Artificial Neural Network Approach for Transient Forced Convective Heat Transfer Optimization

This present research uses artifical neural networks (ANNs) to analyze and estimate the influence of transfer functions and training algorithms on experimentally determined Nusselt numbers, friction factors, entropy generation numbers and irreversibility distribution ratios for nine different baffle plate inserted tubes. Nine baffle-inserted tubes have several baffles with various geometric parameters used in the experiments with a baffle area blockage ratio of two, with different pitch to diameter ratios, different baffle orientation angles and different baffle spacings. The actual experimental data sets were used from previous author’s studies and applied as a input data set of ANNs. MATLAB toolbox was used to search better network configuration prediction by using commonly used multilayer feed-forward neural networks (MLFNN) with back propagation (BP) learning algorithm with thirteen different training functions with adaptation learning function of mean square error and TANSIG transfer function. In this research, eighteen data samples were used in a series of runs for each nine samples of baffleinserted tube. Reynold number, tube lenght to baffle spacing ratio, baffle orientation angle and pitch to diameter ratio were considered as input variables of ANNs and the time averaged values of Nusselt number, friction factor, entropy generation number and irreversibility distribution ratio were determined as the target data. The total 70% of the experimental data was used to train, 15% was used to test and the rest of data was used to check the validity of the ANNs. The TRAINBR training function was found as the best model for predicting the target experimental outputs. Almost perfect accuracy between the neural network predictions and experimental data was achieved with mean relative error (MRE) of 0,000105816% and correlation coefficient (R) that was 0,999160176 for all datasets, which suggests the reliability of the ANNs as a strong tool for predicting the performance of transient forced convective heat transfer applications.


