You are on page 1of 9

Applied Thermal Engineering 27 (2007) 46–54

www.elsevier.com/locate/apthermeng

Performance and exhaust emissions of a gasoline engine


using artificial neural network
Cenk Sayin a, H. Metin Ertunc b, Murat Hosoz c, Ibrahim Kilicaslan c, Mustafa Canakci c,*

a
Department of Mechanical Education, Marmara University, 34722 Istanbul, Turkey
b
Department of Mechatronics Engineering, Kocaeli University, 41040 Kocaeli, Turkey
c
Department of Mechanical Education, Kocaeli University, 41380 Kocaeli, Turkey

Received 11 August 2005; accepted 9 May 2006


Available online 7 July 2006

Abstract

This study deals with artificial neural network (ANN) modelling of a gasoline engine to predict the brake specific fuel consumption,
brake thermal efficiency, exhaust gas temperature and exhaust emissions of the engine. To acquire data for training and testing the pro-
posed ANN, a four-cylinder, four-stroke test engine was fuelled with gasoline having various octane numbers (91, 93, 95 and 95.3), and
operated at different engine speeds and torques. Using some of the experimental data for training, an ANN model based on standard
back-propagation algorithm for the engine was developed. Then, the performance of the ANN predictions were measured by comparing
the predictions with the experimental results which were not used in the training process. It was observed that the ANN model can pre-
dict the engine performance, exhaust emissions and exhaust gas temperature quite well with correlation coefficients in the range of 0.983–
0.996, mean relative errors in the range of 1.41–6.66% and very low root mean square errors. This study shows that, as an alternative to
classical modelling techniques, the ANN approach can be used to accurately predict the performance and emissions of internal combus-
tion engines.
Ó 2006 Elsevier Ltd. All rights reserved.

Keywords: Artificial neural network; Gasoline engine; Octane number; Exhaust emissions

1. Introduction The ON of a fuel required by an engine is determined


according to the engine design and compression ratio.
Since the invent of the internal combustion engine in The weather, driving conditions, and mechanical condi-
the late 19th century, a great number of research studies tions of the engine are some factors that can influence this
have been conducted to improve the engine performance, requirement. Diminished cooling efficiency, problems in
decrease the engine fuel consumption and reduce the fuel systems, ignition troubles and failure of emission con-
unwanted exhaust emissions. Gasoline is the main fuel trols can also change ON requirements [7–9].
for the SI engine, and the octane number (ON) is one of Manufacturers and application engineers usually want
the most important parameters determining the quality of to know the performance of an engine using gasoline with
gasoline. The ON of a gasoline is a measure of its resistance various ONs for the entire range of operating conditions.
to detonation. The effects of ON on exhaust emissions, This requirement can be met either by conducting a com-
engine performance and detonation have been investigated prehensive testing study or modelling the engine operation.
by several researchers [1–6]. Testing the engine under the all possible operating condi-
tions and fuel cases are both time consuming and expen-
*
Corresponding author. Tel.: +90 262 3032285; fax: +90 262 3032203.
sive. On the other hand, developing an accurate model
E-mail addresses: canakci@kou.edu.tr, mustafacanakci@hotmail.com for the operation of a gasoline engine is too difficult due
(M. Canakci). to the complex processes involved. As an alternative, the

1359-4311/$ - see front matter Ó 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.applthermaleng.2006.05.016
C. Sayin et al. / Applied Thermal Engineering 27 (2007) 46–54 47

Nomenclature

a actual output n sum of the weighted inputs


ANN artificial neural network ON octane number
b bias p predicted output
BSFC brake specific fuel-consumption (g (kW h)1) P number of the elements in the input vector
BTE brake thermal efficiency Pe effective power (kW)
cov covariance R correlation coefficient
E expected value RMSE root mean square error
f activation function T engine torque (N m)
LHV lower heating value (kJ kg1) Texh exhaust gas temperature (°C)
MRE mean relative error Tin air inlet temperature (°C)
m_ f fuel mass flow rate (kg s1) wi interconnection weight of the input vector xi
n engine speed (rpm) xi input vector of a neuron
N number of the points in the data set

