Applied Energy 87 (2010) 349355

Contents lists available at ScienceDirect

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

Prediction of torque and specic fuel consumption of a gasoline engine by using articial neural networks
Necla Kara Togun *, Sedat Baysec
Department of Mechanical Engineering, University of Gaziantep, Gaziantep, Turkey

a r t i c l e

i n f o

a b s t r a c t
This study presents an articial neural network (ANN) model to predict the torque and brake specic fuel consumption of a gasoline engine. An explicit ANN based formulation is developed to predict torque and brake specic fuel consumption of a gasoline engine in terms of spark advance, throttle position and engine speed. The proposed ANN model is based on experimental results. Experimental studies were completed to obtain training and testing data. Of all 81 data sets, the training and testing sets consisted of randomly selected 63 and 18 sets, respectively. An ANN model based on a back-propagation learning algorithm for the engine was developed. The performance and an accuracy of the proposed ANN model are found satisfactory. This study demonstrates that ANN is very efcient for predicting the engine torque and brake specic fuel consumption. Moreover, the proposed ANN model is presented in explicit form as a mathematical function. Crown Copyright 2009 Published by Elsevier Ltd. All rights reserved.

Article history: Received 9 April 2009 Received in revised form 22 June 2009 Accepted 11 August 2009 Available online 17 September 2009 Keywords: Gasoline engine Torque Brake specic fuel consumption Neural networks Explicit solution

1. Introduction An engineering phenomenon may embed complicated physical, chemical or electrical theory and may require very complicated arithmetic to describe them, yet, arithmetic emerged may not be solvable in closed form. Articial neural network (ANN) is an alternative technique for providing a relationship between the variable quantities of interest. ANN requires only a set of experimental results, numerical in nature and describes the relation by analyzing them. In other words, it only needs solution examples concerning the problem. ANN techniques require a lot of arithmetic basically of trial-end-error nature, involving numerical differentiation, integration, noise rejection, etc., and are never feasible without fast computation facilities. Advent of digital computers providing high speed arithmetic and vast amounts of data storage has given rise to the application of ANN techniques to many engineering problems. In recent years, this method has been applied to various disciplines including automotive engineering, in forecasting of engine characteristics for different working conditions. The relationship between the temperature of the exhaust gases and fuel consumption of an internal consumption engine has been studied in [1]. ANN approach has been used in another study, to analyze the effect of cetane number on exhaust emissions from the engine [2], and also in [3], to model diesel particulate emission [4,5] are remarkable studies where ANN is used to forecast gaso-

* Corresponding author. Tel.: +90 342 3172534; fax: +90 342 3601104. E-mail address: nkara@gantep.edu.tr (N. Kara Togun).

line consumption and the effects of intake valve timing on engine performance and fuel economy respectively. Similarly in [6] the effect of throttling is studied, taking the intake manifold geometry into consideration. Numerous studies have been undertaken to predict the performance and exhaust emission characteristics of internal combustion engines by using ANNs [711]. Studies done by neural networks and genetic algorithms have been used to predict and reduce diesel engine emissions [12]. Neural networks have been found to be the domain for numerous successful applications of prediction tasks, in modeling and prediction of energy-engineering systems [13], prediction of the energy consumption of passive solar buildings [14], and modeling a burner heated catalytic converter during cold start in a four stroke, spark ignition engine [15]. And also articial neural network technique has been applied to control the air fuel ratio of the engine [16,17] and exhaust gas recirculation control [18]. In this study, an explicit ANN based formulation was developed to predict torque and brake specic fuel consumption of a gasoline engine in terms of spark advance, throttle position and engine speed. Experimental studies were completed to obtain training and testing data. The experimental data from totally 81 test runs was used to train and test the ANN model for predicting torque and brake specic fuel consumption. Inputs for the network were the spark advance, throttle position and engine speed, while the outputs were torque and brake specic fuel consumption. The experimental study to determine torque and fuel consumption characteristics in a gasoline engine is complex, time consuming and costly. It also requires specic instrumentation. To overcome