Introduction
Artifical Neural Networks (ANNs) have been successfully used in many engineering applications to simulate nonlinear complex system without requiring any input and output knowledge such as dynamic control, system identification and performance prediction of thermal systems in heat transfer applications. ANN have been widely used for thermal analysis of heat exchangers during the last two decades. The applications of ANN for thermal analysis of heat exchangers are reviewed in detail [1].
The various network architectures were tested in [2] suggesting feed-forward network with log-sigmoid node functions in the first layer and a linear node function in the output layer to be the most advantageous architecture to use for prediction of helically-finned tube performance. A feed forward ANN approach trained by Levenberg-Marquardt algorithm was developed to predict friction factor in the serpentine microchannels with rectangular cross section has been investigated experimentally [3] hybrid high order neural network and a feed forward neural network are developed and applied to find an optimized empirical correlation for prediction of dryout heat transfer. The values predicted by the models are compared with each other and also with the previous values of empirical correlation [4]. ANN is applied for heat transfer analysis of shell-and-tube heat exchangers with segmental baffles or continuous helical baffles. Three heat exchangers were experimentally investigated. Limited experimental data was obtained for training and testing neural network configurations. The commonly used back propagation algorithm was used to train and test networks. Prediction of the outlet temperature differences in each side and overall heat transfer rates were performed. Different network configurations were also studied by the aid of searching a relatively better network for prediction [5]. ANN is used for heat transfer analysis in corrugated channels. A data set evaluated experimentally is prepared for processing with the use of neural networks. Back propagation algorithm, the most common learning method for ANNs, was used in training and testing the network [6]. The capabilities of an ANN approach for predicting the performance of a liquid desiccant dehumidifier in terms of the water condensation rate and dehumidifier effectiveness is proposed [7]. An application of ANNs to characterize thermo-hydraulic behavior of helical wire coil inserts inside tube. An experimental study was carried out to investigate the effects of four types of wire coil inserts on heat transfer enhancement and pressure drop. The performance of the ANN was found to be superior in comparison with corresponding power-law regressions [8]. This paper describes the selection of training function of an ANN for modeling the heat transfer prediction of horizontal tube immersed in gas-solid fluidized bed of large particles. The ANN modeling was developed to study the effect of fluidizing gas velocity on the average heat transfer coefficient between fluidizing bed and horizontal tube surface. The feedforward network with back propagation structure implemented using Levenberg-Marquardt's learning rule in the neural network approach. Performances of five training functions implemented in training neural network for predicting the heat transfer coefficient [9]. It is reported the results of an experimental investigation to characterize the thermal performance of different configurations of phase change material based pin fin heat sinks. An ANN is developed to determine the optimal configuration of the pin fin heat sink that maximizes the operating time for the n-eicosane based heat sink [10]. A non-iterative method is applied utilizing ANN and principal component analysis to estimate the parameters that define the boundary heat flux. The inversion has been accomplished by employing a non-iterative method using ANN and principal component analysis. The potential use of covariance analysis in reducing the dimensionality of the inverse problem has also been demonstrated [11]. A generalized neural network analysis for natural convection heat transfer from a horizontal cylinder is developed and a threelayer network is used for predicting the Nusselt number. The number of the neurons in the hidden layer was determined by a trial and error process together with cross-validation of the experimental data evaluating the performance of the network and standard sensitivity analysis [12]. Heat transfer correlation developed [13] to assist the heat exchanger designer in predicting the heat transfer coefficient along a horizontal straight circular tube with uniform wall heat flux for a specified inlet configuration in the transition region by using ANN. An application of ANNs was presented to predict the pressure drop and heat transfer characteristics in the plate-fin heat exchangers [14]. A new and detailed three-layer BP network model for prediction of performance parameters on prototype wet cooling towers is developed successfully in this paper, and the improved BP algorithm, the gradient descent algorithm with momentum, is used in [15]. The results of an experimental investigation carried out to characterize the thermal performance of different configurations of phase change material based pin fin heat [16]. In the research ANN approach has been utilized to characterize the thermo hydraulic behavior of corrugated tubes combined with twisted tape inserts in a turbulent flow regime. The experimental data sets have been utilized in training and validation of the ANN in order to predict the heat transfer coefficients and friction factors inside the corrugated tubes combined with twisted tape inserts, and the results were compared to the experimental data [17]. ANNs are utilized to compile values of the mean Nusselt number particularized to binary gas mixtures in the Prandtl number sub-region. Thereafter, these values are used to generate a heat transfer correlation that is obtained from using a combination of available data and predicted values [18]. A linear regression approach was used to correlate experimentally-determined Colburn j-factors and Fanning friction factors for flow of liquid water in helically-finned tubes. The principal finding of the [19] investigation is the fact that in helically-finned tubes both Fanning friction factors and Colburn j-factors can be correlated with exponentials of linear combinations of the same five simple groups of parameters and a constant. The ANNs has been applied for the unsteady heat transfer in a rectangular duct for the prediction of unsteady heat transfer in a rectangular duct [20]. An experimental study has been carried out to investigate the axial variation of inlet temperature and the impact of inlet frequency on decay indices in the thermal entrance region of a parallel plate channel. The investigation was conducted with laminar forced flows.
Despite the fact that comprehensive studies were conducted on heat transfer applications in the literature, the research studies concerning the effectiveness and comparision of different ANN models considering transfer functions and training algorithms in the broader sense are not sufficient. The main focus of the present study is based on the experimental data obtained from author's previous studies [21][22][23] for optimizing transient forced convective heat transfer for turbulent flow in a circular tube with baffle inserts using tangent sigmoid TANSIG function and thirteen training algorithms to predict ANN performances based on mean relative error ( ) and correlation coefficient ( ) for all data sets.

