Iran. J. Chem. Chem. Eng. Vol. 36, No.

2, 2017

Online Composition Prediction of a Debutanizer Column

Using Artificial Neural Network

Nasser Mohammed Ramli

Chemical Engineering Department, Universiti Teknologi Petronas, Bandar Seri Iskandar,
31750 Tronoh, Perak, MALAYSIA

Mohd Azlan Hussain*+; Badrul Mohamed Jan

Chemical Engineering Department, Faculty of Engineering, University of Malaya,50603 Kuala Lumpur,
MALAYSIA

Bawadi Abdullah
Chemical Engineering Department, Universiti Teknologi PETRONAS, Bandar Seri Iskandar,
31750 Tronoh, Perak, MALAYSIA

ABSTRACT: The current method for composition measurement of an industrial distillation column
includes an offline method, which is slow, tedious and could lead to inaccurate results. Among
advantages of using online composition designed are to overcome the long time delay introduced
by laboratory sampling and provide better estimation, which is suitable for online monitoring
purposes. This paper presents the use of an online dynamic neural network to simultaneously
predict n-butane composition of the top and bottom products of an industrial debutanizer columns.
Principal component and partial least square analysis are used to determine the important
variables surrounding the column prior to implementing the neural network. It is due to the
different types of data available for the plant, which requires proper screening in determining
the right input variables to the dynamic model. Statistical analysis is used as a model adequacy test
for the composition prediction of n-butane in the column. Simulation results demonstrated that
the Artificial Neural Network (ANN) can reliably predict the online composition of n-butane of
the column. It is further confirmed by the statistical analysis with low Root Mean Square Error (RMSE)
value indicating better prediction.

KEYWORDS: Principal component analysis; Partial least square analysis; Neural network;
Debutanizer column.

INTRODUCTION
Distillation column is one of the important unit operations distillations are widely used for separating liquid mixtures
in the downstream process industries. In the 20th century, of chemical compounds. In petroleum downstream
* To whom correspondence should be addressed.
+ E-mail: [email protected]
1021-9986/2017/2/153-174 22/$/7.20

153
Iran. J. Chem. Chem. Eng. Nasser Mohammed Ramli et al. Vol. 36, No. 2, 2017

process alone, an approximately two thirds of the energy implementation. Many approaches have been studied
for distillation is consumed where distillation is favorable to build a secondary variable model from readily available
for separating crude oil into petroleum fractions, light measurements such as tray temperatures to replace
hydrocarbon and aromatic chemicals [1]. The separation the quality measurements. A secondary variables approach
of other chemical compounds, often in the presence has been studied for use when quality measurements
of water, is a common practice in the chemical industry. are not available and a multivariate statistical control
The success of a distillation column as a method of proposed [9].
separation is due to its operational flexibility. Inferential Composition of a distillation column requires
estimator is based on heuristic model of the process and an online estimation measurement. Temperature is the main
can be based on availability of measurement and variable that will affect the composition for the column.
multivariate regression technique. This modeling Therefore temperature gives tight control on product
approach is preferred since online prediction can provide composition despite wide variation such as internal reflux
a fast and accurate response. ratio. A simple feedback loop will give good control
In a distillation column, proper control strategies when composition of only one stream is important.
are selected through appropriate implementation. This is When monitoring or controlling both top and bottom
very important because controller has significant effect compositions simultaneously, a simple feedback loop
on product quality, production rate, and energy usage. is unfavorable due to interactions of the two loops.
In a distillation process, controlling a column is challenging With neural network, it is possible to monitor composition
since it involves nonlinearities, dual composition control at both ends of the column. Neural network techniques
and disturbances. In refining industries, the product have been increasingly used for a wide variety of
quality of a debutanizer column is always the main focus applications where statistical method had been
of its operation [2, 3]. Several alternative column traditionally employed. Neural network is able to give
configurations have been developed, to control better prediction of important parameters and be applied
temperature of the column. For example, dual to wide range of problems [10].
temperature control method is achieved by combining Artificial Neural Network (ANN) is a powerful tool to
the middle and top temperature change to control the state model non-linear process in refinery. ANN is a black box
switch of the total reflux and withdrawal during where no priori knowledge for a system is required. ANN
the operation [4]. Composition prediction based on tray has the ability to learn a relationship between the outputs
temperatures is typically used in industries by controlling and the inputs for a system. To develop a process using
tray temperatures [5]. The steady state and dynamic ANN, it requires suitable network architecture and
process could be characterized for product variability appropriate data training. The neural network architecture
prediction. These characteristics were used to generate makes use of many hidden layers for the column and
a linear dynamic tray-to-tray model for a distillation the inputs only consider temperature surrounding
column [6]. the column [11]. To improve product quality in a debutanizer
Dynamic Principal Component Analysis (PCA) column, soft sensor design has been used. The dynamic
consider the dynamic process by introducing time lagged neural model that has been implemented used three steps
variables into the inferential models. The dynamic predictions to evaluate its top product concentration.
methods are suitable for processes with long time delays The output from the first dynamic network are fed to input
and varying throughputs on process variables [7]. of the second dynamic network and the output from
Secondary variables are sensitive to primary variables the second dynamic network are fed to the third dynamic
when implementing inferential estimator in achieving network to obtain the desired product. The approach uses
the column optimum performance. As there are many appropriate lagged inputs including composition to the
possible locations of the temperature sensor for a column, neural network.
the selection of the secondary variables as inputs Real time estimation of plant variables and the
becomes difficult [8]. PCA properties are measured composition are used for monitoring purposes and
by extracting the input variable for the online the number of neurons in the hidden layer for the neural

