Applied Thermal Engineering: Mir Majid Etghani, Mohammad Hassan Shojaeefard, Abolfazl Khalkhali, Mostafa Akbari

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

Applied Thermal Engineering 59 (2013) 309e315

Contents lists available at SciVerse ScienceDirect

Applied Thermal Engineering


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

A hybrid method of modied NSGA-II and TOPSIS to optimize


performance and emissions of a diesel engine using biodiesel
Mir Majid Etghani a, *, Mohammad Hassan Shojaeefard b, Abolfazl Khalkhali b,
Mostafa Akbari b
a
Mechanical Engineering Department, Islamic Azad University, Qaemshahr Branch, Mazandaran, Iran
b
Automotive Engineering Department, Iran University of Science and Technology, Tehran, Iran

h i g h l i g h t s

 Effects of castor oil biodiesel blends have been examined on the diesel engine performance and emissions.
 Modeling engine performance and emissions by articial neural network with back-propagation algorithm accurately.
 2 and 6-objective optimization has been applied by the modied NSGA-II.
 Trade-off optimum design points are determined by applying TOPSIS.

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

Article history: This paper addresses articial neural network (ANN) modeling followed by multi-objective optimization
Received 26 September 2012 process to determine optimum biodiesel blends and speed ranges of a diesel engine fueled with castor oil
Accepted 24 May 2013 biodiesel (COB) blends. First, an ANN model was developed based on standard back-propagation algo-
Available online 3 June 2013
rithm to model and predict brake power, brake specic fuel consumption (BSFC) and the emissions of
engine. In this way, multi-layer perception (MLP) network was used for non-linear mapping between the
Keywords:
input and output parameters. Second, modied NSGA-II by incorporating diversity preserving mecha-
Performance
nism called the -elimination algorithm was used for multi-objective optimization process. Six objectives,
Emissions
Castor oil biodiesel
maximization of brake power and minimization of BSFC, PM, NOx, CO and CO2 were simultaneously
ANN considered in this step. Optimization procedure resulted in creating of non-dominated optimal points
NSGA-II which gave an insight on the best operating conditions of the engine. Third, an approach based on TOPSIS
TOPSIS method was used for nding the best compromise solution from the obtained set of Pareto solutions.
 2013 Elsevier Ltd. All rights reserved.

1. Introduction therefore in the diesel engines, biodiesel is the best candidate as


alternative for petroleum-based diesel fuels. Biodiesel is briey
Today caused by considerable growth of demands, known pe- dened as the monoalkyl esters of vegetable oils or animal fats.
troleum reserves are extremely diminishing and due to the envi- Biodiesel is an ecological, renewable and less polluting fuel, and
ronmental concerns for pollution from exhausted gases and therefore it is environmentally useful [2]. The usage of biodiesel
demanding of greenhouse gas reduction, renewable and alternative does not require any changes in the fuel distribution infrastructure,
fuels in the automotive fuel markets grew fast during the 21st and it is competitive with conventional diesel fuel. Furthermore,
century [1]. biodiesel biodegrades much more rapidly than diesel fuel. So,
Since biodiesel has all the characteristics that an alternative fuel considerable environmental benets are provided [3].
to petrodiesel should have such as technically feasible, economi- Because of experimentally determining the engine performance
cally competitive, environmentally acceptable, and easily available, map for different operating conditions and biodiesel blends are
time and money consuming, ANNs are used in recent studies. ANN
as a computational modeling tool is employed widely to alleviate
the burden of experimental testing. Recently, the application of
* Corresponding author. Tel.: 98 9122393445; fax: 98 2177240362.
E-mail addresses: etghani@iust.ac.ir, m.etghani@yahoo.com (M.M. Etghani),
ANN method to predict performance and emissions of internal
mhshf@iust.ac.ir (M.H. Shojaeefard), ab_khalkhali@iust.ac.ir (A. Khalkhali), combustion engines has gained signicant success. Ghobadian et al.
mr.mostafaakbari@yahoo.com (M. Akbari). [4] modeled a diesel engine using waste cooking biodiesel fuel by

