Digital twins for power plants: three kinds, and which one answers an operating question

The geometric twin is a three-dimensional representation of the plant, usually from laser scanning or a design model. It answers spatial questions: what is behind that wall, will this component fit, how do we plan an outage, how do we train a new operator on a plant they have not stood in. It is genuinely useful and has nothing to say about process behaviour.

The data-driven twin is a statistical model trained on historian data. It learns the relationships that held while the data was being recorded, and it is very good at recognising when current behaviour departs from those relationships. It does not know why any relationship holds.

The hybrid twin combines first-principles models of the physics with data-driven components that account for what the physics does not capture. What that looks like on an electrochemical asset is set out in what a digital twin of an electrolyser actually models. It is the most demanding to build, because it requires the underlying mechanisms to be modelled rather than inferred, and it is the one that remains meaningful outside the conditions in the training data.

All three are sold under the same word, frequently by the same vendor, and the distinction is often not made explicit in a proposal. The third is worth defining properly, which is done in what a hybrid digital twin is.

A data-driven model is an interpolator. Inside the envelope of conditions it has seen, it is accurate and cheap to build. Outside that envelope it still returns an answer, and the answer carries no signal that it is now extrapolating.

A plant following solar or wind spends a large share of its life at operating points that are, by construction, unusual. Partial load, transitions, start-stop cycles and unfamiliar combinations of temperature and load are exactly where the interesting failures originate and exactly where a purely statistical model is least trustworthy.

This is the practical origin of the false-alarm problem. A statistical model flags a departure from its training distribution, which is not the same thing as a fault. Operators learn quickly that most alerts are the model being surprised rather than the plant being wrong, and once that lesson is learned the system is finished regardless of its accuracy on the cases that mattered.

A physics layer changes what an alert means. If a mechanism can be identified that would produce the observed pattern, the alert is about the plant. If no physically possible mechanism fits, the alert is about the measurement chain or the model rather than the process, and the next step is to find which: an unmeasured stream, a boundary drawn in the wrong place, clocks that disagree, or physics the model does not carry. That distinction is what makes the output worth an operator's attention.

Design-stage simulation and operational twins are different products with different requirements, and tools built for one are often proposed for the other. A design tool can take hours to solve and be used by a specialist. An operational twin has to keep up with the plant and be used by someone with a shift to run.

The operational questions worth building for are narrow and concrete. What is the condition of this asset now, including variables no sensor reports directly. What will it be in weeks if nothing changes. What is the best operating point given today's conditions and constraints. And what should be done in the next few hours.

Each of those has a decision attached. A twin that answers none of them, however faithful, produces a display rather than an outcome.

Tag mapping. A plant of any age has thousands of tags with inconsistent naming, missing or wrong units, undocumented sign conventions and duplicates from successive control system upgrades. Establishing what each signal actually represents is unglamorous and routinely consumes more of the schedule than the modelling.

Sensor trust. Models are calibrated against instruments that drift, and a model tuned to a drifting instrument will confidently reproduce the drift. Distinguishing instrument error from process change is a prerequisite rather than a refinement, and it is one of the things a physics layer is good at, because a set of readings can be checked for mutual consistency.

Validation against operation rather than design. A model that matches the design case and has not been tested against a year of real operation, including the awkward periods, has not been validated in any useful sense.

Drift and ownership. Plants change. Setpoints move, equipment is replaced, operating strategy shifts. A twin with nobody responsible for keeping it aligned degrades quietly until it is wrong often enough to be ignored. This is usually an organisational failure rather than a technical one, and it is worth naming in the contract.

No decision attached. The most common outcome is not a wrong model. It is a correct model whose output nobody acts on, because it was never clear which decision it was meant to change.

Start with one asset, one failure mode and one decision. Choose a failure mode that has actually cost the plant something, so the value of earlier warning is already known rather than estimated.

Fix the baseline before starting. How often does this happen now, how much notice is there now, what does it cost now. Without those three numbers the pilot cannot be judged, and pilots that cannot be judged tend to be repeated rather than scaled.

Agree what the output looks like. A number, a chart, an alert and an instruction are different products with different adoption rates. An instruction that names the asset, the cause and the action is used; a chart is looked at once.

Establish data access early, because it is the long pole more often than the modelling. Reading from an existing historian or control system interface without modifying anything is covered separately, and it is usually the first thing a controls engineer wants settled.

Then decide what breadth looks like before you need it. A twin that works on one asset and has a defined route to the next is a programme. One that starts with the whole plant is usually a pilot that never ends.