See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/260211393
Composition Prediction of a Debutanizer Column
using Equation Based Artificial Neural Network
Model
ARTICLE in NEUROCOMPUTING · MAY 2014
Impact Factor: 2.08 · DOI: 10.1016/j.neucom.2013.10.039
CITATIONS
READS
3
205
4 AUTHORS, INCLUDING:
Mohd azlan Hussain
Badrul Mohamed Jan
196 PUBLICATIONS 1,498 CITATIONS
51 PUBLICATIONS 136 CITATIONS
University of Malaya
SEE PROFILE
University of Malaya
SEE PROFILE
Bawadi Abdullah
Universiti Teknologi PETRONAS
41 PUBLICATIONS 22 CITATIONS
SEE PROFILE
All in-text references underlined in blue are linked to publications on ResearchGate,
letting you access and read them immediately.
Available from: Badrul Mohamed Jan
Retrieved on: 14 January 2016
Neurocomputing 131 (2014) 59–76
Contents lists available at ScienceDirect
Neurocomputing
journal homepage: www.elsevier.com/locate/neucom
Composition Prediction of a Debutanizer Column using Equation Based
Artificial Neural Network Model
Nasser Mohamed Ramli a,b, M.A. Hussain b,c,n, Badrul Mohamed Jan b, Bawadi Abdullah a
a
b
c
Chemical Engineering Department, Universiti Teknologi PETRONAS, Bandar Seri Iskandar, 31750 Tronoh, Perak, Malaysia
Chemical Engineering Department, Faculty of Engineering, University of Malaya, 50603 Kuala Lumpur, Malaysia
UMPDEC, University of Malaya, Malaysia
art ic l e i nf o
a b s t r a c t
Article history:
Received 27 May 2013
Received in revised form
17 September 2013
Accepted 28 October 2013
Communicated by J. Zhang
Available online 7 January 2014
Debutanizer column is an important unit operation in petroleum refining industries. The design of online
composition prediction by using neural network will help improve product quality monitoring in an oil
refinery industry by predicting the top and bottom composition of n-butane simultaneously and
accurately for the column. The single dynamic neural network model can be used and designed to
overcome the delay introduced by lab sampling and can be also suitable for monitoring purposes. The
objective of this work is to investigate and implement an artificial neural network (ANN) for composition
prediction of the top and bottom product of a distillation column simultaneously. The major contribution
of the current work is to develop these composition predictions of n-butane by using equation based
neural network (NN) models. The composition predictions using this method is compared with partial
least square (PLS) and regression analysis (RA) methods to show its superiority over these other
conventional methods. Based on statistical analysis, the results indicate that neural network equation,
which is more robust in nature, predicts better than the PLS equation and RA equation based methods.
& 2014 Elsevier B.V. All rights reserved.
Keywords:
Statistical analysis
Neural network
Partial least square analysis
Regression analysis
Debutanizer column
1. Introduction
Distillation column is considered one of the most common unit
operations in the chemical industry. However, its complex behaviour and highly un-predictive nature, has made it as a unit
operation which is complicated and difficult to handle by engineers [1]. Hence it becomes more important to attain the desired
purity of products by manipulating the top and bottom composition of the distillation column accurately. In order to maintain and
control the composition at its optimum value, it is necessary to
predict it with high accuracy and precision, simultaneously with
fast response. Chemical process industries also encounter a lot of
problem in monitoring the debutanizer column. Open loop
instability issues, non-linearity, multivariable issues and the difficulty to measure a certain variable directly are the key factors
complicating the composition prediction. The composition at the
top and bottom respectively for the column is currently measured
using normal laboratory sampling which is tedious and time
consuming. It has been found that the computing time for
composition prediction monitoring by neural network is fast and
accurate compared to normal laboratory while in the industry it
n
Corresponding author at: Chemical Engineering Department, Faculty of
Engineering, University of Malaya, 50603 Kuala Lumpur, Malaysia.
E-mail address: mohd_azlan@um.edu.my (M.A. Hussain).
0925-2312/$ - see front matter & 2014 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.neucom.2013.10.039
normally takes one day to measure the composition by laboratory
sampling. In this context, the need for software-based online
analyzer to provide the speed and accuracy for its measurement
has become incumbent and this research deals with the prediction
of the composition online using equation based artificial neural
network models, and compared with partial least square and
regression models.
In relation to the use of online sensors, an adaptive soft sensor
for online monitoring of melt index (MI), an important variable
determining the product quality in the industrial propylene polymerization (PP) process, has been proposed by Zhang and Liu [2].
The fuzzy neural network (FNN) served as the basic model for its
nonlinear approximation ability using its learning method. To
overcome the difficulty of structure determination of the FNN,
an adaptive fuzzy neural network (A-FNN) is subsequently developed to determine the number of fuzzy rules, where a novel
adaptive method dynamically changes the structure of the model
by the predefined thresholds. In order to get better generalization
ability of the soft sensor, support vector regression (SVR) is
introduced for parameter tuning, where the output function is
transformed into an SVR based optimization problem. The soft
sensors including the SVR, FNN–SVR and A-FNN–SVR models are
compared in detail and the proposed soft sensor achieves good
performance in the industrial MI prediction process.
Three soft sensor models involving radial basis function (RBF),
support vector machine (SVM), and independent component
60
N. Mohamed Ramli et al. / Neurocomputing 131 (2014) 59–76
Nomenclature
actual value
At
xmeamsuredmeasure value
person correlation co-efficient
Cp
xpredicted predicted
product yi ! yi
Di
yi
difference actual and average actual
Ea
actual value
Ep
predicted value
analysis–support vector machine (ICA–SVM)] methods has been
developed by Yan and Liu [3]. The process is to infer the Chemical
Oxygen Demand (COD) of the quench water produced from the
pesticide waste incinerator. An optimization model of COD is
further proposed based on a fore mentioned soft sensor models.
The chaos genetic algorithm is introduced to solve the optimization model. A novel soft sensor model with principal component
analysis, radial basis function neural network (RBF) and multi scale
analysis (MSA) has been proposed by Shi J and Liu [4]. The purpose
is to infer the melt index of manufactured products from real
process variable, where PCA is carried out to select the important
relevance process features and to eliminate the correlations of
input variable, the MSA is to introduce much more information
and to reduce the uncertainty of the system, and RBF networks are
used to characterize the nonlinearity of the process. A black-box
modeling scheme to predict melt index (MI) in the industrial
propylene polymerization process has also been developed by Liu
and Zhao [5]. MI is one of the most important quality variables
determining product specification and influenced by a large
number of process variables. In their work a faster statistical
modeling method has been proposed to predict MI online which
involves fuzzy neural network, particle swarm optimization (PSO)
algorithm, and online correction strategy (OCS).
Furthermore an adaptive soft sensor based on systematic process
key variables has also been proposed for inferential control using
derived adaptive model by Ma Ming et. al. [6]. The key variables are
based on statistical approach of stepwise linear regression. The
online plant measurements are selected as key features to estimate
tardily-detected variables. The parameters of the linear inferential
model are adapted as the online and offline data which are available.
In order to improve the numerical characteristics of the algorithm,
square root filter is used due to the multi-collinearity problem
involved. The soft sensor has been implemented to an o-xylene
purification column. The inferential model predicts accurately the
real plant data which is useful for industrial application in the
distillation column. The statistical stepwise regression technique
was used to infer fast- measuring variables to some key variables
so that the model is easy to maintain. By introducing the concept of
adaptation, the model structure would reflect the current operation
of the plant and the accuracy of the soft sensor could be improved.
In this respect, Artificial Neural Network (ANN) offers as an
alternative powerful and fast tool to model non-linear processes
such as the debutanizer column and which can be utilized as an
efficient soft sensor. ANN has the ability to learn the relationship
between the outputs and the inputs for a system. To develop a
process using ANN, it requires suitable network architecture and
appropriate training data. The literature reported some work on
debutanizer column modeling using neural network. For example,
a nonlinear state space model is used for representing the inputs
and outputs and singular value decomposition (SVD) is used to
remove redundant nodes and model reduction in the work of
Prasad and Bequette [7].
yi
Ea
Ep
s2
Ft
K
MSE
N
R2
T
difference predicted and average predicted
average actual value
average predicted value
variance
predicted value
number of free model parameters
mean square error
number of observation
R squared
number of parameters
The design of dynamic neural network soft sensors to improve
product quality in a debutanizer column has also been reported
using a three step predictive method to evaluate its top product
concentration by Fortuna and co-workers [8]. The approach uses
lagged values of the input and composition in the neural network
prediction. Real time estimation of plant variables such as the
composition are used for monitoring purposes and the number of
neurons in the hidden layer for the neural network was determined by trial and error. The ANN estimator based on Levenberg–
Marquardt (LM) algorithm has been used because it has been
tested for binary as well as multi-component mixture by Singh
and co-workers [9]. The LM algorithm suits very well to both cases
and gives more accurate and sensitive results compared to
Steepest Descent Back Propagation (SDBP) algorithm. For a complex chemical plant having hundreds of parameters, LM approach
work efficiently. By using these parameters, the quality of the
product could be estimated and corrective actions are taken
simultaneously. ANN has also been utilized widely in crude
fractionation section in the oil refinery industry where the output
neural network prediction is the naphtha temperature rather than
composition prediction by Zilochian and Bawazir [10]. Neural
network has in reality been used for a number of chemical
engineering applications involving sensor analysis, fault detection
and nonlinear process control both in simulation and online
implementation, as reported in the literature by Hussain [11].
