Documente Academic
Documente Profesional
Documente Cultură
Energy
journal homepage: www.elsevier.com/locate/energy
University of Karabk, Faculty of Engineering, Department of Mechanical Engineering, 78050 Karabk, Turkey
University of Dzce, Dzce Vocational School of Higher Education, 81500 Dzce, Turkey
c
Yldrm Beyazt University, Faculty of Engineering and Natural Sciences, Department of Mechanical Engineering, 06050 Ankara, Turkey
d
University of Dzce, Faculty of Technology, Department of Manufacturing Engineering, 81620 Dzce, Turkey
b
a r t i c l e i n f o
a b s t r a c t
Article history:
Received 26 June 2012
Received in revised form
17 October 2012
Accepted 27 October 2012
Available online 1 December 2012
This study investigates the use of ANN (articial neural networks) modelling to predict BSFC (break
specic fuel consumption), exhaust emissions that are CO (carbon monoxide) and HC (unburned
hydrocarbon), and AFR (airefuel ratio) of a spark ignition engine which operates with methanol and
gasoline. To obtain training and testing data, a number of experiments were performed with a fourcylinder, four-stroke test engine operated at different engine speeds and torques. The experimental
results reveal that the methanol improved the emission characteristics compared with the gasoline. For
the ANN modelling, the standard back-propagation algorithm was found to be the optimum choice for
training the model. In the building of the network structure, four different learning algorithms were used
such as BFGS (Quasi-Newton back propagation), LM (LevenbergeMarquardt learning algorithm). It was
found that the ANN model is able to predict the engine performance and exhaust emissions with
a correlation coefcient of 0.998621, 0.977654, 0.998382 and 0.996075 for the BSFC, CO, HC and AFR for
testing data, respectively. It was obvious that the developed ANN model is fairly powerful for predicting
the brake specic fuel consumption and exhaust emissions of internal combustion engines.
2012 Elsevier Ltd. All rights reserved.
Keywords:
Gasoline
Methanol
ANN
Engine performance
Exhaust emissions
1. Introduction
It is well known that all vehicles in the world are dependent on
fossil fuels such as gasoline, diesel fuel, LPG (liqueed petroleum
gas), LNG (liqueed natural gas) and CNG (compressed natural
gas). The fossil fuel used in vehicles causes factors threatening
human health such as air pollution, acid rains, smog, built up of
carbon dioxide, changes in the heat balance of the earth, and so on.
In addition, world fossil fuel reserves are also limited and will be
consumed away in the near future [1]. Therefore, it seems that the
use of alternative fuels is inevitable. These fuels include alcohols
(such as ethanol and methanol), ethers, vegetable oils, animal fats,
gaseous fuels and bio-diesel [2]. Among the alternative fuels,
methanol is considered to be one of the most favourable fuels for
engines. The reason is that methanol is liquid fuel and similar to
gasoline and diesel in aspects of usage, storage and transportation.
In addition, it can be produced from widely available raw materials
including coal, natural gas and bio-substances. Methanol is a good
fuel for the spark-ignition engine. Benets such as higher efciency, specic power and lower emissions can be realised with
methanol. Methanol molecule contains 50% oxygen which leads to
combustion speed faster, also its higher laminar ame speed allows
it to be run with rarefaction or more dilute air/fuel mixtures [3].
Methanol fuelled buses had once been introduced to reduce
pollutant emission. However, operational problems have slowed
down the development of methanol-fuelled vehicles, leading to the
phasing out of the methanol-fuelled buses from the market [4].
