Estimation of Biodiesel Properties from Chemical Composition An Artificial Neural Network (ANN) Approach

Transcription

1 Estimtion of Biodiesel Properties from Chemicl Composition An Artificil Neurl Network (ANN) Approch M. I. Jhirul, W. Sendeer, R. J. Brown nd L. Moghddm Biofuel Engine Reserch Fcility, Science nd Engineering Fculty Queenslnd University of Technology (QUT), Brisbne, Austrli Abstrct Biodiesel, produced from renewble feedstock represents more sustinble source of energy nd will therefore ply significnt role in providing the energy requirements for trnsporttion in the ner future. Chemiclly, ll biodiesels re ftty cid methyl esters (FAME), produced from rw vegetble oil nd niml ft. However, cler differences in chemicl structure re pprent from one feedstock to the next in terms of chin length, degree of unsturtion, number of double bonds nd double bond configurtion-which ll determine the fuel properties of biodiesel. In this study, prediction models were developed to estimte kinemtic viscosity of biodiesel using n Artificil Neurl Network (ANN) modelling technique. While developing the model, 7 prmeters bsed on chemicl composition commonly found in biodiesel were used s the input vribles nd kinemtic viscosity of biodiesel ws used s output vrible. Necessry dt to develop nd simulte the network were collected from more thn 0 published peer reviewed ppers. The Neurl Networks Toolbox of MtLb R0 softwre ws used to trin, vlidte nd simulte the ANN model on personl computer. The network rchitecture nd lerning lgorithm were optimised following tril nd error method to obtin the best prediction of the kinemtic viscosity. The predictive performnce of the model ws determined by clculting the coefficient of determintion (R ), root men squred (RMS) nd mximum verge error percentge (MAEP) between predicted nd experimentl results. This study found high predictive ccurcy of the ANN in predicting fuel properties of biodiesel nd hs demonstrted the bility of the ANN model to find meningful reltionship between biodiesel chemicl composition nd fuel properties. Therefore the model developed in this study cn be useful tool to ccurtely predict biodiesel fuel properties insted of undertking costly nd time consuming experimentl tests. Keywords Biodiesel; Kinemtic viscosity; Artificil Neurl Network (ANN); Prediction model I. INTRODUCTION Vegetble oil methyl or ethyl esters, commonly referred to s biodiesel, re renewble liquid fuel lterntive to petroleum diesel. In technicl terms, biodiesel is diesel engine fuel comprised of mono-lkyle esters of long chin ftty cids derived from vegetble oil or niml fts, designted B00 nd meeting the requirements of ASTM D 675or EN 44 []. These mono-lkyl esters re the min chemicl species tht give biodiesel similr or better fuel properties compred with petroleum diesel []. It is lso sfer to hndle, store nd trnsport becuse it is biodegrdble, non-toxic nd hs higher flsh point thn diesel. One of the mjor dvntges of biodiesel is tht it hs potentil to reduce dependency on imported petroleum through the use of domestic feedstock for production [-]. However, s fuel, there re currently severl disdvntges to using biodiesel in diesel engine pplictions. The mjor disdvntge is its higher viscosity compred to conventionl diesel fuel especilly t lower tempertures. These differences minly result from the difference in chemicl compositions between petroleum diesel nd biodiesel. Biodiesels re usully mde from vegetble oils nd niml ft feedstock through chemicl rection clled Trnsesterifiction. In this process the pure oil nd ft is converted from nturl oil (three long chin crbon molecules struck together by glycerin) into three monolkyl esters (three seprted long chin crbon molecules) shown in Fig.. Triglycerides re llowed to rect with n lcohol (mostly methnol) under cidic or bsic ctlyst conditions producing ftty cid ester of the respective lcohol nd free glycerol. After