Performance Optimization
Real-Time Optimization finds the true operating optimum for an asset or plant. An Operational Performance Twin and an AI Optimizer adjust operating variables with the guidance of an AI Agent within real plant constraints.
A simulation of machine and process behavior, built from the plant’s own historical operating data using AI together with first principles of physics. It responds to a change the way the real unit would.
The optimizer carries the objective function (production, yield, profit or energy) along with the constraints: parameter limits and the limits of downstream equipment. It searches for the best achievable operating point inside those bounds. The GenAI Agent identifies the best possible manipulated variables from the AI Optimizer at that moment, and changes them in the Operational Performance Twin to test whether the objective function improves.
After a few iterations between the AI Optimizer and the Operational Performance Twin through the GenAI Agent, the optimum new set points of the manipulated variables are identified, ready to be implemented in the DCS.
In open loop, RTO sends the optimum set points to operators for review and implementation in the DCS. In closed loop, the same set points are pushed directly to DCS controllers or the APC.
A Middle East refiner running a 105 kbpd FCC unit on design-era set points. Feedstock, demand, and plant behavior had all shifted since, leaving the unit short of its true potential, with no safe way to find where optimal was.
A physics- and data-driven plant model paired with an AI Optimizer that pushes independent variables to their limits within constraints. It runs every four hours, and on any major shift in feed or conditions, delivering set points to the DCS engineer in open loop.
$8–9M per year in added benefit, with the model also serving as a what-if tool for feedstock and demand changes.
The operator was losing energy to power curve deviation, the gap between reference and actual output. Closing that gap meant running each turbine harder while staying inside OEM-certified load limits and site curtailment rules.
An Operational Performance Twin replicates turbine performance and behavior using physics plus ML, including the power curve. What-if and waterfall simulation run on the twin before any change goes live, and the AI Optimizer tunes manipulated variables (pitch angle, torque set point, rotor speed, yaw angle, cut-in/out thresholds and ramp-rate schedules) until AEP peaks or a constraint is hit.
3% more energy generated with lower power loss, delivered within OEM load limits and site curtailment rules.
At a 2.6 MTPA plant, the 76-metre rotary kiln (6,000 TPD clinker) is the largest energy consumer. Specific heat consumption sat around 820 kcal/kg clinker, and variable fuel quality and feed chemistry drove burn-zone instability, clinker quality swings and stability-driven interruptions that existing DCS/APC control could not adapt to.
A Kiln Performance Twin (hybrid ML plus physics replica of preheater, calciner, kiln and cooler), an AI Optimizer running Bayesian search for minimum energy per ton within quality and stability constraints, and a GenAI Agent that picks safe, effective levers each iteration from live context and plant SOPs. Optimum set points go to operators open-loop, with what-if analysis on 250+ DCS and lab variables.
2.5% lower specific heat consumption against the 820 kcal/kg baseline, 6% higher TSR without impacting kiln stability, 25–35% less clinker quality variability, 25% better Kiln Stability Index, and throughput held at 6,000 TPD or above with 1.5% upside.
A 3x 40,000 MTPY chlorine plant, where the electrolyzer alone drove ~50% of variable production cost. Cell health and replacement timing added a second exposure: downtime and lost output.
A predictive model of the electrolyzer and brine system paired with an AI Optimizer minimizing energy within real plant constraints, delivering optimal set points each cycle. Cell health models set replacement timing by weighing capex against energy cost, not a fixed schedule.
Energy per ton fell 5%. Production rose 2%, from healthier cells, not harder running. Net benefit: $3M/year, a 20x return, payback under three months.
Plant output drifts below the expected generation curve as soiling, module degradation, inverter derating and tracker misalignment accumulate, and the losses are hard to attribute or act on in real time.
An Operational Performance Twin replicates plant behavior using physics plus ML, including the expected generation curve for the prevailing irradiance and temperature. The AI Optimizer tunes controllable variables (tracker angle, inverter set points, reactive power and cleaning schedules) within grid and OEM constraints, and a GenAI Agent selects the effective levers each iteration, with optimum set points sent to operators open-loop.
About 2% higher energy yield against the reference curve, faster attribution of generation loss, and cleaning and curtailment decisions driven by modeled value rather than fixed schedules.
The HP section ran on design-era set points, leaving urea yield and production below what the equipment and process constraints actually allow.
A Plant Predictive Model, a hybrid simulation replicating HP section behavior using data science and first-principles physics with autotrain capability, paired with an AI Optimizer whose objective function is urea yield and production. The optimizer manipulates NH₃ and CO₂ feed rates, NH₃/CO₂ and H₂O/CO₂ ratios, pre-heater and reactor temperatures, reactor B outlet pressure and recycle stream flow within equipment and safety constraints, iterating until a constraint hits or yield stops improving. Optimum set points go to operators open-loop for implementation in the DCS.
Increased urea yield and production, higher gross profit with lower energy consumption in the optional objective case, and identification of bottleneck areas for debottlenecking.
Enhanced oil recovery on a deepwater subsea tieback depends on how gas lift is injected at the riser base. Operators had no real-time view of the optimum injection rate, and slug formation in the gas lift riser interrupted flow and forced conservative operation well below the production the system could sustain.
Three models run against live SCADA data: an EOR Predictive Model relating oil production to reservoir parameters and choke valve outlet pressure, an AI and physics-based simulation model covering choke outlet through the riser base and gas lift riser to the host platform, and a host platform model for separation and gas processing. A Slug Prediction Model flags slug formation early, identifies the cause, and switches the AI Optimizer’s objective from oil production maximization to gas lift flow while the upset is mitigated. The optimizer manipulates outlet pressure, gas lift flow, host separator pressure and temperature, compressor outlet pressure and other wells’ gas lift flow within gas processing capacity and operating limits, and optimum set points are implemented in SCADA.
2.8% more oil production, slug events predicted and mitigated before they interrupt flow, and injection decisions made from live data rather than periodic sampling at remote locations.
