Performance prediction of rotary solid desiccant dehumidifier in hybrid air-conditioning system using artificial neural network

Applied Thermal Engineering 98 (2016) 1091–1103

Contents lists available at ScienceDirect

Applied Thermal Engineering j o u r n a l h o m e p a g e : w w w. e l s e v i e r. c o m / l o c a t e / a p t h e r m e n g

Research Paper

Performance prediction of rotary solid desiccant dehumidifier in hybrid air-conditioning system using artificial neural network D.B. Jani *, Manish Mishra, P.K. Sahoo Department of Mechanical & Industrial Engineering, Indian Institute of Technology Roorkee, 247667, India

H I G H L I G H T S

• • •

Artificial neural network is used to predict the performance of rotary dehumidifier. Experimental tests are carried out to study the performance of a solid desiccant dehumidifier. The ANN predicted outputs gave high correlation to the experimental data.

A R T I C L E

I N F O

Article history: Received 8 July 2015 Accepted 23 December 2015 Available online 13 January 2016 Keywords: ANN Desiccant wheel Effectiveness Regeneration

A B S T R A C T

Desiccant air conditioning systems are considered as better alternatives to the conventional air conditioning system because of the independent control of temperature and humidity and being environment friendly. An artificial neural network (ANN) model has been developed to predict the performance of a rotary desiccant dehumidifier for different process air inlet conditions. Dry bulb temperature, humidity ratio and flow rate of the process as well as regeneration air streams of dehumidifier and regeneration temperatures are used as inputs to the model. The outputs of the model are outlet dry bulb temperature and humidity ratio of process as well as regeneration air streams of dehumidifier, the moisture removal rate and the effectiveness of the dehumidifier. Moisture removal rate and effectiveness of the dehumidifier are considered as the performance indicators of the system. Experiments are also conducted to investigate the performance of the desiccant wheel and the test results are used as target data to train the ANN model. Performance predictions through ANN are compared with the experimental test results and a close agreement is observed. © 2015 Elsevier Ltd. All rights reserved.

1. Introduction In a typical desiccant based cooling system, the moisture (latent load) in the process air is removed by desiccant dehumidifier and then the temperature (sensible load) of the dried process air is reduced to the desired comfort conditions by sensible coolers. Thus, the latent and sensible loads are handled separately. As compared to traditional vapor compression systems, desiccant systems eliminated the use of chlorofluorocarbons (CFCs) as CFCs are the major contributors to the depletion of ozone layer. The desiccant based cooling system allows larger flow rates of ventilation air to improve indoor air quality. Moreover, it also maintains lower humidity levels and removes air borne pollutants. The desiccant cooling system can be cost effective in combination with waste heat source, very effective in humidity control when latent load is high and in dry cooling coils and ducting to avoid microbial growth. Desiccant based cooling systems are used in several applications such as pharmaceutical

* Corresponding author. Tel.: +91-9428044640; fax: +91-1332-285665. E-mail address: [email protected] (D.B. Jani). http://dx.doi.org/10.1016/j.applthermaleng.2015.12.112 1359-4311/© 2015 Elsevier Ltd. All rights reserved.

plants, supermarkets, theatres, hotels, office buildings, hospitals, health clubs and swimming pools. Several studies have been carried out on desiccant cooling systems by different researchers. Pennington [1] proposed the earliest desiccant cooling cycle by coupling dehumidifier with heat source and evaporative cooler. Similar cycle was proposed by Dunkle [2] using dehumidifier of molecular sieve with additional heat exchanger. Later on Munter [3] improved the performance of the desiccant cooling cycle by introducing parallel passages in the dehumidifier and provided backup of vapor compression system to tackle cooling load if desiccant cooling cycle does not meet the cooling demand. Since then, a number of efforts have been made in order to understand and analyze the performance of desiccant dehumidifiers and the desiccant cooling systems. Important among those were the analogy theory by Banks [4], the pseudo-steady state model by Barlow [5], finite difference method for cross-cooled dehumidifiers by Worek and Lavan [6] and the finite difference method by Maclaine-Cross [7], which are now widely used by other researchers in getting better performance from rotary desiccant dehumidifier [8]. Jurinak [9] studied the effect of desiccant matrix properties on the ventilation cycle performance. Burns et al. [10]

1092

D.B. Jani et al./Applied Thermal Engineering 98 (2016) 1091–1103

combined the desiccant dehumidifier with vapor compression systems for handling moisture removal and the temperature of cooling air separately. Van den Bulck et al. [11] experimentally investigated the effects of process and regeneration air mass flow rates, temperatures, humidity ratios etc. on the performance of solid desiccant dehumidification system. Open cycle solid desiccant cooling system has been numerically modeled by Dhanes and Wiliam [12] to determine the effect of various geometric and operating parameters on the COP of the system. Pesaran [13] proposed the use of solar and waste heat to increase the cycle efficiency. Further, a comparison between various desiccant cooling cycles for air conditioning in hot and humid climates has been carried out by Jain et al. [14]. Desiccant wheel is the most important component of any solid desiccant air conditioning system. Effectiveness and moisture removal capacity (MRC) are mostly used as index to determine the performance of solid desiccant wheel. Different approaches have been used by the researchers to predict the performance of desiccant wheel mainly numerical and experimental ones. Experimental results are always reliable, but consume significant time and effort. Numerical models alone hardly predict the exact results, and should be verified through experimental results. Moreover, for the complex thermal system such as dehumidifier, mathematical models become very complicated to solve and to predict the performance. Thus, instead of solving complex equations numerically and in case limited experimental data are available, faster and simpler solutions are obtained with the artificial neural network [15]. Use of artificial intelligence techniques in preparation of models [16] has become the preferred trend for many researchers [17] in predicting the performance of combined heat and mass transfer processes in the field of desiccant dehumidification because of their ability to learn and to adapt to change with little human interaction. Artificial neural network based model can be opted as a simplified option for predicting the temperature and humidity ratio of air stream at the outlet of desiccant dehumidifier. Cejudo et al. [18] developed a neural network model to calculate the outlet temperature and humidity ratio in silica gel desiccant dehumidifier. It was based on training of a black box model with the experimental measurements. The model used a four-input fouroutput network for calculating the outlet conditions and was successfully validated by experimental results within reasonable agreement. Kalogirou [19] applied artificial neural network model to determine energy conservation in HVAC devices in a building. ANN was used as a tool in the prediction and modeling of building energy systems. Performance of the model was tested with the experimental data obtained from a real system. A multiple layer artificial neural network (ANN) model was applied to study the performance of a liquid desiccant dehumidification system [20]. The experimental results were used to construct and to test the ANN model. The model was further utilized to describe the effect of the inlet conditions of air and calcium chloride solution on the regeneration process. Good agreement was found between the outputs from the ANN model and that from experimental data. It also shows that the proposed model can work well as a predictive tool to complement the experiments. Mohammad et al. [21] also used the artificial neural network (ANN) model for predicting the performance of a liquid desiccant dehumidifier using air and desiccant inlet parameters as inputs to the ANN model. The maximum differences between the ANN and the experimental values were found to be 8.13%. Parmar and Hindoliya [22] also used the concept of artificial neural network to evaluate the performance of a solid desiccant cooling system. Gandhidasan and Mohandes [23] proposed an ANN model to simulate the relationship between inlet and outlet parameters of a dehumidifier. A multilayer ANN model was used to investigate the performance of dehumidifier by using different parameters such as air flow rate, temperature, humidity ratio etc. ANN

