You are on page 1of 8

Use of Artificial

Intelligence in Well-Test
Interpretation
Olivier F. Allain, * SPE, and Roland N. Horne, SPE, Stanford U.

Summary. This paper describes Introduction in the case of a human expert is built almost
techniques for the identification of The use of computers for well-test analysis unconsciously. Techniques were developed
has played an important role in the devel- to replicate this viSUal-perception step.
well-test interpretation models from A major difficulty in the analysis of real
opment and application of new interpretation
pressure-derivative data. Artificial in- techniques. 1 With nonlinear regression, al- data, particularly when the pressure deriva-
telligence (AI) techniques are used to gorithms have been developed that can per- tive is used, is the separation of the true
form automated type-curve matching when reservoir response from signal or differen-
separate the derivative response
given a specific model and a first estima- tiation noise. Again, this is relatively simple
from signal and differentiation noise, for a human observer, but difficult to im-
tion of its parameters. 2-8 Such algorithms
and a rule-based recognition system not only speed the interpretation procedure plement in a computer program. This paper
characterizes the response using a but also increase the confidence in the re- describes an algorithm developed to over-
sults by allowing quantification of the qual- come this problem. The algorithm was able
symbolic representation. to distinguish response from noise correctly,
ity of the match obtained. Correct
application of these methods, however, de- thereby allowing for correct model identifi-
pends on the choice of the interpretation cation.
model. In a manner analogous to a human expert,
An interpretation procedure that uses the the technique constructs an interpretation
pressure derivative 9 - 11 is regarded as the model for the duration of the data by com-
most appropriate tool for the diagnosis of bining the features of the different flow peri-
an interpretation model for well-test data. ods of the response. The adequacy of a
The derivative is not only more sensitive to model is determined by qualitative and quan-
the different flow regimes and flow-regime titative information. Once a model is chosen
transitions but also is more generally ap- its parameters are estimated with a correla~
plicable and therefore provides a substitute tion or an appropriate table.
for all the specialized analyses. In traditional The system developed in this work can
graphical analysis, the choice of an interpre- perform model identification on real or
tation model is essentially a visual process. hypothetical data. Because the method also
Based on a log-log plot of the pressure provides parameter estimation, it can be
derivative, an expert can perform model used with an automated type-curve-matching
identification by recognizing features that analysis, allowing full automation of well-
are specific to particular analytical models. test interpretation.
The combination of the various models then The task of identifying an interpretation
gives an interpretation model. 12 model by use of the derivative of well-test
Although the human visual perception data can be separated into three components:
process is not fully understood, significant (1) observation-extraction of the features
work in AI has been conducted to develop present on the data derivative and represen-
machine vision systems. B - 15 These sys- tation of these features, (2) knowledge of the
models-methods for the construction and
tems operate not on the image itself but on
description of interpretation models, and (3)
a symbolic representation containing the in-
matching-criteria for choosing appropriate
formation needed for the task under consid-
interpretation models for given data.
eration.
This paper describes the development of The development of the specialized rule
techniques to automate the model- base and logical approach for these three
identification step of well-test interpretation components took considerable work before
by use of AI. The computer's choice of a the algorithms were able to interpret real
model is based on the pressure-derivative data reliably. The purpose of this paper is
curve and simulates the visual diagnosis per- to describe this development as an aid to
formed by a human expert. The reasoning others who may wish to attempt a similar
involved in such a diagnosis uses a symbolic approach.
representation of the derivative curve, which
Observation
'Now at Schlumberger Wireline & Testing. Gabon. To build a useful representation of the data
Copyright 1990 Society of Petroleum Engineers derivative, the first stage of the observation

342 March 1990 • JPT


"In a manner
analogous to a human
expert, the technique
constructs an
interpretation model
for the duration of the
data by combining the
features of the
different flow periods
of the response."

