Applied Energy 119 (2014) 330–340

Contents lists available at ScienceDirect

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

Performance and exhaust emissions prediction of a CRDI assisted single

cylinder diesel engine coupled with EGR using artificial neural network
Sumit Roy ⇑, Rahul Banerjee, Probir Kumar Bose
Department of Mechanical Engineering, NIT Agartala, Tripura, India

h i g h l i g h t s

 CRDI/high pressure fuel injection reduces PM and BSFC with the penalty of increase in NOx emissions.
 EGR operation at lowest injection duration case of CRDI operation reduces NOx.
 Artificial neural network modeling of BSFC, BTE, CO2, NOx and PM.
 ANN is capable in predicting performance and emission parameters of the experimental engine.
 MSE, RMSE, MAPE, MSRE, THEIL U2, R, R2, NSE, KGE metrics used as evaluation benchmarks.

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

Article history: The present study explores the potential of artificial neural network to predict the performance and
Received 6 August 2013 exhaust emissions of an existing single cylinder four-stroke CRDI engine under varying EGR strategies.
Received in revised form 8 January 2014 Based on the experimental data an ANN model is developed to predict BSFC, BTE, CO2, NOx and PM with
Accepted 11 January 2014
load, fuel injection pressure, EGR and fuel injected per cycle as input parameters for the network. The
Available online 31 January 2014
study was carried out with 70% of total experimental data selected for training the neural network,
15% for the network’s cross-validation and remaining 15% data has been used for testing the performance
Keywords:
of the trained network. The developed ANN model was capable of predicting the performance and emis-
Artificial neural network
CRDI
sions of the experimental engine with excellent agreement as observed from correlation coefficients
EGR within the range of 0.987–0.999, mean absolute percentage error in the range of 1.1–4.57% with
Engine performance noticeably low root mean square errors. In addition to common correlation coefficients, the present study
Exhaust emissions incorporated special statistical error and performance metrics such as mean square relative error,
forecasting uncertainty Theil U2, Nash–Sutcliffe Coefficient of Efficiency and Kling–Gupta Efficiency.
Low values of MSRE and Theil U2 combined with commendable indices of NSE and KGE proved beyond
doubt the robustness and applicability of the model so developed. Furthermore, the developed ANN
model was capable of mapping the PM–NOx–BSFC trade-off potential of the CRDI operation under EGR
for all cases of actual observations with significant accuracy.
Ó 2014 Elsevier Ltd. All rights reserved.

1. Introduction millennium. Such challenges have demanded paradigm shifts in

Research studies [1–5] have clearly established that the
chronological evolution of emission mandates have continuously
challenged diesel engine design to be contemporary in the present
Abbreviations: ANN, Artificial Neural Network; BDO, baseline diesel operation;
BP, brake power; BSFC, Brake Specific Fuel Consumption; BTDC, before top dead
centre; BTE, Brake Thermal Efficiency; CO2, Carbon-di-oxide; CI, compression
ignition; CRDI, Common Rail Diesel Injection; DI, direct injection; EGR, Exhaust Gas
Recirculation; IC, Internal Combustion; KGE, King–Gupta Efficiency; MAPE, Mean
Absolute Percentage Error; MSE, Mean Square Error; MSRE, Mean Squared Relative
Error; NOx, oxides of nitrogen; NSE, Nash–Sutcliffe Coefficient of Efficiency; PM,
particulate matter; ppm, parts per million; R, correlation coefficient; RMSE, Root
Mean Square Error; Theil U2, Theil uncertainty.
⇑ Corresponding author. Tel.: +91 9402108135; fax: +91 3812346360.
E-mail address: sumitroy@hotmail.de (S. Roy).
diesel engine technology to meet the desired emission directives
on one hand and consumer expectations of superior fuel
economy on the other. Among various contingent technological
advancements, Common Rail Diesel Injection (CRDI) systems have
spearheaded the technological renaissance [6,7]. CRDI systems
have been observed to significantly reduce the specific fuel con-
sumption [8,9] and soot emission precursors [10–12] as compared
to conventional diesel operation, however, with a propensity of
increased NOx formation [7,13,14]. Therefore, a trade-off scenario
arises wherein lower soot emissions and fuel consumption foot-
print of a CRDI system is penalized by higher NOx emissions
[6,13,15]. Exhaust Gas Recirculation (EGR) has been proved to be
a cost effective tool in NOx containment [16–19]. It has been
employed widely due to its simplicity of operation and efficiency

0306-2619/$ - see front matter Ó 2014 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.apenergy.2014.01.044
 S. Roy et al. / Applied Energy 119 (2014) 330–340 331