1359-4311/$ e see front matter  2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.applthermaleng.2013.05.041
310 M.M. Etghani et al. / Applied Thermal Engineering 59 (2013) 309e315

ANN. They considered engine speed, percentage of bio-fuel blend as


the input variables and torque, BSFC, Hydrocarbon (HC) and Car-
bone Monoxide (CO) as the outputs. Kiani Deh Kiani et al. [5] used
ANN to predict the performance and emissions from a spark igni-
tion engine using ethanol-gasoline blends where they considered
engine speed, engine speed and blend percentage as the input
parameters and engine torque, power and exhaust emissions (CO,
CO2, NOx and HC) of the engine as the output parameters. Kara
Togun et al. [6] developed ANN to predict torque and BSFC of a
gasoline engine in terms of spark advance, throttle position and
engine speed. Shivakumar et al. [7] have used ANN for prediction of
performance and emission characteristics of a CI engine using WCO.
In their work, ANN modeling was used to predict BTE, BSEC, Texh,
NOx, HC and Smoke. This technology was also used for modeling of
valve timing in a SI engine [8].
Design of modern engines is driven by many competing criteria. Fig. 1. Schematic diagram of experimental setup.
Reduction of brake specic fuel consumption is common in the
automotive industry to reduce the fuel consumption. Moreover,
because of the environmental concerns, reduction of emissions is set up a multi-objective optimization framework for maximizing
also signicant. On the other hand, increasing the engine power is engine power, minimizing BSFC and emissions. At the last step,
the important goal of engine designers. So, to consider all such TOPSIS is used to nd the best compromise solution.
criteria simultaneously, a complex multi-objective optimization
problem (MOP) must be solved. Many different methods were 2. Experimental methodology
proposed by previous researchers for solving MOPs [9,10]. Non-
dominated Sorting Genetic Algorithm (NSGA-II) proposed by Sri- The experiments were performed on the agricultural direct in-
nivas and Deb [11], which is Pareto based approaches is one of the jection compression engine with specications given in Table 1.
efcient algorithms for solving MOPs. It generates a set of non- An eddy current dynamometer was used for gathering outputs
dominated solutions (Pareto solutions), where a non-dominated of engine. The schematically prepared diagram of experimental
solution performs better on at least one criterion than the other setup is given in Fig. 1. The relative emission parameters from an
solutions. To improve NSGA-II, Nariman-Zadeh proposed modied online and accurately calibrated exhaust gas analyzer AVL DiCOM
NSGA-II which use -elimination algorithm rather than crowding were recorded. The accuracies of the measured parameters and the
factor [12]. This method is employed successfully in many recent uncertainties in the calculated parameters are given in Table 2.
studies [13,14]. After nding out the non-dominated points, it is The experiments were carried out by using pure diesel, B5 (5%
desired to nd some trade-off optimum points compromising biodiesel 95% diesel), B10, B15, B20, B25 and B30 at full loads and
objective functions. For this purpose, technique for ordering pref- various engine speeds (1200, 1400, 1600, 1700, 1800, 2000 rpm).
erences by similarity to ideal solution (TOPSIS) can be used. TOPSIS All tests were completed without any modications on the en-
is based on simultaneous minimization of distance from an ideal gine. The tests were carried out under steady-state condition.
point and maximization of distance from a nadir point. Several Before each test, the engine was warmed up with diesel fuel for
numerical experiments show that NSGA-II determines the Pareto about 15 min until the cooling water temperature was stabilized.
set and TOPSIS nds the best compromise solutions for different Then brake power, BSFC, Nitrogen oxides (NOx), Carbone dioxide
scenarios [15,16]. (CO2), CO and PM were measured. Each test was repeated three
In the present study, optimal operating condition of a diesel times and the results of the three repetitions were averaged. Some
engine fueled by biodiesel is extracted by an approach based on of the experimental results are reported in Table 3.
three steps. At the rst step, a multi-layer perception (MLP)
network is learned by experimental data to model and predict 3. Modeling and prediction using articial neural network
engine power, BSFC and emissions. Such ANN models are then (ANN)
employed in the multi-objective optimization process at the second
step. In this way, modied NSGA-II by incorporating diversity The main focus of this section is modeling and prediction of
preserving mechanism called the -elimination algorithm is used to engine output using ANN based on experimental results. The back-
propagation learning algorithm was used in feed forward with one
hidden layer. Blend percentage and engine speed were considered
Table 1
Engine technical specications.
Table 2
Bore  stroke 100 mm  127 mm The accuracies uncertainties in the results.
Number of cylinders 4
Parameters Accuracies
Volume capacity 3.99 L
Cycle 4 Stroke Load 2 N.m
Aspiration Wastegated turbocharger Speed 5 rpm
Combustion system Fast ram direct injection Time 0.5%
Compression ration 17.25:1 CO 0.01 % Vol
Max. power 61 kW in 2000 rpm CO2 0.01 % Vol
Fuel pump Bosch rotary with Boost control NOx 1 ppm
Governing Mechanical PM 1 mg/m3
Cooling Water, belt driven water pump Calculated results Uncertainty
Weight 265 Kg Power 0.2%
Length  width  height 678.7 mm  655 mm  748.5 mm BSFC 0.22%
M.M. Etghani et al. / Applied Thermal Engineering 59 (2013) 309e315 311