Experimental Setup
The experimental setup illustrated in Figure 1 is used for data gathering for the heat transfer analysis. A detailed description of the experimental setup is avaliable in some of author's previous researches in detail [21][22][23].    The detailed geometrical parameters of baffled tubes were tabulated in Table 1. The heat loss calibration tests were performed before taking measurements on the system for each type of baffle inserted tubes in the following manner. Each baffle inserted tube was completely filled up with insulation materials of multi layer glass wool and constant heat flux was supplied through the pipe wall by means of PLC integrated DC power supply. Type 1  18031  180  31  1  Type 2  18062  180  62  2  Type 3  18093  180  93  3  Type 4  9031  90  31  1  Type 5  9062  90  62  2  Type 6  9093  90  93  3  Type 7  4531  45  31  1  Type 8  4562  45  62  2  Type 9  4562  45  93  3 The average wall temperatures were evaluated at eleven points along the test section in terms of heat flux, difference in wall and ambient temperatures. The time averaged wall temperature variations by time were recorded using data online acquisition system. When the steady state condition is established to insure that external thermal equilibrium can be achieved, heat loss calibration tests for different values of power supply are reported for a steady state case. It was found that the heat loss is directly proportional to the difference between the wall and ambient temperatures. The required constant of proportionality was taken from the previously determined heat loss calibrations. It was observed that the maximum heat loss did not exceed %5 all through the test runs. More detailed explanation of the heat loss calibration technique was given by [21][22][23].

Data Reduction
Data reduction of baffle inserted tubes presented above in Figures 2-4. is avaliable in author's previous researches in detail [21][22][23] for fully developed turbulent flow by using ANNs. The independent parameters are Reynolds number and tube diameter. The Reynolds numbers based on the tube hydraulic diameter are given by, The average fully developed heat transfer coefficients are evaluated as follows, where A is convective heat transfer area. Nusselt numbers and friction factor for fully developed turbulent flow are evaluated by using Eq. (4) and Eq. (5) respectively. (3) where is pressure gradient.
It is considered a circular pipe flow with constant heat flux q ɺ and cross sectional area, A as shown schematically in Figure 5. For an incompressible viscous fluid with mass flow rate, passes through the pipe of lenght of !. In general this heat transfer arrangement is characterised by an control volume on a finite lenght of " pressure gradient is / " $ 0, by a finite wall bulk fluid temperature difference ∆' . For an investigated heat transfer region, the rate of entropy generation per unit lenght expression is /0 . 12 + 345 6 7 / . 8 . 9 : where the first term on the right hand side is the contribution made by heat transfer, while the second term is the contribution due to fluid friction that is The second law requires ( )*+ $ 0 for all real processes. Since classical macroscopic thermodynamics does not provide any theoretical way to calculate entropy generation of irreversible process directly, the only way to determine how much ( )*+ is greater than zero is to use data obtained from experiments. To describe the effects of the flow conditions and geometry parameters of Reynolds number , Prandtl number = , pitch to diameter ratio >/ , baffle orientation angle ?, ratio of smooth to baffled crossection area ( @ /( A and ratio of tube length to baffle spacing !/> on transient forced convection heat transfer for turbulent flow in a circular tube with baffle inserts, time averaged Nusselt number, time averaged friction factor, time averaged entropy generation per unit time and irreversibility distribution ratio are related as follows: In Eq. (7), the Prandtl number =, which should be an important parameter affecting the heat transfer of baffle inserted tube, is defined as: But, Prandtl number has not been separately considered in this investigation because air is only used as working fluid and its Prandtl number in the considered experimental range temperature range remains almost constant. So, Eq. (7) can be simplified as: Similarly, time averaged friction factor and time averaged entropy generation per unit time is related as given below.
An important dimensionless parameter in the second law analysis of convective heat transfer is the irreversibility distribution ratio [25] is: The parameter ∅ describes the relative importance of fluid friction in the total irreversibility of a flow passage. As it is known augmentation entropy generation number W can be rewritten as substituting Eqs. (17) and (16) where Reynolds-Colburn analogy between heat transfer and momentum for turbulent flow is given by Introducing Eq. (21) into Eq. (19), irreversibility distribution ratio is obtained for the turbulent flow as Eq. (22) permits a quick estimate of ∅, without having to calculate the Reynolds number [26]. For the case of W 1 where the the rate of entropy generation per unit lenght of smooth pipe is equal to the rate of entropy generation per unit lenght of bafle inserted augmented pipe irreversibility distribution ratio ∅, can be expressed as For the present case proposed augmentation technique having minimum area under the graph of ∅ versus can be optimally selected in order to yield the maximum reduction in heat exchanger duct irreversibility called irreversibility minimization analysis.

