SPE 56745

A Novel Non-Iterative Method for the Prediction of the PVT Behavior of Reservoir
Fluids
Varotsis N., SPE, Consultant-Oilphase Division of Schlumberger, Gaganis V., Consultant, Nighswander J., SPE, Oilphase
Division of Schlumberger, Guieze P., Schlumberger-Riboud Product Center

Copyright 1999, Society of Petroleum Engineers Inc.

this, other applications are seen in providing accurate esti-
This paper was prepared for presentation at the 1999 SPE Annual Technical Conference and mates of physical properties for use in well test validation and
Exhibition held in Houston, Texas, 3–6 October 1999.
reservoir/production engineering calculations as well as pro-
This paper was selected for presentation by an SPE Program Committee following review of
information contained in an abstract submitted by the author(s). Contents of the paper, as viding for the quality assurance of laboratory generated fluid
presented, have not been reviewed by the Society of Petroleum Engineers and are subject to
correction by the author(s). The material, as presented, does not necessarily reflect any posi-
property data.
tion of the Society of Petroleum Engineers, its officers, or members. Papers presented at SPE
meetings are subject to publication review by Editorial Committees of the Society of Petroleum
Engineers. Electronic reproduction, distribution, or storage of any part of this paper for com- Introduction
mercial purposes without the written consent of the Society of Petroleum Engineers is prohib- Great effort has been spent during the past decades for the de-
ited. 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 velopment of computational methods to predict reservoir fluids
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. PVT phase behavior. Two main approaches to this task are
prevalent. The first involves the use of correlations to provide
Abstract single point predictions of the physical properties of fluids, as
This paper presents a novel approach for the prediction of the functions of measured or estimated data1,2,3 (e.g, specific
complete PVT behavior of reservoir oils and gas condensates. gravities, GOR, saturation pressure). Although in several
The method uses key measurements that can be performed cases, fairly large databases have been employed in the devel-
rapidly either in the lab or at the well site as input to an Artifi- opment of these correlations, their inherent simplicity leads to
cial Neural Network (ANN) architecture. The ANN architec- a low degree of accuracy due to the small number of inde-
ture has been trained by a PVT studies database of over 650 pendent variables taken into account, the number of which
reservoir fluids originating from all parts of the world. rarely exceeds four4. In addition, the developed correlations
Tests of the trained ANN architecture utilizing a validation are often of only regional interest as the data sets utilized for
set of PVT studies indicate that, for all fluid types, most PVT their derivation are applicable to fluids of similar origin.
property estimates can be obtained with a very low mean rela- The other approach, which is currently extremely popular
tive error of 0.5-2.5%, with no data set having a relative error is the use of equation of state models upon which the entire set
in excess of 5%. This level of error is considered better than of calculations is based. This approach requires the knowledge
that provided by tuned Equation of State (EOS) models, which of the extended reservoir fluid compositions as well as other
are currently in common use for the estimation of reservoir physical properties of the fluid to tune against. Unfortunately,
fluid properties. In addition to improved accuracy, the pro- the accuracy of equation of state predictions depends heavily
posed ANN architecture avoids the ambiguity and numerical on the nature of the fluid, on the type of equation selected and
difficulties inherent to EOS models and provides for continu- on the operator-dependent tuning procedure5. The use of pa-
ous improvements by the enrichment of the ANN training da- rameters such as binary interaction coefficients, which could
tabase with additional data. be considered as fitting parameters, demonstrates the subjec-
The developed model has been named PVT Expert*, and is tive character of tuning which is essential for obtaining rea-
shown to be a robust, accurate and quick prediction tool that sonable estimates.
can be used for estimating reservoir fluid properties over the In this work, Artificial Neural Networks (ANNs), have
entire range of operating conditions and fluid types. The main been applied for the prediction of PVT fluid properties. The
application of the model is envisaged to be the rapid genera- application of ANN techniques is rapidly gaining popularity in
tion of PVT data based on field generated measurements. The petroleum engineering type of problems such as reservoir
results show that these reports will be of comparable accuracy characterization6,7, well completion and cementing8, well test
to a full PVT study generated in the laboratory. In addition to interpretation9, well logging10, reservoir simulation11, and spe-
cial cases of PVT physical properties12. Recently ANNs were
* employed for the prediction of the bubble point pressure and
Trademark of Schlumberger
2 [Link], [Link], [Link], [Link] SPE 56745