as compared to the operational and development challenges of
NOx after-treatment systems.
The present day responsiveness to legislative and consumer
requirements is necessitating an extraordinary increase of
dependence on several new degrees of control simultaneously.
Furthermore, the highly nonlinear dependence of the pollutants on
engine operating variables, have made classical one-dimensional
map-based control redundant for efficient engine control [20,21] in
the present day. Advanced on board diagnostic capability and the
ability to reconfigure control on the fly has become an essential crite-
rion in implementing the complex control paradigms. The need of the
day warrants a truly multidimensional, adaptive, learning control
system [22] having credible performance and emissions prediction
capabilities across the entire design range of engine operation.
In contrast to traditional nonlinear identification methods,
Artificial Intelligence (AI) platforms have proved to substantially
reduce the cost, time and complexity associated with engine devel-
opment, performance mapping and control system development
[23–26]. The inherent adaptive strengths of ANN based AI systems
have established itself as a robust system identification and map-
ping tool [22,27–29] in the control paradigm of Internal Combus-
tion (IC) engines. ANNs have the ability in simulating accurate
engine behavior and as such have become an inexpensive basis
of a virtual sensing system for onboard measurements of engine
emissions continuously in real time [30–35]. Early researchers in
the field of IC engines [36–42] were quick to recognize the ability
of ANN platforms in capturing nonlinear trends in complex data to
model the same to any desired accuracy and thus establish their
inherent robustness as ‘universal approximators’ [43,44].
1.1. Motivation of the present study
Injection system as detailed in Table 1. The engine was coupled
to an air-cooled eddy current dynamometer of PowerMagÒ make.
2.1.2. Common rail direct injection setup
The CRDI setup is an attachment to the experimental engine. It
consists of a high-pressure fuel pump, rail, high-pressure fuel
injector and the heart of the system being the Electronic Injection
Controller (EIC). The description of the fuel injection system is gi-
ven in Table 2.
2.1.3. Exhaust gas recirculation setup
The EGR circuit essentially consisted of an EGR control valve,
exhaust control valve, bypass valve, EGR cooler (water-cooled;
double pass), exhaust cooler (water-cooled), digital manometers,
air box orifice meters along with condensate traps. The EGR was
controlled with a digital control valve fitted to the EGR setup.
The EGR fraction was calculated as in Eq. (1) [45]
_ aw=oEGR  m
m _ aEGR
%EGR ¼ ð1Þ
_
maw=oEGR
_ = mass of air.
where ma
2.1.4. Emission analysis instrumentation
The exhaust gases were sampled by a 5 Gas analyzer and an AVL
smoke meter (415S) was used to measure the soot content, present
in the exhaust. The specifications of the emission measuring appa-
ratus are detailed in Tables A1 and A2 in Appendix-A.
NOx was calculated in terms of NO as per the specifications of
the AVL Digas 444. NOx measured in ppm was recalculated in
brake specific units as per the equation given below (Eq. (2)):

_ airþfuel  ½E
An experimental investigation was carried out on an existing ½m PPM  ½ME  1000
½ESPECIFIC ¼ ð2Þ
diesel engine to exploit the synergetic benefits of a CRDI system BP  ½MEXHAUST
by containing the expected increase in NOx emission through
EGR. Subsequent to the experimental investigation, data analysis where E is the pollutant under consideration, [M]E is the molar mass
revealed a marked sensitivity of the output variables with the cho- of the pollutant and [M]EXHAUST is the average molar mass of the
sen inputs under study. Juxtaposing the context of the discussion exhaust gases. The constant 1000 is included in the numerator to
enumerated previously, it was evident that precise controls of convert the mass flow unit from (kg/h) to (g/h).
the input variables were desired to reap the maximum benefits
of the performance–emission trade-off merit posed by the
Table 1
CRDI–EGR system. Consequently a potential domain of study was Experimental engine specification.
established to investigate the benefits of an AI based ANN system
Specification Resources
to model the performance-emission characteristics.
Such a predictive model in real time would then serve as an indis- Make Vidhata
No of cylinder One (1)
pensable tool in predictive controller domains which can be used to
Bore 120 mm
adjudge the trade-off characteristics in real time onboard systems. Stroke 139.7 mm
The applications of ANN are not new in the field of IC engines and Displacement 1580 cc
have been successfully employed in engine management systems, Cooling Water
fuel management systems and in predicting the common pollutants Compression Ratio 18:1

of CI/SI engines. In The present study, an endeavor was undertaken Valve timing
Exhaust valve opening 35 deg before BDC
to bridge the gap by a first-of-a-kind study, incorporating both CRDI
Exhaust valve closing 4 deg after TDC
control parameters and EGR quantification for a direct correspon- Inlet valve opening 4 deg before TDC
dence to emission and performance outputs through predictive Inlet valve closing 35 deg after BDC
modeling. Further such an effort would also stand in good stead
for the establishment of a virtual sensing platform for the difficult
to measure pollutants in on-board diagnostic systems in real time.
Table 2
Specification of the fuel injector.
2. Experimental investigation Specification Resources
Type Common rail injection system
2.1. Experimental setup and methodology Make Bosch
Injection Pressure 10–120 MPa
2.1.1. Experimental engine Number of holes 5 (Symmetric)
Nozzle diameter 0.15 mm
The experiment was conducted on an existing single cylinder
Injection angle 120°
four-stroke CI engine assembled with Common Rail Direct Fuel
332
According to the specification of the smoke meter, specific dry
soot emission can be calculated from Filter Smoke Number (FSN)
by using Eq. (3) [46]
1 mair þ mfuel
Soot ¼  5:32  FSN  exp0:3062FSN  0:001 
0:405 1:2929
where mair denotes air mass flow (kg/h) and mfuel denotes fuel mass
flow (kg/h).
2.1.5. System layout
The layout of the experimental setup used to conduct the exper-
iments is shown in Fig. 1.
2.2. Experimental uncertainty analysis
A comprehensive measurement uncertainty analysis has been
carried out to identify and quantify the error arising in the exper-
imental emission measurements and the derived engine perfor-
mance parameters. The uncertainty was calculated on account of
the employed instrumentation, its calibration, observation accu-
racy and the methodology of experimentation in a given ambient
condition [47–49]. For the present work, the experimental error
analysis consists of two components:
2.2.1. Identification of the measuring accuracy of each device
implemented in the test bed installation
The uncertainty expected during the sampling of the observed
parameters by the corresponding components of engine and emis-
sion analysis instrumentation are enlisted in Tables A1–A3 in
Appendix A, as declared by their respective manufacturers.
2.2.2. Calculating the uncertainty of the computed values of the
various parameters resulting from the post-processing of the measured
data
The combined uncertainty analysis for the performance parame-
ters has been carried out on the basis of the root mean square meth-
od [50]. In this method the total uncertainty DU in the estimated
value of a quantity Q, where Q = f [x1, x2 . . . xn] having individual er-
rors Dx1, Dx2, . . ., Dxn is given by Eq. (4). The calculation of the same
as per this principle has been detailed in Table A4 in Appendix A.
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 2  2
@U @U
DU ¼ Dx1 þ Dx2 þ               
@x1 @x2
Fig. 1. Schematic diagram of experimental setup.
S. Roy et al. / Applied Energy 119 (2014) 330–340
Each recorded value for a given case of engine operation (w &
w/o EGR) was the average value of six (6) consecutive observations
over a sampling span of 120 s. The Total Sampling Uncertainty
(TSU) of each observation set was computed as per Eq. (4) at each
ð3Þ case of engine operation. The sampling uncertainty of the emission
analyzer for each of the respective pollutants and the relative range
of the consecutive observations were taken into account. An exam-
ple calculation of the same for a particular case of engine operation
has been detailed in Table A5 in Appendix A for ready reference.
For a credible viewpoint of the uncertainty analysis, the additional
index of standard deviation of the consecutive samplings has been
computed at each of the designated engine operating conditions.
The average total sampling uncertainty and the average standard
deviation over the entire scope of experimentation have been
reported in Table A6 in Appendix A.
2.3. Experimental results
In the present experimental work, the fuel injection methodol-
ogy was undertaken for 25%, 50%, 75% and full load characteristics
of the engine. The experimental engine when subjected to progres-
sive higher loads then demanded a higher injection rail pressure.
Consequently for each variation of injection duration at the given
load, the injection rail pressure would adjust automatically as
per the specification of the CRDI setup to maintain the constant
speed set. The initial injection duration for each load was set to
approximately 225 bar which was the designed fuel nozzle open-
ing pressure of the experimental engine. The start of injection an-
gle was set to 5° BTDC for all the cases. The injection duration on
the CRDI platform for each load was then reduced at particular
intervals specific to each case of loading. Such reduction of injec-
tion duration was accompanied by a complementary increase of
pressure needed to maintain the set constant speed of 800 rpm.
The reduction in fuel injection duration was continued in steps un-
til a minimum injection duration was reached at that given load,
beyond which the stability of engine operation was compromised.
The fuel injection duration strategy for each case of loading are en-
listed in Table A7 in Appendix A for ready reference.
On preliminary observations of performance and emission
parameters under such CRDI operation showed that increased rail
 2 pressure yielded commendable reduction in soot and BSFC as well
@U
Dxn as increase in BTE than the base-line diesel operation. However,
@xn
such premium of soot and BSFC reduction was accompanied by a
ð4Þ drastic increase in NOx emission. To this end, EGR was inducted
 S. Roy et al. / Applied Energy 119 (2014) 330–340 333

