Académique Documents
Professionnel Documents
Culture Documents
SPE 51086
...................................... (1)
particular model). where,
X=(%
T”7)-L3
‘4PI”’)
........................................... (2)
267
2 W. D. MCCAIN JR., R. B. SOTO, P. P. VALKO, AND T. A. BLASINGAME SPE 51086
This model was fitted to the data set with non-linear “smoothness” for the individual transformations. Although this
regression, the following result was obtained: operation is somewhat hidden in the algorithms (where a “data
5.354891
smoother” code is used), we believe that the ACE algorithm (as
&161488 ~@ . ().74()152 . well as the slightly different GRACE algorithm) does in fact
Pb = 1091.47 [R$O*1465 i
provide a unique and robust minimization of the expected error-
.........................................................................................(3)
-in a non-parametric sense.
where,
The procedure for this approach is given by
.X= (0.013098 T02!23.72)– $.2 x 10-6 AFY2”176124).(4) 1. Calculate the data transforms:
Z1 = fl(xl), 22 = fz(x~, .... Zn = fn(x~ and zo = fo(y).
2. Calculate the transform sum z~ = z] + Z2 + . . . + Zn.
3. Calculate the inverse transform: y = f~l (zo).
The Velarde relation (Eqs. 3 and 4) reproduced the
bubblepoint pressures from the data set used in its creation to an While the ACE and GRACE algorithms do provide a non-
average error of 0.6% and an average absolute error of 11.50A. parametric optimization of the dependent and independent
Fig. 1 is a comparison of bubbiepoint pressures predicted by the variables, this approach does not provide a computation (i.e.,
Velarde relation with the experimental data. predictive) model. Figs. 2 through 5 show the optimal
transformations of the individual data for the four independent
variables. The optimal transform of the data for the dependent
variable, bubblepoint pressure, is given in Fig. 6. This “non-
Non-Parametric Regression: (GRACE algorithm4’5) parametric” regression results in a reasonable match of
Parametric regression--that is, regression where a prescribed calculated and measured bubblepoint pressures. Fig, 7 shows
model is fitted to data, is a robust and effective mechanism for this comparison of calculated and measured bubblepoint
representing a data function. However, it provides little insight pressures for the non-parametric optimization.
into the interrelation of the independent variables, nor does it The ACE and GRACE algorithms do not provide a
provide a “global” minimum expected error of the dependent predictive model, i.e., they do not result in equations. However
and independent variables. simple quadratic polynomials can be fitted to the optimal data
The non-parametric regression approach proposed by tmnsforms, resulting in:
Breiman and Friedman,4 and refined by Xue, et al,5 provides Z~ =1.2206 – 1.8029 X 10-2 (MY)-3.8682 X 10q (API)2 ......
exactly such a “non-biased” mechanism for the purpose of ......................................................................................... (7)
establishing the minimum error relationship between the
dependent - and independent variables. -The method of
22 =3.8649- 6.5080 (yg )+ 1.9670x (yg~ .................. (8)
Alternating Conditional Expectations (ACE)4 is based on the
concept of developing an optimal transformation of each
variable--both the dependent variable, as well as the Z3 = -7.7161 + 2.3604 ~n (T,,,)]- 1.6678x 10-1 [h (T,e,)~ ..
independent variable(s). ......................................................................................... (9)
The ACE approach finds individual transformations of the
independent variables (xIJ2,.. .,xn) which are of the form: 24 = –1.0056x l(!l + 1.6664 [in (R.b )]
.......................(10)
-zI =fl (q] z?- = fz (X2), . .Zn = fn (an) ...................... -1.7682 X 10-2 [In (R~b)]2
In addition, the ACE approach also finds a transformation The inverse transform is defined as:
for the dependent variable, y, of the form ZO=ZI+Z2+Z3 +Z4 ...................................................(n)
20 =fo (y) .....................................................................(6) And the inverse transform, in terms of In@b), is given by
The development of these data transforms uses an intuitively
in (pb )= 7.8421 + 4.9506 x 10-1 [h (20 )]
straightforward requirement--maximize the correlation of the ..................(12)
transformed dependent variabie zo with the transformed
-2.5726 x 10-3 ~n (z. )~
independent variables 21,22,...,zn. The most important aspect of
the ACE4 (or GRACE5) algorithm is that the data The application of these quadratic polynomials to the
transformations are constructed pointwise, rather than optimal transforms determined by non-parametric regression
functional--therefore there is no need to associate a specific does not seriously degrade the quality of the fit to the
algebraic form with the individual data transformations. bubblepoint pressures. Eqs. 7 through 12 reproduce the
The application of the trivial restrictions of zero mean and bubblepoint pressures in the data set to an average error of 3.1
unit variance for the individual transformations results in percent and an average absolute error of 11.8 percerit. Fig. 8
essentially unique solutions. A final requirement is a gives a comparison of bubblepoint pressures calculated with
268
SPE 51086 CORRELATION OF BUBBLEPOINT PRESSURES FOR RESERVOIR OILS--A COMPARATIVE STUDY 3
equations 7 through 12 with the measured bubblepoint approximately 13 percent average absolute error. The average
pressures. absolute error of the neural network model increased from 6
Note that it is not necessary to fit the optimal data transforms percent (for the “original” data) to 25 percent for the
with quadratic polynomials -- equations of any fictional form independent data set. Figs. 10 through 12 show the calculated-
may be used. However, in this case these simple functions are measured comparison. Especially noteworthy is the comparison
adequate. between Figs. 9 and 12 which illustrates that a trained neural
network will not necessarily provide accurate predictions for
Neural Network Modeling: data not involved in the “training.”