formation oil volume factor of crude oils given the reservoir with approximately 400 of the data sets being reservoir oils
temperature, oil API gravity, GOR and gas specific gravity13. and 250 data sets for gas condensate fluids.
In this work, ANNs were employed for the derivation of The full range of PVT properties considered in this data set
the complete set of physical properties describing the PVT be- is summarized in Table 1. Although this range is very large,
havior of both oils and gas condensates including constant the data population is not uniformly distributed but is biased
mass, differential liberation, viscosity, constant volume deple- towards low volatility oils and low condensation gas conden-
tion and standalone separation studies. Wherever applicable sates. This data bias is graphically represented in Figs. 2 & 3.
the properties are given versus the declining pressure for both Judging from this distribution one should expect higher confi-
the mono- and di-phasic regions. Furthermore, single valued dence for the predictions of fluids with properties in the dense
properties such as the thermal expansion factor at initial reser- data region compared to reservoir fluids lying in sparsely
voir pressure are also predicted. populated regions. To account for this data bias and to achieve
uniform performance of the ANN models appropriate trans-
Advantages of the ANN approach formations were selected and utilized. Separate inputs had to
An ANN can be defined as a multi-dimensional function be defined for the oils and the gas condensates due to the dif-
including a large number of parameters which relates input and ferent thermodynamic behavior of the two types of samples.
output data14. A typical 2-hidden layer ANN structure is illus- Input Data. The input data to the ANNs were selected to
trated in Fig. 1. include key measurements which could be rapidly performed
Several advantages can be attributed to ANNs rendering them on the produced fluids either onsite or in the PVT lab. The in-
suitable to applications such as considered here. Firstly, an put data set for the oil ANN models consists of reservoir fluid
ANN learns the behavior of a database population by self- composition, saturation pressure, reservoir temperature, fluid
tuning its parameters in such a way that the trained ANN density at the bubble point, viscosity of the dead oil, flash
matches the employed data accurately. Secondly, if the data molar ratio and flash liquid and gas densities. The input for the
used are sufficiently descriptive15, the ANN provides a rapid gas condensate ANNs consists of reservoir fluid composition,
and confident prediction as soon as a new case, which has not dew point, reservoir temperature, z factor at the dew point,
been “seen” by the model during the training phase, is applied. field GOR and separator and tank liquid densities. In both
Possibly, the most important aspect of ANNs is their ability cases, the required reservoir fluid compositional analysis in-
to discover patterns in data which are so obscure as to be im- cludes concentrations up to C12+.
perceptible to normal observation and standard statistical Output Data. Careful examination of typical PVT data led
methods. This is particularly the case for data exhibiting sig- to the selection of the optimum number of ANNs that had to be
nificantly unpredictable nonlinearities16. Traditional correla- developed to ensure consistency between different sets of data
tions are based on simple models which often have to be (e.g. between liquid relative volume and isothermal com-
stretched by adding terms and constants in order for them to pressibility). Most of the physical properties obtained from
become flexible enough to fit experimental data, whereas neu- PVT studies and contained in the database are described as
ral networks are marvelously self-adaptable. Using a suffi- functions of the declining pressure. These functions are inter-
ciently large database for training, ANNs allow property val- polating forms of a set of experimental points with each prop-
ues to be accurately predicted over a very wide range of input erty described by a certain function type containing a number
data. of parameters which determine the shape17. Unfortunately,
An ANN model can accept substantially more information these parameters can not be employed as ANN output because
as input to the model, thereby, improving significantly the ac- due to the complexity of the interpolating functions, several
curacy of the predictions and reducing the ambiguity of the re- satisfactory fits can be achieved for a given curve employing
quested relationship. Moreover, ANNs are fast-responding entirely different coefficients. To overcome this problem data
systems. Once the model has been “educated” predictions transformations had to be selected to describe numerically the
about unknown fluids are obtained with direct and rapid cal- curves in a ANN comprehensible way.
culations without the need for tuning or iterative computations. Data Transformations. Two of the approaches selected to
Furthermore, an outstanding attribute of the ANNs is their deal with the required data transformations are discussed fol-
capability of becoming increasingly “expert” by retraining lowing.
them using larger databases. Continuous enrichment of the The “curve shape” approach. Wherever applicable the
ANN “knowledge” eventually leads to a predictive model ex- data curves were split into mono and di-phasic regions, so that
hibiting accuracy comparable to the PVT data itself. only monotonic curves were treated. By observing several
curves of a specific PVT property, similarities could be identi-
Development of PVT Expert fied in the curve shape for fluids exhibiting a similar type of
The PVT Expert model was developed using a database con- thermodynamic behavior. For example, lets consider the oil
taining full PVT data for 650 reservoir fluids from around the volume factor curves as functions of declining reservoir pres-
world. This data covered the complete range of fluid types sure (Fig. 4a). By normalizing the diphasic Bo curve in a uni-
tary square, the entire set of Bo curves is included between the
SPE 56745 A NOVEL NON-ITERATIVE METHOD FOR THE PREDICTION OF THE PVT BEHAVIOR OF RESERVOIR FLUIDS 3

