American Oil and Gas Reporter - February 2015 - 55

SpecialReport: Unconventional Resource Science
The study evaluated the impact of parameters such as updip versus downdip deviation of wells, total organic content,
porosity, stimulated lateral length, and cluster spacing on production from wells using an advanced pattern recognition algorithm. The analyses were performed using
production from multiple time intervals
throughout the life of each well.
Given the complex nature of hydraulic
fracture growth and the very low permeability of many shale gas reservoirs' matrix rock, combined with the predominance
of horizontal completions, reservoir simulation tends to be the preferred method
to predict and evaluate well performance. However, limitations in our understanding of the complex phenomenon
also have limited our ability to accurately model shale production, thereby requiring models to incorporate significant assumptions. Making assumptions and modifying different parameters during the
history matching in order to validate the
reservoir simulation model can yield
nonunique sensitivity results in complex
shale reservoirs.
However, the modeling can be constrained reasonably when estimates of matrix permeability are available from special core analyses and microseismic mapping data gathered to determine stimulated reservoir volume and provide insights
into fracture network development. Combining these data with laboratory measurements of unpropped and partially propped
fracture conductivity can improve the insights from reservoir simulation significantly.
In anticipation of these major improvements in shale reservoir simulation, sensitivity analysis of shale formations' production performance, based on
pattern recognition technologies, seems
one of the most reasonable alternatives that
does not rely on any predetermined assumptions.
Pattern recognition is a tool for finding
patterns among nonlinear and interdependent parameters involving the shale gas development process. It has proven capable
of extracting useful information from
large datasets and is used extensively in

many industries.
Given the facts about shale gas reservoirs' complexity, no one should expect
linear and intuitive behavior from such an
unconventional system. Consequently, it
is unrealistic to understand or base decisions off of it using conventional statistics.
Alternately, this study takes advantage of
advanced pattern recognition tools to
generate relevant and reasonable trends between parameters such as the impact of
wells' updip versus downdip deviation,
stimulated lateral length, and cluster spacing on gas production in multiple time intervals. The study focuses on a large
number of horizontal wells in the Marcellus Shale, including some located on
multiple pads with different landing targets, well length and reservoir properties.
Methodology
Discovering trends in data is a quest for
many scientists and engineers. The notion
that shale is a statistical play may be attributed to the fact that the facts do not support many preconceived notions about
shale's storage and flow mechanisms.
Therefore, we set out to examine the
possibility of learning from the data in order to answer some of the questions that
arise during the production process.
These questions relate to the role each
parameter plays during production. This
examination considers both conventional
as well as new pattern recognition technologies from a Marcellus asset's dataset.
The conventional analyses used Cartesian,
semi-log, log-log plots for all parameters
versus some production indicators.
Furthermore, the study examined normalized production indicators, such as production per foot of net thickness or production per foot of completed interval, for
different portions of the field. The results
show shale production is too complex for
meaningful analysis using conventional
statistical analyses. One way to address
these complexities involves an advanced
data mining technology called "fuzzy
pattern recognition," which is based on
fuzzy set theory.
Pattern recognition is a branch of artificial intelligence that classifies or describes observations. It aims to classify

data-discerning patterns-based on either
a priori knowledge or statistical information that has been extracted from the pattern. The patterns to be classified are usually groups of measurements or observations, defining points in an appropriate
multidimensional space. Applying fuzzy
pattern recognition to a limited number of
classes of wells, such as poor, average and
good wells, is called step analysis or
well quality analysis (WQA). A similar
analysis, which treats every well in the
dataset as a potential unique well quality,
results in a continuous curve-rather than
a discrete set of steps-and is called a fuzzy
trend analysis (FTA).
The study applied both WQA and
FTA to a dataset that included more than
120 Marcellus Shale wells, including a
large number of horizontal wells with different production durations. The objective
was to discover hidden but potentially useful patterns and trends in the dataset that
could not be readily inferred from conventional statistical analyses. The technology proved superior to similar data grouping and plotting than is used in many data
mining analyses.
Conventional Analysis
Because of shale's fluid flow complexity, it may be challenging to perform
conventional statistical analysis and expect
linear and intuitive behavior from this kind
of reservoir. In fact, the system's complexity is displayed by the inability of conventional statistical analysis' to recognize any
apparent correlation or pattern between a
few parameters and the wells' best three
months of production. Conventional statistical analysis found no trend or patterns,
even though it was obvious that some parameters-such as higher porosity-resulted in better production performance.
In addition, the best three and 24
months of production for every foot of
stimulated lateral length, total slurry/proppant for different Btu areas, and also for
different horizontal lateral deviations
were plotted to see if conventional statistical approaches could identify any trends.
Although the data were clustered according to Btu areas and deviation types, no
clear trend emerged for either short- or
FEBRUARY 2015 55

American Oil and Gas Reporter - February 2015

Table of Contents for the Digital Edition of American Oil and Gas Reporter - February 2015