for each case of least fuel injection duration at each case of loading.
Cooled EGR operations were performed keeping the EGR outlet
temperature from the cooler constant for all cases of operation at
5%, 10%, 15%, 20%, 25% and 30% recirculation by suitable
adjustments of the cooling water flow rates in the cooler. The study
confirmed EGR as an effective tool in reducing NOx emission but
with a distinct increase in soot emissions and BSFC.
the variation in the output ‘b’ is independent of any variations in
the input ‘a’. A positive correlation coefficient indicates the di-
rect-proportionality between ‘a’ and ‘b’. However, negative correla-
tion coefficient indicates the inverse-proportionality between ‘a’
and ‘b’.
3.2. The ANN Architecture

3. Construction of ANN model The neural network architecture refers to the arrangement of
neurons into layer(s) and the connection patterns between the hid-
3.1. Selection of input and output parameters den layers, the choice of activation functions and most pertinently
the number of hidden neurons in the hidden layers. In the present
Notwithstanding the ‘black box’ label, the success of any ANN study we used a feed-forward neural network model to predict the
model depends upon a clear understanding of the problem under output parameters. A multilayer feed-forward neural network is
consideration with the identification of the most important vari- consists of one input layer, t hidden layer(s) and one output layer.
ables in the process to be modeled. Of the large number of impos- The input layer is interconnected with the hidden layers and sim-
ing factors influencing the performance and formation of ilarly hidden layer is connected with the output layer with the help
pollutants in diesel engines, the final choice of input variables were of synaptic weights [51]. During the training phase of the network,
dictated by the most pertinent contributing factors ascertained synaptic weights are modified on each iteration, in order to learn
after survey of previous similar research studies [17–19] as well the underlying patterns existing in the data. The states of the hid-
as the feasibility of on board measurement of the chosen variables den and output layer determine the output of the network.
in real time. In the present work, an ANN modeling was used to
predict the relationship of BTE, BSFC, CO2, NOx and soot with load, 3.2.1. Choice of network parameters
fuel injection pressure, EGR percentage and fuel injected per cycle For the chosen feed-forward network, a continuous differentia-
as inputs. ble Log-sigmoid activation function was chosen for both the hid-
The correlation between the chosen input variables and desired den as well as for the output neurons. Activation functions serve
outputs in the purview of the present study which has been to introduce nonlinearity into neural networks that makes it more
encapsulated in the correlation matrix (Table 3 and Fig. 2) bear powerful than linear transformations. The logistic sigmoid transfer
testimony to the conflicting influences of the input variables on function has been commonly used as an activation function in
the desired emission and performance output variables. The
correlation matrix in Table 3 signifies the relationship between
the inputs and outputs (considered in the present experimental Table 4
Details of the network parameters developed on the MATLAB Platform.
study) in terms of Pearson Product Moment correlation coefficient
whose value varies between [1, 1]. For example, a zero correla- MATLAB
tion coefficient between ‘a’ and ‘b’ input–output pair signifies that Topology 4 inputs, 5 outputs and 2 hidden layers with 10 hidden
neurons in each layer (4-10-10-5)
Data Training subset: 70% randomly selected observation data
Table 3 (309 patterns)
Matrix of the Pearson Product Moment correlation coefficient of each desired output Validation subset: 15% randomly selected observation data
with chosen input variables. (66 patterns)
Test subset: 15% randomly selected observation data (66
Load Fuel injection EGR (%) Fuel injected/ patterns)
pressure cycle
Activation Log-sigmoid
BSFC 0.941905 0.14627 0.057175 0.985163221 function
BTE 0.944941 0.32642 0.02415 0.802582424 Training Levenberg–Marquardt
CO2 0.877108 0.497537 0.227556 0.72499936 algorithm
NOx 0.64639 0.531601 0.26451 0.726712082 Loss function Minimum MSE
Soot 0.165383 0.85042 0.541958 0.410188053 criteria
Stopping Stop the network training when the validation error starts
criteria increasing

