Applied Energy 87 (2010) 16611669

Contents lists available at ScienceDirect

Applied Energy
journal homepage: www.elsevier.com/locate/apenergy

CNG-diesel engine performance and exhaust emission analysis

with the aid of articial neural network
Talal F. Yusaf a,*, D.R. Buttsworth a, Khalid H. Saleh a, B.F. Yousif b
a
University of Southern Queensland, Faculty of Engineering and Surveying, Mechanical Engineering, Toowoomba, 4350 Queensland, Australia
b
Mechanical Department, Faculty of Engineering, University of Nottingham, Symenih, Malaysia

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

Article history: This study investigates the use of articial neural network (ANN) modelling to predict brake power, tor-
Received 29 March 2009 que, break specic fuel consumption (BSFC), and exhaust emissions of a diesel engine modied to operate
Received in revised form 21 September 2009 with a combination of both compressed natural gas CNG and diesel fuels. A single cylinder, four-stroke
Accepted 6 October 2009
diesel engine was modied for the present work and was operated at different engine loads and speeds.
Available online 6 November 2009
The experimental results reveal that the mixtures of CNG and diesel fuel provided better engine perfor-
mance and improved the emission characteristics compared with the pure diesel fuel. For the ANN mod-
Keywords:
elling, the standard back-propagation algorithm was found to be the optimum choice for training the
CNG fuel
ANN
model. A multi-layer perception network was used for non-linear mapping between the input and output
Engine performance parameters. It was found that the ANN model is able to predict the engine performance and exhaust
Engine emission emissions with a correlation coefcient of 0.9884, 0.9838, 0.95707, and 0.9934 for the engine torque,
BSFC, NOx and exhaust temperature, respectively.
2009 Elsevier Ltd. All rights reserved.

1. Introduction air to fuel (CNG) ratio through the air intake; it was constructed in
house.
Compressed natural gas (CNG) is one of the best alternative Testing modied engines under all possible operating condi-
fuels for Internal Combustion engines (ICE). As reported by tions and fuel cases was considered to be both time consuming
[1,4,8,16] and [16], the use of CNG in IC engines will signicantly and expensive. We have investigated the prospects of using arti-
contribute to reduce environmental pollution; it can also reduce cial neural networks (ANN) to alleviate the burden of extensive
the engine noise and vibration, and improve the engine brake ther- experimental testing. The structure and function of the ANN tend
mal efciency [816]. Replacing diesel fuel with natural gas in pas- to simulate a biological neuron network, which receives inputs
senger cars, buses and heavy truck has a number of further from many sources, combines these inputs in some way, performs
benets: (i) there is a signicant reduction in particulate matter a generally non-linear operation on the result, and then outputs
(PM) emissions; (ii) engine wear and maintenance costs are re- the nal result [12]. The network usually consists of three groups
duced; and (iii) NOx emissions can also be reduced [3,6]. However, of layers: an input layer, a number of hidden layers, and an output
CNG composition is highly variable and strongly dependent on the layer [9,12,13]. Each layer has number of neurons. The number of
supply source [7]. neurons in the input and output layers depends on the input and
In the current study, a commercial high speed single cylinder output experimental data, while the features of the hidden layer
diesel engine was modied to operate on natural gas and diesel are generally regarded as adjustable the hidden layers are gener-
as a dual-fuel system. The engine modications included: (i) reduc- ally tuned until the prediction has the minimum mean square er-
ing the original compression ratio from 22.9:1 into 16:1 to avoid ror. All neurons are connected with other neurons in previous
knock phenomena; (ii) installing a CNG injection system which and subsequent layers. The ANN can be demonstrated to be a use-
consisted of a CNG cylinder, a pressure regulator, and a Venturi ful tool if, as a result of training the ANN with certain experimental
mixer; (iii) modifying the existing diesel injection system; and data, the ANN simulations can then be conrm by independent
(iv) undertaking minor modication related to the air manifold data [1,4,5].
and inlet valve [17]. The Venturi mixer was designed using a nite In this study, we investigate the predictive capability of an ANN
volume commercial CFD package, FLUENT to introduce the correct applied to a dual CNG-diesel fuel compression ignition (CI) engine
operated at the maximum load condition for different engine
* Corresponding author. speeds. In our case, the engine performance parameters and ANN
E-mail address: yusaft@usq.edu.au (T.F. Yusaf). output parameters are: brake power, brake thermal efciency gth,