Due to limitations on the development of methanol-fuelled
engines, lots of studies have been done to improve the engine
performance and to reduce exhaust emissions [5e11]. Yilmaz [12]
performed comparative analysis of biodieseleethanolediesel and
biodieselemethanolediesel blends in a diesel engine. Performance
and emission characteristics of the engine fuelled with biodiesele
methanolediesel and biodieseleethanolediesel were compared to
standard diesel fuel as the baseline. Overall biodieselealcohole
diesel blends showed higher brake specic fuel consumption
than diesel. As alcohol concentrations in blends increase, CO
(carbon monoxide) and HC (unburned hydrocarbon) emissions
increase, while NO (nitrogen oxides) emissions are reduced. Also,
methanol blends were more effective than ethanol blends for
178
Nomenclature
AFR
ANN
BFGS
BSFC
BTE
CNG
CO
f
Ft
HC
i.j
LHV
LM
LNG
LPG
m
MEP
MRE
airefuel ratio
articial neural network
Quasi-Newton back propagation
break specic fuel consumption
brake thermal efciency
compressed natural gas
carbon monoxide
transfer function
fuel type
unburned hydrocarbon
processing elements
lower heating value
LevenbergeMarquardt learning algorithm
liqueed natural gas
liqueed petroleum gas
fuel ow
mean error percentage
mean relative errors
N
NETi
n
NOx
o
p
PSI
R2
RMSE
RP
SCG
t
T
THC
wij
wbi
Xj
engine speed
the weighted sum of the input to the ith processing
element
number of processing elements in the previous layer
nitrogen oxides
output value of ANN
number of pattern
pounds per square inch
determination coefcient
root mean square error
resilient back propagation
scaled conjugate gradient learning algorithm
experimental data
torque
total hydrocarbon
the weights of the connections between ith and jth
processing elements
the weights of the biases between layers
the output of the jth processing element
179
Table 1
Technical specications of the test engine.
Item description
Stery
Engine type
Work type
Cylinder number
Total cylinder volume
Idling speed
Cylinder diameter
Piston stroke length
Compression ratio
Maximum torque
Maximum power
Cooling type
Ford-Escort
4-Stroke
4-Cylinder, sequence type
1297 cc
900 rpm
80.978 mm
62.99 mm
8:1
80 Nm (3000 rpm)
40 kW (5500 rpm)
Water cooled
180
NETi
n
X
wij xj wbi
(1)
j1
f NETi
1
1 eNETi
(2)
(3)
181
Table 2
Prediction accuracy values for CO, HC, BSFC and AFR using four different learning algorithms.
Learning algorithm
Network structure
SCG
SCG
SCG
SCG
SCG
SCG
SCG
SCG
SCG
SCG
SCG
LM
LM
LM
LM
LM
LM
LM
LM
LM
LM
LM
BFGS
BFGS
BFGS
BFGS
BFGS
BFGS
BFGS
BFGS
BFGS
BFGS
BFGS
RP
RP
RP
RP
RP
RP
RP
RP
RP
RP
RP
4-5-1
4-6-1
4-7-1
4-8-1
4-9-1
4-10-1
4-11-1
4-12-1
4-13-1
4-14-1
4-15-1
4-5-1
4-6-1
4-7-1
4-8-1
4-9-1
4-10-1
4-11-1
4-12-1
4-13-1
4-14-1
4-15-1
4-5-1
4-6-1
4-7-1
4-8-1
4-9-1
4-10-1
4-11-1
4-12-1
4-13-1
4-14-1
4-15-1
4-5-1
4-6-1
4-7-1
4-8-1
4-9-1
4-10-1
4-11-1
4-12-1
4-13-1
4-14-1
4-15-1
HC
BSFC
AFR
Training set
Testing set
Training set
Testing set
Training set
Testing set
Training set
Testing set
90.89
93.09
91.97
93.44
92.52
96.06
95.59
94.51
97.24
