Documente Academic
Documente Profesional
Documente Cultură
Summary have been obtained if the well flowed at constant rate for the
Pressure/rate deconvolution is a long-standing problem of well-test duration of the whole test. This is the pressure/rate deconvolution
analysis that has been the subject of research by a number of problem. Pressure/rate deconvolution has been a subject of re-
authors. A variety of different deconvolution algorithms have been search by a number of authors over the past 40 years. Pressure/rate
proposed in the literature. However, none of them is robust enough deconvolution reduces to the solution of an integral equation. The
to be implemented in the commercial well-test-analysis software kernel and the right side of the equation are given by the rate and
used most widely in the industry. Recently, von Schroeter et al.1,2 the pressure data acquired during a test. This problem is ill con-
published a deconvolution algorithm that has been shown to work ditioned, meaning that small changes in input (test pressure and
even when a reasonable level of noise is present in the test pressure rates) lead to large changes in output result—a deconvolved con-
and rate data. In our independent evaluation of the algorithm, we stant-rate pressure response. The ill-conditioned nature of the pres-
have found that it works well on consistent sets of pressure and sure/rate deconvolution problem, combined with errors always
rate data. It fails, however, when used with inconsistent data. Some present in the test rate and pressure data, makes the problem highly
degree of inconsistency is normally present in real test data. unstable. A variety of different deconvolution algorithms have
In this paper, we describe the enhancements of the deconvolu- been proposed in the literature.3–8 However, none of them is robust
tion algorithm that allow it to be used reliably with real test data. enough to be implemented in the commercial well-test-analysis
We demonstrate the application of pressure/rate deconvolution software used most widely in the industry.
analysis to several real test examples. Recently, von Schroeter et al.1,2 published a deconvolution
algorithm that has been shown to work when a reasonable level of
Introduction
noise is present in test pressure and rate data. In our independent
The well bottomhole-pressure behavior in response to a constant- implementation and evaluation of the algorithm, we have found
rate flow test is a characteristic response function of the reservoir/ that it works well on consistent sets of pressure and rate data. It
well system. The constant-rate pressure-transient response depends fails, however, when used with inconsistent data. Examples of
on such reservoir and well properties as permeability, large-scale such inconsistencies include wellbore storage or skin factor chang-
reservoir heterogeneities, and well damage (skin factor). It also ing during a well-test sequence. Some degree of inconsistency is
depends on the reservoir flow geometry defined by the geometry almost always present in real test data. Therefore, the deconvolu-
of well completion and by reservoir boundaries. Hence, these res- tion algorithm in the form described in the references cited cannot
ervoir and well characteristics are reflected in the system’s con- work reliably with real test data.
stant-rate drawdown pressure-transient response, and some of In this paper, we describe the enhancements of the deconvolu-
these reservoir and well characteristics may potentially be recov- tion algorithm that allow it to be used reliably with real test data.
ered from the response function by conventional methods of well- We demonstrate application of the pressure/rate deconvolution
test analysis. analysis to several real test examples.
Direct measurement of constant-rate transient-pressure re-
sponse does not normally yield good-quality data because of our Pressure/Rate Deconvolution
inability to accurately control rates and because the well pressure
In a linear system, the well pressure during a variable-rate test is
is very sensitive to rate variations. For this reason, typical well
given by the convolution integral
tests are not single-rate, but variable-rate, tests. A well-test se-
quence normally includes several flow periods. During one or t
dpu共t − 兲
more of these flow periods, the well is shut in. Often, only the
pressure data acquired during shut-in periods have the quality re-
p共t兲 = p0 − 兰q共兲
0
dt
d. . . . . . . . . . . . . . . . . . . . . . . . . . (1)
quired for pressure-transient analysis. The pressure behavior dur-
ing the individual flow period of a multirate test sequence depends Here, q(t) is the well rate, p(t) is the well bottomhole pressure, and p0
on the flow history before this flow period. Hence, it is not the is the initial reservoir pressure. The pu(t) in Eq. 1 is the rate-
same as a constant-rate system-response function. The well-test- normalized pressure response to constant-rate production, assuming
analysis theory that evolved over the past 50 years has been built that, at the beginning of production, the reservoir is in equilibrium and
around the idea of applying a special time transform to the test the pressure is uniform throughout the reservoir. Eq. 1 is known as
pressure data so that the pressure behavior during individual flow Duhamel’s integral and is an expression of the principle of superpo-
periods would be similar in some way to constant-rate drawdown- sition resulting from the linear character of the system.
