APPLICATION OF ARTIFICIAL NEURAL NETWORK TO PREDICT BRAKE SPECIFIC FUEL CONSUMPTION OF RETROFITTED CNG ENGINE

Transcription

1 Interntionl Journl of Mechnicl nd Mterils Engineering (IJMME), Vol. 4 (2009), No. 3, APPLICATION OF ARTIFICIAL NEURAL NETWORK TO PREDICT BRAKE SPECIFIC FUEL CONSUMPTION OF RETROFITTED CNG ENGINE M. I. Jhirul, R. Sidur nd H. H. Msjuki Deprtment of Mechnicl Engineering University of Mly, Kul Lumpur, Mlysi Emil: md_jhirul@yhoo.com ABSTRACT In this pper the pplicbility of rtificil neurl networks (ANN) is investigted for retrofitted compressed nturl gs (CNG) fueled sprk ignition (SI) internl combustion engine (ICE). A four cylinder crbureted petrol engine is converted to run with NG nd used throughout the work. The neurl networks toolbox of Mtlb 6.5 is used to develop nd test the ANN model on personl computer. An optiml design is completed for the 3 to 2 hidden neurons on single hidden lyer with six different lgorithms: btch grdient descent (GD), resilient bck-propgtion (RP), levenberg-mrqurdt (LM), btch grdient descent with momentum (GDM), vrible lerning rte (GDX), scled conjugte grdient (SCG) in the bck-propgtion neurl network model. The trining dt for ANN is obtined from experimentl mesurements. Engine speed (rpm), throttle position, fuel-ir equivlence rtio (φ) nd torque (N-m) were used in input lyer while brek specific fuel consumption (gm/kwh) ws used s output lyer. Sttisticl nlysis in terms of Root-Men-Squred (RMS), bsolute frction of vrince (R 2 ), s well s men percentge error is used to investigte the prediction performnce of ANN. LM lgorithm with 0 neurons on single hidden lyer in bck-propgtion of ANN model hs shown best result in the present study. The degree of ccurcy of the ANN model in prediction is proven cceptble in ll sttisticl nlysis nd shown in results. So, it cn be concluded tht ANN provides fesible method in predicting specific fuel consumption of CNG driven SI engine. Keywords: Internl combustion engine (ICE), Compressed nturl gs (CNG), Artificil neurl network (ANN) nd Specific fuel consumption (SFC). INTRODUCTION It is well known tht fossil fuel reserves ll over the world re diminishing t n lrming rte nd shortge of crude oil is expected within the next few decdes. The world totl nturl gs (NG) reserve s of Jnury, 2007 ws 6,83 Tscf nd bsed on the current consumption rtes, the estimted totl recoverble gs, including proven reserves is dequte for bout 66.7 yers (IEO,2008). This hs resulted in n incresed interest to use CNG s fuel for internl combustion engines. The merits of CNG s n utomotive fuel over conventionl fuels re mny nd presented comprehensively by Nylund et l. (2002) nd Aslm et l. (2003). Due to some of its fvorble physio-chemicl properties, CNG ppers to be n excellent fuel for the sprk ignition (SI) engine. Moreover, SI engines cn be converted to CNG opertion quite esily with the ddition of second fueling system. CNG hs been used in vehicles since 930 s nd the current worldwide NGV popultion is more thn 4.5 million ccording to the Interntionl Assocition for Nturl Gs Vehicle (IANGV) sttistics nd this figure is fst incresing everydy ( Aslm et l., 2006). To investigte experimentlly the performnce of n engine is complex, time consuming nd costly, especilly for studies, which use mny, different blends. Therefore, mthemticl model is used to predict the performnce nd emissions of the engines. But, the resulting ccurcies my not lwys be stisfctory. One lterntive to the mthemticl model is the experimentbsed pproch, such s rtificil neurl-networks (ANNs). Neurl networks re nonliner computer lgorithms, which cn model the behvior of complicted nonliner processes. For the development of high speed digitl computers, the ppliction of ANN pproch could be progressed t very impressive rte. In recent yers, this method hs been pplied to vrious disciplines including utomotive engineering, in forecsting of engine therml chrcteristics for different working conditions. Some reserchers studied this method to predict internl combustion engine chrcteristics. Artificil neurl network pproch hs been used by Yunwng et l. (Aslm et Al., 2006), to nlyze the effect of cetne number on exhust emissions from engine, Lucs et l. (Yunwng et l., 2003), to model Diesel prticulte emission, Hfner et l. (Lucs et l., 200), for diesel engine control design, Shyler et. l. (2000), in utomotive engine mngement systems, Tn nd Sif, (2000), to model the intke mnifold nd 249

