Ensemble-based reservoir simulation workflows have become popular for estimating prediction uncertainty and optimizing field development objectives under uncertainty. While workflows exist for ensemble generation, ensemble analyses become an increasing challenge for extracting value from ensembles through robust field development guidance. This work presents a self-supervised deep learning modelling approach for classifying 3D-structural features applied to expert-driven field performance optimization under uncertainty. Meaningful representations of subsurface features are essential to efficient machine learning tasks such as classification of categorical outcomes and regression for predicting dependent target variables. Unlabeled data is abundant and used for unsupervised learning which is complex in contrast to the use of labelled data. However, labelling is generally expensive and, in some cases, ill-defined or impractical due to missing modelling approaches. In this work, a self-supervised deep learning modelling approach is applied to learn semantically meaningful representations of subsurface features, e.g.,size and structure of connected volumes. The deep learning model is designed to learn a representation of 3-dimensional subsurface features. Representations are categorized, clustered and ranked for well location decision support. For application demonstration, the self-supervised deep learning modelling approach is embedded in a structured workflow design for guiding well location selection to optimize reservoir delivery performance under subsurface uncertainty. Performance verification and application results are presented for the Olympus semi-synthetic case study. Based on probabilistic success criteria we present an optimized well placement design for which 90% of all producers deliver the economic demand at an 80% probability level or higher. In a second study the workflow is applied to a real field to discuss guidelines of use and to share expected gains in workflow efficiency, result quality and decision support.
Field development planning moves from using deterministically constructed reservoir models to stochastically generated ensembles of models to better capture subsurface uncertainties. This leads to a challenge on how to understand reservoir dynamics and extract insights because of the large volumes of data involved. The objective of the presented system is to demonstrate how cloud-based computing can help to derive actionable insights of this large volume of data through automation, machine learning, and elastic scaling. In this paper we present a cloud-native solution for optimizing and evaluating ensemble-based infill well locations. An opportunity index (OI) is derived from static and dynamic grid properties of reservoir simulation and a connected component search for every realization. Probability maps of OI are constructed to present the likelihood of high-OI areas from all simulations to propose infill well targets for the ensemble. Infill wells are automatically evaluated and ranked by ensemble field production increases. The workflow is deployed in a cloud-based system to leverage elastic scaling of compute resources to cope with the large volumes of data inherent in ensemble models. The result from deploying this new workflow in a mature field in the North Sea shows that optimization of infill well targets at the ensemble level increases the robustness of targets compared to the alternative of selecting one or a few realizations. When randomly selecting two deterministic cases to analyze infill well locations, we observe inconsistency in candidate well locations. In case A, little opportunity is shown in the northern part of the reservoir and no opportunity in the southwestern part, whereas the result from case B shows almost the opposite, having a large area of opportunity in the northern part and a small opportunity in the southwestern part. By applying the proposed solution, only focused areas of the northern part of the reservoir are suitable for candidate well locations. The study time of target identification and evaluation is significantly shortened. The cloud-based deployment removes the company's need to own and manage powerful computer and data-storage infrastructure. In summary, the solution improves the workflow efficiency and provides high-quality results for field development decision making.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.