complete rection, glycerol is removed s byproduct nd esters re known s biodiesel. However, biodiesels re not pure methyl esters wheres smll mount of impurities re exist in commercil biodiesel depending on production process. The most common impurities re includes, free ftty cid, free glycerin, monoglyceride, diglyceride nd triglyceride which eventully ffects the physicl properties of biodiesel [3]. Viscosity is defined s the resistnce to sher or flow which describes the behvior of liquid in motion ner solid boundry such s the wlls of pipe. It is one of the most criticl fetures of fuel which plys dominnt role in fuel spry, fuel-ir mixture formtion nd the combustion process [4-6]. In diesel engine, fuel is spryed into combustion

2 chmber nd then tomized into smll drops ner the nozzle exit. The liquid fuel forms cone-shped spry t the nozzle exit nd its viscosity hs significnt effects on the size of fuel drop nd penetrtion. Therefore, fuel viscosity influences engine combustion, performnce nd emissions, especilly crbon monoxide (CO) nd unburnt hydrocrbon (UHC) [7]. However, very low fuel viscosity is not desirble becuse the fuel then doesn t provide sufficient lubriction for the precision fit of fuel injection pumps, resulting in lekge or incresed wer. Therefore, ll biodiesel stndrds specify the upper nd lower limits of kinemtic viscosity for biodiesel. Fig. Trnseterifiction rection [3] The viscosity of biodiesel is dependent on its complex chemicl composition nd impurities. A recent study showed tht viscosity increses with incresing length of both the ftty cid chin nd lcohol group [8]. As the lengths of the cid nd lcohol segments in the ester molecules incresed, so did the degree of rndom intermoleculr interctions nd consequently viscosity. Reft [9] reported tht shorter ftty cid chins with longer lcohol moieties disply lower viscosity thn ester with longer ftty cid chins nd shorter lcohol moieties. Other fctors tht influence biodiesel viscosity include: number nd position of double bonds, degree of sturtion, moleculr weight, brnching hydroxyl groups nd the mount of impurities, such s unrected glycerides or glycerol etc [3]. In recent yers, ANN modelling techniques hve gined in populrity due to their bility to ccurtely predict from smll dt sets (exmples) rther thn from lrger dt sets requiring costly nd time-consuming studies nd experiments. The most importnt feture of rtificil neurl networks is their bility to solve problems through lerning by exmple, rther thn by becoming involved in the detiled chrcteristics of the systems. This feture of mkes them very useful becuse they works like blck box model nd do not require detil or complete informtion bout the problem, nd when there re only sets of dt inputs nd outputs of the system. Therefore it hs been successfully pplied in vrious disciplines, including neuroscience [0], mthemticl nd computtionl nlysis [], lerning systems [] nd engineering design nd ppliction [3-5]. In this pper, the potentil of ANN modelling techniques in identifying sustinble future genertion biodiesel feedstock re identified bsed on the most recent literture. ANN lso pplied to predict importnt fuel properties of biodiesel. Broutin et l. [6] predicted density from ftty cid profile for plm oil biodiesel. Rmds et. l. [7] used ANN to predict the cetne number of biodiesel bsed on ftty cid profile. ANN lso hs been used to predict viscosity, flsh point, fire point bsed on diesel-biodiesel blend rtio [8]. However these prediction models re limited for specific biodiesel nd experiment condition. No investigtion ws to develop ANN model to predict kinemtic viscosity of biodiesel for wide rnges of biodiesel feedstock nd consider the impurities generlly contined in biodiesel. Such n ANN prediction model hs been developed in this study. The developed ANN model hs been simulted with new input dt nd prediction bility ws presented grphiclly. II. ARTIFICIAL NEURAL NETWORKS (ANN) The foundtion of rtificil neurl networks (ANN) in scientific sense begins with biologicl neuron. In the brin, there is flow of coded informtion (using electrochemicl medi, the so-clled neurotrnsmitters) from the synpses towrds the xon. The xon of ech neuron trnsmits informtion to number of other neurons. Groups of neurons re orgnized into sub-systems nd the integrtion of these sub-systems forms the brin. On the other hnd, n ANN is composed of lrge number of simple processing units clled neurons which re fully connected to ech other through doptble synptic weight. This resembles brin in two spects. Knowledge cn be cquired through trining nd knowledge cn be stored. A widely used ANN model clled the multi-lyer perception (MLP) neurl network, which consist of one input lyer, one or more hidden lyer (s) (middle) in between input nd output lyers nd one output lyer. Ech lyer employs severl neurons (nodes), nd ech neuron in lyer is connected to the neurons in the djcent lyer with different weights. Signls flow into the input lyer, pss through the hidden lyer(s), nd rrive t the output lyer. With the exception of the input lyer, ech neuron receives signls from the neurons of the previous lyer. The incoming signls or input (x ij ) re multiplied by the weights (v ij ) nd summed up with the bis ( bj ) contribution. Mthemticlly it cn be expressed s: n net = X V + b () j The output of neuron is determined by pplying n ctivtion function to the totl input nd clculted using eqution. If the computed outputs do not mtch the known (i.e. trget) vlues, NN model is in error. Then, portion of this error is propgted bckwrd through the network. This error is used to djust the weight nd bis of ech neuron throughout the network so the next itertion error will be less for the sme units. The procedure is pplied continuously nd i ij j

3 repetitively for ech set of inputs until there re no mesurble errors, or the totl error is smller thn specified vlue. The most importnt feture of rtificil neurl networks is their bility to solve problems through lerning by exmple, rther thn by fully understnding the detiled chrcteristics of the systems. This feture mkes it very useful becuse it works like blck box model, nd does not require detil or complete informtion bout the problem, nd cn be utilized when ll tht is vilble re sets of dt inputs nd outputs of the system. It hs nturl propensity to store experientil knowledge nd to mke it vilble for use (Fig. ). Therefore, this nonliner computer lgorithm cn model lrge nd complex systems with mny interrelted prmeters. III. DATA COLLECTION Necessry dt to trin n ANN for kinemtic viscosity prediction of biodiesel were collected from ppers published in recognized interntionl Journls, conferences nd report of renowned reserch centers round the world. The populr scientific nd electronic dtbses including Elsevier, Tylor nd Frncis, DieselNet, Scopus, Springer, Wiley interntionl, Americn chemicl society, IEEE, SAGA Publiction, MDPI etc. were serched for relevnt ppers for this study. More thn 0 ppers were collected mostly published in lst decde which contin experimentl results of chemicl composition nd kinemtic viscosity of biodiesels t 40 C. Fig. shows the bsolute number nd the percentge of the vrious biodiesel dtsets included in the current investigtion. The collected dt comprises biodiesels (ftty cid methyl esters) from 55 different feedstock, 5 biodieselbiodiesel blends nd 59 pure methyl esters. From the Fig. it cn be seen lrge number of feedstock were investigted worldwide for biodiesel production, those includes edible Fig.. Working principle of n ANN [3] Fig. 3 shows severl stges re involved in ANN prediction model developing process. An importnt stge of neurl network is the trining step, in which n input is introduced to the network together with the desired output: the weights nd bis vlues re initilly chosen rndomly nd the weights djusted so tht the