Vous êtes sur la page 1sur 33

,

R13CHT
h

&
17 /32
AUG05 19t ?

SPE
UNSOLICITED .>
PUBLICATIONS

DEVELOPMENT OF V SCOSITY C~!?%ELATIONS FOR CRUDE O LS

by S.A. Kahn, University of Texas

ABSTRACT

This paper presents the development procedure of

empirical equations for estimating oil viscosity at the

bubble point, oil viscosity above the bubble point and oil

viscosity below the bubble point pressure. Oil viscosity at

tihe bubble point is developed as a function of solution gas

oil ratio, gas relative density, oil relative density and

relative temperature. The estimated value of oil viscosity

at the bubble point can then be used to determine both, oil

viscosity above and below the bubble point pressure, given

the pressure and the bubble point pressure of the oil. A

theoretical explanation of the developed correlations is

presented. These correlations should be valid for all types

of crude oils falling within the range of data used in this

study .

LITERATURE REVIEW

One of the earliest studies for viscosity estimation

were conducted by 3eal [1] in 1946. He presented two

correlation charts for determining dead oil viscosity. The

first correlation chart determines dead oil vi~.,<sity at


~=sodawoiPwoloUmanghmn
aqmcn
~f-91-lnOudaltoomsnwtyofPulolwm
fore!atrtwttm
Mctpa@mpuolkmm mmsPEjourm lnbnmw,g~to~~, ~
byttmUuhof(a). Pcmnlon toqmrnmctod toanahtrutotnotwllmttlm
mowofdo.wnmsps~ owt,P,o.aoxe3393s,~,-x
~USA.Tolox7WWSPSDAL


100 ? given the oil API gzavity. This chart was developed

using 655 data points of dead oil viscosity cbtair.ed from

492 oil fields mostly from the Ur:ited Watss. M average

relative error of 29.0% was reported. The second correlation

chart provides est.mates for dead oil viscosity at hiqher

temperatures. An additional data for 98 samples were used to

develop this correlation chart which showed an avera~le

~elative error of 25.6%. The average relative error of all

data points used in both charts was 24.2%.

In 1975, Beggs and Robinson [2] developed a correlation

for dead oil viscosity as a function of temperature and oil

API gravity. They found that a plot of log(T) versus

10g[lOg(~od+l)] revealed a straight line with constant slope

for each API gravity. A total of 460 data points were used

and an average relative error of -0.64% was reported.

A later development in dead oil viscosity correlation

was made by Glaso [3] in 2,980 who also based his correlation

as a function of temperature and oil API gravity. He used

data from six North Sea oil samples.

Another correlation chart developed by Beal [1] was for

oil viscosity below and at the bubble point. This chart

requires the knowledge of dead oil viscosity and solution

gas oil ratio at a given pressure. He used 351 viscosity

observations taken from 29 oil fields.


,, ..,.!
f
i

In 1959, Chew and Connally [4] presented a correlation

for oil viscosity below and at the bubble point as a

function of dead oil viscosi::y and solution gas oil ratio.

This correlation is a much improved form an~ covers a wider

range of dead oil vi$~cosities and solution gas oil ratios.

Their study showed that the relation between log(~b) and

log(i
~~ ) is a straight line at a fixed solution gas o~l

ratic. heir data were taken from the crude oil samples from

the United States, Can~da and South America.

2eggs and Robinson [2] also developed a correlation for

oil viscosity below and at the bubble point as a function of

dead oil viscosity and solution gas oil ratio. Like Chew and

Connaly [4], they also based their correlation on t.9%

straight line relationship between log(~b) and log(~od) at a

constant gas oil ratio, A total of 2,073 observations and

reported an average relative error of -1.83%.

A graphical correlation for estimating oil viscosity

above the bubble point was also presented by Beal [1]. Thi S

correlation is based on the relation between the rate of

change of oil viscosity above bubble point per unit pressure

increase above the bubble point pressure, and the bubble

,point oil viscosity. He used 52 observations taken from 26

crude oil samples showing an average relative error of 2.7%.


*
In 1976, Vazquez and Beggs [5] developed a relation for

oil viscosity above the bubble point as a function of bubble

s?E17U2
point oil viscosity, pressure, and bubble point pressure.

They used 3,143 o.oservations and reported axi aver age

relative error of -7.54%.


..

