Advances in Engineering Software 92 (2016) 15–26

Contents lists available at ScienceDirect

Advances in Engineering Software

journal homepage: www.elsevier.com/locate/advengsoft

Performance evaluation of a petrol station queuing system:

A simulation-based design of experiments study
Masoud Rahiminezhad Galankashi∗, Ehsan Fallahiarezoudar, Anoosh Moazzami,
Noordin Mohd Yusof, Syed Ahmad Helmi
Department of Material, Manufacturing and Industrial Engineering, Faculty of Mechanical Engineering, Universiti Teknologi Malaysia, 81310 UTM Johor Bahru,
Johor, Malaysia

a r t i c l e i n f o a b s t r a c t

Article history: The main goal of this paper was to develop an integrated simulation-design of experiments (DOE) model
Received 24 June 2015 to optimize a petrol station queuing system and sales rate. Initially, the petrol station operating system was
Revised 8 September 2015
simulated using Witness 2014 simulation software©. Then, the responses of simulation were deployed as the
Accepted 18 October 2015
input of DOE. Two-level full factorial experiments with center points were performed where the simulated
model parameter studied were number of pump, number of cashier and inter arrival times (IATs). The re-
Keywords: sponse variables analyzed were queue length and sales rate. The obtained model from experimental design
Design of experiments revealed that number of cashier and inter arrival time were significant in determining the queue length while
Simulation all the factors and their interaction were significantly affecting the sales rate.
Petrol station
© 2015 Elsevier Ltd. All rights reserved.
Queuing system
Service industry

1. Introduction ate the effect of significant factors on queuing system. Proposing a

Service industries recognized that improving and optimizing their
service level is extremely required to increase the efficiency in com-
petitive markets [1]. Customer satisfaction has been announced as
one of the most significant parameters of making competitive ad-
vantage in service industries [2–5]. Many people use petrol stations’
services in their daily life. Petrol companies can increase their profit
through satisfying their customers. In this industry, competitive ad-
vantage can be translated into three perspectives of petrol quality
and service speed as well as price. Since the quality and price are
commonly comparable in majority of the markets, speed service and
consequently queue length can be considered as the most influential
factor on customer satisfaction and correspondingly the revenue [6].
Simulation modeling, mathematical programming and statistics
can be considered as useful tools for all managers, researchers and
practitioners to analyze dynamic systems without interrupting their
operations [7–9]. In addition, simulation modeling of real world cases
and its integration with DOE models is a challenge for many practi-
tioners, managers, and researchers [10–14].
Since the DOE is an active statistical technique, its integration
with simulation can be used to predict a model which can evalu-
∗
Corresponding author. Tel.: +60 7 553 4645.
E-mail addresses: Rgmasoud2@live.utm.my, masoud.rahiminejad@yahoo.com
(M.R. Galankashi), helmi@fkm.utm.my (S.A. Helmi).
simulation-DOE model for the aim of assessing petrol station queu-
ing system and correspondingly sales rate can be beneficial because;
(i) Making critical decisions is a serious challenge to top manage-
ment in different service industries, (ii) Performance of service in-
dustry should be exactly assessed due to numerous restrictions in
time, cost, labor and different sources of energy, and (iii) Using dif-
ferent tools to optimize processes deprived of interrupting the sys-
tem operations and evaluate their influence before implementation
[15–17].
Although the application of simulation in service industries has
been discussed by many authors [18–22], only a very few attempts
have been made to integrate it with design of experiments. Based on
our best of knowledge, there is no integration on evaluating the per-
formance of petrol station queue system.
This study contributes to integrate simulation modeling with DOE
in order to analyze and optimize a petrol station performance. It in-
troduces a new idea for using computer simulation and proposing
different scenarios as the input of DOE. This approach provides a
valuable contribution because it is impossible to stop the operating
system or change the layout due to constraints of cost, time, labor
and many other factors. The novelty of this study lies on integration
of simulation and design of experiments in a petrol station which
will lead to obtain a predictable model in order to optimize the best
scenario for a petrol station considering two perspectives of queue
length and sales rate.

