A continuous learning algorithm for history matching

Abstract

History matching is a classical petroleum reservoir engineering data assimilation process whereby a reservoir model is calibrated to reproduce the behavior of a real reservoir yielding more reliable models for reservoir management, reservoir performance forecast and field development decisions. In this paper, we introduce a continuous learning-from-data algorithm for history-matching problems that generates solutions using the patterns of input attributes identified in the k-best solutions for each variable involved in the process. The proposed algorithm consists of a two-staged optimization strategy, in which each stage handles different types of reservoir uncertain attributes. The proposed learning approach continuously evaluates the data of all-available models and supports the strategic choice of input patterns that can be used in the generation of different and eventually better history-matched models. We apply the proposed algorithm to the UNISIM-I-H benchmark case, a complex synthetic reservoir model based on Namorado field, Campos basin, Brazil. The results outperform the ones from related work for the same benchmark, indicating the efficiency and effectiveness of the proposed strategy towards improving the history-matching quality of an initial set of solutions, with a lower simulation footprint.