These correlations were tested by Sutton and FarshadG

in 1984, who used the data for oil samples from the Gulf of

Mexico. He found that dead oil viscosity was most accurately

determined by Glasos correlation. Beggs and Rcbinsons

9
correlation yielded best results for oil viscosity below and

~ at the bubble point. Oil viscosity above the bubble point

was bes: estimated by Vazquez and Beggss correlation.

Other viscosity prediction methods were presented by

Lohrenz, Bray & Clark [7], Little & Kennedy [8], and

Houpeurt & Thelliez [9]. These methods require the knowledge

of oil composition which makes viscosity prediction more

complex and less useful for practical purposes.

A detailed literature review of these correlations has

been presented by Khan [6]. Recently, Khan et al [7]

presented a comparison of their newly developed correlations

against other available correlations using data from Saudi

Arabian crude oils.

Most of the available correlations in the literature

were developed long time ago using limited data. All these

correlations are complex mathematical expressions and do not

explain the behavior of oil viscosity very clearly.

SPE17132
Furthermore, these correlations do not provide accurate

estimates of viscosity when applied to other are~~. The

objective of this paper is to study the behavior of oil

viscosity in more detail and develop simple and accurate

visccsity corr(:lations utilizing a large set of

experimentally measured viscosity data using simple and

multiple least square and least absolute regression


.
analyses. The data used exclusively belong to Saudi crude

oils, however, the correlations should be valid for all

types of crude oils falling within the range of data used in

this study.

VISCOSITY DATA

This study utilized viscosity data for 75 bottom-hole

samples taken from 62 Saudi oil reservoirs. The bottom-hole

samples had been flash separated to obtain solution gas oil

ratio, gas relative density and oil API gravity. Viscosity

data had been obtained by rolling ball viscosimeter for

various pressures and temperatures. The data were divided

into three different data bases. A total of 150 data points

were used for bubble pcint oil viscosity, 1,503 for oil

viscosity above the bubble point and 1,691 for oil viscosity

below the bubble point pressure. Following are the ranges

of the given data:


Bubble Point Oil Viscosity 0.13 to 17.9 Cp

Oil Viscosity Above Bubble Point 0.13 to 71.0 Cp

Oil Viscosity Below Bubble Point 0.13 to 77.4 Cp

Bubble Point Pressure 107 to 4315 psia

Solution Gas Oil Ratio 24 to 1901 SCF/STB

Gas Relative Density 0.752 to 1.367 (air=l)

Stock Tank Oil Gravity 14.3 to 44.6 API

Temperature 75 to 240 F

Pressure 14.7 to 5015 psia

C02 in Surface Cases 0.03 to 11.07 mol %

N2 in Surface Gases 0.02 to 1.01 mol %

H2S in Surface Gases 0.00 to 9.78 mol %

DEVELOPMENTOF VISCOSITY CORRELAT ONS

Non-linear simple/multiple 1 east square and least

absolute regression analyses were used to develop

correlations for viscosity. Firstly, independent variables

were determined by studying the relationship with the

dependent variable. Secondly, a model was chosen which best

correlated against the experimental data. The least square

and least absolute regression coefficients thus obtained

were fixed one after one to the nearest rounded or fraction

values and the final correlation was formulated. F-

distribution criteria was used to test for model adequacy.

The theory of regression analysis and model. adequacy test

spE 17132
has been presented in the Appendix.

Bubble Point Oil Viscosity

Non-linear multiple regression analysis on 15-;

experimentally determined bubble point oil viscosity data

points showed this viscosity to be a function of solution

gas oil ratio, gas relative density, relative temperature

and oil relative density i.e.


i

(1)

where

Bubble point oil viscosity, cp

Solution gas oil ratio, SCF/STB

141.5/(API + 131.5) = Oil Relative Density

Gas relative density, (air = 1)

(T + 459.67)/459.67 = Relative temperature

Oil gravity, API

Temperature, F

A number of mode 1s were tested to reach this

conclusion. Firstly, the effect of each parameter was

studied individually followed by the combined effect of the

parameters, The model so obtained was then modified

depending on the contribution of each parameter. This

procedure was continued until the final model was reached.

Some of these tested models are:

w----------- .
a18a2ea3Xa4a5
>o~ = aORsgropb (2)

al Za20a3ya4
~ob = aORsgro (3)

al Xa20a3Xa4
~ob = aORsgro (4)
+ asxo

al Za20a3 a4
~ob = aORs g r (b. + bl~o + b2Toz) (5)

alxa20a3 a4
wob = aORs g r
(1-3.) (6)

i where ai and b. represent the regression constants.


