Artificial neural network based modeling of heated ... - CiteSeerX

ARTICLE IN PRESS

Applied Thermal Engineering xxx (2005) xxx–xxx www.elsevier.com/locate/apthermeng

Artificial neural network based modeling of heated catalytic converter performance M. Ali Akcayol

a,*

, Can Cinar

b

a

b

Department of Computer Engineering, Faculty of Engineering & Architecture, Gazi University, Maltepe, 06570 Ankara, Turkey Department of Mechanical Education, Faculty of Technical Education, Gazi University, Besevler, 06500 Ankara, Turkey Received 23 August 2004; accepted 17 December 2004

Abstract Catalytic converters are the most effective means of reducing pollutant emissions from internal combustion engines under normal operating conditions. But the future emission requirements cannot be met by three way catalysts (TWC) as they cannot effectively remove hydrocarbon (HC) and carbon monoxide (CO) emissions from the outlet of internal combustion engines in the cold-start phase. Therefore, significant efforts have been put in improving the cold-start behavior of catalytic converters. In the experimental study, to improve cold-start performance of catalytic converter for HC and CO, a burner heated catalyst (BHC) has been tested in a four stroke, spark ignition engine. The modeling of catalytic converter performance of the engine during cold start is a difficult task. It involves complicated heat transfer and processes and chemical reactions at both the catalytic converter and exhaust pipe. In this study, to overcome these difficulties, an artificial neural network (ANN) is used for prediction of catalyst temperature, HC emissions and CO emissions. The training data for ANN is obtained from experimental measurements. In comparison of performance analysis of ANN, the deviation coefficients of standard and heated catalyst temperature, standard and heated catalyst HC emissions, and standard and heated catalyst CO emissions for the test conditions are less than 4.925%, 1.602%, 4.798%, 4.926%, 4.82% and 4.938%, respectively. The statistical coefficient of multiple determinations for the investigated cases is about

*

Corresponding author. Tel.: +90 312 231 74 00/2123; fax: +90 312 230 84 34. E-mail address: [email protected] (M. Ali Akcayol). URL: http://w3.gazi.edu.tr/~akcayol (M. Ali Akcayol).

1359-4311/$ - see front matter Ó 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.applthermaleng.2004.12.014

ARTICLE IN PRESS

2

M. Ali Akcayol, C. Cinar / Applied Thermal Engineering xxx (2005) xxx–xxx

0.9984–0.9997. The degree of accuracy is acceptable in predicting the parameters of the system. So, it can be concluded that ANN provides a feasible method in predicting the system parameters. Ó 2005 Elsevier Ltd. All rights reserved. Keywords: Artificial neural network; Catalytic converter; Cold start

1. Introduction Because of the growing number of vehicles running all over the world, the problem of urban air pollution has been gained so much importance [1]. The spark ignition engine exhaust gases contain nitrogen oxides (NOx), CO and organic compounds, which are unburned or partially burned HCs. CO and HC occur because the combustion efficiency is lower than 100% due to incomplete mixing of the gases and the wall quenching effects of the colder cylinder walls. The NOx is formed during the very high temperatures of the combustion process [2,3]. Improvements in engine design, microprocessor controlled fuel injection and ignition systems have been substantially reducing the pollutant emissions for two decades in spark ignition engines. However, further reductions in exhaust emissions can be obtained by removing pollutants in the exhaust system. TWCÕs that controls the pollutant emissions of HC, CO and NOx are an effective way to reduce exhaust emissions [3,4]. But the requirements of future emission standards cannot be met by conventional TWC, as they cannot efficiently remove HC and CO from the outlet of internal combustion engines in the cold-start phase [5,6]. The efficiency of a catalytic converter is very dependent on temperature. Until it reaches lightoff temperature at which a converter becomes 50% efficient, 50% to 80% of the regulated HC and CO emissions are emitted from the tailpipe [7–9]. When the engine first starts, both the engine and catalyst are cold. After startup the heat of combustion is transferred from the engine and the exhaust piping begins to heat up. Finally, the temperature is reached within the catalyst that initiates the catalytic reactions. This light-off temperature and the concurrent reaction rate is kinetically controlled; i.e. depends on the chemistry of the catalyst since the transport reactions are fast [2]. In order to reduce cold-start emissions, special techniques have been developed. These techniques are referred to as fast light-off techniques. Among the more successful methods that have been developed for shortening the light-off time are locating of the converter closer to the exhaust manifold, secondary air injection, electrically and burner heated catalysts [1,4,7,10–16]. Traditionally, the processing and understanding of the experimental outputs of the catalytic converter performances was investigated by the researchers. But there are large number of variables and application of complex optimization algorithms for the experimental design makes difficult the direct human interpretation of data derived from high throughput experimentation [17]. Yasgashi et al. [18] presented a simulation technique to optimize the heating pattern of an electrically heated catalytic converter. Koltsakis et al. [19] performed a 2-D model for a TWC to investigate the effects of operating conditions. But none of these models consider the actual catalytic converter performance during cold start. In the last decade, ANNs have been widely used for many different industrial areas such as control, prediction, pattern recognition, classification, speech and vision. ANNs have been trained to solve nonlinear and complex problems that are not exactly modeled mathematically [20]. ANNs

