Accepted Manuscript

Title: A Hybrid Statistical Approach for Modeling and

Optimization of RON; A Comparative Study and Combined
Application of Response Surface Methodology (RSM) and
Artificial Neural Network (ANN) based on Design of
Experiment (DOE)

Author: F.M. Elfghi

PII: S0263-8762(16)30107-1
DOI: http://dx.doi.org/doi:10.1016/j.cherd.2016.05.023
Reference: CHERD 2291

To appear in:

Received date: 8-4-2015

Revised date: 25-5-2016
Accepted date: 26-5-2016

Please cite this article as: Elfghi, F.M.,A Hybrid Statistical Approach for Modeling
and Optimization of RON; A Comparative Study and Combined Application of
Response Surface Methodology (RSM) and Artificial Neural Network (ANN) based
on Design of Experiment (DOE), Chemical Engineering Research and Design (2016),
http://dx.doi.org/10.1016/j.cherd.2016.05.023

This is a PDF file of an unedited manuscript that has been accepted for publication.
As a service to our customers we are providing this early version of the manuscript.
The manuscript will undergo copyediting, typesetting, and review of the resulting proof
before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that
apply to the journal pertain.
 1

Graphical abstract

100 98
97 RBF (3-11-1)
R2= 0.885 2
98 96 R = 1.000
95
96
94
93
94
92
Predicted RON Values

Predicted (Output)
92 91
90

t
90 89
88

ip
88 87
86
86 85
84
84
83

cr
82 82
81
80 80
79
78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98
78 80 82 84 86 88 90 92 94 96 98 100
Experimental (Target)

us
Observed RON Values
Figure 1: Comparison between predicted and
Figure 2: Experimental data plotted against the predicted data
observed RON values for the model developed by RSM model given by ANN model.

an
M
d
p te

Figure 3: 3-D Response surface plot of RON as function of reaction Figure 5: 3-D Response Surface Plot of RON as a function of reaction
ce

temperature and space velocity depicted at fixed total pressure. temperature and total pressure at fixed space velocity.
Ac

Page 1 of 29
 2

A Hybrid Statistical Approach for Modeling and Optimization of RON; A Comparative

Study and Combined Application of Response Surface Methodology (RSM) and Artificial
Neural Network (ANN) based on Design of Experiment (DOE).

Highlights:

 Two models, RSM and ANN have been developed for prediction of RON.


t
ANN methodology showed a very obvious advantage over RSM.

ip
 DOE-ANN approach minimized the need of large quantity of training data.
 RSM-ANN model showed qualitatively and quantitatively the activity
relationships.

cr
us
an
M
d
p te
ce
Ac

Page 2 of 29
 3

A Hybrid Statistical Approach for Modeling and Optimization of RON; A Comparative

Study and Combined Application of Response Surface Methodology (RSM) and Artificial
Neural Network (ANN) based on Design of Experiment (DOE).
F.M. Elfghi*a,b
a
Department of Chemical and Petrochemical engineering, The College of Engineering &

t
Architecture, Initial Campus, Birkat Al Mouz Nizwa, Sultanate of Oman,

ip
b
Department of Environmental Engineering and Green Technology, Malaysia Japan International

cr
Institute of Technology , Universiti Teknologi Malaysia International Campus, Jalan Semarak,

54100 Kaula Lumpur, Malaysia.

us
*Corresponding Author: Fawzi Mohamed Elfghi
Email: felfghi@gmail.com / fawziamin.kl@utm.my
Tel: +60189682307

Abstract
an
The main purpose of catalytic reforming unit is to upgrade low-octane naphtha to high-
M
octane gasoline. In this work, the estimation capacities of the Response Surface Methodology
d

(RSM) and Artificial Neural Network (ANN), to determine the Research Octane Number (RON)
te

of reformates produced from the catalytic naphtha reforming unit were investigated. The article

presents a comparative study and combined application between response surface methodology
p

(RSM) and artificial neural networks (ANN) based on Design of Experiment (DOE) strategy in
ce

the modeling and prediction of the research octane number (RON). In this study, DOE-CCRD

full factorial design was incorporated into the ANN methodology, so the need of a large quantity
Ac

of training data was avoided. ANN methodology showed a very obvious advantage over RSM for

both data fitting and estimation capabilities. Based on the results of analysis of variance

(ANOVA), a multiple determination coefficient of 0.8 and 0.99 were obtained for both RSM and

ANN respectively. It has been found that by employing RSM approach coupled with ANN model

based on DOE strategy, the visualization of the experimental points in three dimensional spaces

can disclose qualitatively and quantitatively the activity relationships. This approach of

Page 3 of 29
 4

combination of RSM-ANN-DOE has revealed its ability to solve a quadratic polynomial model

involving solving, optimization, complexity and difficult relationships especially nonlinear ones

may be investigated without complicated equations involved. The study revealed that, the

maximum RON of 88 was obtained at the optimum conditions offered by RSM. Furthermore, at

the optimal conditions of (T= 521 oC, P=37.6 bar, LHSV = 2.02 h-1), the maximum RON of 98

t
ip
was obtained for the ANN model. However, the models were implemented for the construction

cr
of 3D response surface plots from the ANN and RSM models in order to show the most

effective variables as well as the effects of their interaction on the research octane number.

us
-------------------------------------------------

Key words: Catalytic Naphtha Reforming; Research Octane Number (RON); Central

an
Composite Design (CCD); Response Surface Methodology (RSM); Artificial Neural Network

(ANN).
M
1 Introduction
d

Recently, there has been a huge growth in the transportation of liquid fuels worldwide,
te

mainly because of the mature diesel market in Europe, the massive consumption of motor
p

gasoline in North America and the huge newly growing fuel market in China [1]. No doubt,
ce

petroleum for the coming two to three decades will continue to play a major role in satisfying

the transportation fuel market demands. Fuel composition and consumption, energy efficiency,
Ac

air quality are major concerns nowadays. However, oil based transportation fuels that can now

be produced to comply with stringent environmental standards [2].

