You are on page 1of 9

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 5 ( 2 0 1 0 ) 1 0 0 0 4 e1 0 0 1 2

Available at www.sciencedirect.com

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

Basic tuning of hydrogen powered car and artificial intelligent


prediction of hydrogen engine characteristics

Tien Ho a,*, Vishy Karri b


a
School of Engineering, University of Tasmania, GPO Box 252-65, Hobart, Tasmania, 7001, Australia
b
Australian College of Kuwait, P.O. Box 1411, Safat 13015, Kuwait

article info abstract

Article history: Many studies of renewable energy have shown hydrogen is one of the major green energy
Received 19 June 2009 in the future. This has lead to the development of many automotive application of using
Received in revised form hydrogen as a fuel especially in internal combustion engine. Nonetheless, there has been
21 January 2010 a slow growth and less knowledge details in building up the prototype and control meth-
Accepted 12 March 2010 odology of the hydrogen internal combustion engine [1]. In this paper, The Toyota Corolla 4
Available online 21 April 2010 cylinder, 1.8l engine running on petrol was systematically modified in such a way that it
could be operated on either gasoline or hydrogen at the choice of the driver. Within the
Keywords: scope of this project, several ancillary instruments such as a new inlet manifold, hydrogen
Hydrogen powered car fuel injection, storage system and leak detection safety system were implemented.
Engine characteristics prediction Attention is directed towards special characteristics related to the basic tuning of hydrogen
Artificial neural networks engine such as: air to fuel ratio operating conditions, ignition timing and injection timing in
Tuning of hydrogen engine terms of different engine speed and throttle position. Based on the experimental data,
a suite of neural network models were tested to accurately predict the effect of different
engine operating conditions (speed and throttle position) on the hydrogen powered car
engine characteristics. Predictions were found to be 3% to the experimental values for all
of case studies. This work provided better understanding of the effect of hydrogen engine
characteristic parameters on different engine operating conditions.
Crown Copyright ª 2010 Published by Elsevier Ltd on behalf of Professor T. Nejat Veziroglu. All
rights reserved.

1. Introduction The converted hydrogen powered car operating parameters


were measured regarding to engine speed, throttle position,
The Hydrogen & Allied Renewable Technology (HART) research ignition timing, injection timing, air to fuel ratio and manifold
program (University of Tasmania) has been built up the prototype air pressure. These various useful engine parameters were
four cylinders manual Toyota Corolla to use hydrogen as a fuel in measured for a number of engine operating conditions through
its internal combustion engine. There are several constraints to an Advanced Dash Logger (ADL) supplied by Motec. The engine
be taken into consideration for building appropriate engine speed was measured by using the standard crank angle sensor,
tuning maps to run on hydrogen fuel such as: elimination of cam angle sensor, as supplied on the engine in its original form.
knock, pre-ignition, self-ignition and backfire; smooth operation The throttle position was measured by using the standard
of the engine with desired power output requirements while throttle position sensor, as supplied on the engine by its
minimize the emission and fuel consumption [1] Table 1 shows manufacturer. The air to fuel ratio (lambda value) was
the specifications of the converted hydrogen car. measured using a Bosch wide-band lambda sensor, the output

* Corresponding author.
E-mail addresses: ntho@utas.edu.au (T. Ho), v.karri@ack.edu.kw (V. Karri).
0360-3199/$ e see front matter Crown Copyright ª 2010 Published by Elsevier Ltd on behalf of Professor T. Nejat Veziroglu. All rights reserved.
doi:10.1016/j.ijhydene.2010.03.054
i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 5 ( 2 0 1 0 ) 1 0 0 0 4 e1 0 0 1 2 10005