process is to distinguish between the true the leftmost and rightmost points is at least data derivative with a symbolic representa-
reservoir response and the noise present on 0.2. Compute the segment corresponding to tion in terms of the following shapes: up,
the curve. This distinction is almost uncon- So' and determine its quality, Qo. down, rruvcimum, minimum, plateau, and
scious for the human observer, but needs to 2. Constitute a new set SI by including valley.
be made explicit in a computerized proce- in So the first point on its left. Compute the A plateau is a flat segment preceded by
dure. This step corresponds to the produc- corresponding segment and determine its an up; a valley is a flat segment preceded
tion of a sketch, as described by Hom, 15 quality, Q\. by a down. A flat segment has a slope that
and consists of simplifying the original data, 3. Constitute a new set S2 by including is smaller than 0.1 in absolute value. This
keeping only the significant trends. in So the first point on its right. Compute convention is similar to that of Startzman
The algorithm developed here to produce the corresponding segment and determine its and Kuo,16 who looked at symbolic
the sketch is based on the method presented quality, Q2. representation of log data. Consecutive iden-
by Lowe. 16 4. Take the maximum among Qo, Q1, tical symbols (e.g., up, up) are not allowed
and Q2. If this maximum is Qo, then the in the representation; therefore, a single
Derivative Sketch. The algorithm described best segment is the one corresponding to symbol may represent several segments. Be-
is applied to a presentation of the pressure So. Otherwise, take SI or S2 to be the set cause choosing an interpretation model in-
derivative on log-log coordinates. The with the maximum value and Q 1 or Q2 to volves not only qualitative but also
derivative is obtained with the central be the corresponding qUality. This length- quantitative considerations, it is necessary
scheme described by Bourdet et al. 11 This ens the chosen set in one direction. Final- to record which segment(s), or which
scheme uses a point somewhere before (left) ly, go to Step 2 to seek a still longer point(s) of the sketch, correspond(s) to a par-
and one somewhere after (right) the point segment. ticular symbol. This information is used in
of interest, computes the two correspond- Rather than stopping the extension of a the matching process as a complement to the
ing slopes, and takes their weighted mean. segment as soon as the quality decreases, the symbolic representation.
The noise can be reduced by imposing a algorithm actually checks further, skipping The representation chosen may seem sim-
minimum distance, called the differential in- at most three points, to see whether the qual- plistic at first in that it does not generally
terval, between the point where the deriva- ity would still increase. This enables the allow us to distinguish changes in slope. Be-
tive is calculated and the left and right removal of sharp peaks in regions that are cause real systems seldom behave exactly
points. Thus, the three points-central, left, not very noisy, which can be created by iso- like analytical models, however, the data
and right-are not necessarily adjacent to lated erroneous data points. may include significant shapes that we do
each other. The proximity of the late-time The overall description of the algorithm not want to consider for model identifica-
data points on a log-log scale and the small producing the sketch can now be given as tion. If we chose a more faithful represen-
relative change between them may cause follows. tation, we would have to deal with all the
scatter on the derivative if the differentia- 1. i=1. anomalies from the beginning of the iden-
tion interval is taken to be zero (adjacent 2. Find the best segment at point of tification procedure. On the other hand, a
points). For this reason, and because of ex- Index i. simpler representation allows for a top-down
perience with the examples considered dur- If the last point in the corresponding set procedure, basing a first diagnosis on an
ing this study, a value of 0.2 log cycles was is the last data point, then go to Step 3. overall, not-too-detailed behavior and even-
used in all cases. Otherwise, take i to be the last point in the tually augmenting this diagnosis by con-
To produce the sketch, we approximate set. Go to Step 2. sidering left-behind features on the different
successive sections of the derivative by 3. Consecutive segments are intersected. sections.
straight-line segments. These segments are Because we start from the first point on From the original pressure-derivative
obtained by least-squares fit. The number the curve, segments are preferentially ex- curve for a given test, we extracted two
of segments required to represent the main tended to the right, except for the eventual different characterizations. One is purely
trends of the curve is based on a notion of last points. symbolic and gives a qualitative description
quality, explained below. If we consider a When large scatter is avoided by taking of the response in terms of shapes; the other
subset of points and the straight-line segment the derivative with a high enough differen- is in terms of segments and gives quantita-
obtained by least-squares fit to them, we can tiation interval, this algorithm successfully tive information about these shapes.
assign to it a quality as follows: removes the noise from the derivative, keep-
ing only the main trends present on the Knowledge of Models
Q=L/a . ....................... (1)
curve. These main trends correspond to the An interpretation model is usually obtained
With this measure of quality, we define true reservoir response and constitute the ba- by combining several components that
the notion of best segment at Point M as sis for model identification. produce the observed shapes on the differ-
follows. ent parts of the derivative. 12 For a com-
1. Constitute Set So with enough points Symbolic Representation. From the sketch puterized procedure to exhibit an identical
around Point M that the distance between obtained with this method, we substitute the behavior, it needs to be provided with some

