English abstract
Porosity, permeability, and fluid saturations are considered critical petrophysical
parameters for the reservoir because of their significant impact on the oil and gas reserve
estimation as well as petroleum economics. For identifying reservoir properties,
traditional methods rely on lab measurements, wireline logs, or empirical correlations
between other parameters such as formation resistivity with hydrocarbon saturation,
density with porosity, which are either expensive, time-consuming, or do not meet the
accuracy requirements. All the foregoing limitations are overcome by the new horizon for
integrating machine learning techniques as a new way for predicting these properties.
Therefore, this work will develop a new prediction model by combining advanced data
analytics with visualization approaches for pre-processing the logs collected in the
Sultanate of Oman's Gharif Reservoir. The relationship between the output and input
variables was investigated using descriptive statistics, then the Extreme Gradient Boosting
(XGBoost) regression model was applied to predict reservoir porosity, water saturation,
and permeability. Then, to minimize the dimensionality of the input features, various
feature selection techniques were evaluated, including SelectKBest, Recursive Feature
Elimination, Random Forest, and Principal Component Analysis. To improve the
accuracy of the results, a novel custom ensemble model of Random Forest and Recursive
Feature Elimination with an enhanced feature engineering technique was presented. The
proposed unique ensemble model appears to outperform the classic XGBoost, revealing
the immense potential of ensemble modeling for reservoir characterization. The new
custom ensemble model reduced the characteristics of the high-dimensional dataset while
also achieving low MAE and RMSE prediction errors of 0.0037 & 0.023 for permeability
and saturation respectively