Reformulated gasoline comprises of different mixed components such as reformates

containing over 60 % aromatics and additives such as MTBE to enhance the Research octane

number (RON). Since the octane number measures the maximum compression ratio in engines,

there has been a consistent increase in the demand for more efficient fuel of high octane number

which contributes to the improve of engine antiknock property and improve the combustion

Page 4 of 29
 5

quality to take place perfectly [3, 4]. However, the need to use of higher compression ratios in

engines [5] and therefore production of motor fuel with an even greater octane number requires

either severe process conditions such as temperature, pressure and space velocity or the

use of novel effective catalysts that able to reform straight-long-chain alkanes into branched and

cycled hydrocarbons with higher RON. However, catalytic naphtha (CNR) reforming process at

t
ip
high reaction temperature will result in increasing the coke formation on the catalyst, which

cr
tends to deactivate the catalyst rapidly [6], and using low Liquid Hourly Space Velocity (LHSV,

h-1) will result in decreasing reforming efficiency.

us
Operating conditions such as temperature (T, C), pressure (P, bar), space velocity

(LHSV, h-1) and H2:HC (v/v) are effective variables that play a key role on the performance of

an
any catalytic process. CNR is frequently investigated in terms of the above mentioned variables

using one-variable-at-a-time strategy (OVAT) to categorize and optimize these variables [7].In
M
catalytic naphtha reforming (CNR), the RON can be altered in a wide range either by the

modification of the texture properties of the catalyst (addition of different promoters) or by the
d

optimization of key process variables. In this regard, several publications have been reported
te

pertaining to the optimization of operating conditions using one-variable-at-a-time-approach

p

(OVAT) [8–10]. However, OVAT has some major drawbacks: the experimental space is not
ce

explored very well and the solution may be missed if there are interactions between the

variables. Moreover, OVAT is expensive and time consuming, especially when a huge number
Ac

of parameters are to be examined simultaneously [11]. Nevertheless, the experimental data

provided by OVAT offer some interesting information on the influence of each process variable

independently and can assist in determining the experimental ranges of these parameters for

further modeling and optimization studies using different optimization tools.

To solve this complexity, an alternative statistical approach, e.g. full/fractional factorial

design, canonical analysis, response surface methodology (RSM) and artificial neural network

(ANN) should be adapted to the optimization of industrial processes [10, 11]. RSM is a set of

Page 5 of 29
 6

empirical techniques devoted to estimating interactions and quadratic effects. It also gives an

idea of the local shape of the response surface. RSM and ANN approaches are practically used

to reveal the best value of the response, find out improved or optimal process settings, and

troubleshoot process problems and weak points [11–14]. A number of literatures have reported

that, parameters such as reaction temperature, total pressure and space velocity affected the

t
ip
research octane number (RON) of the reformates and the catalytic performance of the naphtha

cr
reforming process in terms of catalytic activity, selectivity and stability and RON [15].

However, the combined effect of the above mentioned variables with each other has not been

us
investigated.

an
The evaluation of a pilot plant performance for catalytic naphtha reforming is

complicated because of the competing of varies successive reactions taking place subsequently
M
such as dehydrogenation and dehydroisomerization of naphthenes to aromatics, dehydrogenation

of paraffins to olefins, dehydrocyclization of paraffins and olefins to aromatic, isomerization or

d

hydroisomerization to isoparaffins, isomerization of alkylcyclopentanes, and substituted

te

aromatics and hydrocracking of paraffins and naphthenic to lower invaluable hydrocarbons

p

[16,17]. The above subsequent/successive reactions occurred because the composition of the real
ce

feedstock is sophisticated and very complex as it contains hundreds of hydrocarbons, hence, the

use of model compound as a reactant feed is highly recommended as well as the development
Ac

and adaptation of detailed experimental procedures using Design of Experiment (DOE)

approach based on Central Composite Design (CCD) technique coupled with different modeling

and optimization tools such as RSM and ANN are extremely important in order to attain reliable

data [18- 20].

Lately, response surface methodology (RSM) and artificial neural networks (ANN) have

been gaining much attention because they have been successfully implemented as modeling and

optimization tools to solve complex and highly nonlinear problems [21, 22]. The RSM and ANN

Page 6 of 29
 7

are powerful mathematical tools that are used for modeling, simulation and optimization of

different processes [23, 24]. Both approaches do not require precise terminology of the physical

implication of the system or route under examination. As a result, RSM and ANN relate to

modeling methods that dealing with the development of non-parametric simulative models. In

fact, these models estimate the practical relationships between independent variables (input) and

t
ip
the dependent variables (response) of the process using real observed data. In addition to that,

cr
the models are utilized to estimate the optimum set of independent parameters to maximize or

minimize the response (output) of dependent variables [25, 26].

us
Even though RSM haven’t show many disadvantages, and has effectively been

implemented successfully in many studies to investigate and optimize many chemical and

an
biochemical operations, it is tough to recognize that it is suitable/fit to all optimization and

modeling of research studies. However, RSM-based models are accurate only for a narrow range
M
of input process parameters. On the other hand, the development of higher order RSM models

requires a larger number of experiments to be performed. This poses a limitation on the use of
d

RSM models for highly non-linear processes, such as catalytic naphtha reforming. These
te

constraints have led to the model development based on artificial neural networks [27].
p

However, it is worthy to mentioning here that ANNs have a black box temperament. Even
ce

though, ANNs discover qualitative interaction in a multi-dimensional experimental space due to

the shortage of appropriate visualization techniques, these interactions qualitatively still keep on
Ac

concealed for researchers and co-workers. Indeed, by employing RSM coupled with ANN’s.

