Page 172 - Intelligent Digital Oil And Gas Fields

Components of Artificial Intelligence and Data Analytics 131

4.3 APPLICATIONS TO DIGITAL OIL AND GAS FIELDS
4.3.1 Machine Learning and Predictive Analytics

This section presents a few state-of-the-art and emerging applications of data
mining and ML with predictive analytics in the E&P industry. The focus is
on the applications of production optimization because they correlate more
firmly with the data-driven reservoir management and are significant areas of
application in DOF systems.
Mohaghegh et al. (2015) have introduced the concept of surrogate res-
ervoir models (SRM) as a “smart” proxy for numerical reservoir simulation
models. The SRM is an ensemble of multiple ML technologies, including
pattern recognition and intelligent agents, which are trained to learn and
consequently mimic the behavior of fluid flow physics using data generated
by a numerical simulation model; however, SRMs run at extremely high
speeds and complete the simulation run in a fraction of a second.
Mohaghegh et al. (2015) have deployed the SRMs to increase field produc-
tion and optimize choke size schedule—all without drilling new wells.
Bravo et al. (2014) have defined intelligent agents as computational sys-
tems comprising multiple active components that are capable of making
decisions and taking actions autonomously. The intelligent agents are suited
to processing large amounts of data in distributed environments and can also
communicate and collaborate with each other to reach common objectives.
For example, Zangl et al. (2011) have used intelligent agents in the form of
self-learning expert systems to construct a holistic workflow for autonomous
history matching. Fig. 4.12 (from Zangl et al., 2011) demonstrates a sche-
matic of a hierarchical learning exercise of a history matching agent that,
instead of accomplishing the change of state in sequential manner, splits a
complex task (minimizing the Objective) into a set of isolated and repeatable
subtasks executed at different layers of corrective actions.
In an SPE webinar, Saputelli (2015) has presented several varieties of
supervised and unsupervised ML techniques (see Table 4.4) with optimiza-
tion as emerging trends of predictive and prescriptive data analytics applica-
tions in E&P. However, ANNs are also being used in innovative
applications. For example, Shirangi (2012) have built fast proxy models
by combining ANNs and SVR models to solve a robust production optimi-
zation problem and used unsupervised ML (k-means clustering and MDS;
see Table 4.4) to select an optimal set of representative reservoir model real-
izations. Recently, ANNs are being actively used in a variation known as a