A hydrocracker decides whether a refinery turns heavy feed into premium diesel and jet, or loses value to hydrogen overruns and product giveaway. It is also hard to control. Feed composition fluctuates, catalyst beds degrade at different rates, and the reaction dynamics resist the linear, fixed models that conventional advanced process control (APC) depends on. Manual adjustments and delayed lab results leave operations teams reacting hours or even days behind the process.
A controller learned from the plant's own historian data works differently. It updates setpoints minute by minute within safe operating limits and keeps the unit on its targets as conditions change. It runs on top of the existing APC and DCS, with operators able to take over at any time.
Data foundation and baseline KPIs
The model learns from the plant's history, so the data comes first. Collect at least six months of one-minute historian data and lab results. That window captures catalyst aging, feed swings and seasonal severity shifts the model has to understand.
The critical tags are:
- reactor temperatures and pressures
- H₂-to-hydrocarbon ratio
- catalyst bed ΔP
- feed and product assays
- heater duty
- recycle gas purity
- lab sulfur data
Most of these signals already sit in the distributed control system (DCS), the APC and the lab information system. The model writes supervisory setpoints on top of them without disturbing existing interlocks.
Set a clean baseline before starting: conversion, net hydrocracker margin per barrel, off-spec barrels and catalyst run length. Data cleansing and alignment continue throughout, so noisy tags do not drive poor decisions. Perfect data quality is not a prerequisite to begin.
Real-time reactor optimization
From historical data, the model learns how variables interact to affect conversion and product quality. That includes subtle interactions between temperature profiles, hydrogen ratios and catalyst behavior that are hard to see from the board.
After validation, a controller trained by reinforcement learning calculates temperature, quench and hydrogen targets and writes them to the APC or DCS within a defined safety hierarchy. It works to hold conversion while minimizing hydrogen cost as feed quality shifts. Feature-importance maps show operators which variables drove each move, so they can follow the model's decisions and check them.
Extending catalyst life
Nitrogen inhibition, metals deposition and uneven temperature profiles take activity from hydrocracker catalyst over time, forcing higher severity and shorter runs. A model trained on the unit's minute-by-minute history can spot early signs of those deactivation pathways. It can then adjust operating conditions to curb local hot spots and hold conversion without excess severity.
Four health indicators guide it:
- ΔT across each catalyst bed, which reveals developing hot spots before they accelerate deactivation
- Reactor pressure-drop progression, which signals the onset of fouling or coking
- Sulfur slip in product streams, which indicates changes in catalyst activity
- Hydrogen consumption per barrel, which tracks efficiency as the catalyst ages
With these signals, the optimizer recalculates the profit curve continuously. It eases severity when the marginal hydrogen cost outweighs the incremental conversion, and firms conditions again as cycle economics shift.
Because the model learns the catalyst's aging signature, maintenance windows and recycle gas cleanup can be scheduled ahead, and guardrails can tighten as the catalyst nears end of life. At each turnaround, the model's constraint weights should be updated to reflect the activity of fresh catalyst.
Feedstock flexibility
Every new crude batch shifts density, Conradson carbon residue (CCR), sulfur or nitrogen enough to upset a hydrocracker tuned for yesterday's feed. With feed descriptors such as density, CCR, sulfur, nitrogen and aromatics as inputs, the model adjusts reactor temperature, hydrogen partial pressure and quench rates within minutes of a feed change. Manual trims driven by delayed lab data take far longer.
Before running an unfamiliar feed, operators can benchmark recent performance, introduce the new blend gradually, monitor live conversion and product quality, and check results against economic targets. Handling these adjustments automatically lets a refinery run a wider crude slate without giving up stability, which matters more as crude prices and supply become volatile. For the wider picture, see feedstock flexibility.
Reducing transition and off-spec losses
Every transition, whether a feed change, a new product slate or catalyst regeneration, pushes the unit away from steady state and invites off-spec product. Conventional controllers chase the disturbance, and by the time adjustments arrive, giveaway is already in the tank.
A learned model turns each transition into a planned, predictive sequence:
- Preload setpoints that reflect the coming feed and the constraint hierarchy.
- Ramp deliberately, respecting pressure, temperature and compressor limits.
- Update in real time as actual conditions diverge from the forecast.
- Review afterward, so the next transition starts from a better plan.
With live product prices as an input alongside sensor data, the model can steer the ramp toward the most valuable product cut.
Early warning and reliability
A model that has learned the unit's normal behavior can surface subtle deviations before conventional alarms trigger. That gives operators time to act before off-spec product or equipment stress. Typical detectors cover:
- Heat exchanger fouling, flagged by a drop in exchanger UA, the overall heat-transfer coefficient times area, within a shift
- Compressor surge precursors, flagged when suction pressure oscillations exceed normal variation
- Accelerated catalyst deactivation, flagged by upward trends in reactor temperature differentials
- Abnormal temperature distribution, flagged by unusual bed-to-bed deviations
These alerts can feed maintenance planning, so work orders or inspection rounds are scheduled before a small irregularity becomes downtime. Combined with live constraint checks, they help preserve catalyst life, protect rotating equipment and support plant reliability.
How the Imubit Platform applies
Imubit applies Map, Model and Control to hydrocracker yield as part of its work in refining. Map traces lost conversion, hydrogen overruns and transition giveaway to the recurring decisions behind them. Model learns the unit's nonlinear dynamics, catalyst aging and feed response from its own history with Deep Learning Process Control® (DLPC®), so engineers can see the result of a change before they make it. Control takes those decisions to closed loop, usually after a period in advisory mode while trust builds, with the controller writing setpoints to the existing APC and DCS, inside limits set by plant engineers.