Several other bubblepoint pressure correlations from the
MatlabTMwas used to build an “intelligent” software packageh literature were tested with the independent data base. Table ~
for data characterization. The software has three integrated and Figs. 13 through 15 show how predictions of the Standing,
modules: a) a preprocessing module for multivariate statistical Vasquez and Beggs,7 and Kartoatmodjo and Schmidts equations
analysis; b) a neural network module; in this case the (all non-linear regression type models) compare with the
backpropagation algorithm with the Levenberg-Marquardt measured bubblepoint pressures from the independent data base.
procedure as an optimization method for convergence, and c) a Other proposed bubblepoint pressure correlations2’9’1 0’1*were
post-training analysis module for evaluation of the performance also tested with the independent data set. In each case the
of the trained network by calculation of the errors for the average absolute error was larger than those given in Table 6.
training, validation, and testing data sets. Unfortunately, the best results for any correlation, whether
The 728-sample data set was divided into three subsets for prepared by non-linear regression, non-parametric regression, or
training, validation and testing. Half the data was used for the neural network, can only predict bubblepoint pressure (given the
training subset and one-quarter each for the validation and four commonly available independent variables) to an average
testing subsets. The training and validation data subsets are absolute error of about 13 percent. This means that errors of 25
used iteratively to develop the neural network model. The percent or greater are possible for a given situation.
model is “trained” with the training data and the validation data Bubblepoint pressure is used, either directly or indirectly, in
are used to determine when the “training” is optimal. The all oil property correlations. 12’3Thus the errors in estimates of
quality of model is then tested with the testing data subset. bubblepoint pressure will propagate throughout all estimates of
The final neural network model has four input nodes and the other fluid properties such as oil formation volume factor, oil
two hidden layers with 5 and 5 nodes respectively. A complete viscosity, oil density, etc.
description of the trained neural network model is given in Correlations for these other oil fluid properties are
Table 2. reasonably accurate given accurate values of bubblepoint
The neural network model showed very good performance pressure. It appears that an accurate bubblepoint pressure
for prediction of bubblepoint pressure. The bubblepoint correlation is not possible (given the usually available input
pressures predicted by the trained model agreed with the 728 data). Thus, two options are available: regular measurement of
sample data set to within 0.3 percent average error and 6.0 average reservoir pressure (pressure buildup tests, etc.) or
percent average absolute error. The comparison of calculated obtaining a representative sample of the original reservok fluid
and measured bubblepoint pressures is in Fig. 9. and measuring bubblepoint pressure and other properties in the
After the neural network was “trained,” the weight and bias laborato~.
vectors were incorporated into Fortran-90 and Visual Basic
interfaces so that the results could be used in a practical manner. Conclusions
Table 3 shows a comparison of the quality of fit of the three
models with the bubblepoint pressures used in their 1. The best possible correlations of bubblepoint pressure;
development. given the usual input data; solution gas-oil ratio at
bubblepoint, stock-tank oil gravity, separator gas specific
Validation of the Predictive Models gravity, and reservoir temperature; are accucate to an
average absolute error of about 13 percent. This means that
An independent data set consisting of data from 547 PVT predicted values of bubblepoint pressure could be in error
studies was used to test the three models. None of these data by 25 percent or more in some instances.
were included in the 728 sample data set used in developing the 2, Errors this large will cause unacceptably high errors in the
models. Table 4 gives the ranges of data in the independent data prediction by correlation of the other oil fluid properties of
set. Notice that the maxima and minima in the several variables interest: oil formation volume factor, oil density, oil
are very similar in the two data sets (Table 1, Table 4). viscosity, and oil compressibility.