network produces the desired output. After trining, the weights contin meningful informtion, contrry to the initil stge where they re rndom nd meningless. When stisfctory level of performnce is reched, the trining stops, nd the network uses the weights to mke decisions. Fig. 3. Proposed flow chrt of n ANN prediction model development non-edible vegetble oils, west cooking oils, beef tllow, chicken fts, fish oils, lge etc. It is lso interesting to see tht, s representtive of ctul biodiesel, mny investigtions hve used pure methyl esters which re mostly rtificilly trough chemicl process. The most populr feedstock of biodiesel investigted worldwide is soyben (.36%) followed by plm (6.5%), sunflower (6.5%), cnol (5.%), rpeseed (4.83%) s edible oil feedstocks. Among the non-edible oil, most investigted feedstock is Jthrop found in 5 ppers. During dt collection stge, cre hve been tken to only include dt from those reserchers who ctully mesured themselves the vlues reported, following the interntionlly ccepted experimentl stndrds. Altogether, 35 dt sets of biodiesel chemicl compositions nd corresponding kinemtic viscosity were selected to use in this study. The min chemicl compositions in biodiesels re mono-lkyle esters of ftty cids which re feedstock dependent nd vry significntly from to the next. Therefore wide rnge of ftty cid profiles found in the collected dt tht re tbulted in Tble I with their structure nd nme. These ftty cids re stright-chin compounds rnging in size from 8 4 crbons nd re minly the three

4 types, sturted, mono-unsturted nd poly-unsturted. In the sturted cid, no hydrogen cn be dded chemiclly nd they contin only single bonds, wheres in mono-unsturted ftty cids, one hydrogen cn be dded nd contin one double bond. Similrly, in poly-unsturted ftty cids, more thn one hydrogen cn be dded nd contin multiple double bonds. In generl, ftty cids re designted by two numbers: the first number denotes the totl number of crbon toms in the ftty cid nd the second is the number of double bonds. For exmple, 4: designtes Myristoleic cid which hs 4 crbon toms nd one double bond. The most common ftty cid found in biodiesel smples re Oleic (C8:) followed by Steric (8:0), Linoleic (C8:), Plmitic (C6:0) nd Linolenic (C8:3) cid esters. Fig. 5 shows tht these ftty cid esters re found in lmost every biodiesel smples. It is interesting to see in Fig. 5 tht the Oleic (C8:) nd Linoleic (C8:) not only represented in most of the biodiesel smples, but re lso highest in verge weight percentge in the biodiesel smples which re bout 40% nd 3% respectively. On the contrry, n verge bout of 7.5% nd 6.5% of Linolenic (C8:3) nd Steric (8:0) cids methyl esters presents in the smples. This is lso reflected in the verge vlues of chin length, sturted ftty cid esters, monounsturted esters nd poly-unsturted methyl esters show in Tble II. Aprt from ftty cid methyl esters, other chemicl compositions usully found in the biodiesel re minly unrected triglyceride, monoglyceride, diglyceride, triglyceride nd free ftty cid represented s cid vlue. Although the mounts of these impurities re smll s shown in Tble II, these re importnt fctors in determining biodiesel physicl properties. Therefore ll interntionl biodiesel stndrds specify the upper limit of these impurities in order to ensure the qulity of biodiesels. Fig. 5: Number nd verge mount of ech ftty cid methyl ester found in the smples. TABLE I. STRUCTURAL FORMULAE FOR FATTY ACIDS METHYL ESTER FOUND IN BIODIESEL SAMPLES Acid chin C:N Type b CL c Chemicl structure Cprilic C8:0 SFA 8 CH 3(CH ) 6COOH Cpric C0:0 SFA 0 CH 3(CH ) 8COOH Luric C:0 SFA CH 3(CH ) 0COOH Myristic C4:0 SFA 4 CH 3(CH ) COOH Myristoleic C4: MUFA 4 CH 3(CH ) 5 CH=CH (CH ) 5 COOH Pentdecenoic C5:0 SFA 5 CH 3(CH ) 3COOH Plmitic C6:0 SFA 6 CH 3(CH ) 4COOH Plmitoilic C6: MUFA 6 CH 3(CH ) 6 CH=CH (CH ) 6 COOH Meptdecenoic C7:0 SFA 7 CH 3(CH ) 5COOH Steric C8:0 SFA 8 CH 3(CH ) 6COOH Oleic C8: MUFA 8 CH 3(CH ) 7 CH=CH (CH ) 7 COOH Linoleic C8: PUFA 8 CH 3(CH ) 4 CH= CHCH CH =CH (CH ) 7 COOH Linolenic C8:3 PUFA 8 CH 3(CH ) CH=CHCH CH= CHCH CH=CH(CH ) 7 COOH Archidic C0:0 SFA 0 CH 3(CH ) 8COOH Gondonic C0: MUFA 0 CH 3(CH ) 7 CH=CH (CH ) 9 COOH Behenic C:0 SFA CH 3(CH ) 0COOH Erucic C: MUFA CH 3(CH ) 7 CH=CH (CH ) COOH Lignoceric C4:0 SFA 4 CH 3(CH ) COOH C: the number of crbon toms nd N: the number of double bonds of crbon toms in the ftty cid chin. b SAF: sturted ftty cids, MUAF: Mono unsturted ftty cids nd PUFA: Poly unsturted ftty cid. c CL: Chin length of hydrocrbon of respective methyl ester TABLE II. MAXIMUM, MINIMUM AND AVERAGE VALUES OF VARIABLES OTHER THAN INDIVIDUAL FATTY ACID METHYL ESTERS Vribles Mximum Minimum Averge Averge chin length (ACL) Sturtion (SFA, wt.%) Mono-unsturtion (MUFA, wt.%) Poly-unsturtion (PUFA, wt.%) Free glycerin (wt.%) Monoglyceride (wt.%) Diglyceride (wt.%) Triglyceride (wt.%) Acid vlue (AV) Fig. 4. Number nd percentge of biodiesel dtsets investigted in this study

5 Fig. 6. Kinemtic viscosities of the smples with biodiesel stndrds Due to the vrition in chemicl composition, the kinemtic viscosities of biodiesels differ significntly which is evident in Fig. 6. The kinemtic viscosity of biodiesels collected in this study rnges from.4 to 7. cst with the men vlue of 4.39 cst. Most of the viscosity dt re within the cceptble rnge of USA (ASTM 67) biodiesel stndrd which is from.9 to 6 cst. However, the Europen biodiesel stndrd (EN 44) impose more strict kinemtic viscosity regultion for biodiesel which rnges from 3.5 to 5 cst nd mny biodiesels shown in Fig. 6 fil to meet this stndrd IV. PREDICTION MODEL DEVELOPMENT The kinemtic viscosity prediction model ws developed for biodiesel using n ANN which represents mthemticl reltionship between input nd output prmeters of system s such blck box model. The selection of input prmeters which contribute to the output is therefore crucil tsk. It is lso desirble to minimize the number of input prmeters for n ANN system in order to reduce the computtionl time. In generl, the best input prmeters re being selected bsed on n understnding of the physics of problem. Published literture suggested tht the kinemtic viscosity of biodiesel is function of chemicl composition- the ftty cid profile of methyl ester nd impurities. In the present study, 7 vribles re used s input prmeters in developing the ANN for viscosity prediction. These prmeters include: mss percentge of 8 ftty cid methyl ester tht is commonly found in the biodiesel nd 9 prmeters listed in Tble. Among 35 dt sets tht hve been collected from literture, 37 sets re used in the ANN model trining process wheres other 5 dt sets re rndomly selected for simultion. The three dt sets needed to develop ANNs: for trining, vlidtion nd testing the network. The usul pproch is to prepre single dt-set, nd differentite it by rndom selection. The lerning lgorithm clled the feed-forwrd bck-propgtion ws pplied for the single hidden lyer shown in the Fig. 7. The Neurl Network ws optimized using the MATLAB Version R0 Neurl Network Toolbox. In the trining stge, to define the output ccurtely, the number of neurons in hidden lyer nd lerning lgorithm hs been optimized using tril nd error methods. In the optimistion process, networks were trined with lgorithms vilble in the softwre. These re, Byesin Regulriztion (BR), Qusi-Newton (QN), Resilient Bckpropgtion (RB), Scled Conjugte Grdient (SCG), Levenberg-Mrqurdt (LM), Conjugte Grdient with Powell Restrts (CGPR), Fletcher-Powell Conjugte Grdient (FPCG), Polk-Ribiére Conjugte Grdient (PRCG), One Step Secnt (OSS), Vrible Lerning Rte Grdient Descent (VLRGD), Grdient Descent with Momentum (GDM) nd Grdient Descent (GD). For ech lgorithm, the prediction ccurcy of the model ws observed for 0 to 60 neurons in the single hidden lyer with 5 intervls. It ws found tht prediction ccurcy improved with the number neurons in the