Professional Documents
Culture Documents
Application of Artificial Neural Network For Predicting Performance of Solid Desiccant Cooling Systems A Review - Compress
Application of Artificial Neural Network For Predicting Performance of Solid Desiccant Cooling Systems A Review - Compress
Application of Artificial Neural Network For Predicting Performance of Solid Desiccant Cooling Systems A Review - Compress
a
Department Gujarat Technological University, GTU, Ahmedabad, India
b
Department of Mechanical & Industrial Engineering, Indian Institute of Technology Roorkee, 247667, India
A R T I C L E I N F O A BS T RAC T
Keywords: In present study, an attempt has been made to review the applications of artificial neural network (ANN) for
ANN predicting the performance of solid desiccant cooling systems. Different types of neural networks are applied to
Desiccant cooling model the solid desiccant cooling systems. With use of experimental data, an ANN model was developed which
Dehumidifier is based on different algorithms. Available experimental data were divided into two categories for training and
COP
testing of the ANN model. Later on, trained ANN model was tested for predicting the performance of system
Regeneration
based on various input and output parameters such as air stream flow rates, temperatures and humidity ratios,
pressure drop, dehumidifier effectiveness, cooling capacity, regeneration temperature, power input, coefficient
of performance etc. So, present review proposes the use of ANN based model to simulate the relationship
between inlet and outlet parameters of the system. The ANN predictions for these parameters usually agreed
with the experimental values with higher correlation co-efficient. The previous studies show that ANNs can be
used with a higher precision in guessing the performance of solid desiccant cooling systems. This review is
useful for making opportunities to further research of ANNs and its feasibility which is becoming common in the
coming days.
1. Introduction experimental study. So, manufacturers may employ the ANN technique
for evaluating the performance of solid desiccant cooling that can save
The performance of solid desiccant cooling systems is evaluated in both engineering effort and funds. Earlier many review studies were
terms of energy analysis (first law) and exergy analysis (second law) of reported with applications of ANN in the field of drying processes [1],
thermodynamics using conventional approaches i.e. analytical, numer- forecasting [2], atmospheric sciences [3], solar photovoltaic systems
ical and experimental methods. The conventional analytical and [4] and modeling of energy systems [5–7]. Ding [8] summarized
numerical approaches involve complex equations and assumptions, various simulation techniques for modeling and performance estima-
whereas experimental studies are rigorous and costlier. During last two tion of conventional VCR systems.
decades, the use of artificial intelligence systems in the field of space Following the cited literature, it is seen that there is no specific
cooling is increasing gradually to solve the complicated problems. review has been reported on applications of ANN for solid desiccant
Artificial intelligence systems include areas such as artificial neural cooling systems. The present review set out more broadly about up to
network (ANN), genetic algorithms, fuzzy logic and various hybrid date study covering the applications of ANN in the performance
systems, which combine two or more techniques. The main advantage forecasting of solid desiccant cooling systems by prediction of para-
of ANN as compared to the other artificial intelligent systems is its meters in terms of cooling capacity, dehumidifier effectiveness, regen-
accuracy, speed, simplicity and ability to model a multivariable eration temperature, power input and COP of the system. Moreover,
problem for solving complex relationships between the variables and the limitations with ANN models are also highlighted. The present
establishing non linear relationships by means of training network. review can be impacted in term of application of ANN for performance
ANN model can forecast the desired output of the system using limited estimation of desiccant cooling system requiring less formal statistical
training data. Moreover, ANN overcomes the limitations of traditional training, ability to implicitly detect complex nonlinear relationships
approaches by extracting the required information using training data, between dependent and independent variables, ability to detect all
which has not required any specific analytical equations or exhaustive possible interactions between predictor variables and the availability of
⁎
Corresponding author.
E-mail addresses: dbjani@rediffmail.com (D.B. Jani), mmishfme@iitr.ac.in (M. Mishra), sahoofme@iitr.ac.in (P.K. Sahoo).
http://dx.doi.org/10.1016/j.rser.2017.05.169
Received 29 November 2016; Received in revised form 10 March 2017; Accepted 20 May 2017
1364-0321/ © 2017 Elsevier Ltd. All rights reserved.