predictions for these parameters were validated with the experimental results with a reasonable degree of accuracy. A neural network model based on experimental results for predicting the outlet conditions of a solid desiccant wheel, such as temperature and humidity ratio, was proposed by Koronaki et al. [24]. An artificial neural network was developed by Mohammad et al. [25] to predict the performance of a dehumidifier. ANN predictions for the parameters like temperature, humidity ratio, flow rate etc. were compared with the experimental results. Alosaimy [26] also compared the artificial neural network (ANN) model and the experimental data for all the output variables of desiccant dehumidifier with respect to the air inlet temperature and a good agreement was reported. Uckan et al. [27] proposed an artificial neural network (ANN) model to predict the performance of a solid desiccant dehumidifier made up of silica gel in terms of dry bulb temperature and specific humidity of air at the outlet. It shows that the predicted values obtained using this model are very close to the experimental values and the errors are within the acceptable limits. To the best of authors’ knowledge, none of the previous researchers predicted the performance of solid desiccant dehumidifiers in terms of moisture removal rate and effectiveness using artificial neural network (ANN). The previous studies have considered the ANN prediction only for the outlet air stream parameters namely temperature and humidity ratio based on inlet air stream parameters and experimental measurements in desiccant dehumidifier used in desiccant dehumidification system with evaporative cooling. In the current study, a neural network model has been developed using a neural network toolbox of MATLAB© [28] with feed forward back propagation method based on the experimental results for rotary solid desiccant dehumidifier fitted in a solid desiccantvapor compression hybrid air-conditioning system. The effect of ambient on dehumidifier performance was also studied in detail. Experiments have been performed to evaluate the performance at varying inlet temperature, humidity ratio and flow rate. The experimental data are used for training and prediction of the ANN model and for validation against the experimental results. The proposed ANN model can efficiently predict the outlet temperature, humidity ratio, moisture removal rate and effectiveness of desiccant dehumidifier within the range of experimental test data for varying process air inlet conditions.

2. System description A test room, with the following dimensions: 3 m × 3 m × 3 m, has been selected for the study. The sensible and the latent cooling loads [29] are taken as 1.371 kW and 0.391 kW, respectively. Sensible heat ratio (SHR) has been obtained as 0.78 [30]. Flow rates of the process air stream and the regeneration air stream are measured as 322.7 m3/ hr and 196.8 m3/hr, respectively. The comfort conditions are taken as 50% relative humidity and 26 °C dry bulb temperature [31]. The schematic diagram and the photographic view of solid desiccant and vapor compression hybrid air-conditioning system have been shown in Figs. 1 and 2, respectively. The return room air at state 1 passes through the rotary desiccant dehumidifier. Its moisture is adsorbed significantly by the desiccant material and the heat of adsorption raises its temperature up to state 2. The hot and dry air is first sensibly cooled in an air-to-air heat exchanger (2–3) and then in cooling coil of VCR system up to state 4. In the regeneration air line, ambient air at state 6 enters the air-to-air sensible heat exchanger and cools the supply process air. Consequently, its temperature rises when exiting from sensible heat exchanger at state 7. At this point, it is heated to reach temperature at the state point 8, which is high enough to regenerate the desiccant material. Moist air at the outlet of dehumidifier is exhausted to atmosphere at state 9.

D.B. Jani et al./Applied Thermal Engineering 98 (2016) 1091–1103

1093

Test room 5 Return air

Conditioned air

Condenser

Expansion valve Desiccant wheel

Compressor Supply fan

3

1

4

2

Exhaust air

Heater 8

VCR sensible cooling coil

7

9

Ambient air 6

Exhaust fan

+

Heat recovery wheel

Fig. 1. Schematic diagram of solid desiccant – vapor compression hybrid air-conditioning system.

The rotary desiccant dehumidifier used is 360 mm diameter and 100 mm width. The rotational speed of the dehumidifier is kept constant at 20 rph. Synthesized metal silicate is the desiccant material used in desiccant wheel. 3. Measurements Experiments are carried out by simultaneous measurement of temperature, relative humidity, pressure drop and flow rate with the help of multifunctional temperature, humidity and velocity digital transmitters connected via Masibus-85XX micro-controller based scanner with control panel, to control and operate the system. All the sensors are connected to a central computer via data acquisition unit. The inaccuracies in measurement of temperature, relative humidity and flow rate are found ±0.3K @ 296K, ±2.0%, ±3.0%, respectively. Energy meter is also used to measure the electrical power consumption of the system. The measurements were carried out once the temperature and humidity of the system attain steady state condition. Humidistat is fitted inside the test room to control the dehumidifier operation according to the room humidity. Temperature controller is also fitted inside the test room to control the compressor operation through relay, so as to maintain the room temperature.