performance and exhaust emissions of an engine can be coupled to an inclined manometer for measuring mass flow
modelled using ANNs. This new modelling technique rate of the intake air. A surge tank was used for damping
can be applied to estimate desired output parameters out the pulsations produced by the engine, thus forming
when enough experimental data is provided. Therefore, a steady flow of intake air. The engine speed was continu-
ANNs allow the modelling of physical phenomena in ously monitored using a photoelectric tachometer. The
complex systems without requiring explicit mathematical engine torque was measured with a hydraulic dynamome-
representations. ter having a range of 0–100 N m. The CO and HC emis-
The ANN approach has been applied to predict the per- sions were measured using a Bilsa 320 model exhaust gas
formance of various thermal systems [10–18]. The use of analyzer, which was calibrated before each test. The fuel
ANNs for modelling the operation of internal combustion consumption of the engine was determined by monitoring
engines is a more recent progress. This approach was used the fuel level in the measurement container. The tempera-
to predict the performance and exhaust emissions of diesel tures of the intake air and exhaust gas were measured using
engines [19,20] and the specific fuel consumption and fuel- K-type thermocouples. Some features of the instrumenta-
air equivalence ratio of a diesel engine [21]. The effects of tion are summarized in Table 1, and further details of the
valve-timing in a spark ignition engine on the engine per- experimental setup can be found in Sayin et al. [24].
formance and fuel economy was also investigated using The fuels tested in the engine were commercial grade
ANNs [22]. gasoline fuels with 91, 93, 95 and 95.3-ON. Details of the
This study deals with ANN modelling of the perfor- fuel properties are given in Table 2. The BSFC, BTE,
mance and exhaust emissions of a gasoline engine. In order exhaust gas temperature, and exhaust emissions such as
to gather data for the proposed ANN model, an engine CO and HC have been investigated. The tests were per-
was fuelled with 91, 93, 95 and 95.3-ON gasoline fuels formed under steady-state conditions. The fuels were tested
and tested at various engine speeds and torques. The exper- in random order, each test was repeated 3 times and the
imental data from totally 96 test runs was used to train and results of the three repetitions were averaged. The engine
test the ANN model for predicting the brake specific fuel speed was changed between 1000 and 2400 rpm with inter-
consumption (BSFC), brake thermal efficiency (BTE), the vals of 200 rpm, while the engine torque was kept at 20, 30
emissions of carbon monoxide (CO) and unburned hydro- and 40 N m for each engine speed. In each test, engine oil
carbon (HC), and the exhaust gas temperature of the temperature, coolant temperature, fuel consumption, air-
engine. flow rate, and exhaust emissions such as CO, CO2 and
HC were recorded systematically. The value of air-fuel
2. Description of the experimental setup and testing ratio was kept constant through the experiments, which
procedure was around 14:1. The uncertainty of the air-fuel ratio
was found as ±0.3. The ignition timing during the experi-
The experiments were performed on a four-cylinder ments was kept constant at all test conditions.
four-stroke Fiat DKS 1600 cc gasoline engine which origi-
nally requires fuel with 91-ON. The engine has a cylinder 3. Experimental results
bore of 84 mm, a stroke of 71.5 mm, a compression ratio
of 9.2:1 and a maximum power output of 55.2 kW [23]. Test results for the BSFC at various speeds and constant
The test engine and instrumentation are shown in Fig. 1. loads are shown in Fig. 2. The minimum BSFC at 40 N m
The engine was equipped with a conic edge orifice meter constant load was observed as 376 g (kW h)1 for 91-ON
48 C. Sayin et al. / Applied Thermal Engineering 27 (2007) 46–54

Fig. 1. The schematic layout of the experimental setup.

Table 1 Table 2
Characteristics of the instrumentation Properties of the gasoline fuels with different ONs used in the tests [25]
Measured Instrument Range Uncertainty Octane number 91 93 95 95.3
variable Density at 0.738 0.741 0.745 0.765
Temperature Type K 0–500 °C 2 °C 15 °C (g cm3)
thermocouple Reid vapor 59.34 59.72 60.03 60.03
Air mass Orifice-inclined 0–20 mm H2O 0.5 mm H2O pressure (kPa)
flow rate manometer Distillation 198.9 201.0 209.0 210.0
Fuel mass Measurement 0–1000 cc 10 cc temperature
flow rate container (99% vol.) (°C)
Engine speed Digital tachometer 10–100,000 rpm ±2% Tetra alkyl 0.104 0.222 0.340 0.0005
Engine torque Hydraulic 0–100 N m ±2% lead (gPb l1)
dynamometer Lower heating 43 932 43 642 43 304 43 961
CO emission Exhaust gas analyzer 0–100% ±2% value (kJ kg1)
HC emission Exhaust gas analyzer 0–9999 ppm ±3%