http://dx.doi.org/10.1016/j.advengsoft.2015.10.004
0965-9978/© 2015 Elsevier Ltd. All rights reserved.
16 M.R. Galankashi et al. / Advances in Engineering Software 92 (2016) 15–26
2. Literature review
Simulation has been practiced to numerous segments including
services, manufacturing, healthcare, defense and public services [23–
27]. It is known as one of the most proper commonly used approaches
in the field of operations management [28–32]. The appropriateness,
suitability and relevance of simulation methods is a significant is-
sue to study in applied real-world applications, chiefly as there is
a rising necessity to address the difficulties of the entire enterprise
[19,33–36].
In any service organization, managers are mostly concerned about
the customers and their profit simultaneously [13,37,38]. The impor-
tant structures of a standard queuing system contain line structure,
demand group, entrance and service procedures, and queue disci-
pline [36,39,40]. The majority of existing studies in service industry
focus on maximizing the customer satisfaction. The type and qual-
ity of demand or quantity of customers, serving priority, the tolerable
queue length, the bearable waiting time are the major factors which
can affect customer satisfaction. The waiting time and queue length
are two important issues that play significant roles in customer in-
sight about the service quality. Though, to have an optimal service
configuration, both the customer satisfaction and enterprise revenue
should be considered simultaneously. Consequently, numerous deci-
sions should be evaluated to attain the best likely scenario which is
tolerable from both the customer’s and service provider’s view point
[41–44].
Different approaches have been applied to progress service qual-
ity and subsequently customer satisfaction in petrol service indus-
try. Previously, Cornillier, Boctor [45] developed an exact algorithm
for the petrol station replenishment problem. Moazzami et al. [9], fo-
cused on simulation, modeling and analysis of a petrol station where
a petrol station behavior was simulated as one of the most significant
sections of service industry.
Design of experiments (DOE) is a mathematical, statistical and
systematic technique in order to determine the relationship between
affecting process factors and the output of that process. In other
words, it is utilized to find cause-and-effect and interaction be-
tween parameters where in one-factor-at-a-time approach not pos-
sible. Analysis of DOE results is essential to manage process inputs in
order to optimize the process output [46–48].
Cheng and Kleiinen [49], established optimal DOE with simula-
tion models of nearly saturated queues. The application of computer
simulation have been suggested and executed to solve the problems
of variation in incorporated manufacturing systems. Though, a sim-
ulation model merely acts as a device in investigating performance.
Tsai [31], focused on assessment and optimization of joined manu-
facturing system operations with the aim of experimental design in
computer simulation. The results show that this approach can con-
sider the assessment and optimization of operating situations in mul-
tifaceted systems concurrently.
A comparison of experimental designs for simulation-based sym-
bolic regression of manufacturing systems was provided by Can
and Heavey [50]. The objective was to identify a robust sam-
pling approach. In a very recent study, Li et al. [13] developed a
Simulation-based experimental design and statistical modeling for
lead time quotation. This work developed a simulation-based sta-
tistical method to offer high-quality and responsive forecast of a
new job’s flow time through the system, that reduces the ability
of precisely quoting lead times in real time. The method incorpo-
rated analytical queuing analysis, design of experiments, and statisti-
cal modeling.
Previous studies on this field show that the simulation results can
be used as an input to design of experiment. Simulation and DOE are
some tools to analyze the behavior of a system. In this study the be-
havior of the petrol station queuing system was simulated and the
model outputs were used as the raw data of DOE.
According to the previous literature, this is the first attempt to an-
alyze a petrol station queuing system using an integrated simulation-
DOE model. Trying to fill the gap in the literature, this paper
proposes an integrated simulation-DOE model to analyze and opti-
mize the queuing system of a petrol station by considering differ-
ent scenarios, in order to assist managers to efficiently manage their
enterprise.
3. Case study
The case study is located in Skudai, Johor, Malaysia. It consists of
two main platforms. Each of these platforms comprises four fuel dis-
pensers. Every fuel dispenser comprises two nozzles that pump two
types of Octane 95 and Octane 97 fuels. There is also a supermarket
in which customers do their shopping (if any) while refueling their
cars. Two cashiers work in this station; one of them performs cus-
tomers’ payments for those who just do the refueling process. The
other cashier performs the shopping related affairs as well as pay-
ments for the fuel. Customers have the chance of paying the fuel price
(Octane 95 or 97) either by cash or with a credit card.
3.1. Petrol station layout and process mapping
The petrol station layout is shown in Fig. 1. ‘D1–D8’ indicates the
number of pumps. Conceptual model is used to map the whole pro-
cess. The whole process and inputs flow through the output is de-
picted in the process map as Fig. 2. The process commences with cars
arrival to the station and chooses the desired pump among 8 available
options. Their selection would be influenced by the least quantity of
cars in the queue and distance to the counters if they cannot find an
empty queue.
The next process is the payment task which involves paying
money refers to desired fuel type (petrol 95 or 97). Once the cashier
received the payment, the pump will be replenished exactly as the
payment quantity (Octane 95 or 97). Refueling process starts once the
drivers comeback to the selected pump and pick the desired nozzle
(Octane 95 or 97). Afterward, the pump will refill the fuel tank equal
to payment. Once the refueling process ended, drivers put the nozzle
back and leave the system.
4. Model development
The problem in this case can be classified into two categories: first,
it is related to the service level which petrol station provide to the
customers and the second one, refers to the pumps and cashiers idle
times. In addition, based on the observations, the IATs, fuel and pay-
ment choice varies depending on the time and the type of cars. This
problem worsens in unusual times such as rush hours or holidays.
Consequently, top managers may be dissatisfied with the efficiency of
the petrol station due to its crowdedness or under capacity usage. In
addition, managers have to pay for maintenance, cashiers and many
other expenses which can be avoided by optimizing the petrol station
queuing system.
4.1. Model assumption
Following assumptions are considered in simulation modeling:
• All customers have 4 choices to select from; petrol types (Octane
95 or 97) and payment method (by cash or credit card).
• No customer leaves the system after entering the Queue.
• Shopping from petrol station supermarket was considered in the
system.
• There is no jockeying in the system (changing the Queue lines).
• The observation process (data collection) was completed in nu-
merous days of the week and different hours of the day.
• Some fluctuations in the petrol rates were ignored.
 M.R. Galankashi et al. / Advances in Engineering Software 92 (2016) 15–26 17

Fig. 1. Petrol station layout.

Fig. 2. Process mapping diagram.

4.2. Data collection
The required data was collected and recorded regarding to the
process map. Stop-Watch technique was deployed for collecting and
recording the required data. All data were collected in 4 h periods
(from 4:00 A.M. to 11:59 P.M.) during 4 weeks. This long-term pe-
riod was helpful to consider the rush hours (weekends, holidays and
etc.) and consequently the normality of gathered data. After gather-
ing enough sets of interrelated data, a probability distribution was
fitted to prove the process randomness. The process of data collec-
tion did not interrupt the operating system.
The inter arrival time is the time between each car arrival into the
petrol station and the next. The cars were followed to check whether
they refuel Octane 95 or 97 to record their IATs, to be distinctly con-
sidered in simulation model. Cashier time is the next element of the
model. There were two cashiers in the model. Their process time was
recorded to be used in cashiers’ elements of model. Refueling time is
the duration time when the drivers pick the nozzles till put it back.
The cycle time of each pump is founded to be a dependent variable of
refueling quantity and independence of total required time of leaving
the system. Therefore, the fuel consumption data was collected ac-
cordingly. Each fuel type sales quantity (based on liter) was collected
to be used for verification process (see Section 6).
EASYFIT© software was used for data analysis and fit the proper
probability distribution. The collected data of each activity was
put into EASYFIT© and tested for numerous dissimilar continuous
18 M.R. Galankashi et al. / Advances in Engineering Software 92 (2016) 15–26

Table 1
Probability distributions fitting of collected data.