95.69
96.96
93.86
92.76
95.08
96.33
97.02
98.98
97.56
98.54
98.80
98.59
98.69
92.13
93.24
93.05
93.62
97.13
97.71
95.39
97.05
98.13
98.51
98.69
90.78
91.76
90.11
91.86
91.25
92.05
93.08
93.04
93.01
94.00
92.89
84.50
84.37
89.37
86.13
87.88
85.59
90.06
83.32
87.13
84.93
86.56
87.97
84.26
91.44
85.98
83.96
82.57
85.02
83.10
81.60
84.38
82.85
88.93
88.45
90.66
88.17
86.92
86.18
84.56
77.84
76.58
74.92
75.07
87.59
90.47
89.67
87.37
88.05
89.00
89.07
85.46
87.71
84.82
86.28
94.63
96.16
97.23
96.37
97.82
97.71
97.85
97.57
98.90
99.20
98.90
95.69
97.73
98.07
98.41
98.74
99.37
99.38
99.45
99.27
99.37
99.34
96.29
96.12
97.52
97.22
97.87
97.89
98.29
98.88
99.32
99.20
99.30
95.50
95.29
96.11
96.91
96.44
97.13
96.99
97.02
96.72
98.02
97.41
93.20
94.53
93.66
94.89
93.48
94.43
96.26
91.59
93.37
92.16
91.87
94.15
95.05
95.14
92.11
90.33
89.47
87.07
86.62
89.82
83.99
88.24
91.97
92.00
92.78
94.52
91.89
92.38
92.73
91.66
91.43
88.80
90.26
94.33
94.45
93.76
93.08
95.13
94.46
94.08
95.85
94.99
96.48
93.90
97.34
98.06
98.59
98.48
98.63
99.03
98.68
98.48
99.01
99.02
98.57
98.05
98.49
98.57
98.74
98.77
98.78
98.95
98.97
99.08
99.12
98.91
95.04
98.02
97.84
98.69
98.56
98.76
98.83
98.95
99.01
98.99
98.83
96.14
96.90
98.02
98.18
97.74
97.80
97.99
98.26
98.29
98.67
98.05
97.03
96.91
97.40
96.08
94.75
96.38
96.14
95.41
95.60
94.21
94.03
96.05
96.79
96.11
95.37
94.71
94.56
95.88
95.35
95.58
93.62
94.83
95.12
96.56
96.45
94.58
96.13
96.72
96.87
95.22
94.44
95.63
95.63
94.02
95.74
95.57
95.71
95.57
96.51
96.66
96.68
96.27
95.54
95.21
97.22
97.29
98.42
98.59
98.40
98.35
98.43
98.67
99.11
98.90
99.19
98.37
98.49
99.22
99.09
98.81
99.24
99.40
99.22
99.29
99.25
99.26
98.17
98.09
98.45
98.71
98.77
99.17
98.93
99.21
99.31
99.31
99.35
96.88
96.84
97.71
97.61
98.18
98.27
98.27
98.17
98.49
98.19
98.30
93.12
94.00
95.95
95.25
95.08
94.97
96.31
93.43
94.68
94.53
94.06
95.03
96.00
94.77
95.51
93.46
94.01
94.81
93.48
94.30
93.77
93.97
94.62
94.93
95.44
94.58
94.56
94.35
96.72
95.28
93.29
93.59
92.87
94.73
94.56
94.76
94.59
95.87
94.80
94.11
95.09
95.08
94.97
92.52
monoxide, unburned hydrocarbon, break specic fuel consumption and airefuel ratio were normalised dividing by 5, 5000, 35,
26, 8, 1500, 4600 and 20, respectively. The digits for the fuel type
to be entered into the articial neural networks were determined
0
11=2
X
2
1
@
tj oj A
RMSE
p
j
P
j tj oj tj 100
MEP %
p
R2 1
Fig. 3. ANN architecture with a single hidden layer for carbon monoxide.
2 !
P
j tj oj
P 2
j oj
(4)
(5)
(6)
182
lower heating value is 2.24 times less than the lower heating value
of gasoline. Therefore, when the engine is operated on methanol, its
brake specic fuel consumption will be approximately two times
higher than gasoline. As no modication was applied to the engine,
engine performance was negatively affected as higher octane value
of methanol was not beneted. Moreover, in case of insufcient
intake manifold heating the fuel enters the cylinders as liquid. The
negative aspects appearing can easily be seen when looking at the
performance characteristic curves.
