Ensemble Forecasting in Twins
Running many perturbed twin copies forward produces a distribution of futures, quantifying forecast uncertainty rather than a single guess.
Many futures, not one
A single forward run of a twin gives one prediction, with no sense of how reliable it is. Ensemble forecasting runs many copies of the twin forward from slightly different starting states and parameters, sampled from the current uncertainty, and reads the spread of the results as the forecast uncertainty. The technique is borrowed from weather prediction, where it transformed forecasting into a probabilistic discipline.
How it works
- Sample the current state and uncertain parameters to seed a set of members
- Run each member forward under the scenario of interest
- Read the mean as the central forecast and the spread as the uncertainty
- Read the fraction of members crossing a limit as the probability of that outcome
Answering probabilistic questions
Because the output is a distribution, an ensemble answers questions a single run cannot: what is the chance a temperature limit is exceeded in the next hour, or how likely is a component to reach end of life before a planned outage. These probabilistic answers are what decision-makers actually need, since operating a plant is managing risk, not chasing a point estimate.
The cost, and how surrogates pay it
Running hundreds of members is expensive with full physics. Surrogates make ensembles practical by reducing each member to a fast evaluation, and because members are independent they parallelize cleanly. The ensemble Kalman filter uses the same idea for state estimation, so ensembles serve both assimilation and forecasting. See ensemble Kalman filter and surrogate models.
In the Kronos twins
Ensemble forecasting lets the Hyperion breeder and the burner twins report the probability of crossing operational and life limits under candidate operating plans, carrying forward the honest uncertainty that pre-construction models carry. Today the ensembles run on design models and validated surrogates; after the machines operate, the starting-state uncertainty will shrink as real data tightens the estimates. See uncertainty quantification and what-if scenario simulation.