Generative AI is moving faster than any technology the energy sector has seen, yet almost none of it runs inside a live control room. GridOperandi is a Horizon Europe proposal to change that, and we are building the consortium now.
GridOperandi – Taking state-of-the-art GenAI into large-scale operation across Europe’s transmission, distribution and consumer ecosystems
On 3 June 2026 the European Commission adopted its Strategic Roadmap for Digitalisation and AI in Energy as part of the European Technological Sovereignty Package. It sets the direction to 2030, starts an annual Energy Digitalisation Forum, and backs an AI model for energy trained on European data and built by European companies.
The pressure comes from both sides of the meter. The International Energy Agency projects that electricity demand from data centres will more than double by 2030, so the same grids that must carry the AI boom are expected to use AI to run better.
The cost of not seeing the grid in real time is now documented, too. The expert panel investigating the 28 April 2025 blackout in Spain and Portugal found that key voltage-control equipment was switched by hand and that operators lacked a real-time view of reactive-power margins. Data gaps blocked parts of the investigation.
In three years, generative AI went from chat demos to agents that read documents, call software tools and write code. For the energy system, four uses stand out.
Assistants that explain an alarm, summarise a contingency or find the right operating procedure in seconds.
Generated fault, weather and demand scenarios that let models train on rare events no grid wants to experience.
Time-series models pre-trained on huge datasets and adapted to load, solar, wind and price forecasting.
Code and energy-saving apps generated from a description, which shortens the path from idea to deployed tool.
Two trends make these usable on a real grid. Models are getting small: distillation and quantisation shrink large models until they run on a substation gateway. And Europe is building its own compute through the AI Factories network, backed by the GenAI4EU initiative, which mobilises close to €700 million across Horizon Europe, Digital Europe and the EIC.
The caution is justified. Generative models can produce confident wrong answers, consume a lot of energy, and need data that grid operators cannot simply hand over. In a control room, a fluent mistake is worse than no answer.
GridOperandi does not start from zero. It builds on a decade of EU-funded work, and joins it up.
Federated local digital twins of the electricity system, demonstrated in eight pilots across 11 countries. Its twins are the sandbox GridOperandi’s models are tested in.
The two projects funded to prototype a generative-AI-powered digital spine for the EU energy system. GridOperandi takes the next step the call asks for: large-scale operation.
A pan-European foundation model for grids. GridOperandi is the frugal, edge layer that can run such models in operation, rather than a competitor to them.
Cyber-physical resilience tools for transmission and distribution operators, which feed the security tests in GridOperandi’s licence gates.
Edge AI already running in a distribution substation in Granada, Spain: proof that AI at the grid edge is possible today.
Ask a grid operator why an AI tool that worked in a pilot is not in the control room, and the answers come down to four gaps.
There is no shared way to decide when a model is safe enough to advise an operator, or to act on its own under supervision.
Large models cannot run on thousands of substations and chargers, and nobody measures how much energy the AI itself consumes.
SCADA, meter and market data cannot leave a company or cross a border, so models stay trained on one grid.
Who approves a model, who maintains it, and how the AI Act’s rules for critical infrastructure are met are still open questions.
One idea runs through the project: AI that grid operators can license, afford and keep running. It starts with how a model earns its place in operation.
Every model passes four gates before it touches a live grid. Each gate has measurable go/no-go criteria, and the licence file doubles as the AI Act evidence pack.
Around the licence, GridOperandi combines ten technologies that each answer one of the gaps.
Small reflex models answer in milliseconds on substation and charger gateways for forecasting, anomaly detection and local control.
Distillation from open foundation models, 4- and 8-bit quantisation, pruning, ONNX Runtime.
Larger generative models are called only when an operator needs an explanation, a scenario or a procedure, so expensive reasoning is used sparingly.
Small open-weight LLMs such as EuroLLM, LoRA fine-tuning, retrieval with citations.
Edge agents learn locally and share only model updates, so a cross-border swarm of operators improves together while raw data stays home.
Federated learning with secure aggregation (e.g. Flower), incremental and continual learning.
Language agents call power-flow simulators for every number they report, so answers come from grid physics, not from the model’s guess.
Tool calling to PowSyBl or pandapower, verified outputs, refusal when confidence is low.
Models that understand topology and limits generalise across grids instead of memorising one.
Graph neural networks on CIM/CGMES models, physics-informed neural networks.
Reinforcement-learning agents steer EV charging, storage and heat pumps, with hard safety constraints they cannot break.
Multi-agent reinforcement learning with constraint shields, hardware-in-the-loop before field use.
Every model reports the energy it uses per decision, and is licensed only if it saves more energy than it consumes.
Joules-per-inference metering, compute budgets per site, training on EU AI Factories.
Generative models create faults, storms and extreme demand days, so models are tested on what the grid rarely sees but must survive.
Generative scenario engines validated against the digital twin.
Every component speaks the standards operators already use and connects to the Common European Energy Data Space.
IEC CIM/CGMES, IEC 61850, FIWARE NGSI-LD, SAREF, data-space connectors, Common Reference Framework apps.
Cyber tests are part of every licence gate, and the toolkit is released as open source for anyone to reuse.
IEC 62443 and 62351, NIS2, the network code on cybersecurity, release through LF Energy.
Choose the kind of organisation you are. You will see what you gain from the project and what we would ask from you in return.
Pilots decide whether a proposal like this wins. Lighthouse sites are in discussion in Spain, Romania, Bulgaria, Slovenia, Austria and the Netherlands. We are looking especially for these:
A DSO in another regulatory context, so community and grid models learn across borders.
Low inertia, weak connectivity, high renewables: the hardest test for frugal edge AI.
Data centres and AI Factories willing to treat compute as a flexible grid resource.
Ports, bus depots, truck charging hubs and airports where electrification meets congestion.
Cooperatives or communities with hundreds to thousands of members ready to use energy-saving apps.
Industrial parks, irrigation pumping or cold storage that can shift load without hurting production.
A strong pilot has access to an operational grid or real assets, at least twelve months of historical data, someone who can sign a commitment letter, and willingness to run models in shadow mode before anything goes live.
Leave your email and organisation. The proposal team at ATHENA RC will get back to you to discuss your role. Questions? Write to nikolaos.pavlidis@athenarc.gr.
GridOperandi is a proposal in preparation and has not been funded. Proposal coordinator: ATHENA Research Center (ATHENA RC). Contact: nikolaos.pavlidis@athenarc.gr.