154
Iran. J. Chem. Chem. Eng. Online Composition Prediction of a Debutanizer Column ... Vol. 36, No. 2, 2017

network is determined by trial and error method [12]. over a 4 year period in the oil refinery industry. PCA
The Levenberg–Marquardt (LM) algorithm for neural network is one of the criteria to describe a remarkably simple
training has been used because it is suitable for binary approach to multivariate analysis based on projection
as well as multi-component mixture. The LM algorithm methods. The projection approach can be adapted to
is more suitable compared to Steepest Descent Back a variety of data analytical objectives such as summarizing
Propagation (SDBP) algorithm in both cases and give and visualizing a data set, multivariate classification and
more accurate and sensitive results. The LM approach discriminant analysis and finding quantitative
has proven worked efficiently in complex chemical plants, relationship among the variables. Projection methods
having hundreds of parameters [13]. Neural network can be made robust to outliers, deal with non-linear
gives better performance than the conventional control relationship and adapt to drift in multivariate process
loop and inferential control by developing a model based data. Furthermore the application of principal component
on neural control for single composition. The strategy and partial least square analysis are used to determine the
is used to compensate the upsets in the operating pressure, important variables surrounding the column prior
feed flow rate, and feed composition. To keep the content to implementing the neural network which is vital. This is
of the key component in the distillate stream, the because there are different types of data available for
performance of the neural network has to be precise [14]. the plant, which requires proper screening to determine
ANN has also been applied in a crude fractionation the right input variables to the dynamic model. In addition,
section. Back propagation algorithm is used on real time this work also involves a single dynamic neural network
data and the output of the neural network prediction is model with lagged inputs to predict the top and bottom
the naphtha temperature and not the composition prediction. composition simultaneously. Statistical analyses are used
However, the proposed method is only used for the as a model adequacy test for composition prediction of
product quality determination [15]. n-butane in the column. Neural network modeling is a good
An optimization framework to obtain optimal strategy for large industrial application when online
operation of dynamic processes under process-model estimation is required for monitoring purposes. The
mismatches has been developed. In order to model these online measurement provide good estimate to overcome
mismatches, neural network have been utilized. delay introduced by laboratory sampling.
To demonstrate the technique, a batch distillation process
is used. For simplicity, they consider binary PLANT AND DEBUTANIZER COLUMN
batch distillation with only one specified product [16]. DESCRIPTION
A framework is proposed to optimize the operation of The plant under study is for offshore crude oil
batch columns. The proposed framework uses to produce high value petroleum products for domestic
ANN based process model to be employed by the optimizer. and export markets. The plant consists of a refinery process
The optimization of a pilot-plant middle-vessel batch and involves in condensate fractionation and reforming
column has been considered. The maximum-product aromatics processes. The feed stocks of the oil refinery
problem is formulated and solved by optimizing the column are crude oil while the products are petroleum products,
operating parameters, such as the batch time, the reflux liquefied petroleum gas, naphtha and low sulphur waxy
and reboil ratios. ANN based model is capable of residue. The refinery has two main process units, which
reproducing the actual plant dynamics with good accuracy, are the Crude Distillation Unit (CDU) and Catalytic
and allows a large number of optimization studies to be Reforming Unit (CRU). The crude oil is preheated using
carried out with little computational effort [17]. heat exchangers to 190oC – 210oC. Subsequently, the
The main contribution of the work in this paper is stream is then further heated in a furnace at 340 oC –
the use of online closed loop and open loop data for training 342oC to achieve a desired temperature for better
the neural network. The close loop data has been separation at the CDU. The crude is then routed to the
extracted to obtain the open loop data, which is then used CDU before being split into a number of fractions which
for training, validation and testing. In this study, column are Heavy Straight Run Naphtha (HSRN) as overhead
information was obtained from the actual measurement vapor, untreated kerosene, straight run kerosene and

155
Iran. J. Chem. Chem. Eng. Nasser Mohammed Ramli et al. Vol. 36, No. 2, 2017

straight run diesel. From the crude tower, there are Table 1: Column specification.
3 sides cut streams, which are drawn to a stripper column.
Number of tray of the column 35
The stripper consists of naphtha stripper, kerosene
stripper and diesel stripper. Feed tray - stage number 23
The feed of Heavy Straight Run Naphtha (HSRN)
from the CDU is mixed with hydrogen from the reformer Type of tray used Valve

and heated up to the reaction temperature using a heater

Column diameter 1.3 meter
and fed into a pretreater catalytic reactor. The products
from the reactor is sent to the pretreater stripper. Column height 23.95 meter
The bottom product of the stripper is sent to the reforming
unit as feed. The treated naphtha is heated to reaction Condenser type Partial

temperature and fed to the reforming reactors. Effluent

Feed mass flowrate 44106 kghr-1
from the reactor is cooled and collected in a reformer
separator. One part of the gas, which is separated, is sent Feed temperature 113 0C
to an absorber. In the absorber, hydrogen gas is purged
and recycled to the pretreater heater as the hydrogen Feed pressure 823.8 kPa

make-up for the raw naphtha feed. The liquid fraction

Overhead vapor mass flowrate 11286 kghr-1
is pumped into a stabiliser. The reformate is withdrawn
from the stabiliser bottom and cooled before being Overhead liquid mass flowrate 5040 kghr-1
transferred to storage. Overhead vapor from the stabiliser
are cooled, condensed and recovered from the stabiliser Condenser pressure 823.8 kPa

reflux drum. In the current work, we are focusing on the

Reboiler pressure 853.2 kPa
debutanizer column.
The debutanizer column is located at the CDU
section. The unit is used to recover light gases and LPG to strip the light component. The reboiler temperature
from the overhead distillate before producing light is controlled by the debutanizer reboiler control valve while
naphtha. The light gases mainly C2 are used to refine fuel the bottom product level is controlled by the debutanizer
gas and mixed LPG to LPG storage. The unit operation bottom level controller. Table 1 outlines the column
at the LPG section includes the Deethanizer and specification and Table 2 describes the tag name
Debutanizer columns. The recovery unit is used to surrounding the column. Fig. 1 shows the flow chart
process unstable product from the top unit section for the refinery process and Fig. 2 shows the column
together with unstable LPG from the reformer. configuration for the debutanizer column. The objective
Deethanizer bottom product enters as feed to the of this case study is to estimate the top and bottom
debutanizer column. The debutanizer column has compositions of n-butane for the debutanizer column
35 valve trays (one liquid pass). Low boiling point using a single neural network model based on a mix of
components rise up the tower in contact with the internal industrial and simulation data. The methodology applied
reflux. The high boiling point of heavy component flow will be described in the next section.
down in contact with vapor produced in the debutanizer
reboiler. Overhead vapor is then condensed by the METHODOLOGY
debutanizer condenser. The overhead system is set and Data generation
controlled by the debutanizer overhead pressure control Although most online open loop response data from
valves which have two split ranges control. the plant surrounding the column is available, some
Part of the collected condensed hydrocarbon of the variables in open loop surrounding the column
is routed to the top of the debutanizer as reflux. The flow are not available. Such data include Temp 5, Pressure 1 and
is measured by the reflux flow meter. The debutanizer compositions at both ends of the column. Therefore plant
bottom section is provided by the debutanizer reboiler process simulation of the debutanizer column is performed

156
Iran. J. Chem. Chem. Eng. Online Composition Prediction of a Debutanizer Column ... Vol. 36, No. 2, 2017

Table 2: Tag name description of the column.

Tag Description Units
o
Temp 1 Debutanizer top temperature C
o
Temp 2 Debutanizer bottom temperature C
o
Temp 3 Debutanizer receiver bottom temperature C
o
Temp 4 Light Naphtha temperature after condenser E 1 C
o
Temp 5 Reboiler outlet temperature to column C
o
Temp 6 Debutanizer feed temperature C

Level 1 Debutanizer level %

Level 2 Debutanizer condenser level %

Level 3 Debutanizer level indicator %

Level 4 Condenser level indicator %

Flow 1 Light Naphtha flow to storage m3/hr

Flow 2 LPG flow to storage m3/hr

Pressure 1 Debutanizer receiver overhead pressure kPa

Fig. 1: Flow chart for the refinery process.

157
Iran. J. Chem. Chem. Eng. Nasser Mohammed Ramli et al. Vol. 36, No. 2, 2017

Table 3: Controller setting and set point.