from which was then wired into the Motec Engine Control Unit
Table 2 e Engine parameters and appropriate source of
(ECU) M400 as well as was connected to the Motec ADL. The measurement sensors [1,3,6].
manifold air pressure was measured by an aftermarket Bosch
Parameter Designation Sensor/Source
Manifold Air Pressure (MAP) sensor. The output from this was
directly wired into the Motec ECU. The ignition timing or igni- 1. Engine RPM RPM Stock ECU
tion advance was the value of angle (degree before top death 2. Throttle Position TP Stock ECU
3. Mass Air Flow Amass MAF Sensor via stock ECU
centre, oBTDC) to start to ignite the hydrogen fuel while the
4. Manifold MAP MAP Sensor via stock ECU
injection timing which was included end of injection angle and
Air Pressure
injection duration (hydrogen fuel injector pulse width). These 5. Fuel Actual Pulse FAPW Motec ECU lookup table
parameters can be utilised and controlled by the Motec ECU width
and were stored by the ADL for later analysis. The hydrogen 6. Ignition advance ladv Motec ECU lookup table
pressure within the fuel rail was measured using a pressure 7. Exhaust Gas EGT1 EOT sensor via Motec ECU
gauge. The flow rate of hydrogen was calculated using the Temperature
8. Lambda La1 Lambda sensor via PLM
pressure differential across the hydrogen fuel injectors
9. Intake Air AT Stock ECU
(between the hydrogen fuel rail pressure and the inlet manifold
Temperature
pressure) and the injector pulse width time [1,3] Table 2 shows 10. Engine ET Stock ECU
the engine parameters and appropriate source of measure- Temperature
ment sensors. 11. Output Power PWR Dynamometer
Engine parameters were logged at a rate of 2 Hz through
a MoTeC ADL Unit. Results were found to be adequate at that
rate and increasing the logging rate did not further improve following procedures at different desired engine’s speed and
accuracy of recorded data. The ADL has a total of 1 MB of load points as shown below:
memory with logging time being dependant on the number of
parameters being logged as well as logging frequency. The 1. Firstly, the engine’s fuel requirements were adjusted by
vehicle was then placed on a dynamometer while a 5-way gas tuning injection pulse width in order to bring them into line
analyser was installed on the exhaust port. The gas analyser with the desired air to fuel ratio values.
was a commercial unit from OTX-SPC. The set up for the gas 2. Secondly, the engine’s ignition timing was adjusted in
analyser is similar to that for a normal vehicle with the order to achieve minimum ignition advance for maximum
exception of additional moisture traps required due to the high brake torque (MBT).
water vapour content in the exhaust stream [1e3]. Data logging 3. Thirdly, the engine’s injection timing (injection start angle)
was set up for the device with a frequency of 2 Hz to match the was tuned and the procedure as in step 2. However, this
input frequency. The catalytic converter was removed from the control parameter only needs to be tuned for engine speed
exhaust system of the vehicle in order to measure the true site.
(uncontrolled) emissions data from the vehicle. 4. Finally, the ignition timing was checked in order to ensure
that no changes were required.

2. Brief description of experimental Note that the methodology proposed here could be a generic
methodology and neural network model application for all hydrogen internal combustion engines
involving basic tuning control parameters utilise electronic fuel
The basic tuning of hydrogen powered car has been conducted injected vehicle.
within the investigation purposes of this paper with the With the complexity of the configuration and tuning tech-
niques of an aftermarket ECU for a new fuel such as hydrogen,
the work to make the engine to start properly is a very difficult
Table 1 e Specifications of the conversion vehicle [1,2]. task. Meanwhile in the basic terms, the providing of the correct
Manufacturer Toyota amount of hydrogen at the right time with appropriate ignition
timing needs all of ECU functions to work correctly. Idle
Model Corolla
condition was seemed to be the most difficult task for tuning
Series Ascent
Body Type Hatchback for hydrogen engine which requires load to be placed on.
Year of Manufacture 2002 Moreover, the mixture of air/fuel ratio or fuel quality needs to
Type Inline, 4 Cylinders, DOHC, VVTi be measured to keep the mixture is neither too rich nor too lean
Total Displacement 1794 cm3 which are the main reason of exhaust gas excessive heat or
Compression Ratio 10.0:1 severe engine damage and increasing NOx emission gases [4].
Fuel Type Unlead Petrol RON 91 or Higher
A best tuning engine will show the best performance for any
Maximum Power Output 100 kw @ 6000 RPM
Maximum Torque 171 Nm @ 4200 RPM
given conditions including rich mixture at high power, lean or
Length 4385 mm stoichiometric when idle or slow speed. Beside that, hydrogen
Width 1695 mm injection systems should be tuned correctly to precisely control
Height 1475 mm the mixture air/fuel ratio to suit different operating conditions.
Wheelbase 2600 mm Therefore, this has online instrumentation to measure the
Driven wheels Front wheel drive
input parameters were all installed for subsequent data
Transmission 5 speed manual
acquisition and modelling.
10006 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 5 ( 2 0 1 0 ) 1 0 0 0 4 e1 0 0 1 2

