On Optimal Selection of Objective Grouping for Multiobjective History Matching

Abstract

Summary Multiobjective history matching has gained popularity in the last decade. It provides an ensemble of diverse set and good matched models that should lead to improved forecasting. Moreover, in some cases, multiobjective history matching provides faster and more-robust convergence than the single-objective approach. In multiobjective, objective components (usually groups of them) guide the algorithm to different areas of objective space that lead to a diverse set of optimal solutions. These algorithms are widely established and well-developed for problems with two or three objectives. Under an increasing number of objective components, such as in a reservoir model with multiple wells and production data, multiobjective-history-matching performance (convergence speed and match quality) can deteriorate. One effective approach is grouping objective components to reduce the number of objectives. However, the existing literature does not present sufficient information on appropriate grouping techniques and ways of combining objective components. We present a novel technique to group the objective components depending on analysis of the nonparametric-conflict information obtained from a set with a limited number of initial solutions. By grouping the objectives depending on the conflict between them, we aim to achieve better performance in history matching. We apply this framework to history matching of an industry-standard reservoir model and a real-field case study. We also perform history-matching runs of groupings with different degree of conflict, and then analyze the performance among them with the statistical-significance test. Our extensive simulation results show that the proposed conflict-based strategy can be used as a guideline to help select a grouping of the objective components in multiobjective history matching optimally. By calculating the conflict between objectives a priori, we can identify which grouping scheme will result in a better performance. This technique can significantly improve the fitness quality of the matched model given the same number of flow simulations, and can also obtain a diverse set of models.