Kc Ti Td Set point

Flow 1 0.5 30 0 44.64 m3/hr

Reflux flow 0.285 50 0 24 m3/hr

Temp 5 0.4 80 20 135.7 oC

Pressure 1
PC

PT E1 Debutanizer to other Column

Level 2
Temp 1 Condenser
LT LC Flow 2
Feed flow FC FT FC FT

FC LPG to storage
FT
Reflux flow
Temp 3
Debut feed Debutanizer Level 1
Feed
LT Temp 5 LC
Temp 6 TT TC
Flow 1 FC FT

E2
Light Naphtha to storage
E3

Temp 4
Temp 2

Fig. 2: Debutanizer column configuration.

to obtain the unavailable open loop data sets from since the controller setting and set point data are
the plant. The online and simulated close loop response available. Table 3 shows the set point and controller
of the compositions of n-butane at the top and bottom settings for Flow 1, reflux flow and Temp 5.
of the column were also established to obtain the unmeasured The PID equation is used to determine the controller
data. The simulation environment consists of a worksheet output (manipulated variable) derived from reference [18]
and Process Flow Diagram (PFD). The worksheet as given below;
contains information on each flow and heat stream
 1 t de(t) 
involved in the simulation. Comparison between the close  MV  K c e(t)  0 e(t)  D  (1)
loop responses in simulation to the actual plant from  i dt 
laboratory sample is performed to evaluate the deviation
The tuning parameter is used to determine the process
between the simulated and actual compositions,
gain, Kp [19] as shown below;
to ascertain that the simulation data available closely
resemble the actual online industrial data. K p  3K c (2)
Online close loop data for top and bottom
compositions n-butane are available for the column which Equations 1 and 2 were used to extract the close loop
are obtained from laboratory sample (gas chromatographs) data to determine the process variable Temp 5 and
while the close loop data could be extracted to open loop manipulated variable of the reboiler and reflux

158
Iran. J. Chem. Chem. Eng. Online Composition Prediction of a Debutanizer Column ... Vol. 36, No. 2, 2017

between the variables, small number of principal

Determine the error components will summarize a majority of the variation
e= set point – measurement process variable
in X. To analyze the changes in the original data space,
changes occurring within the principal components
Determine Kc, Ti and Td
should be used. From the PCA, the important variables
surrounding the column are determined. Partial Least
Square (PLS) is used to relate the important variables
Calculate the controller output change of the from PCA with respect to the top and bottom compositions
manipulated variable, MV
of n-butane. PLS regression is a method that generalizes
and combines features from principal component analysis
Calculate change of the process variable, and multiple regressions. It is normally useful to predict
PV a set of dependent variables (Y) from a large set of
independent variables or predictors (X).
The data set which obtained from open loop online
and simulation are combined together. SIMCA-P is used
Obtain open loop to perform PCA and PLS analysis for the debutanizer
data
column. There are 2 important variables, which are
the primary and observation variable. The primary variable
consists of 23 variables surrounding the column while
Fig. 3: Flow chart to extract close loop to obtain open loop data. the observation variables are the top and bottom n-butane
compositions. The observation variable is the number of
respectively are called MV2 and MV3. The manipulated observations established once the worksheet has been developed.
variable reboiler flow rate could be obtained from Temp 5 Then PCA model is fitted to these data.
which is applied to regulate the particular process Table 4 shows the important variables involved
variable. Equation 2 has been used to determine the in the PCA analysis for open loop response of the reboiler
process gain from the controller gain, Kc. Process variable flow rate and reflux flow rate data set obtained from the plant
for Flow 1, reflux flow and Temp 5 can be determined and simulation. The simulated data are Temp 5, Pressure
when the process gain, Kp and the change of the 1 and open loop compositions. The rest of the variables
manipulated variable, MV are calculated. Fig. 3 shows are obtained from the actual plant. The variables ‘a’ to
the procedure to extract the close loop data. There are ‘m’ are the important variables surrounding the
three types of data. The first is the online open loop data. debutanizer column where the variables ‘n_top’ to
The second is the data extracted online in close loop and ‘r_bottom’ are the compositions at the top and bottom of
the last is the simulation data. These three data are the column.
combined together for training, validation and testing of From the PCA, variables which are not important
the neural network model. are excluded while the important variables are analyzed
again with respect to the top and bottom compositions n-butane
Principal component (PCA) and Partial Least Square (PLS) using PLS. For PCA, component contribution plot are
Analysis used to analyze all the important variables surrounding
PCA and PLS analysis are used prior to utilizing the column.
neural network to determine the important variables Table 5 shows the important variables involved in the
to be analyzed for composition prediction as this is crucial PLS analysis where the manipulated variable is the
since there is a large number of variables surrounding reboiler flow rate. Table 6 shows the important variables
the column. The PCA and PLS analyses will also determine involved in the PLS analysis where the manipulated
the inputs to the neural network. variable is the reflux flow rate. Data is arranged
PCA is used to analyze all variables surrounding according to Tables 5 and Table 6 for PLS analyses
the column outlined in Table 4. If a correlation exists to further study the importance of the variables surrounding

159
Iran. J. Chem. Chem. Eng. Nasser Mohammed Ramli et al. Vol. 36, No. 2, 2017

Table 4: Important variables for PCA model.

Temp 1 a Debutanizer top temperature

Temp 2 b Debutanizer bottom temperature

Temp 3 c Debutanizer receiver bottom temperature

Temp 4 d Light Naphtha temperature after condenser E121

Temp 5 e Reboiler outlet temperature to column

Temp 6 f Debutanizer feed temperature

Level 1 g Debutanizer level

Level 2 h Debutanizer condenser level

Level 3 i Debutanizer level indicator

Level 4 j Condenser level indicator

Flow 1 k Light Naphtha flow to storage

Flow 2 l LPG flow to storage

Pressure 1 m Debutanizer receiver overhead pressure

Component 1 n_top Top composition propane

n_bot Bottom composition propane

Component 2 o_top Top composition i-butane

o_bot Bottom composition i-butane

Component 3 p_top Top composition n-butane

p_bot Bottom composition n-butane

Component 4 q_top Top composition i-pentane

q_bot Bottom composition i-pentane

Component 5 r_top Top composition n-pentane

r_bot Bottom composition n-pentane

Table 5: Important variables for PLS model (reboiler flow rate as manipulated variable).

