Hydrocracking sits at the center of refinery economics. The unit turns heavy vacuum gas oil into diesel and jet fuel, and the feed it receives keeps getting harder to predict, with heavier cuts, higher nitrogen and more cracked material. Margins are tight and product demand shifts. Control has to keep pace with all of it while respecting hydrogen cost and catalyst life.
How a hydrocracker works
A hydrocracker processes heavy vacuum gas oil through high-pressure, high-temperature reactors with hydrogen recycle and product fractionation. Each reactor holds catalyst beds with quench points between them to control the exothermic reactions.
The main control variables are:
- reactor inlet and outlet temperatures
- hydrogen partial pressure
- overall reactor pressure
- liquid hourly space velocity (LHSV)
- quench hydrogen rates
- bed temperature rise
- catalyst age
Bed pressure drop (ΔP) is the real-time fouling indicator. A spike points to a catalyst problem.
Performance is measured by conversion, middle distillate yield, hydrogen-to-hydrocarbon ratio, ΔAPI and energy efficiency. A useful model of the unit needs historian tags, lab results, catalyst histories and detailed feed characterization.
The unit's behavior is nonlinear, and it changes over time as catalyst ages and feed changes. That makes it a good fit for a model learned from the unit's own data, which can weigh yield against hydrogen cost continuously inside the operating boundaries.
Operating and safety constraints
Those boundaries are not negotiable. Any model on a hydrocracker must respect:
- the maximum temperature rise allowed for each catalyst bed. Older reactor shells can be damaged if a zone creeps past its design limit.
- reactor pressure limits that keep hydrogen in solution and protect welds.
- hydrogen purity high enough to avoid catalyst poisoning.
- linear velocity in high-H₂S piping between 10 and 20 ft/s to limit ammonium bisulfide corrosion, even as quench flows vary.
- product specifications for diesel sulfur, jet aromatics and gasoline nitrogen, which are hard limits.
Equipment alarms on pressure drop or flow bypass any optimization logic and trigger immediate operator or emergency action. Catalyst life adds another layer of constraint. Exceeding the allowed temperature rise by a few degrees can shorten a cycle or force an unplanned shutdown.
The model must also honor emergency depressurization protocols. If temperatures surge or hydrogen leaks, the priority moves at once from optimization to safe shutdown.
Where conventional control struggles
Hydrocracker control has relied on PID loops and advanced process control (APC). Both work well near the conditions they were tuned for, and hydrocracking moves away from those conditions often.
Steady-state assumptions
Control models are built around relationships that hold near a steady operating point. Large, fast swings in feed composition break those assumptions, and the controller adapts slowly to transients such as a spike in nitrogen or asphaltenes.
Catalyst aging
Conventional designs do not model deactivation ahead of time. They correct temperature or yield drift after it appears, and they are insensitive to uneven aging across beds. The result is frequent manual intervention to correct upsets.
Interacting variables
Temperature, pressure and hydrogen interact. Modern APC is built for multivariable control, but its models are identified around a set of operating conditions and lose accuracy as feed and catalyst move the unit away from them. In practice the unit runs with conservative margins, and much of the information in the plant's data goes unused.
How a learned model closes the gap
A learned approach adds inferentials, simulation and a trained controller on top of the existing control layers.
- Real-time inferentials. Models trained on years of plant data turn historian tags into immediate estimates of sulfur, aromatics and conversion, so the unit does not wait on lab samples. They capture nonlinear relationships that static correlations do not represent well.
- Simulation before action. A dynamic model that combines first-principles understanding with live plant data lets engineers test what-if scenarios before a change reaches the distributed control system (DCS). Engineers can see the result of a change before they make it, and search for operating points that balance reactor temperature against hydrogen cost.
- A reinforcement learning controller. A controller trained by reinforcement learning writes setpoints to the APC or DCS minute by minute, adjusting for ΔP excursions, catalyst aging and the trade-off between yield and hydrogen cost. Feature-importance views show operators which variables drove each move.
The aim is a controller that operators can inspect and verify, running inside the unit's safe envelope.
Where the value comes from
On a hydrocracker, tighter control shows up in middle distillate yield, hydrogen use and energy. Lower hydrogen consumption per barrel reduces exposure to hydrogen price swings and to carbon costs. Holding ΔP and quench rates in their preferred ranges supports steadier operation, fewer manual trims and catalyst cycles that are not cut short by temperature excursions. An anomaly detection layer can flag developing equipment problems early, which protects uptime.
Good practice for an AI project on a hydrocracker
- Start with people. Form a cross-functional team from operations, controls, planning and data science, so each model is grounded in the plant's reality.
- Define safety boundaries early. Use early workshops to surface hard limits. Keep emergency interlocks independent of the optimizer and consistent with validated safety requirements, so operators know the new logic cannot violate reactor temperature or pressure limits.
- Make the model transparent. Share feature-importance plots, and let operators run side-by-side drills, operator against model, before the loop is closed.
- Set KPIs up front. Define conversion, yield, hydrogen use and energy KPIs, and baseline at least ninety days of historical data so improvement can be measured without debate.
- Clean the data first. Poor data quality derails projects more often than algorithm choice does. A historian quality pass that filters sensor drift, fills sampling gaps and reconciles lab results should come before modeling.
- Keep testing the model. Hold a rolling six-month blind test set to catch overfitting, and retrain periodically so the model reflects catalyst aging and new feed slates.
- Address human factors. Operators accept the system sooner when training shows how it supports their daily decisions. Hold walk-throughs on every shift, publish a weekly performance scorecard, and share cases where the model flagged a developing ΔP excursion before the alarms did.
- Verify integration early. Prove read and write connectivity to the DCS during discovery, before contracts are signed. Confirm that data flows to and from the DCS without custom middleware, and start with one optimization objective before expanding.
- Value the project per barrel. Price the project on measurable margin per barrel rather than abstract ROI.
- Plan monitoring before handoff. Agree who watches the model around the clock, so an expert is available when the system flags unusual behavior at 2 AM.
How the Imubit Platform applies
Imubit applies Map, Model and Control to the hydrocracker as part of its work in refining. Map traces lost distillate yield and excess hydrogen use to the recurring decisions behind them, such as reactor temperatures, quench rates and hydrogen partial pressure. Model learns the unit's nonlinear dynamic behavior from its own history, guided by first principles, with Deep Learning Process Control® (DLPC®). Control takes those decisions to closed loop. A controller trained by reinforcement learning writes setpoints to the existing APC and DCS minute by minute, with operators able to take over from the board.