1

The study of these models revealed that Pob is not a

function of the bubble point pressure (Pb) since this

pressure itself is a function of the remaining four

parameters Rs, #q, go and Or as shown by Standing [8], Glaso

[3] and A1-Marhoun [9] . Oil relative density (Zo), however,

was found to have a very strong contribution. Eq. (6) proved

to be the most representative model, Table 1 shows the

results of the model adequacy test for this equation. An

F. value of 1,470 was found which is much greater than the

F-table value of 3.32 for a model adequacy of 99%.

The regression coefficients were determined both by

least square and least absolute methods and are shown in

Table 2. The final values of the coefficients were chosen

by carrying out a step-wise regression analysis, i.e., one


@
coefficient was fixed to the nearest rounded or fraction

value and the remaining other coefficients were determined


J,

....... Al.. ... ,.: .-m---..~. -. -. ., .? . -- .- - ..- .y


.
9 ,,--*=.mr.***,m~.*..m,~-f.f
.- ,., . ..*. . . ,.<.,... !

.~

by regression. The next coefficient was then fixed ant, ths

regression was carrieci out to determine the remaining

coefficients. This process was repeated until the constant

was determined last of all. Statistical analysis was


a.
carried out at each step 20 avoid any instability problem.

Xhefinal correlation is presented in Eq.(7) as follcws:

(7)
Mob =
m ors (l-l.)

oil viscosity Above Bubble Point

Once bubble point oil viscosity has been determined,

the only parameters governing the behavior of oil vircosity

above the bubble point are the bubble point pressyre, and

pressure i.e.:

Va = (~ob,pb,p) (8)

where

Ma = Oil viscosity above bubble point, cp

wob = Bubble point oil viscosity, cp

Bubble p{~int pressure, psia


b =
P = Pressure, psia

\
Some of the studied models are given below:

Pa = Mob (p/pb)a (9)

aO+alBob+a2Vob2
Pa = Mob (p/~J (lo)

~pE I?1J2
Pa = vo~ + a (P-Fb) (11)

Pa = ~ob + a Mob (P-Pb) (12)

Va = wob + aO+alVob+a2Pob2) (p-pb) (13)

a(P-Pb)
Pa = Vob e (14)

All these models were chosen to satisfy the condition

that l.ia
= Vob at P however, Eq.{14) revealed the best
= b
correlation. Model adequacy test for this equation (Table 1)

showed an F. value of 2,536,041 being far greater than the

F-table value of 6. 63 for 99% model adequacy. Results from

simple regression analysis on 1,503 data points are shown in

Table 3. Using the recommended values, the following

correlation is formulated:

9.6x10-5(P-Pb)
Pa = Mob e (15)

Oil Viscosity Below Bubble Point

The liberation of solution gas occurs when pressure is

lowered below the bubble point pressure and as a result, the

solution gas oil ratio changes continuously. Once bubble

point viscosity has been determined for an oil sample, then

solution gas oil ratio becomes a function of pressure below

the bubble point. Therefore, oil viscosity below bubble

point can be represented by a function of bubble point oil

SPEIU32
viscosity, bubble point pressure, and pressure i.e.:

Mb =
f(vob#Fb#p) (16)

where
yb = Oil viscosity below bubble point, cp

}Iob = Bubble point oil viscosity, cp

Bubble point pressure, psia


b =
P = Pressure, psia

Following are some of the tested models:

v~ = 11~ (Iwb)a (17)

a(P-Pb)
Mb = Mob t? (18)

Mb = Vob (p/pb) al ea2(p-pb) (19)

These models also satisfy the condition that Ub = Wob at

P Eq.(17) over-estimates the value of viscosity


= b
whereas Eq.(18) shows an opposite trend. Hence both of these

equations are not representative mode 1s; however, a

combination of both models in Eq.(19) proved to be

representative. Model adequacy test in Table 1 for this

equation shows an F. value of 17,402; a value much larger

than the F-table value of 4.61 indicating a model adequacy

of 99X. Table 4 shows the results for regression analysis

on 1,691 Jiscosity data points which resulted in the

following correlation:

sp~17132
-0.14 ~-2.5x10-4(P-Pb) (20)
lJ~ = Po~ (WjJ

DISCUSSION

For a crude oil above the bubble point pressure,

viscosity decreases with decrease in pressure until it

ruaches a minimum value at the bubble point pressure. This

decrease in viscosity is due to the volumetric expansion of

oil. Reducing the pressure below the bubble point causes an

increase in viscosity due to liberation of the solution gas

which predominates the viscosity decrease due to volumetric

expansion of oil and results in net increase in viscosity

until dead oil viscosity is reached at atmospheric pressure.