Temp 1 a Debutanizer top temperature

Temp 2 b Debutanizer bottom temperature

Temp 5 e Reboiler outlet temperature to column

Flow 1 k Light Naphtha flow to storage

Component 3 p_top Top composition of n-butane

p_bot Bottom composition of n-butane

160
Iran. J. Chem. Chem. Eng. Online Composition Prediction of a Debutanizer Column ... Vol. 36, No. 2, 2017

Table 6: Important variables for PLS model (reflux flow rate as manipulated variable).
Temp 1 a Debutanizer top temperature

Temp 2 b Debutanizer bottom temperature

Temp 5 e Reboiler outlet temperature to column

Flow 2 l LPG flow to storage

Pressure 1 m Debutanizer receiver overhead pressure

Component 3 p_top Top composition of n-butane

p_bot Bottom composition of n-butane

the column. For PLS analysis, variable important plot using statistical analysis and it is described in next section.
are used to determine variables which are important The number of inputs to the network is 12 while
with respect to the n-butane composition. From the PCA the outputs are 2 (compositions of top and bottom).
and PLS analyses, component contribution plot and variable The transfer function to train the network is purelin (linear)
important plot are used to identify the variables that are for the entire layer and the networks are trained to predict
important to be selected the right inputs for neural simultaneously the top and bottom compositions of
network. n-butane. Table 7 shows the important variables in the
neural network where the data set are combined with
Neural network design the manipulated reboiler flow rate and reflux flow rate
The online compositions at the top and bottom of changes for n-butane after performing the PCA and PLS
the column in the refinery is currently measured using analyses.
normal laboratory sampling. This is tedious and Fig. 4 shows the neural network architecture for
the results could not be obtained immediately therefore the n-butane composition prediction where the inputs for
neural network are used as a benchmark because it is able the neural network are from mv2 (k) to p_bot (k-1) while
to predict the composition faster with more accuracy and the outputs are the variable p_top (k+1) and p_bot (k+1). Fig. 5
precision and could also handle non-linearities in the shows the general procedure in developing the neural
process variable surrounding the column as proposed network architecture. Prior to implementing the neural
in this study. network, the data are arranged by combining the open
Open loop response of the reboiler and reflux data, loop response from the simulation and online data.
which includes the composition of n-butane, are used The data set are then trained until the network reaches
to develop the dynamic neural network architecture. its epoch and meet its performance criteria. The data set
The selected input variables to the network including are also validated and tested after the network is trained.
the composition of n-butane are time delayed while Since the extracted close loop data are available, the data
the outputs are the future predictions of n-butane. are replaced as inputs to the neural network in the
The type of dynamic network used for training, validation validation and test set by maintaining the actual
and testing the data set are nonlinear autoregressive architecture that are trained for the open loop response.
network with exogenous inputs (NARX) with series-
parallel architecture. The training algorithm used is the
Levenberg-Marquardt method. Early stopping criteria Model adequacy test for neural network to determine
are implemented to train the network while the performance the hidden layer
function used is the mean square error. To determine the number of neurons in the hidden
These data sets are partitioned into three sets, layer, the following criteria is analyzed to determine the
which are the training, validation and test set. In the accuracy performance of the network using statistical
network, the number of layers used is 3 with only 1 analysis. The error is estimated using the Root Mean
hidden layer. The number of hidden layer is determined Square Error (RMSE) method given as,

161
Iran. J. Chem. Chem. Eng. Nasser Mohammed Ramli et al. Vol. 36, No. 2, 2017

Table 7: Important variables for neural network prediction.

MV2 mv2 (k) Manipulated reboiler flow rate

mv2 (k-1) Lag MV2

MV3 mv3 (k) Manipulated reflux flow rate

mv3 (k-1) Lag MV3

Temp1 a (k) Debutanizer top temperature

a (k-1) Lag Temp 1

Temp 5 e (k) Reboiler outlet temperature to column

e (k-1) Lag Temp 5

Component 3 p_top (k) Top composition n-butane

p_top (k-1) Lag top composition

p_bot (k) Bottom composition n-butane

p_bot (k-1) Lag bottom composition

p_top (k+1) Future predictions n-butane top composition

p_bot (k+1) Future predictions n-butane bottom composition

mv2(k)

mv2(k-1)
mv3(k)
mv3(k-1)

a(k)

a(k-1)
e(k)

e(k-1) p_top(k+1)

p_top
p_top(k)
p_bot(k+1)
p_top(k-1)

p_bot
p_bot(k)
p_bot(k-1)

Fig. 4: Neural network architecture for n-butane.

162
Iran. J. Chem. Chem. Eng. Online Composition Prediction of a Debutanizer Column ... Vol. 36, No. 2, 2017

Obtain data for training and

Validation in Excel interface with
MATLAB

Choose suitable input/output

to the network

Choose dynamic neural network

architecture Reevaluate the neural network
architecture by changing the
Initialize of weight and biases value hidden layer and transfer function

Train the neural network with No

early stopping training algorithm until
training parameter goal has been met
Validation
Validate the trained neural network Sufficient ?
With validation data set
Yes
Final neural network architecture

Fig. 5: Procedure of the neural network architecture.

(x measured  x predicted ) 2 T
 (y t  f t )
2
RMSE  (3)
N R 2  1 t L (7)
T 
 (y t  y)
2
Correct Directional Change (CDC) measures the
t L
capacity of a model to accurately predict the subsequent
actual change of a predicted variable. The formula of Mean Absolute Percentage Error (MAPE) is the
CDC is defined as: measure of accuracy in a fitted time series value given as:

100 N 1 N | Ft  A t |
D (4) MAPE   100% (8)
N i i N i 1 A t

where formula of Di is defined as: Pearson correlation coefficient (Cp), measures

the goodness of fit of the regression given as:
yi  yi
NS
The best known information criteria are the Akaike  (E p, j  E p, j )(E a, j  E a, j )
j1
Information Criterion (AIC) and Bayesian information Cp  (9)
NS NS
criteria (BIC):
 (E p, j  E p, j ) 2  (E a, j  E a, j ) 2
j1 j1
2K
AIC  MSE  2 (5)
T The number of neurons in the hidden layer is determined
from a range of 8 to 40. Fig. 5a shows the profile
log(N)2 2K
BIC  MSE  (6) of the RMSE with respect to number of hidden nodes.
T
Using the statistical analysis described above,
The coefficient of determination is defined as: with the following the deviation between actual and

163
Iran. J. Chem. Chem. Eng. Nasser Mohammed Ramli et al. Vol. 36, No. 2, 2017

RMSE profile of n-butane composition prediction by neural network are determined

0.001
6
with the following set of criteria; low RMSE, CDC
botto to
0.001
m p approaching 100, small AIC and BIC, R2 approaching 1,
4
0.001 low MAPE and CP approaching 1. Equations 3 – 9
2
RMSE

0.00 are obtained from reference [20-22]. Based on the set criteria,
1
0.000 the best neural network architecture is determined.
8
0.000 The number of hidden nodes for the optimum neural network
6 model has been found to be 22.
0.000
4
0.000
2 RESULTS AND DISCUSSION
0
8 12 16 20 24 28 32 36 40 Step test reboiler flow rate
Number of hidden nodes
To generate the input-output data for the neural
network training, various step changes are applied to the
Fig. 5a: Profile of the RMSE with respect to number of
inputs to obtain the corresponding outputs. The inputs for
hidden nodes. the system in this case is the reboiler flow rate and reflux
flow rate. Figs. 6 to 9 show some of the step tests for
Step test Temp 1 the reboiler flow rate data sets. The step test of the reboiler
145 62
flow rate, which is the manipulated variable, is generated
by using multi amplitude rectangular pulse. The step test
Reboiler flow rate (m3/hr)

