Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

SPE 68235

Quality Assurance Tool for PVT Simulator Predictions


Varotsis N., SPE, Consultant-Oilphase Division of Schlumberger, Gaganis V., Consultant, Nighswander J., SPE, Oilphase
Division of Schlumberger

Copyright 2001, Society of Petroleum Engineers Inc.


This innovative approach was successfully tested against a
This paper was prepared for presentation at the 2001 SPE Middle East Oil Show held in large set of studies "unseen" by the ANN model. The ability of
Bahrain, 17–20 March 2001.
providing confidence for the accuracy of the PVT predictions
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
and of assessing their quality, significantly upgrades the
presented, have not been reviewed by the Society of Petroleum Engineers and are subject to applicability of the PVT simulator as a valuable reservoir
correction by the author(s). The material, as presented, does not necessarily reflect any
position of the Society of Petroleum Engineers, its officers, or members. Papers presented at management tool.
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 commercial purposes without the written consent of the Society of Petroleum Engineers is Introduction
prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300
words; illustrations may not be copied. The abstract must contain conspicuous An ANN can be defined as a multi-dimensional function that
acknowledgment of where and by whom the paper was presented. Write Librarian, SPE, P.O.
Box 833836, Richardson, TX 75083-3836, U.S.A., fax 01-972-952-9435.
includes a large number of parameters which relate input and
output data1. Advantages such as its ability to learn the
behavior of a database population by self-tuning its
Abstract parameters, the performance of direct and rapid calculations,
The currently available PVT simulators predict the physical as well as its capability of becoming increasingly “expert” by
properties of reservoir fluids with varying degrees of accuracy retraining, render this tool suitable for applications such as the
depending on the type of the model utilized, the nature of the prediction of PVT properties.
fluid and the prevailing conditions. Nevertheless, they all The application of ANN techniques is rapidly gaining
exhibit the significant drawback of lacking the ability to popularity in petroleum engineering problems such as
estimate the quality of their answers. reservoir characterization2,3, well completion and cementing4,
Artificial Neural Networks (ANNs), trained by large PVT well test interpretation5, well logging6, reservoir simulation7,
databases, are increasingly utilized to provide accurate and prediction of equilibrium k-values8. Recently, ANNs were
predictions of physical properties mainly due to their ability to employed for the prediction of the bubble point pressure and
learn from experience. The utilization of such models offers formation oil volume factor of crude oils given the reservoir
the unique capability of estimating the quality of their temperature, oil API gravity, GOR and gas specific gravity9.
predictions as their degree of competence can be evaluated for The PVT prediction model that was presented in our
each unknown test case. The accuracy of the ANN based PVT previous work10 is an ANN based simulator which has been
simulators depends heavily on the density of the database developed using a database containing PVT data of 650
compositional mapping around the coordinates of the reservoir fluids from around the world measured in the
unknown reservoir fluid. Unknown test cases found outside laboratory. This data covers the complete range of fluid types
the available training space may lead to poor predictions. and compositions as well as the reservoir operating conditions
In this work, a quality assurance tool is presented that is with approximately 400 of the data sets belonging to reservoir
integrated to the PVT Expert, which is an ANN based PVT oils and 250 data sets refering to gas condensate fluids.
properties prediction model. In the case of an unknown fluid The type of the PVT properties considered and the range
this tool qualifies the predictions of the PVT simulator based covered in this data set is summarized in Table 1. Although
on the evaluation of the affinity of the test case with the each PVT property in this data set varies over a very wide
training data sets contained in the utilised database. range, the data population is not uniformly distributed and, as
Subsequently, the competence with which the ANN model has expected, is biased towards low volatility oils and gas
learned the general trend in the area around any new test case condensates of low liquid yield. To account for this data bias
is assessed numerically through the development of an ANN and to achieve a uniform performance of the ANN models,
sextuple model per property. Finally, the utilization of the appropriate non-linear transformations were selected and
average predicted value eliminates the risk of considerable utilized.
deviations. Key measurements which could be rapidly performed on
the produced fluids either onsite or in the PVT lab were

