Page 265 - Intelligent Digital Oil And Gas Fields

Integrated Asset Management and Optimization Workflows 215

nongeological and nonphysical features in the reservoir model. Moreover,
the lessons learned are not properly applied to create a realistic reservoir
model, and the perceived “history-matched” models are of low predictive
value.
As reservoirs and assets mature and data acquisition and processing
methods evolve and become more sophisticated, the acquired reservoir data
grow substantially in terms of quantity and complexity. Particularly with the
expansion of DOF projects that use automated and data-driven IAMod and
IAM workflows, the need to solve large-scale, high-resolution modeling
problems, quantify the inherent model uncertainty for more reliable predic-
tion, and optimize their performance are becoming prevalent in the E&P
industry. The challenges resulting from integrating multiple scales of data
with uncertainties in physical parameters and processes make imperative
the use of efficient model parameterization, advanced inversion and optimi-
zation algorithms, with utilization rapidly evolving HPC architectures.
Within the last three decades, the oil industry has gained traction in
developing and implementing stochastic, population-based algorithms in
reservoir characterization and simulation workflows. The applications of
simulated annealing (SA), an algorithm that was originally developed for
solving combinatorial optimization problems, first emerged in the oil and
gas industry in the early 1990s in areas from stochastic reservoir modeling
to optimization of well-scheduling and -placement (Deutsch and Journel,
1994; Ouenes et al., 1994) and have endured through the introduction of
advanced SA algorithms, such as very fast simulated annealing (VFSA), with
recent expansion of unconventional exploration (Sui et al., 2014).
Another important advance in oil and gas stochastic modeling was the
introduction of techniques for the design of experiments (DoE), which
was originally developed in agriculture in the late 1920s (Salsburg, 2001).
DoE modeling has been primarily used for the rapid quantification of uncer-
tainty using proxy models with response surface analysis (RSA) and various
forms of designs (e.g., latin hypercube, Box-Behnken, etc.) in AHM
(Cullick et al., 2004; Alpak et al., 2013), sensitivity analyses (Fillacier et al.,
2014), and risk evaluation (Sazonov et al., 2015). In the early 2000s the
E&P industry started to see an expansion of ensemble-based, Bayesian infer-
ence and model inversion, using, for example, the evolutionary algorithms
(Schulze-Riegert and Ghedan, 2007), the ensemble Kalman filter (EnKF)
(Evensen, 2009), and recently a complementary data assimilation approach,
the ensemble smoother (ES) and multiple data assimilation (MDA) by
Emerick and Reynolds (2012, 2013) and Maucec et al. (2016, 2017) and