Predição do teor de óleos e graxas em água produzida em uma plataforma de petróleo, utilizando redes neurais artificiais

Abstract

The use of physicochemical analyses in industrial processes is an inseparable reality for the contemporary industry. Such determinations are efficient and necessary for verifying production quality, establishing consistent standards, and promoting safety. Furthermore, companies committed to sustainability strive to minimize their environmental impact, particularly concerning the emission of pollutants and changes that could alter the dynamics of the ecosystems where they operate. It is the responsibility of government authorities to regulate actions related to environmental preservation, aiming to harmonize practices and ensure the minimization of the mentioned environmental impacts. On August 8, 2007, CONAMA Resolution No. 393 was enacted, addressing the continuous discharge of process or production water in offshore oil and natural gas platforms. Article 5 stipulates: "The discharge of produced water must comply with a simple monthly arithmetic mean concentration of oils and greases (TOG) of up to 29 mg/L, with a maximum daily value of 42 mg/L." (CONAMA, 2007). The central objective of this study was to develop a predictive method for estimating TOG values, using process variables from an oil treatment plant and online TOG measurements. The study was divided into stages encompassing data acquisition and variable selection, database analysis and validation, definition of TOG classification limits, data balancing, and predictive modeling. Initially, the selection of the most relevant variables was carried out with the support of specialists, ensuring that the model accounted for parameters pertinent to the process. The database underwent a detailed exploratory analysis to address inconsistencies and validate the behavior of the variables concerning gravimetric TOG. Subsequently, limits were defined to create binary classes (compliant/non-compliant), and balancing techniques were applied to ensure a consistent dataset. The predictive modeling employed artificial neural networks, resulting in a robust model with 97.3% accuracy, validated through statistical metrics and generalization tests. The main contributions of this study include identifying the variables with the greatest impact on the process and proposing an alternative strategy based on data science for TOG monitoring. Additionally, the study emphasizes the importance of digital transformation in optimizing production processes and promoting sustainable practices.