Because the lower heating value of methanol is lower than the
lower heating value of gasoline, if the same power and rotation
value are aimed, the ow rate of the methanol sent to cylinders has
to be increased. This case serves as one of the basic reasons of
excessive increase in brake specic fuel consumption. With it is
a high evaporation temperature, methanol is heated with the preheat process applied on suction manifold and an attempt is made to
partially meet the energy needed. Although the manifold was
heated from outside, it can be clearly seen from engine performance that this temperature was less than sufcient. Therefore, the
Fig. 4. Comparison of CO, HC, BSFC, and AFR values for gasoline and methanol engine.
183
Table 3
Statistical data for CO, HC, BSFC and AFR using four different algorithms.
Goal
Sequence
Learning algorithm
Network structure
Training set
Testing set
RMSE
R2
MEP
RMSE
R2
MEP
CO
1
2
3
4
LM
BFGS
RP
SCG
4-7-1
4-7-1
4-6-1
4-11-1
0.020667
0.027081
0.035288
0.014658
0.996902
0.994665
0.990877
0.998443
4.922218
6.950387
8.239697
4.414095
0.049213
0.050708
0.046600
0.051715
0.977654
0.976326
0.979911
0.977706
8.564867
9.343023
9.531393
9.944888
1
2
3
4
RP
SCG
LM
BFGS
4-14-1
4-11-1
4-7-1
4-8-1
0.012874
0.010802
0.011896
0.015316
0.999332
0.999529
0.999430
0.999054
1.975396
2.146407
1.927521
2.775601
0.018734
0.021846
0.027593
0.037676
0.998382
0.997842
0.996507
0.993543
3.515423
3.735359
4.855570
5.481618
1
2
3
4
SCG
BFGS
LM
RP
4-7-1
4-11-1
4-6-1
4-12-1
0.007156
0.004075
0.006477
0.007890
0.999634
0.999881
0.999700
0.999555
1.413558
1.170858
1.505776
1.744768
0.010976
0.012582
0.013977
0.010183
0.998621
0.998207
0.997738
0.998841
2.595133
3.126220
3.210128
3.321998
1
2
3
4
BFGS
SCG
RP
LM
4-11-1
4-11-1
4-9-1
4-8-1
0.006147
0.011275
0.012429
0.005067
0.999872
0.999569
0.999477
0.999913
1.073933
1.568444
1.819480
0.911500
0.032479
0.037634
0.028903
0.038115
0.996075
0.994873
0.997047
0.995011
3.277073
3.694142
4.130722
4.486403
HC
BSFC
AFR
Fig. 5. Matching of the experimental and ANN values for testing sets of CO, HC, BSFC and AFR.
184
CO
HC
1
1 e0:47$F126:5857$F20:6243$F30:3029$F40:1169$F530:1875$F62:5677$F71:3785
(7)
1
1 e1:6669$F10:7857$F22:3483$F32:3565$F41:8671$F50:9765$F63:7981$F71:2198$F80:7510$F91:9372$F101:2220$F112:5052$F121:9013$F131:8961$F143:6045
(8)
BSFC
AFR
1
1 e1:3444$F114:2160$F21:051$F33:7197$F412:3523$F52:5808$F610:3370$F71:3444
(9)
1
1 e5:5227$F16:3870$F211:4906$F33:1129$F48:8843$F50:8038$F61:7763$F73:9820$F84:6401$F914:6868$F109:6594$F110:4704
(10)
which includes some performance features comparable to biological neural networks. The use of an ANN model is considered as
a practical and reliable approach for non-linear problems is to test
the prediction ability of specic fuel consumption, air fuel ratio
and exhaust emissions for a gasoline/methanol engine. The input
parameters of the network are fuel type, engine speed, torque and
fuel ow and its output parameters are also carbon monoxide,
unburned hydrocarbon, break specic fuel consumption and aire
fuel ratio. In this study, a computer program has been developed
in MATLAB platform to predict CO, HC, BSFC and AFR of the engine.