60
Temperature (°C)

144
is important to determine the effect and fluctuations of
58
143 the process variable when performing changes to the
56 reboiler flow rate.
142
54
Step test reflux flow rate
141 52 Figs. 10 to 13 describe some of the step tests for
Reboiler.Flow Temp 1
the reflux flow rate data which demonstrates the effect
140 50
and the fluctuations of the process variable when performing
1 51 101 151 201 251 301 351 401 451 501
changes to the reflux flow rate. The fluctuations of the
Time (min) process variables will start to increase and decreases
as the step test of the reflux flow rate changes. The step test
Fig. 6: Temp 1 Debutanizer top temperature. of the reflux flow rate, which is the manipulated variable,
is generated by using multi amplitude rectangular pulse.
Step test Level 1
145 80
Validation of online data with simulation data for
75
144 composition
Reboiler flow rate (m3/hr)

70 Figs. 14 and 15 represent the compositions at the top

Level (%)

143
and bottom of n-butane from there Root Mean Square
65
Error (RMSE) could be calculated for top and bottom
142
60 compositions are 0.0251 and 0.008184 respectively.
141 It indicates that there is small deviation between the online
55
Reboiler.Flow Level 1 and simulation data. This implies that the simulation and
140 50 close loop data agree well with each other including
1 51 101 151 201 251 301 351 401 451 501 the variables that are not available from the open loop
response for the plant. Once the close loop has been
Time (min)
verified, then the open loop response for variable that
Fig. 7: Level 1 Debutanizer level. is not available from plant can be obtained. The same step

164
Iran. J. Chem. Chem. Eng. Online Composition Prediction of a Debutanizer Column ... Vol. 36, No. 2, 2017

Step test Flow 1 Step test Level 1

145 45 22.5 64

40 22
Reboiler flow rate (m3/hr)

Reflux flow rate (m3/hr)

144
35

Flow rate (m3/hr)

21.5
30 60

Level (%)
143 21
25
20 20.5
142
56
15 20
141 10
Reboiler.Flow 19.5 Reflux.flowrate Level 1
Flow 1
5
140 0 19 52

1 51 101 151 201 251 301 351 401 451 501 1 51 101 151 201 251 301

Time (min) Time (min)

Fig. 8: Flow 1 Light Naphtha flow to storage. Fig. 11: Level 1 Debutanizer level.

Step test Pressure 1 Step test Flow 1

145 850 22.5 36
Reboiler.Flow Pressure 1
Reboiler flow rate (m3/hr)

22
Reflux flow rate (m3/hr)

144 800

Flow rate (m3/hr)

21.5
Pressure (kPa)

32
143 750 21

142 700 20.5

28
20
141 650
19.5 Reflux.flowrate Flow 1

140 600 19 24
1 51 101 151 201 251 301 351 401 451 501 1 51 101 151 201 251 301

Time (min) Time (min)

Fig. 9: Pressure 1 Debutanizer receiver overhead pressure. Fig. 12: Flow 1 Light Naphtha flow to storage.

Step test Temp 1 Step test Pressure 1

22.5 61 22.5 850

22 22
60
Relux flow rate (m3/hr)

800
Reflux flow rate (m3/hr)

Temperature (°C)

21.5 21.5
Pressure (kPa)

59
750
21 21
58
20.5 20.5 700
57
20 20
650
19.5 56 19.5 Reflux.flowrate Pressure 1
Reflux.flowrate Temp 1

19 55 19 600
1 51 101 151 201 251 301 1 51 101 151 201 251 301

Time (min) Time (min)

Fig. 10: Temp 1 Debutanizer top temperature. Fig. 13: Pressure 1 Debutanizer receiver overhead pressure.

165
Iran. J. Chem. Chem. Eng. Nasser Mohammed Ramli et al. Vol. 36, No. 2, 2017

Top composition n-butane Light naphtha flow to storage

0.35 60
Composition (mole fraction)

0.3 simulation online 50

Flow rate (m2/h)

0.25
40
0.2
30
0.15
20
0.1
extract close loop
10
0.05

0 0
0 2000 4000 6000 8000 10000 12000 0 100 200 300 400 500 600

Time (min) Time (min)

Fig. 14: Top composition n-butane. Fig. 16: Flow 1.

Bottom composition n-butane Reboiler outlet temperature to column

0.10 500 150
Composition (mole fraction)

0.09 extract close loop

145
0.08 400

Temperature (°C)
Temperature (°C)

0.07
140
0.06 300
0.05 135
0.04 200
0.03 130

0.02 simulation online 100

125
0.01
0.00 0 120
0 2000 4000 6000 8000 10000 12000 0 100 200 300 400 500 600

Time (min) Time (min)

Fig. 15: Bottom composition n-butane. Fig. 17: Temp 5.

size for the manipulated variable for reboiler and reflux has a large error because the variation between
from plant are performed as inputs to obtain the the fluctuations of the process variable and its set point
fluctuation of the process variable as outputs in the is large compared to Temp 5. The controller setting for
simulation. the plant affects the error to bring the process variables
close to its set point.
Extracted close loop data
Figs. 16 and 17 show the fluctuations of the two PCA analysis
process variables that are in close loop response Step test reboiler flow rate
compared to the extracted closed loop. The fluctuations Component contribution plot shows all of the
of the variables show a variation between close loop and important components of the step test reboiler flow rate
data extracted from close loop. Flow 1 is controlled using for the top liquefied petroleum gas product and bottom
PI and Temp 5 is controlled by using PID. The step light naphtha product. The variables are outlined in Table 4.
change for the Manipulated Variable (MV) for Temp 5 is The values of R2 (variation explained) and Q2 (variation
larger than Flow 1. This is because Temp 5 has small error predicted according to cross validation) are also shown
with respect to time compared to Flow 1. Flow 1 for each variable. The variables with high values of R2

166
Iran. J. Chem. Chem. Eng. Online Composition Prediction of a Debutanizer Column ... Vol. 36, No. 2, 2017

have large loading values for the selected component. Step test reboiler variable. M1 (PCA-X), All variable
The Q2 values indicate the reliability of these R2 and 1.0
loading values. R2VX[3]
The component is considered less important if all of 0.8 Q2VX[3]

the variables have low values of R2 and Q2 in a

0.6
component. This is applicable to variables c, d, f, g, h, i, j,
l and m as depicted in Fig. 18. The highest value of R2 0.4
and Q2 for variable b is 0.48 and 0.44 respectively. From
the plot, it could be concluded that variables a, b, e and k 0.2

have high values for R2 and Q2 indicating that these

0.0
components are important for composition prediction of
a b c d e f g h i j k l m n_top
n_bot
o_top
o_bot
p_top
p_bot
q_top
q_bot
r_top
r_bot
n-butane with respect to the reboiler flow rate.
Var ID (primary)