Analysis of collected data Distribution Analysis of collected data Distribution

95 cash+ shop (IAT) 97 cash (IAT)

Neg. exp (5) Neg.exp (25.22)
Sample size=178 Sample size=150
Kolmogorov–Smirnov Kolmogorov–Smirnov
Statistic= 0.04891 Statistic=0.04377
P-value= 0.76923 P-value=0.92386
Anderson–Darling Anderson–Darling
Statistic= 0.71988 Statistic=0.39626
Chi-squared Chi-squared
Statistic= 2.025 Statistic=3.3975
P-value= 0.95844 P-value=0.84595

95 cash (IAT) 97 card+shop (IAT)

Neg.exp (3.5) Neg.exp (155.25)
Sample size=200 Sample size=40
Kolmogorov–Smirnov Kolmogorov–Smirnov
Statistic= 0.06186 Statistic=0.07846
P-value = 0.41169 P-value =0.95009
Anderson–Darling Anderson–Darling
Statistic= 0.78033 Statistic=0.42903
Chi-squared Chi-squared
Statistic= 4.5179 Statistic=3.1995
P-value = 0.71856 P-value =0.66926

95 card+shop (IAT) 97 card (IAT)

Neg. exp (36.5) Neg. exp (130.28)
Sample size=146 Sample size=56
Kolmogorov–Smirnov Kolmogorov–Smirnov
Statistic=0.05202 Statistic=0.08733
P-value =0.80148 P-value =0.75355
Anderson–Darling Anderson–Darling
Statistic=0.45279 Statistic=0.71192
Chi-squared Chi-squared
Statistic=5.643 Statistic=3.654
P-value = 0.58199 P-value =0.60023

95 card (IAT) Cashier 1 (process)

Neg. exp (45) N (0.6757,0.0985)
Sample size=76 Sample size=200
Kolmogorov–Smirnov Kolmogorov–Smirnov
Statistic= 0.13023 Statistic=0.0467
P-value = 0.38318 P-value =0.75746
Anderson–Darling Anderson–Darling
Statistic= 0.71699 Statistic=0.42349
Chi-squared Chi-squared
Statistic= 5.9798 Statistic=6.8572
P-value =0.30819 P-value =0.4439

97 cash+shop (IAT) Cashier 2 (process)

Neg. exp (38.39) Normal (1.1684,0.1)
Sample size=89 Sample size=200
Kolmogorov–Smirnov Kolmogorov–Smirnov
Statistic=0.10842 Statistic=0.03626
P-value =0.22917 P-value =0.94656
Anderson–Darling Anderson–Darling
Statistic=1.0888 Statistic=0.32444
Chi-squared Chi-squared
Statistic=10.354 Statistic=2.8708
P-value =0.11051 P-value =0.89669
 M.R. Galankashi et al. / Advances in Engineering Software 92 (2016) 15–26 19

Fig. 3. Schematic of the model.

Table 2
Model verification results.

No. Station % Busy Replication time Utilized time Mean No. of operations No. of operations % Difference
Name (min) (min) (min) (Manually) (Witness report)

1 P1 7.71 1440 111.024 1.356 72 70 2.780

2 P2 9.21 1440 132.624 1.356 98 108 9.440
3 P3 8.03 1440 115.632 1.356 85 89 4.706
4 P4 8.75 1440 126.000 1.356 93 87 6.451
5 P5 13.29 1440 191.376 1.356 141 147 3.990
6 P6 11.42 1440 164.448 1.356 121 120 1.060
7 P7 8.33 1440 119.952 1.356 88 89 1.130
8 P8 7.46 1440 107.424 1.356 79 82 3.797
Average % difference 4.17

probability distributions. Chi-square, Anderson–Darling and Kol-
mogorov–Smirnov tests were used simultaneously to analyze the
data to fit the proper probability distribution. Table 1 tabulates the
best probability distributions of each element.
4.3. Model construction
Different components of the model and their interactions were
constructed in Witness© 2014 software. The model sequence and
process was checked with process mapping to demonstrate the pre-
cise relations, orders of different elements and actions. The model
elements include their fitted distributions supplied by EASYFIT© to
run the whole model in Witness© environment. The scheme of the
simulated model in Witness environment is depicted in Fig. 3.
5. Model verification
Verification aims to check the accuracy of developed model. It
compares the software representation of the model with the real
world operating system. Subsequent to construction of model, its be-
havior was checked visually, to be reasonably similar to real-world
condition [51].
Total number of operations was selected to inspect the model’s
reasonableness and verification. Initially, it was computed manu-
ally and subsequently compared to the result of simulation model.
Table 2 illustrates that the variation percentage is less than 5% (the
satisfactory level) which is desirable.
6. Model validation
Validation is apprehensive with construction of the precise model.
It is used to conclude that a model is a precise illustration of the actual
system [52]. Validation is typically attained by calibrating the model.
This procedure is repetitive till the accuracy of model to be satisfac-
tory. The required number of replications was found by Eq. (1) which
was proposed by Ahmed [53] prior to the model validation. The num-
ber of replications to obtain a desirable accuracy is four replicates or
greater.
 2
S(m)tm−1,(1−( α ))
N(m) =
2
(1)
X̄(m)ε
Where N(m) is number of simulation runs to achieve the desired level
of accuracy; X̄(m) is the mean estimate of an initial number of runs
20 M.R. Galankashi et al. / Advances in Engineering Software 92 (2016) 15–26

Table 3
Validation of model.

Replicate X(m) S(m)

1 2 3 4 5

Actual daily sales rate (l) 10261 10933 10971 11074 11118 10871.4 349.34

Table 4 7. Experimental design

Model validation.