Table 3 representing inputs, neurons in hidden layers, and outputs,


Samples of experimental results with input and output parameters. respectively. Table 4 shows its criteria. The learning algorithm used
No. Inputs Outputs in the study is LevenbergeMarquardt (LM), activation function is
Speed Blend Power BSFC PM NO CO CO2
logistic sigmoid (logsig) transfer functions and epoch numbers is
(rpm) (%) (hp) (g/kW-h) (mg/m3) (PPM) (%) (%) 20,000.
1 1200 0 56 241.4 521.9 1380 0.43 12.4
2 1400 0 65 237.6 160.8 1254 0.37 11.4 4. Multi-objective optimization
6 2000 0 82 255.7 41.13 926 0.31 8.9
7 1200 5 55 245.8 488.6 1396 0.42 12.2
10 1700 5 72 243.9 58.79 1193 0.322 9.7 Multi-objective optimization, which is also called multi criteria
11 1800 5 76.1 250.9 47.78 1081 0.312 9.1 optimization or vector optimization, dened as nding a vector of
12 2000 5 79.8 263.9 36.78 970 0.302 8.5 decision variables satisfying constraints to give acceptable values
25 1200 20 52 255.5 361 1428 0.32 11.5
to all objective functions [19,20]. In these problems, there are
26 1400 20 60 250.1 111 1364 0.29 9.1
27 1600 20 64.1 252.37 73.5 1311 0.28 8.5
several objectives or cost functions (a vector of objectives) to be
28 1700 20 66.9 260.3 34.61 1258 0.27 8.15 optimized (minimized or maximized) simultaneously. These ob-
29 1800 20 70.4 271.2 31.04 1140 0.262 7.9 jectives often conict with each other so that improving one of
30 2000 20 74.3 288.2 27.45 1022 0.255 7.64 them will deteriorate another. Thus, there is no single optimal
37 1200 30 50.1 258.7 327 1445 0.28 10.8
solution as the best with respect to all the objective functions.
38 1400 30 57.7 252.63 102 1409 0.27 8
39 1600 30 62.2 254.5 65.47 1345 0.265 7.49 Instead, there is a set of optimal solutions, known as Pareto
40 1700 30 64 264.22 27.15 1281 0.26 7.38 optimal solutions or Pareto front [21e25] for multi-objective
41 1800 30 67.65 276.8 25.63 1165 0.25 7.3 optimization problems. The concept of Pareto front or set of
42 2000 30 71.3 300.4 24.12 1050 0.24 7.21 optimal solutions in the space of objective functions in MOPs
stands for a set of solutions that are non-dominated to each other
but are superior to the rest of solutions in the search space. This
as input layer components, while the Power, BSFC, PM, NOx, CO and means that it is not possible to nd a single solution to be superior
CO2 were considered as output layer components of the ANNs. In to all other solutions with respect to all objectives so that changing
the ANN model, 34 values of which 42 experiments were used for the vector of design variables in such a Pareto front consisting of
training the network and 8 values were selected randomly to test these non-dominated solutions could not lead to the improvement
the performance of the trained network. of all objectives simultaneously. Therefore, such a change will lead
The input layer neurons receive information from the outside to deteriorating of at least one objective. Thus, each solution of the
environment and transmit them to the neurons of the hidden layer Pareto set includes at least one objective inferior to that of another
without performing any calculation. The hidden layer neurons then solution in that Pareto set, although both are superior to others in
process the incoming information and extract useful features to the rest of search space. Such problems can be mathematically
reconstruct the mapping from the input space. The neighboring dened as:
layers are fully interconnected by weights. Finally, the output layer Find the vector X * x*1 ; x*2 ; .; x*n to optimize
neurons produce the network prediction to the outside world [17].
One of the most important tasks in ANN studies is to determine FX f1 X; f2 X; .; fk XT (4)
the optimal network architecture which is related to hidden layers
and neurons in it. Generally, the trial-and-error approach is used. In subject to m inequality constraints
this study, the best architecture of the network was obtained by
trying different hidden layers and neurons. The trial started on gi X  0; i 1 to m (5)
hidden layer with ten neurons, and the performance of each
network was checked by three standard criteria, R, RMSE and MAPE and p equality constraints
(R is the correlation coefcient, RMSE is the root mean square error,
hj X 0; j 1 to p (6)
and MRE is the mean relative error). These criteria are dened as
following equations respectively:
where X * <n is the vector of decision or design variables, and
v FX<n is the vector of objective functions, which must each be
u P 2 !
u
j tj  oj
u either minimized or maximized. However, without loss of gener-
R t1  P  2 (1) ality, it is assumed that all objective functions are to be minimized.
j oj Such multi-objective minimization based on Pareto approach can
be conducted using some denitions.
 X  1=2 A vector U u1 ; u2 ; .; uk <k is dominant to vector
1=p tj  oj 2 (2) V v1 ; v2 ; .; vk <k (denoted by U3V) if and only if