Other than simple full throttle ramp-up power testing runs,


all testing and data recorded were done under steady-state
conditions, according to the requirements of ISO 15550:2002
[5], including the requirement where the engine must be tested
at its proper operating temperature at the required operating
speed corresponding to desired throttle position Fig. 1 shows
the hydrogen car on dynamometer testing.
The artificial neural networks (ANN) as predicting model
was chosen for this research because of the following reasons
[7e15]: their strength of mathematical foundation and exten-
sive prediction application, the ability to model non-linear
Fig. 2 e Map of three layer back-propagation neural
process, adaptive learning, self-organization, real time opera-
network as a virtual sensor for prediction of hydrogen car
tion, and ease of insertion into existing technology.
engine characteristics.
Back-Propagation Neural Network (BPNN) is a systematic
method for training multi-layer Artificial Neural Networks
(ANNs). The learning process in a multi-layer network proceeds 2.2. Step 2: Activation
the same way as for a perceptron. The network will be provided
with a training set of input patterns to the input layer. Then the Activate the back-propagation neural network by applying
network propagates the input pattern from layer to layer until inputs x1( p), x2( p),., xn( p) and desired outputs yd,1( p),
the output pattern is computed and generated by output layer. yd,2( p),., yd,n( p).
If there is an error between actual and desired output pattern, (a) Calculate the actual outputs of the neurons in the
the weights are adjusted to reduce this error [8]. Once the error hidden layer:
" #
is calculated, it will be propagated backwards through the X n

network from the output layer to the input layer. The weights yj ðpÞ ¼ sigmoid xi ðpÞwij ðpÞ  qj
i¼1
are modified using delta rule as the error is propagated to
improve the network performance. Fig. 2 depicts the map of Where n is the number of inputs of neuron j in the hidden
three layer back-propagation neural network as a virtual layer and sigmoid is the sigmoid activation function.
sensor for prediction of hydrogen car engine characteristics. (b) Calculate the actual outputs of the neurons in the
The general training algorithm of the back-propagation output layer:
2 3
neural networks can be defined as following [8]. Xm
yk ðpÞ ¼ sigmoid4 xjk ðpÞwjk ðpÞ  qk 5
j¼1
2.1. Step 1: Initialisation
Where m is the number of inputs of neuron k in the output
Set all the weights and threshold levels of the network to layer.
random numbers uniformly distributed inside a small range:

2.3. Step 3: weight training


 
2:4 2:4
 ;þ
Fi Fi Update the weights in the back-propagation network propa-
gating backward the errors associated with output neurons.
Where Fi is the total number of inputs of neuron i in the This step is done by following procedures:
network. The weight initialisation is done on a neuron-by-
neuron basis. (a) Calculate the error gradient for the neurons in the output
layer:
Calculate the weight corrections
Update the weights at the output neurons.
(b) Calculate the error gradient for the neurons in the hidden
layer:
Calculate the weight corrections
Update the weights at the hidden neurons.

Table 3 e Basic tuning lambda table.


Lambda RPM

Throttle Position 1500 2000 3000 4000


100% 1.41 1.55 1.49 1.47
75% 1.56 1.62 1.58 1.64
50% 1.75 1.75 1.76 2.02
25% 1.86 1.71 1.76 2.06
Fig. 1 e Hydrogen car on Dynamometer Testing [3,7].
i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 5 ( 2 0 1 0 ) 1 0 0 0 4 e1 0 0 1 2 10007

