Closed-Loop Physical AI for LNG

Every team on an LNG train makes sound decisions against its own constraint. Operators hold margin against the hottest afternoon, the treating section runs extra reboiler duty to stay clear of the carbon dioxide limit, dehydration beds switch on a timer set for the worst case, and compressors carry recycle to stay away from surge. Each call is reasonable on its own. Together they hold the train below what it has already shown it can produce, and the losses concentrate in a few places.

• Refrigeration that does not follow the conditions. On a mixed-refrigerant train, the best composition of nitrogen, methane, ethane and propane moves with the feed, the load and the weather, while the target is often set for the season. On a cascade train, the same is true of how load is shared between the refrigerant circuits.
• Compressor recycle. Parallel machines drift apart with fouling and wear. One reaches minimum flow and opens its recycle valve while its neighbor still has room, and driver power goes into flow that adds no refrigeration.
• Pretreatment run with margin. Amine circulation and reboiler duty are held high against feed swings, and molecular sieve beds regenerate on fixed timers. The energy is spent whether or not the gas needed it, and regeneration gas couples the two sections.
• Upstream conditions that reach the cold end. A small shift in acid gas loading or scrub column separation shows up later as a wider temperature approach in the main cryogenic heat exchanger, where its cause is hard to see.
• Boil-off gas in transients. Loading, cool-down and mode changes produce vapor faster than the recovery system is staged for, and the excess goes to fuel or flare.

The same gap shows up across safety, reliability, margin, compliance and competency. Compressor trips and moisture breakthrough are reliability exposures, fuel gas and flaring set the emissions per tonne of LNG, and the judgment that keeps a train steady through a hot afternoon sits with a small number of experienced board operators.

Map

Map traces each gap to the recurring decision behind it: who makes it, how often, against which KPI and with what information. On an LNG train these include the refrigerant composition target, how load is shared between compressors, the rundown temperature, amine circulation and reboiler duty, the timing of dehydration bed switches, and tank pressure targets through a loading. The result shows which gaps carry the most impact and which decisions are ready to model.

Model

The models are built with Deep Learning Process Control® (DLPC®). Process engineers build them in DLPC® Builder from the plant's own historian, with no Python and no machine-learning background. A Deep Learning Predictor learns a nonlinear dynamic model of each section from its history, guided by first principles.

Years of process control leave a train with an operating history that looks much the same from one month to the next, and that is a real risk for any model that learns from data alone. Each model is built for one train on that train's own data, guided by first principles, and your engineers review every relationship it learns before anything moves. Where the history is held too tight to learn from, the work says so.

Operations, planning, maintenance and process engineering then work from the same model, so a composition change or a new rate target can be judged on its predicted effect before anyone commits to it. DLPC® Sim App runs the controller offline, so a move can be tested without touching the plant. See the result of a change before you make it.

Control

A Deep Learning Controller, trained by reinforcement learning across millions of simulations, learns the control policy for the section. It runs on-premise and writes setpoints to the APC or DCS minute by minute, inside the envelope your engineers set. The DCS, and the APC where there is one, keep handling regulatory control and stabilization, and anti-surge protection and the safety systems keep their roles. The controller recalculates the targets above them as ambient temperature, feed and equipment condition change.

Operators can take over at any time. Plants typically start in advisory mode, comparing recommendations with their own calls, and move to closed loop as the model earns trust.

Liquefaction and refrigeration

Refrigeration sets the output of the train. On a mixed-refrigerant train, the composition of the refrigerant decides how efficiently compressor power becomes cooling duty in the main cryogenic heat exchanger. The composition moves slowly, it follows from the make-up, the inventory in the loop and the split between the light and heavy streams, and its best value shifts between day and night and between summer and winter. A model of the cycle moves the composition targets and the compressor loading together, watches and predicts compressor discharge conditions and exchanger temperatures, and keeps each inside its limits.

On a cascade train, each refrigerant is a single component, and the question becomes how the load is shared between the circuits and how each compressor is loaded as ambient temperature and feed change. The method is the same. See how process dependencies limit LNG capacity.

Refrigerant compressors

Recycle is the clearest loss in refrigeration service. When a recycle valve opens, the compressor keeps consuming driver power while the recirculated flow adds no refrigeration. The best loading pattern keeps the whole system away from avoidable recycle while keeping capacity where the train is most likely to bind, even if individual machines carry unequal loads. The controller works above the dedicated anti-surge protection, which keeps its role. See refrigerant compressors in LNG plants and compressor surge.

Acid gas removal

The treating section has to hold carbon dioxide and hydrogen sulfide in the treated gas below the limits the cold section can accept. Circulation rate, reboiler duty and lean amine temperature are usually set with margin against feed swings and foaming. A learned model of the absorber and regenerator trims that margin toward the lowest reboiler duty that still holds the specification. See acid gas removal for LNG and gas processing and amine treating in LNG service.

Dehydration and mercury removal

Molecular sieve beds dry the gas ahead of the cold section, and guard beds remove mercury. Fixed cycle timers regenerate beds that still have capacity and spend fuel and adsorbent life doing it. A model of bed behavior, and of its links to the treating section upstream, supports switching by condition and steadier regeneration. See molecular sieve dehydration in LNG operations and mercury removal and guard bed management.

Storage, boil-off gas and loading

Tank pressure, boil-off gas compressors, the recondenser or reliquefaction, and the loading line respond to each other through every operating mode. Coordinating compressor staging and tank pressure targets through a loading recovers more vapor and sends less to flare. See LNG storage tank pressure and boil-off gas and reducing boil-off gas losses.

Imubit has over 100 closed-loop controllers deployed worldwide, in refineries, chemical plants and kilns. That is the largest install base in the industry, backed by 7+ years of closed-loop engagement in production, and it sits beside 300+ applications deployed, from inferentials to closed loop.

An LNG train poses problems those units already pose. A refrigeration compressor reaches its limit sooner as ambient temperature rises. A composition cannot be measured fast enough to steer by. Handles have to move together to hold a limit. At one refinery's naphtha and jet complex, with three fractionators, closed loop on top of existing APC delivered a 2 to 3% increase in jet recovery. See proof from production.