Métodos de aprendizado de máquina aplicados ao cálculo de permeabilidade equivalente em reservatórios de petróleo fraturados

• Autor(a) principal: Elizeu, Iury Coimbra
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações do LNCC
• Texto Completo: https://tede.lncc.br/handle/tede/378
• topic: Aprendizado por computador; Permeabilidade; Fraturas (Geologia); Reservatórios de petróleo; CNPQ::ENGENHARIAS::ENGENHARIA CIVIL::GEOTECNICA

We conducted an exploratory study of various machine learning methods to predict the equivalent permeability value in cells associated with a discretization mesh occupied by a fractured rock in oil reservoir models. These models are essential for predicting fluid flow regimes in reservoirs. Among other objectives, the simulation can provide relevant information to maximize production and extend the life of reservoirs. Fractures, structures present in many types of rocks, play a crucial role in the flow regime, as they significantly influence the fluid flow and, consequently, production estimates. Different numerical and analytical models have been developed over the past few decades to incorporate fracture networks into reservoir simulations. However, methods based on the explicit treatment of the fracture network tend to considerably increase the computational cost of the simulation, while analytical solutions suffer accuracy deterioration in real applications. In this paper, we propose an exploratory analysis of machine learning methods to efficiently predict the equivalent permeability of cells in fractured reservoir models. To achieve this goal, we built a database with 10,000 examples of 3D fractured cells. This database contains geometric information of the fractures present in the cells, similar to the information used in Oda’s analytical method, as well as the equivalent permeability value, obtained through numerical simulation of single-phase flow using the Non-conforming Discrete Fracture Model (NDFM) method. We used different regression techniques, such as Decision Tree, Random Forest, XGBoost, Multilayer Perceptron Network, Multilinear Regression, and Grammatical Evolution, to train and validate the models. We then evaluated the model predictions by comparing them with results from numerical simulations. The results show that machine learning models can calculate equivalent permeability with significantly lower computational cost than the NDFM numerical method and with greater accuracy than Oda’s analytical method, especially in scenarios where there are a larger number of isolated fractures and/or fractures with a low level of connections among them and which do not create a path between the cell edges. These results highlight the potential of this approach for the oil industry.