Bayesian averaging sensitivity analysis of reservoir heterogeneity and anisotropy of carbon dioxide assisted gravity drainage of a large clastic oil reservoir

Abstract

Identifying the specific reservoir properties that influence oil production is an essential part of field development planning. Distinguishing the relative influence of variables and eliminating those that have no influence assists the design of enhanced oil recovery processes. The conventional approach to identify the most influential variables is the multiple linear regression (MLR) step-wise elimination. Bayesian averaging (BA), a stochastic method, offers a novel alternative approach by generating fifty models compared to each other to select the best one with of the most influential parameters based on the highest R2 and posterior probability with the lowest value of Bayesian information criterion (BIC). BA and MLR step-wise elimination are compared in the evaluation of a simulated immiscible carbon dioxide-assisted gravity drainage (CO2-AGD) system applied to the multi-layered, heterogeneous, upper sandstone oil reservoir of the South Rumaila field. A history-matched compositional simulation covering the nearly 57 years of prior production is used for sensitivity analysis of reservoir variables in relation to CO2-AGD implementation over 10 future years with 22 additional CO2 injection wells and 11 new production wells. Oil is displaced downwards towards the new horizontal production wells installed above the oil water contact. The influential variables evaluated are permeability (K), permeability anisotropy (Kv/Kh) and porosity (ɸ) considered separately across multiple heterogeneous reservoir layers. BA and step-wise elimination methods identify K as the most influential variable in all reservoir layers. Kv/Kh is moderately influential in the production layers and underlying water zone but not in injection or transition reservoir layers. ɸ in all reservoir layers has no influence on oil recovery from the CO2-AGD producing wells. More specifically, the BA-based influential parameters were determined after selecting the best model among the fifty generated models. The best model achieved that highest values of R2 of 0.848 and posterior probability of 0.234, and the lowest value of BIC of −149 with 12 identified influential geological parameters across the reservoir layers. Based on validation tests, varying K, Kv/Kh. and ɸ substantially across all reservoir layers confirms these findings and agrees more closely with the BA-derived reduced variable models than those generated by MLR step-wise elimination.