Partial least square regression (PLSR) together with artificial
neural network (ANN) with back propagation (BP) algorithm has
also been proposed by Xuefeng [12]. The neural networks were
trained to extract the quantitative information from the training
samples for a preflash tower. Hybrid Artificial Neural Network
(HANN) was employed to develop the naphtha dry point soft
sensor which is the most important intermediate product concentration soft sensor in the p-xylene (PX) oxidation reaction. An
optimization framework to obtain optimal operation of the
dynamic processes under process-model mismatches has been
developed by Mujtaba and Hussain [13]. In order to model these
mismatches, neural network have been utilized in the batch
distillation process for a binary batch distillation with only one
specified product. In another work by Greaves and co-workers, a
framework has been proposed to optimize the operation of batch
system and utilize an artificial neural network (ANN) based
process model in the optimization of the pilot-plant middle-vessel
batch column [14]. The maximum-product problem is formulated
and solved by optimizing the column operating parameters, such
as the batch time, reflux and reboil ratios. The ANN based model
was capable of reproducing the actual plant dynamics with good
accuracy, and allows a large number of optimization studies to be
carried out with little computational effort.
Partial Least Square (PLS), an extension of PCA provide model
parameters with diagnostic tools where by increasing the number
of X variables, it could improve the precision of the PLS model [15].
In the literature there also exists some modeling work of
N. Mohamed Ramli et al. / Neurocomputing 131 (2014) 59–76
a debutanizer column using PLS. For example, dynamic partial
least square regression is used in the inferential model for
composition prediction in a multicomponent distillation column
by Kano et. al [16]. Past sampling times measurement are used as
input variables to interpret the dynamic process. PLS was also used
to predict the composition profile in a simulated batch distillation
column by Zamprogna et. al. [17]. The inputs are temperature
measurements and the output is the composition in the distillate
and bottom streams. The estimator performance is evaluated
based on the pre-processing of the calibration and validation data
sets. The number of measurements used as sensor inputs, consist
of lagged measurements. A simple augmentation of the conventional PLS regression approach is based on the development and
sequential use of multiple regression models.
A soft sensor for a chemical process using PLS that could handle
correlations for a number of process variables and nonlinearities
based on the smoothness concept has also been proposed by Park
and Han [18]. The proposed method was to build a soft sensor for a
distillation column based on multivariate smoothing by using local
weighted regression. There were two different type of cases
applied for the distillation column which are the nonlinear and
linear behavior and use for online measurement to estimate the
important variables such as temperature and composition. Process
monitoring using modified PLS through an independent component analysis (ICA) approach has also been developed by Zhang
Yingwei and Zhang Yang [19]. The method make use of the kernel
to the ICA-PLS to solve the non-linearity in the data set and the
original algorithm are modified by giving the regression coefficient
matrix and residual matrix to the ICA-PLS to reduce computation
time. An application of PLS as a soft sensor has been developed to
predict the melt flow index using measured process variable for an
industrial autoclave reactor by Sharmin and co-workers [20].
Detailed first principle model for free radical polymerization is
not an easy task since there are large reactions and kinetic
parameters involved. Multivariate regression model are used to
solve this problem and the melt index can be successfully
predicted using these statistical tools.
A multivariate statistical soft sensor for online estimation of
product quality in an industrial batch polymerization process has
also been proposed by Facco et. al.[21]. For each estimation, PLS
sensors are designed, and their performance is evaluated against
actual plant data. The estimation are evaluated by augmenting
the process variable with lagged measurement. The projection
method, using PLS regression are used to design a soft sensor for
the online estimation of the resin quality properties. Multivariate
statistical (MVS) techniques have been proven to be an excellent
tool for analyzing and monitoring of processes where the process
data are huge. Online soft sensor was proposed by using three
different methods in terms of just in term learning (JITL) which are
based on PLS, support vector regression (SVR) and least squares
support vector regression (LSSVR) by Ge and Song [22]. The real
time performance strategy is to enhance the online efficiency of
the JILT based soft sensor for a distillation column. The JILT
methods are suitable for real time performance. The modeling
efficiency of SVR is not difficult because it only requires a quadratic
programming optimization and the efficiency could be improved
by the LSSVR.
A least squares support vector machines (LS-SVM) soft-sensor
model of propylene polymerization process has been developed by
Shi and Liu to infer the MI of polypropylene [23]. Considering the
use of cost function without regularization might lead to less
robust estimates, the weighted least squares support vector
machines (weighted LS-SVM) approach for the propylene polymerization process is further proposed to obtain a robust estimation of the melt index. Reliable estimation of melt index (MI) for
the production of polypropylene has also been proposed by Shi
61
and coworkers [24]. Propylene polymerization process is highly
nonlinear and characterized by multi-scale nature with huge
number of variables and information which are highly correlated
and derived at different sample rates from different sensors. A
novel soft-sensor architecture based on radial basis function networks (RBF) combining independent component analysis (ICA) as
well as multi-scale analysis (MSA) is proposed to infer the MI of
polypropylene from other process variables.
A RBF (radial basis function) neural network soft-sensor model
for the polypropylene process has been developed by Li and Liu to
infer the MI from a number of process variables [25]. Since the PP
process is complicated for the RBF neural network with a general
set of parameters, a new ant colony optimization (ACO) algorithm,
N-ACO, and its adaptive version, A-N-ACO, which aimed to
optimize the structure parameters of the RBF neural network,
respectively. An optimal soft sensor, named the least squares
support vector machines with Ant Colony-Immune Clone Particle
Swarm Optimization (AC-ICPSO-LSSVM), has also been proposed
by Jiang and coworkers which combines the advantages of the
high accuracy of LSSVM and the fast convergence of PSO [26].
Furthermore, the immune clone (IC) method is introduced into the
PSO algorithm to make the particles of ICPSO diverse and enhance
global search capability for avoiding the premature convergence
and local optimization of the conventional PSO algorithm.
Another novel chemical soft-sensor approach for the prediction
of the melt index (MI) in the propylene polymerization industry
has been developed by Jiang and co-workers using accurate
optimal predictive model of the MI values with the relevance
vector machine (RVM) method [27]. The RVM is employed to build
the MI prediction model and a modified particle swarm optimization (MPSO) algorithm is introduced to optimize the parameter of
the RVM, after which the MPSO-RVM approach is developed. An
online correcting strategy (OCS) is further carried out to update
the modeling data and to revise the model’s parameter selfadaptively whenever model mismatch happens.
In this paper we demonstrate the use of a single ANN to predict
the composition of n-butane for the top and bottom of a debutanizer column simultaneously and compare it with predictions
using PLS and regression analysis. One of the significant and novel
contribution of this work is the use of an equation based neural
network model whereas other works, mention previously, utilize
neural network as a black box model only. The use of an equation
based neural network is more reliable and robust than the
conventional method and at the same time gives better prediction
than the other methods such as the PLS and regression analysis.
This equation based approach is also a concrete, fast and practical
way of utilizing neural network models as a soft sensor for this
system. Furthermore, we utilize a combination of online data both
open loop and closed loop as well as simulated data and further
validate these data using the closed loop system. This further
enhanced the reliability and online capability of the NN model
when applying it online as a software sensor.
The paper is organized in various sections. Section 2 contains
the description of the column and plant, and Section 3 describes
the theoretical background and Section 4 describes the methodology for the online composition prediction. Finally Section 5 is the
overall analysis for the online composition prediction.
2. Description of Crude Oil Processing Plant and Debutanizer
Column
The crude oil processing plant as seen in Fig. 1, consists of a
refinery process, condensate fractionation and reforming aromatics section. The feedstocks of the refinery process are mainly
crude oil while the products are petroleum products, liquefied
62
N. Mohamed Ramli et al. / Neurocomputing 131 (2014) 59–76
Fig. 1. Block diagram for the oil refinery industry.
petroleum gas, naphtha and low sulphur waxy residue. The
refinery has two main process units, which are Catalytic Reforming
Unit (CRU) and Crude Distillation Unit (CDU). The Crude Oil
Terminal provides the feedstock and the crude oil is preheated
using heat exchangers within the range of 190 1C – 210 1C. It is
then further heated in a furnace to 340 1C – 342 1C before being
routed to the CDU. The crude oil is separated into a number of
fractions, which are heavy Straight Run Naphtha as overhead
vapour, untreated kerosene, straight run kerosene and straight
run diesel. From the crude tower, there are 3 sides cut streams,
which are drawn to a stripper column and the stripper consists of
a kerosene stripper, naphtha stripper and diesel stripper.
From the CDU, the pretreater feed Heavy Straight Run Naphtha
(HSRN) is mixed with hydrogen from the reformer and heated up
to the reaction temperature using a heater and fed into the
pretreater catalytic reactor. The reactions involved are denitrification and desulphurization, which will protect the reformer catalyst
from poisoning. The product from the reactor is transferred to the
pretreater stripper while the feed to the reforming unit is the
bottom product of the stripper and the feed to the reformers
reactors is the treated naphtha, which is heated to the reaction
temperature. Effluent from the reactor is collected in a reformer
separator where it is cooled. Some portion of the gas which is
separated, is recycled to the reactor feed stream while the other
portion is transferred to an absorber. In the absorber, at the raw
naphtha feed, hydrogen gas is purged and recycled to the pretreater heater. The feed into the LPG absorber is liquid phase
where it is drawn off and the liquid fraction is pumped into a
stabiliser. Before being sent to storage, reformate is withdrawn
from the stabiliser bottom for cooling. From the stabiliser reflux
drum, overhead vapours from the stabiliser are cooled, condensed
and recovered.