As a result, neural networks have proven themselves in


Table 4 e Injection pulse width.
practice for accurate performance prediction. The project was
Injector Pulse Width (ms) RPM developed with eleven specific algorithms which are available
Throttle Position 1500 2000 3000 4000 from MATLAB back-propagation package as listed below:
100% 9.68 9.79 10.26 12.35
75% 9.01 9.25 9.75 10.76  LevenbergeMarquardt;
50% 8.44 8.60 8.76 8.65  Gradient descent;
25% 6.89 6.88 4.97 3.38
 Gradient descent with momentum;
 Gradient descent with adaptive learning rate;
Note that the weight correction Dwi( p) at iteration p is  Gradient descent with momentum and adaptive learning
calculated by delta rule as below: rate;
Dwi ðpÞ ¼ a$xi ðpÞ$eðpÞ  Resilient back-propagation;
 Scaled conjugate gradient;
Where  Conjugate gradient with Fletcher-Reeves updates;
 Conjugate gradient with Polak-Ribiere updates;
a: the learning rate, a positive constant less than unity.  Conjugate gradient with PowelleBeale restarts;
e( p): the error at iteration p and it equal to the actual output is  One step secant.
Y( p) minus the desired output is Yd ( p).
xi( p): input xi at iteration p Through extensive experimentation covering a compre-
hensive range of data performances, the suite back-
And update the weight is calculated by: propagation neural networks with LevenbergeMarquardt
algorithm were proven them-self is the most accuracy algo-
wi ðp þ 1Þ ¼ wi ðpÞ þ Dwi ðpÞ rithm when performing the project prediction tasks. The
developed prediction neural network models were accurately
2.4. Step 4: iteration predict hydrogen car engine ‘s characterists in almost cases
with average percentage root mean square error less than 3%.
Increase iteration p by one, go back to Step 2 and repeat the The prediction includes mass air flow, manifold air pressure,
process until the selected error criterion is satisfied. fuel actual pulse width, engine temperature, exhaust gas

Fig. 3 e Torque versus speed and load at desired lambda.


10008 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 5 ( 2 0 1 0 ) 1 0 0 0 4 e1 0 0 1 2

Fig. 4 e Ignition advance versus engine speed and load.

temperature and NOx emissions as a function of various engine ISO 15550:2002 [5]. This ensured that the obtained results
speed and load. would suitable for subsequent scientific analysis and testing.
The effect of hydrogen as a fuel compare with gasoline engine
3. Results and discussion parameters (such as: brake power, brake torque, brake mean
effective pressure, brake specific fuel consumption, brake
The engine testing procedures outlined were in compliance thermal efficiency) and the effect of hydrogen as a fuel
with the requirements of the relevant International Standard, compare with gasoline on emission characteristics have been

Injecting Timing
450.0

400.0

350.0
End of injection (Degree BTDC)

300.0

250.0

200.0

150.0

100.0

50.0

0.0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Engine speed (RPM)

°BTDC

Fig. 5 e End of injection angle.


i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 5 ( 2 0 1 0 ) 1 0 0 0 4 e1 0 0 1 2 10009

