Page 177 - Intelligent Digital Oil And Gas Fields

136 Intelligent Digital Oil and Gas Fields

from existing well-treatment data, it is possible to find patterns and signifi-
cant variables that affect the extreme values of pumping job-pause time
(JPT) in a particular region, and what is the most critical value causing frac-
ture screen outs. They performed four case studies on a database that
included data from 200,000 fracturing and data-acquisition jobs from all
over North America, since 2004. The data in the database included: com-
pilation of general well and job information; job-level summary data;
pumping schedule stage-level summary data; pumping schedule individual
stage data, which included additives, wellbore and completion data, event
log data, and equipment data. When mining such complex and extensive
databases, the dependencies and correlations among the variables are mostly
nonlinear, hidden, and highly nonintuitive. By failing to address such intri-
cate data root-cause relations, frustration can occur when the well opera-
tional conditions are thought to be understood but unexpected behavior
occurs. This can lead to severe under-performance and economic failure
of individual wells, even though the generic data indicates identical forma-
tions, similar geologic conditions, and similar completion techniques, as
“similar” wells performed significantly better. Similar high dependence
on the nonlinear intra-variable effects and potentially negative consequences
on the optimization of hydraulic fracturing jobs in shale plays are also
reported by Cipolla (2015).
To address the matter Maucec et al. (2015) have built classification trees
to predict occurrences of fracture screen outs [as categorical response vari-
ables (Fig. 4.16)] and regression trees to predict the JPT as a continuous
variable.
They used k-fold cross-validation to assess the misclassification probabil-
ity for the classification tree and MSE for the regression tree (Fig. 4.17) and
performed pruning to optimize the tree depth. In addition, they introduce
CART enhancements (Maucec et al., 2012, 2013) by mapping the data
points normal (mean¼0; variance¼1) domain using normal score trans-
form (NST) and kernel k-means clustering to identify the variability of cor-
related variables and further reduce the sample size. Both enhancements
were found to improve the root-cause prediction capability of decision trees
by reducing the mean prediction error.
Further examples of predictive modeling with decision trees can be
found in Singh (2015) where they are used for root-cause identification
and production diagnostics of gas wells with plunger lift and in Schuetter
et al. (2015) for production optimization in unconventional reservoirs.