Predictive models with the process physics built in
Anticipating a failure, a rise in consumption or a batch out of tolerance takes a model that knows how the equipment behaves, not just what happened before. We build models that combine the laws of the process with the SCADA, MES and maintenance data the plant already has, and that explain every alert.
Why most predictive models never make it onto the plant floor
A model trained only on history gets it right as long as the plant resembles that history. The alloy changes, a different scrap batch comes in, a burner gets replaced, and the model keeps predicting the same as always. Nobody notices until the alert arrives late, or doesn't arrive at all.
The second reason is that the prediction doesn't say why. An 87% failure probability with no reason behind it doesn't help the shift supervisor decide whether to stop the line. And the third is that nobody monitors the model after it goes live, so it ages in silence.
How we build a model that actually gets used
Physics first
The furnace's energy balance, the tool's wear curve or the press's thermal dynamics go into the model as constraints. The data fine-tunes what the physics doesn't fix: losses, burner efficiency, non-linear effects.
Alerts with a reason
Every alert carries the variable that has deviated, how much margin is left and what to check. Whoever receives it can verify it against the process before acting.
Built on what you already have
Reading SCADA, historian, MES and CMMS over the usual protocols (OPC UA, Modbus, exports). No new sensors, unless the diagnosis shows a signal is genuinely missing.
Monitored in production
The model's error is checked every week against what happened. When it drifts, it's retrained. That's part of the service, not an add-on.
What gets predicted on a plant floor
Gas and electricity consumption
How much a furnace or a thermal line is going to consume with the product coming up, and what setpoint would bring it down. It's the foundation of our energy efficiency projects.
Failures in critical equipment
Deviations in temperature, vibration, pressure or electrical consumption that precede a breakdown in furnaces, presses, compressors or motors, with days of lead time to plan the shutdown.
Rejects and defects
Which combination of temperature, time, material and tooling is going to produce an out-of-tolerance part, before it's made. It shows up most in forging, casting and heat treatment.
Production and demand
How much is going to be produced and how much will be ordered, feeding into shift and energy-purchasing planning.
A case with real plant data
At Reinosa Forgings & Castings, GASAI's predictive gas-consumption model draws on the furnaces' SCADA and years of operating records. It runs alongside the existing control system and is organised in three layers: a per-load consumption model, a furnace-occupancy planner and a deviation monitor that flags problems before they show up on the bill. It was validated across the five furnaces of a main forging line as part of SODERCAN's INNOVA programme. Read the case study.
From history to a model in operation
We check whether the records are good enough: variables, gaps, frequency, consistency between SCADA and maintenance logs. If they aren't, we say so and propose what to record. The seven signals we look at.
We pick a furnace, a press or a line and build the model with its physics and its data. It's checked against the operator before any alert goes live.
The model alerts and explains; the plant decides. We measure how many alerts were useful and how many weren't.
It's rolled out to similar equipment and the error is monitored every week. Retrained whenever the process changes.
Related: Method Energy efficiency Frequently asked questions Glossary
Common questions
What's the difference between a normal predictive model and one with the physics built in?
A model that only learns from historical data gets it right as long as the plant resembles that history; when the feedstock, the format or the equipment changes, it keeps predicting the same as always and nobody finds out until it fails. A model with the process physics built in (the furnace's thermal balance, the tool's wear curve, the press's dynamics) knows why the signal is changing and alerts when something falls outside what's physically expected.
What data do you need to get started?
SCADA, historian or MES records from recent months, maintenance and quality logs, and a conversation with whoever runs the equipment. Before modelling starts, we check whether that data is good enough: variable consistency, gaps, sampling frequency. If it isn't, we say so and propose what to record.
How does the prediction reach the plant floor?
As an alert with a reason: which equipment, which variable has deviated, how much margin remains and what to check in the next shift. It's integrated into the panel the plant already uses, or into a per-shift report. It doesn't replace the control system.
What happens when the model stops getting it right?
Its error is monitored every week against what actually happened. When the deviation crosses an agreed threshold, it's retrained on the new data. That monitoring is part of the service, because a plant changes and an unmaintained model ages in silence.
Does the data leave the plant?
It can stay inside, on a plant server or in a private cloud, depending on each client's systems policy.
45-minute exploratory meeting
We pick a process, review what signals exist, and decide whether it's worth a diagnosis. No cost, no obligation.
Request the meetingSee the method