JPT • March 1990 343


basic models (or parts of models) along with 2. Nondestructive combinations: describe the shapes present on this curve
a representation of their derivative and with quantitatively.
shape 1+shape 2-shape 1+ shape 2.
rules describing what combinations are pos-
sible and their effects. Matching
Combination Constraints. Not all arbitrary
combinations of the models listed earlier are A particular model (or combination of
Models. The flow regimes that can occur
physically possible, and we need to restrict models) qualitatively matches part of the
during a well test usually have been divided
the search for interpretation models by ex- data if the data representation on this part
into three main categories \8: early time
pressing certain constraints. In this work, is equal to a sub-sequence of the model
(representative of the well and its surround-
the following constraints were found nec- representation. The part of the data that is
ings), middle time (the average reservoir
response), and late time (boundary effects). essary. considered ~ two attributes: a list of shapes
In this study, we consider the most common 1. No heterogeneous or boundary effects and a list of segments. The corresponding
cases for these categories. can occur before an early-time regime. part of the model also has two attributes: a
1. Early time-well with wellbore storage 2. The wellbore-storage and vertically list of shapes and a list of regimes. The
and skin, vertically fractured well. fractured well models are mutually ex- equality of the two shape representations es-
2. Middle time-infinite-acting radial clusive. tablishes a correspondence between the seg-
flow, pseudosteady state, and transient dou- 3. Only one double-porosity transition can ments. on the data and the regimes of the
ble porosity. be present. model. This correspondence can be used to
3. Late time-single sealing fault, inter- 4. Nothing can occur after pseudosteady determine whether the model is quantitative-
secting sealing faults, constant-pressure state. ly acceptable. To do so, a new series of facts
fault, closed reservoir. 5. No more than two sealing-fault effects. describing certain regimes and regime com-
The characteristic derivative shapes as- This is not necessarily an exhaustive list; binations is included.
sociated with the models in those three other constraints may be required depend-
categories and.their combinations have been ing on the implementation of this approach. Quantitative Constraints. Earlier, we em-
presented by several authors, 10,11,\9-2\ and phasized that a single symbol in the data
we can represent them symbolically with the Regime Description. Much more is in- representation may correspond to several
vocabulary introduced earlier. (For the first volved in the description of analytical segments in the sketch of the derivative.
category, instead of an early-time effect models than is given in the previous section. Therefore, when a regime corresponds to
alone, we append the infinite-acting radial The different shapes that the derivative of a particular symbol in the matching proc-
flow regime, which therefore will not occur these models exhibits correspond to partic~ ess, it may actually correspond to several
in the middle-time effects.) ular regimes that have quantitative charac- segments. Keeping this in mind, we can now
1. Wellbore storage and skin, infinite teristics; e.g., fixed slope values. To account express the quantitative characteristics that
homogeneous medium (up, 1TU1Ximum, for this information, we first associate a re- must be observed. The thresholds added
down, valley) or (down, valley) or (up, gime description with the representation of around the theoretical values are based on
plateau). each model. Both lists have the same num- the experience with the examples considered
2. Vertically fractured well, infinite ber of elements and correspond to each other in this study. They may require further
homogeneous medium (up, plateau). pairwise. For instance, to the representation tuning.
3. Pseudosteady-state double-porosity of wellbore storage and skin in an infinite 1. Among the segments corresponding to
transition (down, minimum, up, plateau). wellbore storage, the one with slope closest
homogeneous medium (with a high enough
4. Transient double-porosity transition
value of CD e2s ) we could associate the list to one should have a slope at least equal to
(up, plateau). 0.7.
(wellbore storage, maximum of the hump
5. Sealing fault (up, plateau). 2. For wellbore storage and skin with low
after wellbore storage, transition after well-
6. Constant-pressure fault (down ).
bore storage, infinite-acting radial flow) to value of CD e 2s (no hump on the deriva-
7. Pseudosteady state (up).
the list (up, 1TU1Ximum, down, valley). tive), all the segments corresponding to the
It is necessary to extend the representa-
When an interpretation model is built by initial up should have a slope smaller than
tions containing extrema because the sketch
combining the representations of several 0.7.
may flatten the corresponding shape.
models, their regime descnptions must also 3. For a vertically fractured well, all the
be combined to characterize the global segments corresponding to the initial up
Influence of Combinations. Because an in-
response. This is done by adding rules for should have a slope smaller than 0.7.
terpretation model is defined as a combina-
tion of the basic models listed earlier, we shapes and consequences for regimes to the 4. In case of transient double porosity, the
can construct its representation by combin- combination. The two main consequences dimensionless value of the derivative at the
. ing the representations of its components. with the corresponding rules are the fol- bottom of the heterogeneous transition
To do so, we must define the influence of lowing: should not be lower than 0.2.
combining a shape with another. 5. The upward trend caused by a no-flow
valley +up- minimum +up effect should not contain segments with
Given Models M \ and M2 and the re-
spective representations, (Sl, Si . . . S\) and and r\ +r2-gt(r\)+r2, slopes greater than 0.7.
(Sr, Si· ..Sp, the representation f~r the where gt(r\) means that the level corre-
6. The segments corresponding to pseu-
dosteady state for a closed reservoir should
model obtained by combining M2 with M\ sponding to the minimum is higher than the
on the shape Sf is given as level that would have corresponded to the
increase in slope to a value greater than 0.7.
7. There must be consistency between the
[Sl·· .Sl_\, comb(Sl,Sf)' Sr· .S~], valley.
different stabilization levels observed on the
................. (2) plateau +down -1TU1Ximum +down curve-radial flow, less than radial flow,
greater than radial flow, stabilization caused
where comb(S} ,Sf)represents the shape(s) and r\ +r2-lt(r\)+r2, by a no-flow boundary, etc.
obtained by combining Sf with Sf. The where It(r\) means that the level corre-
different possibilities for those combinations sponding to the 1TU1Ximum is lower than the Matching Procedure. We consider for the
are expressed by the following rules. level that would have corresponded to the moment that the data representation includes
1. Destructive combinations:
plateau. An interpretation model is charac- only features representative of the true reser-
valley +down - nothing terized by two different attributes. The first voir response that can be matched with the
one is in terms of shapes and qualitatively analytical models available to the program
valley +up - minimum + up
describes the derivative curve; the other is or their combinations. The aim is to find all
plateau + down -1TU1Ximum + down. in terms of regimes and can be used to interpretation models that qualitatively and

344 March 1990 • JPT


" .. . the first stage of the observation process
Is to distinguish between the true reservoir
response ~nd the noise present on the curve.
This distinction Is almost unconscious for the
human observer, but needs to be made explicit
In a computerized procedure."