A typiral viscosity versus pressure plot of a crude oil is

shown in Fig. 1.

It can be noticed in Fig. 1 that viscosity decreases

almost linearly with reduction in pressure above the bubble

point pressure. Below the butible point pressure, the

viscosity increases abruptly with the reduction in pressure,

the viscosity curve having a concave upward trend. The

viscosity at the bubble point pressure is a point at which

the viscosity trend changes.

All viscosity correlations presented in the literature

estimatfi dead oil viscosity a? the starting point.

Thereafter, viscosity below or at the bubble point pressure


,
~.

- . .. - -

is estimated using the predicted value of dead oil

viscosity. Finally, viscosity above the bubble point

pressure is estimated using the predicted value of bubble

point oil viscosity. It means that the estimation of

viscosity above the bubble point involves a three step

procedure which gives rise to high errors. Another drawback

is that none of the correlations in the literature takes the

effect of solution gas relative density into account which

is an important factor for viscosity estimation.

In this study, a completely new approach has been

developed for viscosity estimation. The first step is the

estimaion of bubble point oil viscosity (Eq.7) which is the

point of trend chang~, in the plot of viscosity versus

pressure (Fig.1). Next, correlations for viscosity above and

below the bubble point pressure (Eqs.15 and 20) are

developed as a second step. No third step is needed at all.

The effect of gas relative density on bubble point oil

viscosity has been determined as is given by Eq.(7). This

equation shows that the bubble point oil viscosity is

directly ,proportional to the square root of gas relative

density. Furthermore, oil relative density and relative

temperature have been used while formulating this


~ correlation instead of oil gravity in OAPI and temperature

in F respectively. It has been done in order to have each

parameter as a dimensionless ratio and thus a simple

sPE 17132

I
.

relation has been obtained. It is shown that the bubble

pOIIIt oil VlsCOSlt~ is inversly proportional to the cube

root of solution gas oil ratio, 4.5th power of relative

temperature and cube of (l-TO).

The general form of Eq.(15) is given by Eq.(14) which

shows that the oil viscosity above bubble point behaves as

an exponential function of the increment in pressure above

the bubble point pressure with a as the viscosity

coefficient for crude oils. Equation (14) is the solution

of the following differential equation:

a = ( l/p) (dV/dP ) (21)

Equation (8) is similar to the oil compressibility equation

[10] and it can be concluded that oil viscosity and oil

compressibility exhibit similar behavior above the bubble

point pressure since change in viscosity is only due to the

volumetric expansion of oil.

However, below the bubble point pressure there occur

two simultaneous phenomena, i.e., volumetric expansion of

oil and liberation of solution gas. Consequently, the

viscosity relation found below the bubble point (Eq.20) is

different from the viscosity relation above the bubble point

(Eq.15) by a factor of P/Pb. The factor P/Pb is involved due

to the changing solution gas oil ratio below the bubble


point. The general form of Eq.(20) is given by Eq. (19) and

SPE 17132
. .

-.

since the solution gas ratio can be reasonably approximated,

below the bu~ble point pressure, by [20]:

P/Pb = Rs/Rsb (22)

the contribution due to solution gas oil ratio can be


replaced by a function of pressure as done in Eq. (19).
Here, al is the coefficient for the contribution due to

changing solution gas oil ratio and a is the viscosity


2
coefficient governing the contribution due to volumetric

expansion of oil. Using Eq.(22) and replacing P/Pb in


Eq. (19),

~b = Bob (R/Rs# al ea2(p-pb) (23)

or
a2(P-Pb)
~b = ob e (24)

where
al
nb = vb/(Rs/R=b) (2s)

and
al
ob = lJob/(Rsb/Rsb) = wob (26)

In general,
al
n = Y/(Rs/R~J (27)

Equation 24 is now the solution of the following


differential equation:

a2 = (1/u)(dn/dP) (28)

SPE 17132
~
s

which is similar to Eq.(21) except for ~ instead of ~. This


shows that below the bubble point pressure, n behaves
exactly as v above the bubble point pressure. The viscosity

COefflClentS, however, are different; la!!above and a2 below

the bubble point.