Trademark of Schlumberger
2 N.VAROTSIS, V. GAGANIS, J.NIGHSWANDER SPE 68235

selected as input data for the developed ANNs . For reservoir Two hidden layers employing different number of neurons
oil fluids, the input data set consists of reservoir fluid in each layer were utilized and the networks were repeatedly
composition, saturation pressure, reservoir temperature, fluid trained until the optimum structure was obtained. Low
density at the bubble point, viscosity and density of stock tank approximation errors on both the training and the prediction
oil and flash gas density. The input for gas condensates sets indicates satisfactory training.
consists of reservoir fluid composition, dew point, reservoir The performance obtained from the trained ANNs was
temperature, z factor at the dew point, field GOR and tank subsequently double-checked against the validation dataset
liquid density. which consisted of studies “unseen” by the model. Recently,
The model provides the full set of properties obtained from an additional set of 161 oil PVT studies was used to test the
a full PVT study including the constant mass study, the performance of the models. The average, standard deviation
differential vaporization or the depletion study, the separation and the maximum error values for oils, for both the training
test and the viscosity study. Curves such as the isothermal and validation sets are listed in Table 2. Generally it can be
compressibility, the oil volume factor, the gas-to-oil ratio, the seen that the average relative error is less than 5%, while the
viscosity, the oil phase density, the gas phase z factor, the standard deviation is of the same order indicating that the
cumulative produced fluid and the retrograde liquid deposit prediction errors are rather evenly spread along the range
are predicted in the full pressure range. spanned by the PVT properties.
To allow the ANNs predict for each fluid full pressure Table 3, presents the performance of the ANN models for
range PVT properties with sufficient accuracy, the data curves gas condensate properties. Their accuracy indicates that the
were split into mono and di-phasic regions. The remaining performance of the ANN models for these fluids is of the same
monotonic curves were further normalized in the unitary order of magnitude as for the oils with all of the relative error
square. This corresponding states-like approach attributes averages ranging between 0.5-2.5%.
similar shapes to the transformed physical property curves of Table 4 presents the performance of the developed ANN
reservoir fluids exhibiting similar thermodynamic behavior models compared to that of the Standing13, Al-Mahroun14 and
even if the actual values of the original curves differ. For Vazquez & Beggs15 correlations for the prediction of the oil
example, by normalizing the diphasic Bo curve in a unitary formation volume factor at bubble point. The comparison test
square, the entire set of Bo curves is included between the is performed on 161 oil PVT studies. Clearly, the average
curve shape types of the very low and very high volatility oils relative error of the ANN models is almost half that obtained
(Fig. 1). The transformed curves obtained are consequently by the correlations.
discretized to obtain a set of numbers which can be Comparisons of typical predictions of the oil volume factor
reproduced by an ANN. The curve endpoints are predicted by and the depletion retrograde liquid condensation curves along
single-output ANNs. The transformed curves proved to be the pressure of unknown fluids (validation set) to the lab
much easier for the ANNs to learn than the original ones. It is reported curves are illustrated in Figures 2 and 3.
noted that special inversible non-linear transformations are
required to turn curves such as the liberated gas specific The Quality Assurance Tool
gravity into monotonic ones. A very important inherent drawback of every single PVT
For some of the single valued properties it was found that prediction model available so far (correlation, Equation-of-
they could be associated with a known measurement. For State based simulator, etc.) is that no method can be available
example, the reservoir oil density at atmospheric pressure is with which to estimate the quality of its predictions when
related to the measured density at the bubble point. Instead of applied to an unknown fluid. On the contrary, for an artificial
aiming directly at the original property, the difference (or the neural networks based simulator, direct links can be
ratio) of its value with respect to a related property value is established between the features of the unknown test case and
predicted. Using this technique, even large deviations of the the data population used for the training on the one hand and
difference/ratio predictions result in only small errors of the the numerical stability of the model on the other.
original property. Simple prediction quality estimation algorithms have been
The available PVT database was randomly split into proposed by Courieou16 and Liu17. The former, assumed that
training (80%), testing (10%) and validation (10%) subsets. the predictions for an unknown case lying in the convex hull
The last two data subsets were used for the evaluation of the defined by the multidimensional training datapoints will
ANN performance. The objective of the training of the ANN always be safe. However, when this method is applied to a
models was to learn the underlying thermodynamic behavior training space disposing regions of low datapoints density it
of the training dataset rather than individual data points can fail to correctly qualify the predictions.
(overtraining)11. The objective of this study was to develop a Quality
For the ANN models training, an extended version of the Assurance Tool following a twinfold approach. Firstly, to
classic back-propagation algorithm with momentum was develop a similarity test for estimating the affinity of any new
used12. More specifically, the sigmoid slopes were allowed to reservoir fluid test case with the training database cases. The
be tuned during the training phase, thus increasing the ANN accuracy of the prediction of the thermodynamic behavior of
flexibility. The tuning is performed using an update law which any unknown fluid using the ANN approach depends on its
tends to minimize the approximation error. similarity with the available training fluid datapoints.
SPE 68235 QUALITY ASSURANCE TOOL FOR PVT SIMULATOR PREDICTIONS 3