TABLE 1-FLOW PERIOD 4 (BUILDUP)

Elapsed Time Pressure Elapsed Time Pressure Elapsed Time Pressure


(hours) (psi) (hours) (psi) (hours) (psi)
0.0 3,816.99 6.39802 x 10- 2 3,867.10 8.70821 x 10- 1 3,892.10
6.2213 x 10- 4 3,817.96 6.64622 x 10- 2 3,867.74 9.08096 x 10- 1 3,892.43
1.24723 x 10- 3 3,819.73 6.95703 x 10- 2 3,868.41 9.45360 x 10- 1 3,892.60
1.86215 x 10- 3 3,821.64 7.26673 x 10- 2 3,869.12 9.82624 x 10- 1 3,892.91
2.48825 x 10- 3 3,823.40 7.57754 x 10- 2 3,869.91 1.01990 3,892.97
3.10317x10- 3 3,824.92 7.88836 x 10- 2 3,870.32 1.05716 3,892.77
3.72927 x 10- 3 3,826.96 8.19917 x 10- 2 3,871.00 1.11307 3,893.91
4.34419 x 10- 3 3,828.28 8.50998 x 10- 2 3,871.52 1.16897 3,893.57
4.97029 x 10- 3 3,829.52 8.88229 x 10- 2 3,872.07 1.22487 3,893.80
5.58521 x 10- 3 3,830.67 9.13049 x 10- 2 3,872.48 1.28078 3,893.64
6.21130 x 10- 3 3,831.71 9.44131 x 10- 2 3,873.00 1.33668 3,893.97
6.83740 x 10- 3 3,832.61 9.75212 x 10- 2 3,873.54 1.39258 3,894.15
7.45232 x 10- 3 3,833.39 1.00618x10- 1 3,873.84 1.48574 3,894.24
8.07842 x 10- 3 3,834.31 1.03726 x 10- 1 3,874.33 1.57892 3,894.64
8.69334 x 10- 3 3,835.18 1.06834x10- 1 3,874.73 1.67209 3,894.76
9.31944 x 10- 3 3,835.92 1.11798x10- 1 3,875.47 1.76525 3,895.13
9.93436 x 10- 3 3,836.81 1.16148x10- 1 3,875.88 1.85843 3,895.20
1.05605 x 10- 2 3,837.58 1.20810x10- 1 3,876.51 1.95160 3,895.38
1.11754x10- 2 3,838.36 1.25472 x 10- 1 3,876.83 2.13794 3,895.65
1.18015x10- 2 3,839.11 1.30437 x 10- 1 3,877.44 2.32427 3,895.92
1.24276x10- 2 3,839.94 1.34786x10- 1 3,877.89 2.51061 3,896.22
1.30425x10- 2 3,840.68 1.39750x 10- 1 3,878.25 2.69696 3,896.54
1.36686 x 10- 2 3,841.53 1.44099x10- 1 3,878.64 2.88329 3,896.62
1.42836x10- 2 3,842.07 1.53423 x 10- 1 3,879.37 3.06963 3,896.81
1.49097 x 10- 2 3,842.79 1.62737 x 10 -1 3,880.07 3.34914 3,897.18
1.55246x10- 2 3,843.41 1.73291x10- 1 3,880.86 3.62865 3,897.36
1.61507x10- 2 3,843.97 1.81374x10- 1 3,881.45 3.90815 3,897.36
1.67656 x 10- 2 3,844.56 1.90687 x 10- 1 3,881.92 4.18766 3,897.78
1.73917 x 10- 2 3,845.11 2.00001 x 10- 1 3,882.35 4.46717 3,898.08
1 .80178 x 10 - 2 3,845.72 2.0fj325 x 10- 1 3,882.77 4.74668 3,898.22
1.86327 x 10- 2 3,846.35 2.113638 x 10- 1 3,883.19 5.02619 3,898.45
1.92588 x 10- 2 3,846.78 2.27951 x 10- 1 3,883.63 5.30570 3,898.79
1.98737 x 10- 2 3,847.49 2.37276 x 10- 1 3,884.03 5.58521 3,898.80
2.11147x10- 2 3,848.40 2.46589 x 10- 1 3,884.39 5.86471 3,899.11
2.17408x10- 2 3,848.91 2.55902 x 10 -1 3,884.70 6.14422 3,899.23
2.23558 K 10- 2 3,849.32 2.65227 x 10- 1 3,884.99 6.42373 3,899.48
2.29819 x 10- 2 3,849.99 2.74540 x 10- 1 3,885.34 6.70324 3,899.62
2.42229 x 10- 2 3,850.73 2.93178x10- 1 3,885.90 6.98275 3,899.81
2.54639 x 10- 2 3,851.61 3.11804x10- 1 3,886.12 7.26226 3,900.03
2.67049 x 10- 2 3,852.36 3.30442 x 10 -1 3,886.69 7.54177 3,900.16
2.79459 x 10- 2 3,853.16 3.49079 x 10- 1 3,887.09 7.91441 3,900.37
2.91981 x10- 2 3,853.98 3.58393 x 10- 1 3,887.39 8.28716 3,900.69
3.04391 x10- 2 3,854.57 3.67706 x 10- 1 3,887.52 8.65980 3,900.69
3.16802 x 10- 2 3,855.39 3.86343 x 10- 1 3,887.94 9.03244 3,900.86
3.29212 x 10- 2 3,855.96 4.04981 x 10- 1 3,888.34 9.40519 3,901.08
3.41622 x 10- 2 3,856.67 4.23607 x 10 -1 3,888.58 9.77783 3,901.38
3.54032 x 10- 2 3,857.14 4.42245 x 10- 1 3,888.93 10.1505 3,901.43
3.66442 x 10- 2 3,857.87 4.60883 x 10- 1 3,889.20 10.5232 3,901.75
3.78853 x 10- 2 3,858.42 4.79509 x 10- 1 3,889.41 10.8959 3,901.79
3.91263 x 10- 2 3,858.89 4.98147x10- 1 3,889.31 11.4549 3,902.17
4.03785 x 10- 2 3,859.48 5.16784 x 10- 1 3,889.46 12.0139 3,902.31
4.16195 x 10- 2 3,859.99 5.35411x10- 1 3,889.72 12.5730 3,902.58
4.34754 x 10- 2 3,860.74 5.54048 x 10- 1 3,889.91 13.1320 3,902.77
4.5342(5 ~ 10- 2 3,861.48 5.72686 x 10- 1 3,890.20 13.6910 3,903.09
4.72097 x 10- 2 3,862.03 6.09950 x 10- 1 3,890.69 14.2500 3,903.19
4.90656 x 10- 2 3,862.87 6.47214 x 10- 1 3,891.14 14.8090 3,903.45
5.09327 x 10- 2 3,863.31 6.84489 x 10- 1 3,890.80 15.3680 3,903.45
5.34147 x 10- 2 3,864.30 7.21754 x 10- 1 3,891.38 15.9271 3,903.71
5.58968 x 10- 2 3,864.99 7.59018x10- 1 3,891.75 16.4861 3,903.78
5.90049 x 10- 2 3,865.96 7.96293 x 10- 1 3,892.04 17.0451 3,904.16
6.14870 x 10- 2 3,866.47 8.33557 x 10- 1 3,892.31 17.6059 3,904.22

