Human control of the electricity system is becoming increasingly unfeasible. Thousands of new distributed assets, new market mechanisms, and increasingly complex regulations means that no humans can optimise the system, and certainly not one that needs to be balanced in real time.
AI and advanced data algorithms offer the promise of handling this level of complexity, models can handle billions of parameters and make reasoned decisions on the best actions to take that satisfy the constraints that they are asked to work to.
Agentic AI can take over many of the manual human tasks, for example, needing to check whether an alarm is the same one that goes off when the sun is in a particular position, or whether this is something that needs investigation.
Generative AI can help design better electricity grids, make decisions on which assets to place where, vulnerabilities in the system can be designed out, and a huge range of scenarios can be bench tested before they meet the real world.
The European Union’s AI act has a clause that defines electricity (and other utilities) as critical systems and hence as high risk. This places stringent requirements on producers and operators of AI systems connected to our networks.
A fear with autonomous systems is that they will cause a crisis by acting together, the algorithm is designed to reduce power when the price increases, all systems rapidly drop consumption, and the system frequency spikes due to excess power.
There’s normally a protecting ‘air gap’ between the systems that control our networks and other systems. But what happens if the hacking has been done at the point that the AI was trained? We can’t look inside the model and check to see that it wasn’t trained to take a specific action on a particular date and time.
The reality is that we need AI to manage the coming complexity, but we also need expertise and experience in deploying it. It would be a brave operator that connected AI without a significant body of evidence that it was safe to do so.