published in a separate paper [1]. Within this paper, attention


Table 5 e Back-propagation neural networks
is directed towards special characteristics related to the basic architecture.
tuning of hydrogen engine such as: air to fuel ratio operating
Types of networks Three layer
conditions, ignition timing and injection timing in terms of
feed forward
different engine speed and throttle position. The discussion back-propagation
on the obtained results has been presented as below:
Hidden layer 10 Neurons
Output layer 1 Neuron
1. Specific properties of hydrogen internal combustion engine
Transfer function ‘tansig’, ‘purelin’
car. Training algorithm LevenbergeMarquardt
2. Artificial neural networks as an intelligent approach to Weight/bias learning function ‘learngdm’
predict hydrogen engine’s characteristics. Performance function ‘mse’
Number of epochs between 100
showing the progress
Maximum number epochs to train 3000
3.1. Specific properties of hydrogen internal combustion
Performance goal 0.0001
engine car Learning rate 0.1
Learning rate increase multiplier 1.05
Typically, the torque generated by hydrogen engine was found Learning rate decrease multiplier 0.8
to be in the ranges between 50% and 65% of gasoline engine for Momentum constant 0.95
the same engine operating conditions. The research results with
basic tuning have shown that hydrogen fuelled engine has the desired engine speed and load and data could be recorded
a peak torque of 100 N m at engine speed 4000 RPM and lambda at the stable engine operating condition corresponding to the
value around 1.47, with a secondary peak of 94 N m at 2000 RPM desired lambda value. Table 3 shows the lambda values cor-
corresponding with lambda value around 1.55. The brake responding to different hydrogen engine operating conditions.
specific fuel consumption for the hydrogen fuelled engine These results have been found to be matched with the
ranges between 14.4 and 15.2 g/kwh. At wide open throttle, expected results from Verhelst and Sierens [16,17]. However,
hydrogen engine is more efficient than gasoline engine due to this was the richest mixture formation at the maximum
lean air to fuel ratio while at other tested operating points was desired engine speed and throttle position that had been
lower than that of the gasoline engine. The maximum brake tested due to the requirement of knock limitation, pre-ignition,
mean effective pressure of hydrogen engine was found to be self-ignition, backfire and constraints of fuel economy, NOx
55.8 MPa at 4000 RPM and wide open throttle while it was emissions and excessive hot engine and exhaust gas temper-
93.8 MPa for gasoline-fuelled engine at the same operating ature. Therefore, the true maximum torque may be achieved
conditions. These hydrogen engine performance results are by using a complete research design of experiment engine
corresponding to the desired engine operation with slightly lean calibration to consider a richer air to fuel ratio mixture. Fig. 3
compare to stoichiometric air to fuel ratios (lambda ¼ 1). shows the hydrogen engine torque versus speed and load at
However, the lean operation of hydrogen fuelled engine has desired lambda.
resulted in extremely low emission of NOx gases and virtually
negligible of other emission gases such as carbon dioxide, 3.1.2. Injection pulse width
carbon monoxide and hydrocarbon. The emission of NOx The injection duration or injection pulse width (IJPU) is the
increases markedly as the lambda value decreases toward main ECU parameters to control how much hydrogen would be
unity, and has a minimum at a lambda value of around 1.87. The injected into the manifold inlet runners. Therefore, it is the
emission of carbon dioxide from the hydrogen engine is virtu- principle control of the air/fuel ratio and the exhaust lambda
ally negligible, being between 0.05% and 0.29%, compared with measured value and should be iteratively and carefully
between 14.44% and 14.58% from the gasoline-fuelled engine. adjusted so that the desired lambda value can be reached at
While the emission of carbon monoxide from hydrogen engine individual desired load points, since any “significant” too lean
was extremely low, at 0.005e0.020%, compared to 0.326e0.767% mixture in high load conditions could damage the engine
for the gasoline engine, the emission of hydrocarbon from the quickly. The MoTeC “Quick Lambda” function tuning software
hydrogen engine was lower than those of gasoline engine, at was used to help the tuner to correct the difference of the desire
20e37 ppm, compared to 50e152 ppm for the gasoline engine [1]. and current measured lambda values. The basic tuning average
result for injection pulse width is shown in the Table 4 below.
3.1.1. Lambda table
The lambda value is defined as the ratio between air to fuel
ratio and stoichiometric air to fuel ratio operating condition.
Table 6 e Prediction results for hydrogen engine
The tuning process of an electronic fuel injected in terms of
charateristics.
different engine speed and load can take up to over 100 indi-
Averages AMass MAP FAPW EGT 1 ET NOx
vidual points which is time consuming. However, it was
decided that the basic tuning of the hydrogen car engine Error (value) 0.0096 0.0749 0.0011 0.5014 0.0059 2.8157
should be examined the normal operating condition at around STD (value) 0.0852 0.5478 0.0376 2.5441 0.4492 15.865
desired engine speed of 1500, 2000, 3000, and 4000 RPM and the Error (%) 0.046 0.0695 0.0285 0.2401 0.0058 0.3207
RMS E (value) 0.0852 0.5478 0.0376 2.5441 0.4492 15.865
desired throttle at around 25%, 50%, 75% and 100% wide open
RMS E (%) 0.4106 0.6243 0.4313 1.338 0.5181 2.736
throttle. The engine’s load would be gradually adjusted so that
10010 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 5 ( 2 0 1 0 ) 1 0 0 0 4 e1 0 0 1 2

30 100

28 95

90
26

M anifold A ir P res s ure (K P a)


85
M as s A ir F low (m g)

24
80

22 75

20 70

65
18
60
16
Actual 55 Actual
Predicted Predicted
14 50
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Number of Testing data Number of Testing data

11 260
Actual Actual
Predicted Predicted
10 240
E x haus t gas tem perature (D eg. C)
F uel A c tual P uls e W idth (m s )

220
9

200
8

180

7
160

6
140

5 120
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Number of Testing data Number of Testing data

90 2000
Actual
Actual
1800 Predicted
Predicted
89
1600
E ngine tem perature (D eg. C )

