Applied Thermal Engineering 50 (2013) 63e70

Contents lists available at SciVerse ScienceDirect

Applied Thermal Engineering journal homepage: www.elsevier.com/locate/apthermeng

Artificial neural networks for automotive air-conditioning systems performance prediction Haslinda Mohamed Kamar a, Robiah Ahmad b, *, N.B. Kamsah a, Ahmad Faiz Mohamad Mustafa a a b

Faculty of Mechanical Engineering, Universiti Teknologi Malaysia, 81310 Skudai, Johor, Malaysia UTM Razak School of Engineering and Advanced Technology, UTM International Campus, Jalan Semarak, 54100 Kuala Lumpur, Malaysia

h i g h l i g h t s < An experimental rig was developed for an AAC system and the data was used to determine ANN model. < The 4e3e3 configuration neural network model was optimized and used to predict the AAC performance. < The predicted outputs gave high correlation to the experimental data in predicting the AAC system performance. < The range values of error index, MSE and RMSE are 0.65e1.65%, 1.09  105e9.05  105 and 0.33e0.95% respectively.

a r t i c l e i n f o

a b s t r a c t

Article history: Received 2 February 2010 Accepted 21 May 2012 Available online 4 June 2012

In this study, ANN model for a standard air-conditioning system for a passenger car was developed to predict the cooling capacity, compressor power input and the coefficient of performance (COP) of the automotive air-conditioning (AAC) system. This paper describes the development of an experimental rig for generating the required data. The experimental rig was operated at steady-state conditions while varying the compressor speed, air temperature at evaporator inlet, air temperature at condenser inlet and air velocity at evaporator inlet. Using these data, the network using LavenbergeMarquardt (LM) variant was optimized for 4e3e3 (neurons in inputehiddeneoutput layers) configuration. The developed ANN model for the AAC system shows good performance with an error index in the range of 0.65 e1.65%, mean square error (MSE) between 1.09  105 and 9.05  105 and the root mean square error (RMSE) in the range of 0.33e0.95%. Moreover, the correlation which relates the predicted outputs of the ANN model to the experimental results has a high coefficient in predicting the AAC system performance. Ó 2012 Elsevier Ltd. All rights reserved.

Keywords: Automotive air-conditioning (AAC) Artificial neural network (ANN) Mathematical modeling Coefficient of performance (COP)

1. Introduction Automotive air-conditioning (AAC) system is used to maintain comfortable condition in the compartment of passenger cars. To achieve such condition, the panel outlet for air flow direction, volume, velocity and temperature has to be adjustable over a wide range of climatic and driving conditions. The compressor of the AAC systems is belt driven by the engine hence its speed is directly governed by the engine speed and causes the cooling capacity of the system to vary with the engine speed. The AAC system must be capable of lowering the air temperature in the passenger compartment quickly and quietly. These conditions make the analysis of the AAC system becomes complicated compared to that

* Corresponding author. Tel.: þ60 3 26154333; fax: þ60 3 26934844. E-mail addresses: [email protected] (H.M. Kamar), [email protected] (R. Ahmad), [email protected] (N.B. Kamsah). 1359-4311/$ e see front matter Ó 2012 Elsevier Ltd. All rights reserved. doi:10.1016/j.applthermaleng.2012.05.032

of the stationary air-conditioning system, as described by Bhatti [1] and Kargilis [2]. Traditionally, the performance of an AAC system has been done experimentally by many researchers such as Rubas and Bullard [3] on determination of COP of refrigerant cycle for AAC system, Ratts and Brown [4] on effects of the R-134a refrigerant charge level on the performance of the AAC system, Al-Rabghi and Niyaz [5] on COP determination using two types of refrigerants, R-12 and R-134a, as a function of compressor speed and Kaynakli and Horuz [6] on determination of the optimum operating conditions for AAC system using refrigerant R-134a. Hosoz and Direk [7] carried out experimental work on an AAC system operating in both air-conditioning and heat pump modes, with varying compressor speed and air temperatures at the inlets of both outdoor and indoor coils. With the advancement in computer technology and availability of simulation software, computer simulation methods have become increasingly popular for analyzing the performance of AAC systems. For example is by Jung et al. [8] studied on working with different

64

H.M. Kamar et al. / Applied Thermal Engineering 50 (2013) 63e70

Nomenclature a AAC ANN b COP e EI f _ m MSE n N OSA pi

activation function automotive air-conditioning artificial neural network bias coefficient of performance expected value error index activation function mass flow rate (kg/s) mean square error sum of weighted inputs compressor speed (rpm) step-ahead prediction input vector

