Você está na página 1de 10

Chemical Engineering and Processing 45 (2006) 568–577

Building inferential estimators for modeling product quality


in a crude oil desalting and dehydration process
S. Abdul-Wahab a,∗ , A. Elkamel b , C.R. Madhuranthakam b , M.B. Al-Otaibi c
a Mechanical & Industrial Engineering Department, Sultan Qaboos University, P.O. Box 33, Al Khoud, P.C. 123, Muscat, Oman
b Department of Chemical Engineering, University of Waterloo, 200 University Avenue West, Waterloo, Ont., Canada N2L 3 G1
c Kuwait Oil Company, Safat, Kuwait

Received 8 July 2005; received in revised form 7 January 2006; accepted 9 January 2006
Available online 23 February 2006

Abstract
Desalting/dehydration plants (DDP) are often installed in crude oil production units in order to remove water-soluble salts from an oil stream.
This paper describes the development of simple inferential estimators for product quality of the desalting/dehydration process. The inferential
estimators were constructed to capture the relationship between the product quality of the plant and the process input variables. Five input process
variables that are known to influence product quality were considered. These include temperature, settling time, mixing time, chemical dosage,
and dilution rate. The product quality of the desalting/dehydration process was identified by the salt removal and water cut efficiencies. Hence,
inferential estimators were used to infer the salt removal and water cut efficiencies from the five input process variables. These inferential estimators
were constructed based on the application of both multiple linear and principal component analysis as well as non-linear regression. The results
indicate that the settling time and dilution water were the common variables in estimating both the salt removal and water cut efficiencies. On
the other hand, temperature contributed insignificantly in predicting the two efficiencies. Furthermore, the inferential model predictions were
compared with the experimental readings. It was found that the actual dependence of the performance of the desalting/dehydration process on
process parameters could not be described only by linear relationships. Addressing the non-linearity of the process variables overcame the problem
of inaccurate predictions. Future studies based on the use of computational intelligence techniques and design of experiments to get better models
are suggested as well as the use of response surface methodologies to determine the set of parameters that will optimize the process efficiencies.
© 2006 Elsevier B.V. All rights reserved.

Keywords: Inferential estimators; Product quality; Principal component analysis; Desalting/dehydration process

1. Introduction associated with a process and these are easily measured on-line.
Due to the nature of chemical and process engineering systems,
A major problem in process industries is to measure variables the states of many of the secondary variables reflect the states of
that define the product quality. These variables, called primary primary variables [1]. This means that changes in some of the
or quality variables, quantify the productivity or the specifica- secondary variables are indicative of changes in product quality.
tions upon which the product is sold (e.g., purity, physical or Thus, by monitoring suitable secondary variables, it is often pos-
chemical properties). They are often the ones that are difficult sible to “infer” the state of the quality variables. The technique
to measure on-line. Inferential estimations are designed to over- uses easily obtainable measurements to generate estimates of
come measurement problems of the primary variables and so product quality. It has also been called “sensor-data fusion” and
that they can be used to enhance process operation and pro- “soft-sensing” [1].
ductivity. They allow process quality to be inferred from other In general, inferential estimations are often used in process
easily made plant measurements. There are usually other sec- industries in place of direct on-line measurement of controlled
ondary variables such as temperature, pressure, and flow that are variables, whenever direct measurements are expensive, unre-
liable or add significant delays [2]. Additionally, they are key
technologies for producing high quality products when on-line
∗ Corresponding author. Tel.: +968 24141360; fax: +968 24413416. analyzers of product quality are not available [3]. They can be
E-mail address: sabah1@squ.edu.om (S. Abdul-Wahab). used to obtain an estimate of product quality, using a model of

0255-2701/$ – see front matter © 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.cep.2006.01.004
S. Abdul-Wahab et al. / Chemical Engineering and Processing 45 (2006) 568–577 569