JPT • March 1990 345


2
10
''.r-------------, + Original derivative
u
>
."
os
-0- Sketch

.~
-8 1
10
~
"
, £
1~10.j...
.• --,,~
.• --,,~
•• --,,~.,--,,~,-~-___l". 0
10
4
16 '6 '
Elapsed time (Ius)

Fig. i-Example data: Flow Period 4 Fig. 2-Example derivative and sketch (differentiation interval =0.2).
(buildup).
quantitatively match the data. This is done In a general case, the identification pro-
recursively, starting from the first shape in gram should not blindly trust the sketch of
the data representation, as follows. the derivative. Rather, it should be able to
I. Find a model allowed at this point such recognize anomalies and to modify its
that its first shape is the same as the one ob- representation of the response accordingly.
served and any quantitative checks applica- To allow for such behavior. the original
ble for the regime corresponding to this representation of the sketch needs to be
shape are satisfied. Go to Step 2. searched for the following: (I) hump (up,
2. Stop if there is no shape left on the maximum, down), (2) trough (down, mini-
data. Otherwise, go to Step 3. mum, up), and (3) late-time trend-up,
3. Go to next shape. If there is no shape down, valley, or plateau-when the last seg-
left in the model, find a new model allowed ment does not belong to a hump or a trough.
TABLE 2-FLOW PERIODS at this point such that its first shape is the Instances of these categories that may
same as the one observed, and any quantita- need to be removed are characterized for
Duration Flow Rate tive checks applicable for the regime cor- humps and troughs, on the basis of width,
Test (hours) (STB/D) responding to this shape are satisfied. depth, and sharpness. The late-time trend
1 3.8 800 Append it to the interpretation model built is characterized on the basis of length and
2 3.3 2.500 so far. Go to Step 2. If the next shape of slope.
3 16.45 830 the model matches the one observed and ap- The matching procedure must then be
4 18 o plicable quantitative checks are satisfied, modified according to the following rule,
then go to Step 2. Otherwise. find a model which is applied each time a new shape in
allowed at this point that, when combined the data representation must be matched. If
with the current shape, produces the one ob- the interpretation model built so far must be
TABLE 3-RESERVOIR AND served and for which applicable quantitative augmented to match the shape; the shape be-
FLUID PROPERTIES longs to an instance of one of the above
checks are satisfied. Take the combination
as new interpretation model. Go to Step 2. categories; the characteristics of this instance
B 1.5 suggest removal; and removing it would al-
Ct. psi- 1 2 x 10- 5 With a backtracking procedure that causes
the system to re-evaluate any assumption low continuation with the same model, then
h. ft 7
remove it from the representation, correct
'w. ft 0.29
0.3
made in the satisfaction of the goal, all the
'possible models will be found. the representation, and continue.
lL.cP
<P 0.05
Nonideality. Although the algorithm Parameter Estimation
producing the sketch of the derivative has To achieve a full automation of the interpre-
been shown to be sufficiently powerful in tation procedure once a model is proposed,
TABLE 4-PARAMETER ESTIMATES most of the cases considered during this it is necessary to find first estimates for its
study (provided that the derivative is taken parameters. These estimates constitute the
(Poltop) match • psi- 1 0.075 with a safe differentiation interval), it would starting point for an automated type-curve-
k. md 566 be too strong an assumption to consider that matching analysis. They are not intended to
[(toICD)/Atlmatch' hours- 1 140 the sketch contains only ideal features. For be the final result of the interpretation.
(C De 2s ), 83.6 instance, anomalous behaviors can occur The correspondence between the regimes
C. bbl/psi 0.028 during buildups, such as the effects of phase in the model description and the segments
s -3.11 redistribution in the wellbore,22 that should 'in the sketch of the derivative is used to es-
w 0.217 not be considered for model identification. timate the parameters. This is done with
Xe -2s 3.9x10- 5 very little computation with either correla-
2s
Technical or human errors, like an errone-
(C oe )'+m 18.2 ous estimation of the flowing pressure, may tions or tables built for this purpose from
also result in anomalies. 23 type curves. All these relations are based on
Some of these effects may not affect the drawdown responses but can be used for
analysis if the representation chosen is very buildups as well because only a first esti-
simple. In a buildup case with wellbore mate of the value is required.
storage, if the flowing pressure is too low,
the derivative will exhibit an inflexion be- Example
fore reaching the expected 45 0 line. The A program based on the methods developed
succession of these two trends will be con- in this study was written in PROLOG (ex-
sidered as a single up for the identification cept for the algorithm producing the sketch,
and will be diagnosed as wellbore storage which was written in FORTRAN). The use
because one of the trends has a unit slope. of this program is illustrated on a buildup
346 March 1990 • JPT
10
2

