Playbook · updated September 1, 2026

Energy: the optimization nobody has to approve capex for

In shortML attacks energy three ways without capex: setpoint optimization (finding the low-energy settings that still hold quality), sequence/scheduling optimization (avoiding peak-tariff and idle-load waste), and anomaly detection on utilities (compressed-air leaks, stuck valves, equipment running out of spec). Program-style results run 5–15%; published ML results reach −20% on optimized lines.

1 · Baseline the spend

The energy calculator turns MWh and tariff into the annual number. Sub-meter by line if you can; models optimize what they can see, and one plant meter hides everything.

2 · The three ML levers

Setpoints: for many processes several setting combinations hold quality — they differ in kWh. Models trained on history find the low-energy corner of the quality envelope; published results around −20% (vendor-reported) come from this lever. Sequences: startup order, idle management, tariff-aware scheduling — the unglamorous lever that often pays first. Utility anomalies: compressed-air leaks and stuck valves are invisible in the bill and obvious to a learned baseline.

3 · Energy rides along

The strategic point: energy optimization uses the same data and the same platforms as quality and scrap work. Plants that deploy process ML for scrap get the energy lever nearly free — which is why 'energy only' pilots are usually the wrong scope.

Run the numbersPrice this loss with the What's on the table in energy? calculator, then take the ROI workbook into the budget meeting.