Fig. 2. Effect of chosen input parameters on desired output variables as per their
respective correlations. Fig. 3. Architecture of the proposed ANN.
334 S. Roy et al. / Applied Energy 119 (2014) 330–340
multilayer perception model, because it is a differentiable, contin-
uous and non-linear function [38,52]. The advantages of employing
Log-sigmoid activation function in the feed forward neural archi-
tecture is that it provides a form of automatic gain control (it is
self-limiting) and its output cannot grow infinitely large or small.
The activation function is given by
1
f ðxÞ ¼ ð5Þ
1 þ ex
Mean Square Error (MSE) was chosen as the loss function to be
minimized, as MSE possesses the very satisfying properties of con-
vexity, symmetry, and differentiability and an excellent metric in
the context of optimization [38,40,52,53].
3.2.2. Choice of learning rule
Subsequent to having determined the activation function, it was
then necessary to apply an appropriate learning rule which is
essentially a performance criterion involving the simultaneous
and incremental adjustment of network weights in such a way that
the activation functions gradually assume features that help collec-
tively approximate the desired response.
In this fully connected, feed-forward architecture, the mean
square error of the network is minimized by a gradient descent
approach, often known as back propagation learning scheme. The
back-propagation learning algorithm is most popular for multi-
layer perception. Usually back-propagation neural networks with
Levenberg–Marquardt learning algorithm has been utilized for
accurate predictions [53–55]. The second-order methods such as
the Levenberg–Marquardt algorithm provide faster solutions
because of the incorporation of an extra second derivative of error
Fig. 4. Overall correlation coefficients of the developed network.
information and automatic internal adjustments that are made to
the learning parameters.
3.2.3. Determination of training and testing data
Training neural network architecture can be seen as a nonlinear
optimization problem in which the task is to find out the set of
parameters i.e. synaptic weights such that the network output is
as close as to the desired output. The present study consisted of
four input and five output parameters. There were 441 values in
the experimental dataset and for every input and output parameter
49 cases existed.
Conferring to the studies [38,40,52,53,56] it was discovered that
different ratios for training and testing data were used. In the
present study, 70% (309 patterns) of total experimental data was
randomly selected for training the neural network, 15% (66 pat-
terns) for the network’s cross-validation (to avoid over-fitting)
and remaining 15% (66 patterns) data has been used for testing
the performance of the trained network.
3.3. Statistical evaluation of output parameters
In order to evaluate the prediction performance of the proposed
ANN model, we have considered correlation coefficient (i.e. R) and
coefficient of determination (i.e. R2); as in line with the correlation
matrices adopted in similar studies [40,42,52,53].
The absolute fraction of variance (R2) is defined as
!
Pn 2
2 i¼1 ðt i  oi Þ
R ¼1 Pn 2
ð6Þ
i¼1 ðoi Þ
 S. Roy et al. / Applied Energy 119 (2014) 330–340 335

Majority of the research work [41,57–60] on the application of The developed ANN model is limited to the experimental dataset
ANN on performance emission modeling in IC engines, have evalu- obtained on the research engine that was used in this project as
ated their model against some statistical error measures namely detailed in Table 1.
that of Mean Square Error (MSE), Root Mean Square Error (RMSE)
and Mean Absolute Percentage Error (MAPE). As regards the liter- 3.5. Model evaluation
ature related to this work [40,56,61] for the performance of the
developed ANN model a MAPE of 5% has been taken as a limit for In view of the shortcomings of model evaluation criteria by the
the output parameters. The error described by RMSE and MAPE commonly used error and correlation metrics as outlined in Sec-
are defined, respectively as follows: tion 3.3, the proposed model has been evaluated against additional
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi error and performance metrics keeping in mind the necessity of
u n
u1 X scale independence and avoidance of ambiguous interpretation
RMSE ¼ t ðt i  oi Þ2 ð7Þ arising out of infinite, undefined or zero values arising in error
n i¼1
measurements.
n  
1X 
t i  oi   100 3.5.1. Relative error measure
MAPE ¼  ð8Þ
n i¼1 ti  Mean Squared Relative Error (MSRE) (Eq. (9)) comprises the
mean of the squared of the error made by the model predicted val-
where ‘n’ is the number of pattern in the data set, ‘t’ is the actual ues relative to the observed values and helps to provide a scale
output and ‘o’ is the predicted output value [42,62]. independent measure of model sensitivity towards larger relative
errors that will occur at low magnitudes of model estimation
3.4. Selection of optimal ANN architecture [64]. For each of the error metrics, a perfect model score would
be zero.
Selection of optimal ANN architecture is essential in order to  2 
1 Xn 
harness the maximum advantage of the computational intelligence  ti  oi 
MSREj ¼   ð9Þ
of the network. In regression problems, single hidden layer was n i¼1 ti 
j
found to be sufficient to learn the data [40,53,57] but some authors
have also suggested the usage of two hidden layers [38,56,63]. The
number of hidden neurons in the hidden layers which are the units
of nonlinear adaption wherein an increased number of hidden neu-
rons may be used in order to achieve more ‘degrees of freedom’
and allow the network to store more complex patterns. However,
in the effort to find an optimum network topology, there must be
a balance in increasing the number of hidden layers and the num-
ber of hidden neurons therein such that the total network weights
do not impose a high computational cost and avoid generalization
loss during the testing phase of the network. Previous studies uti-
lizing ANN in performance-emission modeling in diesel engines
have widely used a single hidden layer but with the expense of a
correspondingly high number of neurons therein to match the
highly complex and nonlinear dependence of the input parameters
to the emission and performance outputs.
In the current study, we have performed an optimal network
topology search consisting of up to two hidden layers. As stated
above the pragmatism of the choice of the hidden layer was
appreciated in view of the pitfalls of single layer topology and
the compromise of increasing hidden layers beyond the chosen
value; at the expense of generalization loss and computational
burden. The choice of search of up to two hidden layers was also
concurrent with the studies envisaged in the research works of
[38,56].
The number of hidden nodes varied from two to twenty-five for
each of the optimal iterations [40,53]. Subsequent to the selection
of the network topology, the neural network iterations were
performed subject to the loss function and stop training criteria as
detailed in Table 4 on the data set. The criterion of MSE was used
as the loss function to be minimized during the topology search. To
prevent over-training during the training phase, we have stopped
the network training when the validation error starts increasing.
This avoids the over-learning of the network. ANN architecture
with four neurons in input layer, two hidden layer with ten
neurons in each layer and five neurons in output layer
(4-10-10-5) was found to be the optimal architecture as shown
in Fig. 3.
The process of choosing the optimum topology in the present
study was performed on MATLAB software, which was also chosen
as our iteration solver. Table 4 exhibits the network model details. Fig. 5. (a and b) Comparison of ANN predicted BSFC with measured data.
336 S. Roy et al. / Applied Energy 119 (2014) 330–340