Secondly, to develop a stability test for assessing the degree of the two studies inputs weighted by the sensitivities of the
competence with which each ANN learned during its training ANN output y around the new case input.
the underlying trend of the required PVT properties in the
∂y new
neighborhood of the unknown test case. i c = ∑ ∆j
j ∂j x new .........................(2)
Similarity Test
This test is based on the principle that the existence of j = p b , Tres , MWRF , ρ b , ρ STO , MWFG , MWSTO
neighboring training studies similar enough, from the point of
The required partial derivatives of the trained ANN model
view of the input, to the new test case can ensure that the
in Eq. 2 for the evaluation of the sensitivities-weighted
ANN models will interpolate safely among “similar” training
distance are based on the formulation proposed by Choi and
data sets and hence their predictions can be trusted as accurate.
Choi18 and are evaluated at the new case input. Let us consider
The principle is illustrated in a simplistic 2-D plot in
as an example that the conditions sub-index is evaluated firstly
Figure 4, where a white inner circle having the new test fluid
for two volatile fluids exhibiting molecular weights of 35 and
in its center defines the “similarity region”, in the sense that if
40 gr/mole respectively and secondly for two black oils of
training data sets are positioned inside this circle, confidence
molecular weights of 250 and 255 gr/mole respectively.
could be provided for the quality of the test case ANN
Although the molecular weight difference is the same in both
predictions. If training data sets are only positioned in the
cases, in the former this difference is evidently multiplied by a
outer grey circle, limited “similar” information was available
larger sensitivity leading thus to a much larger sub-index value
for the ANN to acquire during its training. Finally, if training
than in the latter.
data sets are found only in the area outside the white and grey
The relative error sub-index ir. It counts for the relative
circles, it is assumed that insufficient knowledge was available
error with which the ANN learned the training dataset
locally for the training of the ANN.
“similar” case due to insufficient training and/or experimental
To evaluate the positioning of each one of the training
errors. The value of ir. would be equal to zero in the ideal case
cases in the multi-dimensional space with respect to the new
that the training of the ANN was perfect. This index is used to
test case, the following should be quantified for each pair of
eliminate the risk of rendering as a "neighbor" a training case
studies:
for which the model displays significant prediction deviation.
• the similarity of the fluids compositions The output sub-index io. It counts for the closeness
• the differences of their additional inputs (reservoir between the values of each physical property predicted for the
temperature, saturation pressure etc.) weighted by factors new test case and of the training dataset neighboring case.
related to the sensitivity of the predicted property with This index is defined as follows :
respect to each input
• the prediction error with which the ANN model has y new − y train
learned during the training procedure the properties of io = .........................................................(3)
y new
any neighboring cases as this error could affect the
predictions for the new test case. The four sub-indices are compared to predefined
To mathematically formulate the principle described thresholds tc, tx, tr, to using the following transformation:
above, a similarity index s is defined such that the larger the
value of the index, the closer the training dataset study is  ij 
î j = 101 − , j = c, x , r , o ..............................................(4)
considered to be to the new test case. This single similarity  tj 
 
index is constructed from the following four sub-indices :
The composition sub-index ix. It counts for the It can be readily seen that the transformed sub-indices
dissimilarity between the compared studies with respect to values of Eq.4 can take either positive or negative values
their reservoir fluid compositions. It is defined as the standard indicating either a close or a poor match respectively. For
deviation of the differences of the compositional inputs of the instance, î x = 10 implies that there is a perfect reservoir fluid
two studies. composition match between the new test case and the database
{ }
i x = std ∆x j , j = Ν 2 , ... , C n + .........................................(1) fluid, while îo < 0 implies that the difference of their outputs
For similar compositions displaying small differences, this is larger than the permitted threshold. Obtaining at least one
sub-index exhibits low values. On the other hand, it displays negative sub-index indicates that the training case is not close
large values in such cases when a black oil is compared to a enough to be qualified as “similar”.
volatile one. It is noted that the average of the differences of Once the sub-indices are calculated, the global similarity
the compositional inputs equals zero, since both compositions index s is produced by their harmonic mean as indicated in the
sum to 100%. following equation :
The conditions sub-index ic. It counts for the dissimilarity
between two studies with respect to their input, other than { }
s = 4 î c ⋅ î x ⋅ î r ⋅ î o ⋅ min sgn(î c ), sgn(î x ), sgn( î r ), sgn( î o ) .(5)
fluid composition. This index is defined as the difference of
4 N.VAROTSIS, V. GAGANIS, J.NIGHSWANDER SPE 68235