This difficulty could be overcome. As it has stated previously, RSM visualization of the

experimental points in three dimensional spaces can qualitatively and quantitatively disclose the

activity relationships. This approach, in joint utilization of RSM and ANNs, data that is intrinsic

in the ANNs can be envisaged [28].

Therefore this work was attempted to model and optimize the process conditions of CNR

process by developing a hybrid statistical approach. A Combined application of ANN falling

Page 7 of 29
 8

under Response Surface Methodology (RSM) were convoyed together and employed

simultaneously to predict maximum RON over Pt-Re bimetallic catalyst via reforming of model

compounds (n-octane). In this study, we report the use of Design of Experiment (DOE)

technique and performing of Central Composite Design (CCD) experiments in order to

minimize the number of experiments required for the prediction of the optimum values of

t
ip
process parameters and their interactions for high research octane number.

cr
2. Experimental Setup

us
2.1 Catalysts Preparation

An industrial applied Pt-Re bimetallic catalyst was used in this work. The catalyst has a

an
surface area of 245.7 m2g-1. The catalyst pellets with a cylindrical geometric shape (1.5–2.0 mm

× 5.0–8.0 mm) and consist of dual function components, Pt (0.17 wt %) active species for
M
hydrogenation and dehydrogenation reactions and Re (0.35 wt %) as a promoter, supported on

chlorinated alumina, Al2O3-Cl (Cl = 1.34 wt%) were used. The amount of the loaded metals was
d

determined by Atomic Absorption Spectroscopy (AAS) and Inductivity Coupled Plasma

te

Spectroscopy (ICP) techniques using (VISTA-PRO CCD SIMULTANEOUS ICP-OES) device.

p
ce

2.2 Catalytic Activity Test

The catalytic reforming tests were performed using a high-pressure catatest unit
Ac

(Geomecahnique, Vinci, France). The unit consisted of a fixed-bed down-flow micro reactor

which was designed to be operated at high pressure up to 150 bar and temperature up to 700oC.

The reactor has an internal diameter of 15 mm and volume of 4-15 ml with an electrical heating-

zone (jacket). An amount of 1g (2.04 ml) catalyst was sandwiched between two layers of silicon

carbide (carbines) inert particles. Prior to the activity test, the catatest unit was pressurized up to

90 bars and maintained for 120 mins under nitrogen flow for leakage test. After confirming that

Page 8 of 29
 9

there was no leak (less than 0.1 MPa/h), the unit was depressurized and the operating conditions

were adjusted to the required activity test conditions for the experimental runs.

2.3 Feed and Product Analysis

The feed consisted of a model compound (n-octane) was supplied by Sigma-Aldrich USA. n-

t
ip
octane has a research octane number (RON) equal zero while iso-octane has an octane number

cr
equal to 100. n-octane was the most suitable structure reference for catalytic naphtha reactions

and it could be assumed as a model compound. The feed and product distribution of liquid

us
hydrocarbons and research octane number (RON) were determined using Near Infra Red (NIR)

Portable Petro- Spec Gasoline Fuel Analyzer. Reformat samples of liquid products were

an
collected for each condition after 45 min equilibrium time every hour for 240 min time-on-

stream (TOS).
M
2.4 Design of Experiments (DOE)
d

A complete description of the process behavior requires a quadratic or higher order polynomial
te

model. Hence, the full quadratic models were established by using the method of least squares,
p

which included all interaction which is termed to calculate the predicted response. The quadratic
ce

model is usually sufficient for industrial applications [29]. For n factors the full quadratic model

is shown in Eq. (1)

Ac

Y = bo + ∑ biXi + ∑ bijXiXj (i,j = 1,2,3,…..,k) (1)

Where, Y is the predicted response or dependent variable, Xi and Xj are the independent

variables, and bi and bj are constants. In this case, the number of independent factors is three

and therefore, k = 3 and Eq. (1) Becomes Eq. (2):

Yu=βo+β1X1+ β2 X2+ β3 X3+ β12 X1X2+ β13X1X3+ β23X2X3+ β11X12+β22X22+β33X32 (2)

Page 9 of 29
 10

With Y being the predicted response, X1, X2 and X3 are the coded form of the input variables for

reaction temperature, operating pressure and space velocity, respectively. The term βo is the

intercept term; β1, β2 and β3 the linear terms; β11, β22 and β33 are the squared terms; β12 β13 β23 are

the interaction terms between the three variables. The selection of these variables with their

defined experimental ranges were carefully chosen based on previous screening tests prior to

t
ip
optimization, and are often used in literatures. The lowest and the highest levels of variables

cr
coded as -1 and +1, respectively, are given in Table 1 including axial star points of (- α and + α),

where α is the distance of the axial points from center and makes the design rotatable.

us
Table 1: Independent variables and their coded and actual values

an Coded levels
M
Independent variable Symbol
-α -1 0.00 +1 +α

Reaction temperature (T,◦C) X1 468 480 495 510 521

d

Operating pressure (P,bar) X2 2 10 20 30 37

te

Space velocity (LHSV,h-1) X3 0.9 1.2 1.5 1.8 2.0

p
ce

In this study α value was calculated using Eq. (3) And was fixed at 1.68 (Rotatable).

α = (F) ¼ (3)
Ac

Where F is the number of points in the cube section of the design (F = 2k, k is the number of

factors). Since we have three factors, the F number is equal to 23 (= 8) points, and α = 1.68.

