CNG exhausts emission modeling: Neural network approach

Kirti Bhandari>, Ch Ravi Sekhar, A M Rao and S GangopadhyayTransport Planning and Environment Division, Central Road Research Institute, New Delhi 110 020

Received 28 February 2005; revised 26 lillie 2006; accepted 26 July 2006

Journal of Scientific & Industrial ResearchVol. 65, December 2006, pp, 1000-1007

, I ,'~ .

, i

Traditional statistical regression and Artificial Neural Network (ANN) modeling techniques were applied toassess theemission characteristics of CNG vehicles (cars and three wheelers) to understand the influence of explanatory parameters ,like vehicle age, vehicle type and air-fuel ratio on emissions of CO, HC, CO2 and °2, For ANN modeling. multilayer feed-forward neural network with single hidden layer was considered and back propagation algorithm was applied for training.ANN model. ANN models are shown better predictive models than the traditional statistical modeling techniques for"predicting the CNG exhaust emissions (CO. He. CO2, and O2),

Keywords: ANN, CNG, Exhaust emissions, Regression

IntroductionCompressed natural gas (CNG) has emerged as a

solution to the depleting crude oil resources as well asto deteriorating urban air quality. CNG (85-99%methane) has many desirable properties as .a sparkignition engine fuel. Clean burning, cheap andabundant in many parts of the world, natural gas hasalready played a significant role in Argentina,Canada, Italy, New Zealand and the US. Natural GasVehicles (NGV's) have a marked advantage' overconventional diesels in terms of less pollution(exhaust emission) as follows: carbon monoxide(CO), 84; nitrogen oxides (NOx), 58; and particulatemater (PM), 97 %. In another study", emissioncharacteristics of in-use CNG vehicles in Delhishowed notable reduction in CO and hydrocarbons(BC) emissions from CNG vehicles especially fromthree wheelers.

Chris Brace.' developed and used successfully threeapplications of Artificial Neural Network (ANN)prediction of diesel engine exhaust emission.Shivanagendra & Khare4 studied ANN based linesource models for vehicular exhaust emissionprediction of urban roadways. Dougherty studiedidentification of driver behavior with advanced,traveler in'f~rmation system. J ankowska6 developedmodels for NOx 'and C02 emissions for pulverizedfuel and circulating fluidized bed combustion boilers.Present study con-elates different explanatory

"Author for correspondenceE-mail: kirti@urban.env.nagoya-u.ac.jp

parameters such as vehicle age, vehicle type and air"fuel ratio on emissions of CO, He, CO2; and O2 fromCNG driven vehicles 'using the traditional modelingand ANN modeling techniques.

Materials and MethodsStudy Area and Type of Vehicles

In-use CNG vehicles consisting of three wheelersand cars in Delhi were. tested for their exhaustemission characteristics in April 2004 by using AVL4000 light Digas exhaust gas analyzer. The tailpipeemissions of CO, HC, CO2, O2 along with lambda(air-fuel ratio) were measured under idle conditions.Category of vehicles, registration number and year ofregistration were also recorded. Exhaust emission dataof cars and three wheelers has been used forapplication of traditional regression and ANNmodeling.

Development of Neural Network (NN) ModelA multilayer feed-forward (Fig. 1) NN model with

single hidden . layer was considered and, backpropagation algorithm was selected for training thenetwork7

-9• Back propagation algorithm is systematic

approach for training multilayer NN. Using. thisalgorithm, two propagation phases, forward . andbackward, are required. Forward pass calculatesnetwork output by propagating the input data through.the network. The network output is then comparedwith the desired output to calculate the error using abackward pass; during the backward pass connectionweights are modified to reduce the target en-or.

BHANDARI et 01: CNG EXHAUSTS EMISSION MODELING: NEURAL NETWORK APPROACH 1001

Input Layer Hidden uy.:r(j) Output Layer(i) (k)

lnout Parameters

w, WiL

Network Weichrs

Fig. I-Multilayer feed-forward neural network modelconfiguration \.

Back Propagation AlgorithmBack propagation algorithm used In the present

study is as follows:

Step IDecide network topology and randomize the initial

weights.