The similarity index s acquires a maximum value of 10 for to qualify the prediction due to the lack of close neighbors.
an unknown fluid exhibiting exactly the same input as one However, the functions realized by six ANN models (dashed
training dataset case which was learned by the ANN model curves) are very close to each other as well as to the true
during its training with a relative error of 0%. A negative underlying data trend. Consequently, the marks which
value of s obtained against all training cases indicates that no correspond to the predictions of the ANN model sextuple are
neighboring fluids were found in the “similarity region”. also very close to each other proving the competence of the
Repeated experiments confirmed that safe predictions were ANNs in that area.
often obtained, although no training study could be identified For evaluating the stability of the ANN sextuple
inside the “similarity region”. It was noted that for these cases, predictions y i for an unknown PVT study and a specific
a certain number of training data points lay in the “limited property, the following indices are utilized :
similarity region” possessing negative similarity index values The maximum span. It is defined as the difference of the
in the [-2,0] range. Thus, when a certain number of training two outmost predictions
cases lie in the limited similarity region, it was proven that the
ANN predictions can be safely trusted. max{y i } − min{y i }
span = ............................................(6)
It should be emphasised that the required tolerances for the avg{y i }
developed similarity test were set strictly enough to safeguard
that warnings will be issued whenever insufficient similarities The maximum deviation. It is defined as the largest
are detected with respect to the input between each test case difference between the predictions and their average
and the database studies. max{ y i − avg{y i } }
Nevertheless, several unknown test fluids were noted for dev = ............................................(7)
which the ANN predictions were accurate enough despite the avg{y i }
fact that their description was considered as terra incognita by Both definitions are graphically illustrated in Figure 6. For
the similarity test. This happened because the neural network qualifying the ANN prediction for a new test case, it is
algorithm managed in these cases to extrapolate correctly the required that both span and dev do not exceed 8% and 5% of
underlying physical trend of the training database population. the average values predicted by the ANN sextuple
Thus, there are cases of unknown studies for which no respectively.
neighboring training case can be located and the similarity test An illustrative example for the prediction of the oil
produces warnings, although the ANN model manages to formation volume factor, Bo at the bubble point for a case
provide accurate predictions. study unknown to the model is provided in Table 5 and
Figure 6. The maximum span is found to be between the
Stability Test ANN2 and ANN6 models predictions and equals 7.06% of the
To overcome this problem, a supplementary approach was predicted values average. The maximum deviation from the
adopted for evaluating the ANN algorithm’s behavior. For any average value occurs for the output of ANN6 and equals
unknown test case falling within an “area of competence” vis- 4.22% of the former. None of the indices exceeds the defined
a-vis the underlying physical behavior and for each specific tolerances, thus rendering the prediction of the above study
output, any number of ANN models obtained by training is stable.
expected to produce close outputs with minimum fluctuation
around their average value. On the other hand, for any
Utilization of the ANN sextuple predictions average
unknown study lying outside the competence of the models, value.
the predictions can vary over a wide range.
A major advantage for the accurate prediction of PVT
For the implementation of this test a numerical method properties which arises from the introduction of the Quality
was selected rather than an analytical one due to the high non-
Assurance Tool is that it offers the opportunity to use the
linearity of the ANN’s structure. Five extra ANN models, per average output of the ANN sextuple instead of a single output.
predicted property, were added to the original PVT Expert In this way, significant prediction errors are avoided in the
model using the same training and prediction data sets. Each
cases of fluids not satisfying the similarity and stability
sextuple represents multiple attempts of best fitting the
criteria. Denoting for some specific property by y , ŷ i and ei
available data population.
This principle is illustrated simplistically in Figure 5 for a the lab measured PVT value, the ANN sextuple outputs and
single variable regression fit. The solid line represents the the predictions deviations respectively, it follows that :
hypothetical data underlying trend of a training dataset. The ŷ i = y + e i .....................................................................(8)
training datapoints enclosed in curves define densely
populated areas which in turn correspond to similarity regions.
The ANN models predictions for any unknown fluid which The average prediction y is given by :
lies in these regions will be accurate as it will be indicated by
1 1
the similarity test.
For test cases lying outside the confidence area such as the
y=
n ∑ ŷ
i
i = y+
n ∑e
i
i → y + b ...............................(9)