0306-2619/$ - see front matter 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.apenergy.2009.10.009
1662 T.F. Yusaf et al. / Applied Energy 87 (2010) 16611669

brake specic fuel consumption BSFC, exhaust gas temperature, The tests were conducted under maximum operating conditions
and emission components. and this was achieved by changing the engine load on the dyna-
mometer. The engine started with no load at constant speed and
2. Experimental investigation the load was then applied gradually until the maximum load was
achieved. This process was repeated for different speeds. The
Experiments were performed to evaluate the performance of a CNG percentage was increased in each test to replace the diesel
single cylinder four-stroke diesel engines modied to operate fuel.
using dual CNG-diesel fuels. The major specications of the engine
are given in Table 1. The detail of the engine modication and some 2.2. Temperature and pressure sensors
useful calculations related to the compression ratio modication
and other major and minor modications are described in [18]. Six temperature sensors and four pressure sensors provided
Fig. 1 illustrates the engine test apparatus and sensor locations. temperature and pressure measurements at key positions in the
All experiments reported in the present work were performed at apparatus. Sensors were installed to measure the temperature of
the maximum load condition. the engine oil, inlet air, exhaust gas and the water inlet and outlet
for the calorimeter. The pressure sensors were used to monitor the
2.1. Eddy current dynamometer pressure at the air inlet, exhaust gas pipe, and lubricant oil tank.

A 150 kW eddy current dynamometer with end radial thrust 3. Experimental results
bearings and a torque beam was locally designed and fabricated
for loading the engine. The dynamometer was supplied with a Tests were performed with the engine operating on diesel fuel,
drive shaft coupling ange and associated coupling hardware to and also in a dual fuel (CNG-diesel) mode. Measured and derived
connect to the engine. experimental data was obtained for the brake engine power, tor-
que, brake thermal efciency, brake specic fuel consumption,
NOx, exhaust gas temperature, CO, and CO2 of both diesel and
Table 1
Specications of the test engine.
dual-fuel system and is presented as a function of engine speed
in Figs. 2 through to 9. Fig. 10 presents the variation in the fraction
Engine type Y170f vertical 4 stroke diesel engine of CNG with engine speed.
Bore 70 mm
Stroke 55 mm
3.1. Engine performance
Displacement 211 mm
Engine speed 3600 rpm (max)
Max. power 3.5 (kW) Figs. 2 and 3 shows the brake power and the torque results of
Continuous power 3.13 (kW) the engine operated on diesel and dual fuel. From the results of
Compression ratio 16:1 the engine tests, it was found that dual-fuel system produced high-
Cooling system Air cooled
er power than that of the diesel fuel at engine speeds ranging from
Combustion system Direct injection
about 1000 to 2800 rpm. For speeds above 2800 rpm, the power

PETRO
PETROL DIESEL
Rotameter

Torqu Air vel. Power Speed

Piezometri PC
Temp Temp Temp
Control system
11 22 33 4 5 6 7 88 99 10
10
ELCB
Flowmeter
T&P
transferal
Exhaust
To the Engine temp

Calorimeter
outlet temp.
Reg. Valve
Combustion
chamber
Strain
CNG

Dynamometer Gauge Oil sump Exhaust

temp. temp. Calorimeter
Calorimet
Inlet
er Temp.
inlet
temp.
SIEMENS Engine

Direct Coupling

Fig. 1. Single cylinder engine test bed.

 T.F. Yusaf et al. / Applied Energy 87 (2010) 16611669 1663

Fig. 2. Power vs. engine speed for diesel and dual-fuel system (CNG% = 6585%).

Fig. 4. Brake thermal efciency vs. engine speed for diesel and dual-fuel system
(CNG% = 6585%).

Fig. 3. Torque vs. engine speed for diesel and dual-fuel system (CNG% = 6585%).

