Aplicação de algoritmos de deep learning como modelos substitutos de simuladores de reservatórios de hidrocarbonetos

Abstract

A significant challenge on Oil and Gas Industry, especially in reservoir engineering and its management context, is related to the capacity of run optimization strategies. It occurs because the current computational time costs are prohibitive and demand too many resources, even for a medium-size simulation. Perhaps the currently scientific solutions, none of them used Deep Learning techniques in low-level granularity for pre-dictions, especially the grid-cell level size approach. This thesis proposes, analyzes,

and states a feature selection, a model design, and a training strategy with the applica- tion of Deep Learning techniques (DNN and CNN), the Design of Experiment, and all

statistical evaluation-based metrics and its graphic tools. This defined process intends

to work as a solution to create a proxy model for reservoir numerical software simula- tion. Four different classical industrial scenarios, including production and injection

wells, are conducted in a diversified temporal sampling to generate proxy models. The achieved results are promising, with accuracy always bigger than 80%, and at specific scenario conditions, it reaches even 99,9% as evaluation criteria.