Replicates Actual daily Simulated daily Variation The obtained results of simulation modeling were used for fur-
sales rate (l) sales rate (l) ther analysis by design of experiments technique. In this case DOE
1 10261 10261.1 0.00097
is deployed as a mathematical and statistical method in order to an-
2 10933 11679.8 6.83000 alyze, model and optimize the performance of the simulated petrol
3 10971 11017.8 0.42266 station. The petrol station investigated in this research belongs to one
4 11074 11661.8 5.30793 of the largest and most reputable petrol companies in Malaysia. The
5 11118 10776.4 3.07249
Company’s management board decided to improve their service level
Average variation 3.12681
by investigating their queuing system and maintain their competi-
tive edge. Initially, the company’s managers and executives agreed to
Table 5 analyze different factors that could potentially lead to a better queu-
Process factors and levels during simulation. ing system and improve the customer satisfaction. The potential fac-
No. Factors/levels Unit Low Center High tors were extracted from previous literature and customer feedbacks.
(−20%) (0) (+20%) Next, all these factors were investigated and finalized to be reviewed
using a survey. Finally, by distributing a survey and discussing its re-
1 Number of pump (Number) 6 8 10
2 Number of cashier (Number) 1 2 3 sults, managers agreed to consider three mostly influential factors in-
3 Inter arrival times (IATs) (Time) −1 0 1 volved number of cashiers, number of pumps as well as IATs. Here, a
full factorial design involving three factors with two (high and low)
levels (23 , 3FI) was used to explore the significance effect of parame-
Table 6
Various elements’ IATs levels during simulation.
ters on queue length and sales rate. Five center points as the original
simulated model were considered in order to test the linearity as-
Elements/level −1 0 +1 sumption. In order to evaluate the performance of the petrol station
95 cash + shop Neg. exp.(4.0000) Neg. exp.(5.0000) Neg. exp.(6.0000) refer to sales rate and queue length, the number of pump, number of
95 cash Neg. exp.(2.8000) Neg. exp.(3.5000) Neg. exp.(4.2000) cashier and inter arrival times (IATs) were investigated as three fac-
95 card + shop Neg. exp.(29.200) Neg. exp.(36.500) Neg. exp.(43.800) tors in the design. The number of pumps and number of cashiers in
95 card Neg. exp.(36.000) Neg. exp.(45.000) Neg. exp.(54.000) current simulated model was 8 and 2 respectively, where the range of
97 cash + shop Neg. exp.(30.712) Neg. exp.(38.390) Neg. exp.(46.068)
97 cash Neg. exp.(20.176) Neg. exp.(25.220) Neg. exp.(30.264)
each factor for new design sets as +20% upper for high level and −20%
97 card + shop Neg. exp.(124.20) Neg. exp.(155.25) Neg. exp.(186.30) lower for low levels. Table 5 tabulated the mentioned factors and lev-
97 card Neg. exp.(104.22) Neg. exp.(130.28) Neg. exp.(156.36) els. The response variables investigated were queue length and sales
rate. The IATs has different nature from the others factors and should
be determined for each entrance type separately, so the IATs values
m; S(m) is the standard deviation estimate of m number of runs; α is in various levels have been defined as a statistical distribution. The
level of confidence; ε is allowable percentage error; and tm−1,(1−( α )) IATs in high level dedicates a lower arrival time of customers while
2 in low level dedicates a faster arrival time. Table 6 illustrates the high
is critical value of the two-tailed t-distribution at a level of signifi-
and low levels of IATs for each element enters to the system sepa-
cance, given m − 1 degrees of freedom. The values of X(m) and S(m)
rately. In order to verify the accuracy of collected data, 5 replicates
is shown in Table 3.
have been considered for each run. Based on the factorial design, 8
Model validation was done after determining the required num-
more models (according to the mentioned runs in Table 7) were nec-
ber of simulation replicates. The actual daily sales rate was compared
essary to simulated in order to supply the input data of DOE. In addi-
with the results of the five simulation replicates. Table 4 tabulated
tion, current simulation model results were used as the center point
the data for validation. It shows that the average variation percentage
data. Design Expert version 9 software© was utilized to develop the
of five simulation replicates is 3.12681 (less than 5%) which is satis-
factorial design and analyze the collected data. Table 7 illustrates the
factory.

Table 7
Completed design layout and experimental results.

Std.Run. No. Run order Factors Response: queue length (number) Response: sales rate (l)

No. Pump-A No. Cashier-B IATs-C R1 R2 R3 R4 R5 R1 R2 R3 R4 R5

1 5 6 1 −1 8 9 13 10 8 10970.2 11452.7 11016.4 11433.3 12410.9

2 7 10 1 −1 5 13 7 16 10 11125.0 11380.7 11693.2 10792.7 12208.7
3 9 6 3 −1 1 2 2 2 1 14083.1 13703.6 13710.4 13271.2 13646.2
4 8 10 3 −1 2 2 2 2 2 13466.3 14182.1 13956.3 13791.1 13444.2
5 1 6 1 +1 2 4 3 2 3 9654.07 9611.73 9409.38 8886.02 8830.78
6 3 10 1 +1 2 3 3 5 2 8515.13 9025.21 9148.83 9198.58 9450.71
7 6 6 3 +1 1 1 1 1 1 9774.84 9125.28 9523.83 9523.83 8745.03
8 2 10 3 +1 1 1 1 1 1 10407.1 11325.5 11866.1 11342.9 11268.9
9 4 8 2 0 1 1 1 2 2 10261.1 11679.8 11017.8 11661.8 10776.4
 M.R. Galankashi et al. / Advances in Engineering Software 92 (2016) 15–26 21

Table 8
ANOVA table (partial sum of squares) for response surface (queue length).

Source Sum of squares DF Mean square F-value Prob > F

Model 499.6 7 71.37142857 23.96802 < 0.0001 Significant

A 0.9 1 0.9 0.302239 0.5859
B 250 1 250 83.95522 < 0.0001 Significant
C 152.1 1 152.1 51.07836 < 0.0001 Significant
AB 0.1 1 0.1 0.033582 0.8556
AC 0.4 1 0.4 0.134328 0.7161
BC 96.1 1 96.1 32.27239 < 0.0001 Significant
ABC 0 1 0 0 1.0000
Curvature 27.78 1 27.78 9.328358 0.0042 Significant
Pure error 107.2 36 2.98
Cor total 634.58 44

Table 9
ANOVA table (partial sum of squares) after model reduction (queue length).