Pe
BTE ¼ ð1Þ
1
gasoline, 388, 394 and 397 g (kW h) for 93, 95 and 95.3- m_ f LHV
ON gasoline, respectively, for 1800 rpm. It can be seen that The maximum BTE was obtained for gasoline with 91-
the trend of the BSFC in the case of 91-ON is lower than ON as shown in Fig. 2. According to the engine require-
the others. Starting at the minimum BSFC point, increas- ment, using gasoline with correct ON will diminish fuel
ing or decreasing speed at constant load increases BSFC consumption for constant effective power, thus causing
due to primarily the reduced engine volumetric efficiency. an increase in BTE.
If an excessively high octane rating gasoline is used, the One of the most important emissions from IC engines is
ignition delay becomes longer and the speed of the flame CO. The level of CO emission in the exhaust of an IC
becomes shorter. This phenomenon causes a reduction in engine varies with the fuel-air ratio. For fuel-rich mixtures,
the maximum pressure and output power, thereby increas- high CO concentrations in the exhaust emissions are
ing the fuel consumption per output power. The obtained observed. Since SI engines often operate close to stoichio-
test results confirmed this statement. metric fuel-air ratio at part load and fuel rich at full load,
As shown in the equation below, the BTE is related to CO emission in the exhaust must be controlled. On the
the effective power (Pe), fuel mass flow rate (m_ f ) and lower other hand, for fuel-lean mixtures, CO concentrations in
heating value (LHV) of the fuel the exhaust vary slightly with air-fuel ratio.
C. Sayin et al. / Applied Thermal Engineering 27 (2007) 46–54 49

900 0.16
at 20 Nm constant load 91
850 91 93
BSFC (g/kWh)
93 0.14
800 95
95

BTE
95.3
750 95.3 0.12

700
0.10
650 at 20 Nm constant load
600 0.08
1000 1200 1400 1600 1800 2000 2200 2400 1000 1200 1400 1600 1800 2000 2200 2400
Engine rpm Engine rpm

750 at 30 Nm constant load 0.16


91
700
91 0.14 93
BSFC (g/kWh)

650 93 95
600

BTE
95 0.12 95.3
550 95.3
500 0.10
450 at 30 Nm constant load
400 0.08
1000 1200 1400 1600 1800 2000 2200 2400 1000 1200 1400 1600 1800 2000 2200 2400
Engine rpm Engine rpm

650 0.24
600
at 40 Nm constant load
0.22
BSFC (g/kWh)

91
550
93 0.20
500 91
BTE

95 0.18
450 95.3 93
400 0.16 95
0.14 95.3
350 at 40 Nm constant load
300 0.12
1000 1200 1400 1600 1800 2000 2200 2400 1000 1200 1400 1600 1800 2000 2200 2400
Engine rpm Engine rpm

Fig. 2. Experimental results of BSFC and BTE at different engine loads.

The CO emissions from the engine were presented in caused by an unburned fuel-air mixture, while the other
Fig. 3. The results indicate that CO values for 95.3-ON gas- source is the engine lubricant and incomplete combustion.
oline were lower than 91, 93 and 95-ON fuels. This is prob- Similar to the CO emission, one of the most important
ably resulted from the tetra alkyl, which is one of the most variables affecting the HC emission in the exhaust is air-
important parameters in the formation of the exhaust emis- fuel ratio. In a cold engine, fuel vaporization is slow and
sion. As shown in Table 2, the amount of tetra alkyl in the the fuel flow is increased to provide an easily combustible
95.3-ON gasoline is lower compared with 91, 93 and 95- fuel-rich mixture in the cylinder during the engine warm
ON gasoline fuels. At maximum constant load and speed up. This causes an increased HC emission. At part-load
(40 N m and 2400 rpm), the minimum CO was found and warm engine conditions, the engine produces lower
0.43% for 95.3-ON gasoline and 0.54%, 0.58% and 0.59% HC emissions because of the leaner mixtures. The engine
for 91, 93 and 95-ON fuels, respectively. The values of speed and spark timing are also important since they affect
CO in all working conditions decrease with increasing not only the quality of combustion but also the combustion
rpm, which is the result of improved combustion process. process. In the case of short combustion process at high
The increase in load could probably augment the volumet- speeds, the amount of unburned HC will also decrease.
ric efficiency and boost the turbulence in combustion cham- The changes in HC emissions as a function of engine
ber, thereby ensuring more homogenous mixture and better speed for three constant loads are also shown in Fig. 3.
combustion. The HC concentrations decrease moderately with increas-
The unburned HC in the exhaust of an IC engine is ing speed and load, which were in a similar trend with
another kind of emissions, which must be controlled. These CO. The experimental results indicate that HC values of
emissions have several different sources. Most of the HC is 95.3-ON gasoline are lower than the other fuels. Probably,
50 C. Sayin et al. / Applied Thermal Engineering 27 (2007) 46–54