pressure behavior. The superposition-time transform commonly Pressure/rate deconvolution aims at reconstruction of the con-
used for this purpose does not completely remove all effects of stant-rate drawdown-pressure response pu(t) along with the initial
previous rate variation. There are sometimes residual superposition reservoir pressure p0 from the pressure and rate data, p(t) and q(t),
effects left, and this often complicates test analysis. acquired during a variable-rate well test. This problem is equiva-
An alternative approach is to convert the pressure data acquired lent to solving Eq. 1 for pu(t) and p0 given p(t) and q(t) measured
during a variable-rate test to equivalent pressure data that would during the well test. Several methods for solution of this integral
equation have been proposed in the literature.3–8 However, these
solution algorithms proved to be unstable and could not tolerate
errors normally present in well-test data.3 A review of several
Copyright © 2005 Society of Petroleum Engineers
solution algorithms described in the literature is given in Ref. 1.
This paper (SPE 84290) was first presented at the 2003 SPE Annual Technical Conference
and Exhibition, Denver, 5–8 October, and revised for publication. Original manuscript re-
ceived for review 29 March 2004. Revised manuscript received 21 December 2004. Paper
von Schroeter Deconvolution Algorithm. Recently, von Schroe-
peer approved 25 January 2005. ter et al.1,2 presented a new deconvolution algorithm that dem-
冋 册 冋 册
skin pressure drop change during the test sequence, then the test
dpu共t兲 dpu共兲 data are inconsistent with Eq. 1. It means that Eq. 1 is not the
z共兲 = ln = ln , . . . . . . . . . . . . . . . . . . . . . . . . (2)
d ln t d correct superposition model for the test data in this case. Process-
ing these test data through a deconvolution algorithm based on Eq.
where ⳱ln(t). In terms of the function z(), the convolution Eq. 1 1 will not produce a physically meaningful response function pu(t).
reduces to As we discuss later, testing of our variant of the deconvolution
ln t algorithm confirms this potential problem. Deconvolution fails
p共t兲 = p0 − 兰q共t − e 兲 e
−⬁
z共兲
d. . . . . . . . . . . . . . . . . . . . . . . . . . . (3) when used on simulated test data that have different wellbore-
storage coefficients during different flow periods. Moreover, this
testing shows that an inconsistency localized at early-time flow-
Selection of z() as a new solution variable ensures that dpu(t)/d lnt is period data (like that caused by changing wellbore storage) does
positive—the necessary condition that a constant-rate system re- not produce a local early-time effect on pu(t). In fact, it severely
sponse must satisfy. However, an unwelcome consequence of this affects the response pu(t) at late time.
variable selection is that the resulting Eq. 3 becomes nonlinear. This is a fundamental problem of the deconvolution algorithm
• Regularization by Curvature. Even with the above selection based on Eq. 1. It precludes us from using it with most real test
of solution variable, the problem is still ill conditioned and is very data. Real test data often show the early-time pressure behavior
sensitive to noise in pressure and rate data. Imposing additional that changes from one flow period to the next. There are two
constraint in the form of a penalty on the curvature of z() enforces possible approaches for resolving this problem. The first approach
some degree of smoothness of the solution and improves condi- is to extend the deconvolution algorithm to a more general data
tioning of the problem. model. The model has to be so flexible that it can accommodate
• Total Least-Squares Formulation. The problem is formu- any variation of wellbore storage and skin that may take place in
lated as unconstrained nonlinear minimization with the objective real well tests. This would be very difficult to do. The second
function defined to include residuals of Eq. 3, possible errors in approach is to apply the current deconvolution algorithm based on
rate data, and regularization curvature constraints. Eq. 1—not to all the pressure data at once, but to the pressure data
There are several other details regarding specific implementa- from individual flow periods. When used in this mode, the data
tion of the algorithm.1,2 The implementation is geared toward the and the data model are always consistent, and deconvolution
test sequences that are structured similar to conventional well tests, should work well. Our testing confirms that this is the case.
with stepwise constant-rate functions q(t). The solution z() is
approximated as a piecewise-linear function on a grid of points i Information Content of Data and Deconvolution Problem.
with constant grid increment ⌬. The values zi of the solution z() Pressure/rate deconvolution is not just a smart data-processing
at the grid nodes i are determined in the minimization process. algorithm that could be applied routinely to any well-test data to
The authors use the variable projection algorithm9 for nonlinear recover the system-response function. Deconvolution has to be
least-squares minimization. tailored to the data that are being deconvolved.