hidden lyer. Fig. 8 shows the results of the network optimistion process. The best results ws found using LM lerning lgorithm with 45 neuron in the hidden lyer. Moreover, mong the lgorithm use in this study, LM lerning lgorithms showed better results followed by GD, OSS nd CGPR. With very high number of neurons in hidden lyer, the computtionl time incresed significntly long with the over fitting problems which eventully reduced the models effectiveness. Neurons in the input lyer hve no trnsfer function. Logistic sigmoid (logsig) trnsfer function hs been used in the hidden lyer while pureliner (purelin) trnsfer function hs been used in output lyer. The optimized ANN structure is shown in Fig. 9. The trining ws stopped fter 53 itertions s shown in Fig. 0 to void over fitting of the network. After stopping the trining the softwre hs identified the best ANN network t 47 itertions. Fig 7. Structure of ANN

6 sttisticl nlysis in terms of bsolute frction of vrince (R ), root men squred (RMS) nd mximum verge error percentge (MAEP). Formuls to clculte the error prmeters re shown in Eqution to 4. R = RMS N I = N ( ) E EM I = i N = = ( E E ) p ( E E ) N p () (3) Fig. 8. Optimistion of ANN model MEP = N i = N E E E p 00 (4) where E -Actul result E p -Predicted result E m -Men vlue N-Number of pttern Fig. 9. Optimised ANN for kinemtic viscosity prediction Actul nd predicted dt shown in Fig. indictes the good prediction bility of the developed ANN for predicting kinemtic viscosity of biodiesel for wide rnges of biodiesel feedstocks. In sttisticl nlysis it ws found tht the bsolute frction of vrince (R ) ws close to unity 0.93, Root-Men- Squred (RMS) error ws 0.8 cst nd mximum verge error percentge (MAEP) ws 4.65%. This is good correltion between the mesured nd predicted kinemtic viscosity. The network hs been trined well nd cn be used to simulte biodiesel kinemtic viscosity over wide rnge of feedstock. However, the prediction ccurcy of the model should be further improved by incresing the number nd rnges of trining dt set. Fig. 0. Trining performnce of ANN model V. SIMULATION The bility of the ANN model to predict kinemtic viscosity hs been verified by simulting it with unknown input vribles. For this purpose, 5 sets of dt selected rndomly from collected dt sets which ws not used during the trining of ANN model. After simultion, the predicted kinemtic viscosity of biodiesel ws compred with the corresponding experimentl vlues nd shown in Fig. 6. Using the results produced by the network, sttisticl methods hve been used to investigte the prediction performnce of NN results. To judge the prediction performnce of network, severl performnce mesures re used. Those include Fig.. Experimentl nd predicted kinemtic

7 VI. CONCLUSION The im of this pper ws to investigte the suitbility of using rtificil neurl networks (ANN) for the prediction of biodiesel kinetic viscosity from its chemicl composition. To trin the network, experimentl dt ws collected from more thn 0 peer reviewed pper published in recognized journl, conferences nd reports. Altogether, 35 dt sets which included biodiesel chemicl compositions nd corresponding kinemtic viscosity t 40 C of biodiesel were collected. The collected dt comprises biodiesels (ftty cid methyl esters) from 55 different feedstocks, 5 biodiesel-biodiesel blends nd 59 pure methyl esters. In the trining stge, to define the output ccurtely, the number of neurons in hidden lyer nd lerning lgorithm ws optimized using tril nd error methods. The best network for this study ws stndrd bck propgtion (BP) neurl network model with LM lgorithms nd 45 neurons in hidden lyer. The performnce of the developed ANN prediction model ws evluted by prediction with the 5 dt sets which were not used in the trining process. The network produced the predicted results in good rgument to the experimentl ones. The overll results show tht the networks cn be used s n lterntive wy for