curve shape types of the very low and very high volatility oils two data subsets were used for the evaluation of the ANN per-
(Fig. 4b). Three major curve shape clusters can then be ob- formance.
served. The first type corresponds to black oils and exhibits For the PVT Expert training, an extended version of the
the shape of an arc. The second one, which corresponds to classical back-propagation algorithm with momentum was
medium volatility fluids, remains almost linear above 500 psi, used18. More specifically, the sigmoid slopes were allowed to
while the third type corresponds to highly volatile near critical be tuned during the training phase thus increasing the ANN
oils and exhibits a sigmoidal shape flexibility. The tuning is performed using an update law which
This corresponding states like approach attributes similar tends to minimize the approximation error.
shapes to the transformed physical property curves of reservoir Two hidden layers were used, the number of neurons in
fluids exhibiting similar thermodynamic behavior even if the each layer was varied and the networks were retrained until the
actual values of the original curves differ. The transformed optimum structure was obtained. Low approximation errors on
curves obtained are consequently discretized to obtain a set of both the training and the prediction sets indicates satisfactory
numbers which can be reproduced by an ANN. Those curve training.
endpoints which are unknown are predicted by single-output The objective for the training of the ANN models was to
ANNs. learn the underlying thermodynamic behavior of the training
A major advantage of this approach is that the PVT prop- dataset rather than individual data points (overtraining). To
erties versus pressure are predicted with equal precision at avoid overfitting, the error of the testing dataset was continu-
both the low and high pressure ends. Curves exhibiting com- ously monitored during training. By the time this error reached
plex shapes such as the gas-to-oil ratio curve of a volatile oil a stationary point, both the learning rate and the momentum
can be determined with a high accuracy. were decreased to the 1/10 of the initial value causing a slight
It is noted that special treatment of the curve-fit approach error reduction At the moment that the prediction error started
was required for the retrograde liquid deposit (RLD %) curves to increase again the training was halted. An example of the
of gas condensates. For these systems, the second pressure typical behavior of both the training and prediction errors
endpoint is not directly available hence, two ANNs are utilized during a training session is illustrated in Fig. 6.
to provide the two pressures at which condensation equals
90% of the maximum RLD %. These pressures are clearly de- Results
fined even for curves which are fairly flat around their maxima The performance obtained from the trained ANN was then
as illustrated in Fig. 5. Two additional ANNs provide the max doublechecked against the validation dataset which consisted
RLD% and the curve shape. of studies “never seen” by the model. The results of these
The “reference value” approach. Several single-output comparisons are provided below.
ANNs had to be built either for the prediction of individual Oil studies. The prediction error results obtained from PVT
single-valued properties (such as the thermal expansion factor) Expert for reservoir oils are illustrated in Table 2. The aver-
or for PVT curve endpoint values such as the oil volume factor age, standard deviation and the maximum error values are
at the bubble point discussed above. For some of the proper- listed for both the training and validation sets.
ties to be predicted it was found that they could be associated Although the performance of the ANNs is provided on a
to a known property. For example, the reservoir oil density at relative basis, for isothermal compressibility and oil viscosity
atmospheric pressure is related to the density at the bubble at saturation pressure it is expressed in absolute value. In this
point and the gas condensate relative density is related to the way, misleadingly large relative errors in cases where the re-
density at the dew point. quired property exhibits extremely low value are avoided. For
The advantage of this approach is that the model provides example, a viscosity prediction of 0.19 cP with respect to a re-
distances from a given value rather than direct predictions of ported value of 0.14 gives an absolute error of 0.05 and a rela-
the requested property. In other words, instead of aiming di- tive error equal to 39.2%!, when this is, in fact, a good predic-
rectly at the original property, the difference (or the ratio) of tion.
its value with respect to a reference property value is pre- Referring to Table 2, it can be seen that in most cases the
dicted. Consequently, even large prediction errors of the dis- mean relative error is between 0.5-2.5%, while the standard
tance lead to small errors of the original property. As a specific deviation is of the same order. This indicates that the error
example, for a fluid with ρb=700 kg/m3 and ρatm= 850 kg/m3, distribution is close to uniform. Furthermore, the prediction er-
an error in the prediction of the difference ρatm-ρb of the order rors are well spread along the range spanned by the PVT prop-
of 10% (135 instead of 150 kg/m3) leads to an estimated ρatm erties. This is demonstrated using reservoir fluid density pre-
of 835 kg/m3 (error of only 1.76%). diction (Fig. 7). This result shows that the ANNs perform
equally well for low, medium and highly volatile oils, and ap-
PVT Expert Training plies across all fluid properties.
The PVT database available was randomly split into a training The comparison of the predicted to the experimental rela-
(80%), a testing (10%) and a validation (10%) subset. The last tive volume and isothermal compressibility curves of a low
volatility oil are illustrated in Fig. 8. In Fig. 9, the perform-
4 [Link], [Link], [Link], [Link] SPE 56745