Deconvolution is a form of inverse problem. There are two
von Schroeter Deconvolution and Real Test Data. Single-phase components to inverse problems: (1) test data and (2) a model for
flow of slightly compressible fluid in porous media is a linear representing the data. An inverse problem can be solved success-
problem. In many cases, solution of this problem can be presented fully only if the data model is consistent with test data. For ex-
in the form of Eq. 1. The p(t) and q(t) given in Eq. 1 are sandface ample, a model is inconsistent if it does not account for a physical
characteristics; they are the pressure and the rate on the reservoir phenomenon affecting the data (i.e., changing wellbore storage
side of the wellbore/reservoir interface. The pressure and rate data and/or skin). It is also possible that inconsistency is caused by the
usually acquired during well tests are not sandface characteristics; data itself, specifically by data errors. For example, having the
they are measured somewhere in the wellbore. Moreover, stepwise wrong rate on the right side of Eq. 1 may overconstrain the prob-
constant-rate test sequences are produced by operating chokes and lem so that it does not have a physically meaningful solution. In
valves normally located at the surface or in the wellbore some- this case, it is necessary to add additional degrees of freedom by
where above the reservoir interval. Therefore, the pressure and rate allowing the rate function on the right side of Eq. 1 to become a
data acquired during well tests have to be corrected for wellbore- model parameter so that it can be adjusted to restore consistency
storage and well-damage (skin) effects before they can be used in between the data and the model.1,2 However, one has to be careful
Eq. 1. With these corrections, Eq. 1 reduces to with this approach as well. Too many additional model parameters
t
may result in the problem that is underconstrained when it does not
dpu共t − 兲
兰q 共 兲
have a unique solution.
p共t兲 = p0 − sf d − qsf共t兲 ⌬ps. . . . . . . . . . . . . . . (4)
dt Implementation of the Deconvolution Algorithm
0
Here, ⌬ps is the pressure drop per unit rate at the wellbore/reservoir The main objective of this work is to evaluate performance and
interface (skin effect). The qsf (t) in Eq. 4 is the sandface well rate. The identify possible limitations of the von Schroeter deconvolution
sandface rate is related to the test-measured rate q(t) as algorithm. So, in our implementation of the algorithm, we incor-
porated all the novel ideas1,2 ascribed to the improved algorithm’s
24 dp共t兲 performance. We should note, however, that our implementation is
qsf 共t兲 = q共t兲 + Cw . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (5) not identical to the original von Schroeter algorithm. Hence,
B dt
strictly speaking, the algorithm performance discussed in this pa-
Here, Cw is the wellbore storage coefficient, and B is the fluid per is that of our variant of the algorithm and not of the original
formation volume factor. In principle, the wellbore storage coef- von Schroeter algorithm.
ficient and the pressure drop caused by skin may change with time, Our implementation is different on a number of points. Instead of
Cw(t), ⌬ps(t). the variable projection algorithm suggested by the authors,1 we use
There is one special case when Eq. 4 reduces to Eq. 1. This is the algorithm for unconstrained minimization by Dennis and Schna-
the case when the wellbore-storage coefficient, Cw, and the pres- bel.10 There also are differences in the definition of objective function
sure drop caused by skin, ⌬ps, are constant and do not change for least-squares minimization and in the minimization parameters.
during the entire well-test sequence. The constant-rate drawdown- We can look at deconvolution as a process of fitting the pressure and
pressure p0. Note that rate plays a dual role. It belongs to the model The integral in Eq. 7 can be computed analytically if q(t) is a
parameters; it is also part of the test data that must be fitted to the stepwise constant function of time and if z() is approximated as
model. As a model parameter, a rate value during the flow period may a linear function on each interval of grid j.1
be different from the corresponding rate measurement for the same The model prediction of a rate constraint is
flow period.
⍀qi共x兲 = xqi. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (8)
Model Parameters. Our algorithm implementation has built-in The model prediction of a curvature constraint is2
flexibility to process through deconvolution the pressure data for
any subset of flow periods of a well-test sequence. We also can 1
⍀ci共x兲 = 共x − 2xzi + xzi+1兲. . . . . . . . . . . . . . . . . . . . . . . . . . . (9)
select the rates during specific flow periods to be included in the ⌬ zi−1
list of model parameters. These are the rates that are adjusted
during least-squares minimization. We denote the vector of these Constraint Error Estimates. The objective function, Eq. 6, is a
rate components as xq⳱{xq1,xq2,…,xqL}. measure of closeness of the model prediction computed for a set of
The pressure data selected for processing define the time in- model parameters defined by vector x to the data that are fitted to
terval on which the result of deconvolution, the function z(), is the model. Eq. 6 combines terms of different physical natures:
determined. The upper limit of this time interval, T, is equal to the pressure, rate, and curvature constraints. These terms are expressed
elapsed time from the start of well rate history through the last in different units and may have numerical values differing by
point of pressure data selected for deconvolution. This then allows orders of magnitude. The parameters i in Eq. 6 are the scale
us to define the uniform grid {1,2,…,N} that is used for re- parameters that normalize individual terms in the sum so that they
construction of function z(). Here, N⳱ln T. The grid is com- would make comparable contributions to the objective function.
