Anda di halaman 1dari 11

IADC/SPE 81639

Development of a probabilistic model to estimate productivity improvement due to


underbalanced drilling
P. V. Suryanarayana, SPE, Blade Energy Partners; S. Rahman, Blade Energy Partners; R. Natarajan, SMU;
R. Reiley, BP
Copyright 2003, IADC/SPE Underbalanced Technology Conference and Exhibition
This paper was prepared for presentation at the IADC/SPE Underbalanced Technology
Conference and Exhibition held in Houston, Texas, U.S.A., 2526 March 2003.
This paper was selected for presentation by an IADC/SPE Program Committee following
review of information contained in an abstract submitted by the author(s). Contents of the
paper, as presented, have not been reviewed by the International Association of Drilling
Contractors or the Society of Petroleum Engineers and are subject to correction by the
author(s). The material, as presented, does not necessarily reflect any position of the IADC,
SPE, their officers, or members. Electronic reproduction, distribution, or storage of any part of
this paper for commercial purposes without the written consent of the International Association
of Drilling Contractors or the Society of Petroleum Engineers is prohibited. Permission to
reproduce in print is restricted to an abstract of not more than 300 words; illustrations may not
be copied. The abstract must contain conspicuous acknowledgment of where and by whom
the paper was presented. Write Librarian, SPE, P.O. Box 833836, Richardson, TX 75083-3836
U.S.A., fax 01-972-952-9435.

Abstract
Improvement in productivity from wells drilled and completed
underbalanced can be substantial. This is especially true in
formations susceptible to damage induced during
overbalanced drilling.
Substantial improvement in
productivity can be the driver for the selection of
ubderbalanced drilling (UBD) candidates. However, the
problem of quantitatively estimating such improvements to
assist in candidate selection remains. This paper presents a
unique probabilistic approach to the estimation of productivity
improvement that can be expected from application of UBD
for a given well candidate. The unique feature of this work is
the use of analog data to influence the predictions of the base
probabilistic model. The approach described uses a reservoir
flow model that incorporates linear and nonlinear inflow
relationships. The model can be applied to horizontal or
vertical completion in either gas or oil wells.
Both
compositional and Black Oil PVT methods are used to
estimate reservoir fluid properties. Damage is modeled as a
skin effect, a relative or absolute permeability reduction, or
uneven flow across the producing section. Statistical
uncertainty in the reservoir parameters is characterized using
different applicable probability distribution functions,
depending upon the natural behavior of the parameter.
Standard Monte Carlo simulations are then used to estimate
the expected productivity improvement. However, review of
analog data from 40 different UBD wells indicates that
theoretical estimates alone are not sufficient to estimate the
expected productivity improvement.
Therefore, the
predictions of the theoretical reservoir flow model are
statistically updated using available analog data. The key
concept employed is one of making a prediction using

Corresponding author

standard reservoir models, incorporating the raw uncertainties


of reservoir parameters, and then modifying the prediction
based on a statistically legitimate combination of analog data
(results from previous UBD wells) and parameter
uncertainties. The resulting updated distribution is more
reflective of expectation from UBD in a given field. Two
statistical updating methods are discussed- a Bayesian
updating approach, and a weighted mixture sampling
approach. Both methods are legitimate for productivity
predictions. Results are given in terms of a productivity
improvement factor (PIF) distribution, the ratio of the
productivity index using UBD to that of the reference
productivity without UBD. The resulting PIF distribution can
then be used in full-field reservoir simulators to evaluate the
economic impact of UBD for a given campaign. The
predictive model runs on an Excel-based program with the
statistical package @Risk, and is easy to use. Two examples
are provided to illustrate the use of this method and the impact
of updating on the PIF distribution. The authors suggest
improvements to the model, including gathering and statistical
treatment of additional analog data, better reservoir models,
and improvements in the statistical updating methods.
Introduction
The benefits of underbalanced drilling (UBD) are well
recognized by the industry and there is increasing increase in
interest in this technology. The benefits from UBD generally
fall into two classes Cost avoidance- including mitigation of conventional
drilling problems such as lost circulation and
differential sticking reduction in stimulation and
cleanup costs, etc.
Value creation- including productivity improvement,
increase in ultimate recovery, and real-time reservoir
characterization while drilling.
Value creation is possibly the most important benefit of
UBD. Improvement in productivity and ultimate recovery due
to UBD can be substantial, and can (and should) in fact, be the
driver behind the choice of UBD. Value creation benefits are
largely a consequence of the fact that UBD reduces or
eliminates much of the invasive damage caused by
conventional drilling. These benefits are also unfortunately
difficult to quantify. Pure value driven justification of UBD
has been plagued by this difficulty. Not surprisingly, such

UBD projects are rare. Indeed, most large UBD campaigns are
motivated by the cost reduction (mitigation or elimination of
drilling problems). Productivity improvement has been a
serendipitous after-the-fact observation.
In recent times, however, with increasing evidence of the
productivity improvement benefit of UBD, several operators
are beginning to consider pure value driven UBD
campaigns. Such campaigns will benefit immensely from a
logical approach to the estimation of the productivity
improvement from UBD. Increasingly, the argument of value
is made to reservoir and asset management groups, and a
rigorous means of establishing the value impact is much
needed in the industry. This is the subject of this paper.
This paper details a probabilistic approach to the
estimation of productivity improvement from the application
of UBD to a candidate well. The approach uses wellestablished inflow equations, and includes consideration of
both oil and gas wells. Vertical, horizontal and inclined wells
are treated separately. Without loss of generality, PVT models
with a wide range of applicability are used for both gas and oil
wells. Damage is usually modeled as a skin effect, and the
skin reduction due to UBD is estimated. Damage may also be
modeled as a reduction in permeability or a reduction in the
reservoir section length contributing to flow. The resulting
improvement in productivity is cast as a productivity
improvement factor (PIF). Uncertainty in the defining
parameters of inflow are considered explicitly as probability
distributions, and a Monte Carlo simulation approach is used
to cast the expected PIF as a probabilistic result.
One of the unique features of this work is the use of
analog data to influence the predictions of the base
probabilistic model. The key concept employed is one of
making a prediction using standard reservoir models,
incorporating the raw uncertainties of reservoir parameters,
and then modifying the prediction based on a statistically
sound combination of analog data (results from previous UBD
wells) and parameter uncertainties. The resulting updated
distribution may be more reflective of expectation from UBD
in a given field. Two techniques of achieving this are
discussed- an empirical Bayesian approach, and a weighted
mixture sampling approach. Zonation of the data is also
discussed.
The model runs on an ExcelTM spreadsheet platform,
with the Monte Carlo simulator @RiskTM as an add-on to
enable statistical analysis and interpretation. The program has
an easy-to-understand interface, and is fairly easy to use.
In what follows, the basic reservoir inflow modelling
approach is first discussed in detail. The assumptions made in
implementing the approach are explained. The statistical
considerations in probabilistic prediction of productivity
improvement are then discussed in two stages: (a) the
statistical consideration of reservoir uncertainties, and (b)
statistically influencing probabilistic reservoir-model based
predictions using available analog data. The statistical
approach used and guidelines for the selection of distributions
are then discussed. The implementation approach and the use
of the program are also described. Two examples are included
to illustrate the use of the program. Finally, the limitations of
the current approach, ways of using it in UBD candidate