ance of the complete Bo prediction (single-output and curve PVT Expert gave average prediction relative errors for
shape predicting ANN) is illustrated.. both reservoir oils and gas condesate fluids between 0.5-2.5%.
An example of the reservoir oil density curve is shown in Among the principal advantages of this approach vis a vis the
Fig. 10. This property usually exhibits complex shapes which Equation-of-State based one are the lack of operator dependent
nevertheless are handled excellently by the curve shape ap- tuning, its non iterative nature, and its enhanced accuracy due
proach. In fact, in all cases, the match between the experi- to the ANNs approximation capabilities. The quality of the
mental and predicted curves is very good. predictions can be continuously upgraded by re-training the
The GOR curve for a medium volatility oil is illustrated in ANN architecture using larger databases.
Fig. 11 and the oil viscosity curve in Fig. 12. Once again, the Several important applications are envisaged for the devel-
match is very impressive. Finally, for a high volatility oil, the oped method. Rapid and accurate PVT studies can be gener-
liberated gas z factor curve, the high pressure endpoint of ated at the wellsite within hours enhancing thus the quality of
which is set at 90% of the saturation pressure, is shown in Fig. onsite fluid properties evaluation services such as the FPE19
13. and widely extending fluid analysis to reservoir fluid samples
Gas condensates. Table 3 presents the results of the ANN recovered by Formation Testers. These readily available qual-
model predictions for gas condensate properties. The accuracy ity values of the physical properties can improve the well test
of the predictions for the physical properties values indicates and reservoir/production engineering calculations upon which
that the performance of PVT Expert for gas condensates is of several important decisions are based on and test the data
the same order of magnitude as for the oils with most of the quality of fluid analysis laboratories.
relative error averages ranging between 0.5-2.5%. The low
standard deviations of the validation set errors give the almost Nomenclature
uniform relative error distribution. This is shown using z- Bo = Formation oil volume factor
factor as an example in Fig. 14. GOR = Gas to Oil Ratio
The prediction of a pair of retrograde liquid deposit (RLD) RLD = Retrograde liquid deposit
curves obtained from constant mass and constant volume de- ρ = Reservoir oil density
pletion studies respectively is illustrated in Fig. 15. The GRD = Gas relative density
agreement with the experimental curves is very good all along µ = Oil viscosity
the depleted reservoir pressure range. CM = Constant Mass study
The same curve for a low RLD% gas is shown in Fig. 16. CVD = Constant Volume Depletion study
In this example, the ANN has predicted the curve shape very
accurately including the tailing near the dew point. Subscripts
A di-phasic z factor curve example is illustrated in Fig. 17. b= bubble point
In this case both endpoints were known, thus requiring only a atm = atmospheric
curve shape to be obtained. Predictions of the cumulative pro-
duced fluid % and the gas relative density curves are provided Acknowledgments
in Figs. 18 and 19 respectively. The authors would like to thank the management of Oil-
phase, Division of Schlumberger EPS, for their permission to
Conclusions publish this paper.
The universal non-iterative PVT and physical properties pre-
diction model PVT Expert for both reservoir oils and gas con- SI Metric Conversion Factors
densates is presented. This model uses key measurements of cp x 1.0* E -03 = Pa • s
the reservoir fluid, characteristic of its thermodynamic behav- psi x 6.894757 E +00 = kPa
ior, as input to an ANN architecture that has been trained by a psi-1 x 1.450377 E +02 =Pa-1
PVT studies database of over 650 fluids originating from dif- References
ferent parts of the world, to generate quality predictions of full 1. Standing M.B.: “Volumetric and Phase Behavior of Oil Field
PVT reports. Hydrocarbon System”, SPE-AIME, 9th Printing (1981).
The non-uniform distribution of the database population 2. Lasater J.A.: “Bubble Point Pressure Correlation”, Transaction
led to the development of several output data transformations AIME, (1958), 213, pp. 379-81.
3. Vasquez M.E. & Beggs H.D.: “Correlations for Fluid Physical
to ensure satisfactory performance of the ANN models all
Property Prediction”, SPE 6719, (1977).
along the range of the property values and of the operating 4. De Ghetto et al. : “Reliability Analysis on PVT Correlations”,
conditions. The shape of the physical properties curves versus SPE 28904, (1994).
pressure is obtained from ANNs predicting the curve projec- 5. J.J. Martin “Cubic Equations of State – Which?”, Ind. Eng.
tion in the unitary square. The utilization of the reference value Chem. Fundam., (1979), Vol.18, No.2, pp. 81-97.
approach kept the prediction errors for the single value points 6. Mohaghegh S., et al: “Petroleum Reservoir Characterization
to a minumum. With the Aid of Artificial Neural Networks”, Journal of Petro-
leum Science & Engineering, (1996), Vol. 16, pp. 263-274.
SPE 56745 A NOVEL NON-ITERATIVE METHOD FOR THE PREDICTION OF THE PVT BEHAVIOR OF RESERVOIR FLUIDS 5