pletely defined by two parameters: the first grid node 1 and the Test data acquired during a well test are always measured with
total number of grid nodes N. Normally, we choose Nⱖ70. some errors. There is no point in fitting the data to a model exactly
We parameterize the function z() by introducing a set of if the data are known with some degree of uncertainty. Each of the
parameters xzi defined as the values of z() at the gridpoints i, parameters i in Eq. 6 characterizes the uncertainty (error band) of
xzi⳱z(i). For any other , the function z() is then approximated the corresponding constraint i. In this sense, i are not free pa-
by linear interpolation of grid-node values xzi. We can reconstruct rameters that one could manipulate at will to control least-squares
the constant-rate drawdown-response function pu(t) from the func- minimization. These parameters must be defined based on prior
tion z() by integrating Eq. 2. To do this integration, we need one knowledge of data quality.
more parameter: the value of pu(t) at the point t1⳱exp(1), where In our implementation of the deconvolution algorithm, we set
1 is the first node of the grid. We denote the pu(t1) as xz0. The pi for all pressure constraints equal to the resolution of the pres-
vector of parameters xz⳱{xz0,xz1,…,xzN} completely defines the sure gauge used in the test. For example, for quartz pressure
functions z() and pu(t) on the grid {1,2,…,N}. gauges, a good value for pi is 0.01 psi. The situation is not so
The complete list of model parameters is the vector straightforward with rate constraints. The rate during flow periods
x⳱{xz0,xz1,…,xzN, xq1,xq2,…,xqL, xp0}, which includes the re- when the well is shut in is equal to zero. Therefore, the rates for
sponse-function parameters, the rate parameters, and the initial- these flow periods are never included in the list of model param-
pressure parameter xp0⳱p0. eters and model constraints. For all other flow periods included in
the list of constraints, we set qi based on our estimate of possible
Model Constraints and Objective Function. Each data point rate errors for corresponding flow periods. For example, larger qi
that has to be fitted to the model defines one constraint. There are may be assigned to flow periods during well cleanup, when the
three types of constraints: pressure constraints, rate constraints, well production stream is not sent through the separator and the
and curvature constraints. We use the vector notation p⳱{p1, rate is not measured. For all curvature constraints, we set ci equal
p2,…,pK} for the pressure constraints. Here, pi are the values to one constant value. We have found that the value ci⳱0.05
of individual pressure points (measurements) selected for decon- works well for all data sets that we have used during the validation
volution, pi⳱p(ti). Similarly, rate constraints form the vector and testing of the deconvolution algorithm. This value has been
q⳱{q1, q2,…,qL}. The vector q has the same rate com- chosen to provide a small degree of regularization while not over-
ponents as the vector xq. Finally, the vector c⳱{c1,c2,…, constraining the problem and creating significant bias.
cN−1} represents curvature constraints. The number of curvature
constraints is N–1, where N is the number of nodes of the grid i. Initial Condition for Nonlinear Minimization. The objective
All the components ci in the vector c are set to zero to produce function defined by Eq. 6 is nonlinear with respect to model pa-
a smooth solution z(). The three vectors form a complete con- rameters x. In our implementation of the deconvolution algorithm,
straints list ⳱{ 1 , 2 ,…, M }⳱{ p , c , q }. The total this is a function of at least 70 variables. This function may have
number of constraints is M⳱K+L+N–1. many local minima, and the minimization process may converge to
We use the notation ⍀i(x) for the model prediction to the con- any one of them. Only convergence to the global minimum pro-
straint i. The model prediction ⍀i(x) of i depends on model duces the best data fit and successful pressure/rate deconvolution.
parameters defined by vector x. Also, let ⳱ {1, 2,…, M} be the The starting point for nonlinear minimization, the initial vector of
vector of error estimates for the corresponding data points repre- model parameters x, affects whether the minimization process con-
sented by the constraints vector . Using these notations, we de- verges to the global minimum. Selection of initial vector x is thus
fine the objective function for least-squares minimization as critically important for successful deconvolution.