the process and some information retained by secondary process ential estimators were constructed by using stepwise multiple
variables, which are more easily and less expensively measured regression, principal component analysis (PCA), and non-linear
on-line, and with a negligible time delay [4]. models. All parameters were applied to a sample of actual
In building inferential estimators, therefore, the objective is Kuwaiti crude oil.
to model the relationship between the primary and secondary
variables. The inferential model can then be used to produce 2. Desalting/dehydration plant operation
estimates of the primary variables at the frequency at which
the easily measured inputs are made [1]. Many techniques can The process under consideration was a desalting/dehydration
be used to tackle the relationships between the primary and process plant. The aim of the plant was to remove water-soluble
secondary variables. The simplest approach to build a relation- salts and entrained water. Water normally contains chlorides of
ship between variables is to carry out multiple linear regression sodium, calcium, and magnesium. When designing a desalter,
(MLR) using the least-squares method. While the method is its type and size are all dependent on a number of opera-
straightforward, the results can be affected by a number of data tional factors such as required pressure, temperature, viscosity
issues [5]. There are also other techniques that range from simple and flow rate, as well as user specifications related to maxi-
graphical techniques (scatter plot, box-plot), multivariate statis- mum salt (PTB) allowed in the product oil stream. A typical
tical methods (e.g., principal component analysis, correlation desalting/dehydration plant operation is comprised of six major
analysis, cluster analysis) to control theoretic, relative gain, and steps: separation by gravity settling, chemical injection, heat-
singular value analyses [1,6]. Other approaches are the use of ing, addition of fresh (less salty) water, mixing, and electrical
expert system methods to detect sensor failures and sensor drifts coalescing.
[7–9], fuzzy logic to interpret sensor state [10,11], and neural Fig. 1 is a process flow diagram of a typical desalt-
nets to construct a non-linear mapping between sensor measure- ing/dehydration plant, which shows the six major steps and the
ments and quality variables [12,13]. main equipment [14,15]. At point no. 1, an emulsion comprised
Oil desalting/dehydration systems are industrial processes for of water and oil flows into a wet tank. Such an emulsion may
removing water-soluble salts from an oil stream. The primary contain up to 25% water cut. As per design, a typical desalt-
objective for an oil desalting/dehydration process is to achieve ing/dehydration plant would meet acceptable crude oil specifi-
sufficient product purity. This can be determined in terms of cations, that is water and salt of the crude must be reduced to
two primary variables: salt removal and water cut efficiencies. 0.10 vol.% and 5.0 PTB, respectively [15]. In order to remove
Producing oil with good product quality is important for the such large quantities of water from the oil stream a two-stage
following reasons: to decrease the flow of salt content to refin- desalting system is used. At point no. 2, the emulsion leaves the
ery distillation feed-stocks, to minimize the energy required for wet tank, where the primary water separation takes place. At this
pumping and transportation, and to reduce scale accumulation, point, a chemical demulsifier is injected into the stream. After
corrosion and lowering of activity of catalyst. Therefore, con- settling, formation water, stream 13, flows to the waste water
structing inferential estimators for modeling a crude oil desalt- treatment plant or is sent to a designated disposal pit.
ing/dehydration process is essential to tackling the relationship Point no. 3 shows emulsion flow from the wet tank to a
between the variations in the processing variables and the final heat exchanger, where heat is recovered from the treated crude
product properties. Knowing these relationships will help in product stream (stream no. 10). The emulsion then flows to
quickly obtaining information on the performance and product a water-bath indirect heater, raising its temperature (point no.
quality of the process so that deviations from normal behavior 4). Water recycled from the 2nd stage vessel (stream no. 5) is
can be detected. injected into the emulsion flow coming out of the heater. This
In this paper, the performance of the desalting/dehydration system, recycling water from the 2nd stage back to the 1st stage,
process system was evaluated by conducting a series of exper- aims at minimizing fresh water consumption where a counter
imental runs with various process conditions. The work was current flow is employed. Hence, the dispersed brine in the crude
undertaken with the objective of building inferential estima- is contacted with fresh water streams each time. At the mixing
tors for modeling the desalting/dehydration process system. The valve (stream no. 6), recycled water and emulsion are agitated
inferential estimators were developed to estimate the product by an induced shearing force. The operation of a mixing valve is
quality of the desalting/dehydration process with five secondary carried out by a simple globe valve where an operator would set
process parameters as predictor variables. The study focused the differential pressure across the valve to be as high as possi-
specifically on identification of the process parameters that affect ble ensuring better mixing of the two fluids. Stream no. 7 leaves
final product quality. The five process parameters that were the mixing valve to enter the 1st stage desalter vessel. Inside
selected are the chemical dosage (ppm), crude temperature (◦ C), the 1st stage-vessel, an emulsion is exposed to a high voltage
wash water flow rate ratio (%), mixing time (min), and settling electrostatic field. The action of the electrostatic field coalesces
time (min). The product quality of the desalting/dehydration the dispersed water phase and gravity causes the enlarged water
process was identified by salt removal and water cut efficiencies droplets to fall and collect in the bottom of the vessel. Effluent
η1 , and η2 , respectively. The study was conducted under various water from the 1st stage, stream no. 11, leaves the system to
experimental conditions to develop inferential estimators that a wastewater treatment plant or the disposal pit. This effluent
can be used to generate η1 and η2 of the desalting/dehydration water contains various impurities and salts removed from the
process from the five secondary process parameters. The infer- water-in-oil emulsion.
570 S. Abdul-Wahab et al. / Chemical Engineering and Processing 45 (2006) 568–577

Fig. 1. A typical desalting/dehydration plant [14]. 1: wet crude flow to wet tank; 2: demulsifier/chemical injection; 3: crude flow to heat exchanger; 4: flow to heater;
5: wash water recycled from 2nd stage vessel; 6: flow to 1st stage desalting mixing valve; 7: mixed fluid to 1st stage vessel; 8: flow to 2nd stage desalting mixing
valve; 9: fresh water from water-water heat exchanger originated from wash water tank; 10: treated crude flow; 11: effluent water from 1st stage desalter vessel to
water treatment plant and/or disposal pit; 12: BS&W Analyser. A signal to diverting valve; 13: formation water settled down at the bottom of wet tank, to water
treatment plant and/or disposal pit.

