The cold-storage model I built at Pegasus began with facility floor plans, refrigeration loads, and component efficiencies. It could estimate energy use under a set of assumptions and show how the estimate changed when those assumptions moved.
Then there is the question after a forecast misses. Did the building admit more heat? Did the equipment perform worse? Did operations change? Several parameter adjustments can make the same electricity trace fit. They imply different interventions and different investments.
This is where my interest in physics-based AI becomes concrete. I want a model that can identify what it still needs to observe before recommending an expensive physical change.
Two explanations, one meter
Take door traffic and insulation. Both can increase the refrigeration load. An electricity meter may not distinguish them. A record of door-open times, outside conditions, and internal temperature can make the hypotheses less interchangeable.
A simplified thermal balance illustrates the ambiguity:
Here C is an effective thermal capacity and U summarizes heat transfer through the envelope. This is an illustrative model, not a calibrated description of a particular facility. If several terms are unknown, improving the fit to T or electricity use can leave their individual values poorly identified.
A flexible learned residual can reduce error while making the diagnosis even harder. It may absorb a persistent operating change that an operator would prefer to see explicitly. The useful representation depends on the decision: forecasting tomorrow’s consumption and choosing a retrofit are not identical jobs.
Choose a measurement by its consequences
Suppose the candidate actions are changing a door schedule and improving insulation. The model should keep a set of plausible explanations, propagate each to the consequences of the actions, and ask which observation could change the preferred choice.
An observation that precisely estimates an irrelevant parameter has little value here. An imperfect door-use trace could be more valuable if it separates the two actions. The measurement policy should account for installation, downtime, technician effort, and the delay before a decision can be made.
There is also a stopping rule. If every plausible model recommends the same intervention, more identification may be unnecessary. If the models disagree but no affordable observation resolves the disagreement, the system should expose that limitation rather than manufacture a precise recommendation.
The physical model is a hypothesis, too
A physics-based structure can rule out nonsense and reduce the amount of data needed. It can also be wrong. A missing heat pathway, a changed control loop, or an unmodeled operating regime can make confident parameter estimates misleading.
The system therefore needs a model-discrepancy check alongside parameter uncertainty. Repeatedly finding that every admissible parameter setting misses the same pattern is evidence to revise the model family. Sending an optimizer back into the same family does not fix a missing mechanism.
I would compare a simple mechanistic model, a flexible predictor, and a hybrid under an equal observation budget. Give each an explicit option to request a measurement and evaluate the resulting intervention. Include a specialist system-identification approach and an operator’s measurement choice.
The experiment I would run
Use several thermal systems with similar geometry and different hidden operating conditions. Fit from limited initial observations. Then allow a fixed budget for additional sensing before choosing an intervention. Hold out a regime that the calibration did not include.
Report total decision cost, including sensing and delay, along with intervention error. A lower forecast error is useful evidence, but it does not settle whether the recommended repair was right. Also report how often the system correctly recognized that the available measurements could not distinguish the candidate explanations.
The energy model is work I did. This active-measurement extension is a proposal. The part I find ambitious is making a computational model responsible for the next physical observation it needs, and then holding it accountable for the decision that observation was supposed to improve.