refrigerant mixtures on AAC system. Joudi et al. [9] also carried out computer simulation on predicting the performance of AAC system with different types of refrigerants and investigating a suitable alternative to refrigerant R-12. Lee and Yoo [10] have developed a mathematical model for the AAC system working with R-134a refrigerant and analyzed the performance of each component of the system. They showed that the difference between the simulation and experimental results was about 7%. Tian and Li [11] developed a mathematical model of an AAC system working with refrigerant R-134a, with a variable speed compressor. They showed that the results of computer simulation were within 11% of the corresponding values obtained from experiments. For the past two decades, artificial neural networks (ANNs) have been used in refrigeration and air-conditioning system for performance prediction. Chang [12] studied on relationship of power consumption, chilled water temperature and cooling water temperatures of chillers in air-condition system. Hosoz and Ertunc [13] used the ANN approach to investigate the performance of an AAC system employing vapor-compression refrigeration circuit using refrigerant R134-a. The ANN model was used to predict the compressor power input, heat rejection rate in the condenser, refrigerant mass flow rate, compressor discharge rate and the coefficient of performance (COP). Besides, Ertunc and Hosoz [14] used both ANN and adaptive neuro-fuzzy inference system (ANFIS) models to predict performance of an evaporative condenser such as the heat rejection rate, temperature of the leaving refrigerant along with dry and wet bulb temperature of the leaving air stream. Yilmaz and Atik [15] proposed ANN model to mechanical cooling system with variable cooling capacity. Shahin [16] recently presented work on using ANN and ANFIS for performance analysis of single stage vapor compression refrigeration system with internal heat exchanger using refrigerants such as R134a, R404a and R407c. A complete review on applications of ANN for energy and exergy analysis of refrigeration, air conditioning and heat pump systems has been done by Mohanraj et al. [17]. Although there are many examples of ANN applications in airconditioning system, quite a few studies being done on AAC. The related works were by Hosoz and Ertunc [13] and recently by Atik et al. [18] which modeled mobile air conditioning system in different amount of refrigerants and in different compressor revolution speed using ANN. In this study, ANN model for an automotive airconditioning (AAC) system was developed which employs a vapor-compression circuit working with refrigerant R-134a. The model was developed using several steady-state input data of the AAC system obtained from the actual experiments. The developed

Q_ L R RMSE S T V _ W wi

cooling effect (kW) correlation performance root mean square error sum of squares temperature ( C) velocity (m/s) compressor power (kW) connection weight

Subscripts a air i inlet evp evaporator comp compressor cond condenser

ANN model was then used to predict the AAC system performance namely cooling capacity of the evaporator, the input power to the compressor and the coefficient of performance (COP) of the AAC system. Their predicted values to the real experimental data were analyzed.

2. Artificial neural networks Artificial neural network (ANN) computational modeling technique was inspired by a biological nervous system as in the human brain. It can be employed to model complex physical systems and does not require an explicit mathematical representation. The fundamental processing element of the ANN is a neuron, while weighted connection serves as the synapse. Each neuron receives input through weighted connections (inputs) and these inputs are combined or summed in a specific manner. The formulation for the sum of the weighted inputs is given by

n ¼

R X

wi pi þ b

(1)

i¼1

where pi and wi are the input vector and the connection weight from the input vector pi, respectively and b is the bias of the neuron. The sum of the weighted inputs with a bias is processed through an activation function, represented by a and the output that it evaluates is

aðnÞ ¼ f

R X

! wi pi þ b

(2)

i¼1

The activation function can be defined in many ways such as threshold function, sigmoid function and hyperbolic tangent function. Typically, the architecture of an ANN model is made up of three basic layers namely an input layer, a hidden layer and an output layer. Fig. 1 shows the architecture of a three-layer ANN model developed for an experimental AAC system. The first layer is an input layer with four neurons, the middle layer is a hidden layer consisting of five neurons and the third layer is an output layer with three neurons. The inputs to the ANN model are compressor speed, Ncomp, air temperatures at evaporator inlet, Ta,i,evp, air temperatures at condenser inlet, Ta,i,cond, and air velocity at evaporator inlet, Va,i,evp. The outputs of the model are the cooling effect, Q_ L , the compressor _ and the coefficient of performance (COP) of the input power, W; system. The ANN model is developed through 2 stages: training

H.M. Kamar et al. / Applied Thermal Engineering 50 (2013) 63e70

65

Other measures of the ANN performance are the mean square error (MSE), root mean square error (RMSE) and error index (EI). The error index is defined as [20]

vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi uPn u ðy  x Þ2 EI ¼ t i ¼ P1n i 2 i i ¼ 1 yi

(7)

The accuracy of the ANN model prediction can be measured using a one-step ahead prediction (OSA) technique in which the predicted output is compared with the previous input and output data. The OSA is defined as,

    yi ðtÞ ¼ fi xðt  1Þ; /; x t  ny ; uðt  1Þ; /; uðt  nu Þ

(8)

where fi(t) is the predicted nonlinear function.