IADC/SPE 81639

selection, and further work to improve upon the approach are


discussed.
Theoretical Basis of Inflow Models
Without loss of generality, reservoir inflow is modeled using
standard inflow equations. A linear inflow relationship based
on Darcys law is used for oil wells. For gas wells, a nonlinear inflow performance relationship (IPR) is used, with
non-Darcy effects included. Both steady state and pseudosteady state flow are considered. Specific relationships are
used for vertical (Darcy inflow) and horizontal wells (JoshiEconomides equation), leading to eight inflow relationships in
all. For brevity, these are not listed here. Craft and Hawkins
(1991), and Economides et. al. (1993) have a good discussion
of the inflow performance equations used in this work.
The approach discussed here can easily be extended to
use other, more sophisticated IPR equations (such as quadratic
IPR, Generalized Vogel IPR, etc.), or even a tabulated IPR.
The inflow equations are expressed in terms of a
productivity index, PI. The PI is a normalized measure of well
productivity, and is usually expressed (in units of flow rate per
unit time per unit of drawdown pressure) as:
PI =

p res

q
p wf

(1)

The term p res p wf is the draw down, or degree of


underbalance in the case of UBD. It is sometimes customary
to further normalize the PI with respect to the length of
reservoir exposed to flow, in which case Eq. (1) is rewritten
with the length of reservoir section appearing in the
denominator. In this work, we use Eq. (1) for PI.
Skin Factor
The skin factor s appears as an additive term in the
denominator of the inflow equations. The skin factor s is a
non-dimensional measure of damage to the reservoir. It is
analogous the film coefficient in heat transfer, and represents
an additional resistance to flow. Any phenomenon that causes
a distortion of the flow lines of the reservoir fluid from normal
to the well direction or a restriction to the flow results in a
positive value of the skin. It can be created by mechanical
causes such as partial completion and inadequate number of
perforations, phase blockage, relative permeability reduction
to the main fluid due to filtrate invasion, turbulence, or
damage to the natural reservoir permeability due to solids
invasion and swelling caused by filtrate invasion. A detailed
discussion of the causes and effects of damage is beyond the
scope of this work, but it is obvious that invasive damage, left
unaddressed, results in positive skin.
Since skin usually accounts for all causes of reduction
(or enhancement) in productivity, it is a composite variable
made up of several components. Depending upon the inflow
equations used, skin includes true damage skin, completion
skin, perforation skin, and non Darcy skin. Total skin effect, s,
for a well consists of a number of components as follows:
s = s d + s c + + s p + s Dq +
s pseudo
(2)

Skin is usually estimated from build-up and draw-down


tests, and the estimate is sometimes augmented by log data

IADC/SPE 81639

and special core tests that establish the absolute and effective
permeabilities of the reservoir.
Non-Darcy Skin
When local flow velocities in the reservoir are high (for
instance, due to high rates of inflow, or due to permeability
damage), flow need no longer be laminar, as implicitly
assumed by Darcys inflow relationships. In such a situation,
the additional inflow losses are modeled as a Non-Darcy
skin. The non-Darcy Skin, which in the above inflow
equations appears for the gas inflow equations is given by
S N D = Dq

(3)

where D, the non-Darcy factor, is calculated as


D=

6 10 5 g k s0.1 h

g rw h 2perf

(4)

Note in Eq. (4) the appearance of the permeability in the


near well-bore damaged region. In this work, the non-Darcy
factor is calculated assuming the near wellbore is undamaged
even in a conventional well. This assumption is made solely
for the purposes of estimating the non-Darcy factor. This
assumption is conservative, in that it does not penalize a
conventional well for near-wellbore damage by imposing a
higher non-Darcy factor, but does penalize the UBD well with
a higher non-Darcy skin as UBD flow rates will generally be
higher.
Horizontal and Inclined Wells
As mentioned earlier, elliptical flow is assumed in horizontal
producing sections. Anisotropy also enters the equations in the
form of the anisotropy index Iani,
I ani = k H kV

(5)

The anisotropy index is assumed known for a given candidate


field (thus, both vertical and horizontal permeabilities are
required inputs for horizontal wells). Indeed, the degree of
anisotropy is often used as the basis for the election to drill
horizontal wells in a given field.
Inclined wells are modeled by using the vertical well
IPR, but modifying the skin to include inclination and partial
completion effects. In this work, the approach presented by
Cinco-Ley et. al. (1975) is used.
Other Reservoir Parameters of Note
Most of the other reservoir parameters used in the above
equations are familiar to production and reservoir engineers.
However, a few of them deserve additional attention, in the
context of UBD wells.
Reservoir pressure is the average reservoir pressure at
the time of applying UBD. In the above equations, especially
for pseudo-steady state applications, this assumption is
customarily made.
Vertical and horizontal permeability usually refer to
the absolute undamaged reservoir permeabilities. It should be