7. Briones M., et al.: “Thermodynamic Characterization of Vola- 13. Gharbi R. and Elsharkawy A.: “Universal Neural Network
tile Hydrocarbon Reservoirs by Neuronal Networks”, Procs. 3rd Based Model for Estimating the PVT properties of Crude Oil
Latin American and Caribbean Petroleum Engineering Confer- Systems”, SPE 38099, (1997).
ence, (1994), pp. 235-243. 14. Haykin S., Neural Networks - A Comprehensive Foundation,
8. Fletcher P. and Coveney P.: “Prediction of Thickening Times of MacMillan College Publ. Company, USA, 1994.
Oil Field Cements Using Artificial Neural Networks and Fourier 15. Fauset L., Fundamentals of Neural Networks, Prentice Hall, NJ,
Transform Infrared Spectroscopy”, Advn Cem Bas Mat, (1995), USA, 1996.
Vol 2., pp. 21-29. 16. Hornik K., “Multilayer Feedforward Networks are Universal
9. W. Sung et al.: “Development of the HT-BP Neural Network Approximators”, Neural Networks, (1989), Vol. 2, pp. 359-366.
System for the Identification of a Well-Test Interpretation 17. Guieze P., Williams J.M. and Varotsis N.: “Equations Deter-
Model”, SPE 30974, (1996). mine Locus of PVT Curves”, (1985), Oil & Gas Journal, Vol.
10. Balan B. et al.: “State-Of-The-Art in Permeability Determina- 83, No. 42.
tion From Well Log Data: Part 1- A Comparative Study, Model 18. Silva F. and Almeida L.: “Speeding Up BackPropagation”, Ad-
Development”, SPE 30978 (1995). vanced Neural Computers, (1990), Vol.3.
11. Ahmed T. et al.: “Application of Neural Network Parameter 19. Varotsis N. and Guieze P.: “Onsite Reservoir Fluid Properties
Prediction in Reservoir Characterization and Simulation – A Evaluation”, SPE 18317, (1988).
Case History: The Rabbit Hills Field”, SPE 38985, (1997).
12. Gharbi R. and Elsharkawy A.: “Universal Neural Network
Based Model for Estimating the PVT properties of Crude Oil
Systems”, In Situ, Vol.20, No.4, pp.367-394, (1996).