3. Description of the experimental setup Fig. 1. The architecture of the ANN model for the AAC system.

stage and testing stage. The network is trained to predict an output based on input data during the training stage. To validate the result, the model is tested using difference sets of input. In this study, the computer program was performed under MATLAB environment using neural network toolbox. 2.1. Model validation The validate the performance of the ANN model, statistical analysis is used, i.e. the correlation coefficient, mean relative error and root mean square. All these methods evaluate the predicted outputs of the ANN model with the actual values of the outputs, usually obtained experimentally. A correlation coefficient (R) is used to indicate the strength of the relationship between the predicted outputs and the experimental results. It is given by Ref. [19]

Sxy R ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffi Sxx Syy

Fig. 2 illustrates the schematic diagram of the AAC experimental rig developed in this study. It consists of three sections namely a vapor-compression refrigeration circuit, a closed air duct for evaporator section, and an open air duct for condenser section. The ducts were designed and constructed according to the British Standard (BS 5141: Part: 1975) for rating of duct mounted air cooling coils [21]. The AAC system was constructed using original components of a Denso air-conditioning system used in a typical compact size car working with refrigerant R-134a. As seen in this figure, the main components of the vapor compression refrigeration circuit are a swash-plate compressor, an evaporator, a condenser, an expansion valve, a receiver drier, a sight glass and insulated interconnecting pipes. The air duct for the evaporator section is a closed loop rectangular conduit with an overall length of 5130 mm. The duct walls between the downstream and upstream of the evaporator coil are thermally insulated using 25 mm thick rock wool material. Air was forced to flow through the duct using a variable speed centrifugal

(3)

where Sxy, Sxx and Syy are respectively given by

!

Pn Sxy ¼

n X

i ¼ 1 xi

xi y i 

n X i¼1

Syy ¼

i¼1

yi Air duct

(4)

i ¼ 1 xi

x2i 

Valve

!2

i¼1

yi 

Ncomp

(5)

n Pn

n X

Condenser

!2

Pn Sxx ¼

i¼1

n

i¼1

Pn

Ta,i,cond

!

n

Compressor

yi

Ta,i,evp

(6)

in which n is the total number of data in the particular data set. The correlation coefficient R ranges between 1 and þ1. When R is close to 1 there is a strong positive relationship between x and y data sets. On the other hand, if it is close to 1 there is a strong negative linear relationship between the data sets. When R is close to zero, either from positive or negative direction, there is a weak or no relationship between x and y data sets.

Air duct

Evaporator

Va,i,evp

Fig. 2. Schematic diagram of the AAC system experimental rig.

66

H.M. Kamar et al. / Applied Thermal Engineering 50 (2013) 63e70

fan. To improve the air quality, an air mixer, a honeycomb air flow straightener and gauge screens were installed at the upstream location of the evaporator coil. A 2 kW electrical heater was used to provide a heating load to the evaporator coil. A steam humidifier was incorporated in the ducting system to inject vapor at a rate of 4 kg/h into the air stream, in order to alter the humidity of the air. The heating and humidification sections were used to adjust the air temperature and humidity at the evaporator coil inlet. The air duct for the condenser is a rectangular open conduit with an overall length of 663 mm. Air was forced to flow through the duct using a constant speed blower which was installed at the upstream location of the air duct. The velocity of the air passing through the condenser was varied by controlling the blower speed using a reducer unit and damper. The dry-bulb temperature of the air at the inlet to the condenser coil was varied using a 2 kW electrical heater. The AAC experimental rig was furnished with data acquisition system consisting of a standard notebook computer with a personal computer memory card international association (PCMCIA) slot, a data acquisition PCMCIA card and a signal conditioning rack. Two pt100 resistance temperature detector (RTD) sensors were used to measure the air temperature at the upstream and downstream locations of the evaporator coil. Another pt100 RTD sensor was used to measure the air temperature at the upstream of the condenser coil. Two air velocity transducers were used to measure velocity of the air flowing through the evaporator and condenser coils. The compressor speed was regulated using a frequency inverter. The temperature of the refrigerant at various locations in the vapor-compression refrigeration circuit was measured using type-K thermocouples that were inserted into the copper tubing. The pressure at the compressor suction and discharge sections was measured using a Bourdon tube pressure gauge. We assumed that there was no pressure loss in the piping hence the evaporating and condensing pressures are equal to the suction and discharge pressures, respectively. The mass flow rate of the refrigerant was measured using a flow meter which was mounted at the exit of the condenser. A sight glass tube was placed before the flow meter to ensure that the refrigerant is always in a sub-cooled state. Some features of the instrumentation are summarized in Table 1. During the experiments, the AAC system was run until a steadystate condition was attained. Four operating parameters were varied during series of experiments, each within their respective range as shown below. The condenser air velocity was held constant at about 1.6 m/s at all time. Evaporator inlet air dry-bulb temperature, Ta,i,evp 25e35  C. Condenser inlet air dry-bulb temperature, Ta,i,cond 32e40  C. Compressor speed, Ncomp 1500e2500 rpm. Evaporator air velocity, Va,i,evp 1.6e3.8 m/s.