fluid shown with a circle in Figure 5, the similarity test fails


SPE 68235 QUALITY ASSURANCE TOOL FOR PVT SIMULATOR PREDICTIONS 5

The error distribution ei is expected to be a normal one19 ( î x = −1.14 ). As expected, the respective similarity indices s
with its center at the mean prediction offset b. The larger the
number of the trained ANN models, the closer is the error are negative indicating that the particular training case should
be considered as lying outside the “similarity region”.
average to the bias b which in turn depends on the ANN’s
accuracy obtained by the training procedure. By enlarging the On the other hand, the ANN sextuple predictions for Bo
training database and repeating the retraining, the bias b value display dev and span equal to 0.97% and 1.81% respectively,
qualifying therefore the confidence and competence of the
can be decreased. Following the example demonstrated in
Table 5, the error of the average predicted output equals trained models. Indeed, the deviation of the prediction with
0.44% which is much lower than the errors of individual ANN respect to the lab measured value is 2.2%. For the GOR, the
models’ predictions fluctuate in a wider range as indicated by
models (e.g 2.40% for ANN2 and4.64% for ANN6). By
averaging the prediction values, large deviations of individual the deviation and span indices exceeding their predefined
ANN models are smoothed. thresholds (7.69%, 12.5%). However, although both the
similarity and stability tests produced warnings, the utilization
The flowchart of the developed Quality Assurance Tool is
presented in Figure 7. Firstly, the amount of information that of the averaging prediction value leads to a reasonably
the ANN has learned during its training in the area around the accurate prediction of 104 (v/v) which is less than 3.1% off
the lab measured value.
unknown test case, which defines its degree of confidence, is
assessed. Secondly, the ANNs degree of competence is The Quality Assurance Tool is also utilized for qualifying
evaluated numerically to define the stability of their the ANN models’ predictions for the gas condensate fluids
properties. Table 8 illustrates its application for the prediction
predictions. Finally, for each PVT property, the ANN sextuple
average output value is adopted as the PVT Expert’s of the maximum retrograde liquid deposit % (RLD) occurring
prediction thus minimizing the risk of error deviations. during a Constant Mass study for an unknown fluid. A training
database study was identified in the “similarity region” as
indicated by the positive values of the sub-indices and of the
Examples
global similarity index (4.77). The ANN sextuple outputs
The performance of the Quality Assurance Tool is illustrated
through two examples of predictions for the oil volume factor fluctuate in a narrow range, leading to dev and span values of
and the gas to oil ratio of a reservoir oil at bubble point 4.18% and 7.74% respectively. The results of both tests render
the ANN models predictions safe and, indeed, the actual
conditions.
The application of both the similarity and the stability tests relative error of the prediction of a randomly selected single
on a new fluid and on its closest training case for the first ANN (ANN1) and of the adopted average value for this
specific fluid were 0.43% and 1.3 % respectively.
example is illustrated in Table 6. The value of the
composition index ix is unique for both calculations. It is
calculated from the fluids compositions shown in the leftmost Results
For validating the method described above, the Quality
columns and equals 5.23, as illustrated in the Similarity Test
section of the Table. The very high values calculated for the Assurance Tool was tested against a genuine set of 161 oil
conditions indices ic (8.73 and 7.53 respectively) are explained fluids of worldwide origin.
For 96.2% of the unknown fluids tested (155 oils), the
by the very close input values (pb, Tres, etc) listed in the
topmost rows of the Table. The neighboring case has been average predicted values for the Bo deviated less than 5%
“learned” by the ANN models with learning errors of 0.98% from the lab measured values. The percentage of the fluids
which were successfully qualified by the similarity and
and 1.62% for the Bo and the GOR respectively, thus leading
to learning error indices ir of 8.05 and 6.77. The global stability tests was 95.6%.
similarity indices s are positive and equal to 7.58 and 6.38 for For the prediction of the GOR, the average predicted
the Bo and the GOR respectively indicating that the values deviated less than 5% from the lab values for 85.1% of
neighboring case is close enough to guarantee high confidence the 161 oils. The percentage of the fluids which were
successfully qualified by the similarity and stability tests was
for the trained ANN models.
From the Stability Test section of Table 5, it can be 73.9%.
readily seen that the outputs of the ANN sextuple fluctuate
within a very narrow range with dev and span displaying Conclusions
values as low as 0.71%, 1.26% and 1.25, 1.73% respectively, An innovative approach for qualifying the predictions of a
thus qualifying the models competence and prediction PVT simulator based on artificial neural networks was
developed. The principle upon which this approach is based
stability. As expected, the ANN prediction relative errors of a
randomly selected single ANN for the unknown case equal on can be applied to all kinds of artificial neural network
0.3% and 0.6% for Bo and GOR respectively rendering the based prediction models utilized in the petroleum industry.
Firstly, the Quality Assurance Tool estimates the “similarity”
predictions accurate. The adopted average value leads to even
lower errors of 0.2% and 0.1%. of the reservoir fluid under test to the characteristics of the
A second example is presented in Table 7. The closest fluids available in the training database by calculating the
similarity indices. Secondly, it assesses the degree of
training database case found for the unknown fluid exhibits
similar input to the latter with the exemption of composition, competence with which each ANN learned during its training
6 N.VAROTSIS, V. GAGANIS, J.NIGHSWANDER SPE 68235