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
3.5.2. Correlation metrics of performance KGE ¼ ðrj  1Þ2 þ ðbj  1Þ2 þ ðaj  1Þ2 ð11Þ
The present study also has been fortified with additional corre-
lation indices in conjunction with the standard measure of ‘NSE’ where;
and ‘KGE’.    
The Nash–Sutcliffe Coefficient of Efficiency (NSE) [65] statisti-
rp ^
p
aj ¼ ;b ¼ ;
cally defined as one minus the ratio of Sum Square Error (SSE) to rt j j ^t j
the statistical variance of the observed dataset about the mean of
the observed dataset (Eq. (10)). It represents an improvement over 3.5.3. Model uncertainty
the R2 since it is sensitive to differences in the observed and mod- The Theil uncertainty popularly known as Theil U2 (Eq. (12))
eled means and variances. was proposed in the works of [67] and is a measure of forecast
( "P #)
n
ðoi  ti Þ2 quality in that the Theil uncertainty returns a standardized mea-
NSEj ¼ 1  Pi¼1 ð10Þ sure that is a combination of the mean error and variance of the
n ^2
i¼1 ðt i  tÞ j errors between the predicted and the observed values. This is
the most pertinent measure for evaluating the construct validity
In order to isolate the bias, variability and correlation compo-
of the models under study, as the system constructs individuals
nents of a computed NSE, the Kling–Gupta Efficiency (KGE) [66]
that are optimized for both the mean error and variance [68].
has been incorporated to improve model evaluation and diagnos-
Lower values indicate better results for a credible model
tics by incorporating error compensation for the bias and variabil-
forecasting ability.
ity components and thus provide a more reliable measure of the
2qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi3
overall agreement between the predicted values by a given model Pn 2
and the observed values than conventional linear correlation co- 6 i¼1 ðoi  t i Þ 7
ðU2THEIL Þj ¼ 4 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pn ffi 5 ð12Þ
efficient (Eq. (11)). The maximum positive score of one represents 2
i¼1 ðt i Þ
a perfect model in each of the above correlation measures. j

Fig. 6. (a and b) Comparison of ANN predicted BTE with measured data. Fig. 7. (a and b) Comparison of ANN predicted CO2 with measured data.
 S. Roy et al. / Applied Energy 119 (2014) 330–340 337

4. Results and discussion relative error of 0.000989 and Theil U2 uncertainty of 0.018361
along with root mean square error of 0.023939 kg/kW h. The mean
Based on the experimental work an Artificial Neural Network absolute percentage error observed was 2.65%. In the category of
(ANN) model was developed to predict BSFC, BTE, CO2, NOx and correlation metrics (Fig. 11), the ANN model displayed excellent
PM. The input parameters were load, fuel injection pressure, EGR overall agreement indices with the experimental observations
(%) and fuel injected per cycle. Prediction of output parameters
using ANN for the experimental engine yielded impressive
correlation statistics i.e. the predictive ability of the developed
network for BSFC, BTE, CO2, NOx and PM was commendable. The
comprehensive correlation coefficient ‘R’ of the selected network
architecture is shown in Fig. 4.
Inspection of Figs. 5–9 indicates the consistent and
commendable concurrency of the network predicted values with
the actual observations for the entire range of operation. This
suggests the inherent sensitivity and robustness of the network
in its proficiency to map the performance and emission values
simultaneously with excellent accuracy independent of the case
of engine operation.

4.1. Model prediction of performance parameters

The comparison of ANN predictions with experimental (target)

data for Brake Specific Fuel Consumption (BSFC) is indicated in
Fig. 5a and b. ANN prediction of BSFC with experimentally
measured data exhibited (Fig. 10) extremely low mean square

Fig. 9. (a and b) Comparison of ANN predicted PM with measured data.