4. Performance parameters The performance of rotary desiccant dehumidifier is evaluated by calculating the moisture removal rate and effectiveness. The moisture removal rate or moisture transfer rate [21] from air to the desiccant surface was defined as follows:

Δm = ma ( Win − Wout )

(1)

where, ma is the mass flow rate of process air at the dehumidifier inlet, Win and Wout are the humidity ratios of process air at inlet and outlet respectively. The effectiveness (εdw) of the dehumidifier is evaluated as the ratio of the change in actual humidity ratio of the air to the maximum possible change in humidity ratio [32]

εDW =

W1 − W2 W1 − W2,ideal

(2)

where W2,ideal is the ideal humidity ratio of the air stream at the outlet of the desiccant dehumidifier. By assuming that the air is completely dehumidified at this point, the value of W2,ideal is taken as zero.

1094

D.B. Jani et al./Applied Thermal Engineering 98 (2016) 1091–1103

D C

5 F

E G 3

Conditioned air supply

4

H G

By pass line

Exhaust air out G

7

9 8

G

2 Ambient air in

Process air in

1

A

B

6 A- Dehumidifier unit B- Heat recovery wheel C- VCR cooling coil

D- Outdoor unit E- Data scanner F- Test room

G- Measuring Instruments H- Bypass flow control

Fig. 2. Photographic view of the experimental system.

5. Artificial neural network model A neural network model consists of large number of processing elements called neurons. They are interconnected by communication links called weights. A simplified ANN model has an input layer, an output layer, and at least one hidden layer. The selection of layer is determined by the form of the network and the method of input data required. A simplified neural network model (Fig. 3) consists of three basic elements; synapses or connecting link, summing node with a squashing function and an externally applied bias to increase or decrease the net input of the activation function. The network performance is determined by the weights and biases value

in every single neuron. The network needs to be trained to give the desired output using input data sets. The outputs from the ANN model are compared with the actual (experimental) output. There may be a difference between the network’s output and the target output. The weights are adjusted such that the error function minimizes the differences between actual experimental outputs and model outputs. This process is continued until the error function comes under the desired tolerance limit. This repetitive process of training and correction of the weights is known as back propagation algorithm. While training the ANN model, the weights and bias that minimize the error between the measured output and the ANN network output are obtained as follows

Fig. 3. Generalized structure of an artificial neural network.

D.B. Jani et al./Applied Thermal Engineering 98 (2016) 1091–1103

⎡N ⎤ Y = F (S ) = F ⎢ ∑ X K w K + b ⎥ ⎣ K=1 ⎦

(3)

For the construction of network architecture, the neural network tool box of MATLAB© has been used. ANN structure can be made by the weights feeding into each layer of layered structure by choosing the weight matrices. The values of these weights are determined using the feed forward back propagation method. Error calculation is iterated using different inputs and outputs during learning until the root mean square error of the network reaches to an acceptable level. The performance prediction through the neural network is done by mean square error (MSE) [21] (in %) between the predicted and the actual (experimental) values as per the following expression N

MSE =

∑X

predicted

− X experimental

i=1

1095

Start

Data input

Data check

Establish the structure of ANN

2

(4)

N

Training parameters The lesser the MSE, the better fit the results will be. The root mean square error [26] is given by

RMSE =

1 N 2 ∑ (ai − p i ) N i=1

(5)

where, ai and pi are number of nodes for actual output and predicted outputs, respectively. The absolute fraction of variance (R2) [33], a statistical indicator that can be applied to multiple regression analysis, is determined from

⎛ N ⎞ (ai − p i )2 ⎟ ∑ ⎜ R 2 = 1 − ⎜ i=1 N ⎟ 2 ⎜ ⎟ ⎜⎝ ∑ ( p i ) ⎟⎠ i=1

(6)

R2 ranges between 0 and 1. A very good fit yields R2 value of 1 or closer, whereas a poor fit results to a value near 0. Fig. 4 illustrates the flow chart that describes the development, training and simulation of an artificial neural network. The experimental results are the input parameters to the model. Neural network understands the underlying correlations in the entered input data and stores them as inter-neuron connection strengths or corrected weights. Based on the number of neurons, the number of iterations and the desired accuracy, the training set and the target set are developed. The network needs to be trained using training data set consisting of a group of input data and corresponding output data. In the present artificial neural network model, seven input parameters are used namely inlet temperature, humidity ratio and flow rate of process air as well as ambient air, and regeneration air temperature while outlet temperature and humidity ratio of process air and regeneration air, moisture removal rate (MRR) and effectiveness of rotary desiccant dehumidifier are fixed as the output parameters, important in performance studies of the rotary desiccant dehumidifier. Training involves the revision of synaptic weights. The network reads and processes each set of input data and produces an output, which is compared with the actual experimental output. Based on the difference between the network output and the target output, the model parameters are adjusted so that the network would exhibit the targeted results. The network performance was largely determined by the weights and bias values in every single neuron. The three-layer back propagation (BP) network is shown in Fig. 5. wj,i represents the weights between input layer vectors and hidden layer vectors, and υk,j represents the weights between hidden layer vectors and output layer vectors. The training process requires a proper set of data i.e. input and target output. The input layer has seven nodes namely inlet temperature, humidity ratio and flow rate of process air and ambient air respectively, and regeneration air temperature.

Train the ANN

Simulate the ANN model

Accuracy calculation

NO Lowest MSE YES Save output

End Fig. 4. Flow chart of the developed simulation model.

