Applications of Machine-Learning in Reservoir Connectivity Detection and Optimization

Abstract

Reliable quantification of well connectivity is a crucial aspect in forming a good understanding of a reservoir, which in turn helps in formulating future development plans such as rate optimization and offset wells. This assumes an even greater importance when applied to high-capital projects such as CO2 EOR and polymer floods. Conventional methods for assessing well connectivity include tracer tests and numerical simulation-based techniques such as streamlines. However, these methods of connectivity detection tend to be either computation-intensive (i.e. numerical simulation) or resource-intensive (such as tracer tests).

This dissertation makes three major contributions related to machine-learning applications for connectivity detection and rate optimization. Firstly, I propose a novel approach for connectivity quantification and rate optimization during a waterflood under geologic uncertainty in reservoir properties such as permeability and porosity. A machine-learning (ML) based approach which is quick and scalable for rate optimization over multiple geologic realizations is proposed instead.

Secondly, a machine-learning framework is built on the statistical recurrent unit (SRU) model that interprets well-based injection/production data into inter-well connectivity without relying on a geologic model. Furthermore, a streamline-based validation procedure is also proposed which provides physics-based backing to the results obtained from data analytics.

Thirdly, this dissertation proposes a workflow that integrates unsupervised machine learning and streamline techniques to select representative geologic realizations based on their flow features. The workflow may be used to identify key wells for implementing optimized rate schedules, while taking into account the uncertainty in the geologic model