Therefore, the total number of experimental combinations should be conducted based on the

same concept of CCD by applying Eq. (4)

2k+2k + no (4)

Where, k is the number of independent variables and no is the number of experiments repeated at

the center point. In this case, no = 2 and k = 3 gives the total number of runs needed as 16. A

Page 10 of 29
 11

matrix of 16 experiments with three factors was generated using the software package,

‘STATISTICA version 8 (StatSoft Inc., Tulsa, USA). The two center points were used to

determine the experimental error and reproducibility of the data. Table 2 shows the complete

design matrix of the experiments performed and the results of RON obtained. The responses

were used to develop an empirical model for RON (YRON). Analysis of Variance (ANOVA) at

t
ip
the 5 % level of significance using Fisher F-test was used in determining the experimental

cr
design, interpretations and analyses of the experimental data. An arithmetical method that sorts

out the components of a given variation in a set of data and provides significance test is called

us
the F-test.

an
M
d
p te
ce
Ac

Page 11 of 29
 12

Table 2 Design of Experiment (DOE) of Central Composite Rotatable Design (CCRD) used in
RSM and ANN analysis together with their responses

X1 X2 X3
Standard Run RON
Pressure LHSV
Temperature, (oC) Level Level Level
(bar) (h-1)

t
1 R 480 -1 10 -1 1.2 -1 88.6

ip
2 R 480 -1 10 -1 1.8 1 88.06

cr
3 R 480 -1 30 1 1.2 -1 82.3

us
4 R 480 -1 30 1 1.8 1 81.4

5 R 510 1 10 -1 1.2 -1 95.3

an
6 R 510 1 10 -1 1.8 1 98.5

7 R 510 1 30 1 1.2 -1 88.7

M
8 R 510 1 30 1 1.8 1 91.2

9 C1 495 0 20 0 1.5 0 85.7

d

10 R 468 -α 20 0 1.5 0 81.3

te

11 R 521 +α 20 0 1.5 0 93.7

12 R 495 0 2.36 -α 1.5 0 91.5

p

13 R 495 0 37.63 +α 1.5 0 88.3

ce

14 R 495 0 20 0 0.97 -α 90.5

Ac

15 R 495 0 20 0 2.02 +α 80.8

16 C2 495 0 20 0 1.5 0 85.7

Page 12 of 29
 13

3 Results and Discussion

3.1 Response Surface Methodology Model (RSM).

The experiments were performed in accordance with the designed matrix as shown in Table 2.

The coefficients of the model developed in Eq. (2) were estimated using multiple regression

analysis technique. The quadratic mathematical model of the RON yield (YRON) is represented in

t
ip
Eq. (5).

cr
Y RON = 805.3954 - 2.9024X1 - 4.1514X2 + 1.8125X3 + 0.0060X1 X2 - 0.0286X1X3 + 0.2629 X2X3+

0.0031X12 + 0.0153 X22 + 0.7218 X32 (5)

us
Where X1, X2 and X3 are the real values of temperature, pressure and LHSV respectively.

an
The mathematical relationships employed for estimation of the ANOVA estimators (i.e. SS,

MS, F-value, R2, R2adj) were usually used in the literature regarding DOE and RSM [30, 31].
M
The fitness of various models (linear, two factorial and quadratic) developed was arbitrated

from the determination of the correlation coefficient (R2). A practical rule of thumb for
d

evaluating the determinant coefficient, R2 is that it should be higher or equal 0.75 [32], which
te

explain the total deviation of the observed values of activity about its mean [33]. The closer the
p

R2 value to unity, the better the model will be, as it will give predicted values closer to the actual
ce

values for the response as shown in Figure 1. In this case, the value of R2 = 0.885 indicates that

there is a good agreement between the observed and predicted values of RON from the fitted
Ac

model

Page 13 of 29
 14

100

98
R2 = 0.885
96 Y = 0.801X+18.11
94
Predicted RON Values

92

t
ip
90

88

cr
86

84

us
82

80

78
78 80 82 84 86 88
an
90 92

Observed RON Values

94 96 98 100
M
Figure 1: Comparison between predicted and observed RON values for the model developed by
RSM.
d
te

The accuracy of the RSM model was checked by ANOVA. Generally, the Fisher test value is a

measurement of variance of data about the mean based on the ratio of mean square of group
p

variance due to error [34]. The F-computed value should be higher than the tabulated value for
ce

the model to be considered as a good predictor model. In this study, the calculated Fisher test

values for the RON model shown in Table 3, are greater than the tabulated F-value (F6, 9, 0.05 =
Ac

3.37) at α = 0.05 in statistic tables [35]. By that virtue, the null hypothesis (H0) was rejected

and it can be deduced that the model has a good prediction of the response.

Page 14 of 29
 15

Table 3: ANOVA Analysis Table for RON Model developed by RSM

Sum of Degree of Mean F(calculated) F(Tabulated)

squares freedom squares value (6,9,0.05)
Sources

S.S. Regression 264.362 9 29.373 5.147 >3.37

t
S.S. Error 34.2352 6 5.705

ip
S.S. Total 298.5972 15

cr
The calculated F-value corresponding to RON model is 5.15 greater than the tabulated F- value

us
(F6, 9, 0.05 = 3.37) which implies that eqn. (5) has a good prediction of the RON model and that

the estimated factor effects are real at 95% confidence level.

an
Table 4, demonstrates the evaluated t-Student’s distribution test and the corresponding p-values.

The single-factor term represents a linear effect of the corresponding factor, while the two
M
factors represent the interaction between the two factors. In addition the second order term
d

represents a quadratic effect towards the response. The p-value serves as a tool to check the
te

importance of each coefficient at a specified level of significance. The higher the t-value or the

smaller the p-value the more significant is the corresponding coefficient. Generally, a p-value of
p

less than < 0.05 is considered to be very significant and contributes largely towards the
ce

response.