2.0 1000
at 20 Nm constant at 20 Nm constant load
1.8 900
1.6 800

HC (ppm)
CO (%)

700
1.4
600
1.2 91 91
500
1.0 93 93
400 95
0.8 95
300 95.3
95.3
0.6 200
1000 1200 1400 1600 1800 2000 2200 2400 1000 1200 1400 1600 1800 2000 2200 2400
Engine rpm Engine rpm

1.8 at 30 Nm constant load 600 at 30 Nm constant load


1.6 500
1.4
400

HC (ppm)
CO (%)

1.2
300
1.0 91 91
0.8 93 200 93
0.6 95 100 95
0.4 95.3 95.3
0
1000 1200 1400 1600 1800 2000 2200 2400 1000 1200 1400 1600 1800 2000 2200 2400
Engine rpm Engine rpm

1.4 500
at 40 Nm constant load at 40 Nm constant load
1.2 400
HC (ppm)
CO (%)

1.0 300

0.8 91 200 91
93 93
0.6 95 100 95
95.3 95.3
0.4 0
1000 1200 1400 1600 1800 2000 2200 2400 1000 1200 1400 1600 1800 2000 2200 2400
Engine rpm Engine rpm

Fig. 3. Experimental results of CO and HC at different engine loads.

this is caused by the lower amount of tetra alkyl lead con- an ANN as a set of connection weights and biases. The
tained in 95.3-ON gasoline compared with 91, 93 and 95.3- sum of the weighted inputs with a bias is processed through
ON fuels. As expected, minimum HC was found 82 ppm an activation function, represented by f, and the output
for 95.3-ON gasoline and 93, 112 and 114 ppm for 91, 93 that it computes is
and 95-ON fuels, respectively, at maximum constant load !
XP
and speed (40 N m and 2400 rpm). f ðnÞ ¼ f wi xi þ b ð3Þ
i¼1
4. Artificial neural networks ANNs can be trained to reach from a particular input to
a specific target output using a suitable learning method
An ANN consists of massively interconnected process- until the network output matches the target. The error
ing nodes known as neurons [26]. Each neuron accepts a between the output of the network and the desired output
weighted set of inputs and responds with an output. Such is minimized by modifying the weights and biases. When
a neuron first forms the sum of the weighted inputs given the error falls below a determined value or the maximum
by [27] number of epochs are exceeded, the training process is
X
P ceased. Then, this trained network can be used for simulat-
n¼ wi xi þ b ð2Þ ing the system outputs for the inputs which have not been
i¼1
introduced before.
where P and wi are the number of the elements and the The performance of the ANN-based prediction is evalu-
interconnection weight of the input vector xi, respectively, ated by a regression analysis between the network outputs,
and b is the bias for the neuron. The knowledge is stored in i.e. predicted parameters, and the corresponding targets,
C. Sayin et al. / Applied Thermal Engineering 27 (2007) 46–54 51

