Holistic Workflow for Autonomous History Matching using Intelligent Agents: A Conceptual Approach

Abstract

The current practices of assisted and automated history matching refer only to specific methods to address the optimization problem in history matching, but never span over the complete workflow. Computational power and the development of sophisticated software should allow us today to capture the whole workflow and automate it rather than just a piece of it (usually the regression method). The approach presented in this paper is based on the principle of domain decomposition. The reservoir history matching (HM) problem is decomposed into smaller and smaller pieces (from reservoir to region, from region to layer) to allow a refinement of the final result and hence the HM quality. The decomposition of the problem can be commenced by introducing time dependent well influence areas. These areas are sub-divisions of previous decomposition processes, as there are lithofacies types, clusters, flow units and vintage intervals. For the time dependent "influence areas", the success of the optimization can be tracked over the sequential iterations as well as across the different realizations. The progress of the most sensitive parameters can be tracked and successful strategies can be applied to other, similar areas of each realization. Finally, the interference of the influence areas has to be accounted for in the selection of the parameter values in each of the regression loops. A smart expert system that includes machine-learning techniques (cognitive agents) is used to drive the global history matching process, control the sub-domains and find the best parameter combination amongst all influence areas. The approach presented in this paper allows an automated history matching process, which is scalable for very large reservoir models. Relying on smart run selection and cluster technology even large models (many hundreds to thousands of wells) can be successfully matched in a reasonable amount of time. As the system is reservoir independent and learning from human experts, less and less user input is necessary over time leading to an autonomous history-matching agent.