Fig. 5. Brake specic fuel consumption vs. engine speed for diesel and dual-fuel
system (CNG% = 6585%).
output of the dual-fuel system was lower than the power output of
the diesel system. A contributing factor here may be the relatively arises primarily because of the higher caloric value of the natural
low ame speed of the CNG-air mixtures in comparison with the gas but the more complete combustion in the dual-fuel system
rate of combustion of the diesel [21]. Another possible contributor may also make a contribution.
to reduced power of the dual-fuel system at the higher speeds may
be knock or pre ignition which is possible in the dual-fuel system
because the CNG mixes with the inducted air prior to cylinder 3.2. Exhaust gas emissions test
compression.
From Fig. 4, the brake thermal efciency of the dual-fuel system 3.2.1. Oxides of nitrogen (NOx)
is comparable to, and sometimes higher than the diesel system Fig. 6 shows that the concentration of the NOx for the dual-fuel
with the average increase of around 13% especially at low speeds. system operated at maximum load conditions is lower than that of
At higher speeds, there is no remarkable difference in the thermal the diesel for speeds ranging between 1000 rpm and 3500 rpm.
efciency. On the other hand, the brake specic fuel consumption High temperatures and the availability of oxygen are two main fac-
(BSFC) of the dual-fuel system was lower than that of the diesel tors in the formation of NOx. According to [18], more NOx will be
system, Fig. 5 the dual-fuel system consumes less fuel in compar- formed when sufcient oxygen is available and the average com-
ison to the same diesel system producing same power output. This bustion temperature is high. Exhaust gas temperatures are related
1664 T.F. Yusaf et al. / Applied Energy 87 (2010) 16611669

Fig. 6. NOx vs. engine speed for diesel and dual-fuel system (CNG% = 6585%). Fig. 8. Carbon monoxide vs. engine speed, for diesel and dual-fuel system
(CNG% = 6585%).

During the burning of hydrocarbon such as CH4 a signicant

amount of CO is formed. During power and exhaust stroke most
of the CO oxidizes to CO2. If there is sufcient oxygen available
and high exhaust temperature, CO will continue to convert to
CO2 until relatively small concentrations of CO are left in the
exhaust.

3.2.3. Carbon dioxides (CO2)

The dual-fuel operating conditions produced less CO2 emissions
than the diesel, Fig. 9. On average, the dual-fuel produced about
30% less CO2 emission than the diesel although there was an in-
crease in CO2 emission as the speed increased for both dual fuel
and diesel fuel cases.
If complete combustion actually occurred for both dual and die-
sel fuels, the exhaust composition of CO2 would depend on the
composition of fuel itself. In the dual fuel cases, there was approx-

Fig. 7. Exhaust temperature vs. engine speed, for diesel and dual-fuel system
(CNG% = 6585%).

to the average combustion temperature and Fig. 7 shows that the

exhaust temperature for dual-fuel is much lower than the temper-
ature when diesel fuel is used.

3.2.2. Carbon monoxide (CO)

At the maximum operating condition, the CO concentration
produced by the diesel engine was much higher compared with
the dual-fuel system, Fig. 8. The reduction of the CO in the exhaust
gas means a more complete combustion is achieved; the lower CO
concentration in the exhaust gas, the better and complete combus-
tion is achieved. The air to fuel ratio was very rich for the diesel
operation at higher torque and power, and this resulted in the in-
creased in CO concentration due to the lack of oxygen or incom- Fig. 9. Carbon dioxide emissions vs. engine speed for diesel and dual-fuel system
plete combustion, particularly after 2200 rpm. (CNG% = 6585%).
 T.F. Yusaf et al. / Applied Energy 87 (2010) 16611669 1665

Subsequent network development was undertaken to improve

the performance of the ANN. using The architecture transfer func-
tion of Linear TanhAxon and the learning rule function of Leven-
berg-Marqua were used for the output layer and hidden layers.
The total number of epochs was 1000 and the MSE was 1e6.
The ANN test started with 12 neurons using double hidden layers
TanhAxon function. The results were repeated for higher numbers
of neurons between 13 and 23. It was found that the most accurate
ANN results were found to be failed in 22 neurons where double
hidden layers approach was used. The accuracy of the results
was signicantly improved in comparison to the single hidden
layer conguration, especially for break thermal efciency, CO,
and NOx. The process used in developing the current model was
adopted from [2,19,20].
Three criteria R, RMSE and MRE were selected to evaluate the
networks to nd the optimum solution as follows:
0 1
P 2
B j t j  oj C
R2 1  B
@ P C
A 1
oj 2
j