Source Sum of Squares DF Mean Square F-value Prob > F

Model 498.2 3 166.07 61.17 < 0.0001 Significant

B 250 1 250 92.08 < 0.0001
C 152.1 1 152.1 56.02 < 0.0001
BC 96.1 1 96.1 35.40 < 0.0001
Curvature 27.78 1 27.78 10.23 0.0027 Significant
Residual 108.6 40 2.715
Lack of fit 1.4 4 0.35 0.12 0.9754 Not significant
Pure error 107.2 36 2.98
Cor total 634.58 44

Std. dev. 1.65 R-squared 0.8210

Mean 3.62 Adj R-squared 0.8076
C.V. 45.49 Pred R-squared 0.7881
PRESS 134.47 Adeq precision 16.204

factorial design and measured queue length and sales rate as the re-
sponse variables.
8. Results and discussion
Table 8 depicts the ANOVA table for response of queue length.
The value of “Prob> F” for model is less than 0.05 that indicates the
model is significant, which is desirable and considerably affect the re-
sponse surface of queue length. In the same manner, the main effect
of number of cashier-(B), inter arrival times (IATs)-(C) and the two-
level interaction of number of cashier and IATs (BC) are significant
model terms. Other model terms can be expressed to be not signifi-
cant. These insignificant model terms can be removed and may result
in an improved model.
The resulting ANOVA table after model reduction (where the in-
significant terms are eliminated from the regression model) is pre-
sented in Table 9. Here the F-value of 61.17 for the model indicates
that the model is still significant, which is desirable. The lack of fit
statistics defined as the ratio of lack-of-fit mean square over the pure
error mean square [54]. The value of lack-of-fit should be insignifi-
cant. Here the lack-of-fit ‘F-value’ is 0.12 which is insignificant and
correspondingly desirable. An R-square value of 0.8210 indicates that
the model is able to explain about 82.10% of the variation on the
response. The Adj-R2 and Predicted-R2 are in reasonable agreement
with each other (difference less than 0.2). ‘Adeq. precision’ measures
the signal to noise ratio. A ratio greater than 4 is desirable [55]. Ratio
of 16.204 indicates an adequate signal. According to the ANOVA re-
sults it can be concluded that this model can be used to navigate the
design space.
The significance curvature implies that there is a nonlinear re-
lationship between factors and queue length. Thus, four further ex-
periments have been added to the previous design for obtaining a
second-order regression model in order to justify the available curva-
ture. Four axial points on the center of significant factors edge were
augmented. Since the number of pump-(A) is insignificant the values
for the runs considered as the center value. The experiments design
for augmentation and the related result tabulated in Table 10.
These four runs were added to the previous 9 runs and the ANOVA
was tested again. The result after insignificant model terms reduc-
tion (not including those required to support hierarchy) is tabulated
in Table 11. The Prob.>F of parameters indicates that the second or-
der effect of number of pump (A)2 is added to the previous significant
terms in quadratic model. As before, the associated statistics such as
R-square, Adj.-R2 and Predicted R2 , and ‘Adeq.precision’ all are desir-
able. The empirical model was determined based on the significant
factors and is presented in terms of coded and actual factors as fol-
lows:
Ŷ = 1.40 + 0.15 × A − 2.06 × B − 1.54
×C + 2.50 × A2 + 1.55 × B × C (2)
Queue Length = 54.20 − 9.925 × pump − 5.160 × chashier
− 4.640 × IAT + 0.625 × pump2 + 1.55
× chashier × IAT (3)
The same procedure is applied on response of sales rate. The
ANOVA table result depicts in Table 12. Here, the Prob.>F values
revealed that the main effects of number of pump-(A), number of
cashier-(B), IATs-(C) and two way interaction of AB, AC, BC and three
way interaction of ABC have significant effects on sales rate. Since
all the main effect and their interaction are significant on response
surface of sales rate, so the model reduction procedure is neglected.
Here, the F-value of 83.36 for the model indicates that the model is
significant, which is desirable. Since the model reduction was un-
necessary, the lack-of-fit statistics are removed from the results. An
R-square value of 0.9419 indicates that the model is able to explain
about 94.19% of the variation on the sales rate. The Adj. R2 and
Predicted-R2 are in reasonable agreement with each other (difference
22 M.R. Galankashi et al. / Advances in Engineering Software 92 (2016) 15–26

Table 10
Augmentation test and results of design.

Std. run Factors Response (queue length)

No. pump-(A) No. cashier-(B) IATs- (C) R1 R2 R3 R4 R5

1 8 2 −1 2 1 1 2 1
2 8 2 +1 1 2 1 2 2
3 8 1 0 1 1 3 2 1
4 8 3 0 1 1 1 1 1

Table 11
ANOVA table (partial sum of square) for response surface model after augmentation.

Source Sum of squares DF Mean square F-value Prob > F

Model 523.9138462 5 104.78 33.216 < 0.0001 Significant

A 0.9 1 0.9 0.2853 0.5950
B 212.18 1 212.18 67.365 < 0.0001
C 118.58 1 118.58 37.647 < 0.0001
A2 96.15 1 96.15 30.567 < 0.0001
BC 96.1 1 96.1 30.558 < 0.0001
Residual 185.84 59 3.15
Lack of fit 73.04 7 10.43 1.8132 0.6852 Not significant
Pure error 112.8 52 2.17
Cor total 709.75 64

Std. dev. 1.77 R-squared 0.7382

Mean 2.94 Adj R-squared 0.7160
C.V. 60.40 Pred R-squared 0.6790
PRESS 227.81 Adeq precision 18.286

Table 12
ANOVA table (partial sum of squares) for response surface (Response: sales rate).