CONCLUSIONS

This study leads to the following conclusions:

1. Viscosity correlations for Saudi crude oils have

been developed. These correlations are simpler and

more accurate than the ones so far presented in the

literature.

2. The effect of gas relative density on oil viscosity

has been determined.

3. Oil viscosity above the bubble point behaves as an

exponential function of the pressure increment above

the bubble point pressure, and thus the behavior of

oil viscosity above the bubble point is similar to

that of oil compressibility.

4. Oil viscosity below the bubble point is governed by

a product of exponential function of difference

between the pressure and the bubble point pressure,

and a factor of P/Pb which takes into account the


-.
ij
;.
,, ,

changing solution gas oil ratio below the bubble

point pressure.

RECOMMENDATIONS

Following are some recommendations to extend this


study:

1. The developed correlations should be tested using

data from other geographical areas. Different


regression constants can be determined for other

crude oils.

2. The effect of non-hydrocarbons on oil viscosity

behavior should be studied.

3. A closer investigation can describe the effect of

different base crudes (aromatic, paraffin, etc.) on

oil viscosity behavior.

NOMENCLATURE

= Regression coefficients
.ai
a = Vector of length n containing ai

a = Vector of length n containing least square solution


.,
A = m x 2(m+n) matrix of independent variables

API = Oil gravity, API

bi = Regression coefficients
g
.. .
B

c = Vector of length 2(m+n) in least absolute method

di = Deviations between observed and estimated values


in least absolute method

di = Defined by Eq.(14)

di~ = Defined by Eq.(14)

E = Defined by Eq.(6)

F = Defined by Eq.(11)

= Defined by Eq.(24)
c)
h(w) = F-distribution curve given by Eq.(22)
i
! i = Observation index

m = Number of observations

n = Number of independent variables

P = Pressure, psia

Bubble point pressure, psia


b
R~ Solution gas oil ratio, SCF/STB

R Solution gas oil ratio at the bubble point,


sb
SCF/STB

SSR Defined by Eq.(25)

SSE Defined by Eq.(26)

T Temperature, E

x Independent variable

x m x (n+l) matrix of independent variables


in least square method

x Least absolute solution Vector of length 2(m+n)

Y Dependent variable

Y Vector of length m containing dependent variables

r Gas relative density at 14.7 psia and 60 F


9
(air = 1)

.,.
.......,-..,..+.
.. ..-. .
r. = 141.5/(API + 131.5) = Oil relative de:~s~ty
at 14.7 psia and 60 F

l(a) = Gamma function defined by Eq.(23)

c Error between the observed and estimated values


in least square method

i Defined by Eq.(64)

Defined by Eq.(62)

Defined by Eq.(63)
nob
Or (T + 459.67)/459.67 = Relative temperature

Oil viscosity above bubble point, cp

Oil viscosity below bubble point, cp

Bubble pint oil viscosity, cp


ob
Dead oil viscosity, cp

Degrees of freedom for first distribution


1
Degrees of freedom for second distribution
2

ACKNOWLEDGEMENTS

Acknowledgement is due to the Ministry of Petroleum and

Mineral Resources of Saudi Arabia for providing the data.

The authors are also grateful to the Department of Petroleum

Engineering at the University of Petroleum and Minerals for

support of this research.

REFERENCES

1. Beal, C., The Viscosity of Air, Natural Gas, Crude Oil and its
Associated Gases at Oil Field Temperatures and Pressures,
Trans. , AIME (1946) 165, pp. 94-112.
.. . . ..... . . ..

2. Beggs, H. D. and Robinson, J. R., Estimatingthe Viscosity


of Crude
Oil Systedk, J. Pet. Tech. (Sept. , 1975) pp.
1140-1141.

3. Glaso, 0 ., Generalised Pressure -Volume-Temperature


Correlations, J. Pet. Tech. (May, 1980) pp. 785-795.

4. Chew, J. and Connally, C. A. Jr., A Viscosity Correlation


for Gas-Saturated Crude Oils, Trans. AIME (1959) 216, pp.
23-25.

5. Vazquez, M. E., Correlations for Fluid Physical Property


Prediction, M. S. Thesis, The University of Tulsa (1976).

6. Khan, S. A., Viscosity Correlations for Saudi Crude Oils,


M. S.Thesis, University of Petroleum and Minerals
(1985).

7. Khan, S.A., A1-Marhoun, M.A., Daffua, S., and Abu-


Khamsin, S.A., Viscosity Correlations for Saudi Crude Oils,
paper SPE 15720 presented at the SPE 1987 Middle East
Technical Conference, Bahrain, March 7-10, 1987.