Fig. 10. Oxygen consumption vs. engine speed for diesel and dual-fuel system
(CNG% = 6585%).
Table 2
Summary of different networks evaluated to yield the criteria of network
imately 20% less carbon in the fuelair mixture than in the diesel performance.
fuel cases, and this explains the majority of the difference between
No. of Activation Training Neurons in Training R
the CO2 emissions in each case. hidden function rule hidden error
layer layer
4. Articial neural network modelling 1 Sig/lin Trainlm 19 1.1000E02 0.989
1 Sig/lin Trainlm 20 1.3000E02 0.987
ANN modelling was used in the present work to predict the 1 Sig/lin Trainlm 21 2.8600E02 0.9714
1 Sig/lin Trainlm 22 2.7600E03 0.99724
relationship of brake power, torque, brake specic fuel consump- 2 Sig/lin Trainlm 22 1.9184E02 0.980816
tion, brake thermal efciency and emission components with the 1 Sig/axon Trainlm 22 9.0100E02 0.9099
engine speed and percentage CNG as inputs. Approximately 70% 1 Soft/lin Traingdx 22 1.2954E01 0.87046
of the total experimental data (220 values) was selected at random 1 Sig/axon Traingdx 22 3.1877E02 0.968123
1 Tanax/lin Traingdx 22 3.1238E02 0.968762
and was used for training purpose, while the 30% was reserved for
1 Sig/lin Trainlm 23 1.6640E02 0.98336
testing. The experimental data set for every output parameter in- 1 Sig/lin Trainlm 24 4.6170E02 0.95383
cludes 20 values, of which 14 values were used for training the net- 1 Sig/lin Trainlm 25 2.5300E02 0.9747
work and six values were selected randomly to test the 1 Sig/axon Trainlm 19 8.4522E02 0.915478
performance of the trained network.
Simulations were performed using MATLAB. A multi-layer per-
ception network (MLP) was used for non-linear mapping between
the input and the output variable. To improve the modeling, sev-
eral architectures were evaluated and trained using the experi-
mental data. The back-propagation algorithm was utilized in
training of all ANN models. This algorithm uses the supervised
training technique where the network weights and biases are ini-
tialized randomly at the beginning of the training phase. The error
minimization process is achieved using gradient descent rule.
There were two input and nine output parameters in the exper-
imental test. The two input variables were engine speed in rpm
and the compressed natural gasdiesel ratio. The nine outputs
for evaluating engine performance are indicated in Fig. 10. The in-
put layer consisted of two neurons and the output layer had nine
neurons. The number of hidden layers and neurons within each
layer can be tuned to suit the complexity of the problem and data
set. In this study, the initial network that was selected consisted of
one hidden layer with a set range of neurons. The activation func-
tion for the hidden layer was selected to be linear function. A linear
function was usually best suited for output layer. However, the re-
sults of this ANN model with one hidden layer were not accurate
enough. It was found that NOx, CO and brake thermal efciency
modeled ANN results were not accurate as the indicated R values
were below 90%. Therefore several other networks with various
functions and topologies were examined. Fig. 11. Experimental and predicted powers vs. engine speed.
1666 T.F. Yusaf et al. / Applied Energy 87 (2010) 16611669

Fig. 12. CNG percentages vs. engine speed at maximum loads for diesel and dual-
fuel system.
Fig. 14. The ANN power prediction vs. the experimental power.
Pn !
j t j  oj
RMSE 2 An independent ANN test was conducted for a range specic of
n
   RPMs for different parameter to establish condence in the ANN
n 
1 X  t j  pj  model. Initial tests were conducted to predict the relationship be-
MRE 100   3 tween the speed and the brake power for different speed values
n 1  tj 
[19]. Fig. 11 shows the relationship between the actual and pre-
Three standard criteria R, RMSE and MRE were selected to eval- dicted power for different speed. As seen in Fig. 11, the predicted
uate the various networks (R is the error, RMSE is the root mean results are nicely matched to the experimental ones. This result
square error, and MRE is the mean relative error). A regression will give an idea to the reader of how far the suggested ANN can
analysis between the network response and corresponding targets go for a reliable prediction.
was performed to investigate the network response in more detail.
Different training algorithms were tested and Levenberg-Marquadt
was selected. R values in Table 2 represent the error coefcient be- 5. Results and discussion
tween the outputs and As. R values did not increase when the num-
ber of neurons in the hidden layer was more than 22. Therefore, a In general, using CNG in IC engines will lead to a leaner and
network with one hidden layer and 22 neurons was selected as the cleaner combustion. It was experimentally demonstrated that
preferred ANN. using CNG led to an increase in the engine brake power, torque