0306-2619/$ - see front matter Crown Copyright 2009 Published by Elsevier Ltd. All rights reserved. doi:10.1016/j.apenergy.2009.08.016

350

N. Kara Togun, S. Baysec / Applied Energy 87 (2010) 349355

Nomenclature ANN BPA BSFC CA LM MAPE articial neural network back propagation algorithm brake specic fuel consumption crank angle LevenbergMarquardt mean absolute percentage error N R SA T TP engine speed correlation coefcient spark advance torque throttle position

these difculties, an ANN can be used for the prediction of performance in a gasoline engine. The proposed ANN approach is quite accurate, fast and practical.

2. Description of the experimental setup In this study, a 1400 cc, four cylinder Fiat gasoline engine is used. Details of the engines specications are shown in Table 1. Fig. 1 shows the schematic representation of the experimental setup and also the measuring devices on it.

Table 1 Test engine specications. Type Engine type Fuel type Swept volume (cm3) Cylinder bore (mm) Cylinder stroke (mm) Compression ratio Number of cylinders Cooling type Fuel supply system Maximum torque Maximum power Engine position Fiat Tofas 131 Four-stroke Gasoline 1297 76 71.5 7.8:1 4 Water cooled in closed circuit Naturally aspirated carburetor 12.5 kgm at 3000 rpm 52.2/70 kW/HP at 5500 rpm Vertical

Calibration checks of the devices were made two times, one before and one after each successive test. To transfer data into a computer, data logger systems were used. All measurements were conducted under steady state conditions. The measurement was not started until engine runs faultless. Since there were many measurement points, three different PC data logger combinations were used to reduce error resulting from the measurements. The functions of computers, measuring equipments and methods of measurement are briey explained as follows to obtain mass ow rate of air, fuel entering the combustion chamber and engine speed. Air ow was determined by difference in pressure. A high volume air tank, which was connected to the engine carburetor with a plastic pipe, was used to draw the air ow. Pressure difference formed by a pipe bowl at the entrance of the air tank was measured by a pressure transducer having analog output. Another analog/digital converter was used to convert this analog output into the digital signal and to transfer the signal to the PC for a denite time interval. Mass ow rate of the fuel was measured by using one way RS 232 interface and a weighing machine having a 0.1 g sensitivity and overloading security. The output of the weighing machine was transferred to the PC by RS 232 interface. These data were read by software for a denite time interval and suitable format for calculations. Engine speed is controlled by the position of the throttle valve. To control the throttle position, the buttery is directly coupled to

Fig. 1. Schematics of test engine and setup: (1) air tank, (2) orice, (3) inclined manometer, (4) data logger, (5) exhaust gas analyzer, (6) dynamometer, (7) transducer, (8) tachometer, (9) exhaust manifold, (10) carburetor, (11) weighing device, (12) fuel container, (13) computer.

N. Kara Togun, S. Baysec / Applied Energy 87 (2010) 349355 Table 2 The measuring uncertainties in engine performance parameters. Engine performance characteristics Specic fuel consumption Torque Maximum error, % 0.31 0.81

351

summation of inputs and corresponding weights u, given as follows:

ui

n X j1

wij xj bi

the shaft of a stepper motor. Position of the stepper motor is controlled in 1.8 increments by the control computer via a conventional software. The experimental work in this investigation was performed at various spark advance, throttle position and engine speed conditions. Before starting the engine, the spark advance has been adjusted to 10 crank angle (CA), which is a predened design value of the engine. To start with, our computer-controlled gasoline engine, which is connected to hydraulic dynamometer, has been loaded, with the 50% throttle position. The engine has been tested in the ranges of 35001500 rpm at intervals of 250 rpm. Torque, fuel ow rate, air ow rate and specic fuel consumption have been recorded. Similarly, these measurements have been repeated for the throttle positions of 75% and 100%. After the measurements, the engine has been left to cool. Then the ignition time has been adjusted to successively 5 and 0 CA respectively and the above procedure has been repeated. The uncertainties in calculated characteristics with respect to measured parameters are shown in Table 2. 3. Neural networks (NN) Haykn denes a neural network as a massively parallel distributed processor. It has an inherent tendency for storing experimental knowledge and making it available for use. It resembles the human brain in two respects: the knowledge is acquired by the network through a learning process, and inter-neuron connection strengths known as synaptic weights are used to store the knowledge [19]. Neural network operates like a black box model, and does not require detailed information about the system. Instead, it learns the relationship between the input parameters and the controlled and uncontrolled variables by studying previously recorded data, in a similar way that a non-linear regression might be performed. Another advantage of using ANNs is their ability to handle large and complex systems with many interrelated parameters. They simply ignore excess input data that are of minimal signicance and concentrate instead on the more important inputs [14]. A neural network is composed of large numbers of highly interconnected processing elements known as neurons. The basic elements of an articial neuron are shown in Fig. 2. Articial neuron consists of weight, bias and activation function mainly. Each neuron receives inputs x1, x2,. . ., xn, attached with a weight xi which shows the connection strength for a particular input for each connection. Every input is then multiplied by the corresponding weight of the neuron connection. A bias bi can be dened as a type of connection weight with a constant nonzero value added to the

The summation ui is transferred using a scalar-to-scalar function called an activation or transfer function, f(ui), to yield a value called the units activation, given as

yi f ui

Activation functions serve to introduce nonlinearity into neural networks which makes it more powerful than linear transformation. A back propagation algorithm (BPA), which is the most widely used training algorithm for the multi layer perception [20], is a gradient descent technique to minimize the error for a particular training pattern. Accordingly, for a given input pattern, a ow of activation is forwarded from the input layer to the output layer via hidden layer(s). Then the errors in the output are initiated. BPA is used to adjust the weights, a small amount at a time, in a way that reduces the error. The training of the network is accomplished by adjusting the weights and is carried out through a large number of training sets and training cycles (epochs). The goal of the learning procedure is to nd the optimal set of weights, which in the ideal case would produce the right output for any input. The output of the network is compared with a desired response to produce an error. Once the ANN is adequately trained, it can generalize to similar cases, which it has never seen. Detailed information about ANN and its working principles can be found in ref [20]. 4. Articial neural network model and parameters The main focus of this study is the prediction of torque (T) and brake specic fuel consumption (BSFC) using ANN based on experimental results described above. Spark advance, throttle position and engine speed have been used as input layer components, while the T and BSFC were used separately as output layer components of the ANNs. In the ANN model, the experimental data set includes 81 values, of which 63 values were used for training the network and 18 values were selected randomly to test the performance of the trained network. In this study, a computer program has been developed and performed under Matlab. The back-propagation learning algorithm has been used in feed forward with one hidden layer. The input layer neurons receive information from the outside environment and transmit them to the neurons of the hidden layer without performing any calculation. The hidden layer neurons then process the incoming information and extract useful features to reconstruct the mapping from the input space. The neighboring layers are fully interconnected by weights. Finally, the output layer neurons produce the network prediction to the outside world. One of the most important tasks in ANN studies is to determine the optimal network architecture which is related to the number of hidden layers and neurons in it. Generally, the trial-and-error ap-

x1 x2 x3 xi

wi1 wi2 wi3 wij

bias
u i = wij x j + bi
j =1 n

Activation Function
y i = f (u i )

Output

. . .

Fig. 2. Basic elements of an articial neuron.

352

N. Kara Togun, S. Baysec / Applied Energy 87 (2010) 349355

Bias

Bias

SA TP

N
Input layer

. . . .

Tork

Output layer

BSFC

Hidden layer
Fig. 3. The architecture of proposed NN model.

proach is used. In this study, the best architecture of the network was obtained by trying different number of hidden layers and neurons. The trial started on hidden layer with seven neurons, and the performance of each network was checked by correlation coefcient (R). The goal is to maximize correlation coefcient to obtain a network with the best generalization. Many different network models were tried and their R values were calculated. The highest correlation coefcient for both T and BSFC was obtained at a network. Based on this analysis, the optimal architecture of the ANN was constructed as 3131 NN architecture for T and 3151 NN architecture for BSFC representing the number of inputs, neurons in hidden layers, and outputs, respectively. The proposed ANN model is given in Fig. 3. The learning algorithm used in the study is LevenbergMarquardt (LM), activation function is logistic sigmoid (logsig) transfer functions and number of epochs is 10,000.

