Applied Thermal Engineering 37 (2012) 217e225

Contents lists available at SciVerse ScienceDirect

Applied Thermal Engineering

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

Prediction of engine performance for an alternative fuel using articial

neural network
irog
Yusuf ay a, *, Adem iek b, Fuat Kara c, Selami Sag lu a
a
University of Karabk, Faculty of Engineering, Department of Mechanical Engineering, Balklarkayas Mevkii, 78050, Karabk, Turkey
b
University of Dzce, Faculty of Technology, Department of Manufacturing Engineering, Konuralp Yerleskesi, 81620, Dzce, Turkey
c
University of Dzce, Faculty of Technical Education, Department of Mechanical Education, Konuralp Yerleskesi, 81620, Dzce, Turkey

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

Article history: This study deals with articial neural network (ANN) modeling to predict the brake specic fuel
Received 17 June 2011 consumption, effective power and average effective pressure and exhaust gas temperature of the
Accepted 7 November 2011 methanol engine. To obtain training and testing data, a number of experiments were performed with
Available online 20 November 2011
a four-cylinder, four-stroke test engine operated at different engine speeds and torques. Using some of
the experimental data for training, an ANN model based on standard back propagation algorithm was
Keywords:
developed. Then, the performance of the ANN predictions was measured by comparing the predictions
Spark ignition engine
with the experimental results. Engine speed, engine torque, fuel ow, intake manifold mean temperature
Methanol engine performance
Articial neural network
and cooling water entrance temperature have been used as the input layer, while brake specic fuel
consumption, effective power, average effective pressure and exhaust gas temperature have also been
used separately as the output layer. After training, it was found that the R2 values are close to 1 for both
training and testing data. RMS values are smaller than 0.015 and mean errors are smaller than 3.8% for
the testing data. This shows that the developed ANN model is a powerful one for predicting the brake
specic fuel consumption, effective power and average effective pressure and exhaust gas temperature of
internal combustion engines.
2011 Elsevier Ltd. All rights reserved.

1. Introduction ignition advances. The best performance at standard ignition

The requirement of energy increases due to industrialization
and continuously growing population in the world. Therefore,
developing and newly developing countries tend to the new energy
sources to compensate their energy necessity. Currently, the main
energy source of the motor vehicles is petroleum products. It is
expected that the petroleum reserves will be consumed away in the
near future. In addition, one of the main causes of air pollution in
the cities is harmful emissions of the motor vehicles which are
operated with petroleum products. As a result, a lot of researchers
have started to search for cheap, renewable and environmentally
friendly alternative fuels such as ethanol, methanol, hydrogen, and
liqueed petroleum gas (LPG), liqueed natural gas (LNG),
compressed natural gas (CNG), electric and vegetable oil based
fuels [1]. Calisir and Gms [2] analyzed the usage of mixture of
methanol and gasoline in spark-ignited engine and studied the
effects of methanol ratio on engine performance in different
* Corresponding author. Tel.: 90 370 4332021; fax: 90 380 5421134.
E-mail address: yusufcay@karabuk.edu.tr (Y. ay).
advance value of engine is acquired by mixture of 15% methanol
and 85% petroleum. Celik et al. [3] investigated the effects of use of
pure methanol as fuel at high compression ratio in a single cylinder
gasoline engine on engine power, brake thermal efciency and
emissions. By increasing the compression ratio from 6/1 to 10/1, the
engine power and brake thermal efciency increased by up to 14%
and 36%, respectively. Moreover, CO, CO2 and NOx emissions were
reduced by about 37%, 30% and 22%, respectively. Li et al. [4]
experimentally investigated the effects of injection time and igni-
tion stroke on engine performance and emissions through
a methanol engine with high pressure, direct injection and spark
ignition. It was found that both injection time and ignition stroke
have signicant effects on methanol engine performance,
combustion process and exhaust emissions. The best compromise
values between thermal effectiveness and exhaust emissions were
acquired at the determined optimal injection time and ignition
stroke. Sayin [5] investigated the effects of mixtures of methanol-
diesel (M5-M10) and ethanol-diesel (E5-E10) on engine perfor-
mance and exhaust emissions. For the experiments, an ordinary
diesel engine with single cylinder and four-cycle was used. The
results showed that break specic fuel consumption and nitrogen

1359-4311/$ e see front matter 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.applthermaleng.2011.11.019
218 Y. ay et al. / Applied Thermal Engineering 37 (2012) 217e225

Nomenclature p number of pattern

Pe effective power
ANN articial neural network R2 absolute fraction of variance
Ape average effective pressure RMS root mean square error
BSFC break specic fuel consumption SCG scaled conjugate gradient learning algorithm t target
f transfer function value
GA genetic algorithm T torque
i,j processing elements Tcw cooling water entrance temperature
LM LevenbergeMarquardt learning algorithm Tin intake manifold mean temperature
m fuel ow Tex exhaust gas temperature
MEP mean error percentage wij the weights of the connections between ith and jth
n number of processing elements in the previous layer processing elements
N engine speed wbi the weights of the biases between layers
NETi the weighted sum of the input to the ith processing Xj the output of the jth processing element
element o output value