Step 2 Forward Pass

i) Compute the hidden layer neuron activation (OH,,)

... (1)

where F( . ) is the Sigmoid Transfer function F( a)= 1/1+ exp -a, Wji=weights from node i(source) tonode j(destination), lpi=input value of node i andpattern p.

ii) Compute output layer neuron activation (00",)

001" = F(net pk) netpk = L W,PH"j ... (2)

where F( . ) is the Sigmoid Transfer function F( a)=1/1 + exp -a, W'i= weights from node j(source) tonode k(destination).

iii) Compute the error for each pattern at each outputnode E"

... (3)

where Tpk=target output value of node k for pattern p,OOpk= actual output value of node for pattern p.

Step 3 Backward Pass

i) Computation of the error signal (D",) at output layerand adjustment of weights between output tohidden nodes

Table I-Training and test data sets used for ANN models

SINo Name of the Number of patternsmodel Training phase Testing phase

Autos Cars Autos Cars

I CO 295 288 98 512 HC 288 269 96 483 CO? and 0, 296 287 98 51

D", = (Tpv 00 p,) 00 pk; ( 1-00 pk) ... (4)

Wkj(New) =Wkj (old) + 1] ~)kOHpk+ cx[Wkiold)-Wki o/d-l) ... (5)

ii) Computation of error signal (D"j) at hidden layerand adjustment of weights between hidden andinput nodes.

~)j = OH pj [J - OH pj} L ~)k Wkj ... (6)W;i(New) =WAold) + 1] ~Jj I pi +cx[Wji(old)- WAo/d-III ···0)

Step 4All the above steps are repeated with new training

patterns. One iteration is over when all trainingpatterns are finished. These iterations continue till theaverage Mean Squared Error (MSE) with in thetolerable limit.

Neural Network TrainingNetwork training represents acqumng knowledge

of predicting emission characteristics of in-use CNGvehicles (three wheelers and cars). Afterpreprocessi ng, datasets (75 %) were considered fornetwork training and remaining (25%) wereconsidered for validation for each individual model.The number of patterns used for network training andvalidation of each individual model are presented atTable I. For ANN model development, neuralsolution software'? has been used. This allows theuser to select an ANN paradigm with architecture andidentify training and test data sets. Training W,,5stopped when the average MSE reached previouslyspecified minimum value (0.00l). After several trials,optimum topology for each model was found out(Table 2).

Neural Network TestingValidation process is conducted to verify the

network reliability. It is achieved by processingtesting sets to the network. These sets should be new

HC model

1002 J SCI IND RES VOL 65 DECEMBER 2006

ParametersCO model

Table 2-Optimum network topology considered for various ANN models

CO? & 07 modelAuto Car

I. Optimum network topology

Number of input neuronsNumber of hidden neuronsNumber of output neurons

II. Optimum network training

Momentum FactorNumber of iterations

Ill. Input data types

All independent variables

IV. Output data types

Dependent variables

23I

231

0.9510000

0.9510000

CO

Auto Car Auto Car

26I

261

262

262

0.9510000

0.9510000

0.9510000

0.9510000

Vehicle age and air fuel ratio (A/F)

HC

Table 3--Performance of ANN model vs regression models

Vehicle type ModelRegression co- efficient(R) during ANN model

training phase

Regression coefficienuk)during regression

modeling

Car COHCCO2

O2COHCCO2

07

Three wheeler

that the network has never been exposed before. Forthis, 25% of data sets were considere ..l. Networkoutput using these sets is compared with t.ie desiredoutput to calculate accuracy rate. If the accuracy rateis low, it means that the network is not properlytrained or the hidden nodes considered in the networktopology are inadequate. Otherwise the network isconsidered to be reliable and ready forimplementation.