ARTICLE IN PRESS


3

eliminate the limitations of the classical approaches by extracting the desired information using the input data. Applying ANN to a system needs sufficient input and output data instead of a mathematical equation. ANN can be trained using input and output data to adapt to the system. Also, ANNs can be used to deal with the problems with incomplete and imprecise input data [21,22]. ANN has successfully been applied to modeling and design of catalytic converters by researchers [23–26]. Botsaris et al. [23] presented an estimation of a TWC performance with ANN. This study was performed using data sets from two kind of ceramic catalysts a brand new and old one on a laboratory bench at idle speed. Huang at al. [24] developed a new method for catalyst design based on ANN. It was developed to simulate the relations between catalyst components and catalytic performance. Rodemerck et al. [27] developed an ANN model for establishing relationships between catalyst compositions and their catalytic performance. In this study, an ANN has been used for modeling a burner heated catalytic converter during cold start in a four stroke, spark ignition engine. The ANN predicted and experimental results are extensively compared under different operating conditions.

2. Measurement of experimental data The experimental study was conducted on a Ford-MVH.416-ZETEC gasoline engine with a TWC. The engine is four-cylinder, four stroke engine with a swept volume of 1597 cc. The general specifications of the engine are shown in Table 1. The catalytic converter used is a two piece TWC of 0.75 l, each with Johnson Matthey (JM) wash coat. Exhaust emission was measured by Sun MGA-1200 type emission analyzer device. Before the experiments the analyzer was calibrated. The schematic view of the test equipments is shown in Fig. 1. Temperatures were measured with temperature measuring system having 1 degree of Celsius accuracy and brand name ELIMKO6000. Thermocouples were NiCr–Ni type and can measure up to 1200 °C. In the experimental study, catalytic converter temperature, HC and CO emission variations were measured during cold-start period of the engine with standard and burner heated catalytic converter, under idle operating conditions (950 rpm). Before the engine was started, the burner was activated to heat up the catalytic converter. The burner was located in front of the catalytic converter (Fig. 1). Liquefied petroleum gas (LPG) was burned in the burner to heat the main catalyst faster. The preheat temperature of the catalyst was 150 °C. Data recording was commenced when the engine was started (0 s) and continued for 1000 s. All the tests were performed for cold Table 1 General specifications of the Ford-MVH 416 ZETEC engine Item

Specification

Displacement (cc) Cylinder arrangement Valve train Bore (mm) Stroke (mm) Compression ratio

1597 In-line 4 DOHC 76 88 10.3:1

ARTICLE IN PRESS

4


Fig. 1. Schematic view of the test equipments.

starts of the engine. Also, the engine was shut down for 10 h before the next test to provide an ambient temperature start.

3. Application of ANN There are many types of ANN architectures in the literature; however, multi-layer feed-forward neural network is the most widely used for prediction. A multi-layer feed-forward neural network typically has an input layer, an output layer, and one or more hidden layers [28]. In multi-layer feed-forward networks, neurons are arranged in layers and there is a connection among the neurons of other layers. The input signals are applied to the input layer, the output layer contributes to the output signal directly. Other layers between input and output layers are called hidden layers. Input signals are propagated in gradually modified form in the forward direction, finally reaching the output layer. One neuron can receive signals from other neurons as input and produce output signal using transfer function. A sigmoid function is widely used for transfer function [29] whose output lies between zero and unity and is defined as f ðxÞ ¼

1 1 þ ex

ð1Þ

During learning, the weights of the neurons are adjusted according to the generalized delta rule which is the learning algorithm for a back-propagation multi-layered feed-forward network. The error is the sum of the squares of the overall errors of the network and is minimized by the generalized delta rule, defined as X ðy p op Þ2 ð2Þ Ep ¼ p

where Ep is the sum of square errors over all output units, p is the index of pattern in the training set, o is the desired output and y is the calculated output of network. The weight modification for a neuron is done in proportion to the gradient of Ep with respect to the neuron weights [30]. In this way, each updated weight in a layer depends on all the error terms of the output layer. Thus, the error of the output layer is propagated back to each layer. Faster learning can be done by changing the learning rate constant, but improper learning rate constant may cause the weights to bounce around the local minima, thus failing to learn properly.