oxide emissions increased while thermal effectiveness, smoke mean temperature and cooling water entrance temperature in the
opacity, carbon monoxide and total hydro carbon decreased. input layer. By this way, prediction of some parameters such as
ANN is the algorithms which are developed in order to explore effective power, exhaust gas temperature, average effective pres-
and generate new knowledge via learning without taking any sure and specic fuel consumption was aimed. Also, the mathe-
external help [6]. In other words, it is a synthetic network which matical models for outputs were obtained using MATLAB program.
imitates biological neural network. ANN and biological neural
network have signicant differences in terms of both architectures
and capabilities. ANN constitutes of a mathematical model for 2. Experimental set up and measurements
prediction of new problems [7]. Recently, ANN has been a notable
and commonly used method for engine performance tests, cutting The engine used in this experiment was a 4 cylinder, 4 strokes,
mechanics, signal processing, data decomposition and image pro- 1.3 L volume Ford-Escort automobile engine. In the hydraulic
cessing [8]. Besides, the method is able to produce new solutions brakes, a DPX1A type braking system which has 100 kW power and
for some problems. ANN was preferred as control strategy due to its 750 rpm rates that can rise up to 200 Nm torque maximum was
high reliability and efciency [9]. The studies of literature show that used. The technical specications of the test engine are given in
ANN has a powerful modeling technique [10e13]. Khandelwal and Table 1. Electronic mean rotation tachometer was used for
Singh [14] studied on the prediction of macerals contents of Indian measuring engine rotation. The sensitivity of the device for
coals from proximate and ultimate analyses using ANN. It was measuring number of rotation is 0.04 rpm. During experiments,
found that coefcient of determination between measured and the engine inlet and outlet temperatures of engine cooling water,
predicted macerals by ANN was quite higher as well as mean the temperature of the motor oil in the pan, the insert temperature
absolute percentage error was very marginal as compared to for every cylinder, temperature of exhaust gases, the temperature
conventional multivariate regression analysis prediction. Rajendra of fuel at the inlet of carburetor, and the temperature of the testing
et al. [15] predicted the optimized pretreatment process parame- environment were measured by three different thermometers,
ters for biodiesel production using ANN and GA. Balabin et al. namely digital, mechanical and mercurial. The measurement
developed an ANN model for analysis of biodiesel density, kine- ranges of digital, mechanical and mercurial thermometers
matic viscosity, methanol and water contents using near infrared were 230  Ce127  C, 0  Ce100  C and 0  Ce250  C respectively.
spectroscopy. The lowest root mean squared errors of prediction of The measurement sensitivities of the digital, mechanical and
the developed ANN for density, viscosity, water percentage, and mercurial thermometers were 0.01  C, 1  C and 1  C,
methanol content were 0.42 kg m3, 0.068 mm2 s1, 45 ppm, and respectively.
51 ppm, respectively [16]. Ghobadian et al. [17] developed an ANN Valve adjustments, point gapping, ignition advance adjustment,
model to estimate diesel engine performance and exhaust emission plug gapping, measurements of compression pressures of each
analysis using waste cooking biodiesel fuel. In their study, it was cylinder and measurement of their ignition voltages were con-
observed that the ANN model can predict the engine performance ducted according to the catalogue values of the engine and tests
and exhaust emissions quite well with correlation coefcient
0.9487, 0.999, 0.929 and 0.999 for the engine torque, SFC, CO and Table 1
HC emissions, respectively. Kiani et al. [18] dealt with ANN Technical specications of the test engine.
modeling of a spark ignition engine which used ethanol- gasoline
Item description Stery
blends to predict the engine brake power, output torque and
Engine type Ford-escort
exhaust emissions (CO, CO2, NOx and HC) of the engine. Results Work type 4-stroke
showed that the ANN provided the best accuracy in modeling the Cylinder number 4-cylinder, sequence type
emission indices with correlation coefcient equal to 0.98, 0.96, Total cylinder volume 1297 cc
0.90 and 0.71 for CO, CO2, HC and NOx, and 0.99 and 0.96 for torque Idling speed 900 rpm
Cylinder diameter 80.978 mm
and brake power respectively.
Piston stroke length 62.99 mm
In this study, the changes in engine performance have been Compression pressure 11 bar (3600 rpm)
observed by using methanol as fuel without any modications on Compression ratio 8:1
a gasoline automobile engine, and the impact of the fuel on engine Maximum torque 80 Nm ((3000 rpm)
performance has been examined. An ANN model was developed by Maximum power 40 kW (5500 rpm)
Cooling type Water cooled
considering the torque, engine speed, fuel ow, intake manifold
 Y. ay et al. / Applied Thermal Engineering 37 (2012) 217e225
were initiated. In the experiments, the rotation value of the engine
was kept at 1100 rpm, and the engine was loaded at 5e70 Nm range
by means of hydraulic dynamometer. Loading was continued at
intervals until the engine reached vibration running limits. The
engine rotation was increased in equal intervals of 300 rpm at
1100e4300 rpm range and loading operations were repeated at
each rotation. Engine test mechanism can be seen in Fig. 1.
In the intervals within which the engine showed regular oper-
ation behavior, environment, fresh insert, exhaust, motor oil, water
inlet and outlet temperatures and atmospheric pressure, oil pres-
sure, mean relative humidity, skewed manometer deviation
installed at the air tank, fuel ow rate, torque value shown by
dynamometer, motor rotation value and exhaust emission values
were observed as test values. The test results obtained in the
experimental study was used to train and test the ANN.
3. Construction of ANN model
3.1. Network structure
ANN consists of articial neural cells (neurons). ANN has three
main layers, namely, input, hidden and output layers. Neurons
(processing elements) at input layer transfer data from external
world to hidden layer [6]. The data in input layer do not process as
the data in the other layers. In the hidden layer, outputs are
produced using data from neurons in input layer and bias, and
summation and activation functions. There can be more than one
hidden layer. In this case, each hidden layer sends outputs to the
following layer.
In the output layer, the output of network is produced by pro-
cessing data from hidden layer and sent to external world. The
summation function calculates net input coming to a cell. For this
reason, different functions are used. The most common one is to
calculate the weighted sum. Inputs (N, T, m, Tin, Tcw) are the
knowledge from other cells or external world to the input cells.
These are determined by examples that network wants to learn.
Weights (w1, w2 .wn) are the values which determine the effect of
Fig. 1. The schematic layout of the experimental setup.
219
input set or another processing element in previous layer on the
processing element. Each input value is multiplied by weight value
which connects it to the processing element and then, it is
combined by summation function. Thus, net input of the network
can be found. Summation function is given in Eq. (1).
Xn
NETi j1
wijxj wbi (1)
Activation function provides a curvilinear match between input
and output layers. In addition, it determines the output of the cell
by processing net input to the cell. Selection of appropriate acti-
vation function signicantly affects network performance.
Recently, logistic sigmoid transfer function has been commonly
used as an activation function in multilayer perception model,
because it is a differentiable, continuous and non-linear function
[19]. For this reason, the logistic sigmoid transfer function was used
as the activation function in this study. This function produces
a value between 0 and 1 for each value of net input. The formula of
the logistic sigmoid function is as follows:
1
f NETi (2)
1 eNETi
According to trial results, ANN architecture with ve neurons in
input layer, six neurons in hidden layers and one neuron in output
layers obtained better predictions for effective power and is shown
in Fig. 2.
3.2. Learning algorithm
There are many learning algorithm in order to determine
weights in ANN. One of the most common learning algorithms is
back propagation. The back propagation method updates the
weights in accordance with difference between available data and
network output. Learning parameter used in the method has
a great importance in order to reach the optimal results. Learning
parameter can be constant or dynamically updated in the model.
220 Y. ay et al. / Applied Thermal Engineering 37 (2012) 217e225

