Four approaches to grid intelligence,
and the one that's missing

Physics-informed neural networks (PINNs), grid foundation models (grid FMs), virtual power plants (VPPs) and flexible loads each solve part of the problem. None of them learns from what happens when it acts.

Machine learning has been part of power systems for years, used for everything from forecasting and state estimation to optimisation, anomaly detection and control. Four approaches are particularly relevant to what comes next.

Physics-informed neural networks, or PINNs, bring knowledge of the physical system into the learning process. Grid FMs aim for something broader: representations that can carry knowledge from one task to another. VPPs coordinate thousands of distributed assets so they can participate in the grid as a single resource. More recently, flexible load has shown that large consumers, particularly data centres, can adjust when and how they use electricity in response to conditions on the network.

These are not competing ideas. Each solves a different and important part of the problem. What remains missing is a system that learns from the physical consequences of its own actions.

Teach the model the rules

A neural network trained on data alone can become very good at reproducing the relationships it has seen without ever learning the physical structure behind them. In a power system, that is a serious limitation. A model may perform well under familiar conditions and become much less reliable as the system moves beyond them, precisely where its behaviour matters most.

PINNs take a different approach. They bring governing equations, conservation laws and boundary conditions into the training objective itself. Matching the observations is no longer enough. The model must also find an explanation for them that remains consistent with the underlying physics.

That principle is an important part of our own approach. But PINNs solve a particular problem: they constrain how a model learns. They do not, by themselves, determine what the model should learn from as the grid changes around it.

Teach the model the system

Grid foundation models start from a broader ambition. Instead of building a separate model for each task, the aim is to learn a representation of the grid that can be useful across many of them. A sufficiently capable model might carry what it learns between different assets, network configurations, locations and timescales, providing a common foundation for forecasting, state estimation, planning and eventually control.

But there is only so much that can be learned by looking backwards. Historical data tells us what happened on the grid under the actions that were actually taken. It cannot tell us how the same system would have responded to an action that was never taken. That becomes increasingly important in a distributed grid, where many of the interventions we may want millions of devices to make have never been tried systematically across the full range of network conditions.

Some of the most valuable training data does not exist yet, because it has to be created.

Coordinate the assets

At the same time, the physical grid itself is changing. More of the electricity flowing through it now passes through devices that can be controlled: EV chargers, batteries, heat pumps, solar inverters, industrial loads and, increasingly, computing hardware.

Virtual power plants, or VPPs, make this fragmented capacity manageable. They bring large numbers of distributed resources together so that their behaviour can be forecast, optimised and dispatched as a portfolio. In effect, they allow thousands of small assets to participate in the energy system as something much larger.

That aggregation is enormously useful, but it is not the same as giving the individual assets an understanding of the grid around them. A VPP might determine how much flexibility a portfolio should provide and when. There is still a more local question beneath it: how should each device respond to the electrical conditions at its own point of connection?

Voltage, frequency and network constraints ultimately manifest in physical places. The conditions seen by one device may be quite different from those seen by another, even when both belong to the same portfolio.

That points to another direction for grid intelligence: not only upward through aggregation, but outward across the network itself.

Make the load flexible

A fourth idea has advanced quickly: large electrical loads do not have to be treated as fixed. Industrial processes, cooling systems, electrolysers and, increasingly, computing can shift, shed or modulate their consumption in response to prices, dispatch instructions or conditions on the grid. Computing is particularly well suited to this. The loads are large, concentrated and growing rapidly, while much of the hardware can already be controlled through software.

A good example is how data centres are increasingly being seen as flexible loads. However, much of the technology currently deployed in this manner is more about understanding the facility itself, rather than the actual electrical grid. Which workloads can be deferred without violating a service commitment? How long can thermal inertia sustain a reduction in cooling? Which systems can be throttled, and in what order, before performance begins to suffer? Over time, the operator can build an increasingly detailed picture of what the facility is capable of doing.

Its view of the grid is usually much simpler. The network appears through signals that the facility can respond to: frequency, price, carbon intensity, a dispatch request or some other measure of system conditions. What the facility does not generally need to learn is how its own actions alter the electrical conditions around it. The intelligence is principally concerned with the behaviour and constraints of the load, not with building a continually improving model of the network beyond the meter.

That distinction becomes more important as flexible load scales. What one facility learns about its workloads, cooling systems and operational limits is valuable, but much of that knowledge belongs to the facility. Deploy the same architecture at a thousand sites and each can become better at managing itself. It does not necessarily follow that the system becomes better at understanding the electrical network connecting them.

The deeper question is whether flexibility always has to be engineered this way, site by site, or whether devices can also learn what their actions do to the network around them.

Learn from action

Large Energy Model® begins where these approaches converge. Physics constrains what the model is allowed to learn. Knowledge has to transfer beyond a single device or task. Distributed assets have to be coordinated, and the devices themselves have to be capable of changing their behaviour.