The output layer has six layers namely outlet temperature and humidity ratio of process air and regeneration air, respectively, including the effectiveness and moisture removal rate of rotary desiccant dehumidifier. Back propagation is a multilayer feed forward network with hidden layers between the input and output. The hidden layer has seven nodes. During training, the weights and biases of the network are iteratively adjusted to minimize the network performance function. The training parameters used for the simulation of ANN model are summarized in Table 1. TRAINLM, MSE, TANSIG are the training, performance and transfer functions respectively in the simulation. Typical performance function that is used for training feed forward back propagation neural network is the network mean square errors (MSE). During training, the weights and biases of the network are iteratively adjusted to minimize the network performance function. LM is the fastest training algorithm for network of moderate size and it has the memory reduction feather to be used when the training set is large. While mu is the iteration step size and epoch is a run through all training input-output sets.

1096

D.B. Jani et al./Applied Thermal Engineering 98 (2016) 1091–1103

i

wj,i

Tproin

j υk,j

k Tproout

Wproin Wproout Vpro Tregout Tambin Wregout Wambin

ε Vreg ∆m Tregin

Input layer

Hidden layer

Output layer

Fig. 5. The structure of trained ANN model for rotary desiccant dehumidifier.

Table 1 Training parameters used for the ANN model. Sr. No.

Training parameter

Type/value

1 2 3 4 5 6 7 8 9 10 11

Training function Performance function Transfer function Epochs min_grad Mu mu_dec mu_inc max_time Goal max_fail

TRAINLM MSE TANSIG 1000 0.000001 0.001 0.1 10 inf 0.000000000000001 50

The trained artificial neural network model is shown in Fig. 6. This shows 7-7-6-6 network structure with 7 hidden layers yielded the good model for the accurate prediction of outputs, with minimum MSE during training because it has higher stability and faster convergence rate. This trained network can be used for simulating the system outputs for the inputs that have been introduced before. 6. Results and discussion 6.1. The correlation analysis of simulated ANN model The ANN model was trained through the back propagation technique with TRAINLM, LEARNGDM, MSE and TANSIG as training,

Fig. 6. The trained ANN model for rotary desiccant dehumidifier.

D.B. Jani et al./Applied Thermal Engineering 98 (2016) 1091–1103

Table 2 Operating parameters used for generating the data. Sr. No.

Operating parameter

Operating range

1 2 3 4

Process air inlet temperature (°C) Process air inlet humidity ratio (g/kg) Process air flow rate (m3/hr) Ambient air inlet temperature (°C) Ambient air inlet humidity ratio (g/kg) Regeneration air flow rate (m3/hr) Regeneration air temperature (°C) (at the state point 8 in Fig. 1)

24.1–27.3 8.37–12.28 249.57–374.36 27.6–30.2 15.17–21.31 148.71–196.83 114.50–141.5

6 7

learning, performance and transfer functions, respectively. Seven parameters namely temperature, humidity ratio and flow rate of process and ambient air streams respectively (at state point 1 and 6 in Fig. 1), and regeneration temperature (at state point 8 in Fig. 1) were employed at the input layer while six parameters namely outlet air temperature and humidity ratio of process and regeneration air streams (at state point 2 and 9 in Fig. 1), effectiveness and moister removal rate of solid desiccant dehumidifier were employed at the output layer. The effectiveness and moisture removal rate are the important parameters in studying the performance of rotary desiccant dehumidifier. The experimental range of the operating parameters used for generating the data is shown in Table 2. The artificial neural network (ANN) model has been trained to estimate the model outputs. The experimental results were used to train the feed forward neural network. While a network is being trained, MATLAB automatically generates a performance curve, regression curve, and a cross-validation error diagram. These can be useful to examine whether the network has properly trained. If these plots are acceptable, the network may be exported to the working directory in MATLAB for simulation. Fig. 7 shows the performance graph of training process. The performance graph describes the plot of mean square error (MSE) against the number of epochs (a run through all training input–output sets) or iterations. Since the validation and test curves are very similar, therefore, no major problem or over-fitting occurred with the training. Moreover, Fig. 7 shows the decrease of error over the epochs or iterations carried out by

1097

the network during training, validation and testing. The neural network training process was terminated at the 60th epoch where 50 consecutive cross validation errors occur. The best performance was obtained at the 10th epoch at which the MSE during validation was found to be 1.8323. The best performance in terms of MSE is found to be at epoch 10 because of the occurrence of minimum MSE of the validating sets. Thus, the performance plot is an important diagnostic tool to plot the training, validation and test errors so as to check the progress of training. Further, the training state of the system showing the gradient, mutation and validation check graphs for ANN is shown in Fig. 8.The magnitude of the gradient and the number of validation checks are used to terminate the training. The gradient will become very small as the training reaches a minimum of the performance. If the magnitude of the gradient is less than 1e-5, the training will stop. This limit can be adjusted by setting the parameter. The gradient fluctuates repeatedly, but overall reduces throughout the training process as shown in plot (a). Plot (b) shows the learning rate (mutation) against increasing numbers of iterations. This plot shows that the network error also fluctuates repeatedly, but overall reduces throughout as training progresses. The number of validation checks as shown in plot (c) represents the cross validation errors (max_fail) occurs during training. When it reaches 60 epochs, the training process gets terminated due to number of validation fail reached 50 i.e. max_fail (as shown in Table 1). Thus, training state plot shows only the variation in training parameters during training process while the performance plot shows the output in terms of performance value. The regression plot between predicted values from ANN and the experimental results are shown in Fig. 9. It depicts the correlation between output and target data. This plot also shows up to what extent the network has learnt from the complex relationships of data. It is found that the experimentally measured values show an excellent match with different outputs of the ANN model. Among the different trials, the R of training results approaches to 1.0 and the corresponding results have the least MSE when the numbers of nodes in hidden layer are 7. The results show that R for training, validation, test and for the combined set are 0.99943, 0.99808, 0.99868 and 0.99912, respectively. Thus, the predicted values are found in

0

Mean Squared Error (MSE)

10

-5

10

Train Validation Test Goal Best

-10

10

-15

10

0

10

20

30

40

Number of epochs Fig. 7. Performance plot.