specic fuel consumption and exhaust gas temperature are given as

SCG:5-7-1, LM:5-5-5-1 and LM:5-7-1 respectively.

3.3. Training and testing data

Determination of percentages of training and testing data has an

important role for building of ANN architecture. When the studies
in literature are analyzed, it is revealed that different ratios for
training and testing data are used [26e29]. The percentage of
training and testing data are 90%:10% [23,30], 85%:15% [27],
80%:20% [17], 75%:25 [25], 70%:30% [24,31]. In this study, 55
experimental results were prepared for training and testing data of
ANN. The ratio for training and testing data was selected as
80%:20%. In this context, 11 data for testing and 44 data for training
were randomly selected.

Fig. 2. ANN architecture with a single hidden layer. 3.4. Normalization of input and output data

In back propagation model, scaling of inputs and outputs

dramatically affects performance of ANN. As mentioned above,
There are various training functions that have been applied by the
logistic sigmoid transfer function was used in this study. One of
previous studies such as Bayesian regularization [20], gradient
the characteristics of this function is that only a value between
descent with adaptive learning rule [21], gradient descent with
0 and 1 can be produced. Input and output data sets are normal-
momentum and adaptive learning rule [22], scaled conjugate
ized before training and testing process. In ANN, normalization
gradient [11,23] and LevenbergeMarquardt [23,24,25]. In order to
can be performed between 0 and 1 [25,30,32] or 1 and 1
acquire the closest output values to experimental results, the best
[11,17,26].
learning algorithm and optimum number of neurons in hidden
 
layer was determined. For this reason, both SCG and LM learning vmin  vi
algorithms and different numbers (3e15) of neurons in hidden nvi 2x 1 (3)
vmin  vmax
layer were used in the built network structure for effective power.
In consequence of trials, the best learning algorithm and network  
vmin  vi
architecture for prediction of effective power became LM: 5-6-1 nvi 0:8x 0:1 (4)
vmin  vmax
(Fig. 2). Determination of the best learning algorithm and
optimal number of neurons for effective power is demonstrated in According to studies in the literature, two formulas are used for
Table 2. The best learning algorithms and ANN architecture for the normalization (Eqs. (3) and (4) [20,21]. In this study, the input
other output parameters, namely average effective pressure, and output values were normalized between 0.1 and 0.9 using

Table 2
Statistical data for the effective power using two different algorithms.