 !
1 X tj  oj 
p
Table 4
 100 (3) Three standard criteria obtained from the ANN.
p 1 tj
R RMSE MAPE
where t is the target value, o is the output value and p is the pattern Train Test Train Test Train Test
number [18].
Power 0.999995 0.999956 0.000772 0.002223 0.216595 0.532047
The goal is to maximize correlation coefcient to obtain a BSFC 0.999998 0.999957 0.001158 0.0047 0.12151 0.51145
network with the best generalization. Many different network PM 0.999995 0.996709 0.000719 0.008131 0.733854 4.15107
models were tried and their R, RMSE and MAPE values were NOx 0.999996 0.999955 0.001303 0.004254 0.169048 0.538951
calculated. Based on this analysis, the optimal architecture of the CO 0.999978 0.999937 0.001472 0.002437 0.396388 0.664076
CO2 0.999983 0.999931 0.001968 0.003807 0.350319 0.685741
ANN was constructed as 2e15e6 neural network architecture
312 M.M. Etghani et al. / Applied Thermal Engineering 59 (2013) 309e315

ci f1; 2; .; kg, ui  vi ^ djf1; 2; .; kg : ui < vi. In other words,


there is at least one uj which is smaller than vj whilst the remaining
us are either smaller or equal to corresponding vs.
A point X * U (U is a feasible region in <k satisfying equations (5)
and (6)) is said to be Pareto optimal (minimal) with respect to all
XU if and only if FX * 3FX. Alternatively, it can be readily
restated as: ci f1; 2; .; kg, cXU  fX * g fi X *  fi X ^
djf1; 2; .; kg : fj X * < fj X In other words, the solution X* is said
to be Pareto optimal (minimal) if no other solution can be found to
dominate X* using the denition of Pareto dominance.
For a given MOP, a Pareto set P* is a set in the decision variable
space consisting of all the Pareto optimal vectors
P * fXUjeX 0 U : FX 0 3FX. In other words, there is no other
X 0 as a vector of decision variables in Uthat dominates any XP *. For
a given MOP, the Pareto front PF is a set of vector of objective
functions which are got using the vectors of decision variables in the
Pareto set P * , that is PF* fFX f1 X; f2 X; .; fk X : XP * g. Fig. 3. NOx vs. power in six-objective optimization.
In other words, the Pareto front PF is a set of the vectors of objective
functions mapped from P*.
Evolutionary algorithms are widely used for multi-objective objective weights. The TOPSIS process for determining the best
optimization because of their natural properties suited for these compromise solution is briey presented as follows:
types of problems. This is mostly because of their parallel or
population-based search approach. So, most of the difculties and 1) Input matrix S, where the element sij is the jth objective value
deciencies in the classical methods in solving multi-objective of the ith alternative (that is, S is composed of the Pareto
optimization problems are eliminated. For example, there is no solutions)
need for either several runs to nd the Pareto front or quantica- 2) Calculate normalized rating ~sij according to the following
tion of the importance of each objective using numerical weights. equation:
In this paper, modied NSGA-II which is used in previous re-
searches is employed for multi-objective optimization process. In sij
~sij q
Pm 2 i 1; .; m; j 1; .; n (7)
the modied NSGA-II code, the -elimination approach was
i1 sij
replaced the crowding distance assignment approach in NSGA-II,
all the clones and/or -similar individuals based on Euclidean norm
of two vectors are recognized and simply eliminated from the 3) Construct the weighted normalized values bs ij using the
current population. So, based on a pre-dened value of as the following equation:
elimination threshold ( 0.001 was used in present paper) all the
individuals in a front within this limit of a particular individual are b
s ij wj  ~sij i 1; .; m; j 1; .; n (8)
eliminated [12e14]. where wj is the weight value of jth attribute and must satisfy:

X
n
5. Technique for ordering preferences by similarity to ideal wj 1 (9)
solution (TOPSIS) j1

TOPSIS is used to rank the given alternatives of the Pareto so- 4) Determine S and S as follow:
lutions got by MOPs introduced by Hwang and Yoon [26]. The basic
concept of TOPSIS determines the positive ideal solution (S) as    
S maxb
s ij jjJ1 ; minb
s ij jjJ2 ; i 1; 2; .; m (10)
well as the negative ideal solution (S-), and then nds the best
compromise solution, which is the closest to S and the farthest    
from S-, from the Pareto set according to the decision makers S minb
s ij jjJ1 ; maxb
s ij jjJ2 ; i 1; 2; .; m (11)

Fig. 2. BSFC vs. power in six-objective optimization. Fig. 4. PM vs. power in six-objective optimization.
M.M. Etghani et al. / Applied Thermal Engineering 59 (2013) 309e315 313

Fig. 5. CO vs. power in six-objective optimization.

Fig. 7. PM vs. BSFC in six-objective optimization.


where J1 is a set of benet attributes. That is, the large value
means better performance, such as the amount of production for a
factory. 6. Six-Objective optimization and using TOPSIS to determine
J2 is a set of cost attributes. The smaller value means better the best trade-off solutions
performance, such as the number of employees for a factory.
To investigate the optimal output in different conditions of en-
5) Develop a distance measure over each criterion to both ideal gine variables (engine speed and blend percentage), six-objective
(D) and nadir (D) optimization problems were solved. Power, BSFC, PM, NOx, CO
and CO2 are chosen as objective functions for this multi-objective
v
uX  2 optimization process. This will allow nding trade-off optimum
u
D
i t bs ij  S j
(12) design points from the view point of all six-objective functions
j simultaneously. The feed forward neural network models got in
section 3 are now deployed in these six-objective optimization
v problems. The optimization problem can be formulated in the
uX  2
u following form:
D
i t bs ij  S j
(13)
j
8
>
> Maximize f 1 Brake Power
>
>
>
> Minimize f 2 BSFC
6) Calculate relative closeness Di for each Pareto solution ac- >
>
>
> Minimize f 3 PM
<
cording to following equation: Minimize f 4 NOx
(15)
>
> Minimize f 5 CO
>
>
D >
> Minimize f 6 CO2
Di i 1; .; m (14) >
>
D D >
> 1000  Speed  2000
:
0  Blend  50

7) Choose the best compromise solution whose relative closeness A population of 80 individuals with a crossover probability of 0.6
Di is the closest to 1. and mutation probability of 0.08 was used in 400 generations for
such 6-objective optimization problem.