Fig. 8. (a and b) Comparison of ANN predicted NOx with measured data. Fig. 10. Comparison of the error metrics.
338 S. Roy et al. / Applied Energy 119 (2014) 330–340
Fig. 11. Comparison of the correlation co-efficient metrics.
wherein it achieved a 98.7116% Pearson Product Moment correla-
tion coefficient (R) and the additional correlation metrics of NSCE
and KGE was 97.2724% and 98.2236% respectively.
Statistical analyses for Brake Thermal Efficiency (BTE) (Fig. 10)
revealed that the developed ANN model had a very low MSRE con-
tent of 0.0007 with Theil U2 uncertainty of 0.014439 along with
RMSE of 0.67502% and MAPE of 1.12% across all the test points.
In the correlation metrics (Fig. 11), the Pearson Product Moment
correlation coefficient (R), NSE and KGE of the ANN predicted val-
ues with the observed data were 0.995355, 0.988768 and 0.97175
respectively. The promise of the ANN predicted BTE is also evident
in the persistent concurrency with the experimental values for all
cases of experimental observations as evident from Fig. 6a and b.
4.2. Model prediction of emission parameters
Fig. 7a and b illustrates the predicted versus experimental data
for CO2 emission. In predicting CO2 emission, analysis showed that
(Fig. 10) the ANN model scored extremely low MSRE of 0.00144
with very low Theil U2 uncertainty of 0.016023 and RMSE content
Fig. 12. Experimental observation of the PM–NOx–BSFC trade-off characteristics.
of 0.106179 vol.% The MAPE observed was 2.54%. The correlation
coefficient (R), NSE and KGE between the predicted CO2 and exper-
imental CO2 was found to be 0.993949, 0.984308 and 0.98206
respectively (Fig. 11).
The MSRE, Theil U2 and RMSE of the ANN model in estimating
the NOx emissions were seen (Fig. 10) to be 0.001741, 0.018963
and 0.431545 g/kW h respectively along with 1.99% of MAPE. In
the correlation metrics (Fig. 11), the correlation coefficient (R),
NSE and KGE of the ANN predicted values with the observed data
were 0.998086, 0.995657 and 0.977861. This is corroborated by
the consistent concurrency of the ANN predicted values with the
experimental observations as seen from Fig. 8a and b for the entire
range of observations.
Similarly, PM emissions were predicted accurately as evident
from the resemblance with the experimental values for all cases
of experimental observations (Fig. 9a and b). In error metrics
(Fig. 10), the developed ANN model scored extremely low MSRE
of 0.003897. The Theil U2 uncertainty and RMSE was observed
0.013379 and 0.043703 g/kW h respectively along with MAPE of
4.57%. In the category of correlation metrics (Fig. 11), the devel-
oped ANN model exhibited excellent overall agreement indices
with the experimental observations wherein it achieved a
99.9387% correlation coefficient (R), 99.8737% NSE and 99.349%
KGE.
4.3. Model validation
The study has been further corroborated by exploring the effi-
cacy of the models in mapping the PM–NOx–BSFC trade-off charac-
teristics simultaneously on a single platform in comparison to the
actual observed trade-off zones. Fig. 12 represents the influence of
diesel injection duration and EGR at given point of loading (labeled
as the fourth parameter) on PM, NOx and BSFC over the entire
spectra of experimental investigation. Zone 1 gives the lowest soot
at the expense of highest NOx while Zone 3 gives the lowest NOx at
 S. Roy et al. / Applied Energy 119 (2014) 330–340
Fig. 13. ANN predicted PM–NOx–BSFC trade-off characteristics.
the expense of the highest soot and BSFC. Zone 2 with the penalty
of slightly higher BSFC gives the optimal trade-off characteristics
with respect to soot and NOx. The soot-NOx–BSFC trade-off front
for the entire CRDI operation under EGR strategies is represented
by BCDE. CDE shows the best trade-off characteristic (best compro-
mise between NOx and soot) that was available under the entire
experimental investigation.
Inspection of Fig. 13 clearly manifest the adeptness of the devel-
oped neural network model as an inherent robust and accurate
mapping tool by the exact reproduction of the identified zones
and the Smoke–NOx–BSFC frontier A, B, C, D, E and F in Fig. 12. This
is a clear indication of the developed network model as a robust
virtual sensing platform ‘on the fly’ Smoke–NOx–BSFC trade-off
mapping tool vital to real time engine control and optimization
strategies.
5. Conclusions
The present study investigates the applicability of ANN to emu-
late the performance and emission observations of a CRDI engine
under varying EGR strategies. The ANN model developed in the
present study employed one input layer with four hidden neurons,
two hidden layers with ten hidden neurons in each layer and one
output layer with five hidden neurons (4-10-10-5). The results of
the proposed ANN model has been evaluated on a comprehensive
statistical platform of error and performance metrics in compari-
son to the actual results of the experimental investigation. From
the error analysis, it was evident that the ANN predicted data
matched the experimental data with high overall accuracy with
correlation coefficient (R) values ranging from 0.987 to 0.999.
The MAPEs were observed to be in the range of 1.1–4.57% with
very low RMSE.
The present study also incorporates innovative error and perfor-
mance metrics against which the model has been evaluated. The
superior performance in the MSRE, Theil U2, NSE and KGE metrics
of evaluation of the developed ANN model establishes its inherent
339
robustness and sensitivity. The excellent generalization capability
observed of the developed model helps establish itself as an
effective forecasting tool which can be employed to emulate actual
engine operation for the desired outputs even in cases of engine
operation outside the purview of the present study.
The present study was thus able to establish an ANN architec-