2 throttle body processes in n utomotive engine. Those studies do not need n explicit formultion of the physicl reltionships of concerned problems. Severl studies hve lso used ANNs in different engineering res (Sozen et. l., 2005). In the existing litertures, it ws shown tht the use of ANN is powerful modeling tool tht hs the bility to identify complex reltionships from input output dt. However, no investigtion to predict engine specific fuel consumption (SFC, gm/kwh), retrofitted CNG fueled IC engine using ANN pproch ppers to hve been published in the literture to dte. Therefore, the present work investigtes the pplicbility of ANN method for predicting the specific fuel consumption prmeter. Figure : Lyout of the experimentl setup 2. EXPERIMENTAL SETUP AND TEST PROCEDURE The lyout of the experimentl setup hs shown in Figure. The test engine hs been converted from gsoline (Proton Mgm) engine nd hs been equipped with bifuelling system. The min specifictions of the test engine re listed in Tble. Tble : Specifictions of the reserch engine Chrcteristic Displcement Compression rtio Bore Stroke Mx output (kw/rpm) Mx torque (Nm/rpm) Crburetor :Proton Mgm2-Vlve :,468 cc :9.2: :75.5 mm :82 mm :64/6000 :22/3500 :Down-drft 2-brrel An AG 50 (Froude Consine) eddy-current dynmometer hs been used for testing the engine. All the electronic equipment, together with its mnipultive controls nd indictors, etc ws mounted on CP Cdet0 control unit. The engine hs been operted t constnt throttle 30%, 40%, nd 50% nd 00% with vrible speed rnge of RPM t constnt increment of 00 RPM. CNG consumption hs been mesured with Kobold gs flow meter (Model WFM 2705). The CNG flow meter ws incorported with engine control system through interfce crds. A PC-bsed dt cquisition nd control system hs been used for controlling ll the opertion regrding the test where every stge ws llowed to run round 6 8 min with updting dt in every 30 s. Torque, power nd fuel consumption hve been mesured to clculte SFC. 3. ARTIFICIAL NEURAL NETWORKS A widely used NN model clled the multi-lyer perception (MLP) NN is shown in Figure 2. The MLP type NN consists 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. The weights, fter trining, contin meningful informtion, wheres before trining they re rndom nd hve no mening (Erol et l., 2004). 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 j = X iv ij + b j () i = The output of neuron is determined by pplying n ctivtion function to the totl input nd clculted using Eqution (Kreider et l., 992). 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 repetitively for ech set of inputs until there re no mesurble errors, or the totl error is smller thn specified vlue. The following procedures hve been executed in ll the models developed; (i) dtbse collection; (ii) nlysis nd preprocessing of the dt; (iii) trining of the neurl network; (iv) testing the trin network; nd (v) using the trined ANN for simultion nd prediction using trined network. An importnt stge of neurl network is the trining step, in which n input is introduced to the 250

3 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. rnge. Neurons in the input lyer hve no trnsfer function. Logistic sigmoid (logsig) trnsfer function hs been used in hidden lyer while pureliner (purelin) trnsfer function hs been used in output lyer. After the successful trining of the network, the network ws tested with the test dt. Using the results produced by the network, sttisticl methods hve been used to mke comprisons. Tble 2: Dt sets used for trining the network Figure 2: Architecturl grph of Multilyer Perception (MLP) with one hidden lyer 4. APPLICATION OF NEURAL NETWORKS IN THE PRESENT STUDY Three dt sets re needed for ANNs: for trining, vlidtion nd testing the network. The usul pproch is to prepre single dt-set, nd differentite it by rndom selection. In this study, the experimentl results mentioned bove were used to trin, vlidte nd test n rtificil neurl-network. Engine speed (rpm), throttle position, fuel-ir equivlence rtio nd torque (N-m) re used in input lyer while brek specific fuel consumption (gm/kwh) used s output lyer. The lerning lgorithm clled the bck-propgtion ws pplied for the single hidden lyer. Btch grdient descent (GD), resilient bckpropegtion (RP), levenberg-mrqurdt (LM), btch grdient descent with momentum (GDM), vrible lerning rte (GDX), scled conjugte grdient (SCG) lgorithms hve been used for the vrints. The Neurl Network hs been optimized using the MATLAB Version 6.5 Neurl Network Toolbox. In the trining stge, to define the output ccurtely, we tried to increse the number of neurons step-by-step (i.e 3 2) in the hidden lyer. Inputs nd outputs hve been normlized in the rnge of (0. 0.9) s NN works efficiently within this 30%T ht RPM Equivlence rtio, φ Torque (Nm) SFC (gm/kwh)