As illustrated in the table, the linear reaction temperature term (X1), has the largest effect on
Ac

RON at the 99.7%, which assurance the level of significance as indicated by the lowest p-value

of (<0.003) and the highest absolute t-value (5.18). Moreover, the linear term of total pressure

(X2) could also be considered as a significant factor affecting the RON (p-value < 0.05) at the

95% confidence level followed by the effect of contact time (LHSV, h-1) (X3) and the quadratic

term (X22) of pressure which appeared a statistical significant at 94% and 91%, respectively.

According to a t- and p- values sorted in Table 4, the signifying order of the independent

Page 15 of 29
 16

variables on the maximum obtained RON is reaction temperature, operating pressure and space

velocity, respectively. However, the other terms are not statistically significant.

Table 4 : Evaluated t-Student’s Distribution Test and the Corresponding p-values

t
Parameter Term Coefficient t-value p-value

ip
β1 X1 -2.9024 5.18284 0.002049a

cr
β2 X2 -4.1514 -2.58383 0.041555 a

β3 X3 1.8125 -2.33331 0.058376 a

us
β22 X2 2 0.0153 2.06595 0.084359 b

β12 X1X2 0.0060 1.07012 0.325713 b

β11

β23
X1 2

X2X3 an
0.0031

0.2629
0.93561

0.93395
0.385583 b

0.386374 b
M
β13 X1X3 -0.0286 -0.15245 0.883828 b

β33 X3 2 0.7218 0.08786 0.932848 b

d

R2
te

0.885
p
ce

3.2 ANN Model

Designing the artificial neural network is the first step in training a neural network to obtain the
Ac

optimal number of neurons in the hidden layer of the neural network. The Central Composite

Rotatable (CCRD) experimental design data and the result used for the construction of the ANN

model are shown in Table 2. The data has been divided into three sets including 70% as a

training data, 15% as testing and 15% as validation data (or 30% for testing and validation data)

of the ANN - model. The model has a feed-forward structure trained using the multilayer

perceptron (MLP) and Radial Basis Function (RBF) networks in this study.

Page 16 of 29
 17

3.3 Effect of Architecture and Topology on Neural Network Performance

3.3.1 Optimal number of hidden neurons

The verification and selection of an optimal neural network architecture and topology is one of

the significant steps for an effective application. The input and output layers consist of a fixed

independent and dependent variables as the neurons, which equals to 3 and 1, respectively. The

t
ip
crucial aspect of this process is the selection of the number of hidden layers and neurons in the

cr
hidden layers [36]. However, there is no any specific procedure to determine the numbers of

these neurons in advance. In order to obtain the optimal number of neurons in the hidden layer

us
of the neural network, a number of neurons in the hidden layer have been investigated by a trial

and error method (manifold trials) to verify the best network architecture. Many neural-network

an
architecture and topologies (number of hidden neurons, connection types, learning algorithms

and transfer functions of input and hidden layers) were examined for the estimation and
M
prediction of the research octane number of octane reforming. Table 5 shows the top six ANN

Models.
d
p te
ce
Ac

Page 17 of 29
 t
ip
18

cr
Table 5: The effect of different Neural Network architecture and topologies on R2 and AAD obtained in training and testing of

us
neural networks

Transfer Transfer Training Training Test

Model Training Connection Test error

an
Index Function Function performance error performance
Name algorithm Type ADD %
(Output) (Hidden) R2 ADD % R2

C8 3-10-1 BFGS24 a MLPFFc Exponential Tanh 0.934 0.007 1.000 0.007

M
0.002
C10 3-11-1 BFGS 36 MPLFF Exponential Tanh 0.99 1.000 0.002

C18 3-12-1 BFGS 20

ed MLPFF Exponential Tanh 0.918 0.009 1.000 0.009
pt
C42 3-10-1 RBFTb RBFd Identify Gaussian 0.956 0.004 1.000 0.004
ce

C48 3-11-1 RBFT RBF Identify Gaussian 1.000 0.000 1.000 0.000
Ac

C52 3-12-1 RPFT RBF Identity Gaussian 0.939 0.007 1.000 0.002

a
Broyden-Fletcher-Goldfarb-Shanno
b
Radial Basis Function Network
c
Multi Layer Precision Feed Forward
d
Radial Basis Function
Broyden-Fletcher-Goldfarb-Shanno, b Radial Basis Function Network, c Multi Layer Precision Feed Forward, d Radial Basis Function.

Page 18 of 29
 19

As shown in Figure 2 below, the increase in the number of hidden neurons attains a specific

point frequently resulted in an improved training efficiency. On the other hand, very small

number of hidden neurons tend to constrain the capability of the neural network to successfully

model the process, as well as too much may give access of too much freedom for the weights to

t
alter and adjust hence, to produce the noise present in learning in the database use in training

ip
[37]. The test results from the two sample experiments, estimated statistically based on the

cr
determination coefficient (R2), as shown in Figure 2. In all the tested cases, the optimal number

us
of hidden neuron layers was 11, with a noticeable rise in the computation time and over fitting

as long as too many hidden neuron layers were used. Then the 3-11-1 topology was selected as

an
the best topology for estimation of RON.
M
d
p te
ce
Ac

Figure 2: Optimal number of hidden neurons tested with two different networks
(MLP & RBF) and two different transfer functions