predicting kinemtic viscosity of biodiesel t different temperture conditions. The bsolute frction of vrince (R), Root-Men-Squred (RMS) nd mximum verge error percentge (MAEP) ws vlues were found 0.93, 0.8 cst nd 4.65% respectively which is within n cceptble rnge of ccurcy. The results of this study lso show tht n ANN hs the bility to lern nd generlize wide rnge of experimentl conditions. Therefore, the usge of ANNs my be recommended to optimize the chemicl composition of biodiesels to optimize fuel qulity for internl combustion engine ppliction. However the network should be further improved by including dditionl robust dt set during the trining process. [5]. S. Lee, D. Tnk, J. Kusk, nd Y. Disho, Effects of diesel fuel chrcteristics on spry nd combustion in diesel engine JSAE, 00, Vol. 3, pp [6]. M. E. Tt, nd J.V. Gerpen, Biodiesel blend detection with fuel composition sensor. Trns ASAE 00, Vol. 9, pp [7]. G. Knothe, nd K. R. Steidle, Kinemtic viscosity of biodiesel fuel components nd relted compounds Influence of compound structure nd comprison to petrodiesel fuel components Fuel, 005, Vol. 84(9), pp [8]. D. P. Geller, nd J, W. Goodrum, Effects of specific ftty cid methyl esters on diesel fuel lubricity Fuel 004, Vol. 83, pp [9]. A. A. Reft, Correltion between the chemicl structure of biodiesel nd its physicl properties. Int J Environ Sci Technol 009, Vol. 6, pp [0]. E. Alkim, E. Gürbüz, nd E. Kılıç, A fst nd dptive utomted disese dignosis method with n innovtive neurl network model Neurl Networks, 0 Vol. 330, pp []. M. A. Cost, A. P. Brg, nd B. R. Menezes, Convergence nlysis of sliding mode trjectories in multi-objective neurl networks lerning Neurl Networks, 0, Vol. 330 pp.-3. []. S. Crrillo, J. Hrkin, L. McDid, S. Pnde, S. Cwley, B. McGinley, nd F. Morgn. Advncing interconnects density for spiking neurl network hrdwre implementtions using trffic-wre dptive network-on-chip routers Neurl Networks 0, Vol. 330, pp [3]. T. Smur, nd H. Hyshi, Directionl spike propgtion in recurrent network: Dynmicl firewll s nisotropic recurrent inhibition Neurl Networks, 0, Vol. 330, [4]. D. Go, Z. Yng, C. Ci, nd F. Liu, Performnce evlution of multilyer perceptrons for discriminting nd quntifying multiple kinds of odors with n electronic nose Neurl Networks 0, Vol. 330 pp [5]. R. C. J. Minnett, A. T. Smith, W. C. Lennon, R. Hecht-Nielsen, Neurl network tomogrphy: Network repliction from output surfce geometry Neurl Networks 0, Vol. 45, pp [6]. S. Broutin, nd M. K. Arou, Estimtion of Vegetble Oil-Bsed Ethyl Esters Biodiesel Densities using Artificil Neurl Networks J Appl Sci, 008, Vol. 8 pp [7]. A. S. Rmdhs, S, Jyrj, C. Murleedhrn, nd K. Pdmkumri, Artificil neurl networks used for the prediction of the cetne number of biodiesel. Renewble Energy 006, Vol. 35 pp [8]. J. Kumr, Bnsl A, nd M. K. Jh, Comprison of Sttisticl nd Neurl Network Techniques in Predicting Physicl Properties of Vrious Mixtures of Diesel nd Biodiesel Proceedings of the World Congress on Engineering nd Computer Science. Sn Frncisco, USA. October 4-6, 000. VII. ACKNOWLEDGMENTS The uthors wish to express their thnks to QUTPRA scholrship, Fculty of Science nd Engineering (SEF), Queenslnd University of Technology (QUT) for proving fund for this study. VIII. REFERENCE []. A. Demirbs, Reltionships derived from physicl properties of vegetble oil nd biodiesel fuels Fuel, 008, Vol. 87, pp []. S. Fernndo, P. Krr, R. Hernnde, nd J. S. Kumr, Effect of incompletely converted soyben oil on biodiesel qulity Energy, 007, Vol. 3, pp [3]. M. I. Jhirul, R. J. Brown, W. Sendeer, Z D Ristovski, nd I. O'Hr, Artificil neurl network pproch in identifying sustinble future genertion biofuel feedstock Energies. Specil issue: Alterntive Fuels for the Internl Combustion Engines (ICE), 03, Vol. 6, pp [4]. C. Y. Choi nd R. D. Reitz, A numericl nlysis of the emissions chrcteristics of biodiesel blended fuels J Eng Gs Turbines Power, 999, Vol., pp