ARTICLE IN PRESS


5

A four layers ANN is applied to the system to predict of catalyst temperature, HC emissions and CO emissions. The ANN structure used in this application is shown in Fig. 2. The ANN has four layers namely, an input, an output, and two hidden layers. The input layer consists of one neuron, the output layer consists of six neurons, and each hidden layers consists of 40 neurons. The input variable in the network is time. The output variables are standard and heated catalyst temperature, standard and heated catalyst HC emissions, and standard and heated catalyst CO emissions. The back-propagation algorithm has been implemented to calculate errors and adjust weights of the hidden layer neurons. In order to avoid long training time or network being trapped in local error minima, various learning rate constants are tried. The ANN structure and number of neurons in each hidden layers have been selected by using an evolutionary algorithm. The evolutionary algorithm tried different ANN structures having various hidden layers and various numbers of neurons in each hidden layers. After that, the selected ANN has been trained for various learning rate and termination criteria. The latter may be either completion of a maximum number of epochs or achievement of the error goal. All of the data have been normalized in the range of [0, +1]. Sigmoid function is chosen for transfer function, with 0.5-threshold value as defined, f ðxÞ ¼

1 1þ

ð3Þ

e4ðx0:5Þ

Figs. 3–5 show a parity plot between experimental and computed data by ANN for standard and heated catalyst temperature, standard and heated catalyst HC emissions, and standard and heated catalyst CO emissions. The predictions have R2-values equal to 0.9988 for standard catalyst temperature, 0.9986 for heated catalyst temperature, 0.9991 for standard catalyst HC emissions, 0.9984 for heated catalyst HC emissions, 0.9997 for standard catalyst CO emissions, and 0.9993 for heated catalyst CO emissions. It can be clearly seen from Figs. 3–5, the developed ANN gives a very accurate representation of R2-values over the all range or working conditions. Since experimental results are very close to the calculated values that can be obtained by using ANN, those cannot be graphically shown together. For this reason, the following equations (Eqs. (4)–(9)) are used to calculate the deviation values, and these values have been shown graphically. Layer 1

Layer 2

1

1

2

2

Catalyst temperature (heated)

3

3

HC emissions (standard)

Catalyst temperature (standard)

Time HC emissions (heated) CO emissions (standard) 40

40

CO emissions (heated)

Fig. 2. ANN architecture used for the system.

ARTICLE IN PRESS

M. Ali Akcayol, C. Cinar / Applied Thermal Engineering xxx (2005) xxx–xxx 240 R2 = 0.9988

220

MeasuredValues

200 180 160 140 120 100 80 60 40 20 20

40

60

80 100 120 140 160 180 200 220 240

Predicted Values

Fig. 3. Comparison of measured and predicted values for the standard catalyst temperature.

480 R2 = 0.9991

440 400

MeasuredValues

6

360 320 280 240 200 160 120 80 40 0 0

40 80 120 160 200 240 280 320 360 400 440 480

Predicted Values

Fig. 4. Comparison of measured and predicted values for the standard catalyst HC emissions.

dTempStd ¼

TempStdANN TempStdexperimental TempStdexperimental

dTempHeated ¼

dHCEmsStd ¼

TempHeatedANN TempHeatedexperimental TempHeatedexperimental

HCEmsStdANN HCEmsStdexperimental HCEmsStdexperimental

ð4Þ

ð5Þ

ð6Þ

ARTICLE IN PRESS


7

2 R2 = 0.9997 1.75

Measured Values

1.5 1.25 1 0.75 0.5 0.25 0 0

0.25

0.5

0.75

1

1.2

1.5

1.75

2

Predicted Values

Fig. 5. Comparison of measured and predicted values for the standard catalyst CO emissions.

dHCEmsHeated ¼ dCOEmsStd ¼

HCEmsHeatedANN HCEmsHeatedexperimental HCEmsHeatedexperimental

COEmsStdANN COEmsStdexperimental COEmsStdexperimental

dCOEmsHeated ¼

COEmsHeatedANN COEmsHeatedexperimental COEmsHeatedexperimental

Fig. 6. Variation of the dTempStd and dTempHeated.