the underlying trend of the required PVT properties in the Petroleum Science & Engineering, (1996), Vol. 16, pp. 263-
neighborhood of the unknown test case. 274.
The development of ANN sextuples per property improves 3. Briones M., et al.: “Thermodynamic Characterization of
greatly the models’ performance by allowing the utilization of Volatile Hydrocarbon Reservoirs by Neuronal Networks”,
Procs. 3rd Latin American and Caribbean Petroleum
the predictions average instead of a single output. The Engineering Conference, (1994), pp. 235-243.
application of such a filter smooths any probable prediction 4. Fletcher P. and Coveney P.: “Prediction of Thickening Times of
deviations which may appear due to insufficient training, thus Oil Field Cements Using Artificial Neural Networks and Fourier
minimizing the risk of obtaining significant predictions errors Transform Infrared Spectroscopy”, Advn Cem Bas Mat, (1995),
for fluids which do not satisfy the similarity and stability Vol 2., pp. 21-29.
criteria. 5. W. Sung et al.: “Development of the HT-BP Neural Network
The PVT Expert was successfully tested against a genuine System for the Identification of a Well-Test Interpretation
set of 161 oil fluids of worldwide origin and in 96.2% and Model”, SPE 30974, (1996).
85.1% of the cases for the Bo and the GOR respectively, it 6. Balan B. et al.: “State-Of-The-Art in Permeability
Determination From Well Log Data: Part 1- A Comparative
provided average predictions deviating less than 5% from the Study, Model Development”, SPE 30978 (1995).
lab values. Such a tool enhances the credibility of several 7. Ahmed T. et al.: “Application of Neural Network Parameter
applications based on predictions. Among them is the Prediction in Reservoir Characterization and Simulation – A
generation within hours of rapid and reliable PVT studies at Case History: The Rabbit Hills Field”, SPE 38985, (1997).
the wellsite thus improving the quality of onsite fluid 8. Habiballah W.A, et al.: “Use of Neural Networks for Prediction
properties evaluation20 and widely extending fluid analysis to of Vapour-Liquid Equilibrium K-Values for Light Hydrocarbon
reservoir fluid samples recovered by Formation Testers. Mixtures ”, SPE 28597, 1994.
9. Gharbi R. and Elsharkawy A.: “Universal Neural Network
Based Model for Estimating the PVT properties of Crude Oil
Nomenclature
Systems”, SPE 38099, (1997).
Bo = Formation oil volume factor 10. Varotsis N., et al: "A Novel Non-Iterative Method for the
GOR = Gas to Oil Ratio Prediction of the PVT Behavior of Reservoir Fluids", SPE
RLD = Retrograde liquid deposit 56745, (1999).
pb = Bubble point 11. Fauset L., Fundamentals of Neural Networks, Prentice Hall, NJ,
Tres = Reservoir temperature USA, 1996.
ρb = Reservoir oil density 12. Cho S., Kim J. : "Rapid Backpropagation Learning Algorithms",
ρSTO = Stock tank oil density Circuits Systems and Signal Processing, (1993), Vol.12, No.2,
pp. 155-175.
MWRF = Reservoir fluid molecular weight
13. Standing, M. B.: ”A Pressure Volume Temperature Correlation
MWSTO = Stock tank oil molecular weight for Mixture of California Oils and Gases,” Drill. & Prod. Prac,
MWFG = Flash gas molecular weight API, Dallas (1947) 275-87.
14. Vazquez, M. and Beggs, H. D.: ”Correlation for Fluid Physical
Subscripts Property Prediction,” SPE 6719, (1980).
b= bubble point 15. Al-Marhoun, M. A.: “New Correlation for formation Volume
STO = stock tank oil Factor of oil and gas Mixtures” Journal of Canadian Petroleum
FG= flash gas Technology (March 1992) 22-26.
RF = reservoir fluid 16. Courieu P.: "Three Algorithms for Estimating the Domain of
Validity of Feedforward Neural Networks", Neural Networks,
(1994), Vol.7, No.1, pp.169-174.
Acknowledgments 17. Liu Y. : "Unbiased Estimate of Generalization Error and Model
The authors would like to thank the management of Selection in Neural Networks", Neural Networks, (1995), Vol.8,
Oilphase, Division of Schlumberger EPS, for their permission No.2, pp.215-219.
to publish this paper. 18. Choi J., Choi C.: "Sensitivity Analysis of Multilayer Perceptron
with Differentiable Activation Functions", IEEE Trans. on
References Neural Networks, (1992), Vo.3, No.1, pp. 101-107.
1. Haykin S., Neural Networks - A Comprehensive Foundation, 19. Anderson T., An Introduction to Multivariate Statistical
MacMillan College Publ. Company, USA, (1994). Analysis, John Willey & Sons, NY, USA, 1984.
2. Mohaghegh S., et al: “Petroleum Reservoir Characterization 20. Varotsis N. and Guieze P.: “Onsite Reservoir Fluid Properties
With the Aid of Artificial Neural Networks”, Journal of Evaluation”, SPE 18317, (1988).
SPE 68235 QUALITY ASSURANCE TOOL FOR PVT SIMULATOR PREDICTIONS 7