Source Sum of squares DF Mean square F-value Prob. > F

Model 122879642.8 7 17554234.68 83.36000 < 0.0001 Significant

A 1939055.773 1 1939055.773 9.207547 0.0045
B 28804362.11 1 28804362.11 136.7766 < 0.0001
C 81523898.04 1 81523898.04 387.1137 < 0.0001
AB 3069184.487 1 3069184.487 14.57393 0.0005
AC 1649094.211 1 1649094.211 7.830673 0.0082
BC 3362604.781 1 3362604.781 15.96723 0.0003
ABC 2531443.352 1 2531443.352 12.02048 0.0014
Curvature 28386.49355 1 28386.49355 0.134792 0.7157 Not significant
Pure error 7581390.362 36 210594.1767
Cor total 130489419.6 44

Std. dev. 458.905 R-squared 0.9419

Mean 11150.4 Adj R-squared 0.9305
C.V. 4.11558 Pred R-squared 0.9092
PRESS 11845922.4 Adeq precision 22.903

less than 0.2). ‘Adeq. precision’ ratio of 22.903 indicates an adequate
signal. According to the ANOVA results it can be concluded that this
model can be used to navigate the design space.
Hence, the curvature in this response is insignificant, it can be
concluded that the relationship between the factors follow a linear
relationship and the obtained linear regression model are able to pre-
dict accurately. The empirical model was determined based on the
significant factors and is presented in terms of coded and actual fac-
tors as follows:
Ŷ = 11159.32 + 220.17 × A + 848.59 × B − 1427.62 × C
+277.0 × A × B + 203.05 × A × C
−289.94 × B × C + 251.57 × A × B × C
Sales Rate = 10092.21 + 133.17 × Pump + 2333.005
× Cashier + 352.61 × IAT − 113.06 × Pump
× Cashier − 150.04 × Pump × IAT − 1296.2
×Cashier × IAT + 125.78 × Pump × Cashier × IAT
Fig. 4(a) depicts the normal probability plot of residual for re-
sponse surface queue length. It is visible that the residual are gen-
erally fall on a straight line which is desirable and implies that the
error are distributed normally. Hence, the residual versus predicted
plot is shown in Fig. 4(b) which has no obvious pattern and the data
are scattered randomly. As well, the diagnostics plots of normal prob-
ability and residual versus predicted value for response of sales rate
are shown in Fig. 5(a) and (b). The data in both responses are desir-
able and it can be concluded that the proposed models are adequate
and there is no need for any further violation assumption.
Fig. 6 (a) and (b) illustrates the 3D surface and counter plots for
queue length response surface at center level for number of pump (8
(4) pumps). The curvilinear is visible in both plots which is in accordance
with the obtained quadratic model. The counter plot revealed that at
any particular levels of number of cashier the best surface finish is
obtainable when the number of pump is at somewhere at the center
of number of pump range. This is compatible with the fact that (num-
ber of pump)2 term is significant. As well Fig. 7 (a) and (b) depicts the
3D surface and counter graphs for response surface of sales rate. It
(5) can be obtained from these graphs that at any different levels of fac-
 M.R. Galankashi et al. / Advances in Engineering Software 92 (2016) 15–26 23

Fig. 4. (a) Normal probability plot, (b) residual versus predicted values plot for response queue length.

Fig. 5. (a) Normal probability plot, (b) residual versus predicted values plot for response sales rate.

Fig. 6. (a) 3D surface for queue length, (b) counter plot of response queue length.

tors the curvature is insignificant and the obtained regression model
is adequate to justify the very slight curvilinear.
Fig. 8 (a) and (b) illustrates the cube plots of responses queue
length and sales rate respectively based on the obtained models. It
is visible that the optimum value of 1.68 ∼ 2cars for queue length is
obtainable when the factors; number of pump and IATs are at low
level while number of cashier is at high level. Hence, the optimum
value of 13759.2 l for sales rate is obtainable when the factors; num-
ber of pump, number of cashier is at high level and IATs is at low
level.
24 M.R. Galankashi et al. / Advances in Engineering Software 92 (2016) 15–26

Fig. 7. (a) 3D surface for sales rate, (b) counter plot of response sales rate.

Fig. 8. (a) Cube plot of predicted response for queue length, (b) cube plot of sales rate.

desirable queue length and sales rate defined within the range of 2–6
and 10,000–13,000 l, respectively.