Learning algorithm Number of neurons Training data Testing data

RMS R2 MEP RMS R2 MEP

SCG 3 0.008062 0.999705 1.639147 0.010739 0.999392 2.669630
SCG 4 0.006625 0.999801 1.459743 0.020445 0.997875 3.600677
SCG 5 0.004767 0.999897 0.889307 0.017605 0.998343 2.853290
SCG 6 0.004766 0.999897 0.773381 0.012831 0.999149 2.761909
SCG 7 0.004767 0.999897 1.046037 0.009219 0.999557 1.913609
SCG 8 0.004763 0.999897 1.126455 0.012590 0.999180 3.443527
SCG 9 0.004765 0.999897 1.092406 0.010965 0.999373 2.763547
SCG 10 0.004765 0.999897 1.181055 0.008954 0.999587 2.346251
SCG 11 0.004763 0.999897 0.824662 0.014268 0.998927 4.441902
SCG 12 0.004735 0.999898 1.046426 0.013959 0.998978 3.462730
SCG 13 0.004761 0.999897 0.913550 0.011692 0.999287 3.186602
SCG 14 0.004829 0.999894 1.260758 0.014243 0.998950 4.429296
SCG 15 0.004764 0.999897 0.937891 0.017505 0.998387 3.512959
LM 3 0.009762 0.999568 2.037289 0.012181 0.999242 3.381476
LM 4 0.007324 0.999757 1.644281 0.012755 0.999157 3.911727
LM 5 0.004719 0.999899 1.074534 0.015989 0.998657 3.916263
LM 6 0.003905 0.999931 0.823561 0.006449 0.999781 1.143801
LM 7 0.004752 0.999898 1.037537 0.017516 0.998340 3.647614
LM 8 0.004542 0.999906 1.052637 0.010299 0.999449 2.646704
LM 9 0.004375 0.999913 0.989547 0.012385 0.999219 3.125261
LM 10 0.004007 0.999927 0.948795 0.011434 0.999302 3.250956
LM 11 0.002652 0.999968 0.680720 0.012744 0.999142 2.703216
LM 12 0.003685 0.999938 0.912282 0.015789 0.998706 4.747974
LM 13 0.004102 0.999924 0.908650 0.011807 0.999273 3.044730
LM 14 0.004647 0.999902 1.075433 0.013764 0.998989 2.499398
LM 15 0.003774 0.999935 0.921759 0.011542 0.999296 3.286815
 Y. ay et al. / Applied Thermal Engineering 37 (2012) 217e225 221

Table 3
Normalization values of inputs and outputs.

No N (rpm) T (Nm) m (g/s) Tin ( C) Tcw ( C) Pe (kW) Ape (kPa) Tex ( C) BSFC (g/kWh)
1 0.1 0.1 0.1 0.9 0.9 0.1 0.1013 0.1703 0.8713
2 0.1 0.2 0.1197 0.8433 0.676 0.1379 0.2 0.2099 0.4403
3 0.1 0.3 0.1516 0.8055 0.676 0.1765 0.3 0.2582 0.3185
4 0.1 0.4 0.1639 0.774 0.612 0.2145 0.3999 0.3154 0.2313
5 0.1 0.5 0.1805 0.774 0.548 0.2531 0.5001 0.3769 0.1836
6 0.1 0.6 0.1909 0.7614 0.516 0.2917 0.6 0.4385 0.1463
7 0.1 0.7 0.212 0.7488 0.452 0.3296 0.6999 0.4912 0.104
8 0.1 0.8 0.2442 0.7425 0.42 0.3682 0.8001 0.522 0.1451
9 0.1 0.9 0.2717 0.7236 0.292 0.4062 0.9 0.5923 0.1848
10 0.1828 0.1 0.1126 0.6354 0.164 0.11 0.1013 0.1 0.782
11 0.1828 0.2 0.1381 0.6228 0.1 0.1592 0.2 0.1747 0.4011
12 0.1828 0.3 0.1864 0.5598 0.132 0.2078 0.3 0.2582 0.3085
13 0.1828 0.4 0.1873 0.5094 0.164 0.2564 0.3999 0.3418 0.2083
14 0.1828 0.5 0.2104 0.4465 0.132 0.3057 0.5001 0.4165 0.1684
15 0.1828 0.6 0.2197 0.4024 0.228 0.3542 0.6 0.4736 0.1313
16 0.1828 0.7 0.258 0.3646 0.1 0.4028 0.6999 0.5308 0.1
17 0.1828 0.8 0.2928 0.2827 0.164 0.4514 0.8001 0.5879 0.1399
18 0.1828 0.9 0.3309 0.2134 0.26 0.5007 0.9 0.6275 0.1119
19 0.2931 0.1 0.1232 0.5913 0.228 0.124 0.1013 0.2275 0.6238
20 0.2931 0.2 0.1839 0.5346 0.228 0.1879 0.2 0.2846 0.3717
21 0.2931 0.3 0.2332 0.5094 0.164 0.2478 0.3 0.3418 0.2736
22 0.2931 0.4 0.2487 0.4717 0.196 0.3143 0.3999 0.4209 0.1947
23 0.2931 0.5 0.26 0.4339 0.164 0.3742 0.5001 0.4516 0.1449
24 0.2931 0.6 0.2999 0.3646 0.26 0.4408 0.6 0.5044 0.1316
25 0.2931 0.7 0.3524 0.3142 0.292 0.5007 0.6999 0.5527 0.1226
26 0.2931 0.8 0.3891 0.2575 0.26 0.5606 0.8001 0.5967 0.1354
27 0.2931 0.9 0.4051 0.1882 0.26 0.6271 0.9 0.6319 0.2016
28 0.4862 0.1 0.1622 0.6165 0.26 0.1485 0.1 0.3462 0.5222
29 0.4862 0.2 0.315 0.5787 0.292 0.2356 0.2 0.4209 0.3774
30 0.4862 0.3 0.404 0.5346 0.26 0.3228 0.3 0.4736 0.3154
31 0.4862 0.4 0.471 0.478 0.356 0.4099 0.3999 0.5396 0.2551
32 0.4862 0.5 0.5033 0.4339 0.324 0.497 0.5001 0.5747 0.2014
33 0.4862 0.6 0.5733 0.352 0.292 0.5841 0.6 0.6495 0.1812
34 0.4862 0.7 0.6205 0.2827 0.388 0.6712 0.6999 0.7286 0.1589
35 0.4862 0.8 0.6434 0.2008 0.356 0.7584 0.8001 0.7813 0.1299
36 0.4862 0.9 0.6707 0.1189 0.388 0.8455 0.9 0.856 0.104
37 0.6241 0.1 0.2382 0.5409 0.356 0.1812 0.1 0.2758 0.6291
38 0.6241 0.2 0.317 0.4843 0.324 0.2877 0.2 0.3769 0.356
39 0.6241 0.3 0.4022 0.4276 0.356 0.4275 0.3 0.4473 0.2847
40 0.6241 0.4 0.5387 0.3646 0.292 0.5206 0.3999 0.5484 0.256
41 0.6241 0.5 0.6071 0.3016 0.324 0.6404 0.5001 0.6407 0.2166
42 0.6241 0.6 0.6421 0.2386 0.26 0.7336 0.6 0.7022 0.2014
43 0.6241 0.7 0.711 0.1882 0.228 0.7869 0.6999 0.7945 0.161
44 0.6241 0.8 0.7756 0.1063 0.292 0.9 0.8001 0.9 0.1475
45 0.7621 0.1 0.3605 0.5472 0.26 0.1832 0.1 0.3418 0.7065
46 0.7621 0.2 0.4695 0.4906 0.292 0.3057 0.2 0.4604 0.4047
47 0.7621 0.3 0.5525 0.4024 0.26 0.4275 0.3 0.5308 0.2881
48 0.7621 0.4 0.6837 0.3331 0.228 0.5493 0.3999 0.5835 0.2847
49 0.7621 0.5 0.7472 0.2512 0.292 0.671 0.5001 0.6495 0.2532
50 0.7621 0.6 0.9 0.1693 0.356 0.7935 0.6 0.7462 0.2012
51 0.9 0.1 0.5446 0.5157 0.228 0.2005 0.1 0.3681 0.9
52 0.9 0.2 0.675 0.4528 0.26 0.3403 0.2 0.4385 0.5056
53 0.9 0.3 0.7336 0.3457 0.26 0.4794 0.3 0.5308 0.3356
54 0.9 0.4 0.7647 0.226 0.228 0.6191 0.3999 0.6363 0.2926
55 0.9 0.5 0.8606 0.1 0.388 0.7582 0.5001 0.7462 0.2278