8. Standing, M. B., A Pressure-Volume-Temperature Correlation


for mixtures for California Oils and Gases, Drill. and Prod.
Prac. , API (1947) pp. 275-287.

9. A1-Marhoun, M. A., Pressure-Volume-Temperature Correlations


for Saudi Crude Oils, paper SPE 13718 presented at the SPE
1985 Middle East Technical Conference, Bahrain, March
11-14, 1985.

10. Craft, B. C. and


Hawkins, M. F., Applied Petroleum
Reservoir Prentice-Hall,
Engineering, Inc. , Englewood
Cliffs (1959) pp. 12il.

11. Amyx, J. W., Bass, D. M. and Whiting, R. L., Petroleum


Reservoir Engineering, McCxaw-Hill Book Company, New York
(1960) pp. 377.

12. Montgomery, D. C. and Peck, E. A., Introduction to Linear


Regression Analysis, John Wiley & Sons, New York (1982) pp.
128.

McMillan
.13. Leon, S. J., Linear Algebra with Applications,
Publishing Co., Inc., New York (1980).

14. Walpole, R. E. and Myers, R.H., Probability and Statistics for


Engineersand Scientists, McMillan Publishing Co., Inc. ,
(1978) pp. 179.

SPE 17132

.
Ill
%
,

S1 Conversion Factors

bbl x 1.590E.1 = ~J
Cp x 1.0 rnPaos
ft x 2.832 E-2 = ~J
psi X 6.89S = kPa

.:
APPENDIX

Use of S1 Units in Viscosity Correlations

Correlations in S1 Units

In S1 units, the developed correlations will be as


follows:

= el.4xlo-5(: b)
Pa ~ob~

= ~obl ~pt/pb, )-o.14 =-3.6X10


IJb -s(P-Pb)

where ..
.-::

API = Oil Gravity, API

P = Pressure, kPa
Pb = Bubble Point Pressure, kPa ,.. ;2

Rs = Solution Gas Oil Ratio, mJ/m3


~1
= Temperature, OC :
.:

g Gas Relative Density {air = I) ..


~= .,...> ,,
::... .. .
..

WE 17132 .
l. ,..
, ~
. ...,.
...
...
..*:
,..:-.:
::
: ....~.-.~?:.y,.
:-.:$
:.;
.
..: n..
.,
....
~ :.
i
t-+.A.,
::,.. .
.;,
.:;:.......
.
.$...+..,.,
. .

r. = 141.5/(API + 131.5) = Oil Relative Density

Ma = Oil Viscosity Above Bubble Point, mPa.s

v~ = Oil Viscosity Below Bubble Point, mPa.s

]iob = Bubble Point Oil Viscosity, mPa.s

Or = (T + 273.16)/273.16] = Relative Temperature

Theory of Regression Ana!ysig

The fundamental concept of regressj.on analysis is to

fit a function of independent variables to a given set of

data points in order to estimate or predict one dependent

variable as accurately as possible. If only one independent

variable is involved then it is called simple regression

analysis whereas the name multiple regression analysis is

implied if more than one independent variable is present.

The chosen combination of independent variables is called a

model which can be linear as well as non-linear.

Types of Regression

Linear Multiple Regression

In linear multiple regression, a linear function of

independent variables is chosen as a model. Consider a set

of observations made on properties y,


1 2 .-- n; y
being the dependent variable and xs the n independent

variables. A linear multiple regression model to fii the

SPE17132
. .

given set of observations is expressed as:

Y = a. + alxl a2x2+ - anxn+&


(29)

where as are coefficients to be determined by the

regression analysis and t is a random error component. The

errors or deviations between the observed and predicted

values are assumed to have a zero mean and an unknown

variance. Another assumption is that the errors are


uncorrelated meaning that the value of one error does not

depend on the other error [12].

Non-Linear Multiple Regression

In this type of regression, a non-lir~e,~r function of

independent variables is assumed. Many non-linear functions

can be linearized by using the appropriate transformation.

Consider the following non-linear model:

= al ~oa2u (l+V) a3
Q s wa4w (30)
a.

Taking log on both sides,

log(Q) = log(ao) + allog(s) + a2U + a310g(l+V)

+ a4W log(W) (31)

Let Y = log(Q), = log(ao)


a.
= log(s), =U
1 2
= log(l+V), and = w log(w)
3 4

I
.

then

Y = O + lX1 + 2X2 + a3x3 + 4X4


(32)

Thus the non-linear relationship in Eq. (30) has been

reduced to a linear relationship whose general form is given

by Eq. (29).