~ Sketch of the derivative


.
" ..-------------,
,
.~ _ Representation ~ 10

>
:g f ,
1
1"
~
J:'"
10

"
. + Flow pmad ...
_ Finl:EIti!mte
au

10
0 "
.,
"
. "
,

16 4 16 1
Elapsed time (hrs)
Fig. 4-Slmulated test with first estimates
for pseudosteady-state double-porosity
Fig. 3-Example sketch and simplified representation (differentiation Interval 0.2). = model.

test (Table 1) in which the well flowed at skin iii a reservoir with pseudosteady-state,
three different rates before shut-in. 10 The double-porosity behavior, (2) well with
pressure change is plotted vs. shut-in time wellbore storage and skin in a reservoir with
on log-log scale in Fig. 1. Table 2 lists the transient double-porosity behavior, and (3)
flow periods used for the analysis. The well with wellbore storage and skin in a
derivative was computed with respect to the homogeneous reservoir with a sealing fault.
superposition time function with two differ- Parameter EstimoJion. We illustrate how "Automated
ent differentiation intervals: 0.2 and O. estimates are obtained in the case of the first Identification.
interpretation model. Table 3 lists the reser-
Model identification was performed on both
voir and fluid properties.
provides a
curves.
Pressure match-obtained assuming that comprehensive list of
the level of the infinite-acting radial-flow all Interpretations that
Differentiation Interval =0.2. The deriva-
period in the data is the average between the
tive obtained in this case as well as the sketch are consistent with
highest and lowest derivative value. This
produced are shown in Fig. 2.
Models. A symbolic representation of the
choice is justified because radial flow is not the data, which Is
present in the response and the model con-
derivative is produced from the sketch: (up, tains regimes that produce shapes below and something even an
plateau, down, minimum, up). A symbol in above the 0.5 line. expert would rarely
this representation may correspond to sever- CDe 2s -obtained by table interpolation
al segments on the sketch. A graphical from the height of the maximum on the
attempt to do."
equivalent is given in Fig. 3 and compared early-time hump above the radial-flow level.
to the sketch. Time match-obtained by table interpola-
In this example, three different interpre- tion from the value of CD e 2s and the time
tation models were proposed by the pro- at which the maximum of the early-time
gram: (1) well with wellbore storage and hump occurs.

2,-------------------------------------------------------------------~
10

+ Original derivative
-0- Sketch

1
10

+ +
o
10 +
+ +
+
+

-1~---------,----------_r--------~r_--------_r----------r_--------~
10 -4 -3 -2 -1 o 2
10 10 10 10 10 10 10
Elapsed time (hrs)

Fig. 5-Example derivative and sketch (differentiation Interval 0). =


JPT • March 1990 347
10
2~----------------------------------------------------------------~

-0- Sketch of the derivative


___ Representation

0
10 _4 -3 -2 -1 o 1 2
10 10 10 10 10 10 10
Elapsed time (hrs)

Fig. 6-Example sketch and simplified representation (differentiation Interval = 0).

