Natural gas accounts for a large share of industrial energy use, according to the U.S. Energy Information Administration. Demand from new data centers and growing LNG exports is tightening supply and amplifying price swings, and combustion emissions face closer scrutiny.
For most process plants, fuel is one of the largest variable costs, so that volatility flows straight to operating margin. The task is to burn less gas per tonne of product without giving up throughput, safety or product quality. That takes an optimization layer above the existing advanced process control (APC), one that can see the trade-offs between units.
Why industrial gas use stays under pressure
Extreme weather, new LNG export capacity and data center construction all pull on the same supply. Cold snaps push consumption up across every sector, and production discipline among shale producers limits how quickly supply can respond. Competition for gas keeps prices firm even in the shoulder months.
Regulators and investors also expect credible decarbonization plans alongside reliable output. Every unit of gas a plant does not burn lowers both the fuel bill and the CO₂ emitted per tonne of product.
The five approaches below each target a different source of excess gas use. Each works on its own, and they compound when applied together.
Find the hidden losses
Natural gas slips away in small pockets that rarely trip an alarm. Burners misfire slightly, compressors run a little harder than the load requires, and heat exchangers foul just enough to raise firing without crossing an alarm limit. Each loss is small, but together they add up to substantial waste that conventional control does not see.
A model that brings sensor streams, lab results and historian archives into one analytical view can search far more tags than a team can trend by hand. It surfaces patterns people miss and flags subtle anomalies before operators would notice them.
Compression trains are a good example. A learned model captures the nonlinear interaction between suction pressure, discharge temperature and recycle flow, and adjusts setpoints so each stage runs near its best efficiency. The setpoints are written to the distributed control system (DCS) with production held as a constraint, so fuel use comes down without giving up throughput. Lower energy intensity shows up directly as lower CO₂ per tonne of product.
Adjust setpoints as conditions change
Static operating procedures leave energy unused because they cannot follow the plant as feed, ambient conditions and equipment condition change. A model that reads live sensor data can calculate new setpoints and write them to the DCS as conditions move, faster and more consistently than manual adjustment.
The work starts by streaming historian data and lab results to the model and training it on the unit's operating history. The model then runs in advisory mode, where operators review its recommendations, before the plant moves to closed loop, inside limits set by plant engineers.
Reinforcement learning (RL) controllers learn a control policy against a model of the unit, and safeguards check every move against operating limits before it is written to the DCS. The aim is tighter specification control, fewer manual interventions and less fuel per unit of product. Success depends on reliable data pipelines, clean integration with existing control systems, and operators who trust the model's recommendations.
Predict and prevent energy spikes
Start-ups, sudden feed changes and equipment trips can push gas demand well above steady-state needs. Those short bursts inflate the fuel bill and raise emissions per tonne of product. They often begin with small shifts in vibration, temperature or power draw that conventional alarms do not catch.
A model fed with those signals can detect drift early, before it forces a trip. It flags the anomaly, recommends a controlled ramp-down where one is needed, and calculates contingency setpoints so downstream units keep running while the root cause is addressed. Each avoided trip removes the fuel burned in the restart that would have followed, and supports day-to-day reliability and lower emissions.
Balance demand across units
Optimizing one furnace or compressor in isolation often pushes energy waste somewhere else in the plant. Heat, steam and fuel move through the whole site, so the model has to see the whole system: utility meters, flow tags and lab results across every unit involved.
With an economic objective on top, the model can search for the combination of setpoints that meets production targets at the lowest gas cost. In practice that can mean shifting load from a boiler running near its excess oxygen limit to a newer, more efficient unit, or timing high-pressure compressor starts to fall in off-peak tariff windows.
Balancing the system also smooths the demand swings that fatigue equipment. The practical hurdles are bringing historian data from every unit into a common context, working across organizational silos, and scaling models from one site to many. Data integration is the place to start.
Give operators clear recommendations
Control rooms stream thousands of data points every second. Operators juggle alarms, shifting feed quality and tight safety constraints, often with their own experience as the main guide. As experienced staff retire, that load falls more heavily on newer hires.
A recommendation layer translates historian and sensor data into clear next steps. It ranks possible moves by economic impact, attaches a confidence level to each, and lets operators see the result of a change before they make it. That guidance speeds up knowledge transfer and helps crews make better decisions faster.
Operators keep the decision, and the model gives every shift the same view of the trade-offs across the unit, which is the basis of human-AI collaboration in the control room. This is also the stage where crews build the confidence to let the controller act in closed loop within limits they set.
Working with existing control systems
This approach complements the existing control infrastructure. It streams data from the historian and lab into plant-specific models, then sends setpoints through the DCS, where operators keep override authority. Regulatory control runs as before. The optimization layer sits above it, coordinates across units, and turns static setpoints into targets that follow the economics. Industrial energy efficiency work gains most when that layer can see plant-wide trade-offs as well as individual loops.
Measuring the reduction
Specific energy consumption per tonne of product is the headline measure. Several others show where the savings come from:
- fuel use per operating hour
- excess oxygen in fired heaters
- compressor specific power
- avoided restart events
- CO₂ per unit of output, for decarbonization tracking and regulatory reporting
Track these alongside production output, so that savings are not coming at the expense of throughput. For the wider picture, see energy management in the process industries.
How Imubit applies this
In refining, chemicals and polymers, cement, and mining and minerals, Imubit applies Map, Model and Control to plant energy use. Map traces excess gas use to the recurring decisions behind it, such as firing targets, steam balance and compressor loading. Model uses Deep Learning Process Control® (DLPC®) to learn each unit's nonlinear behavior from its own history, guided by first principles. Control takes those decisions to closed loop, writing setpoints to the existing APC and DCS, inside limits set by plant engineers. Where energy is the optimization target, deployments have delivered a 15 to 30% reduction in energy consumption. Read more about the Imubit Platform.