noted that some production and reservoir engineers capture the


damage in a reduced near-wellbore permeability rather than a
skin. This is a perfectly valid approach. However, in
choosing this approach, care must be taken that damage is not
redundantly captured in both skin and permeability reduction.
Relative permeability of the hydrocarbon phase is used
for undersaturated reservoirs, and refers to the primary
hydrocarbon phase being produced.
Length of producing section (vertical height or
horizontal length) refers to the length of section contributing
to flow, and not the geometric length of the openhole section.
This is particularly important in considering the length of
producing section for the conventional well. In partially
completed vertical and inclined wells, this is implicitly taken
into account in the Cinco-Ley et. al. correlation. However, in
horizontal wells, it is important to use the actual length of
reservoir producing hydrocarbons. In the absence of
production logs, this can only be estimated. Many reservoir
engineering text books elaborate upon this point (see, for
example, Craft and Hawkins).
PVT Correlations
Using inflow equations to predict inflow requires knowledge
of hydrocarbon properties. For oil wells, the formation volume
factor B0, the gas fraction Rs, and the viscosity of the gas and
liquid fractions are of interest. For gas wells, compositional
PVT is usually of interest, where properties such as gas
compressibility and viscosity are desired. In practice,
hydrocarbon properties are calculated using PVT correlations.
Several PVT correlations have been developed over time,
some of them very specific to the reservoir fluid under
consideration. Gregory (2000) and Craft and Hawkins (1991)
give detailed reviews of the PVT models currently used in the
industry. We present below the PVT models used in this
work. It should be noted that any of the other models can be
easily incorporated into this work.
PVT Model for Oil Systems
For oil and low GOR systems, the Vasquez and Beggs (1980)
Black Oil correlation is used for the calculation of the gas
fraction and formation volume factor. The bubble point is also
calculated (with the assumption that the gas fraction is equal
to the GOR). Although several other Black Oil correlations
have been developed, Vasquez and Beggs is a good generalpurpose correlation applicable for a wide range of crude oils.
Other correlations that are more specific to a given fluid can
be used, and as mentioned earlier, can be easily implemented
in the model.
The viscosity of the black oil is calculated using the
Beggs and Robinson (1975) viscosity model. Accordingly,
the dead oil viscosity is first calculated, and then the viscosity
of the live crude (with solution gas) is calculated based on the
pressure and its relationship to the bubble point.
PVT Model for Gas and Condensate-Gas systems
For gas and condensate-gas systems, the compressibility of the
gas is the key thermodynamic property that is predicted by the
PVT model. In addition, viscosity of the gas is also a critical
transport property. In this work, both a fully compositional as
well as a simplified PVT model based on gas gravity alone are

used.
If the gas gravity is unknown, it is found from a userdefined composition, using the Kay (1936) mixing rule, and
the pseudo-reduced temperature and pressure of the mixture
are found using the Standing and Katz (1942) correlation. At
present, the compositional module allows for the definition of
a compositional system up to C10 (decane) for hydrocarbons,
as well as the typical inorganic gases. When the gas gravity is
given, the Sutton (1985) correlation is used to find the pseudocritical properties.
From the pseudo-critical properties, the Dranchuk and
Abou-Kassem (1975) equation-of-state is used to calculate the
gas compressibility of the mixture. Other equations of state
are available, such as the Soave-Redlich-Kwong (1949, 1972),
or Peng-Robinson (1975). These models can also be used to
find the compressibility. The Dranchuk and Abou-Kassem
correlation, however, is considered reasonably accurate for a
wide range of conditions, including dense-phase gas systems.
It is a fitted equation, which has been shown to have typical
errors in gas compressibility factor of the order of 0.5%.
However, it requires an iterative solution, and a bisection
method is used to numerically search for the solution.
Regardless of whether the compressibility is based on
compositional data or gas gravity, the viscosity of gas systems
is calculated using the Lee, Gonzalez and Eakin (1966)
approach.
The Gas PVT model, due to its compositional nature, is
validated against published data for gas systems. The
validation exercise shows that the results are within 1.6% for
gas compressibility and within 1% for gas viscosity.
Limitations of Gas PVT Model Used
The compositional PVT model for gas and gas-condensate
wells used in this work is adequate for most purposes.
However, one of the key limitations of this model is that
pseudo (or hypothetical) components, or residual components
are not modeled. This limitation can be addressed by
including these components in the compositional model.
Another limitation may be the equation of state, especially if it
is different from the reservoir models used for traditional
reservoir analysis. One of the future improvements of this
work is the inclusion of additional PVT models and equations
of state.
Concluding Remarks on PVT Modeling
It should be noted here that the above PVT assumptions are
fairly generic, and are intended to provide a reasonable
estimate of the productivity improvement for a given fluid.
PVT behavior has a significant impact on production, and
where possible, the base properties of the fluid should be
measured and correctly represented, over the temperature
range of interest. Fortunately, in this work, we estimate the
productivity improvement, which is comparative (ratio of a
UBD well productivity to that of a conventional well of the
same kind). This reduces the impact of PVT errors. Once the
estimate of productivity improvement is obtained using the
method described here, it can be used in a full reservoir
simulation to estimate the economic impact of UBD verses a
conventional well.

IADC/SPE 81639

Estimation of PIF
Productivity improvement from a UBD (or other
alternate) operation can now be estimated. Given all of the
reservoir parameters for the conventional offset and the UBD
alternative, the PI for each of these is calculated. The ratio of
the calculated PI for the UBD case to that calculated for the
conventional offset is then reported as the productivity
improvement factor, PIF. The following steps outline the
typical procedure.
1. Estimate and enter all the reservoir parameters,
taking care to note the parameters of interest
(reservoir pressure, skin, permeabilities, length of
producing section), and the impact of damage on
some of them.
2. Using appropriate PVT models, estimate the
properties of the fluids. Calculated properties can be
checked against known or previously estimated
properties.
3. Using appropriate IPR equations, estimate the
additional skin elements and the PI for each of the
alternatives being compared.
For UBD operations, the typical approach is to
change the skin, by reducing the formation
damage component of skin to zero
The above approach can also be used to evaluate
stimulation treatments. For Acid treatments or
fracture treatments, either the skin can be
modified, or the permeability can be modified to
reflect the improved permeability that results
from the operation.
4. Calculate the ratio of the PIs.
Statistical Considerations
Probabilistic Estimate of PIF
The above approach can be extended to probabilistic estimates
of productivity improvement by taking into account the
uncertainty in each of the parameters appearing in the
reservoir equations. The main benefit of this approach is that
production can be risked based on the uncertainty estimates
for each of the reservoir parameters. Indeed, this is one of the
advantages of the model presented in this work, as
incorporating parameter uncertainties is at best cumbersome in
most full-blown dynamic reservoir simulators.
The uncertainty estimate can be based on available data,
or can be simple range uncertainties (mean value and standard
deviation), which will be treated as normally distributed in the
probabilistic analysis. This is adequate for the current
exercise, where central tendencies and not extreme values are
of interest. Specific distributions are more appropriate for
some of the reservoir parameters, and can be selected by a
user. In particular, permeability, and skin factor are typically
right-skewed (exhibit a tail skewed to the right). Uncertainty
in each of the key parameters that are used to model damage is
discussed briefly below:
1. Length of producing section: The radius of the
wellbore is known with reasonable accuracy, and its
uncertainty is unlikely to have a significant impact on
the results. The length and height of reservoir section