i.e. experimental values. The criterions used for measuring


the network performance are the correlation coefficient,
mean relative error and root mean square error. The corre-
lation coefficient assesses the strength of the relationship
between the predicted and experimental results, and it is
defined as [28]
covða; pÞ
Rða; pÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð4Þ
covða; aÞcovðp; pÞ
where cov(a, p) is covariance between a and p sets that refer
to the actual output and predicted output sets, respectively,
and given by [28]
covða; pÞ ¼ E½ða  la Þðp  lp Þ ð5Þ
where E is the expected value, la is the mean value of a set
and lp is the mean value of p set. Likewise, cov(a, a) and
cov(p, p) are the auto covariances of a and p sets, corre-
spondingly, and given by
2
covða; aÞ ¼ E½ða  la Þ  ð6Þ
2
covðp; pÞ ¼ E½ðp  lp Þ  ð7Þ
Fig. 4. The structure of the artificial neural network for the gasoline
The correlation coefficient ranges between 1 and +1. R engine.
values closer to +1 indicate a stronger positive linear rela-
tionship, while R values closer to 1 indicate a stronger and error with different ANN configurations, the network
negative relationship. was decided to consist of one hidden layer with fifteen neu-
The mean relative error, which shows the mean ratio rons. Furthermore, the activation function in the hidden
between the error and the experimental values, is deter- layer was chosen as tangent sigmoid function.
mined from Then, the five output parameters were predicted using a
N  
1 X  ðai  pi Þ three-layer feed-forward ANN. Utilizing standard back-
MRE ð%Þ ¼ 100 ð8Þ propagation algorithm [26,29], the input vectors and the
N i¼1  ai 
corresponding target vectors from the training set were
where N is the number of the points in the data set. used for training the network. The training procedure
Finally, the root mean square error is obtained from adjusted the weighting coefficients using Levenberg-Mar-
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi quardt algorithm [26,29]. The output of the network was
u
u1 X N
compared to the desired output at each presentation, and
RMSE ¼ t
2
ðai  pi Þ ð9Þ an error was computed. This error was then back propa-
N i¼1
gated to the ANN and used for adjusting the weights such
that the error decreases with each iteration. Consequently,
5. Modelling with the ANN the training procedure approximates a function between
the input and output variables. Then, the input vectors
An ANN model for the gasoline engine was developed from the test data set were presented to the trained network
using the data gathered in steady-state test runs. In the and the output parameters predicted by the network were
model, 70% of the data set was randomly assigned as the compared with the experimental ones for the performance
training set, while the remaining 30% was employed for measurement. The computer code solving the back-propa-
testing the performance of the ANN predictions. gation algorithm and measuring the network performance
The architecture of the ANN for the test engine is sche- was implemented under the MATLAB environment.
matically shown in Fig. 4. The inputs to the ANN are the
lower heating value (LHV) of the tested gasoline, engine 6. Results and discussion
torque (T), engine speed (n), and air inlet temperature
(Tin). Note that the effect of ON on the engine performance The ANN predictions for the output parameters of the
and emissions is represented by the LHV of the fuel. The gasoline engine as a function of the experimental ones are
outputs from the ANN are the BSFC, BTE, CO and HC shown in Fig. 5. In the graphs, the accuracy of the ANN
emissions, and the temperature of the exhaust gas (Texh). predictions was evaluated with the help of a straight line
The performance of an ANN is influenced by the some indicating the perfect prediction and a ±10% error band.
features of the network such as the number of hidden lay- The predicted versus experimental values for the brake
ers and the number of nodes in each hidden layer. By trial specific fuel consumption are indicated in Fig. 5a. The
52 C. Sayin et al. / Applied Thermal Engineering 27 (2007) 46–54

900
22
R = 0.994
R = 0.992
Predicted BSFC (g/kWh)

800 MRE = 2.24 % 20

Predicted BTE (%)


MRE = 2.97 %
RMSE = 17.05 g/kWh +10 %
700 18 RMSE = 0.51 %
+10%
600 16
-10 %
-10% 14
500
12
400
10

400 500 600 700 800 900 10 12 14 16 18 20 22


Experimental BSFC (g/kWh) Experimental BTE (%)
(a) (b)

1.8
800
R = 0.996
1.6 R = 0.992
700 MRE = 6.66 %

Predicted HC (ppm)
MRE = 3.64 %
Predicted CO (%)

1.4 RMSE = 0.04 % RMSE = 19.66 ppm


600 +10%
+10%
1.2 500 -10%
-10% 400
1.0
300
0.8
200
0.6
100
0.6 0.8 1.0 1.2 1.4 1.6 1.8 100 200 300 400 500 600 700 800
Experimental CO (%) Experimental HC (ppm)
(c) (d)

190
R = 0.983
180
MRE = 1.41 % +10%
Predicted Texh (°C)

170 RMSE = 2.96 °C

160

150 -10%
140

130

120 130 140 150 160 170 180 190


Experimental Texh(°C)
(e)

Fig. 5. The ANN predictions for the output parameters vs. experimental values.