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
Reservoir temperature (oK) 311 466 Reservoir temperature (oK) 321 451
Fluid molecular weight (gr/mole) 37 352 Fluid molecular weight (gr/mole) 18 48
Stock tank oil density (kg/m3) 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
Reservoir oil density (kg/m3) 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 MODEL FOR 161 OILS


Training Set Validation Set
Property Avg error St.dev Max error Avg error St.dev Max error
Formation oil volume factor (%) 1.17 1.02 5.40 1.94 2.93 8.34
Gas to Oil Ratio (%) 3.84 3.73 11.94 5.51 7.21 13.11
Liberated gas relative density (%) 1.80 1.70 7.34 2.21 2.13 7.14

TABLE 3 – PERFORMANCE OF THE ANN MODEL FOR GAS CONDENSATES


Training Set Validation Set
Property Avg error St.dev Max error Avg error St.dev Max error
CMS - Relative volume (%) 1.73 1.41 7.47 1.80 1.28 4.70
CMS - Max condensation (% vol) 0.55 0.59 3.34 1.06 1.24 4.35
CVD - Max 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

TABLE 4 – PERFORMANCE OF CORRELATIONS VERSUS THE ANN MODELS FOR THE PREDICTION OF Bo at pb
Standing Al-Mahroun Vazquez & Beggs ANN Models
Average relative error % 2.59 2.24 3.23 1.17

TABLE 5 – SAMPLE ANN MODELS OUTPUTS FOR Bo AT THE pb


ANN 1 ANN 2 ANN 3 ANN 4 ANN 5 ANN 6 Avg Lab
Outputs 1.275 1.325 1.274 1.307 1.315 1.234 1.288 1.294
Error (%) 1.47 2.40 1.55 1.00 1.62 4.64 0.44 -

TABLE 6 – PREDICTION QUALITY ASSURANCE TEST FOR AN OIL


New fluid Training case pb Tres
o
ρb MWRF ρSTO MW STO MWFG
mole% mole% MPa K gr/cm3 gr/mole gr/cm3 gr/mole gr/mole
N2 0.17 0.13 New 10.01 385 731 130 864 215 30.2
CO2 0.24 0.23 Training 10.76 383 732 123 850 188 31.0
H2S 0.00 0.00
C1 45.94 45.74 Stability Test
C2 7.62 8.39 ANN 1 ANN 2 ANN 3 ANN 4 ANN 5 ANN 6 Avg Dev % Span %
C3 3.37 4.41 Bo 1.577 1.592 1.575 1.582 1.595 1.582 1.584 0.71 1.26
iC4 0.83 1.23 GOR 174 174 174 173 171 173 173 1.25 1.73
nC4 2.98 2.35
iC5 1.65 1.30 Similarity Test
nC5 2.10 1.70 ic ix ir io s
C6 3.05 3.03 Bo 8.73 5.23 8.05 9.01 7.58
C7 4.21 4.11 GOR 7.53 5.23 6.77 6.22 6.38
C8 4.41 4.68
C9 3.32 3.64 PVT Results
C10 2.40 2.76 PVT Lab ANN model Err % Avg Err %
C11 1.72 2.09 Bo 1.581 1.577 0.3 1.584 0.2
C12+ 15.99 14.23 GOR 173 174 0.6 173 0.1
8 N.VAROTSIS, V. GAGANIS, J.NIGHSWANDER SPE 68235

TABLE 7 – PREDICTION QUALITY ASSURANCE TEST FOR AN OIL (CORRECTION BY AVERAGING)