formula in Eq. (4). The normalization values for inputs and outputs   X 1=
1
of the network are given in Table 3. RMS
p
j
j j
t  oj j 2 2
(5)

3.5. Statistical evaluation of outputs P 2 !

j tj  oj
R2 1  P  2 (6)
BP training algorithm is a ramp descent algorithm. The BP j oj

P  
algorithm minimizes total error by changing the weights through

its ramp and thus tries to improve the performance of the network. j tj  oj tj x100
The training of the network is stopped when the tested values of MEP (7)
p
RMS stop decreasing and begin to increase. In order to understand
whether an ANN is making good predictions, the testing data that Where t is the target value, o is the output and p is the number of
has never been presented to the network, is used and the results are samples. The results predicted by ANN and experimental results
checked at this stage [33]. The statistical methods of RMS, R2 and were compared to each other after training. RMS (Eq. 5), R2 (Eq. (6))
MEP values have been used for comparisons [10,11,30]. and MEP (Eq. (7)) were used for comparison.
222 Y. ay et al. / Applied Thermal Engineering 37 (2012) 217e225

Fig. 3. Variations of performance values for methanol engine.

4. Results and discussion
4.1. Variations of Pe, Ape, Tex and BSFC depending on engine speed
and torque
The experimental results are shown in Fig. 3. The effective
pressure increases proportionally with increasing torque at
constant engine speed (Fig. 3a). Engine power increases as well as
the engine speed at constant torques. This behavior is the general
characteristic of internal combustion engines. Depending on engine
speed, while rate of increase of power is small at low torque values,
it is found to be higher at high torque values. These two cases can be
explained by the lowest values of pumping and heat loose at
increased engine speed, the optimal value of turbulence, the
maximum value of pressure and temperature inside the cylinder,
the fuel-air ratio being at the same rate in all regions of the
combustion chamber and the increase of lling and work in unit
time. Another remarkable point in Fig. 3a is that the slope of power-
torque curve is maximum at 4000 rpm. This is because of the excess
amount of oxygen in methanol, high thermal efciency due to low
temperature of suction pad and high density, and of the best
condition supplied for combustion at this rate in the combustion
chamber. As it can be seen in Fig. 3b, the average effective pressure,
since it is directly linked to torque, increases linearly. The thermal
efciency increases with the increase in torque and this causes an
increase in the average effective pressure. In Fig. 3c, the exhaust gas
temperatures display an increase depending on the torque at
constant engine speed. The slopes of exhaust gas temperature
curves at each engine speed depending on torque, are very close to
each other as it can be seen from the gure. This can also be
explained by the increase of thermal efciency. While the specic
fuel consumption in engines is maximum at minimum torque
values, it is lowered at other torque values. The reason is that, since
the heating value of methanol is about the half of that of gasoline,
methanol consumption increased and as a result the specic fuel
consumption is high. In Fig. 3d, at lowered torque rates, while the
difference between BSFC values obtained at maximum and
minimum engine speeds is high; this difference becomes low at
increased torque rates. This reveals that at rates where torque is
low, the engine runs unstable and knocking, otherwise it runs with
high-performance. At lowered rates since the thermal efciency is
in the lowest value, in order to have the engine supplies enough
torque, the fuel-air mixture taken into the cylinder should be
increased. The required engine torque due to increased amount of
fuel increases the fuel consumption as well.
4.2. Prediction of engine performance using ANN
The aim of using the ANN model, considered as a practical
approach, is to test the ability to predict specic fuel consumption,
engine power, average effective pressure and exhaust temperature
for a methanol engine. The network has ve input parameters:
engine speed, torque, fuel ow, intake manifold mean temperature
and cooling water entrance temperature.