ANN predictions for the BSFC yield a mean relative BTE, the ANN yields a correlation coefficient of 0.992, a
error of 2.24%, a root mean square error of 17.05 MRE of 2.97% and a RMSE of 0.51%.
g (kW h)1 and a correlation coefficient of 0.994. These A plot of the predicted versus experimental CO emis-
values show that the ANN predicts the BSFC quite well sions is presented in Fig. 5c. Although the amount of the
despite wide ranges of operating conditions. It is clear that combustion air supplied to the engine exceeds the theoret-
the performance of the ANN would have been even better ical amount in all test runs, some CO always appears in the
if a higher number of test runs had been performed to pro- exhaust emissions. The CO emission is caused by the fac-
vide a larger amount of experimental data for the network tors such as incomplete mixing and insufficient time for
training. complete combustion. In predicting CO emissions, the
The ANN-predicted versus experimental values for the ANN results in a correlation coefficient of 0.992, a MRE
brake thermal efficiency are reported in Fig. 5b. For the of 3.64% and a RMSE of 0.04%.
C. Sayin et al. / Applied Thermal Engineering 27 (2007) 46–54 53

Fig. 6. Comparisons of the ANN predictions and experimental results for various test patterns.

The ANN predictions for the HC emission versus the brake thermal efficiency, exhaust gas temperature, and
experimental one are shown in Fig. 5d. Since the combus- emissions of CO and HC. The experimental data for train-
tion process involves a series of complex and fast chemical ing and testing the ANN model was acquired from totally
reactions, the exhaust emissions from the internal combus- 96 steady-state test runs. The developed ANN was a three-
tion engines usually contain some unburned hydrocarbons. layer one utilizing the lower heating value of the gasoline,
Besides polluting the environment, excessive HC emissions engine speed, engine torque and air inlet temperature as
increase operating costs of the engines. For the HC emis- the input parameters. The ANN predictions for the perfor-
sion, the ANN yields a correlation coefficient of 0.996, a mance and exhaust emissions of the tested gasoline engine
MRE of 6.66% and a RMSE of 19.66 ppm. yielded a good statistical performance with correlation
The ANN-predicted versus experimental values for the coefficients in the range of 0.983–0.996, MREs in the range
exhaust gas temperature are indicated in Fig. 5e. The of 1.41-6.66% and quite low RMSE values.
exhaust gas temperature is a measure of the energy loss Comparisons of the ANN predictions and the experi-
accompanying the exhaust gas. Increasing the engine load mental results demonstrate that engines using gasoline with
and speed improve the turbulence and volumetric efficiency various ONs can accurately be modelled using ANNs.
and promote a more homogenous mixture in the combus- Consequently, with the use of ANNs, the performance
tion chamber, thus causing an increase in the exhaust gas and exhaust emissions of the internal combustion engines
temperature. The ANN predictions for the exhaust gas tem- can easily be determined by performing only a limited
perature yield a correlation coefficient of 0.983, a MRE of number of tests instead of a detailed experimental study,
1.41% and a RMSE of 2.96 °C. thus saving both engineering effort and funds.
Comparisons of the ANN predictions and experimental
results for the BSFC, BTE, Texh and for the emissions of References
CO and HC are alternatively presented in Fig. 6. It is seen
that the test patterns consist of the results of 29 tests. [1] I. Batmaz, M. Balci, S. Salman, B. Erdiller, Experimental analysis of
fuel economy and exhaust emissions at petrol engine vehicles, in: First
Although data from these tests were not introduced to
Automotive Technology Congress with International Participation,
the ANN during the training process, the ANN excellently Adana, Turkey, 1997, pp. 95–103.
predicts the five output parameters for the entire range of [2] T. Sakaguchi, Influence of diffusion of fuel-efficient motor vehicles on
the experiments. gasoline demand for individual user owned passenger cars, Energy
Policy 28 (2000) 895–903.
[3] I. Korkmaz, A study on the performance and emission characteristics
7. Conclusions of gasoline and methanol fuelled spark-ignition engines, Ph.D. Thesis,
Istanbul Technical University, Istanbul, Turkey, 1996.
[4] H. Bayraktar, Theoretical investigation of the effect of gasoline-
The ANN approach has been applied to a gasoline ethanol blends on spark-ignition engine combustion and cycles, Ph.D.
engine for predicting the brake specific fuel consumption, Thesis, Karadeniz Technical University, Trabzon, Turkey, 1997.
54 C. Sayin et al. / Applied Thermal Engineering 27 (2007) 46–54

