Evaluating the effectiveness of ensemble machine learning approaches for pore pressure prediction using petrophysical log data in carbonate reservoir

Abstract

Precise estimation of pore pressure (PP) holds significant importance in assessing the geomechanical parameters of reservoirs, playing a crucial role in the planning and execution of drilling and development activities in oil and gas fields. Recognizing its necessity various empirical and intelligent methods have been introduced to enhance the precision of PP prediction. The main objective of this study is to assess the effectiveness of ensemble machine learning (ML) models by conducting a comparative analysis of individual ML models for predicting PP. To identify the most influential input variables for constructing ML models, a feature selection analysis was performed. The findings suggest that a combination of 8-input variables holds the most influence on ML model construction. Three individual ML models namely least-square support vector machine, multi-layer perceptron artificial neural network and decision tree regression (DTR) were employed for PP prediction by using petrophysical log data (8 input variables). The dataset of wells A and B was for training, and testing these models. The results from individual models showed that the DTR algorithm provides the most accurate PP prediction, boasting an \({R}^{2}\) value of 0.972 for training dataset, and an RMSE of 110.698 Psi. The performance of individual models can be enhanced using ensemble models, including simple averaging ensemble (SAE), weighted averaging ensemble (WAE), stacking ensemble (SE), random forest (RF). The results reveal that all ensemble models deliver more accurate PP predictions than individual models. Among them, the RF model stands out with an \({R}^{2}\) of 0.999 for both training and testing datasets. It also demonstrates lower RMSE values of 8.948 Psi, and 21.568 Psi for training, and testing datasets, respectively, making it more accurate than SAE, WAE, SE and individual ML models. Furthermore, the generalization analysis demonstrates that the 8-input variable RF model exhibits excellent performance, providing more accurate PP predictions when applied to the well C dataset within the study area.