was set at 1500, 2000, and 2500 rpm by controlling the motor speed using a frequency inverter. At each compressor speed, the air velocity of the evaporator coil is maintained at the blower speeds 1, 2, 3 and 4. At each blower speed, the air dry-bulb temperature at the inlet of the evaporator was kept at 25, 30 and 35  C by varying the electrical energy input to the electric heater. At each evaporator inlet air temperature, the air dry-bulb temperature at the inlet of the condenser was set at 32, 35 and 40  C, by varying the electrical energy input to the electric heaters. To stabilize the AAC system, the air-conditioner was run 30 min prior to each test. A total of 36 tests were conducted on the AAC system. In each test, the cooling capacity, i.e. the rate of heat removal in the evaporator, Q_ L , _ and the coefficient of performance compressor power input, W (COP) of the AAC system were determined. 4. Performance analysis of the AAC system The vapor-compression refrigeration section of the AAC experimental rig consists of four main components namely the compressor, which is belt-driven by an electric motor, condenser, expansion valve and evaporator. These four components are assumed to be working in steady-state modes at all times since the AAC system was operated in steady-state conditions during the experiments. To simplify thermodynamic analysis on this system, the changes in both kinetic and potential energy of the R134a refrigerant are neglected since they are very small compared with the work and heat transfer interactions. With this simplification, the steady-flow energy equation can be used to determine the cooling capacity of the evaporator, i.e.

  _ ref hevap;i  hevap;e Q_ evap ¼ m

ðkWÞ

_ ref is the mass flow rate (in kg/s) of the refrigerant R134-a, where m hevap,i is the specific enthalpy (in kJ/kg) of the refrigerant at the inlet to the evaporator and hevap,e is the specific enthalpy (in kJ/kg) of the refrigerant at the exit of the evaporator. The power input to the compressor is given by

  _ comp ¼ m _ ref hcomp;e  hcomp;i W

Q_ evap _ comp W

COP ¼

Table 1 Characteristics of the instrumentation.

5. ANN model of the AAC system

Instrument

Range

Uncertainty

Temperature Pressure Humidity Air velocity Mass flow rate Compressor speed Voltage Current

RTD sensors Bourdon gauge Humidity sensor Velocity transducer Flow meter Digital tachometer Digital voltmeter Digital ammeter

25  Ce100  C 0e300 kPa 0e100% RH 0e20 m/s 0e25 kg/s 0e20,000 rpm 0e250 V 0e20 A

0.3  C 1 kPa 1% 0.5% 1% 2% 1% 1%

ðkWÞ

(10)

where hcomp,e is the specific enthalpy (in kJ/kg) of the refrigerant at the exit of the compressor hcomp,i is the specific enthalpy (in kJ/kg) of the refrigerant at the inlet to the compressor. The enthalpies were obtained from property tables for the refrigerant, based on the pressure and temperature readings obtained at the corresponding locations in the refrigeration circuit. The coefficients of performance (COP) of the AAC system is then given by

The tests were run at different compressor speeds with varying evaporator heat load. In the air side tests, the compressor speed

Measured variable

(9)

(11)

The procedure for using an artificial neural network (ANN) to model the experimental AAC system was based on a system identification technique. The development of the ANN model was done in four steps namely acquiring experimental data from the AAC experimental rig, determination of a suitable ANN model for the AAC system, construction of the ANN model structure, and finally validation of the ANN model. Fig. 3 shows a schematic of the structure of the ANN model that was developed for the AAC system.

H.M. Kamar et al. / Applied Thermal Engineering 50 (2013) 63e70

Input Variable Parameter

6.1. ANN model with different data types

Output Win,, QL, COP

AAC system

nc, Ta,i,cond, Va,i,evp, Ta,i,evp

xi

xi

yˆ i ANNs model ui ei

Fig. 3. The ANNs model representation of the AAC system.

Parameter u represents the input signal, y represents the output signal from the model and ei represents the error between the ^i . Since the AAC system actual outputs xi and the predicted outputs y is a nonlinear system, a feed forward multilayered neural network was chosen as the basis for developing the ANN model. A multilayered artificial neural network is typically made up of one or more hidden layers. The input layer acts as a data holder which distributes the inputs to the first hidden layer. The input data is then propagated through the network and each neuron computes its output according to a relation

aðnÞ ¼ f

X

! wi pi þ b

(12)

i¼1

where pi and wi are input vector and the connection weight from the input vector pi, respectively and b is the bias of the neuron. The sum of the weighted inputs with a bias is processed through an activation function, represented by f. The activation function used for the neuron in the hidden layer is a tangent sigmoid function, given by

f ðnÞ ¼

1 1 þ en

(13)

In a system modeling application, the dynamic range of the output data may be greater than 1. Hence the activation function for the output layer of the ANNs model is chosen to be a linear function. The output layer performs a weighted sum of its inputs according to