Treatment of an emulsion is further enhanced in the 2nd- Kuwait Oil Company. The characteristics of the used fresh water
stage-desalting vessel. Stream no. 8 flows to a mixing valve can be found elsewhere [16,17]. The chemical used as a demul-
on the entrance of the 2nd stage vessel. Still containing salt sifier in the experiment was under the trade name Servo CC 3408
water, the emulsion is mixed with fresh water (stream no. 9). supplied by Servo Delden BV (Netherlands). In addition, cen-
The differential pressure across the mixing valve is normally trifuge tubes, stoppers, 100 ml graduated-cylinders, micro ml
around 15 psi. Then, partially treated emulsion is introduced syringe, stopwatch, and gloves and chemical-safety gear were
near the bottom of the 2nd stage and, once more, travels upward employed. Details of the laboratory’s instruments are given in
through electrical voltage grids. In this stage, water droplets Al-Otaibi [16] and Al-Otaibi et al. [17].
are enlarged by means of a high voltage electrostatic field and
separated by gravity. The separated water is collected at the bot-
tom of the vessel and recycled to the first stage desalter (stream Table 1
5), while the treated crude flows from the top of the vessel Characteristics of crude oil samples
(stream no. 10). The latter stream (treated) continues to pass Property Value
through a BS&W analyzer (stream no.12). If the treated crude
Specific gravity (60◦ /60◦ ) 0.864
is within specifications, a signal is sent to the diverting valve
to open the dry tank, otherwise the flow is directed back to the Reid vapour pressure (Psia) 10.5
wet tank. Pour point (F) Less than −30.0
Average API gravity at 60 F 31.7
Viscosity (Cs) at
3. Materials and methods 70 F 17.4
100 F 10.5
3.1. Materials and instruments 130 F 6.79
160 F 4.8
Crude oil was supplied by Kuwait Oil Company (KOC). The Average sulphur content (wt.%) 2.7
characteristics of this oil are illustrated in Table 1. Dilution water Asphaltenes (wt.%) 2.23
used in the experiments was collected from field operations of
S. Abdul-Wahab et al. / Chemical Engineering and Processing 45 (2006) 568–577 571

is what happens in real operation processes. The order, shown


in Fig. 2, in which the factors were varied, was followed in
order to mimic the real process. In such a process, an emulsion
that is introduced into the system is subjected to fresh water
injection followed by chemical dosage. The mixture, emulsion,
fresh water, and chemical is then heated to a certain temperature
and then mixed. The blend, at the final stage, enters a vessel
where settling takes place for a specific time period allowing
water and salt to be drained off. The process, at the final stage,
produces dry or treated crude oil samples, which were tested and
analyzed for salt and water cut.
In each cycle of the experiment, a sample of crude oil to
be tested was taken in a sample tube or graduated cylinder of
about 100 ml. Then, according to the previously set of ranges,
fresh water and chemical demulsifier were added. Crude oil,
fresh water, and chemical were heated and mixed for a specific
time (min). The mixture was then taken to a centrifuge where it
was rotated for settling. From the top of the centrifuge tube, a
volume of dry crude was withdrawn by a micro ml syringe and
transferred to a test beaker. The S/R test was conducted on a
partial volume of that dry crude, about 10 ml. Then, 50 ml was
transferred to a centrifuge for the W/C test.

3.3. Methodology and data collection

In a desalting/dehydration process, there are several param-


eters that can be altered in order to reach an optimum combina-
tion of operating conditions. In this study, five parameters that
were expected to affect the desalting/dehydration process were
investigated. To predict the interactions and the optimum com-
bination of these parameters, different variables were allowed
to vary according to a pre-specified design of experiment. The
variables that were considered in this work included crude tem-
perature (◦ C), mixing time (min), settling time (min), chemi-
cal/demulsifier dosage (ppm), and the amount of fresh water
added in ratio to that of the wet crude’s quantity (%). The desalt-
ing/dehydration process was evaluated by conducting a series
of runs. The experimental design was constructed to include all
possible combinations of different values of these parameters.
Table 2 shows the values covered for each parameter. It can be
seen that there were 980 combinations or runs (2 × 2 × 5 × 7 × 7
Fig. 2. Steps followed in carrying out a single experiment. runs) that were carried out to determine the interactions of all of
these parameters.
3.2. Experimental set-up and procedures Temperature and settling time were the factors that were
found to be the least varied in real processes. Further, it is widely
In carrying out the experiments, samples were first analyzed known how temperature and settling time affect oil viscosity
and tested for salt (S/R) in pounds per thousand barrels (PTB) and the rate of downward settling [16,17]. Accordingly, chemi-
and water cut (W/C) in vol.%. The details of these tests, which cal dosage, fresh water addition, and the mixing time variables
were conducted as per KOC standards, were presented elsewhere were tested in various varied values. On the other hand, tem-
[16,17]. Fig. 2 explains the steps followed in carrying out one perature and settling time were investigated only at high and
cycle of the experiment. First, fresh water is added, followed by low values. Hence, the five parameters were classified into two
the addition of chemical/demulsifier. The mixture is then heated groups: two-point and multi-point variation groups. The two-
in a water bath heater. The heated mixture was then mixed and point group consisted mainly of temperature and settling time.
poured into a 100-ml centrifuge tube and rotated at 1000 rpm. The multi-point group, on the other hand, consisted of chemical
The final step in completing one cycle was to suck out the top dosage, fresh water addition, and the mixing time. It is worth
crude volume in the centrifuge tube and test it for salt (S/R) noting here that the values of settling time and temperature used
and water cut (W/C). The top volume was taken because that in this study were selected from real field experience [14,15].
572 S. Abdul-Wahab et al. / Chemical Engineering and Processing 45 (2006) 568–577

Table 2
Description, nomenclature and values of process parameters
Parameter Description Nomenclature Values Number of runs