Page 19 of 29
 20

3.4 Effect of learning algorithm and transfer function

Many learning algorithms were tested for training neural network models. Among them such

model (RMSE < 0.01, R = 1 and DC = 1) have shown that (RBFT) was the most fit learning

algorithms for prediction of RON as it can be seen in Table 5 [38]. The structure of employed

transfer function influences the neural network learning rate and is instrumental in its

t
ip
performance. In the present report, different transfer functions such as Linear, Gaussian,

cr
Logistic, Tanh and exponential for the hidden and output layers were iteratively tested and

determined. Every ANN model was trained to the minimum mean square error (MSE) of a value

us
lower than 0.00001 has been reached. As shown in the figure below, along with all employed

transfer functions for hidden and output layers, reasonable models (MSE < 0.01, R = 1 and

an
DC=1) were formed by the identity function for output layer and Gaussian function for hidden

layer that among them, the top models have been found by Gaussian function.
M
3.5 Artificial Neural Network analysis of RON in n-octane reforming
d

Radial Basis Function network (RBFT) with Gaussian transfer function (Table 5, C48) was
te

chosen as the optimal ANN in the present study. C48 network consisted of a 3-11-1 topology as
p

shown in Figure 3 below.

ce
Ac

Page 20 of 29
 21

98
97 RBF (3-11-1)
2
96 R = 1.000
95
94
93
92
Predicted (Output)

91

t
90

ip
89
88

cr
87
86
85

us
84
83
82
81

an
80
79
79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98
M
Experimental (Target)

Figure 3: Experimental data plotted against the predicted data given by ANN model.
d
te

The training was satisfied in RMSE = 9.99 E-10, R2=1. In this case study, the training data set,
p

the coefficient of determination (R2) and absolute average deviation (AAD) were 1, whereas for
ce

the testing data set, R2 was 1 as it can be seen in Table 5 above. The comparison between the

predicted and experimental values in training, testing and validating data sets, not only revealed
Ac

capability of ANN in the prediction of known data responses, but also showed the ability of

generalization for unknown data, which imply that empirical models derived from ANN can be

utilized to sufficiently describe the relationship between the input factors and output in

reforming of n-octane.

Page 21 of 29
 22

3.6 Comparison of RSM and ANN predicted values

Table 6 below, shows the predicted output values of RSM and ANN. Though both the models

were based on RSM and ANN performance and offered stable responses in predicting the

combined interactions of the independent variables with respect to the response, yet the ANN

based approach was better fitting to the measured response in comparison to the RSM model.

t
ip
cr
Table 6: Predicted RON by ANN and RSM Models Along With Absolute
Deviation

us
Actual ANN Predicted ANN Absolute RSM Predicted RSM Absolute
RON RON deviation RON deviation
2 88.0 88.0 0.00 85.2 2.8

an
4 81.4 81.4 0.00 80.6 0.7
5 94.3 94.3 0.00 94.0 0.2
6 95.8 95.8 0.00 93.2 2.5
M
9 85.7 85.7 0.00 94.7 0.0
10 81.3 81.3 0.00 81.8 0.5
d

11 93.7 93.7 0.00 94.4 0.7

12 91.5 91.5 0.00 94.1 2.6
te

13 88.3 88.3 0.00 86.9 1.3

14 90.5 90.5 0.00 88.1 2.3
p

15 80.8 80.8 0.00 84.5 3.7

ce

16 85.7 85.7 0.00 85.7 0.0

1 88.6 88.6 0.00 88.7 0.1
Ac

3 82.3 82.3 0.00 83.8 1.5

7 88.7 88.7 0.00 90.5 1.8

3.7 Effect of operating variables

The 3D design response surfaces are the graphical representations of the regression equations.

Figure 4 and 5 represent the 3-D plot of the RON model in an estimated response over the

process conditions (reaction temperature and operating pressure) depicted at a fixed space

velocity an at reaction temperature and space velocity at a fixed pressure respectively. From the

Page 22 of 29
 23

figures its revealed that the RON enhancement of up to 110 as the temperature increased up to

530oC at operating pressure in the range between 5-10 bar (Figure.4) and at a space velocity in

the range of 0.8-2.2 (Figure.5). This can be attributed to the fact that raising the reaction

temperature at low operating pressure, thermodynamically favored the aromatization activity as

well as the formation of Olefin hydrocarbons. However, these hydrocarbons have high octane

t
ip
numbers that tend to enhance the RON of reformat [39-41]. This confirms the results obtained

cr
from Pareto chart shown in Table 4 which implies that the reaction temperature parameter is the

most influential variable on RON in the catalytic naphtha reforming process.

us
an
M
d
p te
ce
Ac

Figure 4: 3-D Response surface plot for the design RON as function of reaction
temperature and operating pressure depicted at fixed space velocity

Page 23 of 29
 24

t
ip
cr
us
an
M
d

Figure 5: Response Surface Plot of RON yield as a Function of Reaction

te

Temperature and LHSV at Operating Pressure = 20 bar

p

Figure 6 displays the 3-D plot of the RON with the pressure and space velocity at fixed reaction
ce

temperature. The inclined profile implies the interacting effect between both factors on the

RON. At low pressure and high contact time (low space velocity) the RON is improved up to
Ac

110. On the contrary, a decrease in RON up to 14% at higher pressure was observed. This

situation can be attributed to the fact that increasing pressure suppressed aromatic formation. It

is worthy to mention here that the changes in LHSV lead to a significant effect on RON. From

(RSM) (statistic software), the optimum maximum point for the RON is 112 when the pressure

is 5 to 10 bar and space velocity is between 0.8 to 1.6 h-1.