Experimental Uncertainity Analysis
The uncertainties of experimental quantities were computed by using the method presented [27]. The uncertainty calculation method used involves calculating derivatives of the desired variable with respect to individual experimental quantities and applying known uncertainties. The general equation presented by [27] showing the magnitude of the uncertainty in ( ) is where (" D , " 4 , … . . " + ) and " + is the variable that affects the results of . The experimental results up to a Reynolds number of 20000 were correlated with a standard deviation of 5% at most. Experimental uncertainties in the Reynolds number, friction factor, and Nusselt number were estimated by the above procedure described [27]. The mean uncertainties are g2.5% in the Reynolds number, g4% in the friction number.
The highest uncertainties are g9% in the Nusselt number for the type 9031. Uncertainties in the Nusselt number range between g5% and 8% for 3000 ≤ ≤ 20000 at the type 18093 and g8% and 10% 3000 ≤ ≤ 20000 at the type 9031, highest uncertainties being at the lowest Reynolds number [21][22][23]. ANN is a numerical model that simulates the human brain's biological neural network ability to learn and recognize complicated nonlinear functions. This learning ability makes the ANN more powerful than the parametric approaches. ANN usage in heat transfer applications is popular because of its functional approximation between the inputs and desired outputs. In this present study a MLFNN with BP learning algorithm [28] has been used. It is simple and high learning rates; therefore it is widely used to train the networks.

Development of Artificial Neural Network
The ANN model was developed for the system with four independent parameters in the input layer (Reynold number, tube lenght to baffle spacing ratio, baffle orientation angles and pitch to diameter ratio), four parameters (time averaged values of Nusselt number, friction factor, entropy generation number and irreversibility distribution ratio) and ten neurons in hidden layer. The architecture of the network for this current study is shown in Figure 6.
Neural network tool in the MATLAB R2011b version is used for ANN modelling of the system. There are fourteen different back propogation (BP) training algorithms in MATLAB ANN toolbox [29]. In this study, multilayer feedforward neural networks (MLFNN) with back propagation (BP) training and validation algorithms were applied for each of thirteen different training functions given in Table 2. TRAINBFG BFGS quasi-Newton method. Requires storage of approximate Hessian matrix and has more computation in each iteration than conjugate gradient algorithms, but usually converges in fewer iterations [39,40]. TRAINGDX Adaptive learning rate. Faster training than TRAINGD, but can only be used in batch mode training.

Normalization of Experimental Data
It is desirable to normalize all the input and output data with the largest and smallest values of each of the data sets, since the variables of input and output data have different physical units and ranges. So, all of the input and output data were normalized between 0.1 and 0.9 due to restriction of sigmoid function [41][42][43] using the below rearranged formula as follows: where the is the measured value, while 5v+ and 5A] values are the minimum and maximum values of found in the train set and also employed data for normalization are given shown in Table 3. TANSIG transfer function gives better results than logarithmic sigmoid function (LOGSIG) according to present investigation as mentioned [44]. TANSIG transfer function is being used as an activation function in the hidden layer of ANN [24] is given as