Two-point variations group


Temperature (◦ C) Temperature of the outlet crude X1 55 ◦ C (low), 70 ◦ C (high) 2
Settling time (min) Settling time X2 1 min (low), 3 min (high) 2
Multi-point variations group
Mixing time (min) Mixing time X3 1, 3, 5, 7 and 9 5
Chemical dosage (ppm) Chemical addition X4 1, 2, 5, 8, 10, 12 and 15 7
Dilution water (%) Fresh water addition X5 1, 2, 3, 4, 6, 8 and 10 7
Total number of runs 980

The parameters of the multi-point variations group were care- and the fresh water addition (0.179). Moreover, the salt removal
fully selected to have values that were practically encountered efficiency was negatively but weakly correlated with the tem-
in real desalting/dehydration process systems. perature of the crude oil (−0.050). Additionally, the correlation
The product quality of the desalting/dehydration process was coefficient of the salt removal efficiency with the chemical addi-
evaluated by determining the salinity and water cut efficiencies tion variable was generally poor and insignificant (0.054).
(η1 , η2 ). These efficiencies were obtained from the collected On the other hand, the correlation coefficients between the
experimental data [18] and are expected to depend on the mixing water cut efficiency and the five process variables were mostly
rate, the wash water rate, the demulsifying chemical dosage, and positive, except for the mixing time (−0.132). The largest in
the rate of heating. The salinity efficiency (η1 ) was calculated positive magnitude were those with the settling time (0.424), fol-
from Eq. (1), whereas water cut efficiency (η2 ) was calculated lowed by fresh water addition (0.259), chemical addition (0.239)
from Eq. (2). and temperature of the crude oil (0.137).
Zout
η1 = 1 − (1) 4.2. Inferential estimator development
Zin
Xout This section presents the development of inferential estima-
η2 = 1 − (2) tors for the determination of the salt removal and water cut
Xin
efficiencies of the desalting/dehydration process as a function
where Zout denotes outlet salt result (PTB, pound of salt per of the secondary input process variables.
thousand barrels of oil); Zin denotes inlet salt result (PTB); Xout
denotes outlet water cut (%); Xin denotes inlet water cut (%). 4.2.1. Linear estimators
The calculations of the salinity and water cut efficiencies As a first approximation, multiple linear regressions were car-
at different experimental conditions were evaluated in order to ried out in order to find inferential estimators for the salt removal
determine the effect of the various parameters on the perfor- and water cut efficiencies with the input process variables as the
mance of the desalting/dehydration process. predictor variables. Two separate regression analyses were per-
formed, the first one for the salt removal efficiency (η1 ) and
4. Results and discussions the second one for the water cut efficiency (η2 ). The number of
process secondary variables that were introduced in each anal-
4.1. Data analysis ysis was five inputs. These include temperature (X1 ), settling
time (X2 ), mixing time (X3 ), chemical concentration (X4 ), and
Table 3 shows the Pearson correlation matrix of the secondary
and primary variables included in this study. This matrix helps Table 3
in deciding whether a significant relationship exists between the Pearson correlation matrix of variables
secondary and primary variables. The last two columns in this Inputs Outputs
table indicate the tendency of the salt removal and water cut effi-
X1 X2 X3 X4 X5 η1 η2
ciencies (primary or quality variables) to change with a change
in various input process parameters (secondary variables). Sta- X1 1 0.000 0.000 0.000 0.000 −0.050 0.137
tistically significant correlation coefficients (p < 0.001) are high- X2 – 1 0.000 0.000 0.000 0.695 0.424
X3 – – 1 0.000 0.000 0.455 −0.132
lighted in bold. As indicated in this table, the salt removal
X4 – – – 1 0.000 0.054 0.239
efficiency was positively correlated with the settling time, mix- X5 – – – – 1 0.179 0.259
ing time and fresh water addition. This result indicated that η1 – – – – – 1 0.431
an increase in salt removal efficiency was associated with an η2 – – – – – – 1
increase in the values of settling time, mixing time, and fresh X1 = temperature, X2 = settling time, X3 = mixing time, X4 = chemical dosage,
water addition. The most significant correlations were those with X5 = dilution water, Y1 = salt removal efficiency (η1 ), Y2 = water cut removal
the settling time (0.695), followed by the mixing time (0.455) efficiency (η2 ).
S. Abdul-Wahab et al. / Chemical Engineering and Processing 45 (2006) 568–577 573