Fig. 13. The structure of ANN for dual fuel (CNG_diesel) and diesel fuel engine.
 T.F. Yusaf et al. / Applied Energy 87 (2010) 16611669 1667

and brake thermal efciency and decrease the brake specic fuel shows that the increase the engine speed is associated with the in-
consumption. The lean combustion improves the completeness of crease of CNG percentage.
combustion and therefore the CO emission was expected to be de- An ANN was developed based on this experimental work to pre-
creased. The experimental results conrmed that by using CNG, the dict the missed data and avoid spending excessive time running
CO was decreased. The oxygen enrichment generated from CNG in- experimental tests (Fig. 13). The results showed that the training
creased the oxygen ratio in the ratio in the charge and led to lean algorithm of Back Propagation was sufcient for predicting engine
combustion. CO2 emission varies with the A/F ratio and CO concen- torque, brake power, brake thermal efciency, specic fuel con-
tration. As a result, CO2 emission increased because of the im- sumption and exhaust gas components for different engine speeds
proved combustion. When the combustion process is closer to and different type of fuel.
stoichiometric, ame temperature increase, therefore, the NOx The predicted vs. experimental values for experimental param-
emission increased. eters are indicated in Figs. 1418. The ANN predictions for the
Fig. 12 illustrates the relationship between the speed and the brake power, engine torque, brake thermal efciency, brake spe-
CNG percentage for maximum operating conditions. This gure cic fuel consumption and exhaust temperature yield a correlation
coefcient (R) of 0.9808, 0.9884, 0.92897, 0.9838, and 0.9934
respectively. It was found the root mean square error (RMSE)

Fig. 15. The ANN torque prediction vs. the experimental torque.

Fig. 17. The ANN prediction for the BSFC vs. experimental data.

Fig. 16. The ANN prediction for the brake thermal efciency (lth) vs. experimental
date. Fig. 18. The ANN prediction for the exhaust temperature vs. experimental values.
1668 T.F. Yusaf et al. / Applied Energy 87 (2010) 16611669

values were 0.1469 kW, 1.8803 N m, 1.2103%, 4.06 kg/kW h and

8.265 C for the brake power, engine torque, brake thermal ef-
ciency, brake specic fuel consumption and exhaust temperature
respectively.
The ANN predictions for the performance of the tested IC engine
yielded a good statistical performance. Comparisons of the experi-
mental results and the ANN predictions demonstrate that IC engine
using CNG-diesel fuel can be accurately simulated using ANN. The
predicted vs. experimental values for the CO emission are indicated
in Fig. 19. The ANN predictions for the CO yield a correlation coef-
cient (R) of 0.9359, root mean square error (RMSE) of 10.26 ppm.
Fig. 20 indicates the predicted vs. experimental values of the CO2
emission with error correlation coefcient (R) of 0.9964 and root
mean square error (RMSE) of 0.3422. The predicted vs. experimen-
tal values for the NOx emission are indicated in Fig. 21. The ANN

Fig. 21. The ANN prediction for the NOx vs. experimental values.

Fig. 19. The ANN prediction for the CO vs. experimental values.

Fig. 22. The ANN predictions for the O2 vs. experimental values.

predictions for the NOx yield correlation coefcient (R) of

0.95707 and root mean square error (RMSE) of 7.0099. Fig. 22
shows the relationship between the predicted vs. experimental
values of O2 emission. The ANN predictions for the O2 yield a cor-
relation coefcient (R) of 0.9705 and root mean square error
(RMSE) of 0.62186.