The debutanizer column is the main column for producing the
main product, which is the liquefied petroleum gas. The debutanizer column is located at the CDU section depicted top right in
Fig. 1. The unit is used to recover light gases and LPG from the
overhead distillate before producing light naphtha. The light gases
Table 1
Column specification.
Number of tray of the column
Feed tray - stage number
Type of tray used
Column diameter
Column height
Condenser type
Feed mass flowrate
Feed temperature
Feed pressure
Overhead vapor mass flowrate
Overhead liquid mass flowrate
Condenser pressure
Reboiler pressure
35
23
Valve
1.3 m
23.95 m
Partial
44106 kghr " 1
113 1C
823.8 kPa
11286 kghr " 1
5040 kghr " 1
823.8 kPa
853.2 kPa
mainly C2 is used to refine fuel gas and mixed with LPG. The feed
to the debutanizer column which has 35 valve trays, is from the
Deethanizer bottom product. The debutanizer condenser condenses the overhead vapor and the debutanizer overhead pressure
control valves with two split ranges controls the overhead system.
The reflux from the top of the debutanizer consists of the collected
condensed hydrocarbon while reboiler section is used to strip
lighter hydrocarbon.
There are three manipulated variables for the column which are
the feed flow rate, reflux flow rate and reboiler flow rate. The feed
flow rate controls the feed to the column, the debutanizer reboiler
control valve controls the reboiler temperature while the debutanizer bottom level controller controls the bottom product level. The
debutanizer reflux control valve controls the ratio of the liquid and
distillate flow rate at the top of the column. This column is a
challenging process because it deals with non-linearity, is a highly
multivariable process, involves a great deal of interactions between
the variables, has lag in many of the control system, all of which
makes it difficult system to be modeled by linear techniques. Hence
non-linear methods such as the neural network equation based
model is highly appropriate for this process. Table 1 outlines the
63
N. Mohamed Ramli et al. / Neurocomputing 131 (2014) 59–76
column specification while Table 2 describes in detail all the
variables surrounding the column. The measured variables are the
Feed flow, Pressure 1 (Debutanizer receiver overhead pressure),
Flow 2 (LPG flow to storage), Flow 1 (Light Naphtha flow to storage),
Level 2 (Debutanizer condenser level), Level 1 (Debutanizer level)
and Temp 5 (Reboiler outlet temperature to column). The top and
bottom compositions of the column are currently measured using
laboratory sampling by gas chromatography. Fig. 2 shows the
column configuration of the debutanizer column under study in
this work.
3. Theoretical background
Artificial Neural Network (ANN) is a popular and reliable tool
when dealing with problems involving prediction of variables in
engineering problems at the present age [14]. It comprises a great
number of interconnected neurons that consists of a series of
layers with a number of nodes. Every node receives a signal from
the network link and the signal is added together before being
applied to a specific transfer function to produce the output. The
signal from the output will be sent to other node until it reaches
the network output. Nodes called neuron are the basic processors
Table 2
Description of the variables for the column.
Tag
Description
Units
Temp 1
Temp 2
Temp 3
Temp 4
Temp 5
Temp 6
Level 1
Level 2
Level 3
Level 4
Flow 1
Flow 2
Pressure 1
Debutanizer top temperature
Debutanizer bottom temperature
Debutanizer receiver bottom temperature
Light Naphtha temperature after condenser E 1
Reboiler outlet temperature to column
Debutanizer feed temperature
Debutanizer level
Debutanizer condenser level
Debutanizer level indicator
Condenser level indicator
Light Naphtha flow to storage
LPG flow to storage
Debutanizer receiver overhead pressure
1C
1C
1C
1C
1C
1C
%
%
%
%
m3/hr
m3/hr
kPa
of neural network. Each connection between two nodes with a real
value is called weight and the values of the weights are obtained
by training a set of input and output correlations. The weights are
adapted by the learning rule and it has long-term memory for the
network.
The advantage of ANN is in their ability to be used as an
arbitrary function approximation mechanism that learns from
observed data. However, using them is not so straightforward
and a relatively good understanding of the underlying theory is
essential. One of the main criteria is the choice of model and this
will depend on the representation of data and its application. The
second criteria is the learning algorithm where there are numerous trade-offs regarding these algorithms. Furthermore selecting
and tuning an algorithm for training on unseen data requires a
significant amount of experimentation to ensure the robustness of
the selected model. If the model, cost function and learning
algorithm are selected appropriately, the resulting ANN can be
extremely robust and gives the correct implementation. It can be
used naturally in online learning and large data set applications.
However the main argument against the widespread use of the
neural network is that it is a black box model and can only be
represented by the NN structure and difficult to be represented by
algorithmic equations which are cumbersome in nature. In this
work, it can be shown that by the appropriate use of the activation
functions and with proper pruning of the weights, an equation
based neural network model can be obtained to be used in the
prediction for the column compositions.
The general equation for the output from the neural network
can be given as (for a 3 layer network)
i
i
i
1
2
3
y ¼ f ðLW 3;i f ðLW 2;i f ðIW 1;i p þ b Þ þ b Þ þb Þ
IW 1;i ¼ input weight at layer 1 (input layer)
b1 ¼ bias values at layer 1
LW 2;i ¼ layer weight at layer 2 (hidden layer)
b2 ¼ bias values at layer 2
LW 3;i ¼ layer weight at layer 3 (output layer)
b3 ¼ bias values at layer 3
p¼ vector inputs to the neural network
Fig. 2. Debutanizer column configuration.
ð1Þ
64
N. Mohamed Ramli et al. / Neurocomputing 131 (2014) 59–76
y¼vector outputs from the neural network
i
f ¼activation function at layer i
This equation based neural network model is more robust and
stable as compared to the black box based model, frequently used
by researchers and practitioners and will be the highlight of our
research work in this paper.
PLS regression is a method that generalizes and combines
features from principal component analysis and multiple regressions. This is very useful in data analyses for system which are
collinear and have incomplete variables. The precision of PLS
model is a function of the number of input variables. This is often
useful in predicting a set of dependent variables (Y) from a large
set of independent variables or predictors (X). PLS has been proven
reliable in process monitoring and optimization prediction. PLS
interpretation could indicate matrix vector multiplication to a set
of bivariate regression. It provides the connection between two
operations in algebra matrix and statistics. PLS has the ability to
provide the foundation of a multivariable system. It could also
demonstrate projection models as long as there is a similarity
between the variables[15]. Based on PLS, the general regression
equation is given as
Y ¼ y þ XW nC þ F
ð2Þ
where y the variable average for Y, W nC are the loading weights
and F is the residual in Y.
The disadvantages of PLS with further increase in the size of the
data sets is that we will start to see inadequacies in these
multivariate methods, both in their efficiency and interpretability.
PLS coefficients are of interest because it could be simplified when
there are several components in the model but the disadvantages
of the coefficients for the PLS equation is that information
regarding the correlation structure among the response is
unknown.
Multivariate regression is the other conventional method used
to obtain the relationship between the input variables, X and the
output variable, Y. The Y can be predicted as a function of X by
using an equation in the following form given as,
Y 0 ¼ a þ bo X o þ b1 X 1 þ::: þ bn X n
Temp 5, Pressure 1 and composition at both ends of the column.
The simulated close loop response of the composition of n-butane
at the top and bottom of the column was also established to
compare with the online close loop data. The steady state for the
column needs to be developed before transition of the steady-state
to the dynamic state. Steady state simulations can be cast easily
into dynamic simulations by specifying additional engineering
details, including pressure/flow relationships and equipment
dimensions. The necessary information such as feed conditions,
feed compositions, reflux ratio, condenser pressure, reboiler pressure etc. have to be provided to the selected unit operation in the
simulation. The simulation data was performed using similar steps
test as in the plant to obtain the fluctuation of the process variable
under open loop response, where the manipulated variables are
reboiler and reflux flow rates.
The data generated for the process is taken for 541 minutes
with 1 minute sampling interval which amounts to a total data of
5410 as will be seen in later sections. These data that are available
from actual plant are large and therefore need to be screened by
performing principal component analysis (PCA) and partial least
square (PLS), where the important variables for the column
are obtained and are used for monitoring the composition of
n-butane. Table 2 outlines all variables surrounding the column.
For each of the step test, PCA is used to determine the important
variables surrounding the column. Once we have determined the
process variables, the important variables affecting the composition of n-butane is further analysed using PLS analysis. The raw
process data generated are scaled down between 0.05 to 0.95
using the following equation:!