^i ¼ y

X

wi pi

67

(14)

i¼1

The effects of number of neurons used in the hidden layer of the ANN model on the performance of the model were investigated. For each number of neuron, two different types of data were used namely normalized and non-normalized data and the performances were measured using performance measures namely MSE, RMSE and EI. For the normalized data, the inputs and outputs were converted into the range of zero to unity, while for the nonnormalized data the inputs and outputs remain unchanged. The results from Table 2 show that the average value of EI decreases with the number of neurons in the hidden layer for both types of data. For a given number of neurons, the average EI for the nonnormalized data is much smaller than that for the normalized data. For the normalized data, the average EI decreases by about 56% as the number of neurons was increased from one to two while the average EI for the non-normalized data decreases by about 71%. As the number of neurons was increased from two to three, the average EI for the normalized data further decreases by about 87% while the average EI for the non-normalized data remains unchanged. The results in Table 2 show some ANN models with non-normalized data giving better performance. However, developing an ANN model with non-normalized data is difficult and time consuming. Furthermore Haykin [22] mentioned about the importance of data normalization to reduce fluctuation. From this investigation, ANN model with two-layer feed-forward networks with three neurons using normalized data is adequate to model the AAC system. 6.2. ANN model with various input parameters The effects of five input parameters are investigated in this section. Five different cases were considered which are denoted as NN1 to NN5 and the details are explained in Table 3. For each analysis, the error index (EI) of the model is determined and results are shown in Table 3. Note that large value of error index indicates that the ANN model is less sensitive and has lower performance. Result in Table 3 shows that the sensitivity of the ANN model is largest hence has lowest performance when the compressor speed was excluded in the analysis. The error index is the lowest when all input parameters were included in the analysis. Based on this finding, we have included all the input parameters into our ANN model. Fig. 4 shows both the optimized ANN model (NN5) output superimposed the actual experimental output data for training and test data sets. It indicates that a good agreement between experimental values and those predicted by the model has been obtained.

6. Parametric study on the ANN model

6.3. ANN model with different number of layers

A parametric study was conducted to investigate the effects of different data types, input parameters and number of layers on the performance of the ANN model of the AAC system. During this study the output results predicted by the model were compared with the results obtained from the experiments. About 70% of the data was dedicated for training and the remaining for capability prediction of the ANN model. The training step was repeated several times in order to avoid local extreme points and to find the optimum performance of the model. The performance was assessed by determining the average values of the mean square error (MSE), the root mean square error (RMSE) and the error index (EI). The learning rate and the number of epochs were set at 0.05 and 1000, respectively. During the training process the weighing coefficients were adjusted by using a back-propagation LevenbergeMarquardt algorithm.

Finally, the effects of number of layers on the performance of the ANN model were carried out and results are summarized in Tables 4 and 5. The result in Tables indicates that a feed-forward ANN model with three nodes in the hidden layer gives a better performance

Table 2 Performance of the ANN model with different data types. Normalized data Neuron Avg. MSE 1 2 3 4 5

Unnormalized data

Avg. RMSE Avg. EI Avg. MSE

0.022 13.39 0.004 5.60 5 0.71 5.81  10 0.82 9.39  105 1.04 1.62  104

0.233 0.102 0.013 0.015 0.018

0.02 2.27 2.33 2.29 3.01

   

104 104 104 105

Avg. RMSE Avg. EI 10.17 1.04 1.14 1.10 0.47

0.035 0.010 0.010 0.010 0.002

68

H.M. Kamar et al. / Applied Thermal Engineering 50 (2013) 63e70

Table 3 Performance of the ANN models with various input parameters.

Table 4 Performance of the ANN model with two layers.

Neural network model Compressor speed Ta,evap Ta,cond Va,evap Error index NN1 NN2 NN3 NN4 NN5

NI I I I I

I NI I I I

I I NI I I

I I I NI I

0.018 0.016 0.014 0.015 0.013

2

Ta,evap e air temperatures at inlet evaporator, Ta,cond e air temperatures at inlet condenser, Va,evap e air velocity at inlet evaporator, NI e not included in the analysis, I e included in the analysis.

with an average MSE of 5.81  105, RMSE of 0.71% and error index of about 0.013. Initially the performance of the ANN model is decreasing with increasing number of nodes in the hidden layer. However, the performance begins to increase when four nodes are used in the hidden layer. This behavior could be due to the complexity that occurs in the neural network model. Table 5 shows results on the performances of the ANN model when the hidden

a

Actual output Predicted output

Cooling Effect, kW

1

Avg. MSE

Avg. RMSE

Avg. IE

1 2 3 4 5

0.022 0.004 5.81  105 9.39  105 1.62  104

13.4 5.60 0.71 0.82 1.04

0.230 0.102 0.013 0.015 0.018

layer was increased up to three layers. The results indicate that the ANN model is able to perform a fairly good predictions but not as good as the ANN model with three nodes in one hidden layer. The feed-forward ANN model with three hidden layers only gives an average MSE of about 1.07  104, RMSE of 0.92% and error index of 0.017 when three nodes are used in the first hidden layer and five nodes in the second hidden layer. These results suggest that an ANN model with one hidden layer and three nodes within the layer is sufficient to represent the AAC system and can be used fairly accurately to predict the thermal performance of the system. Fig. 5 shows plots of the predicted vs. actual outputs for this optimized ANN model for both training and test data sets. The graph shows a close agreement between the outputs of NN model and actual experimental data. 7. Results and discussion