Step test reflux flow rate Fig. 18: Component contribution plot reboiler flow rate.
Component contribution plot shows all of the
important components of the step test reflux flow rate for
Step test reboiler variable. M1 (PCA-X), All variable
the top liquefied petroleum gas product and bottom light
1.0
naphtha product. The values of R2 (variation explained)
R2VX[5]
and Q2 (variation predicted according to cross validation) 0.8
Q2VX[5]
are also shown for each variable. Variables with high
0.6
values of R2 show large loading values for the selected
0.4
component. Q2 value indicates the reliability of these R2
and loading values. 0.2
If all of the variables show low values of R2 and Q2 0.0
in a component, it indicates that the component is less
-0.2
-
important [23]. This is true for variables c, d, f, g, h, i, j
and k as depicted in Fig. 19. The highest value of R2 and a b c d e f g h i j k l mn_top
n_bot
o_top
o_bot
p_top
p_bot
q_top
q_bot
r_top
r_bot

Q2 for variable e is 0.1 and 0.28 respectively. Form the

Var ID (Primary)
plot it indicate that variables b, e, n_top and q_top have
high values of R2 and Q2. It indicates that these Fig. 19: Component contribution plot reflux flow rate.
components are important for composition prediction of
n-butane with respect to the reflux flow rate.
Step test reboiler variable. M1 (PLS),
n-butane VIP [Comp. 1]
PLS analysis for n-butane 1.8
From Fig. 20, the output variable of the PLS analysis 1.6

is the n-butane composition and the input variable is 1.4

the variable Temp 1 to Flow 1 outlined in Table 5. All of 1.2

VIP [1]

1.0
the variables for level have been excluded from the analysis,
0.8
since level will not affect the fluctuations of the top and
0.6
bottom compositions of n-butane.
0.4
From Fig. 20, the importance of these variables is 0.2
determined by having the y axis value which is the 0.0
Variable Important Plot (VIP) more than 0.5. If the value a k b e

of the bar chart for the particular variable is less than 0.5, Var ID (Primary)
the variable is not important and it could be excluded
from the analysis the variables i.e. k, b and e. Fig. 20: Step test reboiler variable important plot.

167
Iran. J. Chem. Chem. Eng. Nasser Mohammed Ramli et al. Vol. 36, No. 2, 2017

From Fig. 21, the output variable of the PLS analysis Step test reflux variable. M4 (PLS),
is the n-butane composition and the input variable are the n-butane VIP [Comp. 1]

variables Temp 1 to Pressure 1 outlined in Table 6. All of 1.6

the variables for the level have again been excluded from 1.4
the analysis. The importance of a variable is considered 1.2
important if its y axis value in the VIP bar chart is more

VIP [1]
1.0
than 0.5 [25]. If the value of bar chart of the particular 0.8
variable is less than 0.5, that means the variable is not 0.6
important and the variable could be excluded namely 0.4
variables l, m and b. 0.2
0.0
a e l m b
Artificial Neural network modeling Var ID (Primary)
Table 8 shows the summary of the neural network
architecture for predicting the compositions of n-butane at Fig. 21: Step test reflux variable important plot.
the top and bottom of the column simultaneously. The data
are partitioned according to the training; validation and test
Manipulated variable reboiler and reflux
set as shown in Fig. 22. Fig. 22 a shows the training neural
145 22.5
network by LM. Figs. 23 to 26 show the top and bottom
Reboiler
compositions prediction of n-butane for validation and 144.5
Reflux 22
testing. The amount of data that are partitioned according to 144
MV (m3/h)

MV (m3/h)
training is 65%, for validation is 18% and test is 17%. The 143.5 21.5
(m /h)

results indicate the RMSE deviation between the open loop

3

143

and the extracted close loop is 8.213  10-10 for the bottom 142.5 21

composition validation and 0.16  10-9 for the top 142

20.5
composition validation. The CDC value for bottom 141.5
composition validation and testing are similar. The CDC 141 20
0 50 100 150 200 250 300
value for top composition validation varies by 1.62 and for
top composition testing varies by 0.32. The CDC value for Time (min)
bottom validation and testing is high calculated to be 100 for
Fig. 22: Manipulated variable reboiler and reflux flow rate.
extract close loop and open loop response. High CDC value
indicates that prediction by NN is very good. Di is equal to 1
if yi  yi is greater than zero. Di is equal to zero if yi  yi Performance is 6.11025e-018. Goal is 1e-006
is negative. The high CDC value indicate Di is equal to 1
105
based on the prediction is larger than Di which is zero. The
CDC for top validation and testing is low. This is because Di 100
Performance

which is zero is larger than Di is equal to 1 and the

subsequent actual change of the predicted variable is low. 10-5

The regression value of R for top and bottom compositions

10-10
test and validation is 1. Thus the prediction between the
actual and simulated is similar. The AIC and BIC values for 10-15
open loop for top composition validation are -572 and -564
respectively. The AIC and BIC values for open loop for 0 1 2 3 4 5 6

bottom composition validation are -357 and -349 6-Epochs

respectively. The values are larger compared to extract
from close loop. The model for extract from close loop Fig 22a: Training neural network by LM.

168
Iran. J. Chem. Chem. Eng. Online Composition Prediction of a Debutanizer Column ... Vol. 36, No. 2, 2017

Table 8: Neural network architecture.

Parameters Description

Network NARX series parallel network

Category With partitioning

Training function Levenberg-Marquardt

Adaptation learning function Gradient descent

Performance function MSE

Epochs 1000

Goal 1e-6

Number of layers 3
Layer 1: Number of Neuron 12
Transfer function Linear
Layer 2: Number of Neuron 22
Transfer function Linear
Layer 3: Number of Neuron 2
Transfer function Linear

Actual and Simulated plot for Top composition Actual and Simulated plot for bottom composition
n-butane:Testing n-butane:Testing
0.126 0.08
Bottom composition (model fraction):

Actual Actual
Top composition (model fraction)

0.125 Simulated 0.075 Simulated

0.124 0.07
0.123 0.065
testing

testing

0.122 0.06
0.121 0.055
0.12 0.05
0.119 0.045
0.118 0.04
0.117
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Time Time

Fig. 23: Actual and simulated n-butane at the top composition Fig. 24: Actual and simulated n-butane at the bottom
for testing. composition for testing.

has low value of the information criteria gives the best compositions are calculated to be 1 and the MAPE should
performance. For the AIC and BIC, low value is preferred be close to 0. The MAPE values for top and bottom
as it indicates better prediction. This is also applied composition validation open loop are 1.04  10-6 and -6.4
to testing data set. The Akaike information criteria related  10-6 respectively. The MAPE values are larger
to the square of residual to the number of free model compared to extract from close loop. This is also applied
parameters. The purpose is to weigh the error of the to testing data set for MAPE values. When having a
model against the number of parameters. The BIC is perfect fit, MAPE is zero. But in regards to upper level
similar to AIC except that it is motivated by the Bayesian the MAPE has no restriction. The percentage error
model selection principles. Cp value is close to 1. The Cp calculated for MAPE is to compare the error of fitted
value for validation and test set for bottom and top time series. Table 9 shows the statistical analysis

