MLEDGE — Federated energy-efficiency supervision for steam boilers

Steam boilers sit at the operational and energy-cost core of most traditional industries, but every plant treats their telemetry as commercially sensitive — and rightly so, given that production volumes and energy intensities can be inferred from it. MLEDGE’s traditional-economy use case set out to prove that boiler operators can still benefit from collective intelligence: trained models that draw on data from competing plants without any of those plants ever exchanging records.

Together with Inmarepro S.L., the industrial-boiler specialist that led the workpackage, Acuratio designed and built the federated machine-learning models that supervise efficiency across four anonymised installations — covering cosmetics, food, industrial laundry, and pharmaceutical production. Real PLC-acquired telemetry from each site (gas flow, feed-water temperature, flue-gas temperature, oxygen probe, steam pressure) is sampled minute-by-minute and used to predict the next five minutes of steam output. After iterating on architecture, the final model chains two LSTMs — one trained on the full feature set to project the first two minutes, then a vapour-only model that extends three minutes further — which together remove the apparent lag that single-shot predictors suffer at longer horizons.

The federated training weights each node’s contribution by dataset size, so installations with sparser operating windows don’t dilute the global model. The pay-off shows up clearly on the harder sites: at installation US_4, whose gas-flow sensor lost data for seven weeks, the federated model improved R² from 0.44 (local) to 0.68 and dropped RMSE from 368 to 277 — a result you simply can’t get from a model that only sees that single boiler. The food-production site (US_2) saw a similar improvement, with R² rising from 0.25 to 0.59.

Acuratio’s contribution covers the data-cleaning pipeline (idle-state filtering, gap interpolation, Modbus-status fault detection with automated SMTP alerts), the LSTM model family, and the federated training and aggregation. On top, Inmarepro can run cross-boiler “what-if” simulations: apply a model trained on best-in-class equipment to another site’s operating conditions and quantify the production gap, turning the federated layer into a concrete business case for energy-efficiency investments. MLEDGE is funded by the Spanish Ministry of Economic Affairs and Digital Transformation through EU NextGeneration-EU; the work was reviewed by IMDEA Networks.