IADC/SPE 81639

Using Analog Data to Modify Predictions


It is apparent that the probabilistic PIF discussed above is
based entirely on theory, without consideration of data. As
such, it reflects only the parameter uncertainty given the
theoretical approach described earlier. Although the theory
used so far is reasonable, it should be noted that all of the
theory is based on established and existing reservoir equations.
One of the ongoing debates in the UBD industry is whether
existing theory can adequately reflect the value that truly

accrues when UBD is used. The debate is strengthened by


noting that in most applications of UBD, post-UBD reservoir
modeling based on established techniques underestimate
actual production increases unless additional parameters such
as permeability or skin are tweaked.
Such an argument is not without precedent. Witness a
period of time when horizontal wells were first introduced into
the industry. In the absence of specific reservoir inflow
equations, early modelers of inclined and horizontal wells
resorted to an empirical adjustment of the skin factor in the
vertical well inflow equations. The Cinco-Ley et. al. approach
to the analysis of inclined wells is a telling illustration of this
approach. Eventually, new (at that time) inflow equations
appeared in the literature to specifically model horizontal well
production, some of which are used in this work.
Thus, until such time as UBD-specific reservoir models
appear in the literature, it behooves any UBD value estimation
technique to heed analog UBD data in making value
predictions. Prediction updating using data is also intuitively
sound. The fundamental question we are posing here is:
We have made a prediction based on our theoretical model, and
someone comes along with data. Our question then is would
we care to change our mind in any way having seen the data?
The answer of course is, it depends, usually on the
confidence in theory, the relevance of the data to the prediction,
and the confidence in the data.

It is therefore reasonable to say that if a statistically


sound approach is developed to influence theoretical
predictions using analog observations of UBD results, the
resulting value predictions may better reflect expectation.
In this section, the statistical treatment of data made
available by BP is first discussed. Two different techniques to
incorporate data in the model predictions are then discussed.
The technique implemented in this work is then presented.
Treatment of Analog Data
BP has provided analog data on 38 wells. Based on
discussions with BP engineers responsible for this data, it
appears that the data have been gathered by a combination of
literature review and polling of personnel involved in previous
UBD projects. Figure 1 shows the histograms from all of the
data.
As Figure 1 shows, data is right-skewed. Although a
majority of the observations are in a narrow PIF range of 1-5,
there are enough higher-end observations to warrant their
consideration in the data distribution.
Histogram of Data
14

12
10
Frequency

can however have significant impact on the results,


especially if the comparable conventional well has
non-producing sections, either due to localized
damage or due to the fact that certain zones are
deliberately shut-off or not perforated. In general,
reservoir engineers should have a reasonable idea of
the range of producing lengths (for a given openhole
length) for conventional wells, particularly in
development fields where offset data and
stratigraphic maps are available.
2. Permeability: Permeability uncertainty data can
usually be obtained from core analyses, or from
existing reservoir analysis and ongoing reservoir
management efforts. Generally, in development
fields, a reasonable estimate of permeability and its
range can be obtained. Destruction of near-wellbore
permeability is a usual consequence of invasive
damage and should be accounted for in the
estimation, especially for openhole horizontal
completions.
3. Skin: Skin uncertainty is usually the most difficult to
obtain. In general, it is fair to assume that good
drilling and production practices are applied in any
given field. Therefore, it is reasonable to expect that
the skin distribution is right-skewed, with a majority
of observations falling in the lower end of the skin
range. Such a distribution is best approximated using
either a log-normal or inverse-Gaussian (IG)
distribution. IG distributions are discussed in more
detail when discussing the statistical consideration of
analog data.
Although the theoretical normal distribution extends
from to , and the IG (or log normal) from 0 to , most
practical distributions have natural truncations. Generally,
such natural truncations reflect the number of observations
made, and may therefore be regarded as a natural limit to the
central tendencies of the population distribution, and treated as
the 3 limits.
Once the uncertainties in all the parameters of interest
are represented by appropriate distributions, a Monte Carlo
simulation is performed, with the PI for each alternative and
the PIF being calculated every iteration. Once a sufficient
number of simulations are performed (a typical lower limit
being 10,000), the resulting PIF distribution gives a
probabilistic estimate of the expected PIF from applying UBD
(in comparison to a conventional well). As would be expected
from the central limit theorem, the resulting PIF distribution is
usually fairly normal in distribution, with possibly some
skewness introduced by the truncations and the skewed skin
distribution.

8
6

4
2

0
1

15

22

29

36

43

50

PIF

Figure 1. Analog Production Data for UBD Wells (Source: BP)

IADC/SPE 81639

Based on the data, it seems reasonable to model the data


from an inverse Gaussian (IG) distribution with parameters
= 1 / , (where is the mean) and the shape-factor . The
validity of this was tested and confirmed using the approach
described by Natarajan and Mudholkar (2001). The shape
factor of an inverse Gaussian distribution is related to the
variance as

2 =

(6)

An IG distribution is more appropriate for this


application than a log-normal since IG increases the
probability of observations in the central region at the expense
of the probability of observations further down the right tail.
Although a log-normal distribution does the same, it tends to
give a somewhat higher probability to observations further
down the tail. In other words, the IG distribution possesses
the feature of exponential decay in both tails. The IG
distribution is also one of the most generalized distributions,
as it can model distributions ranging from exponential to
Weibull by menas of suitable selection of the shape factor. In
recent times, the IG distribution has been used to model
several practical situations, such as failure rates, repair
expectations, and life estimates. It is usually strongly
recommended for modeling and analyzing asymmetrical data.
The applications of IG distribution are discussed further by
Chhikara and Guttman (1982), Chhikara and Folks (1989) and
Mudholkar and Natarajan (2002).
For purposes of statistical modeling, the data is
approximated using a theoretical, right-truncated IG
distribution. The distribution can also be left un-truncated to
the right without significantly affecting the results.
Data Zonation
The data provided were also examined statistically to evaluate
if further zonation is possible. The data includes detailed
description of each subject well. description of rock type,
initial permeability, type of reservoir (gas/oil), type of well
(horizontal/inclined/vertical), as well as whether the well was
drilled and completed underbalanced, or was killed either on
trips or on completion. These criteria were evaluated to
examine if the UBD results were statistically different for each
of them.
To begin with, the number of data points available is
low. Moreover, the variance in the data is fairly large.
Therefore, we find that the UBD results from data sub-sets
based on above criteria were not statistically different from the
results based on the total data set. The only weak exception to
this was the criterion of whether the well was maintained
underbalanced throughout drilling and completion, or it was
killed at some point during UBD operations. It appears that at
least in the mean sense, wells maintained underbalanced
throughout the operations gave PIF results about 3 times that
from wells that were killed at some point during the UBD
operations. However, the number of UBD completed wells
was very small, and therefore not considered sufficient at this
stage to zone the two cases as different IG distributions.

