MACHINE LEARNING APPLICATION TO ASSESS A PROCESS CON- TROL OF A CATALYTIC CRACKING UNIT

Resumo

Catalytic cracking is extensively applied in the downstream oil and gas industry to process an oil charac-
terized by a high value of API degree, in other words, petroleum with a high percentage of heavy hydrocarbons,

resulting in many products with smaller molecular weight. The biggest challenge of this process is the evaluation
and control of catalyst deactivation, a phenomenon characterized by inducing the inactivity of disponible porous

regions of the catalysts that take place the chemical species transfer or reagent to inside the porous, after the chem-
ical reaction discharge to external medium products with a less molecular chain. One alternative workable in this

industry is a continuous regular substitute of deactivated catalyst, using the chemical or pyrolysis reactivation and
returning to the chemical process. To ensure the efficiency is crucial to determine the substitute frequency and
the amount of reactivated catalyst to maintain the maximum yields of the process. To analyze aspects of catalyst
deactivation, a system control project that aims to ensure the conversion obtained at the reactor and the flow of
the reactivated catalyst is required. Since both variables are explicitly important to the problem, it is possible to
define the optimal set-point for system control by monitoring these variables. So, a strategy based on ratio control,

that uses mathematical modeling to obtain the set-point to the conversion and flow of the reactivated catalyst, clas-
sification as control and manipulate variables, respectively. Thus, get the suitable ratio in accord at the set-point

of the process. As a way of evaluating and optimizing the flow of catalyst employed, after the projected control
system, it will apply a model of machine learning, Support Vector Machine (SVM), a supervised method that with
a data set prescribe a hyperplane and assess each point of distance on relation this plane for determining which
class best representing the data. This work aims to study which variables influence the deactivation catalyst and
suggest ways to mitigate and control them.