D.B. Jani et al. Renewable and Sustainable Energy Reviews 80 (2017) 352–366
multiple training algorithms. The objective of the present review is to coming days.
aware about an introduction of an application of artificial neural
network in the field of space cooling energy modeling as a powerful 2. Overview of solid desiccant cooling
data-driven, self-adaptive, flexible computational tool having the
capability of capturing nonlinear and complex underlying character- 2.1. Brief history of solid desiccant cooling
istics of any physical process with a high degree of accuracy. Moreover,
it is found that artificial neural network is very suitable for inverse Different configurations of solid desiccant cooling system have been
energy modeling when the numerical relations between input and proposed earlier by many researchers so far to attain better system
output variables are unknown, and can't be established. Thus, this performance in terms of maintaining a thermal comfort. The earliest
review is useful for making opportunities to further research of use of form of solid desiccant cooling cycle was proposed by coupling
artificial neural network in performance prediction of solid desiccant desiccant dehumidifier with heat source and evaporative cooler [9].
cooling systems and its feasibility which is becoming common in the Similar cycle was proposed later by Dunkle [10] using desiccant wheel
353
D.B. Jani et al. Renewable and Sustainable Energy Reviews 80 (2017) 352–366
made up of molecular sieve with additional heat exchanger to achieve high temperature of hot surrounding.
the better performance than the earlier one. Later on, Munter [11] Dunkle cycle combines the advantages of both the ventilation cycle,
further enhanced the performance of the solid desiccant cooling cycle having supply of large amount of ventilation fresh air with relatively
by introducing parallel passages in rotary dehumidifier and provided low temperature and the recirculation cycle, having enhanced cooling
backup of traditional vapor compression system to handle the sensible capacity. As shown in Fig. 4, an additional sensible heat wheel is
and latent load separately. Since then, number of efforts have been incorporated in system which can provide colder process air with
made for enhancing the performance of rotary desiccant dehumidifiers relatively low temperature for the sensible heat exchanger [23].
used in the desiccant cooling systems. Important among those were the Fig. 5 shows comparison between different solid desiccant cooling
analogy theory suggested by Banks [12], the pseudo-steady state model cycles on psychrometric chart. It was depicted from the psychrometric
by Barlow [13], combined potential technique by Jurinak [14], finite chart that the reactivation temperature required for desorption of
difference method for cross-cooled dehumidifiers [15] and finite rotary dehumidifier in case of recirculation cycle is highest due to
difference method by Maclaine-Cross [16] which are now widely used excess humidity ratio at the dehumidifier inlet in the regeneration air
by other investigators in getting better performance of desiccant side. The temperature of supply room air is observed lowest in case of
cooling cycles [17]. Burns et al. [18] evaluated the performance of recirculation cycle because it reuses 100% building air as a process air
hybrid desiccant cooling cycle used for supermarket and shows better at the dehumidifier inlet.
performance than the traditional VCR system in terms of energy In case of Pennington cycle, maximum amount of fresh ventilated
consumption in humid climate. Van den Bulck et al. [19] have carried air supply is possible because it uses 100% outdoor air as a process air
out experiments to investigate the effects of process and regeneration at the dehumidifier inlet. But, in case of ventilation cycle thermal
air mass flow rates, temperature, humidity ratio etc. on the perfor- coefficient of performance and specific cooling capacity would be
mance of solid desiccant cooling system. Higher regeneration tempera- reduced in comparison with other standard cycles because humidity
ture was recommended by them to reduce the required mass flow rates ratio and temperature of outdoor air are usually higher than that of the
of air streams. Pesaran [20] proposed the use of renewable energy room recalculated air especially in hot and humid surroundings.
(solar heat) for the regeneration of dehumidifier used in solid desiccant
cooling system to increase the overall system efficiency. Comparison
3. Overview of artificial neural network
among various solid desiccant cooling cycles for air conditioning in hot
and humid climates was also carried out by Jain et al. [21].