Combining Data and Theory


Several techniques were studied to identify the best approach
to combining data and theory. Some of the techniques were
too simplistic and statistically unsound, or fairly esoteric and
complex to implement, or not suitable for the current problem.
Of the techniques studied, two that appear to be most suitable
to the current problem are discussed below.
Bayesian Approach
The fundamental question we are posing here of whether we
would care to update our predictions based on data is a
typically Bayesian one. Empirical Bayesian analysis can
therefore be used to answer this question. Suppose sample
observations are drawn from a distribution given by its
pdf f ( x | , ) , where the parameters are not fixed but are
assumed to have a prior distribution. The highly right-skewed
PIF data x = ( x1 ,..., x n ) , is modeled from an inverse
Gaussian distribution with parameters = 1 / , where the
mean parameter is re-parameterized in order to ease Bayesian
analysis. Based on the Monte-Carlo simulations using the
inflow models and parameter uncertainties, we have described
a prior distribution of PIF. However, in classical Bayesian
analysis, it is usually desired to update a characteristic statistic
of the prior distribution [Banerjee and Bhattacharya (1979)],
in this case , the re-parameterized mean.
Since the distribution of the theoretically predicted PIF is
normal, the prior distribution for conditional on can be
assumed to be from a normal distribution say, N( 0 , 1 / ),
where 0 is the re-parameterized mean of the prediction.
Further, it can be assumed that the distribution for is
proportional to 1 / , being a locally uniform reference prior,
see Banerjee and Bhattacharya (1979).
Given the above assumptions, it is our desire to obtain an
updated posterior distribution of PIF.
The likelihood
function l ( , | x1 ,..., x n ) , based on a random sample from
IG( , ) is expressed as
n
x
1
l ( , | x1 ,..., xn ) = c exp{ nv [1 + ( ) 2 ]} 2
2
v
x

In Eq. (7), x =

1
n

xi , v =

i =1

1
n

x
i =1

(7)

1
, and c is an
x

integrating constant of the inverse Gaussian distribution.


Then, P ( , | x1 ,..., xn ) , the posterior distribution of the reparameterized mean conditional on the data is given by
P( , | x1 ,..., xn ) = l ( , | x1 ,..., xn ) P ( | ) P ( ) ,
(8)

( 0 ) 2 } , c1 being the
2
normalizing constant, and P ( ) 1
For the present problem, given that the mean from the
theoretical predictions is normally distributed, and that the

where P( | ) = c1 exp{

IADC/SPE 81639

data is from an IG population, the posterior distribution can be


shown to be given by
P( , | x1 ,... x n ) = K exp{( s + t )}
1
where s = nv[1 +
2

( n +1) 2
2

(9)

1
x ( ) 2
x ] , and t = 1 ( ) 2 .
0
2
v

This is a gamma distribution. The shape factor is still


retained in the above. In the Bayesian approach here, the only
statistic that we are updating is the mean. Therefore, we
assume, as is customary in such analyses, that the entire range
of is possible, and integrate Eq. (18) with respect to , with
ranging from 0 to .
This results in the posterior
distribution of , i.e.

P( | x1,...,xn ) = K1((n +1) / 2)[1+

1
2

nq

n + 0

1+ nx
0 < <

)2 ](n+1) / 2 ,
(10)

Equation (10) represents a left truncated t-distribution (as

ranging from 0 to , and the standard t-distribution is

from to ,) with n degrees of freedom. Here, n is the


number of data points. Although a t-distribution has n-1
degrees of freedom, we have introduced a theoretical
prediction as an additional point in the Bayesian analysis.
The location parameter m of the t distribution is
m = (n + 0 ) /(1 + n x ) ,

and the scale parameter q is given by


x (1 + n x ) 2
,
n[(1 0 x ) 2 + (1 + n x )v x ]
the point of truncation being zero.
q2 =

In other words, a posterior * = ( m) / q has a


Students t-distribution with n degrees of freedom, which is
truncated at (-m/q).
As we are interested in the distribution of the mean of the
posterior PIF, and since Eq. (10) gives us the distribution of
= 1 / , the reciprocal of the Eq. (10) needs to be
considered. The distribution of the reciprocal of Eq. (19) (note
the truncation) is then the distribution of the mean of the
posterior, i.e., an estimate of the updated mean of the PIF,
given data. The highest posterior density (HPD) over a typical
interval can then be obtained to give an estimate of the
distribution over the interval of interest. This is analogous to
the frequentist confidence level. For our problem, the
typical HPD of interest is the interval between 10% and 90%
probability.
While this approach updates a specific statistic of the
prior distribution given data, it does not update the entire
distribution, which is of more general interest in this problem.
Moreover, the Bayesian approach does not incorporate the

variance of the prior PIF distribution (recall that has been


integrated out). The only variance remaining is that of the
data itself. As a result, it treats the theoretical prediction as
only one additional point added to the existing data points.
Thus, it is significantly weighted towards the data. This may
in fact be an appealing feature of this approach (particularly to
frequentists).
Additional work significantly beyond the scope of the
current work will be needed to develop a more general
Bayesian approach that allows the incorporation of the
theoretical variance (which is, after all, representative of the
uncertainty of reservoir parameters), and allows the updating
of the entire PIF distribution (or at least the percentiles of
interest). It is also of interest to consider a point-by-point
Bayesian updating, where each data point is considered
individually alongside the theoretical prediction and the
prediction is incrementally updated. These are subjects of
future work.
Weighted Mixture Sampling Approach
A second approach considered in this work is that of a
weighted mixture sampling approach. As before, our desire
is still to update the theoretical distribution of PIF using actual
field data. And as before, the data is taken to be from an IG
population. In this approach, however, we will obtain an
updated distribution by sampling from both distributions based
on some rule. Statistical considerations warrant that the
proportion of combining the distributions is inversely
proportional to the variance. In this sense, the inverse of
variance is representative of the informational value of the
distribution. This is logical, as it essentially argues that the
less the variance, the more certain, in a sense, we are about
that particular information. Thus, if t is the weight or value
of the theoretical distribution, and d is the weight of the data
distribution, the updated distribution is constructed by
randomly sampling from the theoretical PIF distribution t of
the time, and from the data distribution the remaining time.
The sampling is randomized by basing the choice on the value
of a uniform random variable.
While the above approach is appealing in that it treats the
two distributions in proportion to their informational value, it
has some limitations. In the current problem, the variance of
the data distribution is far in excess of the typical variance
from the reservoir model, except when there are significant
variances in the estimates of reservoir parameters themselves.
Therefore, in typical application, theoretical prediction will
dominate the updated distribution.
Discussion of Updating Approaches
In a sense, therefore, the two approaches discussed here reflect
two extremes from the perspective of the impact of data on the
updated distribution. In the first approach, data dominates the
updated distribution, while in the second, theoretical
prediction typically dominates. Reality, as always, probably
lies somewhere in between. It is possible to arbitrarily assign
different weights to the theoretical and data components to
obtain different results, but unless based on statistical
considerations or extensive experience, such an approach is

IADC/SPE 81639

fraught with error.