0.6 0.4

0.2

-0.2

Training Data

0

20

40

Test Data

60

80

100

120

No. of data

b 1.2

Actual output Predicted output

1

Compressor Power, kW

Nodes

0.8

0

0.8 0.6 0.4 0.2 0 -0.2

Training Data

0

20

40

Test Data

60

80

100

120

No. of data

c 1.2

Actual output Predicted output

1

Coefficient of Performance

Layer

Based on the above parametric study, a suitable ANN model for AAC system was developed. The properties of the ANN model for the AAC system are tabulated in Table 6. Figs. 6e8 are the plots of performance parameters of the AAC system predicted by the ANN model vs. the corresponding values obtained from experiments. Note that in all cases, the correlation coefficients R are very close to unity. This indicates that the ANN model was able to predict the performance parameters of the AAC system with a very good accuracy. These comparisons were made using only the test data set, which was not introduced to the ANN model during the training process. Fig. 6 shows the plot of cooling effects predicted by the ANN model vs. the values obtained from the experiments. The ANN predictions yield a mean square error (MSE) of about 1.09  105 kW, root mean square error (RMSE) of 0.33%, error index of 0.56% and correlation coefficient, R of 0.99. These values show that the ANN model is able to predict the heat absorbed by the refrigerant in the evaporator with a very good accuracy. A plot of the ANN prediction’s of the compressor power vs. the experimental values is shown in Fig. 7. The ANN prediction produces MSE value of about 9.05  105 kW, RMSE of 0.95%, index error of 1.65% and correlation coefficient of 0.9998. Either the compressor power or the cooling effect of the AAC system could have been determined indirectly using the principle of conservation of

0.8

Table 5 Performance of the ANN model with three layers.

0.6

Layer

0.4

3 0.2 0 -0.2

Training Data

0

20

40

Test Data

60

80

100

120

No. of data

Fig. 4. One step-ahead prediction output superimposed on experimental output using NN5 for (a) cooling effect, (b) compressor power, and (c) coefficient of performance.

Node 1

Node 2

Avg. MSE

Avg. RMSE

Avg. IE

3

1 2 3 4 5 6 7 8 9 10

0.022 0.004 2.06  5.32  1.07  2.30  2.50  3.64  4.66  7.06 

13.39 5.60 1.25 2.05 0.91 1.50 1.47 1.74 2.16 2.56

0.23 0.10 0.022 0.037 0.017 0.027 0.027 0.032 0.038 0.046

104 104 104 104 104 104 104 104

H.M. Kamar et al. / Applied Thermal Engineering 50 (2013) 63e70

a

1.4

1.2

Actual output Predicted output

R = 0.99995 MSE = 1.0893e-5 kW RMSE = 0.33% Index Error = 0.56%

1.2

Predicted Cooling Effect (kW)

Compressor Power, kW

1 0.8 0.6 0.4 0.2 0 -0.2

69

Training Data

0

20

40

Test Data

60

80

100

1

0.8

0.6

0.4

0.2

120

No. of data 0

b

Actual output Predicted output

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Fig. 6. ANN prediction of the cooling effect vs. the experimental results.

0.8 0.6 0.4

1

0.2

0.9

0 -0.2

Training Data

0

20

40

Test Data

60

80

100

120

No. of data

c 1.2

Actual output Predicted output

1 0.8 0.6

R = 0.99978 MSE = 9.0506e-5 kW RMSE = 0.95% Index Error = 1.65%

0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1

0.4

0 0

0.2 0 -0.2

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Experimental Compressor Power (kW) Training Data

0

20

40

Test Data

60

80

Fig. 7. ANN prediction of the compressor power vs. the experimental results. 100

120

No. of data

Fig. 5. One step-ahead prediction output superimposed on experimental output using three nodes in a hidden layer for (a) cooling effect, (b) compressor power, and (c) coefficient of performance.

energy if one of these two output parameters is known from the ANN prediction [17]. Instead, these parameters were predicted independently in order to evaluate the performance of the ANN in both categories of parameters. Fig. 8 shows the plot of coefficient of performance (COP) of the AAC system predicted by the ANN model vs. the values obtained from the experiments. The ANN prediction yields MSE value of about 7.28  105, RMSE of 0.85%,

Table 6 Properties of the ANN model for the AAC system. Structure

Feed forward neural network

Algorithm Type of training Learning rate Number of epochs Transfer function Data type Input parameter Network structure

Back propagation LevenbergeMarquardt 0.05 1000 Tangent sigmoid and linear Normalized data All 4 inputs Three nodes in a hidden layer

index error of 1.65% and correlation coefficient R of 0.9997. It can be seen that predicted results are well close to the experimental data and the deviations are small for each performance parameter.

