Two-Stage Scenario Reduction Process for An Efficient Robust Optimization

Abstract

Summary Probabilistic approaches for optimization objectives need a large ensemble size to consider uncertainties, which is often computationally expensive. Our proposed method includes two scenario reduction (SR) techniques applied to geostatistical realizations and reservoir simulation models to handle geological and dynamic uncertainties. The goal is to select a subset of simulation models to be used in an efficient robust optimization (RO). The proposed workflow is summarized in the following sections. Generate total geostatistical (TG) realizations representing grid properties using Latin Hypercube (LH) sampling; Select representative geostatistical (RG) realizations from the TG realizations using an integrated statistical technique named Distance-based Clustering with Simple Matching Coefficient (DCSMC). This section is the first stage of SR; Integrate other uncertainties with the RG scenarios to generate total simulation (TS) models using Discrete Latin Hypercube with Geostatistical models (DLHG); Apply data assimilation process to reduce uncertainty and generate total history-matched simulation (THS) models using a filtering indicator named Normalized Quadratic Deviation with Signal (NQDS); Select representative history-matched simulation (RHS) models from the THS models set using a tool based on a metaheuristic optimization algorithm named RMFinder. This section is the second stage of SR; Perform an RO to maximize NPV as the objective function using the selected RHS models; The novel SR workflow selects the representative scenarios (RG realizations and RHS models) during two steps: (1) RG selection based on static features before the simulation process and, (2) RHS selection based on simulation-based (dynamic) features after the simulation process. The workflow is applied to a fractured synthetic reservoir model named the UNISIM-II-D flow unit-based. To check the computational-time and efficiency of the methodology, we compare two candidate production strategies based on (1) five RHS models obtained from the two-stage SR process considering DCSMC and RMFinder techniques (workflow A), and (2) five RHS models obtained from one-stage SR process using the RMFinder method (workflow B). In workflow A, the SR process is performed gradually during two steps while in workflow B, the SR process is applied all at once in one step. The results show that the distribution of simulation outcomes after RO for the representative scenarios and the total scenarios in workflow A are more similar than workflow B. In addition, the robust production strategy obtained from workflow A is preferred to workflow B because it presents higher chances of high NPV value and lower chances of low NPV value.