Page 24 of 29
 25

t
ip
cr
us
an
M
d

Figure 6 :Response Surface Plot of RON yield as a Function of Operating

te

Pressure and LHSV at Reaction Temperature = 495 oC

p
ce

4 Conclusion

A comparison study of the performance of the RSM and ANN in the estimation RON in octane
Ac

reforming was carried out in this report. Both models provide good quality predictions for the

three independent variables (reaction temperature, operating pressure and space velocity) in

terms of the Research Octane Number; however the ANN methodology showed a very obvious

advantage over RSM as a modeling strategy for data sets showing nonlinear relationships. As a

modeling method, artificial neural network was superior to RSM for both data fitting and

estimation capabilities. Conversely, the need of a large quantity of training data is a drawback of

the neural networks method, in contrast with RSM that give a huge sum of information from a

Page 25 of 29
 26

small number of experiments. This benefit of RSM is due to the application of the design of

experiment (DOE). To conquer the ANN disadvantage, in the present study, we utilize the RSM

methodology and then a statistical experimental design based on central composite rotatable

design CCRD was employed to minimize the number of experiments. Accordingly, ANN may

possibly be a very dominant and flexible tool for modeling the optimization process.

t
ip
cr
References

[1] Kontorovich AE (2009) Estimate of global oil resource and the forecast for global oil

us
production in the 21st century. Russ Geol Geophys 50:237–242.

[2] Speight JG (2011) Chapter 10: Refinery of the future. In: The refinery of the future, William

an
Andrew Publishing, Boston, p 315–340.

[4] Iranshahi, D., Karimi, M., Amiri, S., Jafari, M., Rafiei, R., and Rahimpour, M. R. (2014).
Modeling of naphtha reforming unit applying detailed description of kinetic in continuous
M
catalytic regeneration process. Chemical Engineering Research and Design 92, 1704-1727.

[5] Binjuwair, S., Mohamad, T. I., Almaleki, A., Alkudsi, A., and Alshunaifi, I. (2015). The
effects of research octane number and fuel systems on the performance and emissions of a spark
d

ignition engine: A study on Saudi Arabian RON91 and RON95 with port injection and direct
injection systems. Fuel 158, 351-360
te

[6] HafezKhiabani, N., Fathi, S., Shokri, B., and Hosseini, S. I. (2015). A novel method for
p

decoking of Pt–Sn/Al2O3 in the naphtha reforming process using RF and pin-to-plate DBD
plasma systems. Applied Catalysis A: General 493, 8-16
ce

[7] Elfghi, F. M., and Amin, N. (2015). Influence of tin content on the texture properties and
catalytic performance of bi-metallic Pt–Re and tri-metallic Pt–Re–Sn catalyst for n-octane
reforming. Reaction Kinetics, Mechanisms and Catalysis, 114(1), 229-249.
Ac

[8] Moljord K, Grande K, Tanem I, Holmen A (1996) Pilot reactor testing of the effect of
naphtha boiling point in catalytic reforming. In: Deactivation and testing of hydrocarbon-
processing catalysts, American Chemical Society, p 268–282

[9]. D’Ippolito SA, Vera CR, Epron F, Samoila P, Especel C, Mare´cot P, Gutierrez LB, Pieck
CL (2009) Catalytic properties of Pt - Re/Al2O3 naphtha-reforming catalysts modified by
germanium introduced by redox reaction at different pH values. Ind Eng Chem Res 48:3771–
3778

[10]. Ali SA, Siddiqui MA, Ali MA (2005) Parametric study of catalytic reforming process.
Reac Kinet Cat Lett 87:199–206

Page 26 of 29
 27

[11]Poroch-Seritan M, Gutt S, Gutt G, Cretescu I, Cojocaru C, Severin T (2011) Design of

experiments for statistical modeling and multi-response optimization of nickel electroplating
process. Chem Eng Res Des 89:136–147.

[12] Trinh TK, Kang LS (2011) Response surface methodological approach to optimize the
coagulation-flocculation process in drinking water treatment. Chem Eng Res Des 89:1126-1135

[13] Larentis AL, de Resende NS, Salim VMM, Pinto JC (2001) Modeling and optimization of

t
the combined carbon dioxide reforming and partial oxidation of natural gas. Appl Catal A

ip
215:211–224.
[14] Arami-Niya A, Wan Daud WMA, Mjalli FS, Abnisa F, Shafeeyan MS (2012) Production of

cr
microporous palm shell based activated carbon for methane adsorption: modeling and
optimization using response surface methodology. Chem Eng Res Des 90:776–784

us
[15] Busto M, Vera CR, Grau JM (2011) Optimal process conditions for the isomerization–
cracking of long-chain n-paraffins to high octane isomerizate gasoline over Pt/SO42–ZrO2
catalysts. Fuel Process Technol 92:1675–1684.

an
[16] Elfghi, F. M., Amin, N., & Elgarni, M. M. (2015). Optimization of Isomerization Activity
and Aromatization Activity in Catalytic Naphtha Reforming over Tri-Metallic Modified Catalyst
M
using Design of Experiment Based on Central Composite Design and Response Surface
Methodology. Journal of Advanced Catalysis Science and Technology, 2, 1-17.

[17] Antos, G.A., Aitani, A.M., 1995. Catalytic naphtha reforming, 2nded. Marcel Dekker, New
d

York,
te

[18] Bishara A, Stanislaus A, Hussain SS (1984) Effect of feed composition and operating
conditions on catalyst deactivation and on product yield and quality during naphtha catalytic
reforming. Appl Cat 13:113–125.
p

[19] Antos, G. J. and Aittani, A.M. (2004). Catalytic Naphtha Reforming. Second edition
ce

Book, Revised and Expanded (Chemical Industries). (MARCEL DEKKER,INC. NEWYOUR.

BASEL).