Table 4 The R2 increased markedly when the mixing time (R2 = 0.689)
Inferential estimators for salt removal efficiency (η1 ) using five independent
process variables
and dilution water (R2 = 0.721) variables were added. Table 4
shows that when the three variables (settling time, mixing time
Predictors Constant X2 X3 X5 X4 X1 and the dilution rate) are fitted to the salt removal efficiency
Adjusted R2 0.482 0.689 0.721 0.724 0.726 data, the value of the adjusted R2 becomes 0.721. This means
Change in R2 0.482 0.207 0.032 0.003 0.003 that 72.1% of the variations in salt removal efficiency data can
Estimated regression 50.277 8.040 1.863 0.679 0.129 −0.077 be explained by the variations of these three variables. Adding
coefficient
Standard error 1.760 0.194 0.068 0.064 0.040 0.026
more variables (e.g., chemical dosage and temperature) resulted
in an insignificant further increase in R2 .
X1 = temperature, X2 = settling time, X3 = mixing time, X4 = chemical dosage, The adjusted R2 appearing in Table 5 indicates that the
X5 = dilution water, Y1 = salt removal efficiency (η1 ), Y2 = water cut removal
efficiency (η2 ).
inferential estimator of the water cut efficiency is more highly
correlated with the settling time, dilution, and chemical dosage
variables than with mixing time and temperature. The statisti-
Table 5 cal evaluation of the inferential estimator using only the settling
Inferential estimators for water cut efficiency (η2 ) using five independent process
variables
time produced an R2 of 0.179. When the dilution water vari-
able was added, R2 increased from 0.179 to 0.245. A further
Predictors Constant X2 X5 X4 X1 X3 increase was obtained when the chemical dosage variable was
Adjusted R2 0.179 0.245 0.302 0.320 0.337 added. When the three variables (settling time, dilution rate, and
Change in R2 0.179 0.067 0.057 0.019 0.017 chemical dosage) were fitted to the water cut efficiency data, the
Estimated regression 18.111 6.793 1.365 0.798 0.293 −0.747 value of the adjusted R2 was 0.302. This means that only 30.2%
coefficient
Standard error 3.793 0.417 0.137 0.087 0.056 0.148
of the variations in water cut efficiency data can be explained
by the variations of these three secondary variables. Adding yet
X1 = temperature, X2 = settling time, X3 = mixing time, X4 = chemical dosage, more variables (e.g., mixing time, and temperature) led only to
X5 = dilution water, Y1 = salt removal efficiency (η1 ), Y2 = water cut removal
efficiency (η2 ).
insignificant further increases in R2 .
Principal component analysis was used for filtering the
data so that only the significant secondary process variables,
the dilution wash water (X5 ). Both the salt removal and water responsible for determining the product quality of the desalt-
cut efficiencies were the outputs of the inferential models. The ing/dehydration process system, could be determined.
inferential estimators were developed by using a stepwise mul- The five-predictor variables (i.e., secondary variables) were
tiple regression modeling procedure. The procedure is used to first transformed into an equal number of principal components.
automatically select the independent input parameters that are After transformation, varimax rotation was used to maximize the
of most importance and eliminates those that are of least impor- loading of a predictor variable on one component. In general,
tance. Tables 4 and 5 present the results of the analysis for the salt application of PCA procedures followed by a varimax rotation
removal and water cut efficiencies, respectively. The coefficient produces a ranked series of factors. Table 6 summarizes the
of multiple determination, R2 , gives the proportion of the vari- results of the varimax rotation on the five principal components
ation in the efficiencies explained by the independent variables together with the amount of variance explained by each compo-
in the developed inferential estimators. It measures whether a nent. The higher the loading of a variable, the more that variable
line is a good fit or undetermined. It may vary from the value of contributes to the variation accounted for by the particular prin-
0.0% to a perfect linear fit of 100%. cipal component. In practice, only loadings with absolute values
The adjusted R2 appearing in Table 4 indicates that the infer- greater than 50% are selected for the principal component inter-
ential estimator of the salt removal efficiency is more highly pretation [19,20]. A principal component with an eigenvalue
correlated with the settling time, mixing time and the dilution greater than or equal to one, is usually considered as being of
rate. The statistical evaluation of the salt removal model using statistical significance. From Table 6, it can be seen that the first
only the settling time process variable yielded an R2 of 0.482. PC accounted for 20% of the total variation in the data. It is

Table 6
Rotated principal component loadings
PC1 PC2 PC3 PC4 PC5

X1 −7.24E−17 −3.18E−143 1.466E−16 0.000 1.000


X2 −3.35E−18 −5.93E−19 1.000 0.000 1.466E−16
X3 1.000 −7.46E−17 −3.35E−18 0.000 7.241E−17
X4 −7.46E−17 1.000 −5.93E−19 0.000 3.184E−13
X5 0.000 0.000 0.000 1.000 0.000
Eigenvalue 1.00 1.00 1.00 1.00 1.00
% of Variance 20.00 20.00 20.00 20.00 20.00
Cumulative % 20.00 40.00 60.00 80.00 100.00

X1 = temperature, X2 = settling time, X3 = mixing time, X4 = chemical dosage, X5 = dilution water.


574 S. Abdul-Wahab et al. / Chemical Engineering and Processing 45 (2006) 568–577

Table 7
Inferential estimators for salt removal efficiency (η1 ) using principal components
Predictors Constant PC3 PC1 PC4 PC2 PC5