8.1. Confirmation test

In order to verify the adequacy of the developed models, 5 more

Fig. 9. Overlay plot.
The overlay plot can be developed by imposing the counters for
the various responses. The upper and lower limits have been defined
for both queue length and sales rate response surface and the shaded
area of the overlay plot, as illustrates in Fig. 9. Commonly the limits
for the responses will be selected based on the other constraints. In
this case according to the petrol station executive management, the
confirmation run were designed and the model elements was sim-
ulated as Table 13. The 5 runs conditions have been selected from
those runs that were not previously used but are within the range of
levels defined before. The point prediction tab from the design ex-
pert© software was used in order to predict the responses for queue
length and sales rate with the 95% prediction intervals. The predicted
values and the actual values through simulated models were com-
pared and the percentage error range for queue length and sales rate
are as follows: QL ∼ 0–16.66% and SR ∼ 2.63–4.95% where the values
particularly for sales rate are less than 5% and acceptable.
It can be proposed that the empirical models developed were rea-
sonably accurate. All the actual values for the confirmation run are
within the 95% prediction interval. The 95% prediction interval is the
range in which we can expect any individual value to fall into 95% of
the time [56].
9. Conclusion
This research contributes to incorporate simulation modeling
with DOE to investigate and optimize the performance of a petrol
station. Its novelty lies on integration of computer simulation and
 M.R. Galankashi et al. / Advances in Engineering Software 92 (2016) 15–26
Table 13
Confirmation tests.
No Factors Response, queue length
No. pump No. cashier IATs Predicted value Actual value
1 6 2 −1 6 7
2 10 2 +1 3 3
3 8 1 −1 7 6
4 8 3 +1 1 1
5 10 3 0 2 2
suggesting diverse scenarios as the DOE input. The proposed re-
search framework offers a valuable contribution since it is infeasible
to stop/change the layout of an operating system due to cost, time,
labor and other constraints. This simulation-based design of experi-
ments presented the finding of an empirical survey into the influence
of factors; (A)-number of pump, (B)-number of cashier and (C)-inter
arrival times on the queue length and sales rate of a simulated petrol
station. The procedure starts with simulation modeling of a petrol
station and finally the obtained data were used as the input of DOE.
The ANOVA results express that the IATs is the most effective factor
on both queue length and sales rate. The significances of (number of
pump)2 provides a nonlinear contribution to the investigated param-
eters on queue length. The obtained regression models were reason-
ably accurate and can be used for prediction. The minimum amount
of queue length is obtainable when the number of pumps and IATs
distribution were 6 and the low levels for each simulated elements
respectively, and number of cashiers was 3. As well, the maximum
amount of sales rate is achievable when the number of pumps and
number of cashiers was 10 and 3, respectively, and the IATs distribu-
tion was in low level for each elements based on Table 6. Although
this research aimed to investigate the behavior of the real system,
small parts of data were removed from analysis due to unpredicted
issues (e.g. human errors while collecting the data). This can be in-
terpreted as the limitation of study. As a direction for future research,
the proposed methodology of this study can be applied to other ser-
vice industries in order to optimize their services level.
References
[1] Hurst K. Improving service efficiency and effectiveness: the resource implica-
tions. Int J Healthc Qual Assur 2014;27(1).
[2] Saeidi SP, Sofian S, Saeidi P, Saeidi SP, Saaeidi SA. How does corporate so-
cial responsibility contribute to firm financial performance? The mediating
role of competitive advantage, reputation, and customer satisfaction. J Bus Res
2015;68(2):341–50.
[3] Rogers HP, Strutton D, Doddridge BF. Measuring customer satisfaction with logis-
tics services: an investigation of the motor carrier industry. In: Proceedings of the
1997 Academy of Marketing Science (AMS) Annual Conference. Springer; 2015.
[4] Molina-Azorín JF, Tarí JJ, Pereira-Moliner J, López-Gamero MD, Pertusa-
Ortega EM. The effects of quality and environmental management on compet-
itive advantage: A mixed methods study in the hotel industry. Tour Manag
2015;50:41–54.
[5] Mende M, Thompson SA, Coenen C. It’s all relative: how customer-perceived com-
petitive advantage influences referral intentions. Mark Lett 2014:1–18.
[6] Conway N, Briner RB. Unit-level linkages between employee commitment to the
organization, customer service delivery and customer satisfaction. Int J Human
Res Manag 2014;26(16):2039–61 1–23.
[7] Pinto L, Silva P, Young T. A generic method to develop simulation models for am-
bulance systems. Simul Model Pract Theory 2015;51:170–83.
[8] Tako AA, Kotiadis K, PartiSim. A multi-methodology framework to support facili-
tated simulation modelling in healthcare. Eur J Oper Res 2015;244(2):555–64.
[9] Moazzami A, Galankashi MR, Khademi A. Simulation, modeling and analysis of a
petrol station. Int Rev Model Simul (IREMOS) 2013;6(1):246–53.
[10] Antony J. Design of experiments for engineers and scientists. Elsevier; 2014.
[11] Amiri M, Mohtashami A. Buffer allocation in unreliable production lines based on
design of experiments, simulation, and genetic algorithm. Int J Adv Manuf Tech-
nol 2012;62(1):371–83.
[12] Ajdari A, Mahlooji H. An adaptive exploration-exploitation algorithm for con-
structing metamodels in random simulation using a novel sequential experimen-
tal design. Commun Stat-Simul Comput 2014;43(5):947–68.
[13] Li M, Yang F, Wan H, Fowler JW. Simulation-based experimental design and sta-
tistical modeling for lead time quotation. J Manuf Syst 2014. doi:10.1016/j.jmsy.
2014.07.012.
25
Response, sales rate
Error (%) Predicted value Actual value Error (%)
14.28 12569.8 12246.8 2.63
0 10154.9 9863.90 2.95
16.66 11448.4 11825.3 3.18
0 10290.4 10776.9 4.51
0 12505.1 13156.9 4.95
[14] Ekren BY, Heragu SS, Krishnamurthy A, Malmborg CJ. Simulation based experi-
mental design to identify factors affecting performance of AVS/RS. Comput Indus
Eng 2010;58(1):175–85.
[15] Kouritzin MA, Newton F, Wu B. A graph theoretic approach to simulation and
classification. Comput Stat Data Anal 2014;70(0):281–94.
[16] López-Fidalgo J, Rivas-López MJ. Optimal experimental designs for partial likeli-
hood information. Comput Stat Data Anal 2014;71(0):859–67.
[17] Denman NG, McGree J M, Eccleston J A, Duffull S B. Design of experiments for
bivariate binary responses modelled by Copula functions. Comput Stat Data Anal