50

60

1098

D.B. Jani et al./Applied Thermal Engineering 98 (2016) 1091–1103

Gradient = 0.013238, at epoch 60

5

gradient

10

0

10

-5

(a)

10

Mu = 0.1, at epoch 60

5

mu

10

0

10

-5

10

(b)

Validation Checks = 50, at epoch 60 val fail

50

(c)

0

0

10

20

30 60 Epochs

40

50

60

Fig. 8. Training state plots for (a) gradient (b) mutation (c) validation checks.

good agreement with the experimental values. The selected ANN model also demonstrated a good statistical performance with the standard correlation coefficient in the range of 0.998–0.999, and the mean square error (MSE) for the training and predictions are found to be very low compared to the experimental results. 6.2. Comparison between ANN predictions and experimental results The comparison between the results predicted by artificial neural network and that by the experimental findings is shown in Tables 3 and 4 for process air and regeneration air, respectively. The maximum differences in temperature and humidity ratio at the outlet of the dehumidifier for the process air were found as 1.53% and 11.83%, respectively. The same for the regeneration air stream were found to be 2.45% and 5.35%, respectively. The results may be considered to be within the acceptable limits. Fig. 10 shows the comparison between the ANN prediction and the experimental results for exit temperatures and humidity ratios of process air stream with good agreement between the two. In both the cases, R2 values are found to be good. Fig. 11 shows the comparison between the ANN prediction and the experimental results for exit temperatures and humidity ratios of regeneration air stream with reasonably good match between the two. In both the cases, R2 values are found to be fairly good. The maximum difference in percentage between the results estimated by the ANN model and that by experimental findings for effectiveness and moisture removal rate of the dehumidifier are found as 3.22% and 7.27%, respectively (Table 5), which are considered as reasonable. The model is further tested for the main performance parameters of the dehumidification i.e. the effectiveness and the moisture removal rate (MRR). Figs. 12 and 13 depict the comparison between outputs predicted by ANN and that by experiments for dehumidification effectiveness and moisture removal rate, respectively. The performance of trained artificial neural network in the prediction of effectiveness and moisture removal rate of dehumidifier gives R2 = 0.986 and R2 = 0.766, respectively. The results demonstrate that the predictions of the effectiveness by ANN yield better results. Compared to this, the predictions for moisture removal rate from ANN

shows poor agreement because of relatively higher uncertainties involved in the measurement of mass flow rate apart from that in humidity ratio. Also, instead of measuring the moisture removal directly from the experimental study, eq. (1) uses mass flow rate in calculation, which has still poorer uncertainty. Consequently, this uncertainty influences the training process, thus yielding a slightly poor performance (R2 values) for the moisture removal rate. But, the errors are still in the tolerance, and predicted results are acceptable [24]. The developed ANN model is capable of predicting the outputs for the changes in all input parameters as described earlier.

6.3. Effect of ambient on effectiveness and MRR Now, the effect of ambient air temperature and humidity ratio on the performance of dehumidifier has been observed. Figs. 14 and 15 illustrate the influence of ambient air temperature on dehumidifier effectiveness and moisture removal rate, respectively. Both effectiveness as well as moisture removal rate tends to decrease as ambient temperature increases. This is because as the ambient air temperature increases, the inlet temperature of process air also increases, which in turn decreases the partial vapor pressure of process air at inlet. Due to this, the vapor pressure difference between the air and the desiccant along the channel gets reduced. Since the moisture attraction by the desiccant material from process air is based on the difference in vapor pressure between desiccant material surface in channel and moist air flowing through it, the moisture removal rate and ultimately the effectiveness of the dehumidifier get reduced. Since the adsorption process inside dehumidifier is exothermic, hence it is favored by low temperatures of process moist air. Results also show good agreement between outputs predicted by the ANN model and that by experiments for the dehumidifier effectiveness and moisture removal rate of dehumidifier. We got better agreement by using simulated dehumidifier process air outlet humidity ratio of ANN model instead of using directly predicted ANN results for dehumidifier effectiveness and MRR due to due inaccuracies involved in the ANN model because of selection of hidden layer, learning rate, momentum etc.

D.B. Jani et al./Applied Thermal Engineering 98 (2016) 1091–1103

1099

Fig. 9. The regression plot between ANN predictions and the experimental results for (a) training, (b) validation, (c) test, and (d) combined set.

Figs. 16 and 17 depict the effect of ambient humidity ratio on the desiccant dehumidifier effectiveness and moisture removal rate respectively. With the increase in ambient humidity ratio, moisture removal rate and dehumidifier effectiveness increase. This is

due to the fact that, as the water vapor content in the process air inlet to the dehumidifier increases because of the ambient humidity, the difference of vapor partial pressure between process moist air and desiccant material also increases. This increases the mass

Table 3 Comparison of ANN testing results with the experimental data for the dehumidifier exit conditions of process air stream.

Table 4 Comparison of ANN testing results with the experimental data for the dehumidifier exit conditions of regeneration air stream.

Temperature (°C)

Humidity ratio (g/kg)

Temperature (°C)

Humidity ratio (g/kg)

ANN

Experimental

Difference (%)

ANN

Experimental

Difference (%)

