Characterization of lacustrine shale oil reservoirs based on a hybrid deep learning model: A data-driven approach to predict lithofacies, vitrinite reflectance, and TOC

Abstract

The integration of deep learning technologies into geoscience domains enables the evaluation of rock properties in unconventional shale reservoirs. In particular, combinations of deep learning algorithms can recognize and process details in well log data, leading to unparalleled prediction precision. This research investigates the capabilities of hybrid deep learning models in addressing the challenges related to characterizing lacustrine shale oil reservoirs. These challenges include the classification of lithofacies, as well as the estimation of vitrinite reflectance (R_o; %) and total organic carbon (TOC; wt%). A comparative analysis was conducted between traditional machine learning (ML) algorithms, including Random Forest (RF), XGBoost (XGB), and Support Vector Regression (SVM), and hybrid deep learning models, like Deep Belief Network (DBN) and Long Short-Term Memory Network (LSTM). The results show promising outcomes for these tasks, highlighting the superior reliability and reproducibility of hybrid deep learning models compared to traditional ML algorithms. The combination of DBN + LSTM hybrid models exhibited high predictive accuracy and closely matched the lithofacies, R_o, and TOC values of laboratory measurements. This study presents a rapid and accurate method for estimating reservoir parameters in a complex lacustrine shale oil reservoir in the Songliao Basin, China. It compares traditional machine learning algorithms and hybrid DBN + LSTM models, and its results suggest that combining other deep learning methods as hybrid models could provide additional formation evaluation benefits.