Adjusted R2 0.482 0.689 0.721 0.724 0.726


Estimated regression 75.12 8.045 5.271 2.069 0.622 −0.579
coefficient

loaded heavily on the mixing time variable (X3 ) with no sig-


nificant contributions from the other remaining input variables.
Principal components two, three, four and five loaded heavily
on chemical dosage (X4 ), settling time (X2 ), dilution water (X5 ),
and temperature (X1 ), respectively.
Principal component scores were then used as independent
variables in a stepwise regression analysis to determine the orig-
inal independent process variables that are significant to the
variation of the quality variables. The results of the analysis Fig. 3. Predicted salt removal efficiency vs. experimental salt removal efficiency.
are summarized in Tables 7 and 8 for salt removal and water cut
efficiencies, respectively.
For the salt removal efficiency data (Table 7), using the first dictor variables in the regression analysis lead to the following
three principal components (PC3, PC1, and PC4) yielded an inferential estimator:
adjusted R2 of 0.721. Inclusion of additional components did
not significantly improve the coefficient of determination (new η2 = 18.111 + 6.793X2 + 1.365X5 + 0.798X4 (4)
adjusted R2 of 0.726). Thus, application of principal compo-
nent regression analysis yielded three components for the salt The efficiencies (η1 , η2 ) estimated by Eqs. (3) and (4) were plot-
removal efficiency. By comparing Tables 6 and 7, the method ted against experimental values and the results are presented
identified settling time (X2 ) from PC3, mixing time (X3 ) from in Figs. 3 and 4, respectively. It can be seen from Fig. 3 that
PC1, and dilution water (X5 ) from PC4. Therefore, the results of the predicted and experimental values of salt removal efficiency
both multiple linear regression and principal component regres- were clustered around the diagonal that entails the validity of the
sion techniques affirmed that the settling time, mixing time derived predictive salt removal model. The adjusted R2 value was
and dilution water variables were the most important factors in 0.721. Thus, there was a significant correlation between the pre-
estimating the salt removal efficiency. Using these variables as dicted and the experimental values of salt removal efficiency. In
predictor variables in the regression analysis, the derived infer- contrast, the points of Fig. 4 tended not to be clustered around the
ential estimator for the salt removal efficiency was obtained as: diagonal, an indication of model inadequacy (R2 = 0.302). Thus,
the estimation of water cut efficiency was not consistent with
η1 = 50.277 + 8.040X2 + 1.863X3 + 0.679X5 (3) the corresponding experimental values and the derived inferen-
tial estimator was not able to produce reasonably accurate water
For the water cut efficiency data (Table 8), using the first three
cut efficiency predictions. This leads us to consider non-linear
principal components (PC3, PC4, and PC2) yielded an adjusted
models.
R2 of 0.302. Inclusion of additional components did not signifi-
cantly improve the coefficient of determination (new adjusted R2
of 0.337). Thus, application of principal component regression
analysis yielded three components for the water cut efficiency.
By comparing Tables 6 and 8, the method identified settling
time (X2 ) from PC3, dilution water (X5 ) from PC4, and chemi-
cal dosage (X4 ) from PC2. Therefore, the results of both multiple
linear regression and principal component regression techniques
affirmed that the settling time, dilution water and chemical
dosage variables were the most important process factors in
predicting the water cut efficiency. Using these variables as pre-

Table 8
Inferential estimators for water cut efficiency (η2 ) using principal components
Predictors Constant PC3 PC4 PC2 PC5 PC1

Adjusted R2 0.179 0.245 0.302 0.320 0.337


Estimated regression 58.964 6.797 4.158 3.837 2.201 −2.115
coefficient
Fig. 4. Predicted water cut efficiency vs. experimental water cut efficiency.
S. Abdul-Wahab et al. / Chemical Engineering and Processing 45 (2006) 568–577 575

Table 9
Values of the parameters used for calculating η1 from Eq. (5)
Parameter Value

β0 75.28
β1 −2.99
β2 7.66
β3 13.81
β4 −0.25
β5 3.52
β12 0.98
β13 −3.02
β14 −0.47
β15 0.08
β23 −2.01
β24 −0.17
β25 −0.89
β34 −0.40
β35 −0.54
β45 −0.84
β1212 −0.41 Fig. 5. Comparison of the salt removal efficiency obtained from experiments
β1313 −5.16 and the non-linear model.
β1414 1.37
β1515 −0.14
β2323 −1.07
A similar analysis for the water cut efficiency generated a
β2424 −0.30 non-linear model with a maximum R2 of 0.64. Many different
β2525 −0.18 trials were used in an attempt to increase R2 . It was found that
β3434 0.13 splitting the data of the variable X3 (mixing time) into two groups
β3535 0.29 improved the R2 . Hence, the mixing time data was split into
β123 4.24