TABLE 1 – RANGE OF THE PVT DATABASE PROPERTIES

Oils Gas Condensates
Inputs Min Max Inputs Min Max
Bubble point (MPa) 0.25 46.70 Dew point (MPa) 10.86 72.62
o o
Reservoir temperature ( K) 311 466 Reservoir temperature ( K) 321 451
Fluid molecular weight (gr/mole) 37 352 Fluid molecular weight (gr/mole) 18 48
3
Stock tank oil density (kg/m ) 786 999 Field GOR (vol/vol) 675 75280
PVT Properties PVT Properties
Formation oil volume factor (vol/vol) 1.04 7.69 CMS - Retrograde liquid deposit (% vol) 0.01 42.94
Gas to oil ratio (vol/vol) 3 2010 CVD - Retrograde liquid deposit (% vol) 0.01 40.36
3
Reservoir oil density (kg/m ) 446 946 Cumulative produced fluid (% mole) 48.6 87.5
Gas phase relative density 0.561 3.078 Gas relative density 0.622 1.597
Gas phase z factor 0.686 1.011 z factor 0.789 1.516
Oil viscosity (cP) 0.10 47.33

TABLE 2 – PERFORMANCE OF THE ANN MODELS FOR OILS

Training Set Validation Set
Property Average Standard Max Average Standard Max
error deviation error error deviation error
Formation oil volume factor (%) 1.17 1.02 5.40 1.52 1.42 4.29
-6 -1
Isothermal compressibility (10 psi ) 1.07 0.98 6.72 1.06 0.92 5.25
Gas to Oil Ratio (%) 3.84 3.73 11.94 6.82 7.35 10.05
Reservoir oil density (%) 0.45 0.39 2.07 1.06 0.87 3.36
Oil viscosity (cP) 0.30 0.81 6.00 0.40 0.63 2.39
z factor at 90% of pb (%) 2.11 1.74 6.54 1.84 1.49 5.48
Liberated gas relative density (%) 1.80 1.70 7.34 1.99 1.70 9.30
6 [Link], [Link], [Link], [Link] SPE 56745

TABLE 3 – PERFORMANCE OF THE ANN MODELS FOR GAS CONDENSATES

Training Set Validation Set
Property Average Standard Max Average Standard Max
error deviation error error deviation error
CMS - Relative volume (%) 1.73 1.41 7.47 1.80 1.28 4.70
CMS - Maximum condensation (% vol) 0.55 0.59 3.34 1.06 1.24 4.35
CVD - Maximum condensation (% vol) 0.29 0.27 1.25 0.38 0.33 1.01
Cumulative produced fluid (%) 0.76 0.61 3.26 0.95 0.74 2.16
Gas relative density (%) 2.09 2.00 7.84 2.19 1.59 6.45
z factor (%) 0.87 0.83 3.59 1.03 0.87 4.50

100 100

80 90

60 80

Studies
%

40 70

20 60

Input Output
0 50
layer layer
1st hidden 2nd hidden 5 10 15 20 25 30 35 40
layer layer Fig. 3-Distribution of the max RLD % database values
Fig. 1-A typical 2-hidden layer neural network structure.
4.0
100

300
3.0

95
Medium volatility
Studies

200
%
2.0

High volatility
90
100
Low volatility
1.0
0 10 20 30 40 50
Pressure (MPa)
0 85
(a)
2 3 4 5 6 7 8
Fig. 2-Distribution of the oil volume factor database values
SPE 56745 A NOVEL NON-ITERATIVE METHOD FOR THE PREDICTION OF THE PVT BEHAVIOR OF RESERVOIR FLUIDS 7

1.2 1.6

1.4
1.0
1.2
Low volatility
0.8 1.0
%
0.8
(Bo-1)/(Bopb-1)

0.6
0.6

0.4
0.4
Medium volatility
0.2
0.2 0.0
High volatility
kg/m3 < 750 < 800 < 850 < 900 < 950
0.0 Fig. 7-The distribution of the prediction set relative errors of ρatm
0.0 0.2 0.4 0.6
p/pb
0.8 1.0 1.2 versus ρatm is fairly uniform.