"
actual value " min value
scaled value ¼
ð0:95 " 0:05Þ
max value " min value
þ min value
ð4Þ
Hence the actual value is then given by,
actual value
ðscaled value " min valueÞ ! ðmax value " min valueÞ
¼
ð0:95 "0:05Þ
þ min value
ð5Þ
ð3Þ
where Y’ is the predicted variable on the Y variable, a is the slope
representing the predicted change in Y for a one unit increase in Xo
[28]. The performance of regression analysis methods in practice
depends on the form of the data generating process, and how it
relates to the regression approach being used. Since the true form
of the data-generating process is generally not known, regression
analysis often depends to some extent on making assumptions
about this process. These assumptions are sometimes not testable
if a large amount of data to be utilised. Regression models for
prediction are still useful even when the assumptions are moderately violated, although they may not perform optimally. However,
the main disadvantage in many applications, of these regression
methods, is that it could give misleading results when causality
exists on the observation data.
4. Methodology
4.1. Model data generation
Although most online open loop response from the plant
surrounding the column is available, some of the variables in open
loop surrounding the column are not available. In this work,
dynamic simulation of a debutanizer column is performed using
the plant process simulator HYSYS to obtain the unavailable data
sets from the plant where the variables that are not available are
4.2. Neural network, Partial least square (PLS) and Regression
Analysis (RA) data sets
One of the objective of this work is to develop composition
predictions online using neural network, partial least square and
regression analysis. The composition at the top and bottom for the
column in the refinery is currently measured using normal
laboratory sampling. Therefore neural network, PLS and RA are
used as alternative online methods to predict the composition as
they are expected to produce more robust, stable and precise
results at a faster period.
Open loop responses of the reboiler and reflux data set, which
include the composition of n-butane, are used to develop the
dynamic neural network architecture. The selected input variables
to the network are time delayed including the composition of
n-butane since the models are dynamic in nature and the outputs
are the future predictions of n-butane. The numbers of past values
for each input variable are considered to be only 1. These past
values are determined by trial error method and it is found that
this past value for each variable gives the optimum performance
and also reduces the complexity of the dynamic model. The type of
dynamic network used for this case is the Nonlinear Autoregressive Network with Exogenous inputs (NARX) while the training
algorithm used is the Levenberg-Marquardt method. In addition,
the adaptation learning function with momentum is used and the
performance function evaluated is the mean square error criteria.
65
N. Mohamed Ramli et al. / Neurocomputing 131 (2014) 59–76
required inputs and outputs are fed to the regression analysis, it will
calculate the predicted output, the equation for RA and the residual
analysis. The regression is based on multivariate linear equation and
these input variables are generally shown in Eq. 3 in terms of the X
variable.
The data sets are partitioned into 2 sets which are classified as
training and validation sets with 65% data for training and 35% for
validation. The network training and validation are achieved by
using the mean square performance with specified number of
epoch (training cycle). The number of inputs to the network is 10
and the outputs are 2 and the transfer function is linear for all the
entire layers.
The architecture consists of 3 layers which are the input, hidden
and output layer. The weights and biases value used in the neural
network equation are obtained after training and validation of the
neural network. The hidden nodes are selected by trial and error
method. The neural network is trained with an initial guess of the
hidden nodes at 8 and then the number of hidden nodes is increased
by a factor of 2 till the hidden nodes achieves a value of 40. The Root
Mean Square Error (RMSE) is then monitored and the one with the
lowest RMSE value is selected for determining the final number of
hidden nodes. Fig. 3 shows the profile of RMSE with the change in the
number of hidden nodes in the hidden layer. Analysis of variance
(ANOVA) for NN is also done by using the Statistical Toolbox in
MATLAB using the F test statistics method. In this work, the number of
neurons which gives optimum predictions of the outputs is found to
be 10 nodes as seen in the Fig. 3.
Table 3 shows the important variables involved for the neural
network where the open loop responses of the reboiler flow rate and
reflux flow rate data set are obtained from plant and simulation. The
simulated data is the composition of n-butane and the rest of the
variables are obtained from actual plant data. The inputs for the neural
network are obtained from mv2(k) to p_bot(k-1) while the outputs are
the variable p_top(kþ1) and p_bot(kþ1) decided by data pretreatment using PCA and PLS as mention earlier.
Multivariate data are measured based on observations and
variables from the step tests in the input variables and the data
generated for PLS is similar to the data generated for NN. PLS
analysis are performed using the multivariate software called
SIMCA-P. There are 2 important variables classified which are
the primary variable and the observation variable. The primary
variable consists of 10 variables surrounding the column and the
observation variables are the top and bottom n-butane composition. Once the work set has been developed, the PLS model will be
fitted with the Partial Least Square equation and it involves the
loading weight and residual in terms of the composition of nbutane and average value of the composition of n-butane.
The data generated for Regression Analysis (RA) is also similar to
the data generated for NN and PLS. The data for regression are
analyzed using the data analysis tool in Excel. The important elements
of the RA modeling is the range of inputs and outputs of the data
analyzed where the confidence level is set at 95%. Once all the
4.3. Model adequacy test for NN, PLS and RA models
The performances and comparison of the predictions by the
different methods are determined using the Root Mean Square
Error method. (RMSE) given by;
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ðxmeasured " xpredicted Þ2
RMSE ¼
ð6Þ
N
Correct Directional Change (CDC) measures the accuracy of a
model in its prediction of the subsequent actual change of a
predicted variable. The formula of CDC is given below as;
CDC ¼
100 N
∑D
N i i
ð7Þ
where the formula of Di is defined as:
Di ¼ yi ! yi
The best known information criterion is the Akaike information
criterion (AIC) and Bayesian information criteria (BIC) which is
given below as;
AIC ¼ MSE þ s2
BIC ¼ MSE þ
2K
T
ð8Þ
log ðNÞs2 2K
T
ð9Þ
Table 3
Variables involved in the PLS analysis, regression analysis and neural network.
mv2 (k)
mv2 (k-1)
mv3 (k)
mv3 (k-1)
f (k)
f (k-1)
p_top (k)
p_top (k-1)
p_bot (k)
p_bot (k-1)
p_top (k þ1)
p_bot (kþ 1)
Manipulated reboiler flow rate
Lag mv2
Manipulated reflux flow rate
Lag mv3
Debutanizer feed temperature
Lag feed temperature
Top composition n-butane
Lag top composition
Bottom composition n-butane
Lag bottom composition
Future predictions top composition n-butane
Future predictions bottom composition n-butane
RMSE profile of n-butane
0.0012
0.001
RMSE
0.0008
0.0006
0.0004
0.0002
0
8
10
12
14
16
18
20
22
24
26
28
30
32
34
Number of hidden nodes
bottom
top
Fig. 3. Profile of the RMSE with respect to number of hidden nodes.
36
38
40
66
N. Mohamed Ramli et al. / Neurocomputing 131 (2014) 59–76
The coefficient of determination which also determines the
measure of fit is defined as below;
5.1. Step tests for reboiler flow rate
ð10Þ
1 N jF t " At j
! 100%
∑
N i ¼ 1 At
ð11Þ
∑Tt ¼ L ðyt "y t Þ2
Mean Absolute Percentage Error (MAPE) is measure of accuracy
in a fitted time series value, given by;
MAPE ¼
Pearson Correlation Coefficient (Cp), measures the goodness of
the regression fit: the closer the value to one indicate higher
accuracy as given below;
S
∑N
j ¼ 1 ðE p;j " E p;j ÞðE a;j " E a;j Þ
C p ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 NS
2
S
∑N
j ¼ 1 ðE p;j "E p;j Þ ∑j ¼ 1 ðE a;j "E a;j Þ
ð12Þ
Theil’s Inequality Coefficient (TIC), measures the model evaluation for the difference between output model and the actual
output is considered as the error given below;
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
^ 2
∑N
i ¼ 1 ðyi " yi Þ
TIC ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2
^2
∑N
∑N
i ¼ 1 yi þ
i ¼ 1 yi
ð13Þ
Based on the statistical analysis described above, the criteria for
its acceptable performance is decided on the deviation between
actual and composition prediction by NN, PLS and RA established
as follows; low RMSE, CDC approaching 100, small AIC and BIC, R2
approaching 1, low MAPE, CP approaching 1 and low TIC value. Eqs.
6–10 are obtained from [29], Eq. 11 are obtained from [30], Eq. 12
are obtained from [31] and Eq. 13 obtained from [24] for this work.
5.2. Online close loop composition validation and simulation
Figs. 8 and 9 represent the differences between online and
simulation of the top and bottom composition of the n-butane in
the column under normal operating condition. The calculated Root
Mean Square Error (RMSE) for the top and bottom composition is
0.0251 and 0.0082 respectively and the Mean Square Error for top
and bottom compositions is 0.00063 and 6.697 ! 10 " 5 respectively for n-butane. These result shows that there is a small
Step test Temp 1
145
Reboiler flow rate
(m3/hr)
Figs. 4–7 show some of the step tests of the reboiler flow rate data
sets. In order to generate the input-output data for the neural network
training, various step changes are applied to the inputs to obtain the
corresponding outputs in which the inputs for this system is the
reboiler flow rate. The step test of the reboiler flow rate which is one
of the manipulated variable are generated by using multi amplitude
rectangular pulse [32]. The step test is important to observe the effect
and the fluctuations of the process variable when performing changes
to the reboiler flow rate. The fluctuations of Temp 1, Flow 1 and
Pressure 1 (see Figs. 4–6) increases and decreases as the reboiler flow
rate changes as shown in these figures. Level 1 (see Fig. 7) has no effect
to the fluctuations as the step test of the reboiler flow rate changes
which indicates that level does not effect the composition of n-butane.