Table 10
Values of the parameters used for calculating η2 from Eq. (6)
4.2.2. Non-linear models
In order to capture the non-linear relationships and inter- Parameter Value when X3 = 1–5 Value when X3 = 7–9
actions of the various parameters involved in the desalt- β0 70.91 70.33
ing/dehydration system, the cubic mixture model (also known β1 6.29 0.39
as Scheffe’s model) was used. More details on mixture models β2 5.82 −2.99
can be found in Montgomery [21]. The cubic mixture models β3 14.65 −16.14
β4 6.23 15.31
were used in order to find out the non-linear models for the salt β5 14.51 13.52
removal and water cut efficiencies with the input variables as β12 −1.53 −4.17
the predictor variables. Eq. (5) was found to give the best model β13 8.99 −2.55
for the salt removal efficiency, with an R2 value of 0.88. Table 9 β14 −3.62 −4.42
shows the values of the parameters used in Eq. (5). β15 −3.82 0.434
β23 −2.41 4.84
β24 2.62 1.99
η1 = β0 + β1 X1 + β2 X2 + β3 X3 + β4 X4 + β5 X5 β25 2.42 0.50
+ β12 X1 X2 + β13 X1 X3 + β14 X1 X4 + β15 X1 X5 β34 −4.77 −22.46
β35 −9.90 −9.15
+ β23 X2 X3 + β24 X2 X4 + β25 X2 X5 + β34 X3 X4 β45 −2.57 −0.89
β1212 3.14 0.23
+ β35 X3 X5 + β45 X4 X5 + β1212 X1 X2 (X1 − X2 ) β1313 −9.04 2.14
β1414 0.89 −0.16
+ β1313 X1 X3 (X1 − X3 ) + β1414 X1 X4 (X1 − X4 ) β1515 −5.63 −3.64
β2323 3.81 0.60
+ β1515 X1 X5 (X1 − X5 ) + β2323 X2 X3 (X2 − X3 ) β2424 −2.85 −0.18
+ β2424 X2 X4 (X2 − X4 ) + β2525 X2 X5 (X2 − X5 ) β2525 −0.29 −3.06
β3434 −5.03 10.50
+β3434 X3 X4 (X3 − X4 ) + β3535 X3 X5 (X3 − X5 ) β3535 −9.58 8.53
β123 3.82 9.07
+ β123 X1 X2 X3 (5) β1234 6.43 0.09
β1235 4.41 −2.58
A cross plot for this model, showing the predicted versus the β2345 2.00 −0.88
actual experimental values of the salt removal efficiency, is given β3451 −2.02 −0.16
β4512 2.72 1.01
in Fig. 5. Comparison of Figs. 3 and 5 clearly illustrates the β12345 1.45 −0.45
superiority of the non-linear model.
576 S. Abdul-Wahab et al. / Chemical Engineering and Processing 45 (2006) 568–577

two groups. The first group contains the data where the mixing
time was in the range of 1–5 min whereas the second group
contains the data for which the mixing time was in the range of
7–9 min. Accordingly, a modified special cubic equation with
four and five interaction variables was used to model the water
cut efficiency. It should be noted that higher order terms are
frequently necessary in these kind of mixture models because of
complex phenomena and the experimental region is frequently
the entire operability region and is therefore large, requiring
an elaborate model [21]. The results indicated that Eq. (6) was
found to give the best model for the water cut efficiency, with
an R2 value of 0.82 for X3 ranged between 1 and 5 min, and
R2 = 0.78 for X3 varied between 7 and 9 min. Table 10 shows the
values of the parameters used in Eq. (6) for both X3 = 1–5 and
X3 = 7–9 min.

η2 = β0 + β1 X1 + β2 X2 + β3 X3 + β4 X4 + β5 X5 + β12 X1 X2
+ β13 X1 X3 + β14 X1 X4 + β15 X1 X5 + β23 X2 X3 Fig. 7. Comparison of the water cut efficiency obtained from experiments and
the non-linear model for X3 = 7–9 min.
+ β24 X2 X4 + β25 X2 X5 + β34 X3 X4 + β35 X3 X5
+ β45 X4 X5 + β1212 X1 X2 (X1 − X2 ) + β1313 X1 X3
5. Conclusions
× (X1 − X3 ) + β1414 X1 X4 (X1 − X4 ) + β1515 X1 X5
× (X1 − X5 ) + β2323 X2 X3 (X2 − X3 ) + β2424 X2 X4 The aim of this study was to develop interferential estimators
for the salt removal and water cut efficiencies of the desalt-
× (X2 − X4 ) + β2525 X2 X5 (X2 − X5 ) + β3434 X3 X4 ing/dehydration process (product quality) in terms of five pro-
× (X3 − X4 ) + β3535 X3 X5 (X3 − X5 ) + β123 X1 X2 X3 cess secondary parameters as predictor variables (input). These
include temperature (X1 ), settling time (X2 ), mixing time (X3 ),
+ β1234 X1 X2 X3 X4 + β1235 X1 X2 X3 X5 chemical dosage (X4 ), and dilution water rate (X5 ). The primary
+ β2345 X2 X3 X4 X5 + β3451 X3 X4 X5 X1 or quality variables were the salinity and water cut efficiencies
(η1 , η2 ).
+ β4512 X4 X5 X1 X2 + β12345 X1 X2 X3 X4 X5 (6) The linear inferential estimator for salt removal, which con-
tains three process terms (settling time, mixing time, and the
Figs. 6 and 7 present cross plots of the predicted and actual exper-
dilution water), was selected as the best estimator for salt
imental data of the water cut efficiency for X3 ranging from 1 to
removal efficiency. When these three secondary variables were
5 min and X3 ranging from 7 to 9 min, respectively. Comparison
fitted to the salt removal efficiency data, the value of the adjusted
of Figs. 4, 6 and 7 clearly demonstrated the superiority of the
R2 was calculated to be 0.721. On the other hand, the linear infer-
non-linear model.
ential estimator for water cut efficiency, which contains three
secondary process terms (settling time, dilution water, and chem-
ical dosage), was selected as the best estimator for predicting
water cut efficiency. When these three secondary variables were
fitted to the water cut efficiency data, the value of the adjusted
R2 was calculated to be 0.302. Inclusion of additional predictors
did not significantly improve the coefficient of determination for
both the salt removal and the water cut efficiencies.
The study also showed that principal component regression
(PCR) combined with varimax rotation, was a good variable
selection technique for identifying the most appropriate subset
of process variables to include in estimating the salt removal
and water cut efficiencies. A variable selection method based
on principal component regression analysis was applied that
yielded three process variables for the salt removal efficiency
and three for the water cut efficiency. For the salt removal effi-
ciency, the PCA method identified settling time, mixing time,
and dilution water from the five secondary process variables.
Fig. 6. Comparison of the water cut efficiency obtained from experiments and For the water cut efficiency data, the three variables of set-
the non-linear model for X3 = 1–5 min. tling time, dilution water and chemical dosage were selected.
S. Abdul-Wahab et al. / Chemical Engineering and Processing 45 (2006) 568–577 577

