Accurate knowledge of crude oil pressure–volume–temperature (PVT) properties is fundamental for industrial and academic applications. A recent study presents advanced compositional models developed through various machine learning (ML) techniques to efficiently and reliably predict the solution gas-oil ratio (Rs), one of the most significant PVT properties.
Traditional methods for determining Rs are not only time-intensive but also costly, with results heavily influenced by sampling quality. The researchers utilized a comprehensive database containing 1,154 data points to develop these models, showcasing the potential of modern data-driven techniques to streamline the prediction process.
Among the ML algorithms tested, the extra trees (ET) algorithm emerged as the most capable, achieving an average absolute percent relative error (AAPRE) of approximately 3%. This high level of accuracy positions ET models as powerful alternatives to conventional Rs estimation approaches, which had limitations related to accuracy and efficiency.
The study also evaluated Rs predictions based on seven different equations of state (EoS). Among them, the Schmidt-Wenzel (SW) EoS was identified as the most accurate conventional method, achieving an average error of about 11%. This comprehensive analysis highlights the advantages of the ML models relative to traditional techniques, providing insights for future research and applications.
Fossil fuels and particularly petroleum continue to have vast impacts on multiple aspects of life and industry. The accurate characterization of PVT properties, including Rs, is indispensable for reservoir engineering calculations. These include well testing, reservoir simulations, reserve estimations, and economic assessments. Proper data around these parameters benefits not just industry but academic studies focusing on fuel properties.
While traditional methods, such as differential liberation (DL) and constant mass expansion (CME) tests are considered reliable, they also entail substantial investments of time and cost. The research community has been exploring alternative approaches including equations of state, artificial intelligence, and empirical correlations to develop Rs models.
The study lays out findings from previous attempts to predict Rs using differing methodologies, including efforts to incorporate artificial intelligence techniques to address traditional model shortcomings.
Emerging from this recent study, proposals included novel models based on tree ensembles and neural networks. Researchers noted the need for advanced machine-learning techniques to account for compositional variations affecting Rs behavior, leveraging comprehensive data as input to refine predictive capabilities.
The choice of ML models was prompted by their proficiency in capturing non-linear relationships within the input data, which are often oversimplified within conventional EoS models.
Key findings revealed the robustness of ET models, validating their predictive reliability across temperature ranges and confirming their capability to accurately visualize the physical relationships between Rs and pressure. A trend analysis performed during the study corroborated this finding, showcasing how ET models could track Rs variations corresponding to changing pressure conditions.
Evaluative measures of the models' performance included formal statistical assessments such as standard deviation, correlation coefficient, and root mean square error. Results indicated the tree-based models performed with greater efficiency compared to conventional EoS.
The models were tested with both normal datasets containing detailed molecular characteristics and grouped datasets, where components were clustered for simplicity. The latter approach is particularly optimal for real-time applications and field-level predictions where rapid estimations are requisite.
Despite the successes, the study also acknowledged the computational complexity limitations of ML models, which may hinder deployment under resource-constrained environments. Nonetheless, the study presents significant advancements and substantial contributions to the existing PVT modeling field, particularly with regards to Rs prediction.
The study concludes with strong recommendations for extending the employment of machine learning models for various PVT parameters, promoting their scalability and efficiency across the petroleum industry. Their development has potential advantages including cost reductions and optimized field operations.
This study could serve as a benchmark for future research aimed at integrating such advanced techniques more broadly and effectively within reservoir engineering and other related applications, providing fertile ground for continued advancements.