ð7Þ ð8Þ ð9Þ

ARTICLE IN PRESS

8


The deviations for standard and heated catalyst temperature, standard and heated catalyst HC emissions, and standard and heated catalyst CO emissions are illustrated in Figs. 6–8. According to the results, maximum deviations in standard catalyst temperature is 4.925%, in heated catalyst temperature is 1.602%, in standard catalyst HC emissions is 4.798%, in heated catalyst HC emissions is 4.926%, in standard catalyst CO emissions is 4.82%, and in heated catalyst CO emissions is 4.938%. Table 2 shows the minimum and maximum deviations for each of the output. These results prove that the proposed ANN can be used successfully for the prediction of catalyst temperature, HC emissions and CO emissions for the system.

Fig. 7. Variation of the dHCEmsStd and dHCEmsHeated.

Fig. 8. Variation of the dCOEmsStd and dCOEmsHeated.

ARTICLE IN PRESS


9

Table 2 Max and min deviations of catalyst temperature, HC emissions and CO emissions Output

Min/Max

Time

Deviations (%)

Experimental value

Standard catalyst temp. Standard catalyst temp. Heated catalyst temp. Heated catalyst temp. Standard catalyst HC emissions Standard catalyst HC emissions Heated catalyst HC emissions Heated catalyst HC emissions Standard catalyst CO emissions Standard catalyst CO emissions Heated catalyst CO emissions Heated catalyst CO emissions

Min Max Min Max Min Max Min Max Min Max Min Max

666 35 477 17 371 567 47 12 1000 580 666 115

0.0289632 4.9258398 0.0231151 1.6021841 0.0352935 4.7982165 0.5531468 4.9265346 0.000654 4.82 0 4.9381818

220 108 219 157 248 85 355 460 0 0.01 0 0.22

4. Conclusions In this study, an artificial neural network is used for prediction of catalyst temperature, HC emissions and CO emissions in the catalytic converter. HC and CO emission variations were measured during cold-start period of the engine with standard and burner heated catalytic converter under idle operating conditions at 950 rpm engine speed. Before the engine was started, the burner was activated to heat up the catalytic converter. LPG is burned in the burner to heat the main catalyst faster. The preheat temperature of the catalyst was 150 °C. Data recording was commenced when the engine was started (0 s) and continued for 1000 s. All the tests were performed for cold starts of the engine. Also, the engine was shut down for 10 h before the next test to provide an ambient temperature start. The deviations for catalyst temperature, HC emissions and CO emissions for different time are obtained by using ANN. The maximum deviations are 4.925% for standard catalyst temperature, 1.602% for heated catalyst temperature, 4.798% for standard catalyst HC emissions, 4.926% for heated catalyst HC emissions, 4.82% for standard catalyst CO emissions, and 4.938% for heated catalyst CO emissions. The statistical coefficients are above 0.99. This degree of accuracy shows that the proposed ANN can be used to obtain the experimental data. To sum up, this study is considered to be helpful in predicting the performance of the catalytic converter.

References [1] A. Caraceni, V. Cioffi, F. Garofalo, A. Senatore, G. Vittorioso, C. Barberio, G. Saroglia, Emission control technologies for EU stage IV + EOBD on small cars (Part I) pre-screening of potential solutions, SAE Paper 199901-0775, 1999. [2] R.M. Heck, R.J. Farrauto, Automobile exhaust catalysts, Applied Catalysis 221 (1–2) (2000) 443–457. [3] G.C. Koltsakis, A.M. Stamatelos, Catalytic automotive exhaust aftertreatment, Progressive Energy Combustion Science 23 (1997) 1–39. [4] T. Kirchner, G. Eigenberger, On the dynamic behavior of automotive catalysts, Catalysis Today 38 (1) (1997) 3– 12.