The difference is where the learning continues. Large Energy Model treats the measured consequence of an action as a source of new information about the network.

Its basic unit of experience is simple: the state of the network before an action, the action itself, what happened afterwards, and the electrical, spatial and temporal conditions in which it took place. Historical observations show how the grid has behaved. Carefully bounded interventions can reveal something more: how the system responds when its behaviour is deliberately changed. Accumulated over time and across many points of connection, those responses can help the model learn how controllable actions interact with the network, rather than only recognising patterns in its past behaviour.

A node senses the conditions around it, takes a bounded action within its permitted operating envelope, and measures what follows. That response becomes another piece of training data.

No one has to label the outcome. The grid already did.

Physics supervises the learning

Learning from intervention makes physical constraints more important, not less. A sufficiently expressive neural network can find patterns that fit the measurements remarkably well while bearing little resemblance to the physical system that produced them.

Large Energy Model® therefore brings known physical relationships into the learning process itself. Measurements from participating nodes provide evidence about what is happening on the network, while relationships such as power and energy balance provide constraints on how those observations can be explained. The model is trained not simply to reproduce what it sees, but to do so in a way that remains consistent with the physics of the network it is trying to learn.

This is different from training an unconstrained model and checking its outputs against physical limits afterwards. When physics is present during training, it influences the representations the model develops in the first place. That can become particularly valuable when the network encounters combinations of conditions that are poorly represented in its historical data.

However, there are limits to what this provides. Physical supervision does not prove that an individual prediction is correct. It does not guarantee stability, nor does it make a neural network inherently safe. What it provides is an inductive bias towards explanations that are compatible with what is already known about the physical system.

That distinction becomes much more important when a model is asked not only to predict the grid, but to act on it.

Intelligence doesn't imply authority

Understanding the physics does not guarantee a good decision. In critical infrastructure, the ability to learn and the authority to act are different things.

Each node therefore operates within a defined control envelope. How much freedom it has to adapt depends on the conditions around it and the operating margin available at the time. When those margins are comfortable, the system can permit a wider range of bounded actions. As they tighten, that range contracts. Beyond defined limits, adaptation gives way to deterministic protection, established control logic and safe fallback behaviour.

This means that control authority is never simply handed to the model. It is conditional, bounded and revocable, with the permissible action space changing as conditions on the network change. Physics shapes what the system learns. Governance decides what it is allowed to do.

Intelligence at the point of connection

Placing intelligence close to where the measurements are made means that time-sensitive decisions do not always have to travel to a central model and back. Each node sees only a small part of the network, but it does not exist in isolation. The nodes are coupled by the electrical system itself. When one device changes its behaviour, some of the resulting physical effects may be measurable at other points on the network.

None of this removes the need for communications, protection systems, operators or markets. Nor does physical coupling, by itself, guarantee that many distributed controllers will behave coherently or remain stable. The point is narrower, but important: digital communication is not the only way in which one part of an electrical network can carry information about what is happening elsewhere.

Some information about the behaviour of the system is already travelling through the physics.

Sitting beneath the portfolio

Large Energy Model® does not need to replace a VPP. It can sit beneath one.

A VPP brings distributed assets together so their flexibility can be managed as a portfolio and offered to markets or system operators. Large Energy Model® works closer to the physical network, allowing participating devices to respond according to the electrical conditions and constraints they encounter at their own points of connection. The two layers can therefore solve different parts of the same problem: one coordinates the resource as a portfolio, while the other determines how that flexibility is expressed across the network.

VPPs aggregate capacity. Large Energy Model® distributes intelligence.

Deployment becomes part of learning

The same problem confronts grid foundation models. The effect of an intervention depends on where it happens, the topology of the network, its operating state, the devices nearby and the conditions at that moment. No historical dataset is likely to contain every combination that matters. Some knowledge can only be acquired by continuing to observe the physical network as it changes.

That changes the role of deployment. Instead of treating training as something that happens before a model enters the field, deployment becomes a source of new experience. Nodes sense, act within defined limits and measure what follows. Those observations can then inform subsequent learning under physical supervision. Over time, the deployed system builds a record of interventions and responses that did not exist beforehand, because the experiences themselves had not yet occurred.

The significance of this becomes clearer at the edge of the network. Millions of electrical devices are being deployed as loads, generators, storage and controllable resources. Increasingly, they are also computers. They can measure local conditions, alter their behaviour and observe what follows. Because they are connected to the same electrical system, the consequences of those actions may also be visible beyond the device that initiated them.

At sufficient scale, this begins to look less like a model operating above the infrastructure and more like a learning system embedded throughout it. Individual nodes see only fragments of the network, but together they generate experience across different devices, locations, operating conditions and moments in time. Physics provides a common structure for learning from that experience, while bounded control authority determines how it can be acquired safely.

The progression is from predicting the grid, to representing it, to coordinating it, and ultimately to learning through it.