The optimum network structures and statistical parameters of
ANN models for four learning algorithms were given in Table 3. It
is apparent from Table 3, the prediction performances for both
training and testing sets of HC, BSFC and AFR showed that all the
approaches provide a quite satisfactory accuracy. Their R2 values
are more than 0.99. R2 value of testing set of CO is only about 0.97
while R2 value of training set of CO is only more than 0.99. The
best prediction results were obtained by LM, RP, SCG and BFGS
learning algorithms for CO, HC, BSFC and AFR respectively. But, in
general, SCG learning algorithm gave optimal or near optimal
results for all engine values. The LM learning algorithm had the
highest speed compared with the other learning algorithms and it
reached to optimal solutions with small number of neurons in
hidden layer.
Fi
1
1 eEi
(11)
where Ei is the weighted sum of the inputs, and is calculated via the
equations in Tables 4e7 respectively. The data ow was completed
with the weights between the layers. The weight values of the input
and hidden layers are given in the Tables 4e7. Here, the effect of the
parameters that are at the input layer (fuel type, engine speed,
torque, and fuel ow) on the BSFC, CO, HC and AFR can be observed.
Table 4
Weights between input layer and hidden layer for carbon monoxide.
Ei w1 x Ft w2 x N w3 x T w4 x m qi
i
w1
w2
w3
w4
qi
1
2
3
4
5
6
7
20.5567
3.2179
8.8508
7.4488
2.3375
2.3561
3.0032
18.8115
2.6905
1.6973
5.2651
16.5006
2.5278
4.2582
43.5460
0.6381
1.6597
0.1136
34.2741
0.6399
0.9201
137.1448
2.1837
9.6879
10.3693
0.3408
2.0305
0.4256
64.8508
2.3224
1.5509
1.0095
16.1222
1.9868
0.9356
w1
w2
w3
w4
qi
1
2
3
4
5
6
7
8
9
10
11
12
13
14
2.3233
20.7017
37.6477
5.4834
20.8279
0.6687
3.1090
5.7232
2.1559
15.8216
9.5964
2.0175
5.1409
2.1905
14.8439
6.4541
1.1580
9.3051
3.4757
0.7165
0.1522
4.0079
50.9460
0.2831
1.2197
7.6085
11.3877
6.5494
8.8019
37.3731
0.5470
4.8038
12.8066
3.3239
8.3508
16.6259
1.9872
12.6757
3.8456
9.4049
0.1544
12.0538
2.0713
13.3634
10.9087
3.8739
53.1614
27.3673
2.5015
2.5595
39.0036
7.5211
5.6979
1.0777
1.7356
2.4864
7.9305
13.1148
28.9961
1.2853
0.3983
3.3342
4.3130
5.8816
12.8126
9.0443
1.9857
2.8794
9.4755
4.1366
Table 6
Weights between input layer and hidden layer for break specic fuel consumption.
Ei w1 x Ft w2 x N w3 x T w4 x m qi
i
w1
w2
w3
w4
qi
1
2
3
4
5
6
7
6.1367
0.9215
6.2304
8.2761
13.3334
2.0337
10.8711
8.6065
6.1443
7.5374
4.2407
9.6042
0.4886
0.8904
2.7542
9.7505
5.2516
2.2491
21.8834
0.1355
0.2756
1.1512
3.2887
2.1166
0.5642
11.6602
8.3803
2.6254
11.4027
4.3706
0.8362
1.4162
5.6802
0.3306
1.6316
185
References
Table 7
Weights between input layer and hidden layer for air-fuel ratio.
