Comparative Evaluation of Machine Learning and Bayesian Deep Learning Methods for Estimating Ultimate Recovery in Shale Well Reservoirs
Saeed Beheshtian, Sara Kishan Roodbari, Hamzeh Ghorbani, Mohamadreza Azodinia, Mohamed Mudabbir, Annamária R. Várkonyi-Kóczy
Abstract
The Estimated Ultimate Recovery (EUR) is an important parameter in oil and gas production resources, particularly in hydrocarbon-rich shale wells. This study shows EUR prediction using three algorithms: Group Method of Data Handling (GMDH), Lasso, and Bayesian Deep Learning (BDL). The dataset includes 2200 data records from 15 shale gas wells, divided into 70% for training, 15% for testing, and 15% for validation. After the evaluation of three algorithms, the BDL algorithm has higher accuracy than the other algorithms. This algorithm has advantages such as providing quantitative prediction uncertainty assessment, resilience to outliers, adaptability to complex relationships, intrinsic regularization, transfer learning, ensemble formation, handling of heteroscedasticity, integration into active learning, and improved model interpretability. The results obtained using the test data for the BDL algorithm include a Root Mean Square Error (RMSE) of 4.2691 m3 and an R-squared (R2) value of 0.9879. Based on the findings of this study, it is suggested that researchers utilize this algorithm to predict key parameters.