For the Bayesian approach, additional statistical work
will be needed, as mentioned earlier, to generalize the
approach to allow a more intuitively acceptable updating of
the prior predictive distribution. It is our strong
recommendation that such work be vigorously undertaken.
Such work will also benefit from better, higher quality data.
The basic weighted mixture approach discussed above
must also be modified to improve the predictive capability of
the approach. As noted earlier, the approach relates the
informational value of the distribution to the inverse of the
variance, and therefore will always tend to be (unfairly, it
would seem) swayed towards theory. This completely misses
the possibility that the theory itself is incorrect or incapable of
accurately predicting UBD PIF. In many probabilistic design
approaches, this aspect is often captured in a term known as
the model error.
While the inverse of the variance is a reasonable basis
for valuing the data, it is a flawed basis for valuing theory.
This is because the variance as calculated from the Monte
Carlo simulations is a result of the uncertainties in the
reservoir parameters. It is adequately reflective of the
estimated uncertainties in the parameters, but it does not by
itself convey any information about the ability of the model to
accurately predict UBD PIF within the assumed parameter
uncertainties. Intuitively, then, it seems that the informational
value in theory as reflected by the variance in the predicted
distribution should somehow be devalued or discounted to
reflect its inaccuracy. This can be achieved by capturing the
model inaccuracy in a model error term that will then have to
be separately calibrated.
Suppose that the predictions of the model (Theoretical
PIF) for a given well can be compared to the actual results
from applying UBD to that well for a suitably selected anchor
point on the distribution.
Suppose that several such
comparison can be made, giving an estimate of the distribution
of discrepancy of the theory with respect to data, for any
selected statistic (say the mean, or some preferred percentile).
Then the variance of the error distribution, 2 , can be
estimated, and the proportion of the theoretical distribution in
the updated distribution can be taken to be

t* = t

(11)

where

d2 + t2

( t + )2 + d2

(12)

The factor , which by Eq. (12) is always less than 1,


may be treated as a discounting or devaluing factor that
devalues the proportion of the theoretical distribution in the
updated PIF distribution in a manner proportional to the error
in the theory itself.
For purposes of illustration in this work, the error factor
is set to 0.9. Further work is clearly in order to characterize
and calibrate the error factor . One beneficial aspect of this
approach is that the characterization of the error factor
improves with each successive application of the model and
comparison to actual UBD results.

Discussion of Probabilistic Estimation of PIF


The reservoir model used here is fairly simplistic in that it
does not consider dynamic reservoir behavior, multi-well
developments, well-to-well interference, and geophysical
features that affect production (compartmentalization, barriers
to flow, special geological structures, etc.). Also note that the
above model only predicts the PIF, i.e., sand-face flow, and
not wellhead deliverability. Prediction of wellhead
deliverability requires the use of an upflow model, which is
provided by most standard production engineering simulators.
To some extent, the ratio of the UBD and conventional
well PIs dampens the effect of error in theoretical predictions
themselves.
The resulting PIF distribution can also
incorporate to some extent the experience of the reservoir
engineers and production engineers close to the problem, and
thus capture valuable risk information that may be used as a
basis of candidate selection. Even though the results are for PI
and not deliverability, for comparable wells, deliverability is
clearly proportional to the PI. For mature fields and tight (low
permeability) reservoirs, PI is directly indicative of
productivity since flow is usually inflow limited and not
upflow limited.
Moreover, for purposes of candidate
selection, improvement in PIF directly translates to
improvement in production and hence improvement in NPV.
Because this model only predicts the PIF (and cannot
capture ultimate recovery benefits) and because the model
does not include sophisticated reservoir models, it is
recommended that the resulting PIF values be used in a fullblown dynamic reservoir simulator to evaluate field-level
economics with multi-well campaigns.
The use of
comprehensive and integrated dynamic simulators to plan the
development approach and estimate field economics is
common in most operating companies. Using the P(10), P(50)
and P(90) values as a guide to the assumptions made in the
field-specific reservoir model, field level development and
economic impact of UBD can be estimated for an upside,
expected value and downside scenario.
Such a reservoir-based economic evaluation has many
benefits.
It can be used as a reasonable basis for
candidate selection
It can be used to estimate the number of wells
needed and can impact the development plan of
the field (possibly reducing the number of
wells needed for UBD, and thus adding to
NPV)
It can also capture the ultimate recovery
improvement possible using UBD. As
discussed earlier, this can be substantial.
It makes a probabilistic argument that
intrinsically appeals to decision makers and
asset management, and changes the paradigm
of UBD candidate selection from a drilling
technique to a reservoir exploitation technique.
Implementation of Solution Approach
The solution approach is implemented using an Excel-based
spreadsheet program, with an @Risk add on to enable easy
consideration of a wide range of probability distributions.
Excel is chosen as the platform owing to its familiarity to

IADC/SPE 81639

engineers. Both the updating approaches are applied, although


the examples discussed here use the weighted mixture
sampling approach.