4 30%Tht Tble 3: Dt sets used for vlidtion 30%Tht RPM Equivlence rtio, φ Torque (Nm) SFC (gm/kwh) %Tht Tble 4: Dt sets use for test network Engine Speed RPM Equivlence rtio, φ Torque (N-m) SFC (gm/kwh) MEASURES OF PREDICTION PERFORMANCE 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 sttisticl nlysis in terms of Root-Men-Squred (RMS), bsolute frction of vrince (R 2 ), s well s men error percentge vlues []. Those re defined bellow: i = N 2 ( E p ) 2 i = ( ) R = (3) I = N 2 E M I = i N = i = = ( E ) RMS (4) N i = N E p Men % Error = 00 (5) N i= E where E -Actul result E p -Predicted result E m -Men vlue N-Number of pttern The coefficient of multiple determintions R 2 compres the ccurcy of the model to the ccurcy of trivil benchmrk model. A perfect fit would result in n R 2 vlue of nd very good fit ner. 6. RESULTS AND DISCUSSIONS The im of using the Artificil Neurl Network (ANN) is to test the bility to predict specific fuel consumption of retrofitted CNG engine. The network hs four input prmeters: Engine speed (rpm), throttle position, fuel-ir equivlence rtio nd torque (N-m) nd one output prmeter: specific fuel consumption. The experimentl dt set includes 84 vlues, of which 44 vlues were used for trining network, 20 vlues were used for vlidtion nd 20 vlues were selected rndomly to test the performnce of the trined network. The experimentl results re shown in Tbles -3. All of the input nd output vlues were normlized into the rnge Vlidtion vlues were used for erly stop of trining nd to void over tinting. After 30 trining cycles the level of error ws stisfctory nd further cycles hd no significnt effect on error reduction of testing. This cn clerly be seen in Fig. 3. In the trining, n incresed number of neurons (from 3 to 2) re used in the hidden lyer to define the output ccurtely for the btch grdient descent (GD), resilient bckpropegtion (RP), levenberg-mrqurdt (LM), btch grdient descent with momentum (GDM), vrible lerning rte (GDX), scled p 2 252

5 conjugte grdient (SCG) lgorithms. The testing ccurcy of trined networks is shown in Figs performnce ccurcy nd consistency with chnging number of hidden neurons. The best network ws found to be the LM lgorithm with 0 hidden neurons. In Tble 4, the sttisticl vlues of the outputs for this lgorithm hve been shown for the trining, vlidtion nd testing dt. The ctul nd predicted outputs of trining nd testing hve been shown grphiclly in Figs The ANN predictions for the BSFC yield men reltive error of 0.224%, root men squre error of gm/kwh nd correltion coefficient of These vlues show tht the ANN predicts the BSFC quite well despite wide rnges of operting conditions. It is cler tht the performnce of the ANN would hve been even better, if higher number of test runs hd been performed to provide lrger mount of experimentl dt for the network trining Figure 3: Network trining cycles Tble 4: Performnce of optimized network R 2 RMS Men (gm/kwh) Error (%) Trining Vlidtion Test RMS Number of hidden neurons RP LM GDX SCG Frction of vrince RP LM GDX SCG Number of hidden neurons Figure 4: R 2 vlue of test for different lgorithms with incresing number of hidden lyer. The performnces of GD nd GDM were not in stisfctory level nd the sttisticl vlues were out of the rnge for Figs The ccurcies of lgorithms RP, GDX nd SCG hve shown good result but not consistent with hidden neurons. LM lgorithm hs shown good Figure 5: RMS vlue of test for different lgorithms with incresing number of hidden lyer. Men percentge error (%) RP LM GDX SCG Number of hidden neurons Figure 6: Men percentge error of test for different lgorithms with incresing number of hidden lyer 253