(b) 33
Fig. 4-Three pairs of Bo curves sharing similar Bo @ pb but differ- Predicted
ent pb values (a). There is a perfect match for each pair of the 30
normalized curves in (b). Each pair is indicated by a solid and a Experimental
dotted line. 27
-6 -1
IC (10 psi )
24
Ambiguous
max RLD %
pressure 21
Maximum RLD %
18
90% of max RLD %
A B
15

12
25 30 35 40 45
Pressure (MPa)
(a)
pd 1.000
Predicted
Fig 5-The pressures at which 90% of the max RLD% occurs 0.995
Experimental
(points A, B) are predicted by two separate ANNs. A third ANN 0.990
predicts the value of the max RLD%.
0.985
0.980
Learning RV
0.975
rate
d ti 0.970
Training Error Training 0.965
termination
0.960
0.955
0.950
25 30 35 40 45
Prediction
Pressure (MPa)
E
(b)
Training Epochs Fig. 8-Comparison of the predicted isothermal compressibility (a)
Fig. 6-The training procedure is terminated at the time the predic- and the relative volume (b) curves to the PVT Lab ones. The PVT
tion error increases rapidly. Lab IC curve was obtained by differentiating the relative volume
curve.
8 [Link], [Link], [Link], [Link] SPE 56745

2.2 0.8
Predicted
Predicted
Experimental 0.7
2.0
Experimental
0.6
1.8
0.5
Bo (vol/vol)
1.6
0.4
µ (cP)
1.4 0.3

1.2 0.2

0.1
1.0
0 15 30 45 0.0
Pressure (MPa) 0 5 10 15 20 25
Pressure (MPa)
Fig. 9-Comparison of the predicted reservoir oil Bo curve to the
PVT Lab one Fig. 12-Comparison of the predicted viscosity curve to the PVT
Lab one.
870
1.00
Predicted
Predicted
Experimental
Experimental
850 0.97

ρ (Kgr/m3)
0.94
830

0.91
810 z factor

0.88

790
0 15 30 45 0.85
Pressure (MPa) 0 5 10 15 20
Pressure (MPa)
Fig. 10-Comparison of the predicted reservoir oil density curve to
the PVT Lab one. Fig. 13-Comparison of the predicted gas phase z factor curve to
the PVT Lab one. The first curve point is at 90% of the saturation
pressure (21.95 MPa).
350
1.2
Predicted
300
Experimental
1.0
250

200 0.8
GOR (vol/vol) %
150 0.6

100
0.4
50
0.2
0
0 5 10 15 20 25 30
0.0
Pressure (MPa)
< 0.8 < 0.85 < 0.9 < 0.95 <1 < 1.05 < 1.1 < 1.4
Fig. 11-Comparison of the predicted gas to oil ratio curve of a
medium volatility oil to the PVT Lab one. Fig. 14-The distribution of the prediction relative errors of z factor
versus z factor is fairly uniform.
SPE 56745 A NOVEL NON-ITERATIVE METHOD FOR THE PREDICTION OF THE PVT BEHAVIOR OF RESERVOIR FLUIDS 9

24 90
80 Predicted
CMS RLD %
20 Experimental
70

16 60
CVD RLD %
50
12 CPF (% mole)
40

8 30
20
Predicted
4
Experimental 10

0 0
10 15 20 25 30 0 5 10 15 20 25
Pressure (MPa) Pressure (MPa)

Fig. 15-Comparison of the predicted RLD % curves from CM and Fig. 18-Comparison of the predicted cumulative produced fluid
CVD studies to the PVT Lab ones. The marked points correspond curve to the PVT Lab one.
to points A and B of Fig. 5.
1.2 1.1

Predicted Predicted
1.0
Experimental Experimental
1.0
0.8

0.6 0.9

CVD RLD % GRD

0.4
0.8
0.2

0.0 0.7
5 10 15 20 25 30 35 5 10 15 20 25 30 35
Pressure (MPa) Pressure (MPa)
Fig. 16-Comparison of the predicted CVD RLD % curve of a low Fig. 19-Comparison of the predicted gas relative density curve to
liquid deposit gas to the PVT Lab one. the PVT Lab one.
1.00
Predicted
Experimental
0.95
z factor

0.90

0.85

0.80
0 5 10 15 20 25 30
Pressure (MPa)
Fig. 17-Comparison of the predicted CVD z factor curve to the
PVT Lab one.