Fig. 6. CO2 vs. power in six-objective optimization. Fig. 8. NOx vs. BSFC in six-objective optimization.
314 M.M. Etghani et al. / Applied Thermal Engineering 59 (2013) 309e315

functions. Thus, there are some points in each plane that may
dominate others in the same plane. However, these individuals are
all non-dominated when considering all six objectives simulta-
neously. If the design variables (Engine speed and Biodiesel percent
in blends) are selected, based on the Pareto sets, it leads to the best
possible combination of those six objectives. Moreover, it can
readily be observed that the results of such six-objective optimi-
zation include those of the two-objective optimization and,
therefore, provide more optimal choices for the designer. Such an
overlay graph in which the results of the independently got two-
objective optimization lie on the boundary of the six-objective
optimization indeed exhibits the correctness of the Pareto frontiers.
It is now desired to nd a trade-off optimum design points out of
all non-dominated six-objective optimization process compro-
mising all objective functions. This can be achieved by TOPSIS
Fig. 9. CO vs. BSFC in six-objective optimization.
method described in section 6. TOPSIS method is applied on the
individuals obtained in six and some of two-objective optimization
process separately. In this way, equal weights are used to consider
non-preference for all objective functions. Consequently, optimum
design point A is got by applying TOPSIS on the results of two-
objective optimization problems of power and BSFC. Similarly,
considering conjugation of Power or BSFC with the emissions as
objective functions and applying TOPSIS, optimum design points B
to I are got. Finally TOPSIS is employed to get trade-off optimum
point P from the six-objective optimization problem. Design points
A to P are depicted in Figs. 2e10 and the values of design variables
and objective functions for these points are depicted in Table 5.
Optimum points J to O represented in Table 5 and Figs. 3e10 are
the results of the single objective optimization which show the best
values of each objective. It means that if someone looks for improve
on only one of the six-objective functions, one of the points J to O.
As it is expected, biodiesel percent reduction in blend to 0 and
Fig. 10. CO2 vs. BSFC in six-objective optimization. engine speed increase to about 2000 rpm, maximum power was
achieved (See point K). Moreover, for gaining minimum NOx, en-
gine speed was set to the lower bound (point L).
A total 1274 non-dominated optimum design points were gotten Moreover, another trade-off design point Q can be simply
during optimization process. The non-dominated individuals of the recognized from Figs. 2e10. This point is closely near to the Pareto
six-objective optimization in the plane of power and BSFC together fronts in all planes and also has good power about 50 kW. This issue
with the results of a separately run two-objective optimization of shows that it is worthy from the view point of all objective
the same two objectives were depicted in Fig. 2. Such non- functions.
dominated individuals of six and two-objective optimization Consequently, there are some important optimal design facts
were shown on the other planes in Figs. 3e10. It should be noted among the objective functions which discovered by the multi-
that there is a single set of individuals as a result of the six-objective objective optimization of the ANN meta-models obtained using
optimization that are shown in different planes of objective the experimental analysis of the diesel engine. Such important

Table 5
The best compromise solution determined by TOPSIS.