1400
88
1200
N O x (ppm )

87 1000

800
86
600

400
85
200

84 0
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Number of Testing data Number of Testing data

Fig. 6 e Comparison of actual and prediction performance results for hydrogen powered car engine’s characteristics.
i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 5 ( 2 0 1 0 ) 1 0 0 0 4 e1 0 0 1 2 10011

3.1.3. Ignition timing 3.2.1. Mass air flow prediction (AMass)


The ignition timing was set to that which was required It can be seen that very good prediction results were
minimum ignition advance for maximum brake torque (MBT) obtained for mass air flow prediction with the percentage
at the various engine speeds and loads. The process to tune up average root mean error (%ARMS) was 0.4106 and average
the ignition advance was firstly hold the engine at a given deviation was 0.0852.
dynamometer load at the desired engine speed and load, and
the ignition timing was advanced from 5 to 50 before top 3.2.2. Manifold air pressure (MAP)
death center (BTDC). As the ignition timing was advanced, the The prediction result was obtained for prediction of manifold
torque generated would be increased result in gradually air pressure with %ARMS error was 0.6243% and deviation was
increase in engine speed. Therefore, the extra load at the 0.5478.
desired speed and load. There was a point where the peak
torque was achieved and further increase in spark timing 3.2.3. Fuel actual pulse width (FAPW)
produced either no change or decrease in engine speed [1e3]. The prediction result was obtained for prediction of FAPW
If knock occurs at any stage, the ignition advance should be with %ARMS error was 0.4313% and deviation was 0.0376.
set at the nearest point where there is no knock. The best
ignition advance will control the best efficiency of power 3.2.4. Exhaust gas temperature (EGT1)
extraction per stroke. Fig. 4 depicts ignition advance versus The prediction result was obtained for prediction of EGT1 with
engine speed and load. %ARMS error was 1.338% and deviation was 2.5441.

3.1.4. Injection timing 3.2.5. Engine temperature (ET)


The basic tuning of hydrogen internal combustion engine car The prediction result was obtained for prediction of engine
was also carried out the tuning of injection timing so that temperature with %ARMS error was 0.5181% and deviation
best fuel efficiency at specific desired air to fuel ratios can be was 0.4492.
achieved. The injection timing setting was set so that
hydrogen can be started to inject a few degrees after the 3.2.6. Emission of NOx
exhaust valve had closed to avoid the backfire problem. The The prediction result was obtained for prediction NOx emis-
injection timing was altered so that the minimum time sion with %ARMS error was 2.736% and deviation was 15.865.
required after the closing of the exhaust valve can be ach- The Table 6 and Fig. 6 below summarises the prediction
ieved without abnormal engine operating. This basic desire performance of each hydrogen powered car engine‘s
tuning methodology would allow more time for hydrogen to characteristics.
enter the combustion chamber at high load conditions while
also minimising the chance of igniting hydrogen due to
remaining hot gases in combustion chamber [2]. The default
setting of Motec ECU for injection timing is “end of injection” 4. Conclusion
which is the degrees BTDC that injection must be finished by.
Note that this control parameter only needs to be tuned for From the basic tuning of hydrogen internal combustion
engine speed site. The tuning result of injection timing is engine, the main engine control parameters such as: lambda,
shown in Fig. 5. ignition timing, injection pulse width and injection timing
were presented. In addition, the excellent prediction results of
hydrogen car engine‘s characteristics were achieved in almost
3.2. Neural networks for prediction of hydrogen engine cases with average percentage root mean square error less
characteristics than 3%. However, further study on the specific features of
the use of hydrogen in internal combustion engine car as well
The neural networks created were included 2 inputs which are as its best tuning intelligent techniques should be conducted
engine speed (rpm) and load (% throttle position) and 6 output by thoroughly investigate and modelling these basic tuning
parameters (as shown in Fig. 2 above), which is cover data in the statistical engine analysis tool so that the optimi-
a comprehensive range of data variation through many sation and trade-off study between power requirement, fuel
different testing engine conditions. The back-propagation consumption and emissions can be applied. Different opti-
neural networks are built up within MatLab neural network misation case study can be done included: intake cam and
toolbox [18], which is used to appraise the predictive models. exhaust cam phase, spark timing schedules, best injection
In each of the studied prediction hydrogen powered car timing, best fuel quality, best fuel pressure as a function of
parameters, the large number of data (1400 data) was used for load and speed.
a training set and 50 data were used for a testing set. The best
back-propagation neural networks architecture is shown in
Table 5 below.
In order to provide a measure of accuracy of the predic- Acknowledgements
tions, several parameters are used such as: average error
(value), standard deviation (value), average error (%), average The authors are deeply grateful to Dr. Sergio Giudicci, Hydro
root mean square error (Value), average root mean square Tasmania Pty. Ltd. for financial support and all of the Centre
error (%). for Renewable Energy and Power Systems research members
10012 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 5 ( 2 0 1 0 ) 1 0 0 0 4 e1 0 0 1 2