Results and DiscussionTotal data sets are randomly separated in training

and testing data sets (Table 1). Training data sets wereused for ANN model development. After gettingrequired MSE, training phase of NN model is stoppedand corresponding synaptic weights between input tohidden and hidden to output layers are used for thevalidation of NN model, for this testing data has beenused. These data sets are unseen to the NN model.Traditional regression analysis also has been carriedout for the same data adopted for ANN modeldevelopment. Thus ANN models are better predictive

Regression co- efficient(R) during ANN model

testing phase

0.680.200.670.960.690.120.830.95

0.710.310.570.('0.650.120.820.93

0.400.130.630.890.420.120.800.88

model than regression models in all the cases, exceptfor CO2 for cars (Table 3).

Significant correlation was observed byconsidering CO-ANN model for cars as shown byregression coefficient (R) value of 0.71 as comparedto the value of 0.65 in case of three wheelers (Fig. 2).But CO- regression models failed to correlate thesignificance between input and output variables. Thisimplies that the explanatory variable such as vehicleage and air-fuel ratios (lambda) have significantcorrelations with the CO emissions.

Tn the case of HC model, both ANN and regressionmodels are unable to correlate between explanatoryvariable and the prediction of He. This issubstantiated by the insignificant R-values of 0.31 incase of cars and 0.12 in case of three wheelers(Fig. 3). This implies that the explanatory variablesuch as vehicle age and air-fuel ratios (lambda) do nothave any significant correlations with the HCemissions.

In case of CO2 (cars), ANN and regression modelsexplains that the explanatory variable such as vehicle

BHANDARI et al: CNG EXHAUSTS EMISSION MODELING: NEURAL NETWORK APPROACH 1003

1~------------------------------~-----------------.--CO....... CO Output_ANN0.8

CO Output_Regression

(a)

.... ,

0.6""

~OAa. ..•..::l00.2 . ,... '." . '"" .,. ,:... .' .; L·JiLH" .. ' ~:':: ':.. - '. \ ..., ,

''''"'''1''--=':''' . , .10 19 28 37 46 55 64 73 82 91

_OAL--------------------------......I

Exemplar

08

1(b)

06·.\ ..

04

1

-

0.2~.•.• I s«.::lQ,.•..::l0

21 45 49

-0. t'