[20] Silva, G. F. d., Gandolfi, P. H. K., Almeida, R. N., Lucas, A. M., Cassel, E., and Vargas, R.
Ac

M. F. (2015). Analysis of supercritical fluid extraction of lycopodine using response surface

methodology and process mathematical modeling. Chemical Engineering Research and Design
100, 353-361.

[21] Liao, C.-C., and Chung, T.-W. (2013). Optimization of process conditions using response
surface methodology for the microwave-assisted transesterification of Jatropha oil with KOH
impregnated CaO as catalyst. Chemical Engineering Research and Design 91, 2457-2464.

[22] Eppinger, T., Wehinger, G., and Kraume, M. (2014). Parameter optimization for the
oxidative coupling of methane in a fixed bed reactor by combination of response surface
methodology and computational fluid dynamics. Chemical Engineering Research and Design
92, 1693-1703

Page 27 of 29
 28

[23] Elfghi, F. M., & Amin, N. (2013). Optimization of hydrodesulfurization activity in the
hydrotreating process: canonical analysis and the combined application of factorial design and
response surface methodology. Reaction Kinetics, Mechanisms and Catalysis, 108(2), 371-390.

[24] Abbasi, M., Soleymani, A. R., and Parssa, J. B. (2014). Operation simulation of a recycled
electrochemical ozone generator using artificial neural network. Chemical Engineering Research
and Design 92, 2618-2625

[25] Khayet, M., Cojocaru, C., and Essalhi, M. (2011). Artificial neural network modeling and

t
response surface methodology of desalination by reverse osmosis. Journal of Membrane

ip
Science, 368(1-2), 202-214.

cr
[26] Soleimani, R., Shoushtari, N. A., Mirza, B., and Salahi, A. (2013). Experimental
investigation, modeling and optimization of membrane separation using artificial neural network
and multi-objective optimization using genetic algorithm. Chemical Engineering Research and

us
Design 91, 883-903

[27] Desai, K. M., Survase, S. A., Saudagar, P. S., Lele, S. S., and Singhal, R. S. (2008).
Comparison of artificial neural network (ANN) and response surface methodology (RSM) in

an
fermentation media optimization: Case study of fermentative production of scleroglucan.
Biochemical Engineering Journal, 41(3), 266-273.

[28] Tompos, A., Margitfalvi, J. L., Tfirst, E., and Végvári, L. (2003). Information mining using
M
artificial neural networks and "holographic research strategy". Applied Catalysis A: General,
254(1), 161-168.

[29] Franceschini, G. and Macchietto, S. (2008). Model-based design of experiments for

d

parameter precision: State of the art. Chemical Engineering Science. 63:4846- 4872.
te

[30] Bezerra, M.A., Santelli R.E., Oliveira, E.P., Villar, L.S., Escaleira, L.A., 2008. Response
surface methodology (RSM) as a tool for optimization in analytical chemistry, Talanta. 76(5),
965–977.
p

[31] Bhatti, M. S., Kapoor, D., Kalia, R. K., Reddy, A. S., and Thukral, A. K. (2011). RSM and
ce

ANN modeling for electrocoagulation of copper from simulated wastewater: Multi objective
optimization using genetic algorithm approach. Desalination, 274(1-3), 74-80.

[32] Cetina, D. M., Giraldo, G. I., and Orrego, C. E. (2011). Application of response surface
Ac

design to solvent, temperature and lipase selection for optimal monoglyceride production.
Journal of Molecular Catalysis B: Enzymatic, 72(1-2), 13-19

[33] Trinh, T. K., & Kang, L. S. (2011). Response surface methodological approach to optimize
the coagulation–flocculation process in drinking water treatment. Chemical engineering research
and design, 89(7), 1126-1135.

[34] Hamzaoui, A. H., Jamoussi, B., and M'nif, A. (2008). Lithium recovery from highly
concentrated solutions: Response surface methodology (RSM) process parameters optimization.
Hydrometallurgy, 90(1), 1-7.

Page 28 of 29
 29

[35] Shih, I. L., and Shen, M. H. (2006). Application of response surface methodology to
optimize production of poly-[var epsilon]-lysine by Streptomyces albulus
IFO 14147. Enzyme and Microbial Technology, 39(1), 15-21.

[36] Zahedi, G., Mohammadzadeh, S., and Moradi, G. (2008). Enhancing Gasoline Production
in an Industrial Catalytic-Reforming Unit Using Artificial Neural Networks. Energy & Fuels,
22(4), 2671-2677.

[37] Nazghelichi, T., Aghbashlo, M., and Kianmehr, M. H. (2011). Optimization of an artificial

t
neural network topology using coupled response surface methodology and genetic algorithm for

ip
fluidized bed drying. Computers and Electronics in Agriculture, 75(1), 84-91.

cr
[38] Chen, S., Cowan, C.F.N., and Grant, P.M. 1991. Orthogonal Least Squares Learning
Algorithm for Radial Basis Function Networks, Ieee Transactions On Neural Networks. 2(2),
302-309.

us
.
[39] Han, W.-Q., and Yao, C.-D. (2015). Research on high cetane and high octane number fuels
and the mechanism for their common oxidation and auto-ignition. Fuel 150, 29-40.

an
[40] Rankovic, N., Bourhis, G., Loos, M., and Dauphin, R. (2015). Understanding octane
number evolution for enabling alternative low RON refinery streams and octane boosters as
transportation fuels. Fuel 150, 41-47.
M
[41] Elfghi, F. M., and Amin, N. (2014). Applying response surface methodology to assess the
combined effect of process variables on the composition and octane number of reformat in the
process of reducing aromatization activity in catalytic naphtha reforming. Reaction Kinetics,
Mechanisms and Catalysis, 111(1), 89-106.
d
p te
ce
Ac

Page 29 of 29