169
Iran. J. Chem. Chem. Eng. Nasser Mohammed Ramli et al. Vol. 36, No. 2, 2017

Table 9: Statistical analysis of n-butane open loop and extract close loop.
Parameter Extract Open loop
rmse_bottom_validation 2.52E-09 3.35E-09
rmse_top_validation 1.15E-09 1.31E-09
CDC_bottom_validation 100 100
CDC_top_validation 17.30 15.68
R_bottom_validation 1 1
R_top_validation 1 1
AIC_bottom_validation -352.80 -357.77
AIC_top_validation -561.06 -572.12
BIC_bottom_validation -344.99 -349.89
BIC_top_validation -553.26 -564.24
MAPE_bottom_validation -4.67E-06 -6.4E-06
MAPE_top_validation 9.23E-07 1.04E-06
Cp_bottom_validation 1 1
Cp_top_validation 1 1
rmse_bottom_testing 2.48E-09 3.33E-09
rmse_top_testing 1.15E-09 1.32E-09
CDC_bottom_testing 100 100
CDC_top_testing 16.32 16
R_bottom_testing 1 1
R_top_testing 1 1
AIC_bottom_testing -344.941 -349.38
AIC_top_testing -479.41 -487.47
BIC_bottom_testing -337.29 -341.66
BIC_top_testing -471.76 -479.74
MAPE_bottom_testing -4.498E-06 -6.20E-06
MAPE_top_testing 9.46E-07 1.07E-06
Cp_bottom_testing 1 1
Cp_top_testing 1 1

Actual and simulated plot for top composition Actual and simulated plot for top composition
n-butane: validation n-butane: validation
0.127 0.075
Actual
Top composition (model fraction):

Top composition (model fraction):

Actual
Simulated 0.07 Simulated
0.126
0.065
0.125
0.06
validation

validation

0.124 0.055

0.123 0.05

0.045
0.122
0.04
0.121
0.035

0.12 0.03
0 10 20 30 40 50 60 0 10 20 30 40 50 60

Time Time

Fig. 25: Actual and simulated n-butane at the top composition Fig. 26: Actual and simulated n-butane at the bottom
for validation. composition for validation.

170
Iran. J. Chem. Chem. Eng. Online Composition Prediction of a Debutanizer Column ... Vol. 36, No. 2, 2017

Actual and simulated plot for top composition Actual and simulated plot for top composition
n-butane: Testing n-butane: Validation
0.22 0.22
Online
Top composition (model fraction):

Online

Top composition (model fraction):

0.2 Neural 0.2 Neural
0.18 0.18
0.16
0.16
0.14

validation
Testing

0.14
0.12
0.12
0.1
0.1
0.08
0.08
0.06
0.06 0.04
0.04 0.02
0 20 40 60 80 100 120 0 20 40 60 80 100 120

Time Time

Fig. 27: Actual and simulated n-butane at the top composition Fig. 29: Actual and simulated n-butane at the top composition
closed loop for testing. close loop for validation.

Actual and simulated plot for top composition Actual and simulated plot for top composition
n-butane: Testing n-butane: Validation
0.09 0.1
Top composition (model fraction):

Online Online
Top composition (model fraction):

0.085 Neural Neural

0.08 0.09

0.075
0.08
0.07
validation
Testing

0.065 0.07
0.06
0.055 0.06

0.05
0.05
0.045
0.04 0.04
0 20 40 60 80 100 120 0 20 40 60 80 100 120

Time Time

Fig. 28: Actual and simulated n-butane at the bottom Fig. 30: Actual and simulated n-butane at the bottom
composition close loop for testing. composition close loop for validation.

of n-butane with partition open loop data and extracted architecture for online validation is similar as shown in
close loop data. From the statistical analysis, the Table 8. The actual composition for n-butane is obtained
extracted close loop performs better than the open loop from the laboratory measurement in the refinery plant.
response because RMSE value is low, CDC is high, AIC The required data collected surrounding the column
and BIC is low and MAPE is low. consists of the closed loop data. The purpose for validate
online is to monitor the composition for its precision and
Validate online based on close loop data accuracy. The training data are trained until the epoch
Validate online prediction for the n-butane is and the performance has achieved and then the network
implemented to estimate the composition of n-butane is validated and tested.
as the output for the column. The data of the composition Figs. 27-30 show the top and bottom compositions
prediction are partitioned to three sets, which are training, prediction of n-butane for validation and testing.
validation and testing. Only the validation and testing The results indicate that the RMSE is low at 3.73  10-9
sets are shown in this paper. The neural network for the bottom composition validation data set and 7.16  10-9

171
Iran. J. Chem. Chem. Eng. Nasser Mohammed Ramli et al. Vol. 36, No. 2, 2017

Table 10: Statistical analysis for composition n-butane for the top composition validation data set. On the other
validation and testing based on closed loop data. hand the CDC is high for the top and bottom
Parameter Online compositions for validation at 98.16. The high CDC
values indicate Di is equal to 1 based on the prediction
rmse_bottom_validation 3.73E-09 is larger than Di which is zero. The CDC is also high
rmse_top_validation 7.16E-09
for testing data sets. The regression value of R is 1, thus
the prediction between the actual and simulated is similar for
CDC_bottom_validation 98.16 validation and testing data sets. The AIC and BIC values
for online top composition validation are -505 and -494
CDC_top_validation 98.16
respectively. The AIC and BIC values for online bottom
R_bottom_validation 1 composition validation are -796 and -785 respectively.
For the AIC and BIC, low value is preferred as it
R_top_validation 1
indicates better prediction also applies to testing data set.
AIC_bottom_validation -796.17 The MAPE values for top and bottom composition
validation online are 5.53 10-6 and 5.93  10-6
AIC_top_validation -505.38
respectively. The MAPE should be close to 0 and this
BIC_bottom_validation -785.36 also applies to testing data set. Cp values are close to 1.
The Cp value for validation and test set for bottom and
BIC_top_validation -494.58
top composition are calculated to be 1. Table 10 shows
MAPE_bottom_validation 5.93E-06 the statistical analysis for the n-butane composition
prediction validation and testing based on close loop data.
MAPE_top_validation 5.53E-06
Figs. 31-32 show the residual analysis of n-butane for
Cp_bottom_validation 1 testing and validation data set for top and bottom
composition. From the analysis it can be concluded that
Cp_top_validation 1
the residual is very small and hence the prediction
rmse_bottom_testing 4.20E-09 obtained for the neural network estimation is highly accurate.

rmse_top_testing 8.06E-09 CONCLUSIONS

This paper presents a case study in utilizing a neural
CDC_bottom_testing 98.14
network model to estimate the top and bottom
CDC_top_testing 98.14 compositions of a debutanizer column. The use of online
close loop data, open loop data and simulation data
R_bottom_testing 1
makes the model robust and highly suitable for online
R_top_testing 1 use. PCA and PLS analyses have also been found
to facilitate the correct and right inputs of the variables
AIC_bottom_testing -750.06 for the column since a large amount of data are available
AIC_top_testing -461.92 from the industry.
The neural network prediction of n-butane gives high
BIC_bottom_testing -739.30 accuracy and the error between the prediction and actual
BIC_top_testing -451.160
composition is small and this will ensure that neural
network can also be used as an inferential estimator for
MAPE_bottom_testing 6.65E-06 composition estimation online. From the statistical
analysis for top and bottom compositions n-butane,
MAPE_top_testing 6.23E-06
it indicates the RMSE is small for extract from close loop
Cp_bottom_testing 1 compared to open loop response. Therefore the extract
from close loop perform is better than the open loop
Cp_top_testing 1
response.