New fluid Training case pb Tres ρb
o
MWRF ρSTO MW STO MWFG
mole% mole% MPa K gr/cm3 gr/mole gr/cm3 gr/mole gr/mole
N2 1.40 0.63 New 10.01 385 731 130 864 215 30.2
CO2 4.70 0.48 Training 10.76 383 732 123 850 188 31.0
H2S 0.00 0.00
C1 17.28 20.77 Stability Test
C2 5.14 7.81 ANN 1 ANN 2 ANN 3 ANN 4 ANN 5 ANN 6 Avg Dev % Span %
C3 9.48 8.52 Bo 1.433 1.422 1.448 1.433 1.441 1.436 1.436 0.97 1.81
iC4 1.83 1.36 GOR 99 104 106 109 96 109 104 7.69 12.5
nC4 6.05 5.15
iC5 2.30 1.79 Similarity Test
nC5 2.67 3.02 ic ix ir io s
C6 3.94 4.32 Bo 6.05 -1.14 9.39 9.29 -4.95
C7 6.30 5.84 GOR 5.14 -1.14 4.52 5.68 -3.5
C8 6.15 6.61
C9 4.77 4.81 PVT Results
C10 3.96 4.40 PVT Lab ANN model Err % Avg Err %
C11 2.55 3.43 Bo 1.469 1.433 2.5 1.436 2.2
C12+ 21.47 21.05 GOR 107 99 7.5 104 3.1

TABLE 8 – PREDICTION QUALITY ASSURANCE TEST FOR A GAS CONDENSATE


New fluid Training case pb Tres
o
z MWRF ρSTO
mole% mole% MPa K - gr/mole gr/cm3
N2 0.12 0.36 New 27.68 399 0.95 23.7 755
CO2 5.08 4.79 Training 30.16 397 0.97 22.6 757
H2S 3.70 2.11
C1 81.62 83.81 Stability Test
C2 2.31 2.55 ANN 1 ANN 2 ANN 3 ANN 4 ANN 5 ANN 6 Avg Dev % Span %
C3 1.54 1.52 RLD 2.76 2.62 2.60 2.76 2.73 2.81 2.71 4.18 7.74
iC4 0.35 0.35
nC4 0.84 0.81
iC5 0.40 0.35 Similarity Test
nC5 0.52 0.45 ic ix ir io s
C6 0.68 0.57 RLD 6.38 2.65 3.84 7.95 4.77
C7 0.63 0.54
C8 0.57 0.45
C9 0.45 0.36 PVT Results
C10 0.36 0.29 PVT Lab ANN model Err% Avg Err %
C11 0.23 0.20 RLD 2.76 2.75 0.43 2.71 1.3
C12+ 0.62 0.49
SPE 68235 QUALITY ASSURANCE TOOL FOR PVT SIMULATOR PREDICTIONS 9

1.2
1.0
Predicted
1.0
Experimental
0.8
0.8
Low volatility

0.6 0.6
(Bo-1)/(Bopb-1)

CVD RLD %
0.4
0.4
Medium volatility
0.2
0.2

High volatility 0.0


5 10 15 20 25 30 35
0.0 Pressure (MPa)
0.0 0.2 0.4 0.6 0.8 1.0
p/pb Fig. 3-Comparison of the predicted CVD RLD % curve of a low
liquid deposit gas to the PVT Lab one.
Fig. 1-Three pairs of Bo curves sharing similar Bo @ pb and pb
values. There is a perfect match for each pair of the normalized
curves. Each pair is indicated by a solid and a dotted line.

2.2
Predicted Similarity region

2.0 Experimental
Limited similarity region

1.8
s=10
Bo (vol/vol) s=0
1.6 s=-2
s<-2

1.4
Fig. 4-Simplistic illustration to demonstrate the establishment of
1.2 the similarity test.

1.0
0 15 30 45
Pressure (MPa)

Fig. 2-Comparison of the predicted reservoir oil Bo curve to the


PVT Lab one. Similarity regions
Outputs

Prediction point

Training point

Inputs

Fig. 5-Single-variable regression fit to demonstrate the principle


of the stability test.
10 N.VAROTSIS, V. GAGANIS, J.NIGHSWANDER SPE 68235

Lab value
1.30

Average b
Span

Dev
1.25

1.20
ANN 1 ANN 2 ANN 3 ANN 4 ANN 5 ANN 6

Fig. 6-Definition of maximum span, deviation and the average


output value for the prediction of Bo at pb.

New fluid
input

PVT Expert

Similarity Training Stability


check Database check

Establishment of Establishment
similarity of stability

Qualified
average
predictions

Fig. 7-Flowchart of the Quality Assurance Tool.

You might also like