The step test for the reflux flow rate, the other manipulated variable, is
also done in the same way but only the step tests for the reboiler flow
rate are shown in this paper.
62
60
144
58
143
56
142
54
141
52
140
1
50
99
148
197
246
295
344
393
442
491
Temperature (0C)
bt Þ2
∑Tt ¼ L ðyt " y
50
540
Time (min)
Reboiler.Flow
Temp 1
Fig. 4. Temp 1 Debutanizer top temperature.
Step test Flow 1
Reboiler flow rate
(m3/hr)
145
144
143
142
141
140
1
50
99
148
197
246
295
344
393
Time (min)
Reboiler.Flow
Flow 1
Fig. 5. Flow 1 Light Naphtha flow to storage.
442
491
45
40
35
30
25
20
15
10
5
0
540
Flow rate (m3/hr)
R2 ¼ 1 "
5. Results and Discussion
67
N. Mohamed Ramli et al. / Neurocomputing 131 (2014) 59–76
Step test Pressure 1
850
144
800
143
750
142
700
141
650
140
1
50
99
148
197
246
295
344
393
442
491
Pressure (kPa)
Reboiler flow rate
(m3/hr)
145
600
540
Time (min)
Reboiler.Flow
Pressure 1
Fig. 6. Pressure 1 Debutanizer receiver overhead pressure.
80
144
75
70
143
65
142
60
141
Level (%)
Reboiler flow rate
(m3/hr)
Step test Level 1
145
55
140
1
50
99
148
197
246
295
344
393
442
491
50
540
Time (min)
Reboiler.Flow
Level 1
Fig. 7. Level 1 Debutanizer level.
Top composition n-butane
Composition (mole fraction)
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
0
2000
4000
6000
8000
10000
12000
Time (min)
simulation
online
Fig. 8. Top composition n-butane close loop.
deviation between the online and simulation data and the purpose
of the close loop response is to validate between the online and
simulation data. Once the close loop results has been verified with
the simulation results, then the open loop response for the
variables that is not available from the plant could be obtained
in simulation, based on the same step size of the manipulated
variable from the plant, which involve variables such as Temp 5,
Pressure 1 and composition. The combined data consisting of the
plant and simulation data are then used to developed the neural
network model, represented by the equations as will be shown in
the next section. Similar data sets are also used to generate the PLS
and regression models for comparison with the neural network
predictions for the top and bottom n-butane compositions.
5.3. Neural network, PLS and RA modeling
5.3.1. Neural network Equation-based model
As mention in section 4.2, the final configuration of the neural
network model obtained from the training and validation exercise is
given to be of a 10-10-2 network. By applying the general Eq. (1) for
this network with the linear activation function, we get the following
equation for the top and bottom composition prediction of n-butane
68
N. Mohamed Ramli et al. / Neurocomputing 131 (2014) 59–76
Bottom composition n-butane
Composition (mole fraction)
0.10
0.09
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0.00
0
2000
4000
6000
8000
10000
12000
Time (min)
simulation
online
Fig. 9. Bottom composition n-butane close loop.
where y1 refers to top composition and y2 refers to the bottom
composition;
" #
h
i
y1
1
1
2
y¼
¼ LW 2;1 f IW 1;1 p þ b þ b
ð14Þ
y2
The values of the inputs weights IW 1;1 , layer weightLW 2;1 , b1
and b2, obtained after validation are given in the Appendix. A.
Here p is the inputs to the neural network and for this case
study is given by the vector,
h
p ¼ mv2ðkÞ mv2ðk "1Þ mv3ðkÞ mv3ðk " 1Þ f ðkÞ f ðk "1Þ ptop ðkÞ
iT
ptop ðk " 1Þ pbot ðkÞ pbot ðk " 1Þ
On applying the values of the respective weights and biases for
the validated optimum neural network model for Eq. (14) and with
further pruning of the values, we get the following equation to
represent the neural network model for the composition prediction as in equation below ie;
"
#
y1
y¼
y2
$
%
" 0:29 0:15 0:37 0:23 0:38 0:40 " 0:50 0:97 0:12 " 0:31
¼
p
" 0:09 0:006 0:31 " 0:10 0:02 " 0:019 " 0:42 " 0:12 0:36 " 0:08
$
" 0:28
" 0:22
%
5.3.2. PLS model
After validation, The equation of PLS for prediction of n-butane
at top composition is given as
2
IW 1;1 ¼ input weight at layer 1 (input layer)
b1 ¼biases value at layer 1
LW 2;1 ¼ layer weight at layer 2 (hidden layer)
b2 ¼biases value at layer 2
þ
complex structure of the neural network, normally difficult to use
in an online measurement system.
ð15Þ
This Eq. (15) is obtained by simplifying the general Eq. 1 by
considering only the hidden layer with inputs weights IW 1;1 , and
the output layer with the layer weight LW 2;1
Initially the matrix input IW 1;1 is multiplied with the input
vector, p and added to biases value b1. Since the activation function
of f1 is determined as unity, the resulting matrix is then multiplied
to layer weight 2, LW 2;1 and added to biases value at layer 2, b2. By
pruning out the small resulting values, the equation is then
simplified to the version in Eqn (15).
This Eq. (15) is a multi input multi output equation based
representation of the neural network model for composition
prediction of the debutanizer column. This equation is robust in
nature and can be easily used as an online estimation for
composition in the column, without having to resort to use of
2
3
mv2 ðkÞ
6
7
6 mv2 ðk " 1Þ 7
6
7
6 mv3 ðkÞ
7
6
7
6
7
6 mv3 ðk " 1Þ 7
6
7
6 f ðkÞ
7
6
7
Y 1;PLS ¼ 0:1335 þ 6
7
6 f ðk " 1Þ
7
6
7
6 p_top ðkÞ
7
6
7
6
7
6 p_top ðk " 1Þ 7
6
7
6 p_bot ðkÞ
7
4
5
p_bot ðk " 1Þ
3
" 0:003
6 0:0007
7
6
7
6
7
6 " 0:0006 7
6
7
6 " 0:001
7
7
2
3 6
0:07
6
7
6 " 0:001
7
6 " 0:07 7 6
7
6
7 6 " 0:0007 7
6
7 6
7
6 " 0:06 7 6
7
6
7 6 " 0:0004 7
6 0:06 7 6
7
6
7 6 " 0:000076 7
6
7 6
7
6 " 0:06 7 6
7
6
7 6
7
6 " 0:11 7 þ 6 0:0003
7
6
7 6
7
7
6
7 6 0:001
7
6 0:06 7 6
7
6
7 6 0:002
7
6 " 0:01 7 6
7
6
7 6
7
6
7 6 0:004
7
4 0:68 5 6
7
6 0:018
7
6
" 0:83
7
6
7
6 0:004
7
6
7
6:
7
6
7
6
5
4:
ð16Þ
" 0:0003
and the equation of PLS for predictions of n-butane at the bottom
composition is given as,
2
2
3
mv2 ðkÞ
6
7
6 mv2 ðk " 1Þ 7
6
7
6 mv3 ðkÞ
7
6
7
6
7
6 mv3 ðk " 1Þ 7
6
7
6 f ðkÞ
7
6
7
Y 2;PLS ¼ 0:05276 þ 6
7
6 f ðk " 1Þ
7
6
7
6 p_top ðkÞ
7
6
7
6
7
6 p_top ðk " 1Þ 7
6
7
6 p_bot ðkÞ
7
4
5
p_bot ðk "1Þ
3
" 0:004
6 " 0:004 7
6
7
6
7
6 " 0:004 7
6
7
6 " 0:003 7
7
2
3 6
0:002
6
7
6
7
6 " 0:0007 7 6 " 0:002 7
6
7 6 " 0:001 7
6
7 6
7
6 " 0:0012 7 6
7
6
7 6 " 0:0011 7
6 0:0004 7 6
7
6
7 6 " 0:0002 7
6
7 6
7
6 " 0:004 7 6
7
6
7 þ 6 0:001
7
6 0:002
7 6
7
6
7 6
7
6
7 6 0:002
7
6 0:06
7 6
7
6
7 6 0:002
7
6 0:17
7 6
7
6
7 6
7
6
7 6 0:001
7
4 1:64
5 6
7
6 " 0:0002 7
6
7
" 0:073
6
7
6 " 0:0017 7
6
7
6:
7
6
7
6
7
4:
5
ð17Þ
0:0012
The F residual for PLS equation consists of 301 data points for
top and bottom composition.
69
N. Mohamed Ramli et al. / Neurocomputing 131 (2014) 59–76
Table 4
ANOVA of the n-butane top composition for NN model.
Table 5
ANOVA of n-butane bottom composition for NN model.
