Real time optimisation of the cat cracker.

The optimum moves, and a fixed target lags behind it.

The most profitable way to run a fluid catalytic cracker moves constantly. The optimum shifts with feed quality, catalyst activity and the relative prices of gasoline, LPG and distillate, and it sits against several competing limits at once. In practice operators hold fixed targets chosen for a typical day and leave margin on the table to stay clear of the wet gas compressor, the regenerator air blower and product quality. Conventional two step real time optimisation reconciles a steady state model and then solves an economic problem on a slow cycle, so by the time it advises a move the feed and catalyst have already drifted and the recommended point lags the true optimum.

A surrogate of the unit and an agent that trades the objectives.

EntroMetrix builds a physics informed surrogate of the unit from the plant historian and pairs it with a reinforcement learning agent that continuously re optimises set points across the competing objectives. The surrogate learns the riser and regenerator response from operating data and is checked against the mechanistic model before it is trusted. The agent trades gasoline and LPG yield, unit margin, energy, coke and carbon, and honours the full operating envelope through safe reinforcement learning rather than a penalty added after the fact. It returns a Pareto set of feasible operating points, from which the plant chooses one.

Built from the historian, checked against the rigorous model.

The optimiser is built from the unit's own historian. Tags across the riser, regenerator, fractionator and gas plant train a surrogate of how the unit responds, and the surrogate is checked against a mechanistic model before it is trusted. The operating envelope and the economics are set with the plant's engineers, and the assumptions are written down before any advice is given.

On conversion units of this kind the surrogate reproduces the rigorous model to better than one percent and re optimises every set point in under a second, against the minutes a two step optimiser needs, so it stays with the optimum as the feed drifts rather than lagging behind it. Across validation runs the policy held every hard limit without a single violation, because the limits are built into the agent rather than added as a penalty afterwards.

Advises, does not act: offline first, with the operator in the loop.

The work is retrospective and offline first. The agent is trained and tested on the plant's own history and validated before it advises, and every recommendation is delivered with the yield, margin, energy and carbon it implies, so the operator stays in the loop. The same surrogate and agent transfer to the catalytic reformer and the hydrocracker with unit specific retraining, so one method covers the conversion block.