w and A-obtained with the relations Conclusions those languages or by using an expert-
presented by Bourdet et al. 10 With an AI approach, it has been possible system shell.
Table 4 gives the values obtained. The to automate the identification step in the in- Parameter estimation is obtained with very
corresponding multirate type curve, taking terpretation of well-test data. This paper de- little computation and can be used to pro-
into account the flow-rate history (Table 2), scribed the logical steps required to separate vide initial estimates for an automated type-
was generated and differentiated with the reservoir response from the signal noise curve-matching analysis.
respect to the superposition time function. and to recognize characteristics of the reser- Automated identification produces allthe
Both type curve and original data are shown voir model from the extracted symbolic interpretation models that, according to the
in Fig. 4. representation. Definition of the rule base knowledge of the program, match the data.
requires considerable care and experimen- The measure of quality given by an auto-
Differentiation Interval=O. In this case, tation. We hope that the experience summar- mated type-curve-matching analysis for the
the derivative was obtained by use of the two ized in the paper can reduce the time different models can constitute the basis for
pressure data points immediately adjacent required for others to apply this kind of ap- automatically deciding which is most appro-
to the point at which the pressure derivative proach. priate.
was required. Using central differences in The identification process was based on Automated identification is important for
this way is known to produce extremely the pressure-derivative curve and separated two major reasons. First, it provides con-
noisy derivative values (as seen in Fig. 5) into three different components. sistency and reduces misinterpretations
and is not recommended-the approach was I. Observation-extraction of the infor- caused by associating flow models with
used in this case as a test of the AI al- mation from the data derivative by produc- inappropriate parts of the data. Second, it
gorithm's ability to handle very noisy data .. tion of the sketch of the derivative and provides a comprehensive list of all interpre-
The sketch obtained from these noisy deriva- symbolic representation of this sketch. tations that are consistent with the data,
tive data is shown on the same plot (Fig. 5) 2. Knowledge of the models-construc- which is something even an expert would
and is seen still to contain anomalies. The tion and description of interpretation rarely attempt to do.
initial representation of the sketch is quite . models, including basic models, influence
complex: (up, plateau, down, minimum, up, of combinations, combination constraints, Nomenclature
maximum, down, minimum, up, plateau, and regime description. B = FVF, RB/STB [res m 3 /stock-
down, minimum, up, maximum, down, mini- 3. Matching-identification of appropri- tank m 3 ]
mum, up, maximum, down, valley, up, max- ate interpretation models through qUalitative C r = total compressibility, psi- 1
imum, down, minimum, up, maximum, match and quantitative constraints. [kPa -I]
down). With the rules described earlier, A program based on the methods de- C = wellbore storage, bbl/psi
however, the program recognizes the scribed in this paper was written in the AI [m 3 /kPa]
anomalies and, by successively removing language PROLOG. The program has been CD = dimensionless storage constant
them, constructs the final representation of able to perform model identification on real h = formation thickness, ft [m]
the curve: (up, plateau, down, valley, down, data even when they were very noisy. The i = index
valley, up). The graphical equivalent of this use of PROLOG was important to this work k = permeability, md
representation is shown in Fig. 6 along with in that this language performs "backtrack- L = distance between the leftmost
the original sketch. ing," which means that all possible solutions and rightmost points
Only one model is identified in this case, are determined. This was one of the major
p = pressure, psi [kPa]
a well with wellbore storage and skin in a goals of the work: to provide a mechanism
p D = dimensionless pressure
reservoir with pseudosteady-state, double- by which no feasible interpretation fails to
ilp = pressure change, psi [kPa]
porosity behavior. The two other models ob- be included (as might occur in a manual
tained in the previous case are not proposed analysis). This backtracking feature is the
Q = quality
here. This is because the representation ex- principal reason for the choice of PROLOG r = regime
hibits a stabilization after the early-time over such languages as LISP, C, or FOR- r w = wellbore radius, ft [m]
hump (Fig. 6). The stabilization found TRAN, although it is conceivable, as with s = van Everdingen-Hurst skin
occurs in a very noisy region and may there- any AI application, that the method could factor
fore be erroneous. be implemented with greater difficulty in Si = shape