Point Objectives Speed (rpm) Bio percent Power (kW) BSFC (g/kW-h) CO2 (% Vol) CO (% Vol) NOx (ppm) PM (mg/m3)

Two-objective optimization A Power-BSFC 1000.0 35.35 53.16 214.07 14.15 0.24 566 601.15
process Topsis B Power-PM 1356.6 2.15 41.24 235.28 11.36 0.40 1363 0.84
C Power-NOx 1000.2 37.83 50.45 205.10 14.49 0.25 383 600.01
D Power-CO 1901.6 45.45 40.36 316.49 1.92 0.08 1316 58.56
E Power-CO2 1917.4 46.60 38.13 325.03 1.48 0.07 1429 62.47
F BSFC-PM 1312.1 0.34 31.68 227.56 11.66 0.47 1500 8.00E-05
G BSFC-NOx 1000.2 38.52 49.35 207.55 14.41 0.24 365 599.89
H BSFC-CO 1846.3 46.37 38.37 316.88 1.79 0.075 1381 54.85
I BSFC-CO2 1858.5 48.35 35.11 329.13 1.22 0.06 1555 63.05
Single objective optimization J BSFC 1353.9 36.1 29.46 203.78 12.17 0.36 761 0.013
K Power 1996.1 0 61 255.0 8.92 0.31 927 47.5
L NOx 1000.3 39.41 48.36 211.37 14.21 0.24 355 600.39
M PM 1299.4 7.37 29.83 232.85 11.28 0.46 1539 0.00003
N CO 1897.5 49.67 32.68 336.64 0.86 0.05 1659 62.25
O CO2 1969.6 49.43 30.92 340.45 0.69 0.06 1662 67.80
Six-objective optimization P 6-objective 1394.8 37.96 47.23 216.11 11.37 0.30 580 27.79
process TOPSIS
- Q e 1831.3 35.894 67.02 249.0671 7.64429 0.21931 658 27.82
M.M. Etghani et al. / Applied Thermal Engineering 59 (2013) 309e315 315

