Resumo:
The produced water generated by the primary oil processing carried out by offshore oil
platforms, which has a total of oil and grease (TOG), is usually reinjected or disposed
into the open ocean. This disposal is monitored by environmental regulatory agencies
that determine maximum TOG values. In Brazil, the gravimetric method is that
homologated for measuring TOG, which must be carried out in onshore laboratories.
Due to the logistics of transferring samples from the platform to the laboratory, the
measurement result is available approximately 20 days after the day of collection. This
work proposes the development of a predictive model of gravimetric TOG (TOG-G) from
process variables, added to a variable extracted from the response variable, which can
be used offshore and in real time, to more quickly guide possible preventive or corrective
actions in order to avoid its non-compliance. For this, the observations were grouped
into clusters associated with TOG-G ranges, through which the base balancing was
performed. Training and test sets were generated and a classifier was built for the cluster
according to the most significant process variables for the prediction of TOG-G,
identified through linear regression. Subsequently, the TOG-G was modeled from the
significant process variables and the cluster. The results obtained for the test set were
evaluated by means of Mean Absolute Error (MAE), Mean Absolute Percent Error
(MAPE), coefficient of determination (𝑅
2
) and Pearson's correlation coefficient (𝜌), and
showed to be superior both to the forecasts generated from the predictive model
developed from the same forecasters, but disregarding the cluster, as to the real values
of the spectrophotometric TOG (TOG-S) measurements, which constitutes the real-time
method currently used as a reference in the platform. To validate the gains in accuracy
with the proposed method, it was also applied to a classical set of linear regression for
predicting fish weight. Thus, the inclusion of the cluster information in the TOG-G model
proved to be an innovative and efficient approach to increase the accuracy of its
prediction from information available on the platform, which may considerably benefit
the oil industry in terms of process control.