An artificial neural network consists of large number of processing
elements called neurons. They are joined by connecting links called
2.2. Principle of solid desiccant cooling
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
A solid desiccant is a hygroscopic material made up of silica gel, Li-
the network architecture and the method of input data required. A
Cl and Molecular sieves. Solid desiccant cooling system works on
generalized neural network model (Fig. 6) consists of three primary
principle of adsorption of water vapor from air. In solid desiccant
elements; synapses or connecting link, summing node with a squashing
cooling system, the excess water vapor present in the process air is
function and an externally applied bias to increase or decrease the net
removed by passing through rotary dehumidifier. The temperature of
input of the activation function. The network performance is deter-
this warm and dried process air is then lowered further to the desired
mined by value of weights and biases in every single neuron. The
room comfort conditions by use of sensible air-to-air heat recovery
network needs to be trained with the help of measured data sets to give
wheel and cooling coils. Dehumidifier needs to be regenerated by
the desired output using input data sets [24].
supplying hot regenerated air via heater to remove adsorbed water
Multi layer feed forward networks (MLFFN) consists of an input
vapor (Fig. 1). So, desiccant cooling system comprises mainly four
layer, followed by one or more hidden layers and an output layer.
components, namely desiccant wheel, sensible heat recovery wheel,
Multiple layers of neurons with nonlinear transfer functions make the
cooling unit and the reactivation heater.
network to learn linear and nonlinear relationships between input and
output vectors. Back propagation (BP) learning algorithm is used
2.3. Types of solid desiccant cooling cycles
widely to train the MLFFN. The network is trained with selected
number of neurons in the hidden layer, momentum factor, learning
Solid desiccant cooling system which was introduced first by
rate and transfer function. Radial biased function network (RBFN)
Pennington is shown in Fig. 2. It was also known as ventilation cycle
having three layers as input layer, hidden layer followed by an output
[22]. The moist ambient air (state 1) dehumidifies (state 2) by passing
layer. The interconnections between input layer and hidden layer form
through a rotary dehumidifier owing to adsorption and converted into
hypothetical connections and between the hidden layer and output
warm and dry air. It is then cooled sensibly by passing through sensible
layer form weighted connections. The Gaussian transfer function is
heat recovery wheel (2-3). Then after, it is cooled further up to comfort
used to operate the weighted inputs to produce neuron output. The
supply conditions (state 4) by passing through direct evaporative
activation function used in hidden layer is radial base function. The
cooler. Return room air (state 5) is cooled and humidified by passing
linear transfer function is used in output layer. RBFN has faster
through another direct evaporative cooler (state 6). This air is then
sensibly heated with process air by passing through sensible heat
recovery wheel (state 7). This warm air is then heated up further to the
required regeneration temperature of desiccant material used in the
desiccant wheel by passing through regeneration heat source. After
reactivating the desiccant wheel, this air is then exhausted to surround-
ing (state 9). A small portion of return air steam (state 7) bypasses the
regeneration heat source in order to reduce the regeneration heat
consumption.
To enhance cooling capacity, recirculation cycle shown in Fig. 3
which is a modified form of Pennington (ventilation) cycle, is config-
ured to reuse the room return air as a dehumidifier process air inlet in
severe hot and humid climatic conditions [22]. This is because due to
the fact that this lowers the energy consumption to lower the extreme Fig. 2. Pennington (ventilation) cycle.
354
D.B. Jani et al. Renewable and Sustainable Energy Reviews 80 (2017) 352–366
square nodes. The second layer is called product layer and every node
in this layer is a fixed node marked by a circle node and labelled by M.
The third layer is a normalized layer and every node in this layer is a
fixed node, marked by a circle node and labelled by N. The fourth layer
is the de-fuzzy layer having adaptive nodes and marked by square
nodes. The fifth layer computes the overall system output as the
summation of all incoming signals [102–104].
Outputs from the ANN model are compared with the actual output.
Fig. 4. Dunkle cycle.