Table 4
Statistical values for methanol engine.

Goal Learning algorithm Number of neurons Training data Testing data

RMS R2 MEP RMS R2 MEP

Pe LM 6 0.003905 0.999931 0.823561 0.006449 0.999781 1.143801
APe SCG 7 0.004765 0.999916 1.528305 0.0068 0.999831 2.074398
Tex LM 7 0.004397 0.99993 0.795259 0.013531 0.999282 3.140767
BSFC LM 5e5 0.004737 0.999825 1.234845 0.011913 0.998521 3.761731
 Y. ay et al. / Applied Thermal Engineering 37 (2012) 217e225 223

Fig. 4. Matching of the experimental and ANN model values for testing sets of Pe, Ape, Tex and BSFC.

Statistical values obtained for each output parameter is given in
Table 4. As shown in Table 4, it was shown that R2 are close to 1 for
both training and testing data. Similarly, RMS and mean error
percentage are fairly low. All MEP results for training and testing
data are within acceptable error limits (5%). Predictive ability of
the network for the specic fuel consumption, engine power,
average effective pressure and the exhaust temperature was found
to be fairly satisfactory.
The highest MEP values for effective power, average effective
pressure, exhaust gas temperature and specic fuel consump-
tion was calculated as 3.25% at 4000 rpm engine speed and 15
Nm torque, 12.40% at 3000 rpm engine speed and at 5 Nm tor-
que, 9.73% at 1400 rpm engine speed and for 15 Nm torque
and 8.19% for 3500 rpm engine speed and 20 Nm torque,
respectively.
In the training period, for Pe, Ape, Tex and BSFC, the mean
relative error was found to be less than 1.6%. The maximum errors
in the testing period between predicted and measured values were
found to be in the range 3e13%. The mean relative error during
testing was found to be 3.76% in predicting specic fuel
consumption, 1.14% in predicting effective power, 2.07% in pre-
dicting average effective pressure and 3.14% in predicting exhaust
gas temperature. Matching of the experimental values and the
values predicted by ANN for testing sets of Pe, Ape, Tex and BSFC
was shown in Fig. 4. As shown in Fig. 4, predictive ability of the
network for the Pe, Ape, Tex and BSFC was very good.
The equations of the effective power, average effective pressure,
exhaust gas temperature and brake specic fuel consumption are
given in Eqs. (8)e(11). Also, Pe, Ape, Tex and BSFC of the methanol
engine can be accurately calculated by these formulas.

1
Pe (8)
1 e1:4607F10:0141F20:0038F30:8217F42:4808F59:4737F61:6080

Table 6
Table 5 Weights between input layer and hidden layer for average effective pressure.
Weights between input layer and hidden layer for effective power.
I_ 1
I_ Fi
Fi
1 1 ew1 Nw2 Tw3 mw4 Tin w5 Tcw qi
1 ew1 Nw2 Tw3 mw4 Tin w5 Tcw qi
w1 w2 w3 w4 w5 qi
w1 w2 w3 w4 w5 qi
1 2.4183 1.8406 2.4448 6.0472 5.4174 12.4959
1 1.8753 0.6018 1.0001 8.3148 4.9101 2.6209 2 2.6993 2.9009 4.9763 3.3074 4.1692 3.9604
2 2.4974 2.7309 2.9669 1.0380 0.6452 3.8879 3 3.5661 3.0140 2.9753 3.9556 1.9582 5.9395
3 1.9683 2.5111 4.5373 1.4289 3.0554 1.3872 4 0.5856 2.6737 3.0100 6.2950 4.3751 6.6687
4 6.2361 0.2589 2.3992 2.8393 0.3491 3.7052 5 0.3896 0.7600 6.2336 5.6204 0.7717 1.1852
5 3.4094 3.6861 3.2303 0.1049 1.0419 2.1847 6 0.2484 4.2772 0.3037 0.2534 0.7706 0.0734
6 0.6313 3.8983 5.2719 2.7041 1.5751 4.0664 7 3.1196 4.2079 0.7541 3.8453 1.9351 3.1314
224 Y. ay et al. / Applied Thermal Engineering 37 (2012) 217e225