6 SFC(gm/kWh) Actul Predicted Figure 7: Comprison of ctul nd predicted vlues for BSFC of trining dt Tble 5. The weights (C) between input lyer nd hidden lyer for BSFC i E i = C T ht + C 2 N s + C 3 φ + C 4 T + C 5 C C 2 C 3 C 4 C CONCLUSION SFC (gm/kwh) Actul Predicted Figure 8: Comprison of ctul nd predicted vlues for BSFC test dt. The formultions of the outputs obtined from the weights re given using Eqs BSFC = + e ( F 0.048F F3.4424F F F F F F F ) Where F i = ( i =,2,3.0) cn be clculted ccording to eqution (7). ACKNOWLEDGEMENT (6) The im of this pper hs been to show the possibility of using the neurl networks for predictions of dul fuel engine performnce. The network produces the predicted results of brke specific fuel consumption prllel to the experimentl ones. The RMS error vlues re smller thn 0.05 gm/kwh, R 2 vlues re bout nd men error smller thn 0.25%, which my esily be considered within the cceptble rnge. A bck propgtion (BP) neurl network model with GD, RP, LM, GDM, GDX nd SCG lgorithms hve been studied in single hidden lyer. Number of neurons on hidden lyer lso vried to optimize network. In most cses, the best results were obtined from the LM lgorithm. On the other hnd GD nd GDM lgorithoms showed very poor prediction performnce. The overll results show tht the networks cn be used s n lterntive for predicting the performnces of CNG fueled internl combustion engine. The result of this study shows tht ANN hs bility to lern nd generlize wide rnge of experimentl conditions. Therefore, the usge of ANNs my be highly recommended to predict the engine performnce insted of hving to undertke complex nd time-consuming experimentl studies. The uthors would like to thnk the Ministry of Science nd Technology nd University of Mly (IRPA F i = + e i (7) 03-30) for providing the fund to crry out the work reported in this pper. The uthors would lso like to express their grtitude to ny one who hd ever Where E i is the weighted sum of the of the input nd is given bycontributed to their work either directly or indirectly eqution s seen in the Tbles

7 REFERENCE Aslm M.U, Msjuki H.H, Klm M.A, Abdesselm H., Mhli T.M.I, Amlin M.A., An experimentl investigtion of CNG s n lterntive fuel for retrofitted gsoline vehicle. Fuel,. vol 85, pp Aslm MU, Msjuki HH, Mleque MA, Klm MA, Mhli TMI, Zinon Z, Introduction of nturl gs fueled utomotive in Mlysi. Proc. TECHPOS , UM, Mlysi. Erol Arckhoglu, Abdullh Cvusoglu, Ali Erisen Thermodynmic nlyses of refrigernt mixtures using rtificil neurl-networks. Applied Energy. vol. 78, pp Interntionl Energy Outlook, Interntionl Energy Outlook Energy informtion dministrtion. Wshington, DC: Deprtment of Energy; Dter 5/08/2008 Kreider JF, Wng XA Artificil neurl networks demonstrtions for utomted genertion of energy use predictors for commercil buildings. ASHRAE Trnsctions.. vol. 97(), pp Nylund N.O., Lurikko J., Ikonen M., Pthwys for nturl gs into dvnced vehicles. IANGV (Interntionl Assocition for Nturl Gs Vehicle) Edited Drft Report Shyler P. J, Goodmn M., M T., 2000 The exploittion of neurl networks in utomotive engine mngement systems, Eng. Appl. Artif. Intell. vol. 3, pp Sozen A, Arckhoglu E., Prediction of solr potentil in Turkey. Appl Energ, vol. 80, pp Tn Y., Sif M., Neurl-networks-bsed nonliner dynmic modeling for utomotive engines, Neurocomputing. vol. 30, pp Yunwng D., Meilin Z., Dong X., Xiobei C., An nlysis for effect of cetne number on exhust emissions from engine with the neurl network, Fuel. vol 8, pp Lucs A., Durn M., Crmon, M. Lpuert M, 200. Modeling diesel prticulte emissions with neurl networks, Fuel, vol. 4: pp