兺冋 册
In our implementation of the deconvolution algorithm, we be-
i − ⍀i共x兲 2
M
1
⌽共x兲 = . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (6) gin minimization with the vector of model parameters defined as
2 i=1 i follows: (1) the parameter xz0 is initialized to unity; (2) the re-
maining response function parameters xzi for i=1,2,…,N are set
Model Predictions. At any time during a test, the model- to the same constant value, xzi⳱c1 (the expression for the con-
predicted pressure is given by the right side of Eq. 3. In our stant c1 is given in Ref. 1); (3) the rate parameters xqi are initialized
notations, ⍀pi(x) is the model pressure prediction at the time ti, to the corresponding values of rate constraints; and (4) the initial
p(ti). One can show that if the time corresponding to the first grid pressure parameter xp0 is estimated based on the record of test
configuration to accurately predict the pressure behavior over 590 of the results from several deconvolutions. The process is more
hours of data sequence. This is illustrated in Fig. 18, in which a model akin to test analysis than to automatic data processing.
that reproduces the PBU pressure in Fig. 17 does not honor the 3. We demonstrated performance and application of the deconvo-
pressure over the entire data sequence. lution algorithm to several real test examples. The new analysis
Deconvolution analysis in this case offers a way to determine technique allows us to develop additional insights into pressure-
characteristic system behavior over a much longer time interval. transient behavior and to extract more information from well-
Fig. 19 presents the constant-rate drawdown response derived by test data than is possible by using conventional well-test-
deconvolution of the last PBU pressure data. The response func- analysis techniques.
tion is defined on the time interval of 557 hours. The half-slope 4. Deconvolution unmasks the effects of rate variation and pre-
derivative trend in Fig. 17 that develops between 5 and 30 hours sents characteristic system pressure behavior on a longer time
levels off to a lesser slope at later time, indicating that the channel interval compared to conventional analysis techniques. In some
widens further away from the well. Fig. 19 presents a model match cases, this longer time span of characteristic pressure behavior
of the deconvolved constant-rate drawdown-response function. allows interpretation of the same test data in terms of a larger
The same model reproduces the pressure for the entire sequence; radius of investigation. Conversely, one can obtain the same
this is shown in Fig. 20. information as before with shorter test duration, less flaring, and
rig-time savings.
Conclusions 5. We have found deconvolution analysis to be extremely useful
1. The von Schroeter deconvolution algorithm works well on test for early detection of closed reservoir behavior. Pressure/rate
pressure and rate data that are consistent with the convolution deconvolution also opens new opportunities for analysis of per-
model defined by Eq. 1. It fails, however, when used with in- manent gauge pressure data. It is a valuable addition to the suite
consistent data. Changes of wellbore storage or skin during a of techniques used in well-test analysis.
well-test sequence are examples of possible inconsistencies.
These inconsistencies are often present in real test data. There- Nomenclature
fore, the von Schroeter deconvolution algorithm (in the original B ⳱ formation volume factor, RB/STB, RB/Mscf
form described by its authors) cannot be used reliably for analy- Cw ⳱ wellbore-storage coefficient, RB/psi
sis of real well tests. K ⳱ number of pressure constraints
2. This problem is resolved if a deconvolution algorithm based on Eq. L ⳱ number of data-model rate parameters
1 is used with the pressure data from an individual flow period. Our M ⳱ total number of data-model constraints
testing showed that when used in this mode, the algorithm per-
forms reliably with real test data. Deconvolution of test pressure
data one flow period at a time requires comparison and evaluation
Fig. 18—Test Example 3: Conventional analysis approach. The 8. Thompson, L.G. and Reynolds, A.C.: “Analysis of Variable-Rate Well-
model derived from analysis of the last PBU data in Fig. 17 fails
Test Pressure Data Using Duhamel’s Principle,” SPEFE (October
to honor the pressure data over the entire data sequence.
1986) 453.
9. Bjorck, A.: Numerical Methods for Least Square Problems, Soc. for
6. Fair, P.S. and Simmons, J.F.: “Novel Well Testing Applications of Industrial and Applied Mathematics (SIAM), Philadelphia, Pennsylva-
Laplace Transform Deconvolution,” paper SPE 24716 presented at the nia (1996).
1992 SPE Annual Technical Conference and Exhibition, Washington, 10. Dennis, J.E. and Schnabel, R.B.: Numerical Methods for Uncon-
DC, 4–7 October. strained Optimization and Nonlinear Equations, Prentice-Hall Inc.,
7. Thompson, L.G., Jones, J.R., and Reynolds, A.C.: “Analysis of Pres- Englewood Cliffs, New Jersey (1983).
sure Buildup Data Influenced by Wellbore Phase Redistribution,” 11. Press, W.H. et al.: Numerical Recipes: The Art of Scientific Computing,
SPEFE (October 1986) 435. Cambridge University Press, New York City (1992).