1
APe (9)
1 e1:5118F10:2032F20:7779F33:4992F40:0108F56:7585F63:8696F72:8952

1
Tex (10)
1 e5:2115F17:5846F20:8095F35:4005F48:9348F55:2034F60:8767F75:0572

1
BSFC (11)
1 e9:4341F18:0574F210:1680F34:4145F40:4963F51:4014

Table 7 values of a methanol engine. To train the network, N, T, m, Tin and

Weights between input layer and hidden layer for exhaust gas temperature. Tcw are used as the input, while the outputs are BSFC, Tex, Ape and
i 1 Pe. This study deals with ANN modeling of a methanol engine to
Fi
1 ew1 Nw2 Tw3 mw4 Tin w5 Tcw qi predict the brake specic fuel consumption, effective power,
w1 w2 w3 w4 w5 qi average effective pressure and exhaust gas temperature of the
1 0.2622 0.8638 1.2079 7.3785 0.7152 2.1422 engine. Using some of the experimental data for training, an ANN
2 10.2475 4.2641 13.1398 4.2328 3.6590 4.5660 model based on standard back propagation algorithm for the
3 5.6124 15.6497 4.7771 1.4393 2.4866 4.8388 engine was developed. Then, the performance of the ANN predic-
4 13.0878 0.3090 2.1506 4.5552 2.4497 6.9391 tions were measured by comparing the predictions with the
5 6.1636 5.4309 3.8861 2.5760 0.0146 2.2605
experimental results which were not used in the training process.
6 10.6552 3.7820 0.9610 1.1763 3.6155 4.3587
7 0.6216 6.3461 0.9591 0.5975 1.2864 7.6510 Engine speed, engine torque, fuel ow, intake manifold mean
temperature and cooling water entrance temperature were used as
the input layer, while brake specic fuel consumption, engine
Table 8 power average effective pressure and exhaust manifold tempera-
Weights among layers for break specic fuel consumption. ture were used separately as the output layer. After training, it was
found that the R2 values are closely 1 for the training and testing
Weights between rst and second hidden layers
1 data. RMS values are smaller than 0.015 and mean errors are
Fi
1 ew1 N1 w2 N2 w3 N3 w4 N4 w5 N5 qi smaller than 3.8% for the testing data. The results may easily be
j w1 w2 w3 w4 w5 qj considered to be within the acceptable limits. The relationship
1 8.5021 5.8863 0.1896 6.9624 2.1129 3.7662
between fuel properties and engine performance can be deter-
2 1.9392 11.1112 2.4918 6.2377 17.0421 6.0910 mined at different engine speeds by using the network. Therefore,
3 2.9427 13.0911 3.3345 0.2694 6.6699 9.4252 the usage of ANNs may be highly recommended to predict the
4 2.4800 8.0603 6.6686 8.4274 0.9775 2.2013 engines performance and exhaust gas temperature instead of
5 4.8419 4.0994 4.1877 1.0585 1.0323 1.6204
complex and time-consuming experimental studies. This study
Weights between input layer rst hidden layer shows that, as an alternative to classical modeling techniques, the
1
Nj ANN approach can be used to accurately predict the problems of
1 ew1 Nw2 Tw3 mw4 Tin w5 Tcw qj
internal combustion engines. Moreover, as an alternative to clas-
i w1 w2 w3 w4 w5 qj
sical modeling techniques, the ANN approach can be used to
1 0.6018 0.4427 4.8668 9.3818 16.6715 12.3387 accurately predict the performance, temperature and other
2 1.6242 4.9997 1.5911 0.1435 2.8806 2.5529
3 6.4301 3.5437 4.3163 6.9783 3.2267 1.0562
parameters of internal combustion engines.
4 3.7890 14.6413 5.4238 2.2726 0.9534 0.9113
5 15.1118 4.9085 6.7407 5.0553 7.3873 4.6286
References