ARTICLE IN PRESS

10


[5] R.F. Farrauto, R.M. Heck, Catalytic converters: state of the art and perspectives, Catalysis Today 51 (3–4) (1999) 351–360. [6] Y. Zhenming, Z. Jinsong, C. Xiaoming, L. Qiang, X. Zhijun, Z. Zhimin, Microwave enhanced exhaust conversion of internal combustion engines, Applied Catalysis 34 (2) (2001) 129–135. [7] W.W. Pulkrabek, Engineering Fundamentals of the Internal Combustion Engine, Prentice Hall Inc., USA, 1997. [8] W.R. Moore, R. Mondt, Predicted cold start emission reductions resulting from exhaust thermal energy conversation to quicken catalytic converter lightoff, SAE Paper 931087, 1993. [9] C.A. Amann, Control of the homogeneous-charge passenger car engine-defining the problem, SAE Paper 801440, 1980. [10] K. Horie, H. Takahasi, S. Akazai, Emission reduction during warm-up period by incorporating a wall-wetting fuel model on the fuel injection strategy during engine starting, SAE Paper 952478, 1995. [11] E. Korin, R. Reshef, D. Tshernichovesky, E. Sher, Reducing cold-start emission from internal combustion engines by means of a catalytic converter embedded in a phase-change materials, Journal of Automobile Engineering 213 (D) (1999) 575–583. [12] J. Kaspar, P. Fornasiero, N. Hickey, Automotive catalytic converters: current status and some perspectives, Catalysis Today 77 (4) (2003) 419–449. [13] Y. Shimasaki, H. Kato, F. Abe, S. Hashimoto, T. Kaneko, Development of extruded electrically heated catalyst system for ULEV standards, SAE Paper 971031, 1997. [14] F.J. Hanel, E. Otto, R. Bru¨ck, Electrically heated catalytic converter (EHC) in the BMW ALPINA B12 5.7 switchtronic, SAE Paper 960349, 1996. [15] V.S. Sendil, K.B. Jeyachandran, K. Bhaskar, Experimental investigation of emission control from spark ignition engine using electrically heated catalyst, SAE Technical Paper 2001-01-2000, 2001. [16] C. Cinar, T. Topgul, The investigation of the effect of catalytic convertor on exhaust emissions during heating period in fuel injection gasoline engines, 7th International Combustion Symposium, Ankara, Turkey, July 17–18, 2002. [17] J.M. Serra, A. Corma, A. Chica, E. Argente, V. Botti, Can artificial neural networks help the experimentation in catalysis? Catalysis Today 81 (3) (2003) 393–403. [18] T. Yasgashi, K. Yoshizake, T. Nagami, S. Sugiuram, T. Yoshigama, K. Ohsawa, New technology for reducing the power consumption of electrically heated catalysts, SAE Paper 940464, 1994. [19] G.C. Koltsakis, P.A. Konstantinidis, A.M. Stamatelos, Development and application range of mathematical models for 3-way catalytic converter, Applied Catalysis B: Environmental 12 (1997) 161–191. [20] M. Mohandes, S. Rehman, T.O. Halawani, Estimation of global solar-radiation using artificial neural networks, Renewable Energy 14 (1–4) (1998) 179–184. [21] J.S.R. Jang, C.T. Sun, E. Mizutani, Neuro-fuzzy and Soft Computing: a Computational Approach to Learning and Machine Intelligence, Prentice-Hall International, 1997. [22] J.S.R. Jang, C.T. Sun, Neuro-fuzzy modeling and control, The Proceedings of the IEEE 83 (1995) 378–406. [23] P.N. Botsaris, D. Bechrakis, P.D. Sparis, An estimation of three-way catalyst performance using artificial neural network during cold start, Applied Catalysis A: General 243 (2) (2003) 285–292. [24] K. Huang, X.L. Zhan, F.Q. Chen, D.W. Lu¨, Catalyst design for methane oxidative coupling by using artificial neural network and hybrid genetic algorithm, Chemical Engineering Science 58 (1) (2003) 81–87. [25] T. Hattori, S. Kito, Neural network as a tool for catalyst development, Catalysis Today 23 (4) (1995) 347–355. [26] T. Hattori, S. Kito, Artificial intelligence approach to catalyst design, Catalysis Today 10 (2) (1991) 201–211. [27] U. Rodemerck, M. Baerns, M. Holena, D. Wolf, Application of a genetic algorithm and neural network for the discovery and optimization of new solid catalytic materials, Applied Surface Science 223 (1–3) (2004) 168–174. [28] S.A. Kalogirou, Applications of artificial neural-networks for energy systems, Applied Energy 67 (2000) 17–35. [29] A. Palau, E. Velo, L. Puigjaner, Use of neural-networks and expert systems to control a gas/solid absorption chilling machine, International Journal of Refrigeration 22 (1999) 59–66. [30] K.S. Reddy, M. Ranjan, Solar resource estimation using artificial neural-networks and comparison with other correlation models, Energy Conversion and Management 44 (2003) 2519–2530.