Table 3 Range of inputoutput parameters in trainingtesting phase and normalization values. Parameters Range of values Min SA TP (%) N (rpm) T (N m) BSFC (g/kW h) 0 50 1500 40.7 248.41 Max 10 100 3500 92.8 458.49 10 100 3500 30 500 Normalization value

5. Analysis results The ANN model developed in this study is used to predict the T and BSFC based on the spark advance, throttle position and engine speed. A total of 63 samples were used for training the network

both T and BSFC. The proposed ANN model for T and BSFC prediction had correlation coefcients of 0.9997 and 0.9971, respectively, for training data sets, 0.99505 and 0.98331, respectively, for testing data sets. Moreover, MAPE of the T prediction was about 0.2912 and 1.74 for the training and testing set, respectively. Similarly, MAPE of the BSFC prediction was about 1.0186 and 2.7588 for the training and testing set, respectively. As it is seen these MAPE are fairly reasonable. Figs. 4 and 5 also demonstrated that the ANN was quite successful in learning the relationship between the different input parameters and the outputs (T and BSFC). The result

T 1e

30

0:7479 0:6726 3:0776 0:7687 0:6354 1:5951 0:3399 0:5024 0:2028 1:9807 3:1996 0:1289 0:1742 0:3153 1eU1 1eU2 1eU3 1eU4 1eU5 1eU6 1eU7 1eU8 1eU9 1eU10 1eU11 1eU12 1eU13

and other 18 (randomly) were used as a test set. Data set was normalized using a simple normalization method. The range of the samples and normalization values are given in Table 3. The performance of the proposed ANN model was plotted in Figs. 4 and 5 for both T and BSFC, respectively. It was observed that a high prediction. Capability was achieved for both training and testing data sets of T and BSFC. Therefore, the ANN appears to have a high generalization capability. The statistical values for both T and BSFC in training and test sets are given in Table 4. As seen in Table 4, a high correlation coefcient (R) and a low mean absolute percentage error (MAPE) were obtained for the training and testing data sets for

of testing phase in Figs. 4 and 5 showed that the ANN was capable of generalizing between input variables and output reasonably well. Figs. 6 and 7 demonstrate the percentage errors of training and testing sets for the predicted T and BSFC. ANN applications are treated as black-box applications in literature. However, this study opens this black box and introduces the ANN application in a closed form solution. This study aims to present the closed form solution of T and BSFC based on the trained ANN parameters (weights and biases) as a function spark advance (SA), throttle position (TP) and engine speed (N). Using weights and biases of trained ANN model, engine torque can be given as follows:

N. Kara Togun, S. Baysec / Applied Energy 87 (2010) 349355

353

a 100
90 80 R (NN) 70 60 50 40 40

R=0.9997

500 R=0.9971 450 400 350 300 250 250

50

60

70 R (EXP)

80

90

100

R (NN)

300

350 R (EXP)

400

450

500

Fig. 4. Prediction of NN and actual values for training sets: (a) T and (b) BSFC.

a 100
90 80 R (NN) 70 60 50 40 40

R=0.99505

500 R=0.98331 450 400

R (NN) 50 60 70 R (EXP) 80 90 100

350 300 250 200 200

250

300

350 R (EXP)

400

450

500

Fig. 5. Prediction of NN and actual values for testing sets: (a) T and (b) BSFC.

Table 4 Statistical parameters of train and test sets. MSE Test set Train set Brake specic fuel Test set consumption (BSFC) Train set Torque (T) 0.002 0.0001 0.0005 0.0001 MAPE 1.74 0.2912 2.7588 1.0186 Correlation coefcient R 0.99505 0.9997 0.98331 0.9971

where Ui are given in Eq. (7) for the BSFC.