optimal design points could not be found without the multi- [6] N. Kara Togun, S. Baysec, Prediction of torque and specic fuel consumption of
a gasoline engine by using articial neural networks, Appl. Energy 87 (2010)
objective optimization approach of such diesel engine.
349e355.
[7] Shivakumar, P. Srinivasa Pai, B.R. Shrinivasa Rao, Articial Neural Network
based prediction of performance and emission characteristics of a variable
7. Conclusions
compression ratio CI engine using WCO as a biodiesel at different injection
timings, Appl. Energy 88 (2011) 2344e2354.
Feed forward neural network approach was used successfully to [8] M. Glc, Y. Sekmen, P. Erduranl, M. Sahir Salman, Articial neural-network
derive models of the performance and emissions of the diesel en- based modeling of variable valve-timing in a spark-ignition engine, Appl.
Energy 81 (2005) 187e197.
gine fueled with castor oil biodiesel blends. The derived models [9] S. Orun Mert, Z. zelik, Y. zelik, I. Diner, Multi-objective optimization of
were then used in an evolutionary multi-objective Pareto based a vehicular PEM fuel cell system, Appl. Thermal Eng. 31 (2011) 2171e2176.
optimization process so that some interesting and informative [10] P. Ahmadi, I. Dincer, Thermodynamic and exergoenvironmental analyses, and
multi-objective optimization of a gas turbine power plant, Appl. Thermal Eng.
optimum design aspects revealed for the diesel engine with respect 31 (2011) 2529e2540.
to the control variables of biodiesel percent and engine speed. For [11] N. Srinivas, K. Deb, Multiobjective optimization using nondominated sorting
getting better results, common NSGA-II algorithm was modied. in genetic algorithms, Int. J. Evol. Comput. 2 (1994) 221e248.
[12] A. Jamali, N. Nariman-zadeh, A. Darvizeh, A. Masoumi, S. Hamrang, Multi-
This optimization led to the discovering of some important trade- objective evolutionary optimization of polynomial neural networks for
off points among those objective functions. Such multi-objective modelling and prediction of explosive cutting process, Int. J. Eng. Appl. Art
optimization of diesel engine could unveil very important design Intel 22 (2009) 676e687.
[13] A. Khalkhali, N. Nariman-zadeh, A. Darvizeh, A. Masoumi, B. Notghi, Reli-
trade-offs between conicting objective functions which would not
ability-based robust multi-objective crashworthiness optimisation of S-sha-
have been found otherwise. Furthermore, it was shown that the ped box beams with parametric uncertainties, Int. J. Crashworth 15 (4) (2010)
results of six-objective optimization include those of two-objective 443e456.
[14] N. Nariman-Zadeh, M. Salehpour, A. Jamali, E. Haghgoo, Pareto optimization of
optimization in terms of Pareto frontiers and provide, accordingly,
a ve-degree of freedom vehicle vibration model using a multi-objective
more choices for optimal design. Finally, some trade-off optimum uniform-diversity genetic algorithm (MUGA), Eng. Appl. Art Intel. 23 (2010)
