Métodos de aprendizagem de máquina aplicados à ciência do petróleo
Carregando...
Data
2025-03-17
Autores
Barboza, Maria Carolina da Vitória Alvarenga
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Federal do Espírito Santo
Resumo
This study aims to present a new machine learning approach to classify crude oil samples based on their physicochemical properties, such as sulfur (S) concentration, total acid number (TAN), and API gravity (American Petroleum Institute). Crude oil is a complex mixture predominantly composed of carbon and hydrogen substances, along with heteroatomic elements such as nitrogen, oxygen, and sulfur. This complexity makes precise analysis essential, especially to avoid problems throughout the production chain. Proposed method seeks to overcome the limitations of traditional techniques, which are often time-consuming, require large sample volumes, and use excessive solvents. As a promising alternative, spectroscopic techniques have been employed for crude oil characterization, and machine learning methods have demonstrated high efficiency in analyzing complex mixtures. These methods offer faster and more accurate exploration of chemical variability compared to conventional approaches. This study, 196 crude oil samples, varying in sulfur content, TAN, and API gravity, were analyzed. The use of SVM (Support Vector Machine) ensembles was explored as a powerful approach to improve classification performance by reducing the variability of individual models, increasing robustness against overfitting, and enabling better generalization than a single model. To evaluate performance, criteria such as sensitivity, specificity, error rate, Matthews correlation coefficient, and accuracy were used, comparing SVM ensemble models with PLS-DA and standard SVM. The results demonstrated that the combination of NIR spectroscopy (Near Infrared Spectroscopy) with SVM ensemble models is an efficient and reliable method for the simultaneous qualification of sulfur content, TAN, and API gravity in crude oils. This is because SVM ensembles tend to perform better, reducing overfitting. Moreover, they create more robust models, reduce variance, and increase model stability.
Descrição
Palavras-chave
SVM ensemble , Petróleo , NIR