ANN

Experimental

(%) Difference

ANN

Experimental

(%) Difference

56.828 57.077 60.825 56.627 57.232 53.978 53.999 58.763 58.426 58.336

56.5 56.2 60.7 56.8 56.8 53.3 53.4 58.3 58.3 58.4

0.57 1.53 0.20 −0.30 0.75 1.25 1.10 0.78 0.21 −0.10

3.8941 2.9457 1.5762 3.3873 4.8392 4.0340 3.9404 2.2684 2.2820 2.3882

3.7 2.8 1.5 3.3 5.0 4.3 4.2 2.0 2.1 2.2

4.98 4.94 4.83 2.57 −3.52 −6.59 −6.04 11.83 5.34 4.11

49.8098 49.8197 54.4847 49.3109 46.1575 47.9042 47.6178 49.6567 49.4317 49.1021

50.3 48.6 54.5 48.1 46.5 47.2 47.4 49.1 49.1 49.2

−0.98 2.44 −0.02 2.45 −0.74 1.47 0.45 1.12 0.67 −0.19

24.0072 24.6192 27.3438 24.4292 25.1282 23.3962 22.474 23.6263 23.5375 23.3846

23.7 23.3 27.6 24.2 25.5 23.3 23.1 24.0 23.4 23.4

1.27 5.35 −1.22 0.93 −1.47 0.11 −2.82 −1.58 0.37 −0.45

D.B. Jani et al./Applied Thermal Engineering 98 (2016) 1091–1103

62 61 60 59 58 57 56 55 54 53 52

Exit temperature of process air (oC) R² = 0.9832 Experimental

Experimental

1100

50

(a)

55 60 ANN predictions

Exit humidity ratio of process air (g/kg)

6 5 4 3 2 1 0

R² = 0.9865

0

65

2

4

6

ANN predictions

(b)

Fig. 10. Comparison of the ANN predictions with experimental results for (a) exit temperature, and (b) exit humidity ratio, of process air stream.

Table 5 Comparison of ANN testing results with the experimental data for the dehumidifier effectiveness and moisture removal rate. Moisture removal rate (kg/hr)

ANN

Experimental

(%) Difference

ANN

Experimental

(%) Difference

68.8648 72.4128 87.3896 68.7506 56.1922 61.7057 63.1342 82.3511 81.3699 79.6699

69.8 74.7 86.9 69.6 57.3 61.3 61.1 82.2 80.5 79.2

−1.35 −3.15 0.56 −1.23 −1.97 0.65 3.22 0.18 1.06 0.47

2.8252 2.9551 2.5911 2.9686 2.6050 2.7384 2.7279 2.6831 2.6825 2.6392

3.0 3.1 2.7 2.9 2.7 2.9 2.7 2.6 2.6 2.5

−6.18 −7.27 −4.20 −0.72 −3.64 −5.90 −1.99 −0.25 3.07 3.37

Experimental

Dehumidifier effectiveness (%)

Dehumidifier effectiveness 90 85 80 75 70 65 60 55 50

R² = 0.9866

50

60

70 ANN predictions

80

90

Fig. 12. Comparison of the ANN predictions for dehumidifier effectiveness with experimental results.

6.4. Effect of flow rate on effectiveness and MRR

55 54 53 52 51 50 49 48 47 46 45

Exit temperature of regeneration air (oC) R² = 0.9425

45

(a)

The effect of variation in process air flow rate on the desiccant dehumidifier effectiveness and moisture removal rate is graphically shown in Figs. 18 and 19, respectively. As observed from the behavior, both the dehumidifier effectiveness as well as moisture removal rate, decreases with successive increment in the process air flow rate. This is because as we increase the process air flow rate, the contact time between inlet air and desiccant surface decreases leads to poor adsorption of moisture. Results also show a good agreement between the outputs predicted by ANN and that by experiments for the dehumidifier effectiveness and moisture removal rate. We got better agreement in both cases using the

Experimental

Experimental

transfer potential between moist air and desiccant material. Consequently, diffusion of water vapor droplets from the former to the latter rises and the moisture removal rate of the dehumidifier, successively increases. Results also show a good agreement between the outputs predicted by ANN and that by experiments for the dehumidifier effectiveness and moisture removal rate. Instead, better agreement is achieved by using the simulated dehumidifier process air outlet humidity ratio of ANN model for both the effectiveness and moisture removal rate of dehumidifier than using the directly predicted ANN results. This is due to the model inaccuracy, which is caused by the selection of hidden layer, learning rate, momentum etc.

50 ANN predictions

55

Exit humidity ratio of regeneration air (g/kg)

29 28 27 26 25 24 23 22

R² = 0.903

22

(b)

24

26

28

ANN predictions

Fig. 11. Comparison of the ANN predictions with experimental results for (a) exit temperature, and (b) exit humidity ratio, of regeneration air stream.

D.B. Jani et al./Applied Thermal Engineering 98 (2016) 1091–1103

90 88

R² = 0.7669

2.6

2.7 2.8 ANN predictions

2.9

3

Dehumidifier effectiveness (%)

Experimental

Moisture removal rate (kg/hr) 3.3 3.2 3.1 3 2.9 2.8 2.7 2.6 2.5 2.5

1101

Fig. 13. Comparison of the ANN predictions for moisture removal rate with experimental results.

Experimental ANN Wout simulated

86 84 82 80 78 76 74 72 14

simulated dehumidifier process air outlet humidity ratio of ANN model instead of using directly predicted ANN results for the dehumidifier effectiveness and MRR due to comparatively higher inaccuracies involved in the ANN model. This may be because of selection of hidden layer, learning rate, momentum etc.

16 17 18 Ambient humidity ratio (g/kg)

19

20

Fig. 16. Influence of ambient humidity ratio on dehumidifier effectiveness.

76

2.85

74

Experimental ANN Wout simulated

72

Experimental ANN Wout simulated

2.80

70 MRR (kg/hr)

Dehumidifier effectiveness (%)

15

68 66 64

2.75

2.70

62 2.65

60 58 56 27.0

2.60

27.5

28.0 28.5 29.0 Ambient temperature (oC)

29.5

14

30.0

16 17 18 Ambient humidity ratio (g/kg)

19

20

Fig. 17. Influence of ambient humidity ratio on moisture removal rate.

Fig. 14. Influence of ambient air temperature on dehumidifier effectiveness.

88

2.68 2.66

Experimental ANN Wout simulated

2.62

Experimental ANN Wout simulated

86

Dehumidifier effectiveness (%)

2.64

MRR (kg/hr)

15

2.60 2.58 2.56 2.54 2.52 2.50

84

82

80

78

2.48 2.46 27.0

27.5

28.0 28.5 29.0 Ambient temperature (oC)

29.5

Fig. 15. Influence of ambient air temperature on moisture removal rate.