Regression Statistics
Regression Statistics
Multiple R
R Square
Adjusted R Square
Standard Error
Observations
0.9943
1.00
0.9884
0.0018
301
ANOVA
df
SS
MS
F
Regression
Residual
Total
10
290
300
0.0906
0.0010
0.0917
0.0090
3.5399E-06
2562.012
Multiple R
R Square
Adjusted R Square
Standard Error
Observations
0.9526
1.00
0.9383
0.0060
301
Significance F
ANOVA
df
SS
MS
F
Significance F
2.3449E-276
Regression
Residual
Total
10
290
300
0.0467
0.0106
0.0573
0.0046
3.6617E-05
127.565
5.7343E-100
Y _1;RA ¼ 0:0008mv2 ðkÞ " 0:0007mv2 ðk " 1Þ þ 0:0004mv3 ðkÞ
"0:0006mv3 ðk " 1Þ " 0:0011f ðkÞ þ 0:0019 f ðk " 1Þ
þ 1:01p_top ðkÞ
"0:051p_top ðk " 1Þ þ 0:002p_bot ðkÞ
"0:01p_bot ðk " 1Þ " 0:078
ð18Þ
each other. The standard deviation s could also be determined
from the MS of the residual and has the value of 6.05 ! 10 " 3.
The analysis for top and bottom composition based on ANOVA
is used to determine the hypothesis between the actual and
predicted value of n-butane composition. The F test in ANOVA
provides a single test of the hypothesis that all the population is
assume to be equal. The F test was used to access differences for a
set of two group where the two groups are the regression and
residual.
Y 2;RA ¼ 0:0019mv2 ðkÞ " 0:0018mv2 ðk " 1Þ " 0:002mv3 ðkÞ
5.4. Comparison NN, PLS and RA
5.3.3. Regression model
For the regression model, the equations for the top and bottom
prediction n-butane are described below;
þ0:001mv3 ðk " 1Þ þ0:004f ðkÞ " 0:006 f ðk " 1Þ
þ 0:30p_top ðkÞ " 0:23p_topðk " 1Þ þ 0:81p_bot ðkÞ
"0:059p_bot ðk " 1Þ þ 0:27
ð19Þ
5.3.4. Analysis of variance (ANOVA) results for neural network model
5.3.4.1. Top composition. From Table 4, the adjusted R2 is smaller
than R2 value since to the number of cases is relatively small and
the number of predictor variables is relatively large. There is a total
of 301 samples data observations. The sum of square regression is
calculated to be 0.0906 and the total sum of square is calculated to
be 0.0917. The multiple R is calculated based on the square root of
ratio between these 2 values. The multiple R is proportional to the
total variance in the actual and predicted value. The standard error
shows the ratio between the standard deviation to the square root
of number of observations. The degree of freedom (df) is the
variation between the sample size and number of groups with
confidence level 95%.
The sum of square (SS) consists of regression, residual and total.
It is explained by the difference between each group mean and the
overall mean. The value of mean squares (MS) are obtained from
the ratio of the sum of the square (SS) to the degree of freedom
(df). The F value is obtained from the ratio of MS of regression to
MS of residual. From the ANOVA analysis outlined in Table 4, the F
value obtained is 2562. It indicate that the between estimate
groups is more than 2562 times the within group estimate. The
standard deviation (s) may also be determined from the MS of
residual and the s value is 1.88 ! 10 " 3.
5.3.4.2. Bottom composition. Table 5 also shows that the R2 value is
greater that the adjusted R2 due the number of cases which is
small and the number of predictor variables is large. The samples
data observation consists of 301 data points. From the ANOVA
analysis obtained in Table 5, the F value is 127. It indicates that the
between groups estimate is more than 127 times the within group
estimate. The significance F value is relatively very small so
therefore the different population mean are recorded. The F
value is larger than 1.83, which indicates that all the variables
involved for composition prediction is important and related to
Fig. 10 shows the observed versus predicted values of the top
composition of n-butane as predicted by the neural network
equation. It is apparent that all the points fall close to the 45
degree line. The calculated RMSE for the NN equation is 6.6 ! 10 " 7
were the square regression of one indicates excellent fit of data.
Fig. 11 shows the composition line plot of the actual and neural
network equation for n-butane top composition. Fig. 12 shows the
observed versus predicted values of the bottom composition of
n-butane from the NN equation. It is apparent that all the points
fall close to the 45 degree line. The calculated RMSE for the NN
equation is 3.88 ! 10 " 7. Fig. 13, shows the composition line plot of
the actual and neural network equation for the n-butane bottom
composition. The CDC value for top composition is calculated to be
at 26.33 and for bottom composition is calculated to be 100 where
high CDC value indicates better prediction. The regression value of
R for top and bottom composition is 1 and thus the prediction
between the actual and simulated is similar. The Cp value for
bottom and top composition are calculated to be 1 and the MAPE
for top and bottom are calculated to be 0.0005 and 0.00132
respectively. The TIC values for bottom and top composition are
calculated to be 3.56 ! 10 " 6 and 2.45 ! 10 " 6 respectively.
Fig. 14 shows the observed versus predicted values of the top
composition of n-butane from using the PLS equation. It is
apparent that all the points fall close to the 45 degree line. The
calculated RMSE for the PLS equation is 0.002 with R2 is 0.9851 but
the scattered data points around the regression line are an
indication of poor prediction. Fig. 15, shows the composition plot
of the actual and PLS equation n-butane top composition. Fig. 16
shows the observed versus predicted values of the bottom composition of n-butane from PLS equation. The calculated RMSE for
the PLS equation is 0.0059 and the value of the R2 is 0.8117. Again
scattered data points around the regression line are an indication
of poor prediction by the PLS equation. Fig. 17, shows the
composition plot of the actual and PLS equation n-butane bottom
composition. The CDC value for top composition is calculated to
17.66 and for bottom composition is calculated to be 56.66. The
regression value of R for top and bottom composition is 0.99 and
0.9 respectively with Cp value is almost close to 1. The Cp value for
70
N. Mohamed Ramli et al. / Neurocomputing 131 (2014) 59–76
NN equation top composition n-butane
Predicted composition (mole fraction)
0.2
R2 = 1
0.19
0.18
0.17
0.16
0.15
0.14
0.13
0.12
0.11
0.1
0.1
0.11
0.12
0.13
0.14
0.15
0.16
0.17
0.18
0.19
0.2
Actual composition (mole fraction)
Fig. 10. Prediction versus actual value neural network equation top composition n-butane.
Neural network prediction top composition n-butane
Composition (mole fraction)
0.25
0.2
0.15
0.1
0.05
0
0
50
100
150
200
250
300
Time (min)
Actual
NN eq
Fig. 11. Prediction and actual value for top composition n-butane.
NN equation bottom composition n-butane
Predicted composition (mole fraction)
0.09
0.08
2
R =1
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
Actual composition (mole fraction)
Fig. 12. Prediction and actual value equation based neural network bottom composition n-butane.
bottom and top composition are calculated to be 0.9 and 0.99
respectively and the MAPE for top and bottom are calculated to be
0.034 and 0.97. The TIC values for bottom and top composition are
calculated to be 5.51 ! 10 " 2 and 7.9 ! 10 " 3 respectively.
Fig. 18 shows the observed versus predicted values of the
n-butane top composition using regression analysis equation.
Most of the data points falls close to the 45 degree line but with
more scatter than the neural network case. The calculated RMSE
for the regression equation is 0.0021 and the value of the R2 is
0.9888. Fig. 19 shows the composition plot of the actual and RA
equation of the n-butane top composition. Fig. 20 shows the
observed versus predicted values of the n-butane bottom
71
N. Mohamed Ramli et al. / Neurocomputing 131 (2014) 59–76
Neural network prediction bottom composition n-butane
Composition (mole fraction)
0.09
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
0
50
100
150
200
250
300
Time (min)
Actual
NN eq
Fig. 13. Prediction and actual value for bottom composition n-butane.
PLS equation top composition n-butane
Predicted composition (mole fraction)
0.2
R2 = 0.9851
0.19
0.18
0.17
0.16
0.15
0.14
0.13
0.12
0.11
0.1
0.1
0.11
0.12
0.13
0.14
0.15
0.16
0.17
0.18
0.19
0.2
Actual composition (mole fraction)
Fig. 14. Prediction versus actual value equation based PLS top composition n-butane.
PLS prediction top composition n-butane
Composition (mole fraction)
0.25
0.2
0.15
0.1
0.05
0
0
50
100
150
200
250
300
Time (min)
Actual
PLS eq
Fig. 15. Prediction and actual value for top composition n-butane.
composition using regression analysis equation. The points are
scattered as shown in the figure by the RA equation. This indicates
poor prediction by the RA equation. The calculated RMSE for the
normal regression equation is 0.0064 and the value of the R2 is
0.8148. Fig. 21 shows the composition plot of the actual and RA
equation n-butane bottom composition. The CDC value for top
composition is calculated to 17.33 and for bottom composition is
calculated to be 56.66. The Cp value for bottom and top composition are calculated to be 0.89 and 0.99 respectively. The MAPE for
top and bottom are calculated to be 0.058 and 2.67. The TIC values
for RA prediction bottom and top composition are calculated to be
5.46 ! 10 " 2 and 6.86 ! 10 " 3 respectively.
The Akaike information criteria (AIC) is related to the square of
residual to the number of free model parameters. The purpose is to
weigh the error of the model against the number of parameters.
The BIC is similar to AIC except that it is motivated by the Bayesian
model selection principles. The AIC values depend on the mean
square error, the variance, the number of free model parameter
72
N. Mohamed Ramli et al. / Neurocomputing 131 (2014) 59–76
PLS equation bottom composition n-butane
Predicted composition (mole fraction)
0.08
0.07
0.06
R2 = 0.8117
0.05
0.04
0.03
0.02
0.01
0
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
Actual composition (mole fraction)
Fig. 16. Prediction versus actual value equation based PLS bottom composition n-butane.
