Data-driven oil production strategy selection under uncertainties

Abstract

This study presents a user-friendly tool to assist in selecting oil production strategies when facing high levels of uncertainty in a real case. Specifically, we address the challenge of determining optimal well-bore positioning and control parameters under uncertain geological conditions, aiming to maximize production efficiency while managing computational complexity. The model deals with decision-making factors and geological data, represented by high-dimensional maps traditionally handled through intensive numerical methods. The production strategy goes through robust optimization based on decision variables in set $P$, such as well-bore positioning, and uncertainties associated with 3D reservoir properties in set $R$, such as porosity and permeability. The method combines two sets of $P$ variables, emphasizing positioning and control guidelines. The technique employs representative scenarios to find a generally applicable strategy considering $P\times R$ mixtures. The variables $R$ describe a real and heterogeneous reservoir in the pre-salt area in Brazil. The method focuses on critical information through dimensionality reduction while guaranteeing faster, more accurate, robust decisions and balancing efficiency with effectiveness. We rely upon machine-learning, such as Gradient Boosting Regression with few-shot training strategies. The SHapley Additive exPlanations and feature importance allow the model interpretation, enabling us to understand how the well-bore positioning impacts the response. The method is integrated into the optimization loop to work alongside the simulator, and both methods work in tandem as a fast metaheuristic system supported by a slow numerical one. The method improves the computational footprint by 76%.