These variables were then used as predictor variables in two [4] E. Zamprogna, Process Monitoring and Control using Artificial Neural
separate multiple regression analyses to obtain the final infer- Networks and Other Advanced Techniques, University of Padova, Italy,
2005.
ential estimators for the salt removal and water cut efficiencies.
[5] S. Albert, H. Hiden, A. Conlin, E.B. Martin, G.A. Montague, A.J. Mor-
The adjusted coefficients of determination for the salt removal ris, Inferential quality assessment in breakfast cereal production, J. Food
and water cut efficiencies models were 0.721 and 0.302, respec- Eng. 50 (2001) 157–166.
tively. Therefore, the results of the PCA are in agreement with [6] M.J. Piovoso, K.A. Kosanovich, J.P. Yuk, Process data chemometrics,
those of the regression analysis. IEEE Trans. Instrum. Measurement 41 (1992) 262–268.
In addition, the existence of the non-linear relationships [7] S.A. Abdul-Wahab, C. Bakheit, S.M. Al-Alawi, Principal component
and multiple regression analysis in modelling of ground-level ozone
and interactions between the processes variables involved in and factors affecting its concentrations, Environ. Model. Software 20
the desalting/dehydration process system were investigated by (2005) 1263–1271.
using cubic mixture models. It was found that addressing the [8] M. Kramer, B.L. Palowitch, Expert system and knowledge-based
non-linearity of the process variables overcame the problem of approaches to process malfunction diagnosis, in: AIChE National Meet-
inaccurate predictions, especially, for predicting the water cut ing, Chicago, 1985.
[9] Y. Ikuta, K. Hamanaka, R. Funakoshi, A. Kako, Operation instruction
efficiency. system, in: Presented at 987 AIChE National Meeting, 1987.
This study was a preliminary research into the secondary [10] P.J. King, E.H. Mamdani, The application of fuzzy control systems to
process variables contributing to salt and water cut removal industrial processes, Automatica 13 (1977) 235–242.
efficiencies (product quality) in the desalting/dehydration pro- [11] R.M. Tong, M.B. Beck, A. Latten, Fuzzy control of the activated sludge
cess system. Further studies need to be undertaken to better wastewater treatment process, Automatica 16 (1980) 659–701.
[12] K. Watanabe, I. Matsuura, M. Abe, M. Kubota, D.M. Himmelblau, Incip-
understand the effects of other process variables on the two ient fault diagnosis of chemical processes via artificial networks, AIChE
efficiencies. Future work includes the development of extended J. 35 (1989) 1803–1812.
correlations for the two efficiencies that will take into account [13] Y.H. Pao, Adaptive pattern recognition and neural networks, Addison-
all process variables affecting the system, including the applied Wesley, New York, 1990.
voltage. Future studies to capture the highly non-linear relation- [14] KOC, Dehydration-desalting principles: a training guide, Production per-
sonnel development division, production and export operations group,
ships and interactions of the various parameters, based on the Kuwait, 1987.
use of intelligent computing and design of experiments, are sug- [15] TPL Company, Desalters phase IV (GCs 1, 3, 6, 8 and 21). Project
gested. Surface methodologies can be useful in optimizing the Report for Kuwait Oil Company, Contract 92G026, 1992.
process. [16] M. Al-Otaibi, Experimental investigation of Kuwaiti crude oil desalt-
ing/dehydration process, M.S. Thesis, Department of Chemical Engi-
neering, Kuwait University, 1999.
References [17] M. Al-Otaibi, A. Elkamel, T. Al-Sahhaf, A.S. Ahmed, An experimental
investigation of crude oil desalting/dehydration process, J. Chem. Eng.
[1] M. Willis, M. Tham, Advanced Process Control, Department of Chem- Commun. 190 (2003) 65–82.
ical and Process Engineering, University of Newcastle Upon Tyne, UK, [18] M. Al-Otaibi, Modelling and optimizing of crude oil desalting process,
1994. Ph.D. Thesis, Loughborough University, 2004.
[2] J.V. Kresta, T.E. Marlin, J.F. MacGregor, Development of inferential [19] I.T. Jolliffe, Principal Component Analysis, Springer, New York, 1986.
process models using PLS, Comput. Chem. Eng. 18 (1994) 597–611. [20] S.A. Abdul-Wahab, A. Elkamel, A.S. ALYahmadi, M.A. Al-Weshahi,
[3] H. Kamohara, A. Takinami, M. Takeda, M. Kano, S. Hasebe, I. Troubleshooting for brine heater of the MSF plant using a fuzzy logic-
Hashimoto, Product quality estimation and operating condition moni- based expert system, Desalination, submitted for publication.
toring for industrial ethylene fractionator, J. Chem. Eng. Jpn. 37 (2004) [21] D.C. Montgomery, Design and Analysis of Experiments, 5th ed., John
422–428. Wiley & Sons Inc., 2001.

Você também pode gostar