30.0

76 240

260

280 300 320 Process air flow rate (m3/hr)

340

Fig. 18. Influence of process air flow rate on dehumidifier effectiveness.

360

1102

D.B. Jani et al./Applied Thermal Engineering 98 (2016) 1091–1103

2.90

2.80

Experimental ANN Wout simulated

2.85

2.78

Experimental ANN Wout simulated

MRR (kg/hr)

MRR (kg/hr)

2.76

2.80

2.75

2.74

2.72

2.70

2.65 240

2.70

260

280 300 320 Process air flow rate (m3/hr)

340

2.68 110

360

Fig. 19. Influence of process air flow rate on moisture removal rate.

The effect of variation in regeneration temperature on the desiccant dehumidifier effectiveness and moisture removal rate has been illustrated in Figs. 20 and 21, respectively. It is shown that both the effectiveness as well as moisture removal rate of dehumidifier gets increases with increase in regeneration temperature. This is due to the fact that with higher regeneration temperature air could remove more moisture during desorption process, as the pressure difference between hot regeneration air and desiccant surface tends to increase. Similarly, during the dehumidification the desiccant could adsorb more vapor from process air. So, when regeneration temperature increases, more moisture will get adsorbed, which increases the moisture removal rate. Increase in moisture removal rate causes an increase in dehumidifier effectiveness. This is also clear from equations 1 and 2 as described earlier. Results also show a good agreement between the outputs predicted by ANN and that by experiments for dehumidifier effectiveness and moisture removal rate. We got better agreement by using the simulated dehumidifier process air outlet humidity ratio of ANN model instead of using directly predicted ANN results for dehumidifier effectiveness and MRR

82

Dehumidifier effectiveness (%)

Experimental ANN Wout simulated

78

An artificial neural network (ANN) model has been developed to predict the performance of a rotary desiccant dehumidifier for the variations in process and regeneration air stream inlet conditions and regeneration temperature. Moisture removal rate and effectiveness of the dehumidifier are considered as performance indicators. Experimental runs have also been performed and the results are compared with the ANN predictions. Based on experimental and ANN results, the following conclusions were drawn:

•

74

•

72

• 120 125 130 Regeneration temperature (oC)

135

140

7. Conclusions

•

115

135

due to inaccuracies involved in the ANN model because of selection of hidden layer, learning rate, momentum etc. The accuracy of artificial neural network (ANN) model greatly relies on the network structure, amount of training and testing data, the training and testing characteristics and on the selection of learning as well as performance function. Moreover, the variations in the dimensionality of the data set and the network architecture, specifically, the number of hidden units and layers have a significant effect on accuracy of ANN model. The accuracy of the model for better prediction can be further improved by expanding the experimental database for training i.e. the size of the training data set, discriminating variables and by using the proper nature of the training and testing sets as well as network parameters. It can also be done by using swarm intelligent techniques to update the weights, noise reduction in the target data, stable data by use of cross validation techniques etc. The artificial neural network architecture is found to have slightly higher accuracies by using the larger and more complex networks.

76

70 110

120 125 130 Regeneration temperature (oC)

Fig. 21. Influence of regeneration temperature on moisture removal rate.

6.5. Effect of regeneration temperature on effectiveness and MRR

80

115

140

Fig. 20. Influence of regeneration temperature on dehumidifier effectiveness.

The maximum differences in temperature and humidity ratio at the outlet of the dehumidifier for the process air were found as 1.53% and 2.45% respectively. The same for the regeneration air stream were found to be11.83% and 5.35%, respectively. The maximum percentage difference between the ANN predictions and the experimental values for MRR and the effectiveness of the dehumidifier was found to be 7.27% and 3.22%, respectively. The results indicate that the accuracy of the ANN model was satisfactory and coincide with the experimental data. The correlation coefficient (R) and mean square error (MSE) are used to assess the performance of ANN model. This ANN model demonstrates a good statistical performance with the correlation coefficient in the range of 0.998–0.999 are found to be very close to unity, and the MSE values for the ANN training and

D.B. Jani et al./Applied Thermal Engineering 98 (2016) 1091–1103

•

•

predictions were very low relative to the range of the experiments while assessing the performance of rotary desiccant dehumidifier. The ANN model can be efficiently used to predict the performance of dehumidifier in terms of moisture removal rate and effectiveness. It also provides a theoretical basis for a detailed heat and mass transfer analysis of desiccant wheel at different operating conditions. The accuracy of prediction depends on the type of model containing the characteristics of a particular combination of layers and nodes as well as on the database for training. The accuracy can further be improved by expanding the experimental database for network training.

This study reveals that the rotary desiccant dehumidifier alternatively be modeled using ANNs within a high degree of accuracy. Because this new approach requires only a limited number of tests instead of an exhaustive experimental study or dealing with complicated mathematical models, manufacturers may employ the ANN technique for evaluating the performance of rotary desiccant dehumidifier can save both engineering effort and funds. Nomenclature a Actual output (experimental output) ANN Artificial neural network b Bias Air stream mass flow rate [kg/hr] ma MRR Moisture removal rate [kg/hr] MSE Mean square error p Predicted output (network output) R The correlation coefficient RMSE Root mean square error SHR Sensible heat ratio T Temperature [°C] TANSIG Tan-sigmoid transfer function TRAINLMLevenberg–Marquardt back propagation VCR Vapor compression refrigeration The weight between hidden layer vectors and output layer υk,j vectors W Humidity ratio [g/kg] w Synaptic weights The weight between input layer vectors and hidden layer wj,i vectors V Air stream flow rate [m3/hr] X Input signal Y Output Greek letters Δm Moisture removal rate [kg/hr] ε Effectiveness of dehumidifier Subscripts amb Ambient air DW Desiccant wheel in Inlet i,j,k The number of nodes pro Process air out Outlet reg Regeneration air 1,2, etc. Reference state points