172
Iran. J. Chem. Chem. Eng. Online Composition Prediction of a Debutanizer Column ... Vol. 36, No. 2, 2017

Residual analysis n-butane testing set Residual analysis n-butane validation set
0.00E+00 0.00E+00
0 20 40 60 80 100 0 20 40 60 80 100
-1.00E-09
Residual (mole/fraction)

-2.00E-09

Residual (mole/fraction)
-2.00E-09

-4.00E-09 -3.00E-09
-4.00E-09
-6.00E-09 -5.00E-09
-6.00E-09
-8.00E-09
-7.00E-09

-1.00E-08 -8.00E-09
Test top Test bottom -9.00E-09 Validation top Validation bottom
-1.20E-08 -1.00E-08
Time (min) Time (min)

Fig. 31: Residual analysis n-butane test data set. Fig. 32: Residual analysis n-butane validation data set.

Acknowledgment Received : Jun. 1, 2014 ; Accepted : Jun. 20, 2016

The authors would like to acknowledge PETRONAS
for providing the required data and information for the
REFERENCES
research and Universiti of Malaya for providing the
[1] Seader J.D., Header E.J., “Separation Process
research grant. (PS107/2010B)
Principle” John Wiley and Sons, 2nd edition (2006).
[2] Guptal S., Ray S., Samanta A.N., Nonlinear Control
Nomenclature of Debutanizer Column Using Profile Position
At Actual value Observer, Computers and Chemical Engineering,
xmeamsured Measure value 33: 1-36 (2008).
Cp Person correlation co-efficient [3] Hori E.S., Skogestad S., Selection of Control
xpredicted Predicted value Structure and Temperature Location for Two
Di Product yi  yi Product Distillation Columns, Chemical Engineering
yi Difference actual and average actual Research and Design, 85: 293-306 (2007).
Ea Actual value [4] Peng B., Li Xi., Sheng M., Song S., Zhao G., Li Xi.,
Ep Predicted value Study of Dual Temperature Control Method on
Cyclic Total Reflux Batch Distillation, Chemical
yi Difference predicted and average predicted
Engineering and Processing, 46: 769-772 (2007).
Ea Average actual value [5] Luyben W.L., Evaluation of Criteria for Selecting
Temperature Control Trays in Distillation Columns,
Ep Average predicted value
Journal of Process Control, 16: 115–134 (2006).
2 Variance [6] Enagandula S., Riggs James B., Distillation Control
Ft Predicted value Configuration Selection Based on Product
yt Data set certain value Variability Prediction Control Engineering Practice,
K Number of free model parameters 14: 743-755 (2006).
ft Associated model value [7] Liu X., Kruger U., Littler T., Xie L., Wang S.,
MSE Mean square error Moving Window Kernel PCA for Adaptive
y Mean of observed data Monitoring of Nonlinear Processes, Chemometrics
N Number of observation and Intelligent Laboratory Systems, 96: 132-143
R2 R squared (2009).

173
Iran. J. Chem. Chem. Eng. Nasser Mohammed Ramli et al. Vol. 36, No. 2, 2017

[8] Zamprogna E., Barolo M., Seborg Dale E., Optimal [21] Wan J., Huang M., Ma Y., Guo W., Wang Y., Zhnag H.,
Selection of Soft Sensor Inputs for Batch Distillation Li W., Sun X., Prediction of Effluent Quality of a
Columns Using Principal Component Analysis, Paper Mill Wastewater Treatment Using Adaptive
Journal of Process Control, 15: 39-52 (2005). Network-Based Fuzzy Inference System, Applied
[9] Chen G., McAvoy Thomas J., Piovoso Michael J., Soft Computing, 11: 3238–3246 (2011).
A Multivariate Statistical Controller for Online [22] Sharma R., Singh K., Singhal D., Ghosh R., Neural
Quality Improvement, Journal of Process Control, Network Application for Detecting Process Faults in
8: 139-149 (1998). Packed Towers, Chemical Engineering and
[10] L Smith Cecil, “Industrial Process Control” Processing, 43: 841–847 (2004).
Proceedings of the workshop American Institute of [23] Eriksson L., Johansson E., Kettaneh-Wold N.,
Chemical Engineers, 70-75 (1979). Trygg J., Wilstrom C., Wold S., “Multi and
[11] Singh V., Gupta I., Gupta H.O., ANN Based Megavariate Data Analysis Part I Basic Principles
Estimator for Distillation-Inferential Control, Chemical and Applications”, (Umetrics Academy 2nd edition,
Engineering and Processing, 44: 785–795 (2005). (2006).
[12] Fortuna L., Graziania S., Xibilia M.G. , Soft Sensors
for Product Quality Monitoring Debutanizer
Distillation Columns, Control Engineering Practice,
13: 499-508(2005).
[13] Singh V., Gupta I., and Gupta H.O., ANN-Based
Estimator for Distillation Using Levenberg-
Marquardt Approach, Engineering Applications of
Artificial Intelligence, 20: 249–259(2007).
[14] Baratti R., Corti S., A Feedforward Control Strategy
for Distillation Column, Artificial Intelligence in
Engineering, 11: 40-412 (1997).
[15] Zilochian A., Bawazir K., “Chapter 7: Application of
Neural Network in Oil Refineries ” CRC Press (2001).
[16] Mujtab I M, Hussain M A., Optimal Operation of
Dynamic Processes Under Process-Model Mismatches:
Application to Batch Distillation, Computers
Chemical Engineering, 22: 621-624 (1998).
[17] Greaves M A, Mujtaba I.M., Barolo M., Trotta A.,
Hussain M.A., Neural Network Approach to
Dynamic Optimization of Batch Distillation :
Application to Middle-vessel Column, Trans
IChemE, 81: 393-401 (2003).
[18] Seborg Dale E., Edgar Thomas F., Mellichamp
Duncan A., “Process Dynamic and Control”, John
Wiley and Sons, 2nd edition (2003).
[19] Dwyer Aidan O., “Handbook PI and PID Controller
Tuning Rules” World Scientific (2003).
[20] Mohamed Ramli Nasser., Hussain M. A., Mohamed
Jan Badrul., Abdullah Bawadi., Composition
Prediction of a Debutanizer Column using Equation
Based Artificial Neural Network Model,
Neurocomputing, 131: 59-76 (2014).

174