ture, which could consistently emulate actual engine parameters
proficiently even under different modes of CRDI operation with
EGR thereby providing a holistic and robust predictive platform
for virtual sensing in real time optimization strategies for such en-
gine operation.
6. Future scope
In context of the present study, the ANN model can be deployed
in the form of firmware in embedded computing environment. An
interesting aspect of the proposed study would be to design the
hardware for proposed ANN architecture for the purpose of incor-
porating onto the existing ECUs (Electronic Control Units) which
are often found in modern vehicles. Hardware implementation of
logic(s) provides faster computation along with low-power
requirement which are proving for modern VLSI design for better
controlling of IC engine parameters in real time.
Appendix A. Supplementary material
Supplementary data associated with this article can be found, in
the online version, at http://dx.doi.org/10.1016/j.apenergy.2014.
01.044.
References
[1] McGeehan JA, Yeh S, Couch M, Hinz A, Otterholm B, Walker A, et al. On the road
to 2010 emissions: field test results and analysis with DPF-SCR system and
ultra low sulfur diesel fuel. SAE International; 2005.
[2] Johnson TV. Diesel emission control in review. SAE Int J Fuels Lubr
2008;1:68–81.
340 S. Roy et al. / Applied Energy 119 (2014) 330–340
[3] Johnson TV. Review of diesel emissions and control. SAE Int J Fuels Lubr
2010;3:16–29.
[4] Johnson TV. Diesel emissions in review. SAE Int J Eng 2011;4:143–57.
[5] Zhao H. Advanced direct injection combustion engine technologies and
development. Diesel engines, vol. 12. Cambridge, UK: Woodhead Publishing
Limited; 2010.
[6] Badami M, Nuccio P, Trucco G. Influence of injection pressure on the
performance of a di diesel engine with a common rail fuel injection
system. SAE International; 1999.
[7] Suh HK. Investigations of multiple injection strategies for the improvement of
combustion and exhaust emissions characteristics in a low compression ratio
(CR) engine. Appl Energy 2011;88:5013–9.
[8] Balusamy T, Marappan R. Effect of injection time and injection pressure on CI
engine fuelled with methyl ester of Thevetia peruviana seed oil. Int J Green
Energy 2010;7:397–409.
[9] Nagata K, Tanaka Y, Yano K. Technologies of DENSO common rail for diesel
engine and consumer values. Convergence Transportation Electronics
Association; 2004.
[10] Shimazaki N, Tsurushima T, Nishimura T. Dual mode combustion concept with
premixed diesel combustion by direct injection Near top dead center. SAE
International; 2003.
[11] Pickett LM, Siebers DL. Non-sooting, low flame temperature mixing-controlled
DI diesel combustion. SAE International; 2004.
[12] Minato A, Tanaka T, Nishimura T. Investigation of premixed lean diesel
combustion with ultra high pressure injection. SAE International; 2005.
[13] Desantes JM, Benajes J, Molina S, González CA. The modification of the fuel
injection rate in heavy-duty diesel engines: part 2: effects on combustion.
Appl Therm Eng 2004;24:2715–26.
[14] Pierpont DA, Reitz RD. Effects of injection pressure and nozzle geometry on D.I.
diesel emissions and performance. SAE International; 1995.
[15] Payri F, Benajes J, Arrègle J, Riesco JM. Combustion and exhaust emissions in a
heavy-duty diesel engine with increased premixed combustion phase by
means of injection retarding. Oil Gas Sci Technol – Rev IFP 2006;61:247–58.
[16] Reitz RD. Controlling D.I. diesel engine emissions using multiple injections and
EGR. Combust Sci Technol 1998;138:257–78.
[17] Ladommatos N, Abdelhalim S, Zhao H. Control of oxides of nitrogen from diesel
engines using diluents while minimising the impact on particulate pollutants.
Appl Therm Eng 1998;18:963–80.
[18] Hountalas DT, Mavropoulos GC, Binder KB. Effect of exhaust gas recirculation
(EGR) temperature for various EGR rates on heavy duty DI diesel engine
performance and emissions. Energy 2008;33:272–83.
[19] Maiboom A, Tauzia X, Hétet J-F. Experimental study of various effects of
exhaust gas recirculation (EGR) on combustion and emissions of an
automotive direct injection diesel engine. Energy 2008;33:22–34.
[20] Atkinson C, Mott G. Dynamic model-based calibration optimization: an
Introduction and application to diesel engines. SAE International; 2005.
[21] Malaczynski GW, Mueller M, Pfeiffer J, Cabush D, Hoyer K. Replacing
volumetric efficiency calibration look-up tables with artificial neural
network-based algorithm for variable valve actuation. SAE International; 2010.
[22] Kalogirou SA. Artificial intelligence for the modeling and control of
combustion processes: a review. Prog Energy Combust Sci 2003;29:515–66.
[23] Lenz U, Schroeder D. Artificial intelligence for combustion engine control. SAE
International; 1996.
[24] Isermann R, Sinsel S, Schaffnit J. Modeling and real-time simulation of diesel
engines for control design. SAE International; 1998.
[25] Shayler PJ, Goodman M, Ma T. The exploitation of neural networks in
automotive engine management systems. Eng Appl Artif Intell
2000;13:147–57.
[26] Papadimitriou I, Warner M, Silvestri J, Lennblad J, Tabar S. Neural network
based fast-running engine models for control-oriented applications. SAE
International; 2005.
[27] de Lucas A, Durán A, Carmona M, Lapuerta M. Modeling diesel particulate
emissions with neural networks. Fuel 2001;80:539–48.
[28] Ghobadian B, Rahimi H, Nikbakht AM, Najafi G, Yusaf TF. Diesel engine
performance and exhaust emission analysis using waste cooking biodiesel fuel
with an artificial neural network. Renew Energy 2009;34:976–82.
[29] Oğuz H, Sarıtas I, Baydan HE. Prediction of diesel engine performance using
biofuels with artificial neural network. Expert Syst Appl 2010;37:6579–86.
[30] De Cesare M, Covassin F. Neural network based models for virtual NOx sensing
