The optimization of the locations of a large number of wells in an oil or gas field represents a challenging computational problem. This is because the number of optimization variables scales with the maximum number of wells considered. In this work, we develop and test a new two-stage strategy for large-scale oil field optimization problems. In the first stage, wells are constrained to lie in repeated patterns, and the reduced set of optimization variables defines the pattern type and geometry (e.g., well spacing, orientation). For this component of the optimization, we introduce several important modifications, including optimization of the drilling sequence, to an existing well-pattern optimization procedure. The solutions obtained in the first stage are then used to initialize the second stage optimization. In this stage we apply comprehensive field development optimization, where the well location, type (injection or production well), drill/do not drill decision, completion interval for 3D models, and drilling time variables are determined for each well. Pattern geometry is no longer enforced in this stage. Specialized treatments (consistent with actual drilling practice) are introduced for cases where multiple geomodels, used to capture geological uncertainty, are considered. In both stages optimization is achieved using a particle swarm optimization-mesh adaptive direct search (PSO-MADS) method. The two-stage procedure is applied to 2D and 3D models corresponding to different geological scenarios. Both deterministic and geologically uncertain systems are considered. Optimization results using the new procedure are shown to clearly outperform those from the single-stage comprehensive field development optimization approach. Specifically, for the same number of function evaluations, the two-stage treatment provides net present values that exceed those of the single-stage approach by about 15–18% for the cases considered. This suggests that this optimization strategy may indeed lead to improved results in practical problems.

石油或天然气田中大量油井位置的优化是一个具有挑战性的计算问题。这是因为优化变量的数量与考虑的最大井数成正比。在这项工作中，我们开发和测试了针对大型油田优化问题的新的两阶段策略。在第一阶段，将井限制在重复的模式中，并且减少的优化变量集定义了模式类型和几何形状（例如，井距，方向）。对于优化的这一部分，我们对现有的井网优化程序进行了一些重要的修改，包括钻井顺序的优化。然后，将在第一阶段中获得的解决方案用于初始化第二阶段优化。在此阶段，我们将应用全面的现场开发优化方法，在其中确定每口井的井位，类型（注入或生产井），钻探/不钻探决策，3D模型的完井间隔以及钻探时间变量。在此阶段不再强制使用图案几何。针对考虑用于捕获地质不确定性的多个地质模型的情况，引入了专门的处理方法（与实际的钻井操作相一致）。在两个阶段中，均使用粒子群优化-网格自适应直接搜索（PSO-MADS）方法实现优化。两阶段程序适用于与不同地质情况相对应的2D和3D模型。确定性和地质不确定性系统都被考虑。结果表明，使用新程序的优化结果明显优于单阶段综合现场开发优化方法。具体来说，对于相同数量的功能评估，对于所考虑的案例，两阶段治疗提供的净现值比单阶段方法的净现值高约15-18％。这表明这种优化策略的确可以导致实际问题中结果的改善。