A Comparative Study of Ensemble Learning Techniques and Mathematical Models for Rigorous Modeling of Solution Gas/Oil Ratio

Summary The solution gas/oil ratio (Rs) represents the quantity of gas dissolved in oil under reservoir conditions. It is a vital parameter in petroleum engineering, defining the gas content available during production. While many experimental techniques exist for measuring this ratio, they often require considerable time and resources. Thus, mathematical and intelligent models are essential for accurate determination. A total of 720 data points from diverse geographical regions were collected from published studies for this research, using gas-specific gravity, temperature, bubblepoint pressure, and oil API gravity as inputs, with the solution gas/oil ratio as the output. Statistical and physical analyses assessed the impact of parameters on the solution gas/oil ratio, revealing that temperature does not always decrease dissolved gas. Beyond a specific point, known as the inversion temperature, higher temperatures enhance gas solubility. The data set was split, with 80% allocated for training and 20% for testing. The accuracy of the Al-Marhoun model, originally established from 160 data sets from the Middle East, was evaluated using test data, which produced a root mean square error (RMSE) of 468.79 scf/STB. The recalibration of coefficients using 576 training data points and the differential evolution (DE) algorithm led to the formulation of New Model 1. By incorporating the temperature effect and defining the inversion temperature, New Model 2 was developed. Testing results showed that New Model 1 achieved an RMSE of 100.97 scf/STB, while New Model 2 reached 105.1 scf/STB, both showing better accuracy compared with previous mathematical models, including the Al-Marhoun model. Subsequently, machine learning techniques were applied, including the multilayer group method of data handling (GMDH), voting regressor (VR), extra trees (ET), histogram-based gradient boosting regression (HGBR), extreme gradient boosting (XGBoost), and gradient boosting with categorical features support (CatBoost) for the modeling process. Notably, models such as ET, HGBR, XGBoost, and CatBoost effectively captured the temperature effect from the data. The performance of the models was evaluated using statistical and visual analyses. The HGBR model outperformed all others, achieving a training error of 0.0044 scf/STB and an RMSE value of 73.03 scf/STB on the test data, demonstrating its clear superiority among the considered models.