[1] A. Surmen, M.I. Karamangil, R. Arslan, Engine Thermodynamics. Aktuel

where Fi values for Tex, Pe and APe are the activation functions and
were calculated by the equations in Table 5, 6 and 7 respectively. It
was calculated using weights between input and hidden layer. The
weights for engine power, average effective pressure and exhaust
temperature are also given in Table 5, 6 and 7 respectively.
Activation function (Fi) for specic fuel consumption was
calculated using weights between rst and second hidden layers
after calculation of Nj function using weights between input and
rst hidden layers due to two layered ANN architecture in Table 8.
5. Conclusions
The applicability of ANNs has been investigated for the perfor-
mance, average effective pressure and exhaust gas temperature
Publishing, Istanbul, 2004.
[2] A. Calisir, M. Gumus, The effect of gasoline-methanol blends on the engine
performance and exhaust emission in a spark ignition engine. IATS 2009:
Proceedings of the 5th International Advanced Technologies Symposium;
2009 May 13-15; Karabk, Turkey. p. 189e1898.
[3] M.B. Celik, B. Ozdalyan, F. Alkan, The use of pure methanol as fuel at high
compression ratio in a single cylinder gasoline engine, Fuel 90 (2011)
1591e1598.
[4] J. Li, C.M. Gong, Y. Su, H.L. Dou, X.J. Liu, Effect of injection and ignition timings
on performance and emissions from a spark-ignition engine fueled with
methanol, Fuel 89 (2010) 3919e3925.
[5] C. Sayin, Engine performance and exhaust gas emissions of methanol and
ethanolediesel blends, Fuel 89 (2010) 3410e3415.
[6] E. Oztemel, Articial Neural Network. Papatya Publishing, Istanbul, 2003.
[7] G. Zhang, B.E. Patuwo, M.Y. Hu, Forecasting with articial neural networks:
the state of the art, International Journal of Forecasting 14 (1998) 35e62.
[8] A. Negarestani, S. Setayeshi, M. Ghannadi-Maragheh, B. Akashe, Estimation of
the radon concentration in soil related to the environmental parameters by
 Y. ay et al. / Applied Thermal Engineering 37 (2012) 217e225
a modied adaline neural network, Applied Radiation and Isotopes 58 (2)
(2003) 269e273.
[9] M. Taheri, A. Mohebbi, Design of articial neural networks using a genetic
algorithm to predict collection efciency in venturi scrubbers, Journal of
Hazardous Materials 157 (2008) 122e129.
[10] I. Korkut, A. Acir, M. Boy, Application of regression and articial neural network
analysis in modeling of tool-chip interface temperature in machining, Expert
Systems with Applications (2011). doi:10.1016/j.eswa.2011.03.044.
[11] A. Kurt, Modelling of the cutting tool stresses in machining of Inconel 718
using articial neural networks, Expert Systems with Applications 36 (2009)
9645e9657.
[12] J. Porteiro, J. Collazo, D. Patio, J.L. Mguez, Diesel engine condition monitoring
using a multi-net neural network system with nonintrusive sensors, Applied
Thermal Engineering 31 (2011) 4097e4105.
[13] T. Boushaki, S. Guessasma, J.C. Sautet, Predictive analysis of combined burner
parameter effects on oxy-fuel ames, Applied Thermal Engineering 31 (2011)
202e212.
[14] M. Khandelwal, T.N. Singh, Prediction of macerals contents of Indian coals
from proximate and ultimate analyses using articial neural networks, Fuel 89
(2010) 1101e1109.
[15] M. Rajendra, P.C. Jena, H. Raheman, Prediction of optimized pretreatment
process parameters for biodiesel production using ANN and GA, Fuel 88
(2009) 868e875.
[16] R.M. Balabin, E.I. Lomakina, R.Z. Saeva, Neural network (ANN) approach to
biodiesel analysis: analysis of biodiesel density, kinematic viscosity, methanol
and water contents using near infrared (NIR) spectroscopy, Fuel 90 (2011)
2007e2015.
[17] B. Ghobadian, H. Rahimi, A.M. Nikbakht, G. Naja, T.F. Yusaf, Diesel engine
performance and exhaust emission analysis using waste cooking biodiesel
fuel with an articial neural network, Renewable Energy 34 (2009) 976e982.
[18] M. Kiani Deh Kiani, B. Ghobadian, T. Tavakoli, A.M. Nikbakht, G. Naja,
Application of articial neural networks for the prediction of performance and
exhaust emissions in SI engine using ethanol- gasoline blends, Energy 35
(2010) 65e69.
[19] B. Krose, P. van der Smagt, An Introduction to Neural Networks, eighth ed. The
University of Amsterdam, 1996.
[20] E.O. Ezugwu, D.A. Fadare, J. Bonneya, R.B.D. Silva, W.F. Sales, Modelling the
correlation between cutting and process parameters in highspeed machining
of Inconel 718 alloy using an articial neural network, International Journal of
Machine Tools and Manufacture 45 (2005) 1375e1385.
225
[21] C. Sanjay, C. Jyothi, A study of surface roughness in drilling using mathe-
matical analysis and neural networks, International Journal of Advanced
Manufacturing Technology 29 (2006) 846e852.
[22] J.P. Davim, V.N. Gaitonde, S.R. Karmik, Investigations into the effect of cutting
conditions on surface roughness in turning of free machining steel by ANN
models, Journal of Material Processing 205 (2008) 16e23.
[23] M. Nalbant, H. Gokkaya, I. Toktas, G. Sur, The experimental investigation of the
effects of uncoated, PVD and CVD coated cemented carbide inserts and cutting
parameters on surface roughness in CNC turning and its prediction using
articial neural networks, Robotics and Computer-Integrated Manufacturing
25 (2009) 211e223.
[24] T.F. Yusaf, D.R. Buttsworth, K.H. Saleh, B.F. Yousif, CNG-diesel engine perfor-
mance and exhaust emission analysis with the aid of articial neural network,
Applied Energy 87 (2010) 1661e1669.
[25] A. Parlak, Y. Islamoglu, H. Yasar, A. Egrisogut, Application of articial neural
network to predict specic fuel consumption and exhaust temperature for
a Diesel engine, Applied Thermal Engineering 26 (2006) 824e828.
[26] G. Naja, B. Ghobadian, T. Tavakoli, D.R. Buttsworth, T.F. Yusaf,
M. Faizollahnejad, Performance and exhaust emissions of a gasoline engine
with ethanol blended gasoline fuels using articial neural network, Applied
Energy 86 (2009) 630e639.
[27] N. Pasadakis, S. Sourligas, Ch Foteinopoulos, Prediction of the distillation
prole and cold properties of diesel fuels using mid-IR spectroscopy and
neural networks, Fuel 85 (2006) 1131e1137.
[28] G. Liu, L. Wang, H. Qu, H. Shen, X. Zhang, S. Zhang, et al., Articial neural
network approaches on compositioneproperty relationships of jet fuels based
on GCeMS, Fuel 86 (2007) 2551e2559.
[29] E. Jorjani, S.C. Chelgani, Mesroghli Sh, Application of articial neural networks
to predict chemical desulfurization of Tabas coal, Fuel 87 (2008) 2727e2734.
[30] E. Arcaklioglu, I. Celikten, A diesel engines performance and exhaust emis-
sions, Applied Energy 80 (2005) 11e22.
[31] C. Sayin, H.M. Ertunc, M. Hosoz, I. Kilicaslan, M. Canakci, Performance and
exhaust emissions of a gasoline engine using articial neural network,
Applied Thermal Engineering 27 (2007) 46e54.
[32] H. Oguz, I. Saritas, H.E. Baydan, Prediction of diesel engine performance using
biofuels with articial neural network, Expert Systems with Applications 37
(2010) 6579e6586.
[33] Z. Azlan Mohd, H. Habibollah, S. Saan, Prediction of surface roughness in the
end milling machining using articial neural network, Expert Systems with
Applications 37 (2010) 1755e1768.