I )'CO-0.4-

-O.J....... CO Output ANN

- - - CO Output Reqression

Exemplar

Fig. 2--Observed and predicted CO emission for (a) cars (b) three wheelers

1004 J SCI IND RES VOL 65 DECEMBER 2006

(a)1 r--------------------------------------------------,

0.9

0.80.7

0.6

--He....... He Output_ANN

He Output_Regression

-::J.s- 0.5::Jo 0.4

~~~'Vvt~vvtvJJ'u'bk;'V'V'r1O~I I I I I I I I I I I

1 10 19 28 37 46 55 64 73 82 91

Exemplar

0.8~----------------~--~-------------------------------.(b) -He

...... ·He Output_ANNHe Output_Regression

0.7

0.6

0.5

0J.....c...-__ t-- __+-_-+-_-t- __-+-_-+_--+__ +--_-t--_-t-_-+-_

1 17 21 25 29 33 37 41 455 9 13

Exemplar

Fig. 3--Observed and predicted HC emission for (a) cars (b) three wheelers

BHANDARI et a/: CNG EXHAUSTS EMISSION MODELING: NEURAL NETWORK APPROACH 1005

(a)--C02

,nn", CO2 Output_ANN--C02 Output_Regression

O.

0.8

0.1'

0.6

-:::JQ. O.-:::J00.1

0~ __ -+~ __ ~ __ ~~ __ ~~ __ ~ __ -7~ __~ __~~ __~ __~~~1 10 19 28 37 46 55 64 73 82 91

Exemplar

(b)1.2~---------------------------------------------------------.

--C02

...... 'C02 Output_ANN--C020utput_Regreesion

0.2

oj----+_--~---+--~~I~r---+---+---~---+---+--~----~1 37 41 45 4913 17 21 25 29 335 9

Exemplar

Fig. 4--Observed and predicted CO2 emission for (a) cars (b) three wheelers

1006 J SCI IND RES VOL 6S DECEMBER 2006

1.2.----------------------------------------------------.

1

(a)

0.4

0.2 --02- - - - - _.O2 Output_ANN-- O2 OutputRegression

O+----+----~--~----~----~---r----+----+----~----~~1 8219 28 37 46 55 64 73 9110

Exemplar

0.9~-------------------------------------------------------.

0.8

0.7

0.6

--02- _..... O2 Output_ANN

-- O2 Output_Regression

(b)

..._ 0.5::lQ.

'5 0.4'o

0.3

0.2

0.1

O~--~---+----~--+-_v-~--~~~~--4_--_+----~--~--_+~

1 9 13 17 21 25 29 33 37 41 45 495

Exemplar

Fig. 5--Observed and predicted O2 emission for (a) cars (b) three wheelers

BHANDARI et a/: CNG EXHAUSTS EMISSION MODELING: NEURAL NETWORK APPROACH

age and air-fuel ratios (lambda) do not have anysignificant correlations with CO2 emissions. But in thecase of CO2 (three wheelers), these models identifieda significant relation between input and outputvariables. In both the cases, ANN models R values areslightly higher than regression models (Fig. 4).

In prediction of O2 emission, regression models aresignificant as indicated by the significant R-values of0.88 in case of cars and three wheelers (Fig. 5).Compared to this, R-values as given by CO-ANNmodels are 0.96 in case of cars and 0.93 in case ofthree wheelers. This indicates that ANN models arehighly capable of correlating between input andoutput variable than regression models.

ANN models are observed more suitable forcapturing the highest and lowest values of CNGemissions whereas regression models are unable to dothis.

ConclusionsANN models are better models than the traditional

statistical modeling techniques (regressiontechniques) for predicting CNG exhaust emissions(CO, HC, CO2, and O2), Although ANN modeling hasbeen able to predict CO, CO2 and O2 emissions, it hasnot been able to show any significant correlations forHC emissions(R=0.3 for cars and 0.12 for threewheelers) for the adopted explanatory variables ofvehicle age and air-fuel ratio. R values, as given byregression modeling, are also very low (R=0.12 inboth vehicle categories) indicating insignificantcorrelation. The success of NN implementation isdependent not just on the quality of data used fortraining but also on the topology/structure of the NNadopted.

1007

AcknowledgementsAuthors thank Director, CRRI, for permission to

publish this paper. Authors also thank Prof P KSikdar, former Director, CRRI, for his valuablesuggestions and guidance.

References1 Mike F, Noton P, Clark N N & Lyons D W, Evaluation of

natural gas versus diesel in medium-duty buses, SAETechnical Paper Series 2000-01-2822, Warendale.Pennsylvania.

2 Singh A, Sharma N, Sharma K & Bhan C, Emissioncharacteristics of in-use CNG vehicles in Delhi, in Workshopon Land use, Transport and Environment (CIRT Pune incollaboration with Harvard University) 3-4 Dec 2001, 26ppa vai lable at http://www.deas.harvard.edulTransportAsia-/workshop-papers

3 Brace C, Prediction of diesel engine exhaust emission usingartificial neural networks, in Neural Networks ill SystemDesign (Lucas Electrical and Electronic System, Solihull) 10J unc (1998), 1- I 1, http://www.bath.ac.uk/-enscjb-/s591.pdf

4 Shiva Nagendra S M & Khare M, Artificial neural networkbased line source models for vehicular exhaust emissionpredictions of an urban roadway, Tranportation Res Part D,9 (2004) 199-208.

5 Dougherty M, A review of neural networks applied totransport, Transportation Res Part C, 3 (1995),247-260.

6 Jankowska A, Neural models of air pollutants emissions inpower units combuston process in AI-METH 2003- ArtificialIntelligence Methods, (Gliwice, Poland) 5-7 November(2003),141-144

7 Ravisekhar C H, Master dissertation mode choice analysisusing neural network, Department of Civil Engineering,University of Roorkee, Roorkee, India, 1999.

8 Hegazy T. F::z;n P & Moselhi 0, Developing practical neuralnetwork applications using back-propogation, J MicroComputers ill Civil Engg, Vol. 9 (2), (1994) 145-159.

9 Jain A K & Mao 1, Artificial neural networks: A tutorial,lEES COIliPUMag, (1996) 31-44.

10 Neural solution software. developed by Neuro Dimension;www.nd.com