as well as Intelligent Hydrogen Car project for sharing ideas [3] Lim, J., Development of a hydrogen car and emissions
and concept along the way. modelling using artificial intelligence tools, In School of
Engineering, University of Tasmania, Hobart.
[4] Butler DA. In: Enhancing automotive stability control with
artificial neural networks. Hobart: School of engineering,
Nomenclature University of Tasmania; 2006.
[5] ISO 1550:2002(E). Internal combustion engines e
determination and method for the measurement of engine
UTAS University of Tasmania, Australia power e general requirements, I.S. Organisation, Editor; 2002.
HART Hydrogen & Allied Renewable Technology [6] Ho T, Karri V. Fuzzy expert system to estimate ignition
research program timing for tuning of hydrogen car. Lecture Notes in
Computer Science November 2008;2008(5264):570e9.
AS Australian Standard
[7] Karri, V. and T. Ho. Predictive models for emission of
ECU Engine Control Unit hydrogen powered car using various artificial intelligent
L Engine Load tools, in neural computing and applications. November 2008.
N Engine Speed [8] Negnevitsky M. Artificial intelligence: a guide to intelligent
RPM Round Per Minute systems. Harlow, Essex: Addison Wesley; 2001.
ADL Advanced Dash Logger [9] Maren AJ, Harsten CT, Pap RM. Handbook of neural
MAP Manifold Absolute Pressure computing applications. Academic Press; 1990.
[10] Huang SH, Zhang HC. Artificial neural networks in
ppm Part Per Million
manufacturing: concepts, applications and perspectives,
ARMSE Average Root Mean Square Error IEEE transactions on components. Packaging and
o
BTDC Degree Before Top Death Centre Manufacturing Technology 1994;17(2):212e28.
MBT ¼ Maximum Brake Torque [11] Zurada JM. Introduction to artificial neural systems. St. Paul:
AMass Mass Air Flow West Publishing Company; 1992.
MAP Manifold Air Pressure [12] Khanna T. Foundations of neural networks. Massachusetts:
Addison-Wesley; 1990.
FAPW Fuel Actual Pulse Width
[13] Caudill M, Butler C. Basic Networks. Understanding neural
EGT Exhaust Gas Temperature
networks e computer explorations, vol. 1. Massachusetts
ET Engine temperature Institute of Technology; 1992.
[14] Ho T, Karri V, Madsen O. Prediction of hydrogen safety
parameters using intelligent techniques. International
references Journal of Energy Research; 2008.
[15] Karri V, Ho T, Madsen O. Artificial neural networks and
neuro-fuzzy inference systems as virtual sensors for
[1] Ho T, Karri V, Lim D, Barrett D. An investigation of engine hydrogen safety prediction. International Journal of
performance parameters and artificial intelligent emission Hydrogen Energy June 2008;33(11):2857e67.
prediction of hydrogen powered car. International Journal of [16] Verhelst S, Sierens S. Hydrogen engine specific properties.
Hydrogen Energy June 2008;33:3837e46. International Journal of Hydrogen Energy 2001;26:987e90.
[2] Barrett DST. In: Study of the performance of a four cylinder [17] Verhelst S, Sierens S. Aspects concerning the optimisation of
hydrogen-fuelled internal combustion engine. Hobart, a hydrogen fueled engine. International Journal of Hydrogen
Australia: School of engineering , University of Tasmania; Energy 2001;26:981e5.
March 2007. [18] Inc, M.. MATLAB Neural network toolbox. 2001.

You might also like