An AI-Driven Multiwell Placement and Scheduling Method

Abstract

Abstract Well placement optimization is important for informed field investment decisions. Numerical simulation models are usually utilized for well placement optimization, as they offer a detailed account of subsurface reservoirs. However, a full loop optimization using numerical simulation models makes this process intractable. Instead, many assets and reservoir engineers follow heuristic or trial-and-error approaches, leading to suboptimal results. This paper introduces a novel method for well placement optimization and scheduling based on AI-proxy models. Our approach relies on generating an Ensemble of simulation runs (100s-1000s), varying field-level and well configuration parameters. We follow a loosely coupled process to solve the field control parameters and well placement optimization problems. Output data from the field, well and completion levels are collected automatically from the simulation models. Thereafter, several Machine Learning (ML) meta-models are trained to predict field-level production, well drilling and abandonment schedules, and the performance of production and injection wells. The outcome is selected well trajectories based on opportunity maps generated using the meta-models. The new workflow was first tested on a synthetic reservoir model. A sensitivity study of 275 simulation cases was performed varying reservoir management and well configuration parameters. It was observed that although the petrophysical properties such as rock quality and hydrocarbon saturations were not part of the ML model training, they were well reflected in the opportunity maps. This provided assurance that the model captures the basic static and dynamic controls on well performance. The model accuracy was improved further by adding features such as the perforation top depth. Additional drilling constraints, such as the presence of large faults, had to be considered to ensure the validity of the optimized well locations. The best simulation case from the sensitivity study was selected based on specific economic factors and adopted as a comparison base case for the new well optimization method. The original wells were replaced with an optimized set of well obtained with our deep learning and optimization workflow. This resulted in a 10% increase in total oil production combined with 50% decrease in the amount of water production. Modern cloud/local computational resources can expedite full-loop optimization processes. However, both computation efficiency and elasticity do not address the iterative nature of classic well optimization methods. The method presented in this paper attempts further speedup by leveraging ML.