Regression Criteria
..

The objective of r(!qression analysis is to find the

values of coefficient as in Equation 1 which give the best

possible correlation. Two criteria to determine the best

fit are used in this study and are described as follows:

Least Square Method

In this criterion, coefficients are determined which

minimize the sum of squares of deviations between the

observed and the predicted values. We now apply it to Eq.

(29).

For any given observation i, Eq.(29) can be written as:

Y1=ao+a1xi:+a2xi2 + .......... +anxin+~i (33)

I WE 17132
where

i = 1,2 8....-.m

The function to be minimized is E, where

E = ?(yi-aO )2 = %i)* (34)


alxil-a2xi2- -anxin
i=l i=l

The minimum occurs when the partial derivatives of E with

respect to ai are zeros i.e.

~E/aai = O (35)

Equation (35) is a set of (n+l) simultaneous linear

equations which can now be solved for a. .


1

A convenient and direct approach to determine the least

square solution is through a linear algebraic technique. The

given set of experimental observations can be represented by

the following matrix system:

1 11 12 ln a. Y1

1 21 22 2n al Y~

1 31 32 3n a2 Y3
. . .,. . . . (36)

. . . . . . . . .

. . . . . . . . .

a
1 lm 2m nm n Ym

SPE 17132
I y

-------- .- .
) --. .. .... -.
... ..... ..
----

or

xa=y (37)
.-.

where x is an m x (n+l) matrix of independent variables

containing the observed values, ~ is (n+l) coefficient


vector and y is the right hand side vector of length m

containing the obseved values of the dependent variable. For

m>n, a unique least square solution exists which is given

as [10]:

a = (XTX)-l XT y (38)
-- --

where ~ is the (n+l) vector containing the least square

solution for Eq. (36). In this study, computer programs were

developed to apply this approach for least square solution.

Least Absolute Method

The other approach to the regression analysis is to

minimize the sum of absolute deviations between the obserled

and the estimated values, [12] i.e., F is minimized, where

m
F = Ilyi - aO - alxil - a2Xi2 - ......- ~nxinl (39)
i=l
or
m
F= I~dil (40)
i=l
where
di=yi-aO -alxil-a2xi2 - ...... -ax
n in

Since an absolute function has to be minimized, the

conventional calculus methods are not applicable as with


least squares. It is because F involves an absolute

function whose derivative does not exist. However, the


problem can be transformed to an equivalent linear
programming problem. A linear programming problem consists

of two parts; (1) the Objective Function\ and (2) the

Constraints . The purpose of linear programming is either

to minimize or maximize a given linear Objective Function

subject to the given linear Constraints. Our problem can be

transformed as follows:

The deviations dl can


- be positive as well as negative.

di are splitted into di and d.~~ by the following relation:


1

di = di - di (42)
where

di 2 0 and d ,120
i

The coefficients a. can as well have positive or


1
negative values. The y are splitted as follows:

a. =a. -a. a. 20, a.20 (43)


1 1 1 1 1

SPE In32

I .:.;..
., .-, .. . ... . ...... ..- . . .. . . .. .
Substituting ai in Eq. (41),

II+a lx
x. +
ao-ao 1 il-a1xil+a2xi2-a2 12

. . . . . . . +a X. -antxin+di-di = Yi (44)
n In

where

i =1,2, ......m

The Objective Function is given as:

minimize ! (di+dif) (45)


i=l

which in vector form can be represented as a product of two

vectors as follows:

minimize C X (46)

where

c= [0000 . ..0 0 1 1 1 1 . ..1 1]

x= [ aoaoalal...an ?anf!d~ td~ !~d


~d2...dmdm]T

Where C is a vector of length 2(m+n) with 2n zeros in the

beginning followed by 2m 1s. Vector X has the same lenqth

and contains coefficient ai and deviations di which are to

be calculated.

Using Eq. (44), the Constraints are represented in the

matrix system as follows:


r

AX=y (47)

where

0 0 . . . 0 0
1 1 11 Xll 12 X12 ln Xln 1 1
0 1-1 ... 0 0
1 -1 21 X21 22 X22 2n x2n 0
. . . . . . . . . . . . . . . . .
A=
. . . . . . . . . . . . .
.............
. . . . .
I
II. . . . . . . . . . . . . i)0 0.:.
. . . . .