of compression ignition engines. SAE International; 2011.
[31] Docquier N, Candel S. Combustion control and sensors: a review. Prog Energy
Combust Sci 2002;28:107–50.
[32] Nareid H, Grimes M, Verdejo J. A neural network based methodology for
virtual sensor development. SAE International; 2005.
[33] Alberer D, Re Ld, Winkler S, Langthaler P. Virtual sensor design of particulate
and nitric oxide Emissions in a DI diesel engine. Consiglio Nazionale delle
Ricerche; 2005.
[34] Clark NN, Gajendran P, Kern JM. A predictive tool for emissions from heavy-
duty diesel vehicles. Environ Sci Technol 2002;37:7–15.
[35] Delagrammatikas GJ, Assanis DN. Development of a neural network model of
an advanced, turbocharged diesel engine for use in vehicle-level optimization
studies. Proc Inst Mech Eng Part D: J Automob Eng 2004;218:521–33.
[36] Ouenou Gamo S, Ouladsine M, Rachid A. Diesel engine exhaust emissions
modelling using artificial neural networks. SAE International; 1999.
[37] Hentschel R, Cernat RM, Varchmin JU. In-car modelling of emissions with
dynamic artificial neural networks. SAE International; 2001.
[38] Çay Y, Çiçek A, Kara F, Sağiroğlu S. Prediction of engine performance for an
alternative fuel using artificial neural network. Appl Therm Eng
2012;37:217–25.
[39] Kesgin U. Genetic algorithm and artificial neural network for engine
optimisation of efficiency and NOx emission. Fuel 2004;83:885–95.
[40] Yusaf TF, Buttsworth DR, Saleh KH, Yousif BF. CNG-diesel engine performance
and exhaust emission analysis with the aid of artificial neural network. Appl
Energy 2010;87:1661–9.
[41] Canakci M, Erdil A, Arcaklioğlu E. Performance and exhaust emissions of a
biodiesel engine. Appl Energy 2006;83:594–605.
[42] Sayin C, Ertunc HM, Hosoz M, Kilicaslan I, Canakci M. Performance and exhaust
emissions of a gasoline engine using artificial neural network. Appl Therm Eng
2007;27:46–54.
[43] Hornik K, Stinchcombe M, White H. Multilayer feedforward networks are
universal approximators. Neural Networks 1989;2:359–66.
[44] Hartman E, Keeler JD, Kowalski JM. Layered neural networks with Gaussian
hidden units as universal approximations. Neural Comput 1990;2:210–5.
[45] Pradeep V, Sharma RP. Use of HOT EGR for NOx control in a compression
ignition engine fuelled with bio-diesel from Jatropha oil. Renew Energy
2007;32:1136–54.
[46] Specification of AVL 415S filter smoke number meter. Graz, Austria: AVL List
GMBH; 2007.
[47] Devan PK, Mahalakshmi NV. Performance, emission and combustion
characteristics of poon oil and its diesel blends in a DI diesel engine. Fuel
2009;88:861–7.
[48] Mani M, Nagarajan G. Influence of injection timing on performance, emission
and combustion characteristics of a DI diesel engine running on waste plastic
oil. Energy 2009;34:1617–23.
[49] Kannan GR, Anand R. Experimental investigation on diesel engine with
diestrol–water micro emulsions. Energy 2011;36:1680–7.
[50] Rakopoulos CD, Dimaratos AM, Giakoumis EG, Rakopoulos DC. Investigating
the emissions during acceleration of a turbocharged diesel engine operating
with bio-diesel or n-butanol diesel fuel blends. Energy 2010;35:5173–84.
[51] Reby D, Lek S, Dimopoulos I, Joachim J, Lauga J, Aulagnier S. Artificial neural
networks as a classification method in the behavioural sciences. Behav Process
1997;40:35–43.
[52] Kara Togun N, Baysec S. Prediction of torque and specific fuel consumption of a
gasoline engine by using artificial neural networks. Appl Energy
2010;87:349–55.
[53] Najafi 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 artificial neural network. Appl Energy 2009;86:630–9.
[54] Parlak A, Islamoglu Y, Yasar H, Egrisogut A. Application of artificial neural
network to predict specific fuel consumption and exhaust temperature for a
diesel engine. Appl Therm Eng 2006;26:824–8.
[55] Cortés O, Urquiza G, Hernández JA. Optimization of operating conditions for
compressor performance by means of neural network inverse. Appl Energy
2009;86:2487–93.
_ A diesel engine’s performance and exhaust emissions.
[56] Arcaklioğlu E, Çelıkten I.
Appl Energy 2005;80:11–22.
[57] Mohamed Ismail H, Ng HK, Queck CW, Gan S. Artificial neural networks
modelling of engine-out responses for a light-duty diesel engine fuelled with
biodiesel blends. Appl Energy 2012;92:769–77.
[58] Gölcü M, Sekmen Y, Erduranlı P, Sahir Salman M. Artificial neural-network
based modeling of variable valve-timing in a spark-ignition engine. Appl
Energy 2005;81:187–97.
[59] Cay Y. Prediction of a gasoline engine performance with artificial neural
network. Fuel 2013;111:324–31.
[60] Çelik V, Arcaklioğlu E. Performance maps of a diesel engine. Appl Energy
2005;81:247–59.
[61] Shivakumar, Srinivasa Pai P, Shrinivasa Rao BR. Artificial neural network based
prediction of performance and emission characteristics of a variable
compression ratio CI engine using WCO as a biodiesel at different injection
timings. Appl Energy 2011;88:2344–54.
[62] Sözen A, Arcaklioǧlu E. Solar potential in Turkey. Appl Energy 2005;80:35–45.
[63] Guang-Bin H. Learning capability and storage capacity of two-hidden-layer
feedforward networks. IEEE Trans Neural Networks 2003;14:274–81.
[64] Dawson CW, Abrahart RJ, See LM. HydroTest: a web-based toolbox of
evaluation metrics for the standardised assessment of hydrological forecasts.
Environ Modell Software 2007;22:1034–52.
[65] Nash JE, Sutcliffe JV. River flow forecasting through conceptual models part I —
A discussion of principles. J Hydrol 1970;10:282–90.
[66] Gupta HV, Kling H, Yilmaz KK, Martinez GF. Decomposition of the mean
squared error and NSE performance criteria: implications for improving
hydrological modelling. J Hydrol 2009;377:80–91.
[67] Theil H. Applied economic forecasting. Chicago: Rand McNally; 1966.
[68] Bliemel F. Theil’s forecast accuracy coefficient: a clarification. J Mark Res
1973;10:444–6.