The predictive abilities of the three models were tested using 3. The only options currently available for obtaining accurate
the independent variables from the independent data set and values of bubblepoint pressure are either regular field
comparing the calculated values of bubblepoint pressure with measurement of average reservoir pressures or laboratory
the measured values. Table 5 gives the results. The non-linear measurement with a sample representative of the original
regression and non-parametric models worked reasonably well reservoir oil.
on the independent data set--retaintig their accuracy of
269
4 W. D. MCCAIN JR., R. B. SOTO, P. P. VALKO, AND T. A. BLASINGAME SPE 51086
270
SPE 51086 CORRELATION OF BUBBLEPOINT PRESSURES FOR RESERVOIR OILS--A COMPARATIVE STUDY 5
Maximum Minimum
Reservoirtemperature 327 74
API 55 11.6
Gas gravity 1.367 0.556
GOR 1870 10
Pob 6700 70
The valuesfor the bias ( b 1, b2 and b3) and for the weights(w1, w2, and w3)are:
ZJl={-3.283148798632070.8024301701303761.08076979327138-2.83265706363712 1.68691436705254}
b3=-0.39379012457655
md
~3={-O.25763606379853-0.1027177075256732.50555221016993
-0.334182893301272-2.688487244623}
271
6 W. D. MCCAIN JR., 1?.B. SOTO, F’. P. VALKO, AND T. A. BLASINGAME SPE 51086
~
Standin -6.1 13.2
Vasquez and Beggs7 2.8 13.1
1
6000
~f9#; ~b:~::mgw+’e!mw Cormw.n (728 pmt.)
-1.5~
0 1000 2000 3000 4000 Smo 6000
Measured Bubbie-Point Preesure, ph paia Stock Tank Oil Gravity, “API
Fig. 2.- Grace optimal transformation of oil gravity ~APl).
Fig. l.- Comparison of calculated and measured bubblepoint
pressures for Velarde p~ correlation. Correlation data set.
272
CORRELATION OF BUBBLEPOiNT PRESSURES FOR RESERVOIR OILS--A COMPARATIVE STUDY 7
SPE 51086
37.0
z
~ ~,~
$
~ 6.0
5.5
5.0 ~
●
4,5
~ .“”
40~
-6-7-6-5-4-3-2-10 123
-’”~ Optimal Transform: In@d, Ifl(pda)
4.o 4.2 4.4 4.6 4.8 5.0 5.2 5.4 5.6 5.8 6.0
In(T#, lncF)
Fig. 6.- Grace optimal transformation of bubblepoint pressure, Pb.
20
~ 1.5- t
i
~ l.o - .
~=
z
~ 0.5
a
$? 0.0
~
~
~ .0.5
~ .1,0 ~ \
%*
~
z
0
●*
-1.5- %
-“ “’
. I.&acdOWMTrmdcm040ss
Gravity (728 pints)
(hum mm+ Plcqrm)
-2.0 l,,,,l,,,~l,,~,ib,,,l-1 ““1~
0.5 0.6 0.7 0.8 0.9 1.0 1.1 1,2 1.3 1.4 0 1000 2000 30”04 ‘f&a 5000 m
Fig. 4. - Grace optimal transformation of gas specific gravity Fig. 7.- Comparison of calculated and measured bubblepoint
(air=l .0). pressures for Grace p~ correlation-no model. Correlation data set.
-7~
i loio Z& 30”20 MO 5000 @
ln(Ff.b), ln(scUSTB) ❑ubble-Point Pressure, pb psla
M@aaured
273
8 W, D. MCCAIN JR., R. B. SOTO, P. P. VALKO, AND T. A. BLASINGAME SPE 51086
.’i:i
Z?.QOO
o-p -i +
Qmoo2m 30004000 sooo Oooo aloio2&o&4L”w 5ok EJiLl
MWIswed Bubb!%-Point Pressure, pb psia Measured Bubble-Point Pressure, ph PSla
Fig. 12. - Comparison of calculated and measured bubblepoint
. - Comparison of calculated and measured bubblepoint pressures fOrneural network Pb COITddiOt’k Independent data Set.
~r%!lwes for neural network f)b correlation. Correlation data sat.
w I
, 1 1 I I
274
CORRELATION OF BUBBLEPOINT PRESSURES FOR RESERVOIROILS--ACOMPARATIVESTUDY 9
SPE 51086
6000
U&am KaliorAlwa’lMchtiu Ctm!nlkm
ControlD.m*M6a (54 paints)
s 5ooo -
&
$
L
g 4ooo -
~
n ‘%
r
., . . . . .
0 looozooo3Wo~- ~
Measured Bubble-Point Pressure, ph Psia
275