maintenance base. If inspection workload rises at roughly the same rate as physical infrastructure, grid expansion can place substantial pressure on operating expenditure and engineering headcount.
Predictive analytics offers a way to make maintenance intensity scale according to risk rather than simply according to the number of assets installed.
Can smart meters become sensors for the electricity network?
Smart meters were installed primarily to measure electricity consumption, but their value to network companies increasingly extends beyond billing.
UK Power Networks uses smart-meter data to detect loss of supply and to verify whether repairs have restored electricity to affected properties. Multiple meter readings can also help locate the section of the low-voltage network affected by a fault.
The more interesting use is indirect measurement. UK Power Networks combines readings from smart meters with information from directly monitored transformers and asset databases to estimate utilisation at substations without dedicated monitoring devices.
This could change the economics of grid observability. Rather than installing a specialised sensor on every low-voltage transformer, utilities can use highly instrumented locations to train statistical models that estimate conditions elsewhere.
There are limits. Estimated utilisation is not identical to direct measurement, and smart-meter data cannot reveal every mechanical, thermal or electrical failure mode inside a transformer. The likely model is therefore not sensors versus AI, but selective deployment of high-quality sensors combined with inference across the wider network.
Could AI reduce the number of routine field inspections utilities need?
In some applications, it already can reduce unnecessary inspection effort. Removing field engineers entirely is much less realistic.
Computer vision can process thousands of images and highlight suspected defects, allowing specialists to concentrate on the fraction of assets that appear abnormal. Drones can also reduce climbing, helicopter flying and repeated travel required to gather images from overhead infrastructure.
The economics depend on model quality. A system producing excessive false alarms may simply create a new queue of assets that humans must inspect. More importantly, a false negative involving a serious structural or electrical defect could carry substantial safety and reliability consequences.
That is why the strongest existing deployments generally retain human decision-making. UK Power Networks explicitly says sites predicted to have unusually high utilisation are reviewed by expert planners before network investment decisions are made.
AI therefore appears strongest today as a prioritisation layer rather than an autonomous replacement for grid engineers.
Why are ageing grid assets and poor data difficult problems for AI?
Electricity networks contain equipment installed across many decades, produced by different manufacturers and documented under changing engineering and data standards.
Some assets generate continuous telemetry. Others are inspected periodically. Older equipment may have incomplete records or failure histories stored across different information systems.
Rare failures create another machine-learning problem. Utilities naturally possess enormous quantities of data showing assets operating normally but comparatively few examples of catastrophic transformer, insulator or structural failure. Training models to recognise rare events can therefore be difficult.
National Grid’s FoSMo programme directly addresses this problem. National Grid says individual operators may not possess enough examples of unusual defects to train sufficiently robust computer-vision models, while combining anonymised learning across multiple network operators could improve the available dataset.
The challenge is that equipment behaviour is also site-specific. Coastal corrosion, vegetation, temperature, flooding, maintenance history and equipment design all influence failure patterns.
Grid AI consequently requires engineering context rather than simply larger datasets.
What cybersecurity and regulatory risks come with AI-controlled power grids?
Risk changes substantially depending on what the artificial-intelligence system is allowed to do.
Software that ranks drone photographs for an inspector creates one class of risk. An AI system directly operating safety-critical electricity infrastructure creates a considerably higher one.
Cybersecurity therefore becomes important not only because utilities possess sensitive data but because software increasingly influences physical equipment. Compromised inspection data could distort maintenance priorities, while interference with a system involved in operational grid control could potentially have direct physical consequences.
European regulation recognises that distinction. The current consolidated European Union AI Act lists AI systems intended to operate as safety components in the management and operation of electricity supply among the critical-infrastructure uses that can be classified as high-risk.
However, the timetable changed in July 2026. Regulation (EU) 2026/1744 postponed application of the principal Chapter III requirements for high-risk AI systems classified under Article 6(2) and Annex III until December 2, 2027. The earlier version of this feature incorrectly implied those full requirements were already applying in August 2026.
That delay does not make regulation irrelevant. It gives utilities and technology providers more time to prepare for requirements concerning areas such as risk management, data governance, documentation, human oversight, accuracy, robustness and cybersecurity when qualifying systems fall within the high-risk framework.
Will AI eventually predict grid failures before the lights go out?
In selected situations, that is already beginning to happen, although not with the certainty implied by the word “predict”.
PG&E has used machine-learning outputs to identify transformer abnormalities early enough for proactive intervention and has moved elements of the approach into production. UK Power Networks uses machine learning to estimate loading across transformers it does not directly monitor. E.ON uses AI-assisted predictive maintenance and image analysis, while Enel uses machine learning and computer vision in grid-maintenance processes. National Grid and its partners are now attempting to build a common AI foundation for visual asset inspection across several British electricity networks.
The commercial effect may ultimately be less about predicting the precise moment of failure and more about changing the unit of maintenance decision-making.
A traditional asset-management programme asks when a transformer was installed, when it was last inspected and when its next maintenance interval arrives. A data-rich grid can increasingly ask which transformer is behaving unusually, which line has deteriorated fastest and which inspection today is most likely to prevent tomorrow’s interruption.
That transition becomes more valuable as power systems absorb renewable generation, electric vehicles, heat pumps, storage, industrial electrification and rapidly growing data-centre loads. Utilities are being asked to install more infrastructure while maintaining reliability and controlling operating costs.
Artificial intelligence cannot remove ageing equipment, storms, corrosion or human error from the electricity system. It can potentially make the enormous maintenance problem more selective.
For utilities, that may be the real breakthrough.
The most valuable prediction may not be exactly when a transformer will fail. It may simply be knowing which transformer an engineer should inspect first.
Discover more from Business-News-Today.com
Subscribe to get the latest posts sent to your email.