1

Predicted Coefficient of Performance

Coefficient of Performance

0

Experimental Cooling Effect (kW)

Predicted Compressor Power (kW)

Cooling Effect, kW

1

R = 0.99965 MSE = 7.2817e-5 kW RMSE = 0.85% Index Error = 1.65%

0.8

0.6

0.4

0.2

0

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Experimental Coefficient of Performance

Fig. 8. ANN prediction of the coefficient of performance vs. the experimental results.

70

H.M. Kamar et al. / Applied Thermal Engineering 50 (2013) 63e70

8. Conclusion An ANN model based on the back-propagation algorithm has been developed in this study to predict the performance parameters of an experimental AAC system. The ANN model contains one hidden layer comprising of five nodes. Training and testing data set for the ANN model were obtained from steady state tests conducted on the AAC experimental rig. The trained ANN model was then used to predict the cooling load, compressor power input and the coefficient of performance of the AAC system experimental rig. The performance of the ANN model was assessed using the mean square error, root mean square error, error index and the correlation coefficient. The ANN model was found to be capable of accurately predicting the performance parameters of the AAC system. For all the performance parameters, the error index of the ANN model is in the range of 0.65e1.65%, MSE between 1.09  105 and 9.05  105 and root mean square error in the range of 0.33e0.95%. The correlation coefficients for all performance parameters are found to be very close to unity, indicating that the ANN model can predict the performance parameters of the AAC system with a very good accuracy. Acknowledgements The authors would like to acknowledge the supports from Universiti Teknologi Malaysia and fund provided by Ministry of Science, Technology and Innovation, Malaysia throughout this study under IRPA Vot 74062. References [1] M.S. Bhatti, Riding in comfort, part II: evolution of automotive air conditioning, ASHRAE Journal 41 (9) (1999) 44e52. [2] A. Kargilis, Design and Development of Automotive Air Conditioning Systems, ALKAR Engineering Company, 2003. [3] P.J. Rubas, C.W. Bullard, Factors contributing to refrigerator cycling losses, International Journal of Refrigeration 18 (3) (1995) 168e176. [4] E.B. Ratts, J.S. Brown, An experimental analysis of the effect of the refrigerant charge level on an automotive refrigeration system, International Journal of Thermal Science 39 (2000) 592e604. [5] O.M. Al-Rabghi, A.A. Niyaz, Retrofitting R-12 car air conditioner with R-134a refrigerant, International Journal of Energy Research 24 (2000) 467e474.

[6] O. Kaynakli, I. Horuz, An experimental analysis of automotive air conditioning system, International Communications in Heat and Mass Transfer 30 (2003) 273e284. [7] M. Hosoz, M. Direk, Performance evaluation of an integrated automotive air conditioning and heat pump system, Energy Conversion and Management 47 (2006) 545e559. [8] D. Jung, B. Park, H. Lee, Evaluation of supplementary/retrofit for automobile air conditioners charged with R12, International Journal of Refrigeration 22 (1999) 558e568. [9] K.A. Joudi, A.S. Mohammed, M.K. Aljanabi, Experimental and computer performance study of an automotive air conditioning system with alternative refrigerants, Energy Conversion and Management 44 (2003) 2959e2976. [10] G.H. Lee, J.Y. Yoo, Performance analysis and simulation of automobile air conditioning system, International Journal of Refrigeration 23 (2000) 243e254. [11] C. Tian, X. Li, Numerical simulation on performance band of automotive air conditioning system with variable displacement compressor, Energy Conversion and Management 46 (2005) 2718e2738. [12] Y.-C. Chang, Sequencing of chillers by estimating chiller power consumption using artificial neural networks, Building and Environment 42 (2007) 180e188. [13] M. Hosoz, M.H. Ertunc, Artificial neural network analysis of an automobile airconditioning system, Energy Conversion and Management 47 (2006) 574e587. [14] H.M. Ertunc, M. Hosoz, Comparative analysis of an evaporative condenser using artificial neural network and adaptive neuro-fuzzy inference system, International Journal of Refrigeration 31 (2008) 1426e1436. [15] S. Yilmaz, K. Atik, Modeling of a mechanical cooling system with variable cooling capacity by using artificial neural network, Applied Thermal Engineering 27 (2007) 2308e2313. [16] Arzu Sencan Sahin, Performance analysis of single-stage refrigeration system with internal heat exchanger using neural network and neuro-fuzzy, Renewable Energy 36 (2011) 2747e2752. [17] M. Mohanraj, S. Jayaraj, C. Muraleedharan, Applications of artificial neural networks for refrigeration, air-conditioning and heat pump systems e a review, Renewable and Sustainable Energy Reviews 16 (2012) 1340e1358. [18] K. Atik, A. Aktas, E. Deniz, Performance parameters estimation of MAC by using artificial neural network, Expert Systems with Applications 37 (2010) 5436e5442. [19] M.T. Hagan, H.B. Demuth, M. Beale, Neural Network Design, PWS Publishing Company, Boston, 1995. [20] R. Ahmad, H. Jamaluddin, M.A. Hussain, Application of memetic algorithm in modelling discrete-time multivariable dynamics systems, Mechanical Systems and Signal Processing 22 (2008) 1595e1609. [21] Haslinda Mohamed Kamar, M.Y. Senawi, Empirical correlations for coil performance of automotive air conditioning system, In: 1st International Meeting on Advances in Thermo-Fluids, 26th August 2008, Universiti Teknologi Malaysia, Malaysia. [22] S. Haykin, Neural Networks: A Comprehensive Foundation, PHI, India, 2003.