348 March 1990 • JPT


!M = elapsed time, hours 13. Binford, T.O.: "Visual Perception by Com-
puter," Proc., IEEE Systems Science and Authors
A = interporosity pseudosteady-
state-flow parameter Cybernetics Conference, Miami (Dec. 1971).
p. = viscosity, cp [mPa·s] 14. Brady, J.M.: Computer Vision, North-
Holland Publishing Co., Amsterdam (1981).
(J = standard deviation of error of
15. Horn, B.P.K.: Robot Vision, McGraw-Hili
fit Book Co., New York City (1986).
cp = porosity 16. Lowe, D.G.: Perceptual Organization and
w = storativity ratio Visual Recognition, Kluwer Academic Pub-
lishers, Boston, MA (1985).
Acknowledgments 17. Startzman, R.A. and Kuo, T.B.: "A Rule-
This work has been supported by the Based System for Well Log Correlation,"
SPEFE (Sept. 1987) 311-19.
SUPRI-D research consortium in Innova-
18. Gringarten, A.C.: "Interpretation of Tests in Allain Horne
tions in Well Test Analysis and by Schlum- Fissured Reservoirs and Multilayered Reser-
berger. voirs With Double Porosity Behavior: Theory Olivier F. Allain is a field engineer with
and Practice," JPT (April 1984) 549-64. Schlumberger Wlrellne & Testing In
References 19. Clark, D.G. and Van Golf-Racht, T.D.: Gabon. He holds an MS degree from
1. Ramey, H.J. Jr.: "Pressure Transient Test- "Pressure-Derivative Approach to Transient Ecole Natl. des Ponts et Chaussees In
ing," JPT(July 1982) 1407-13. Test Analysis: A High-Permeability North Paris and MS and Engineer degrees In
2. Earlougher, R.C. Jr.: Advances in Well Test Sea Reservoir Example," JPT (Nov. 1985) petroleum engineering from Stanford U.
Analysis, Monograph Series, SPE, Richard- 2023-39. Roland N. Home Is an associate profes-
son, TX (1977) 5, Chap. 12. 20. Houze, D.P., Horne, R.N., and Ramey, H.J. sor of petroleum engineering at Stan-
3. Padrnanabhan, L. and Woo, P.T.: "A New Jr.: "Pressure-Transient Response of an ford U. He holds BE and PhD degrees
Approach to Parameter Estimation in Well Infinite-Conductivity Vertical Fracture in a In theoretical and applied mechanics
Testing," paper SPE 5741 presented at the Reservoir with Double-Porosity Behavior," and a DSc degree in engineering, a/l
1976 SPE Symposium on Reservoir Simula- SPEFE (Sept. 1988) 510-18. from the U. of Auckland. The Stanford
tion, Los Angeles, Feb. 19-20. 21. Proano, E.A. and Lilley, 1.1.: "Derivative U. SPE Student Chapter faculty sponsor
4. Tsang, C.F. etal.: "Variable Flow Well Test since 1981, Home received the 1982
of Pressure: Application to Bounded Reser-
Analysis by a Computer Assisted Matching Distinguished Achievement Award for
voir Interpretation," paper SPE 15861 Petroleum Engineering Faculty. He
Procedure," paper SPE 6547 presented at the presented at the 1986 SPE European Confer-
1977 SPE California Regional Meeting, chaired the 1986-87 Microcomputer Ap-
ence, London, Oct. 20-22. plications Committee and served on the
Bakersfield, April 13-15. 22. Stegemeier, G.L. and Matthews, C.S.:
5. Padrnanabhan, L.: "Welltest-A Program for 1983-87 Editorial Review Committee
"Study of Anomalous Pressure Build-Up Be- and on a Technical Committee for the
Computer-Aided Analysis of Pressure Tran- haviors," JPT (Feb. 1958) 44-50; Trans.,
sients Data from Well Tests," paper SPE 1989 Annual Meeting. Horne is a mem-
AIME, .213. ber of the Textbook Committee and a
8931 presented at the 1979 SPE Annual Tech- 23. Ramey, H.J. Jr.: "Practical Use of Modern
nical Conference and Exhibition, Las Vegas, Technical Committee for the 1990 An-
Well Test Analysis," paper SPE 5878 nual Meeting.
Sept. 23-26.
presented at the 1976 SPE California Regional
6. McEdwards, D.G.: "Multiwell Variable-
1yieeting, Long Beach, April 8-9.
Rate Well Test Analysis," SPEI (Aug. 1981)
441-46.
7. Rosa, A.J. and Horne, R.N.: "Automated 51 Metric Conversion Factors
Type-Curve Matching in Well Test Analy- bbl x 1.589 873 E-OI = m3
sis by Using Laplace Space Determination of cp x 1.0* E-03 = Pa·s
Parameter Gradients," paper SPE 12131 ft x 3.048* E-Ol = m
presented at the 1983 SPE Annual Technical md x 9.869233 E-04 = I'm2
Conference and Exhibition, San Francisco, psi x 6.894 757 E+OO = kPa
Oct. 5-8. psi -I x 1.450 377 E-OI = kPa- 1
8. Barna, J. et al.: "Improved Estimation Al- • Conversion factor is exact.
gorithms for Automated Type-Curve Analy-
sis of Well Tests," SPEFE (March 1988)
186-96; Trans., AIME, 285. Provenance
9. Bourdet, D. et al.: "A New Set of Type- Original SPE manuscript, Use of Artificial
Curves Simplifies Well Test Analysis," Intelligence for Model Identification and
World Oil (May 1983) 95-106.
Parameter Estimation in Well Test In-
10. Bourdet, D. et al.: "Interpreting Well Tests
in Fractured Reservoirs," World Oil (Oct. terpretation, received for review Oct. 2,
1983) 77-87. 1988. Paper accepted for publication Dec.
11. Bourdet, D., Ayoub, J.A., and Pirard, Y.M.: 11, 1989. Revised manuscript received Sept.
"Use of Pressure Derivative in Well-Test In- 11, 1989. Paper (SPE 18160) first presented
terpretation," SPEFE (June 1989) 293-302. at the 1988 SPE Annual Technical Confer-
12. Gringarten, A.C.: "Type Curve Analysis: ence and Exhibition in Houston, Oct. 2-5.
What It Can and Cannot Do," JPT (Jan.
1987) 11-13. JPT

JPT • March 1990 349

You might also like