AI Against Blackouts: How Fraunhofer Wants to Keep the Grid Stable
On 2 June 2026, Fraunhofer presented several AI projects for grid stability. They are meant to relieve operators, cushion the skills shortage and prevent blackouts like the one on the Iberian Peninsula. This article explains what the projects do, which hardware belongs with them and what European companies should watch for on maturity, security and the EU AI Act.
Europe's power grid now needs around 500 manual interventions a day to stay stable, where one or two were enough a few years ago. The reason is the energy transition, with many decentralised, weather-dependent feed-ins instead of a few inertia-rich power plants. On 2 June 2026, Fraunhofer presented several AI projects: eKI4DS, with NVIDIA and TransnetBW, detects critical grid states using machine learning; an AI agent from Fraunhofer IOSB-AST automates routine analyses and cushions the skills shortage, with a field test at three grid operators running since February 2026. Grid-forming inverters, battery storage and bidirectional charging add the stabilising hardware. The blackout in Spain and Portugal on 28 April 2025, when around 15 gigawatts were lost and the frequency dropped from 50 to 47 hertz, shows what is at stake. The central caveat: AI in grid operations is high-risk under the EU AI Act, with obligations from December 2027. The AI stays an assistant; the decision and the governance stay with people.
500 interventions a day: why the grid is at its limit
Europe's power grid now needs around 500 manual interventions a day to stay stable, where one or two were enough a few years ago. The reason is the energy transition: many decentralised, weather-dependent feed-ins replace a few large, inertia-rich power plants. Fraunhofer researchers see the ageing grid as the central hub of the energy transition and turn to AI to keep the growing complexity manageable.
How AI is used in grid control in practice is set out in innobu's piece on agentic AI in energy utilities and grid management . The current push comes from research.
The Fraunhofer projects: how AI supports grid operators
Fraunhofer bundles several AI projects meant to relieve operators and prevent cascading failures. At their core is explainable AI that detects critical grid states early and delivers concrete recommendations, without taking the decision away from people. This division of roles matters: the AI stays an assistant, the operator keeps control.
eKI4DS
With Fraunhofer IEE, NVIDIA and TransnetBW. Machine learning detects dangerous grid states and proposes remedial actions, trained on an open-source model of the German transmission network.
AI agent
From Fraunhofer IOSB-AST. Automates routine analyses and reduces the long onboarding. A field test with three grid operators has run since February 2026, with deployment expected in one to two years.
InterSCADA
An EU project from Fraunhofer IEE. A modular open-source platform for data acquisition, remote monitoring and grid control across European networks, to detect anomalies earlier.
Operations Center
From Fraunhofer IFF. Consolidates data from generation, transport, storage and consumption and generates AI-based training scenarios for crises, such as after the sabotage outage in Berlin.
Operators need one to two years of training. Automated analysis lowers that entry barrier and relieves the teams in day-to-day work.
Dr.-Ing. Dennis Rösch, Fraunhofer IOSB-ASTStability from below: inverters and vehicle-to-grid
AI alone does not stabilise a grid; it needs stabilising hardware. Grid-forming inverters can replace the lost rotational inertia of conventional power plants and actively support frequency and voltage. Combined with battery storage and bidirectional charging, they create a decentralised reserve that can even form island grids in an emergency.
Grid-forming inverters
They actively support frequency and voltage and replace the inertia of large generators. In the SUREVIVE project, Fraunhofer ISE connected the first grid-forming battery system directly to a substation.
Storage and island grid
In the ALene project, Fraunhofer couples inverters with batteries and measures grid impedance in real time. During a large outage this allows a local island grid to be built until the main grid returns.
Vehicle-to-grid
An EV battery of around 90 kilowatt-hours covers about a week of household consumption. Bidirectional charging supports the grid at peak times; pilot projects already ran in Munich and Hamburg.
Hardware and AI belong together: Explainable AI detects a critical state, but physical technology has to act on it. Grid-forming inverters, storage and controllable loads are the levers that make an AI recommendation effective in the first place. Anyone planning AI has to plan the hardware too.
European perspective: learning from the Iberian blackout
The blackout on the Iberian Peninsula on 28 April 2025 shows what is at stake. Around 15 gigawatts of power were lost, about 60 percent of demand at the time, and the frequency dropped from 50 to 47 hertz. The cause was a combination of low system inertia, grid-following rather than grid-forming inverters and missing voltage support. European grid operators are already working on AI tools.
At the European level, AI in grid operations is not new territory. The ENTSO-E association lists concrete examples: TenneT uses the GridOptions tool for congestion management, RTE in France runs the ORIGAMI system to relieve operators, and the Czech operator ČEPS uses AI to detect errors in grid models. The Horizon Europe project AI-EFFECT is building a European testing and experimentation facility for trustworthy AI in the energy sector together with TenneT.