6. Conclusion

Operating a diesel engine on dual fuel CNG-diesel, gives better

brake thermal efciency and lower emissions. It was found that
the training algorithm of back propagation is sufcient for predict-
ing engine torque, specic fuel consumption and exhaust gas com-
ponent for different engine speeds. Analysis of the experimental
Fig. 20. The ANN predictions for the CO2 vs. experimental values. data by the ANN revealed that there is good correlation between
 T.F. Yusaf et al. / Applied Energy 87 (2010) 16611669 1669

the simulated results from the ANN and the measured data, with R [9] Yang KT. Articial neural networks (ANNs): a new paradigm for thermal
science and engineering. J Heat Transfer 2008;130(September):093001 119.
values are ranging between 0.92 and 0.99. Therefore the ANN is
[10] Lu PJ, Zhang MC, Hsu TC, Zhang J. An evaluation of engine faults Diagnostics
demonstrated to be a useful simulation tool in the evaluation of using articial neural networks. J Eng Gas Turbines Power
the engine parameters. Generally speaking, ANNs can provide 2001;123(April):3406.
accuracy and simplicity in the analysis and simulation of the en- [11] Saynic C, Ertunc HM, Hosoz M, Kilicaslan I, Canakci M. Performance and
exhaust emissions of a gasoline engine using articial neural network. Appl
gine performance. Therm Eng 2007;27:4654.
[12] Haykin, S. Neural networks: a comprehensive foundation. New York: Mac-
References millan; 1994.
[13] Chouai A, Laugier S, Richon D. Modelling of thermodynamic properties using
neural networks. Appl Refrig Fluid Phase Equilibr 2002;49:110.
[1] Ghobadian B, Rahimi H, Nikbakht AM, Naja G, Yusaf FT. Diesel engine
[14] Mccann DW. Forecasting techniques a neural network short-term forecast of
performance and exhaust emission analysis using waste cooking biodiesel fuel
signicant thunderstorms. J Weather Forecast 1992;7(September):52534.
with an articial neural network. Renew Energy 2009;34(4):97682.
[15] Ou S, Achenie LEK. Articial neural network modelling of PEM fuel cells. J Fuel
[2] Naja G, Ghobadian B, Tavakoli T, Buttsworth DR, Yusaf TF. Performance and
Cell Sci Technol 2005;2(November):22633.
exhaust emission of a gasoline engine with ethanol blended gasoline fuels
[16] Murata H. Example of introducing compressed natural gas vehicle, In: 7th
using articial neural network. Appl Energy 2009;86:6309.
international conference and exhibition on natural gas vehicles. 1719
[3] Cascetta F, Rotondo G, Musto M. Measuring of compressed natural gas in
October, Yokohama, Japan; 2000.
automotive application: a comparative analysis of mass versus volumetric
[17] Jonsson, O. Introduction of the rst heavy NGV trucks in Sweden. In: 7th
metering methods. Flow Meas Instrum 2008;19:33841.
international conference and exhibition on natural gas vehicles, 1719
[4] Yusaf FT, Shamsuddin A, Ali Y, Ismail AF. Design modication of high speed
October, Yokohama, Japan; 2000.
diesel engine to accommodate compressed natural gas. RERIC Int Energy J
[18] Yusaf T. Experimental investigation of a single cylinder diesel engine as a
1996;18(1):1926.
hybrid power unit for a series hybrid electric vehicle. J Energy Res Technol
[5] Hertz J, Krogh A, and Palmer RG. Introduction to the theory of neural
ASME 2009;131:122037.
computation. Redwood City, NJ: Addison Wesley publishing company, 1991.
[19] Mustafa Glc, Yakup Sekmen, Perihan Erduranli, Sahir Salman M. Articial
[6] Fritz SG, Egbuonu RI. Emission from heavy duty trucks converted to CNG. J Eng
neural-network based modelling of variable valve-timing in a spark-ignition
Gas Turbine Power 1992;114:5617.
engine. Appl Energy 81(2):18797.
[7] Gandhidasan P, Ertas A, Anderson EE. Review of methanol and compressed
[20] Mustafa Glc. Articial neural network based modeling of performance
natural gas (CNG) as alternative for transportation fuels. J Energy Res Technol
characteristics of deep well pumps with splitter blade. Energy Convers Manage
1991;113:1017.
2006;47(1819):333343.
[8] DAmberosio S, Spessa E, Vassallo A. Methods for specic emission evaluation
[21] Duan SY. Using natural gas in engines: laboratory experience with the use of
in spark ignition engines based on calculation procedures of airfuel ratio:
natural gas fuel in IC engines. IMechE Seminar Publication; 1996. p. 3946.
development, assessment, and critical comparison. J Eng Gas Turbines Power
2005;127(October):86981.