Results and Discussion
MATLAB toolbox was used to search better network configuration prediction by using commonly used feed forward back propagation algorithm with thirteen different training functions with adaptation learning function of MSE and TANSIG transfer function. In this research, eighteen data samples were used in a series of runs for each nine samples of baffle-inserted tube. Reynold number, tube lenght to baffle spacing ratio, baffle orientation angle and pitch to diameter ratio were considered as input variables of ANNs and the time averaged values of Nusselt number, friction factor, entropy generation number and irreversibility distribution ratio were determined as the target data. Up to 70% of the whole experimental data was used to train the models, 15% was used to test the outputs and the remaining data points which were not used for training were used to evaluate the validity of the ANNs. As mentioned above the ANN was trained using all possible thirteen different training functions avaliable in MATLAB toolbox. To determine the optimal neural network structure, both the error convergence rates was checked by changing the number of hidden layer and also by decreasing momentum rate ranged from 0.9 to 0.7 in successive decreasement of 0.025 to increase learning rate of the networks. Based on the analysis, it was observed that the optimal number of hidden neurons varies mostly from one training function to another one but the optimal momentum rate was found to be 0.825 for all training functions. TRAINBR training function has shown better performance as compared to other twelve training functions under the constant network parameters. Constructed configuration of TRAINBR network has ten neurons in the hidden layer as shown in Figure 7. The absolute fraction of variance values ( 4 ) and optimal number of hidden neurons for each training function were determined and tabulated in Table 4. Training regression plots for the best training algorithm of TRAINBR are shown in Figure 8. Thirteen different ANN training models have been compared by mean square error (MSE), mean relative error (MRE) and absolute fraction of variance ( 4 ) mathematically expressed as following equations: where Z v is the actual (experimental) value, X v is the predicted (output) value and } is the number of the data. The networks were trained for all thirteen different training functions under same network parameters. The training was continued till the least value of MSE at a definite value of epochs attained for all thirteen different training functions seperately. The use of the MSE is an excellent numerical criterion for evaluating the performance of a prediction tool. Table 5 shows the results for the MRE, MSE and R 2 values for different training algorithms. After analysing all the results, TRAINBR training function has shown best performance as compared to other twelve training functions for predicting the target experimental outputs which has the least MSE value. The graphs in Figures 9-12. generated by using friction factor, Nusselt number, entropy generation number and irreversibility distribution ratio values that appear in all the tested ANN training algorithms with respect to Reynolds numbers respectively.    Best training performance plot for the best training algorithm of TRAINBR is shown in Figure 13. This figure is the performance plots of the mean square error value versus the number of epochs that is iteration numbers. Mean square error decreases with increasing iteration numbers and converges to a steady state value based on the TRAINBR algorithm characteristic as the best training performance is achieved at 246 epochs. The training state of the best training algorithm of TRAINBR is shown in Figure 14. In this graph it is clearly shown that optimized network is developed with mean squared error of 9.99956x10 -9 and sum of squared network parameters found to be 308.7523. The performance goal of optimized network having 89.8305 parameters is achieved in 246 epochs.

Figure 14. Performance plots of the best training function of TRAINBR.
A comparison of predicted values using best training function TRAINBR and the experimental values of the system is given in Table 6 for performance evaluation. The deviation values (MSE, MRE, and 4 ) of thirteen different training functions for estimation of ANNs are presented in Table 7. A well trained ANN model produces small MSE and large 4 values. According to this table, the optimal network configuration which is TRAINBR training function has a lower MSE and higher 4 values. A parity plots of the output layer parameters are drawn to show the performance of optimal ANN TRAINBR training function in Figures 15-18. All of the graphs clearly show that the TRAINBR training function works very well. Based on these figures and MSE values of Table 5, parity plots show the accuracy with which the optimal ANN TRAINBR training function predicts output layer parameters of friction factor, Nusselt number, entropy generation number and irreversibility distribution ratio obtained from the experimental outputs. The coefficient of determination values 4 for best training function TRAINBR has achieved unity for all outputs. The results show that the optimal neural network configuration TRAINBR training function is successful in predicting the solution of transient forced convective heat transfer problems to determine friction factor, Nusselt number, entropy generation number and irreversibility distribution ratio.

Conclusions
In this paper, the performance of transient forced convection heat transfer with nine various baffle inserted tubes have been analyzed to determine optimal training function by using commonly used MLFNN with BP learning function with thirteen different training function with adaptation learning function of mean square error and TANSIG transfer function. The importance of this study is to develop an optimal ANN configuration between thirteen different ANN configurations using an actual experimental data set and to develop an optimal ANN architecture as well.
The ANN architecture consists of four independent parameters in input layer and four dependent parameters in output layer. It is obvious that all of the the training functions are in good agreement with the experimental data set but TRAINBR training function is the best training function for prediction of output layer parameters. Almost perfect accuracy between the TRAINBR neural network training function predictions and experimental data was achieved with mean relative error ( ) of 0,000105816% and correlation coefficient ( ) that was 0,999160176 for all data sets, which suggests the reliability of the ANNs as a strong tool for predicting the performance of transient forced convective heat transfer applications.