PLS prediction bottom composition n-butane
Composition (mole fraction)
0.09
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
0
50
100
150
200
250
300
Time (min)
Actual
PLS eq
Fig. 17. Prediction and actual value for bottom composition n-butane.
RA equation top composition n-butane
Predicted composition (mole fraction)
0.2
R 2 = 0.9888
0.19
0.18
0.17
0.16
0.15
0.14
0.13
0.12
0.11
0.1
0.1
0.11
0.12
0.13
0.14
0.15
0.16
0.17
0.18
0.19
0.2
Actual composition (mole fraction)
Fig. 18. Prediction versus actual value equation based RA top composition n-butane.
and number of parameter. The BIC values depend on the mean
square error, the variance, number of observation, the number of
free model parameter and number of parameter. The AIC and BIC
predicted by NN for top composition is calculated to be 2572 and
2555 respectively while the AIC and BIC for bottom composition
calculated to be 1957 and 1942 respectively. The AIC and BIC
predicted by PLS for top composition is calculated to be 2573 and
2558 respectively. The AIC and BIC for bottom composition
calculated to be 2073 and 2059 respectively. The AIC and BIC
predicted by RA for top composition calculated to be 2580 and
2560 respectively and the AIC and BIC for bottom composition,
calculated to be 2074 and 2058 respectively. These values can be
seen in Table 6, which shows that the neural network equation
with smaller AIC and BIC values, still gives the optimum prediction
even with slight extra parameters in its formulation.
From the statistical analysis outlined in Table 6, NN equation
give better prediction for the n-butane composition than PLS
equation and RA equation as the calculated RMSE is small, CDC
is high, R2 is close to 1, MAPE is close to 0, Cp is close to 1 and TIC
close to zero. The CDC values for NN are larger compared to PLS
73
N. Mohamed Ramli et al. / Neurocomputing 131 (2014) 59–76
Regression prediction top composition n-butane
Composition (mole fraction)
0.25
0.2
0.15
0.1
0.05
0
0
50
100
150
200
250
300
Time (min)
Actual
RA eq
Fig. 19. Prediction and actual value for top composition n-butane.
Predicted composition (mole fraction)
RA equation bottom composition n-butane
0.08
0.07
0.06
2
R = 0.8148
0.05
0.04
0.03
0.02
0.01
0
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
Actual composition (mole fraction)
Fig. 20. Prediction versus actual value equation based RA bottom composition n-butane.
Regression prediction bottom composition n-butane
0.09
Composition (mole fraction)
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
0
50
100
150
200
250
300
Time(min)
Actual
RA eq
Fig. 21. Prediction and actual value for bottom composition n-butane.
and RA and the high CDC indicates that the subsequent actual
change of the predicted variable is high. The R and Cp value for NN
is the optimum performance as the neural network prediction
matches the actual data. The MAPE values indicate that NN
prediction is the optimum as the values are closest to 0 compared
to PLS and RA where the MAPE values are larger. When having a
perfect fit, MAPE is zero. The percentage error calculated for MAPE
is to compare the error of the fitted time series. The difference
between actual value and predicted value divided by the actual
value determine the MAPE. The absolute value is summed for
every value fitted in time and divided again by the number of
fitted points. The TIC values indicate the prediction by NN is the
best as the TIC values are small as compared to PLS and RA. These
statistical analyses proves that the prediction by the proposed NN
model gives optimum performance, better than the other conventional methods.
74
N. Mohamed Ramli et al. / Neurocomputing 131 (2014) 59–76
Table 6
Statistical analysis of NN equation, PLS equation and RA equation for top and bottom n-butane predictions.
rmse_bottom
rmse_top
CDC_bottom
CDC_top
R_bottom
R_top
AIC_bottom
AIC_top
BIC_bottom
BIC_top
MAPE_bottom
MAPE_top
Cp_bottom
Cp_top
TIC_bottom
TIC_top
NN eq
PLS eq
RA eq
3.88E-07
6.6E-07
100
26.33
1
1
1957.26
2572.72
1942.43
2555.89
0.00132
0.0005
1
1
3.56E-06
2.45E-06
0.0059
0.0020
56.66
17.66
0.90
0.99
2073.63
2573.78
2059.8
2558.96
0.97
0.034
0.90
0.99
5.51E-02
7.90E-03
0.0064
0.0021
56.66
17.33
0.89
0.99
2074.26
2580.29
2058.44
2560.46
2.67
0.058
0.89
0.99
5.46E-02
6.86E-03
NN eq
PLS eq
RA eq
5 second
45 second
1 minute
Table 7
Computing time.
Computing time
Residual analysis top composition equation NN, PLS and RA
6.70E-07
Residual composition (mole
fraction)
0.02
NN
0.015
0.01
6.68E-07
PLS
RA
6.66E-07
6.64E-07
0.005
6.62E-07
0
-0.005
0
50
100
150
200
250
300
6.60E-07
-0.01
6.58E-07
-0.015
6.56E-07
NN residual composiiton (mole
fraction)
6.72E-07
0.025
6.54E-07
-0.02
Time (min)
PLS
RA
NN
Fig. 22. Residual analysis for neural network equation, PLS equation and regression analysis equation top composition n-butane.
The difference in computing time using these different
approaches are shown in Table 7 where the NN model takes less
than 5 seconds to compute which is faster than the PLS (45 seconds) and RA method (1 minute). Hence it is suitable for online
measurement since the industrial method takes more than 1 day
to analyse and compute.
5.5. Residual analysis
Figs. 22 and 23 show the residual of the neural network
equation, PLS equation and normal regression equation for top
and bottom composition n-butane respectively. From the plot, the
residual of the neural network equation is smaller compared to the
PLS equation and NR equation. This shows that neural network is
able to predict the top and bottom composition n-butane with
high accuracy with small error compared to the PLS and RA.
Residual analysis is very important to evaluate the deviation
between actual and prediction for all the three models.
6. Conclusion
This paper presents the prediction of the composition of nbutane at the top and bottom of a debutanizer column using the
equation based neural network model which is then compared to
other methods such as PLS and regression analysis. All of the
results gives optimum results in predicting the n-butane compositions but it can be concluded that NN equation gives the best nbutane prediction compared to other models based on the
statistical analyses. This proposed equation based NN model is
useful for online composition prediction since it is robust, versatile
with fast computing time and hence can be easily applied as a soft
sensor for the distillation column. It could also easily be further
75
N. Mohamed Ramli et al. / Neurocomputing 131 (2014) 59–76
Residual analysis bottom composition equation NN, PLS and RA
0.015
4.04E-07
PLS
RA
4.02E-07
4.00E-07
0.01
3.98E-07
0.005
3.96E-07
0
-0.005
0
50
100
150
200
250
3.94E-07
3003.92E-07
3.90E-07
-0.01
3.88E-07
-0.015
-0.02
3.86E-07
NN
NN residual compositon (mole
fraction)
Residual composition (mole
fraction)
0.02
3.84E-07
3.82E-07
-0.025
Time (min)
PLS
RA
NN
Fig. 23. Residual analysis for neural network equation, PLS equation and regression analysis equation bottom composition n-butane.
Table A1
Input weight and biases value for n-butane with partition.
input weight 1,1 for the first layer
" 0.86
" 0.55
0.23
0.70
0.74
0.95
" 0.81
" 0.25
" 0.65
0.13
" 0.82
0.51
" 0.40
" 0.59
" 0.22
0.05
0.59
0.17
0.92
0.11
0.98
0.36
0.11
0.30
" 0.72
" 0.95
0.44
0.25
" 0.47
0.96
b1¼ biases at layer 1
" 0.09
" 0.08
" 0.16
" 0.17
0.34
0.31
0.45
0.16
1.01
" 0.07
0.34
0.97
0.21
0.81
0.63
0.91
" 0.43
" 0.24
" 0.07
0.85
0.97
" 0.62
" 0.19
" 0.13
" 0.77
0.77
" 0.50
0.72
" 0.71
" 0.63
0.96
0.23
" 0.45
0.95
" 0.62
0.08
0.35
" 0.89
" 0.37
" 0.73
0.04
0.15
0.34
0.39
" 0.77
0.68
" 0.80
0.84
0.64
" 0.82
0.65
" 0.36
0.63
" 0.56
" 0.30
0.81
0.36
0.13
" 1.08
0.77
" 0.16
" 0.13
0.71
" 0.40
0.50
" 0.10
0.86
" 0.93
0.14
0.66
layer weight 2,1 for the second layer
0.28
" 0.09
" 0.17
0.24
0.07
0.18
" 0.11
0.53
0.43
" 0.10
" 0.20
0.37
" 0.32
0.86
" 0.84
" 0.57
b2¼ biases at layer 2
" 0.11
0.02
" 0.50
0.24
0.16
" 0.07
applied as an inverse controller in the equation form especially for
nonlinear system, where linear controllers are not able to perform
successfully. This proposed model based NN method is also easier
to visualize and applied for various applications as compared to
method of using the black box neural network structure which is
cumbersome and non-portable in nature. Furthermore it is MIMO
based model that can predict both the top and bottom composition through the use of a single vector equation.
Acknowledgment
The authors would like to acknowledge PETRONAS for providing the required data and information for the research and
University Malaya for providing the research grant (PS107/2010B).