The high share of renewable generation was not the cause of the Iberian blackout; the management of voltage and inertia in the system was.
This framing matters because it counters blanket blame. Renewables as such do not endanger the grid; missing support mechanisms do. This is exactly where grid-forming inverters and AI-based early warning come in, as the piece on virtual power plants and AI battery storage also shows.
Challenges and risks
AI in the grid is no sure thing. It falls under the strictest rules of the EU AI Act and meets an industry that so far uses AI mostly in customer service, barely in the technical core. Companies should soberly weigh four points.
High-risk under the EU AI Act: AI used to manage critical infrastructure is classified as high-risk under Annex III of the EU AI Act. Following the latest deadline shift, the obligations apply from 2 December 2027 and include risk management, documentation, conformity assessment and human oversight.
Maturity gap in the technical core
According to PwC, 79 percent of utilities already use AI, but the focus is on customer service. Grid operations and generation lag far behind. Many utilities test AI in pilots; few are mature in productive use. This gap between ambition and reality is the biggest hurdle.
Security and cyber attack surface
The sabotage outage in Berlin on 3 January 2026 left 45,000 households and more than 2,200 businesses without power for four days and showed that a real-time situational picture of the grid was missing in the crisis. AI can improve that picture, but it also widens the cyber attack surface and creates new dependencies on data quality and models.
Data quality and legacy systems
AI in grid operations needs clean, consistent data and has to be integrated into safety-critical legacy systems. ENTSO-E names data gathering across many IT systems, data quality and integration into existing protection systems as central hurdles. Without a solid data base, every AI recommendation stays uncertain.
People stay responsible
Explainable AI lowers the risk of wrong decisions but does not guarantee correctness. For a high-risk application like grid control, the human has to remain the final authority. Researchers stress this division of roles deliberately, also to secure acceptance among operators.
What companies should do now
Utilities and grid operators should plan AI in grid operations as an assistant, not a replacement for operators, and consider governance from the start. Anyone who builds the data base, explainability and compliance now will be ready in 2027 when the EU AI Act obligations apply. Four steps help.
-
Start with explainable assistance
Begin with systems that recommend rather than switch autonomously. That keeps the operator responsible and the findings traceable, a must for a high-risk application.
-
Build the data base and situational picture
Build a real-time situational picture and consistent data models. Without clean data, every AI fails. The Berlin outage showed how costly a missing picture becomes in a crisis.
-
Plan compliance early
Classify AI applications in grid operations as potentially high-risk and plan documentation, risk management and human oversight. This is cheaper before the first productive run than after.
-
Pilot with research
Use Fraunhofer projects and EU initiatives such as AI-EFFECT as a starting point for your own tests. And plan the hardware: grid-forming inverters and storage make AI recommendations effective.
How regulation and AI come together at utilities is explored in the piece on AI regulation and compliance for energy utilities . For the background on data centers as a grid risk, see the article on the NERC alert and AI data centers .
AI in the grid is a high-leverage tool, but not full automation. The value comes from explainable assistance, a solid data base, stabilising hardware and clear governance. Companies that respect this can manage the growing complexity without giving up control.
Further reading
Frequently asked questions
Grid operators now have to make around 500 interventions a day to keep the power grid stable, where one or two used to be enough. The reason is the energy transition: many decentralised, weather-dependent feed-ins such as wind and solar replace a few large, inertia-rich power plants. This increases complexity and raises the risk of cascading failures and bottlenecks.
AI detects critical grid states earlier and proposes concrete remedial actions to operators. Fraunhofer is developing the eKI4DS project with NVIDIA and TransnetBW, which uses machine learning to spot dangerous conditions and deliver explainable recommendations. An AI agent from Fraunhofer IOSB-AST automates routine analyses and cushions the skills shortage. The AI stays an assistant; the human keeps the decision.
In the Iberian Peninsula blackout on 28 April 2025, around 15 gigawatts were lost, about 60 percent of demand at the time, and the frequency dropped from 50 to 47 hertz. The cause was a combination of low system inertia, grid-following rather than grid-forming inverters and missing voltage support. According to IEEFA, the high share of renewables was not the cause; the management of voltage and inertia was.
Yes. AI used to manage critical infrastructure such as the power grid is classified as high-risk under Annex III of the EU AI Act. Following the latest deadline shift, the related obligations apply from 2 December 2027. They include risk management, technical documentation, conformity assessment and human oversight. Anyone planning AI in grid operations should factor in these requirements from the start.
Utilities should plan AI in grid operations as an assistant, not a replacement for operators, and consider governance early. Sensible steps are explainable assistance systems that recommend rather than switch autonomously, a solid data base with a real-time situational picture, an early EU AI Act classification and links to research projects such as AI-EFFECT. According to PwC, 79 percent of utilities already use AI, but the focus is on customer service, not the technical grid operation.