[5] C. Twu, J. Coon, A generalized interaction method for the predic- [16] E. Arcaklioglu, Performance comparison of CFCs with their substi-
tion of octane numbers for gasoline blends, Simulation Science Inc., tutes using artificial neural network, International Journal of Energy
1998. Research 28 (2004) 1113–1125.
[6] K. Mogi, H. Katsushi, K. Arisawa, H. Kobayashi, Analysis and [17] H.M. Ertunc, M. Hosoz, Artificial neural network analysis of a
avoidance of pre-ignition in spark-ignition engines, JSAE Review 19 refrigeration system with an evaporative condenser, Applied Thermal
(1998) 9–14. Engineering 26 (2006) 627–635.
[7] J.B. Heywood, Internal combustion engine fundamentals, Mc Graw- [18] M. Hosoz, H.M. Ertunc, Artificial neural network analysis of an
Hill, USA, 1984. automobile air conditioning system, Energy Conversion and Man-
[8] P. Sudsanguan, S. Chanchaowna, Using higher octane rating gasoline agement 47 (2006) 1574–1587.
than engine requirement: loss or gain, Research Report of King [19] M. Canakci, A. Erdil, E. Arcaklioglu, Performance and exhaust
Mongkut’s University of Technology, Thailand, 1999. emissions of a biodiesel engine, Applied Energy 83 (2006) 594–605.
[9] S. Chanchaowna, The effect of gasoline octane number on engine [20] E. Arcaklioglu, I. Celikten, A diesel engine’s performance and
performance, Research Report of King Mongkut’s University of exhaust emissions, Applied Energy 80 (2005) 11–22.
Technology, Thailand, 1999. [21] V. Celik, E. Arcaklioglu, Performance maps of a diesel engine,
[10] S.A. Kalogirou, Application of artificial neural-networks for energy Applied Energy 81 (2005) 247–259.
systems, Applied Energy 67 (2000) 17–35. [22] M. Golcu, Y. Sekmen, P. Erduranli, S. Salman, Artificial neural
[11] A. Pacheco-Vega, M. Sen, K.T. Yang, R.L. McClain, Neural network network based modelling of variable valve-timing in a spark ignition
analysis of fin-tube refrigerating heat exchanger with limited exper- engine, Applied Energy 81 (2005) 187–197.
imental data, International Journal of Heat Mass Transfer 44 (2001) [23] FIAT, Automobile Technical Specification, Bursa, Turkey, 1997.
763–770. [24] C. Sayin, I. Kilicaslan, M. Canakci, N. Ozsezen, An experimental
[12] H. Bechtler, M.W. Browne, P.K. Bansal, V. Kecman, Neural study of the effect of octane number higher than engine requirement
networks – a new approach to model vapor-compression heat on the engine performance and emissions, Applied Thermal Engi-
pumps, International Journal of Energy Research 25 (2001) 591– neering 25 (2005) 1315–1324.
599. [25] TUPRAS, Product Specification, Izmit, Turkey, 1996.
[13] M.M. Prieto, E. Montanes, O. Menendez, Power plant condenser [26] M.T. Hagan, H.B. Demuth, Neural network design, PWS Publishing
performance forecasting using a non-fully connected ANN, Energy 26 Company, Boston, USA, 1995.
(2001) 65–79. [27] S. Haykin, Neural networks: a comprehensive foundation, Mc
[14] A. Chouai, S. Laugeier, D. Richon, Modelling of thermodynamic Millan, New Jersey, USA, 1994.
properties using neural networks – application to refrigerants, Fluid [28] C.G. Looney, Pattern recognition using neural networks: theory
Phase Equilibria 199 (2002) 53–62. and algorithms for engineers and scientists, Oxford University Press,
[15] A. Sozen, E. Arcaklioglu, M. Ozalp, A new approach to thermody- New York, USA, 1997.
namic analysis of ejector-absorption cycle: artificial neural networks, [29] F.M. Ham, I. Kostanic, Principles of neurocomputing for science and
Applied Thermal Engineering 23 (2003) 937–952. engineering, Mc Graw-Hill, Singapore, 2001.

You might also like