Table 3: Summary statistics of PPIF distribution for Example 1


Statistical Summary PIF Simulation W History
PIF
Mean of PIF
1.999
PIF

Examples
Following are two examples to illustrate the use of and results
from the PI Calculation program. The first case is a horizontal
gas well and the second is a vertical oil well. Input data,
history data, and simulation results for the two examples are
given in this section. @Risk is used for defining different
input and output distribution functions and for performing
Monte Carlo simulations. Ten thousand simulations are run for
each example.
General Assumptions
Specific input data for each example is discussed later. For
both examples, input distributions were defined for average
reservoir pressure (both conventional and UBD well) and for
conventional skin. In Example 1, UBD skin is taken to be
zero, while in Example 2, a distribution is used for UBD skin.
Distribution of reservoir pressure for both conventional and
UBD well is defined as normal with mean 2700 psi and
standard deviation 100 psi, truncated at 2400 psi and 3000 psi.
The other key assumptions are:
Reservoir height = 180 ft
Wellbore Diameter = 6, Drainage radius =
2500 ft
Reservoir temperature = 220 F
Example 1
Example 1 is a horizontal gas well with pseudo-steady state
reservoir inflow.
Input data are given in two columns for
conventional and UBD well respectively. Table 1 shows the
reservoir specific data assumptions. The gas gravity of the gas
is assumed to be 0.572. For conventional skin, we consider an
inverse Gaussian distribution with = 2 and = 1, where is
the arithmetic mean and is defined by Eq. (6). UBD skin is
assumed to be deterministic at zero.
Table 1: Well specific reservoir information for Example 1
Specific Param eters
Horizontal Reservoir Perm eability (m d)
kH
Vertical Reservoir Perm eability (m d)
kV
Relative Perm eability of G as
k rg
Length of Horizontal Reservoir Section (ft)
L

Conv

UBD

10

10

0.9

0.9

500

500

Results
Table 2 summarizes the results without data-based updating,
while Table 3 summarizes the results with updating.
Table 2: Summary statistics of PIF (Prior PIF) for Example 1
Statistical Summary PIF Simulation W/O History
PIF
Mean of PIF
1.255
PIF

Standard Deviation of PIF


Variance of PIF
Max of PIF
Min of PIF
Percentiles
90% prob of PIF greater than
50% prob of PIF greater than
10% prob of PIF greater than

Standard Deviation of PIF


Variance of PIF
Max of PIF
Min of PIF
Percentiles
90% prob of PIF greater than
50% prob of PIF greater than
10% prob of PIF greater than

3.294
10.848
47.800
0.001
1.065
1.460
2.146

Although the tables show mean and variance, the


percentile results are more useful in this problem. From Table
2, ignoring data and relying entirely on theory, we find that
there is a 50% chance that the PIF is greater than 1.095
(marginally better than conventional) but there is an upside
(10% chance) that the PIF is greater than 1.6. Table 3, where
the distributions are updated using data, suggests that there is
now a 50% chance that the PIF is greater than 1.46, and an
upside (10% chance) of the PIF being in excess of 2.15.
Thus, data has updated the original probabilistic results, giving
a more reasonable expectation.
Example 2
This is a vertical oil well example. Table 4 lists the reservoir
assumptions.
For this example, we define the skin distribution for both
conventional and underbalanced skins as IG distributions, but
with different parameters. For conventional skin, we use an IG
distribution with a mean of 4 and and shape factor of 1. The
UBD skin is also modeled as an IG distribution with a mean of
1, shape factor of 1, and a shift of -2, so that the skin lies
between -2 and 0, with a mean of -1
Table 4: Well specific reservoir information for Example 2

Specific Parameters
Reservoir Permeability (md)
k
Relative Permeability of Oil
kro

Conv

UBD

80

80

0.9

0.9

Results
Tables 5 and 6 are the key results of interest for Example 2.
From Table 5, note that the variance is lower in this case,
although more parameters are risked. From this table, we find
that based solely on theory, there is a 50% chance that the PIF
is greater than 1.1 (marginally better than conventional) but
there is an upside (10% chance) that the PIF is greater than
1.4. The skewness is not as great here as it is in Example 1.
The updated PIF considering analog data are given in Table 6.
This time, the updated PIF suggests that there is now a 50%
chance that the PIF is greater than 1.8, and an upside (10%
chance) of the PIF being in excess of 2.98. The skewness has
increased significantly once data is given some consideration.

0.496
0.246
12.259
0.993
1.029
1.095
1.610

It is noted here that if the empirical Bayesian approach is used


instead, the updated mean PIF for this example is close to 4.0.
However, the updated mean itself has a distribution with variance
close to that of the original data.

10

IADC/SPE 81639

Table 5: Summary statistics of PIF distribution for Example 2


Statistical Summary PIF Simulation W/O History
PIF
Mean of PIF
1.183
PIF
Standard Deviation of PIF
0.261
Variance of PIF
0.068
Max of PIF
5.366
Min of PIF
0.956
Percentiles
90% prob of PIF greater than
1.021
50% prob of PIF greater than
1.097
10% prob of PIF greater than
1.426

Table 6: Summary statistics of PPIF distribution for Example 2


Statistical Summary PIF Simulation W History
PIF
Mean of PIF
2.355
PIF

Standard Deviation of PIF


Variance of PIF
Max of PIF
Min of PIF
Percentiles
90% prob of PIF greater than
50% prob of PIF greater than
10% prob of PIF greater than

3.147
9.901
47.806
0.000
1.131
1.815
2.980

Conclusions
The ability to estimate the value from applying UBD on a well
or in a field is clearly extremely important to the industry. It
underpins and drives candidate selection. This work has
attempted to develop a method based on rigorous reservoir and
statistical principles to enable such estimation.
A probabilistic productivity improvement factor
estimation tool has been built to evaluate UBD candidates and
estimate the value from applying UBD in comparison to a
conventional alternative. The theoretical approach has been
described in detail, and incorporates sufficient rigorous
reservoir mechanics to give reasonable estimates of PI.
Probabilistic treatment is applied in two stages- (a)
probabilistic
consideration
of
reservoir
parameter
uncertainties, and (b) consideration of data in updating the
predictions.
In a perfect world, if the theory were excellent or shown
to be consistently accurate with respect to empirical
observations, we will not need to consider data each time we
make a prediction. This is the case with all mature theories.
In that case, we only need to theoretically predict a
probabilistic PIF where reservoir parameter uncertainties are
considered. However, in the imperfect world we currently
inhabit in the UBD industry, it is important to give due
credence to analog data in any value estimation. Clearly,
relying entirely on either theory or data is inappropriate at the
current stage of development of UBD. Therefore, this work
has concentrated on developing a statistically-based approach
to update theoretical predictions based on data observations.
Two statistical approaches to updating using data are
discussed here. Both methods can be improved. An error
function is also proposed here to improve the predictions. It
should be noted, more importantly, that regardless of
technique, the quality of the data has a significant impact on
the conclusions derived from using these methods. Therefore,
improvement in data quality and acquisition of more data
should naturally form an important basis of the continued
improvement of this work.