Ei w1 x Ft w2 x N w3 x T w4 x m qi
i
w1
w2
w3
w4
qi
1
2
3
4
5
6
7
8
9
10
11
3.4724
0.6950
13.9449
7.6819
5.2721
0.6250
1.6734
16.8867
7.3617
0.4572
6.7352
6.0658
9.5892
8.1972
24.1519
12.4115
5.7735
21.6636
22.9654
5.8099
8.6139
4.9454
5.4282
7.2658
1.6756
0.0617
1.1646
16.3744
16.0527
1.7804
1.5742
0.7837
3.2717
4.9926
9.2055
9.9349
7.9679
7.2202
18.6978
18.8932
11.1511
6.1791
2.9229
3.8012
4.3480
24.3873
1.5099
5.4471
8.8788
16.5786
19.4814
14.2951
2.0851
4.8989
0.7176
4. Conclusions
In this study, an ANN modelling was used for the prediction of
the BSFC, CO, HC and AFR of a spark ignition engine which operates
with methanol and gasoline fuels. For training period, different
training algorithms such as BFGS, LM, RP and SCG back propagation
were used. The best results in the prediction of BSFC, CO, HC and
AFR were obtained by network architecture of 4-7-1, 4-7-1, 4-14-1
and 4-11-1 and the SCG, LM, RP and BFGS learning algorithms,
respectively. In the ANN model, the coefcients of determination of
the BSFC, CO, HC and AFR for both the training and the testing set
were notably close to 1. It was determined that the ANN results
obtained for the BSFC, CO, HC and AFR were within acceptable error
limits (5%). These results show that the learning capacity of the
ANN is quite powerful in the prediction of the BSFC, CO, HC and
AFR. Therefore, the use of ANN is highly recommended for the
prediction of break specic fuel consumption and exhaust emissions without conducting complicated, expensive, and timeconsuming experimental studies. This study shows that an ANN
186
[18] Topcu IB, Saridemir M. Prediction of compressive strength of concrete containing y ash using articial neural networks and fuzzy logic. Comput Mater
Sci 2008;41:305e11.
[19] Ismail HM, Ng HK, Queck CW, Gan S. Articial neural networks modelling of
engine-out responses for a light-duty diesel engine fuelled with biodiesel
blends. Appl Energy 2012;92:769e77.
[20] Tinaut FV, Melgar A, Gimenez B, Reyes M. Prediction of performance and
emissions of an engine fuelled with natural gas/hydrogen blends. Int J Hydrog
Energy 2011;36:947e56.
[21] Yusaf TF, Yousif BF, Elawad MM. Crude palm oil fuel for diesel-engines:
experimental and ANN simulation approaches. Energy 2011;36:4871e8.
[22] Sharon H, Jayaprakash R, Karthigai Selvan M, Soban Kumar DR, Sundaresan A,
Karuppasamy K. Biodiesel production and prediction of engine performance
using SIMULINK model of trained neural network. Fuel 2012;99:197e203.
[23] Yap WK, HoT Karri V. Exhaust emissions control and engine parameters
optimization using articial neural network virtual sensors for a hydrogenpowered vehicle. Int J Hydrog Energy 2012;37:8704e15.
[24] Naja G, Ghobadian B, Tavakoli T, Buttsworth DR, Yusaf TF, Faizollahnejad M.
Performance and exhaust emissions of a gasoline engine with ethanol
blended gasoline fuels using articial neural network. Appl Energ 2009;86:
630e9.
[25] Ghobadian B, Rahimi H, Nikbakht AM, Naja G, Yusaf TF. Diesel engine
performance and exhaust emission analysis using waste cooking biodiesel
fuel with an articial neural network. Renew Energ 2009;34:976e82.
[26] Deh Kiani MK, Ghobadian B, Tavakoli T, Nikbakht AM, Naja G. Application of articial neural networks for the prediction of performance and
exhaust emissions in SI engine using ethanol- gasoline blends. Energy
2010;35:65e9.