Energy was over 35% of product cost at a billet & long-steel producer. MetaDeep's custom ML models, trained on the plant's own historical sensor data, cut energy use 18% and peak-hour energy cost 22%, boosted line productivity 9%, and paid back in under 8 months.
A continuous-casting line ran on heavy energy costs (induction furnace electricity + preheating gas), yet energy decisions were made manually by operators — with no visibility into load changes, peak-hour price fluctuation, or input material quality.
Architecture: IoT sensors / MES data → ingestion → time-series store → prediction models → optimizer → dashboard / signals to controllers.
Three custom models (trained on 2 years of that plant's sensor data):
The optimizer balances load vs energy subject to quality constraints (temperature, melt chemistry, solidification rate).
| Metric | Before | After | Change |
|---|---|---|---|
| Specific energy (kWh/t) | 640 | 525 | −18% |
| Line productivity | baseline | up | +9% |
| Peak-hour energy cost | high | reduced | −22% |
| 24h forecast accuracy | — | 94% | new |
| ROI | — | 8 months | — |
Success came from customizing models on the plant's own data — real equipment behavior rather than industry averages.