1 -1 Xml -Xml xm2 -xm2 . . . Xmn -Xmn o i -i


L

Y =[ Y~ Y2 Y~ . . . Ym IT

Where A is an m x 2(m+n) matrix of independent variables, X

o is the required least absolute solution vector of length

2(m+n) as in Eq. (46), and y is the right hand side vector of

length m containing the dependent variables.

TEST FOR MODEL ADEQUACY

This test provides means to determine how much a model

behaves linearly and is valid only for linear or

linearizable models. The test assumes that if all

regression coefficients are zeroes meaning that there is no

linear relationship among the dependent and the independent

variables, then such model follows an F-distribution [13].

However, if the chosen mode 1 does not follow the F-

SPE I?132
distribution then the hypothesis of regression coefficients

being zeroes, is rejected which implies that the model

behaves linearly. The F-distribution is a ratio between two

distributions and is given as [14]:

vi/2
rml+v2u21(~1/v2) WV12-1
(50)
h(w) =
(vl+v2)/2
r!v@) r(v*/a (l+vlw/2)

where I(a)is the Gamma Function defined as:

I(a)= J us-l e-u du (51)


o

and, and denote the degrees of freedom for


1 2
distribution 1 and 2 respectively.

Now if the model has a linear relationship among the

variables, it will not follow an F-distribution which can be

known by computing I?O value for the given model.


o s
given as

= (ssR/vl)/(ss@J (52)
o

where

m
SSR = X (ie~~-yob~)z (53)
i=l

(54)
i=l

and in our case


sPE 1713?!

,.
(n,
,,,, ,,
s,,:, ?i, , ,,.. ; ;,-.,.:;,:,. :,.
m%!,:. ..!e, . .. . . . 0 ,
1 ?

1 =n (55)

= m-n-1 (56)
2

Hence FO becomes

o = [:(ie~~-yo~~)2/n]/[r(yob~-ye~~)2/(m-n-l )] (57)
i=l i=l

The calculated value of F. is compared with the value

given in standard F-tables [14]. These tables are made at

various percentage points (0.01,0..05 etc.) of che F-

distribution fol- different VI and V2. A value for o.01

percentage point table is denoted as F and


(O.O1,V, ,V5)
indicates that 99% of the area under this part~cu~ar F-

distribution falls towards the left of this value. Thus ,

for a mcdel under consideration, if the calculated value of


is greater than (0.01,v1,V5) then it means that the
o
model falls in the right ha~d side remaining 1% area and

hence does not follow the F-distribution meaning that the

model follows a linear relationship with 99% confidence.

The F. value also gives an indication of goodness of

fit in a magnified form. A perfect fit has an F. value of

infinity. As mentioned above, the lower bound is the value

taken from F-table, below which the model is rejected for

linear relationship,

SPE 1?132
Table 1 - Model Adequacy Test

SSR SSE
I)istri- 1 2 o F(.O1:V1, V2)
bution _ _ (Calculated) (From Tables)

27.71 0.68 4 145 1,470 3.32


ob
Pa 437.43 0.26 1 1,501 2,53~,041 6.63

~b 181.91 8.82 2 ?,688 17,402 4.61

Table 2- Coefficients for Bubble Point Oil Viscosity


I
:
I

Coefficient LS Solution LA Solution Recommended Value

0.087491 0.091537 0.09


a.

-0.317877 -0.336046 -1/3


al
0.556703 0.424521 1/2
a2
-4.789603 -4.476508 -4.5
a3

-3.009681 -2.992938 -3
a4

Table 3 - Coefficients for Oil Viscosity Above Bubble Point

Coefficient LS Solution LA Solution Recommended Value

a 9. 70053 E-5 9.58255E-5 9.6E-5

Table 4 - Coefficients for Oil Viscosity Below Bubble Point

Coefficient LS Solution LA Solution Recommended Value

-0.139469 -0.137972 -0.14


al

-2.59026E-4 -2.55863E-4 -2.5E-4


2

SPE 17132

4
-Dead Oil Viscosity
o.
N
Oil Viscosity
~: Below Bubble Point
i.
z*N Oil Viscosity

P- -
.* At Bubble Point
~;
W d
.+
>= Oil Viscosity Above B&ble Poznt
G.
(

0 40 80 120 160 200 240 280 320 360


Pressure(psia)

Fig, 1 - Viscosity Versus Pressure Plot for Crude Oils

Vous aimerez peut-être aussi