2011;55(4):1509–20.
[18] Jahangirian M, Naseer A, Stergioulas L, Young T, Eldabi T, Brailsford S, et al.
Simulation in health-care: lessons from other sectors. Oper Res 2012;12(1):
45–55.
[19] Tan W, Xu W, Yang F, Xu L, Jiang C. A framework for service enterprise
workflow simulation with multi-agents cooperation. Enterp Inf Syst 2013;7(4):
523–542.
[20] Chetouane F, Barker K, Oropeza ASV. Sensitivity analysis for simulation-based de-
cision making: application to a hospital emergency service design. Simul Model
Pract Theory 2012;20(1):99–111.
[21] Wu DD, Olson DL. A system dynamics modelling of contagion effects in accounts
risk management. Syst Res Behav Sci 2014;31(4):502–11.
[22] Abbott D, Marinov MV. An event based simulation model to evaluate the design
of a rail interchange yard, which provides service to high speed and conventional
railways. Simul Model Pract Theory 2015;52:15–39.
[23] Robinson S, Radnor Z J, Burgess N, Worthington C. SimLean: utilising simulation
in the implementation of lean in healthcare. Eur J Oper Res 2012;219(1):188–97.
[24] Tako AA, Robinson S. The application of discrete event simulation and sys-
tem dynamics in the logistics and supply chain context. Decis Support Syst
2012;52(4):802–15.
[25] Quesnel G, Duboz R, É Ramat. The virtual laboratory environment–an operational
framework for multi-modelling, simulation and analysis of complex dynamical
systems. Simul Model Pract Theory 2009;17(4):641–53.
[26] Gul M, Guneri AF. A comprehensive review of emergency department sim-
ulation applications for normal and disaster conditions. Comput Indus Eng
2015;83(0):327–44.
[27] Mayer RJ, Benjamin PC. Using the Taguchi paradigm for manufacturing system
design using simulation experiments. Comput Ind Eng 1992;22(2):195–209.
[28] Amoako-Gyampah K, Meredith JR. The operations management research agenda:
an update. J Oper Manag 1989;8(3):250–62.
[29] Pannirselvam GP, Ferguson LA, Ash RC, Siferd SP. Operations management re-
search: an update for the 1990s. J Oper Manag 1999;18(1):95–112.
[30] Du Y, et al. Microscopic simulation evaluation method on access traffic operation.
Simul Model Pract Theory 2015;53(0):139–48.
[31] Tsai C-S. Evaluation and optimisation of integrated manufacturing system opera-
tions using Taguch’s experiment design in computer simulation. Comput Ind Eng
2002;43(3):591–604.
[32] Carvalho H, et al. Supply chain redesign for resilience using simulation. Comput
Ind Eng 2012;62(1):329–41.
[33] Jahangirian M, et al. Simulation in manufacturing and business: A review. Eur J
Oper Res 2010;203(1):1–13.
[34] Aboueljinane L, Sahin E, Jemai Z. A review on simulation models applied to emer-
gency medical service operations. Comput Ind Eng 2013;66(4):734–50.
[35] Alrabghi A, Tiwari A. State of the art in simulation-based optimisation for main-
tenance systems. Comput Ind Eng 2015;82(0):167–82.
[36] Thorp M, Moscardini A, Cross M. Interactive discrete event simulation without
programming. Adv Eng Softw (1978) 1982;4(1):26–31.
[37] Abbasian-Hosseini SA, Nikakhtar A, Ghoddousi P. Verification of lean construc-
tion benefits through simulation modeling: a case study of bricklaying process.
KSCE J Civil Eng 2014;18(5):1248–60.
[38] Nikakhtar A, Hsiang SM. Incorporating the dynamics of epidemics in simulation
models of healthcare systems. Simul Model Pract Theory 2014;43:67–78.
[39] Xiaobing P. An application of OR and IE technology in bank service system im-
provement. In: Proceedings of the IEEE international conference on industrial en-
gineering and engineering management, 2008. IEEM 2008; 2008.
[40] Cha J-H, Roh M-I. Combined discrete event and discrete time simulation frame-
work and its application to the block erection process in shipbuilding. Adv Eng
Softw 2010;41(4):656–65.
[41] Madadi N, Roudsari AH, Wong KY, Galankashi MR. Modeling and Simulation of
a Bank Queuing System. In: Proceedings of the fifth international conference
on computational intelligence, modelling and simulation (CIMSim), 2013. IEEE;
2013. p. 209–15.
26 M.R. Galankashi et al. / Advances in Engineering Software 92 (2016) 15–26
[42] Tang Z, Zhang W, Wu P. A holistic framework for engineering simulation platform
development gluing open-source and home-made software resources. Adv Eng
Softw 2014;76:99–109.
[43] iCasas PF, iCasas AF, Garrido-Soriano N, Casanovas J. Formal simulation model to
optimize building sustainability. Adv Eng Softw 2014;69:62–74.
[44] Uğur A, Aydin D. An interactive simulation and analysis software for solving TSP
using ant colony optimization algorithms. Adv Eng Softw 2009;40(5):341–9.
[45] Cornillier F, Boctor FF, Laporte G, Renaud J. A heuristic for the multi-period petrol
station replenishment problem. Eur J Oper Res 2008;191(2):295–305.
[46] Tack L, Vandebroek M. An adjustment algorithm for optimal run orders in design
of experiments. Comput Stat Data Anal 2002;40(3):559–77.
[47] Steibel JP, Rosa GJM, Tempelman RJ. Optimizing design of two-stage experiments
for transcriptional profiling. Comput Stat Data Anal 2009;53(5):1639–49.
[48] Chuang SC, Hung YC. Uniform design over general input domains with applica-
tions to target region estimation in computer experiments. Comput Stat Data Anal
2010;54(1):219–32.
[49] Cheng RC, Kleijnen JPC, Melas VB. Optimal designing of simulation experiments
with nearly saturated queues. J stat plan infer 2000;85(1):19–26.
[50] Can B, Heavey C. Comparison of experimental designs for simulation-based sym-
bolic regression of manufacturing systems. Comput Ind Eng 2011;61(3):447–62.
[51] Kamrani M, Abadi SMHE, Golroudbary SR. Traffic simulation of two adjacent
unsignalized T-junctions during rush hours using Arena software. Simul Model
Pract Theory 2014;49:167–79.
[52] Nikakhtar A, Wong KY, Zarei MH, Memari A. Comparison of two simulation soft-
ware for modeling a construction process. In: Proceedings of the third inter-
national conference on computational intelligence, modelling and simulation
(CIMSiM), 2011; 2011. p. 200–5.
[53] Ahmed KI. Modeling drivers’ acceleration and lane changing behavior. Mas-
sachusetts Institute of Technology; 1999.
[54] Ahmadipourroudposht M, Fallahiarezoudar E, Yusof NM, Idris A. Application
of response surface methodology in optimization of electrospinning process to
fabricate (ferrofluid/polyvinyl alcohol) magnetic nanofibers. Mater Sci Eng: C
2015;50:234–41.
[55] Fallahiarezoudar E, Ahmadipourroudposht M, Yusof NM, Idris A. Influence of pro-
cess factors on diameter of core (γ -Fe2O3)/shell (polyvinyl alcohol) structure
magnetic nanofibers during co-axial electrospinning. Int J Polym Mater Polym
Biomater 2015;64(1):15–24.
[56] Noordin MY, Venkatesh VC, Sharif S, Elting S, Abdullah A. Application of response
surface methodology in describing the performance of coated carbide tools when
turning AISI 1045 steel. J Mater Process Technol 2004;145(1):46–58.