Comparative Analysis of Linear Regression and Artificial Neural Networks for Permeability Prediction in Reservoir Characterization

DOI: 10.14800/IOGR.1312

Abstract

Permeability prediction from well log data is a critical aspect of reservoir characterization, providing essential insights for effective reservoir management and hydrocarbon recovery. This study investigates the efficacy of two distinct modeling approaches—linear regression and artificial neural networks (ANN)—in predicting permeability from well log data. The linear regression models explored include Standard Linear Regression, Interactions Linear Regression, Robust Linear Regression, and Stepwise Linear Regression, while the ANN approach focuses on varying network structures to optimize performance. The Interactions Linear Regression and Stepwise Linear Regression models demonstrated strong predictive capabilities, with Root Mean Square Error (RMSE) values of 4.47 and an R-squared (R²) value of 0.98, indicating a robust fit between the predicted and actual permeability values. However, the ANN model, particularly with a structure of 10 neurons in the hidden layer (n-10), outperformed the linear models, achieving an RMSE of 29.90 and a remarkably high R² value of 0.9996. This underscores the ANN’s superior ability to capture complex, non-linear relationships within the data. The study provides a detailed analysis of model performance, highlighting the strengths and limitations of each approach. The ANN model’s superior accuracy makes it particularly suited for complex reservoirs where non-linear interactions are prevalent, while the Interactions Linear Regression model offers a simpler, more interpretable alternative for less complex scenarios. Based on these findings, the study recommends the adoption of ANN models for intricate reservoir characterization tasks, while linear regression models can be utilized for quicker, more straightforward predictions. Furthermore, comparison of this model with other existing models were made and this study’s model outperformed.