Appendix A
See Table A1.
References
[1] J.F. Canete, S. Gonzalez-Perez, P. Saz-Orosco, Artificial Neural Network Identification and Control of a Lab-Scale Distillation Column using LABVIEW,
International Journal of Intelligent Systems and Technologies 3 (2008)
111–116.
[2] M.M. Zhang, X.G. Liu, A soft sensor based on adaptive fuzzy neural network
and support vector regression for industrial melt index prediction, Chemometrics and Intelligent Laboratory Systems 126 (2013) 83–90.
[3] Z.B. Yan, X.G. Liu, Soft sensing and optimization of pesticide waste incinerator,
Asia Pacific Journal of Chemical Engineering 7 (2012) 635–641.
" 0.63
0.35
0.04
0.55
0.67
" 0.37
0.57
" 0.08
0.29
" 1.08
[4] J. Shi, X.G. Liu, Melt prediction by neural soft-sensor based on multi scale
analysis and principal component analysis, Chinese Journal of Chemical
Engineering 13 (2005) 849–852.
[5] X.G. Liu, C.Y. Zhao, Melt index prediction based on fuzzy neural network and
PSO algorithm with online correction strategy, American Institute Chemical
Engineering Journal 58 (2012) 1194–1202.
[6] K.o. Ma Ming-Da, Wang Jing-Wei, San-Jang, Wu Ming-Feng, Jang Shi Shang,
Shieh Shyan-Shu,Wong David Shan-Hill, Development of adaptive soft sensor
based on statistical identification of key variables, Control Engineering
Practice 17 (2009) 1026–1034.
[7] V. Prasad, B. Wayne Bequette, Nonlinear system identification and model
reduction using artificial neural networks, Computer and Chemical Engineering 27 (2003) 1741–1754.
[8] L. Fortuna, S. Graziania, M.G. Xibilia, Soft sensors for product quality monitoring debutanizer distillation columns, Control Engineering Practice 13 (2005)
499–508.
[9] V. Singh, I. Gupta, H.O. Gupta, ANN-based estimator for distillation using
Levenberg Marquardt approach, Engineering Applications of Artificial Intelligence 20 (2007) 249–259.
[10] A. Zilochian, K. Bawazir, Application of Neural Network in Oil Refineries, CRC
Press, 2001. (Chapter 7).
[11] M.A. Hussain, Review of the application of neural networks in chemical
process control – simulation and online implementation, Artificial Intelligence
in Engineering 13 (1999) 55–68.
[12] Y. Xuefeng, Hybrid artificial neural network based on BP-PLSR and its
application in development of soft sensors, Chemometrics and Intelligent
Laboratory Systems 103 (2010) 152–159.
[13] I.M. Mujtaba, M.A. Hussain, Optimal Operation of Dynamic Processes Under
Process-Model Mismatches: Application to Batch Distillation, Computers
Chemical Engineering 22 (1998) 621–624.
[14] M.A.I.M. Greaves, I.M. Mujtaba, M. Barolo, A. Trotta, M.A. Hussain, Neural
Network approach to dynamic optimization of batch distillation Application to
a Middle-vessel Column, Trans IChemE 81 (2003) 393–401.
[15] L. Eriksson, E. Johansson, N. Kettaneh-Wold, J. Trygg, C. Wilstrom, S. Wold,
Multi and Megavariate Data Analysis Part I Basic Principles and Applications,
2nd edition, Umetrics Academy, 2006.
76
N. Mohamed Ramli et al. / Neurocomputing 131 (2014) 59–76
[16] M. Kano, K. Miyazaki, S. Hasebi, I. Hashimoto, Inferential control system of
distillation compositions using dynamic partial least squares regression,
Journal of Process Control 10 (2000) 157–166.
[17] E. Zamprogna, M. Barolo, D.E. Seborg, Estimating product composition profile in
batch distillation via partial least square, Chemical Engineering Practice 12 (2004)
917–929.
[18] S. Park, C. Han, A nonlinear soft sensor based on multivariate smoothing
procedure for quality estimation in distillation columns, Computer and
Chemical Engineering 24 (2000) 871–877.
[19] Zhang Zhang Yingwei, Yang, Complex monitoring using modified partial least
square method of independent component regression, Chemometrics and
Intelligent Laboratory Systems 98 (2009) 143–148.
[20] R. Sharmin, U. Sundararaj, S. Shah, L.V. Griend, Y.J. Sun, Inferential sensors for
estimation of polymer quality parameters: Industrial application of a PLS based
soft sensor for a LDPE plant, Chemical Engineering Science 61 (2006) 6372–6384.
[21] P. Facco, F. Doplicher, F. Bezzo, M. Barolo, Moving average PLS soft sensor for
online product quality estimation in an industrial batch polymerization
process, Journal of Process Control 19 (2009) 520–529.
[22] Song Zhihuan Ge Zhiqiang, A comparative study of just in time learning based
methods for online soft sensor modeling, Chemometrics and Intelligent
Laboratory Systems 104 (2010) 306–317.
[23] J. Shi, X.G. Liu, Melt index prediction by weighted least square support vector
machines, Journal of Applied Polymer Science 101 (2006) 285–289.
[24] J. Shi, X.G. Liu, Y.X. Sun, Melt index prediction by neural network based on
independent component analysis and multi scale analysis, Neurocomputing
70 (2006) 280–287.
[25] J.B. Li, X.G. Liu, Melt index prediction by RBF neural network optimized with
an adaptive new ant colony optimization algorithm, Journal of Applied
Polymer Science 119 (2011) 3093–3100.
[26] H.Q. Jiang, Z.B. Yan, X.G. Liu, Melt index prediction using optimized least
square support vector machines based on hybrid particle swarm optimization
algorithm, Neurocomputing 119 (2013) 469–477.
[27] H.Q. Jiang, Y.D. Xiao, J.B. Li, X.G. Liu, Prediction of melt index based on
relevance vector machine with modified particle swarm optimization, Chemical Engineering and Technology 35 (2012) 819–826.
[28] R.M. Warner, Applied Statistics, Sage Publication, 2008.
[29] Ramli Siti Aizura, Study of Neural Network for Heat exchanger with development of graphical user interface, Thesis Universiti, Teknologi PETRONAS, 2006.
[30] J. Wan, M. Huang, Y. Ma, W. Guo, Y. Wang, H. Zhang, W. Li, X. Sun, Prediction of
effluent quality of a paper mill wastewater treatment using an adaptive networkbased fuzzy inference system, Applied Soft Computing 11 (2011) 3238–3246.
[31] R. Sharma, K. Singh, D. Singhal, R. Ghosh, Neural network applications for
detecting process faults in packed towers, Chemical Engineering and Processing 43 (2004) 841–847.
[32] J.S. Lim, M.A. Hussain, M.K. Aroua, Control of a hydrolyzer in an oleochemical
plant using network based controllers, Neurocomputing 73 (2010) 3242–3255.
Nasser Mohamed Ramli is a PhD student in the
Chemical Engineering Department, Faculty of Engineering, University of Malaya. He obtained his bachelor’s
degree in chemical engineering from Loughborough
University, United Kingdom and his master’s degree
from University of Queensland, Australia. His area of
research is in artificial intelligence, process modeling
and control.
Dr Mohd Azlan Hussain joined the Department of
Chemical Engineering, University of Malaya in 1987 as
a lecturer and obtained his Ph.D in Chemical Engineering from Imperial College, London in 1996. He is a
member of the American Institute of Chemical Engineers and British Institute of Chemical Engineers. At
present he is holding the post of Professor in the
department of chemical Engineering. His main research
interests are in modelling, process controls, nonlinear
control systems analysis and applications of artificial
intelligence techniques in engineering systems. He has
published more than 250 papers in book chapters,
journals and conferences within these areas at present.
He has also publish and edited a book on “Application of Neural Networks and
other learning Technologies in Process Engineering” published by Imperial College
Press in 2001.
Dr Badrul Mohamed Jan, SPE is a researcher and
academic lecturer attached to the Department of Chemical Engineering, University of Malaya, Malaysia. He
holds a BS, MS and PhD degrees in petroleum engineering from New Mexico Institute of Mining and Technology. Jan’s research areas and interest include the
development of super lightweight completion fluid
for underbalance perforation, ultra low interfacial tension microemulsion for enhanced oil recovery, and
conversion of palm oil mill effluent into super clean
fuel for diesel replacement. He has worked closely with
industry in oil and gas project such as 3 M Asia Pacific
and BCI Chemical Corporation. He has also published
numerous technical conference and journal papers. Jan is the deputy director of
University Malaya Center of Innovation & Commercialization. His responsibilities
include providing an environment at the University of Malaya conducive to
researchers bringing their research outputs to a commercialization-ready level.
Dr Bawadi Abdullah is a Senior Lecturer in the Chemical Engineering Department, Faculty of Engineering,
Universiti Teknologi PETRONAS. He is also a Professional and a Charted Engineer. He obtained his bachelor’s degree in chemical engineering from University of
Wales, Swansea, United Kingdom and master’s degree
from Dalhousie University, Canada. He obtained his
PhD degree from University of the New South Wales,
Australia. He teaches at undergraduate level courses
such as Transport Phenomena, Chemical Engineering
Thermodynamics and Chemical Analysis. His area of
research is reaction engineering.