1103

References [1] N.A. Pennington, Humidity changer for air conditioning, USA Patent No. 2, 700, 537, 1955. [2] R.V. Dunkle, A method of solar air conditioning, Mech. Chem. Eng. Trans. Inst. Eng. 73 (1965) 73–78. [3] C.G. Munters, Inorganic, fibrous, gas-conditioning packing for heat and moisture transfer, U. S. Patent 3, 377, 225, 1968. [4] P.J. Banks, Coupled equilibrium heat and single adsorbate transfer in fluid flow through porous media – I, characteristic potentials and specific capacity ratios, Chem. Eng. Sci. 27 (1972) 1143–1155. [5] R. Barlow, Analysis of the adsorption process and desiccant cooling systems: a pseudo-steady-state model for coupled heat and mass transfer, Technical Report No. SERI/TR-631–1330 (1982), Solar Energy Research Institute, Golden, CO. [6] W.M. Worek, Z. Lavan, Performance of a cross-cooled desiccant dehumidifier prototype, J. Sol. Energy Eng. 104 (1982) 187–196. [7] I.L. Maclaine-Cross, Proposal for a desiccant air conditioning system, ASHRAE Trans. 94 (1988) 1997–2009. [8] B.S. Davanagere, S.A. Sherif, D.Y. Goswami, A feasibility study of solar desiccant air conditioning system – part I: psychrometrics and analysis of the conditioned zone, Int. J. Energy Res. 23 (1999) 7–21. [9] J.J. Jurinak, Open cycle solid desiccant cooling-component models and system simulation (Ph.D. thesis), University of Wisconsin, Madison, 1982. [10] P.R. Burns, R.B. Mitchell, W.A. Bechman, Hybrid desiccant cooling systems in supermarket applications, ASHRAE Trans. 91 (1985) 457–468. [11] E. Van den Bulck, J.W. Mitchell, S.A. Klein, The use of dehumidifiers in desiccant cooling and dehumidification systems, Trans. ASME 108 (1986) 684–691. [12] C. Dhanes, M.W. Wiliam, Parametric study of an open cycle adiabatic solid desiccant cooling system, Energy 9 (1988) 739–747. [13] A.A. Pesaran, T.R. Penney, A.W. Czanderna, Desiccant cooling-state of the art-assessment, DOE-EERE Research Reports, National Renewable Energy Laboratory, Colorado, 1992. [14] S. Jain, P.L. Dhar, S.C. Kaushik, Evaluation of solid based evaporative cooling cycles for typical hot and humid climates, Int. J. Refrigeration 18 (1995) 287–299. [15] R. Rojas, Neural Networks: A Systematic Introduction, Springer, Germany, 1996. [16] S.A. Kalogirou, Artificial neural networks and genetic algorithms in energy applications in buildings, Adv. Build. Energy Res. 3 (2009) 83–120. [17] G. Zhang, B.E. Patuwo, Y.H. Michael, Forecasting with artificial neural networks: the state of the art, Int. J. Forecast. 14 (1998) 35–62. [18] J.M. Cejudo, R. Moreno, A. Carrill, Physical and neural network models of a silica-gel desiccant wheel, Energy Build. 34 (2002) 837–844. [19] S.A. Kalogirou, Applications of artificial neural networks in energy systems, Energy Convers. Manage. 40 (1999) 1073–1087. [20] E.B. Zeidan, A.A. Ayman, A.M. Hamed, Investigation on the effect of operating parameters on the performance of solar desiccant cooling system using artificial neural networks, Int. J. Therm. Environ. Eng. 1 (2010) 91–98. [21] A.T. Mohammad, S.B. Mat, M.Y. Sulaiman, K. Sopian, A.A. Abduljalil, Artificial neural network analysis of liquid desiccant dehumidifier performance in a solar hybrid air-conditioning system, Appl. Therm. Eng. 59 (2013) 389–397. [22] H. Parmar, D.A. Hindoliya, Artificial neural network based modelling of desiccant wheel, Energy Build. 43 (2011) 3505–3513. [23] P. Gandhidasan, M.A. Mohandes, Artificial neural network analysis of liquid desiccant dehumidification system, Energy 36 (2011) 1180–1186. [24] I.P. Koronaki, E. Rogdakis, T. Kakatsiou, Thermodynamic analysis of an open cycle solid desiccant cooling system using artificial neural network, Energy Convers. Manag. 60 (2012) 152–160. [25] A.T. Mohammad, S.B. Mat, M.Y. Sulaiman, K. Sopian, A.A. Abduljalil, Implementation and validation of an artificial neural network for predicting the performance of a liquid desiccant dehumidifier, Energy Convers. Manag. 67 (2013) 240–250. [26] A.S. Alosaimy, Application of evaporative air coolers coupled with solar water heater for dehumidification of indoor air, Int. J. Mech. Mechatron. Eng. 13 (2013) 60–68. [27] I. Uckan, T. Yılmaz, E. Hurdogan, O. Buyukalaca, Development artificial neural network model for the prediction of the performance of a silica gel desiccant wheel, Int. J. Green Energy 16 (2014) 1–42. [28] Matlab, Version 8.4, Neural Network Tool Box, Inc., 2014. [29] K. Kulkarni, P.K. Sahoo, M. Mishra, Optimization of cooling load for a lecture theater in a composite climate in India, Energy Build. 43 (2011) 1573–1579. [30] D.B. Jani, M. Mishra, P.K. Sahoo, Performance studies of hybrid solid desiccantvapor compression air-conditioning system for hot and humid climates, Energy Build. 102 (2015) 284–292. [31] ASHRAE Handbook – Fundamentals, ASHRAE, Atlanta, 2013. [32] K. Daou, R.Z. Wang, Z.Z. Xia, Desiccant cooling air conditioning: a review, Renew. Sustain. Energy Rev. 10 (2006) 55–77. [33] M. Hosoz, H.M. Ertunc, H. Bulgurcu, Performance prediction of a cooling tower using artificial neural network, Energy Convers. Manag. 48 (2007) 1349– 1359.