The methods described here have been implemented on


an Excel platform with Visual Basic programs, and with
@Risk providing the statistical package necessary for
statistical considerations. Two examples have been shown to
illustrate the use of this program.
The tool described here is not meant to be an accurate
estimator of the economic value from applying UBD. It is
meant to be a probabilistic candidate selection tool that gives
statistically useful percentile values of the improvement in PI
that can be expected using UBD. It is our recommendation
that these PI values then be used to guide a comprehensive and
more sophisticated reservoir simulation to perform field-level
economics and estimate the economic impact of UBD in a
probabilistic manner. For example, the P(10), P(50) and P(90)
values of the updated PIF can be used to estimate the P(10),
P(50) and P(90) economic impact from UBD, using a fullblown reservoir simulator.
Further Work
In the opinion of the authors, this is a work in progress, a first
step in solving a vexing problem in UBD. We believe that
furtner work to improve upon the methods proposed here will
immensely benefit UBD candidate selection and contribute to
the growth of the industry. The statistical questions posed by
this application are fundamental and important to both our
industry as well as to statistical science, and merit further
attention. The inflow equations can be improved. Better
characterization of the reservoir can be incorporated. The
predictive accuracy can be improved by continuous data
gathering and evaluation of results. Data volume and quality
also can be improved. We present below some of the areas of
improvement that merit attention.
Establishment of a database with the purpose of
gathering and evaluating data from UBD applications
worldwide. Such a database would significantly
improve the accuracy of the results. It would also
allow zonation so that for any given application, the
most appropriate analog data set can be used from the
data base.
Improvement of the reservoir inflow models and
assumptions used in the theoretical development,
including better PVT and inflow models, and
consideration of permeability as function of axial and
radial coordinates.
Improvements in statistical updating methods, and
consideration of additional updating methods.
Calibration of error distribution and characterization
of the error in current reservoir models.
Improvement to the theoretical models themselvesi.e., development of a reservoir model that more
accurately models the impact of UBD on production
and ultimate recovery. This is, needless to say, one
of the holy grails of our industry, and will need
substantial commitment of time and effort.
Acknowledgements
The authors are grateful to BP for providing the data and
funding for the work reported here, and for permitting its
publication.

IADC/SPE 81639

Nomenclature
Bo = Oil Formation Volume Factor, res bbl/STB
D = non-Darcy coefficient, (STB/d)-1
h = Reservoir thickness, ft
Iani = Index of permeability anisotropy,
dimensionless
k = Absolute Permeability, md (for Vertical Well)
ko = Effective Oil permeability, md
kH = Horizontal Absolute Permeability, md
kro = Relative Oil permeability, dimensionless
krg = Relative Gas permeability, dimensionless
ks = near-wellbore permeability, md
kV = Vertical Permeability, md
L = Horizontal well length, ft
p = Reservoir Pressure, psi
pe = Constant reservoir pressure at the boundary
(re), psi, (used in steady-state calculation)
p = Average reservoir pressure, psi, (used in
pseudo- steady-state calculation)
pwf = Wellbore Pressure, psi
s = skin effect, dimensionless
sd = damage skin
sc+ = skin due to partial completion and slant
sp = perforation skin
s Dq = Non-Darcy skin

spseudo = pseudo skin (phase- and rate-dependent effects)


T = Reservoir temperature, o F
Z = Gas Compressibility, dimensionless
References
1.

Banerjee, A. K., and Bhattacharyya, G. K., (1979),


Bayesian. Results for the Inverse Gaussian Distribution
with an Application, Technometrics, 21, 247-251.
2. Beggs, H. D, and Robinson, J. R., (1975), Estimating the
Viscosity of Crude Oil Systems, J. Petrol. Technology.
3. Chhikara, R. S., and Folks, J. L., (1989), The Inverse
Gaussian Ditribution: Theory, Methodology and
Applications, Marcel and Dekker, Inc., New York.
4. Chhikara, R. S., and Guttman, (1982), Prediction Limits
for the Inverse Gaussian Distribution, Technometrics, 24,
319-324.
5. Cinco-Ley, H., Ramey, H. J., and Miller, F. G., (1975)
Pseudoskin Factors for Partially Penetrating Directionally
Drilled Wells, SPE Paper 5589.
6. Craft, B. C., and Hawkins, M. F., (1991), Applied
Petroleum Reservoir Engineering, Prentice Hall,
Englewood Cliffs, New Jersey.
7. Dranchuk, P. M., and Abou-Kassem, J. H., Calculation of
Z Factors for Natural Gases Using Equations of State, J.
Canadian. Petrol. Technology, July-Sept.
8. Economides, M. J., Hill, D., and Ehlig-Economides, C.,
(1993), Petroleum Production Systems, Prentice Hall
PTR, New Jersey.
9. Gregory, G. A., (2001) Hydrocarbon Fluid Properties and
Phase Behavior, Notes for a Short Course, Neotechnology
Consultants, Canada, 2001.
10. Kay, W. B. (1936), Density of Hydrocarbon Gases and
Vapor at High Temperature and Pressure, Ind. Eng.
Chem., September.
11. Lee, A. L., Gonzalez, M. H., and Eakin, B. E., (1966), The
Viscosity of Natural Gases, J. Canadian. Petrol.
Technology, Vol. 18.

11

12. Mudholkar, G. S., and Natarajan, R., (2002), The Inverse


Gaussian Models: Analogs of Symmetry, Skewness and
Kurtosis, Annals of the Institute of Statistical
Mathematics, Vol. 54, No. 1, 138-154.
13. Natarajan, R., and Mudholkar, G. S., (2001), MomentBased Tests for Inverse Gaussian Models, Technical
Report # 297, Department of Statistical Science, Southern
Methodist University, Dallas, TX (submitted).
14. Peng, D. Y., and Robinson, D. R., (1976), A New TwoConstant Equation of State, Ind. Eng. Chem. Fund., Vol.
15.
15. Redlich, O., and Kwong, J. N. S., (1949), The
Thermodynamics of Solutions. V. An Equation of State.
Fugacities of Gaseous Solutions, Chemical Reviews, v.
44, 1949. Soave, G., (1972), Equilibrium Constants from
a Modified Redlich-Kwong Equation of State, Chem. Eng.
Sci., Vol. 27. (The combination of above two is referred to
as the SRK equation of state).
16. Standing, M. B., and Katz, D. L., (1942), Density of
Natural Gases, Transaction of the AIME, Vol. 146, 1942.
17. Sutton, R. P., (1985), Compressibility Factor for High
Molecular Weight Reservoir Gases, SPE 14265, 1985.
18. Vasquez, M., and Beggs, H. D., (1977, 1980),
Correlations for Fluid Physical Property Prediction, SPE
6719. See also J. Petrol. Technol.

Anda mungkin juga menyukai