U i C 5i SA C 6i TP C 7i N C 8i

where, the constants (Cji) are given in Table 5 for LM algorithm with 13 neurons for T and 15 neurons for BSFC. It should be noted that the proposed explicit formulation of the NN model presented above is valid for the ranges of training set. 6. Conclusion

where the transfer function used for this approach is given as follows:

Fi

1 1 eUi

where Ui are given in Eq. (5) for the torque

U i C 1i SA C 2i TP C 3i N C 4i
Similarly, BSFC can be found as follows:

This study presents a new and efcient approach for the prediction of torque and brake specic fuel consumption of a gasoline engine using ANNs. The database used for NN training is based on experimental results. It used 63 results as data sets to train the network, while 18 results were used as test data from the total of 81 experimental results. To train the network, spark advance, engine speed, and throttle positions are used as the input layer, while outputs are the engine torque and brake specic fuel consumption. Back-propagation NNs are used for the training process. The proposed NN model shows perfect agreement with experimental results.

BSFC 1e

500
0:0417 0:0597 0:0831 0:1335 0:1716 0:1525 0:0784 0:0279 0:0659 0:1134 0:0224 0:1295 0:3636 0:2162 0:3334 0:7102 1eU1 1eU2 1eU3 1eU4 1eU5 1eU6 1eU7 1eU8 1eU9 1eU10 1eU11 1eU12 1eU13 1eU14 1eU15

354

N. Kara Togun, S. Baysec / Applied Energy 87 (2010) 349355

1.5 1 0.5

4 3 2

% Error

% Error

1 0 -1 -2

0 -0.5 -1 -1.5

10

20

30 40 Exp. No.

50

60

70

-3

10

20

30 40 Exp. No.

50

60

70

Fig. 6. Percentage error of training set: (a) T and (b) BSFC.

4 3 2 1

8 6 4

% Error

0 -1 -2 -3 -4 -5 0 2 4 6 8 10 12 Exp. No. 14 16 18 20

% Error

2 0 -2 -4 -6 -8 0 2 4 6 8 10 12 Exp. No. 14 16 18 20

Fig. 7. Percentage error of testing set (a) T (b) BSFC.

Table 5 The weights and biases between input layer and hidden layer for Eqs. (3) and (6). T i 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 C1i 1.2725 1.741 0.0061 1.64 0.343 0.0533 0.8738 2.4591 0.6457 0.0365 0.013 0.7099 0.1032 C2i 0.1925 0.1166 0.0932 0.0867 0.2451 0.1257 0.4051 0.1951 0.2505 0.2125 0.1984 0.1447 0.0824 C3i 0.0034 0.0006 0.0023 0.001 0.002 0.002 0.0049 0.0032 0.0032 0.0033 0.0015 0.0069 0.0052 C4i 11.8911 27.5182 3.597 4.9378 15.8767 0.4814 5.5608 5.0804 28.016 30.006 12.8857 24.1371 9.4364 BSFC C5i 0.1653 0.3659 0.7941 0.2102 1.4948 0.857 1.1732 0.6916 0.9153 1.3535 0.2564 0.534 0.7501 1.1861 0.8518 C6i 0.1768 0.1835 0.2073 0.0125 0.0218 0.0686 0.1451 0.0152 0.144 0.1177 0.1639 0.071 0.0331 0.1221 0.0531 C7i 0.005 0.0057 0.0018 0.0063 0.005 0.0049 0.003 0.0053 0.0042 0.0027 0.0054 0.0053 0.0041 0.0012 0.0041 C8i 9.222 18.5103 13.0825 11.3466 2.7369 12.0769 5.6508 19.3996 24.4002 13.664 7.3443 10.0671 13.0804 12.5888 15.8735

The proposed ANN model for torque and brake specic fuel consumption had correlation coefcients of 0.9997, 0.9971, respectively, for training data sets, 0.99505 and 0.98331, respectively, for testing data sets. Moreover, MAPE of the T prediction was about 0.2912 and 1.74 for the training and testing set, respectively. Similarly, MAPE of the BSFC prediction was about 1.0186 and 2.7588

for the training and testing set, respectively which may easily consider within the acceptable range. The explicit formulation of torque and brake specic fuel consumption based on the proposed NN model is also obtained and presented. As a result the proposed NN model has strong potential as feasible tools for prediction of torque and brake specic fuel consumption. And also the engine