design points were determined and presented by applying TOPSIS 543e551.
method on the non-dominated solutions obtained through six- and [15] W.S. Lee, L.C. Lin, Evaluating and ranking the energy performance of ofce
building using technique for order preference by similarity to ideal solution,
two-objective problems. Designers can use these points in the Appl. Thermal Eng. 31 (2011) 3521e3525.
engine map for getting better performance. [16] M. Boix, L. Pibouleau, L. Montastruc, C. Azzaro-Pantel, S. Domenech, Mini-
mizing water and energy consumptions in water and heat exchange net-
works, Appl. Thermal Eng. 36 (2012) 442e455.
Acknowledgements [17] S. Haykin, Neural Networks: A Comprehensive Foundation, Mac-millan, New
York, 1994.
[18] Y. ay, A. iek, F. Kara, S. Sagiroglu, Prediction of engine performance for an
The authors would like to acknowledge Iranian Fuel Conserva- alternative fuel using articial neural network, Appl. Thermal Eng. 37 (2012)
tion Company for supporting us in doing this research. 217e225.
[19] C.A. Coello Coello, A.D. Christiansen, Multiobjective optimization of trusses
using genetic algorithms, Comp. Struct. 75 (2000) 647e660.
References [20] A. Osyezka, Multicriteria Optimization for Engineering Design, Design Opti-
mization, Academic Press, NY, 1985, pp. 193e227.
[1] A. Demirbas, Biodiesels e A Realistic Fuel Alternative for Diesel Engines, [21] C.M. Fonseca, P.J. Fleming, Genetic algorithms for multi-objective optimiza-
Springer, London, 2008. tion: Formulation, discussion and generalization, in: S. Forrest (Ed.), Proc.. of
[2] G. Knothe, J.V. Gerpen, J. Krahl, The Biodiesel Handbook, AOCS Press, Illinois, the Fifth Int. Conf. on Genetic Algorithms, Morgan Kaufmann, San Mateo, CA,
2005. 1993, pp. 416e423.
[3] J. Starbuck, G.D.J. Harper, Run Your Diesel Vehicle on Biofuels, McGraw-Hill, [22] C.A. Coello Coello, D.A. Van Veldhuizen, G.B. Lamont, Evolutionary Algorithms
New York, 2009. for Solving Multi-objective Problems, Kluwer Academic Publishers, NY, 2002.
[4] B. Ghobadian, H. Rahimi, A.M. Nikbakht, G. Naja, T.F. Yusaf, Diesel engine [23] V. Pareto, Cours deconomic ploitique, Rouge, Lausanne, Switzerland, 1896.
performance and exhaust emission analysis using waste cooking biodiesel fuel [24] D.E. Goldberg, Genetic Algorithms in Search, Optimization, and Machine
with an articial neural network, Renew. Energy 34 (2009) 976e982. Learning, Addison-Wesley, New York, 1989.
[5] M. Kiani Deh Kiani, B. Ghobadian, T. Tavakoli, A.M. Nikbakht, G. Naja, [25] A. Toffolo, E. Benini, Genetic diversity as an objective in multi-objective
Application of articial neural networks for the prediction of performance and evolutionary algorithms, Evol Comput 11 (2) (2003) 151e167.
exhaust emissions in SI engine using ethanol-gasoline blends, Energy 35 [26] C.L. Hwang, K. Yoon, Multiple Attribute Decision Making e Methods and
(2010) 65e69. Applications, Springer-Verlag Press, Heidelberg, 1981.

You might also like