(530e) An Intelligent Agent for On-Line Optimization of Gasoline Blending

An intelligent agent for on-line optimization of
gasoline blending

C. G. Pernalete ^1,2, J. Dávila ²

¹Gerencia Técnica de Refinación.
PDVSA Intevep. Venezuela.

² Centro de Simulación y Modelos
(CESIMO). Universidad de los Andes. Venezuela

Abstract

Gasoline blending is a batch process where different intermediate
products of the refining process are mixed in order to produce a final product
to meet some quality specifications. The most important standard specifications
for gasoline include: RVP (Reid Vapor Pressure), AKI (Anti knock index), RON
(Research octane number), and D-86 distillation as leading properties. It is
important to notice that each feedstock differs from the others on its
properties and cost. Therefore, in order to reach a finished product that
respects the aforementioned constraints and using the available volume of each
feedstock, an optimization problem must be resolved, normally off-line, that
is, previously and without intervention in the actual blending process in real
time.

In this off-line optimization process two important assumptions are
normally made: firstly, that the feedstock quality parameters are constant
within a certain period of time for the blending and, secondly, that the models
that describe the blending processes are linear or, at most, require only some
non-linear adjustments. These assumptions, together with the possible
heterogeneity of each feedstock inside the tank, caused by compounds separation
because of density differences, often cause the production of a blend to end at
some distance from the expected quality , generating high additional costs
because of re-blend or quality giveaway: products with better quality than
required.

In this work, it was designed and developed an optimizer that
implements a novel strategy for executing real time optimization in gasoline
blending processes. In order to account for the problems of process-model
differences, the optimizer in this work uses blending models automatically
obtained from machine learning experiments, specifically learning tasks with
Support Vector Machines (SVM). For this purpose we took advantage of the large
amount of operational blending data available in a Venezuelan refinery to
produce different grades of gasoline.

The basic elements of any strategy for optimization of the gasoline
blending process are: 1) the distributed control system (DCS), which executes
the control actions on the plant, 2) the optimizer, which makes the
calculations using current information from the plant, 3) the on-line analyzer,
which determines on-line the feedstock and product properties, and 4) the tank
information system (TIS).  All of them must and have been considered in order
to implement the solving application. All these operational constraints suffice
to consider the on-line optimization system as a complex software application
that must cleverly interact with other elements of the system during the
optimization process. Considering that an intelligent agent is a computational
system located in an environment and that is able to behave in an autonomous
and flexible manner in such environment in order to reach its designed
objectives, it was decided to develop the gasoline blending optimization system
as an intelligent agent.

Summarizing, there is evidence that the modeling of blend properties
for an specific refinery using support vector machines, is a promising
technique as the average error in our experiments is always below that obtained
by conventional methods. Also, we can show that the agent developed and tested
in this work can deal efficiently with the calculation of the specific
optimization problems. In our experiments, a solution was always found and the
convergence time was in the neighborhood of one minute. The agent was also
tested in simulations to evaluate its capacity to deal with perturbations in
the product quality during the blending process and the results are enticing.
Based on conservative estimated explained in the paper, we estimate that, by
using  this intelligent agent for the on-line optimization of gasoline blending
in a medium capacity refinery of the National Refining Circuit of PDVSA,  the
system could generate savings in the order of 5MM$/year.