N. Kara Togun, S. Baysec / Applied Energy 87 (2010) 349355

355

torque and brake specic fuel consumption can be determined using models with ANN methodology. References
[1] Parlak A, Islamoglu Y, Yasar H, Egrisogut A. Application of articial neural network to predict specic fuel consumption and exhaust temperature for a diesel engine. Appl Therm Eng 2006;26:8248. [2] Yuanwang D, Meilin Z, Dong X, Xiaobei C. An analysis for effect of cetane number on exhaust emissions from engine with the neural network. Fuel 2002;81:196370. [3] Lucas A, Durn A, Carmona M, Lapuerta M. Modeling diesel particulate emissions with neural networks. Fuel 2001;80:53948. [4] Nasr GE, Badr EA, Joun C. Backpropagation neural networks for modeling gasoline consumption. Energy Convers Manage 2003;44:893905. [5] Glc M, Sekmen Y, Erduranli P, Salman MS. Articial neural network based modeling of variable valve-timing in a spark ignition engine. Appl Energy 2005;81:18797. [6] Tan Y, Saif M. Neural network based nonlinear dynamic modeling for automotive engines. Neurocomputing 2000;30:12942. _ [7] Arcaklioglu E, elikten I. A diesel engines performance and exhaust emissions. Appl Energy 2005;80:1122. [8] Thompson GJ, Atkinson CM, Clark NN, Long TW, Hanzevack E. Neural network modeling of the emissions and performance of a heavy duty diesel engine. P I Mech Eng D-J Aut 2000;214:11126. [9] Ycesu HS, Sozen A, Topgl T, Arcaklioglu E. Comparative study of mathematical and experimental analysis of spark ignition engine performance used ethanol-gasoline blend fuel. Appl Therm Eng 2007;27: 35868.

_ [10] Sayin C, Ertunc HM, Hosoz M, Kilicaslan I, Canakci M. Performance and exhaust emissions of a gasoline engine using articial neural network. Appl Therm Eng 2007;27:4654. [11] Krijnsen HC, Kooten WEJ, Calis HPA, Verbeek RP, Bleek CM. Prediction of NOx emissions from a transiently operating diesel engine using an articial neural network. Chem Eng Technol 1999;22(7):6017. [12] Alonso JM, Alvarruiz F, Desantes JM, Hernndez L, Hernndez V, Molt G. Combining neural networks and genetic algorithms to predict and reduce diesel engine emission. IEEE T Evolut Comput 2007;11:4655. [13] Kalogirou SA. Applications of articial neural-networks for energy systems. Appl Energy 2000;67:1735. [14] Kalogirou SA, Bojic M. Articial neural-networks for the prediction of the energy consumption of a passive solar-building. Energy 2000;25: 47991. [15] Akcayol MA, Cinar C. Articial neural network based modeling of heated catalytic converter performance. Appl Therm Eng 2005;25:234150. [16] Wang SW, Yu DL, Gomm JB, Page GF, Douglas SS. Adaptive neural network model based predictive control for air fuel ratio of SI engines. Eng Appl Artif Intel 2006;19:189200. [17] Yazdanpanah MJ, Kalhor A. Air/fuel ratio control in SI engines using a combined neural network and estimator. IEEE CCA 2003:34752. [18] Hongwei C. Exhaust gas recirculation control in a spark-ignition LPG engine using neural networks. WCICA Dalian China, 2006; 2123 June 2006, 633235. [19] Haykin S. Neural networks: a comprehensive foundation. New York: Macmillan; 1994. [20] Eyercioglu O, Kanca E, Pala M, Ozbay E. Prediction of martensite and austenite start temperatures of the Fe-based shape memory alloys by articial neural networks. J Mater Process Tech 2008;200:14652.