Artificial neural networks for automotive air-conditioning systems (2 ...

Artificial neural networks for automotive air-conditioning systems (2).pdf. Artificial neural networks for automotive air-conditioning systems (2).pdf. Open. Extract.

481KB Sizes 1 Downloads 395 Views

Recommend Documents

Using artificial neural networks to map the spatial ...
and validation and ground control points (GCPs) to allow registration of the ..... also like to thank Dr Arthur Roberts and two anonymous reviewers for their.

Using artificial neural networks to map the spatial ...
The success here to map bamboo distribution has important ..... anticipated that a binary categorization would reduce data transformation complexity. 3.2.

Loudness Model using Artificial Neural Networks
Universidad Tecnológica de Chile, Brown Norte 290, Santiago, Chile, [email protected] c. Universidad Tecnológica de Chile, Brown Norte 290, Santiago, Chile, [email protected]. English Translation by A. Osses. ABSTRACT: This article prese

pdf-0946\artificial-neural-networks-in-cancer-diagnosis-prognosis ...
... apps below to open or edit this item. pdf-0946\artificial-neural-networks-in-cancer-diagnosi ... gement-biomedical-engineering-from-brand-crc-press.pdf.

[PDF Online] Fundamentals of Artificial Neural Networks
Online PDF Fundamentals of Artificial Neural Networks (MIT Press), Read PDF .... over 200 end-of-chapter analytical and computer-based problems that will aid ...

Loudness Model using Artificial Neural Networks
Various 3-layered feed-forward networks were trained using the Quasi Newton. Backpropagation algorithm with 2,500 training epochs for each network, having ...

Impact of missing data in evaluating artificial neural networks trained ...
Feb 17, 2005 - mammographic Breast Imaging and Reporting Data System ... Keywords: Diagnosis; Computer-assisted; Mammography; Breast neoplasms .... The Lypaponov energy function was used as a measure of the network stability. ..... Electrical and Ele

Introduction to Artificial Neural Systems by Jacek M Zurada.pdf ...
Introduction to. Art if icial. Neural Sy s. JACEK M. ZURADA. Professor of Electrical Engineering and of. Computer Science and Engineering. WEST PUBLISHING ...

Towards cortex sized artificial neural systems
implementation on a cluster computer is not communication but computation bounded. ...... multiple times came out as a slower communication solution. In Fig.

Introduction to Artificial Neural Systems by Jacek M Zurada.pdf ...
Professor of Electrical Engineering and of. Computer Science and Engineering. WEST PUBLISHING COMPANY. St. Paul New York Los Angeles San Francisco.

Neural Networks - GitHub
Oct 14, 2015 - computing power is limited, our models are necessarily gross idealisations of real networks of neurones. The neuron model. Back to Contents. 3. ..... risk management target marketing. But to give you some more specific examples; ANN ar

Learning Methods for Dynamic Neural Networks - IEICE
Email: [email protected], [email protected], [email protected]. Abstract In .... A good learning rule must rely on signals that are available ...

Neural Networks and Fuzzy Systems by Kosko.pdf
Neural Networks and Fuzzy Systems by Kosko.pdf. Neural Networks and Fuzzy Systems by Kosko.pdf. Open. Extract. Open with. Sign In. Main menu. Displaying ...

Neural Networks and Fuzzy Systems by Kosko.pdf
construction loans wa. If you were not redirected automatically, click here. Page 1 of 2. Page 2 of 2. Neural Networks and Fuzzy Systems by Kosko.pdf. Neural Networks and Fuzzy Systems by Kosko.pdf. Open. Extract. Open with. Sign In. Main menu. Displ

Artificial Neural Network for Mobile IDS Solution
We advocate the idea that mobile agents framework enhance the performance of IDS and even offer them new capabilities. Moreover agent systems are used in ...

Larochelle - Neural Networks 2 - DLSS 2017.pdf
Page 5 of 40. Larochelle - Neural Networks 2 - DLSS 2017.pdf. Larochelle - Neural Networks 2 - DLSS 2017.pdf. Open. Extract. Open with. Sign In. Main menu.

Artificial Neural Network for Mobile IDS Solution (PDF Download ...
Agents is defined as a distinct software process, which. can reason independently, and ..... James P. Anderson Company, (Fort Washington, Pennsylvania, 1980). [2] D. E. Denning, An .... [44] CISCO, http://www.cisco.com.AccessedMarch2008.