There may be small variation between the network’s output and the
target output. The weights are adjusted such a way that the error
function minimizes the variation between actual outputs and model
outputs. This process is continued until the error function comes under
the desired tolerance limit [25]. 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 which minimize
the error between the measured output and the ANN network output
are obtained [26] as follows
⎡ N ⎤
Y = F (S) = F ⎢ Σ XK wK + b⎥
⎣K=1 ⎦ (1)
Neural network tool box of MATLAB© was used for the formation
of network architecture. ANN structure can be made by inserting the
weights into each layer of layered structure by selecting the weight
matrices. The values of these weights are determined using the feed
forward backpropagation (BP) method. Error calculation is iterated
using different inputs and outputs during learning until the root mean
square error (MSE) of the network reaches to an acceptable level. The
performance estimation through the neural network is done by mean
square error (MSE) [27] between the estimated and the actual values as
per the following relation
N
Σ X predicted − X experimental 2
i=1
MSE =
N (2)
Lesser the MSE is, better fit the results will be. The root mean
square error [28] can be given by
Fig. 5. Comparison between different solid desiccant cooling cycles.
1 N
convergence, smaller extrapolation errors and has higher reliability RMSE = Σ (ai−pi )2
N i =1 (3)
compared to MLFFN. Generalized regression neural network (GRNN)
made up of four-layer feed forward neural network which is based on where, ai and pi are number of nodes for actual output and estimated
the non-linear regression theory consisting of input layer, pattern layer, outputs, respectively. The absolute fraction of variance (R2) [29], a
summation layer and output layer. GRNN is a memory-based feed statistical indicator that can be applied to multiple regression analysis,
forward networks as it consists of estimation of probability density is obtained as
functions. Adaptive neuro fuzzy interface systems (ANFIS) consisting ⎛ N ⎞
of nodes and directional links, which combines the learning capabilities ⎜ Σ (ai − pi )2 ⎟
i =1
of a neural network and reasoning capabilities of fuzzy logic. ANFIS R2 = 1−⎜ N ⎟
⎜⎜ Σ ( p )2 ⎟⎟
architecture made up of fiver layers. The first layer in the network is ⎝ i =1 i ⎠ (4)
called fuzzy layer. The adjustable nodes in this layer are represented by
2 2
R ranges between 0 and 1. A very good fit yields R value of 1 or
355
D.B. Jani et al. Renewable and Sustainable Energy Reviews 80 (2017) 352–366
closer, whereas a poor fit results to a value near 0. streams at the dehumidifier inlet and outlet respectively.
Fig. 7 illustrates the flow chart that describes the development, Dehumidifier effectiveness (εdw) can be expressed by [32]
training and simulation of an artificial neural network. The experi-
w1 − w2
mental measurements are input parameters to the model. Neural εdw =
w1 − w2,ideal (6)
network understands the underlying correlations in the entered input
data and stores them as inter-neuron connection strengths or corrected where w2,ideal is the ideal humidity ratio of the process air stream at exit
weights. Based on number of neurons, number of iterations and the of the rotary desiccant dehumidifier, it is to be taken zero by assuming
desired accuracy, the training set and the target set are developed. The that the air is completely dehumidified at this point.
network needs to be trained using training data set consisting of a Effectiveness of sensible heat wheel (εhw) can be determined by [33]
group of input data and corresponding output data. In an artificial
neural network model, different input parameters are used such as T2 − T3
εhw =
temperature, humidity ratio or relative humidity, dehumidifier pres- T2 − T6 (7)
sure drop, flow rates of process as well as regeneration air streams at
different state points in the cycle, while moisture removal rate (MRR), where the subscripts correspond to the state point in the solid desiccant
dehumidifier effectiveness, cooling capacity, power consumption, coef- cooling system.
ficient of performance (COP) of the system are used as output Effectiveness of evaporative coolers (εec) can be evaluated by [34]
parameters. Training involves the revision of synaptic weights. The T3 − T4 T − T6
network reads and processes each set of input data and produces an εec1 = and εec2 = 5
T3 − Twb,3 T5 − Twb,5 (8)
output, which is compared with the actual (measured) 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 mainly determined by
the weights and bias values in every single neuron.
The three-layer back propagation (BP) ANN network for rotary
dehumidifier is shown in Fig. 8 as a sample study. 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 valid data i.e.
input and target output. The input layer in the present sample case has
seven nodes namely inlet temperature, humidity ratio and flow rate of
process air and ambient air respectively, and regeneration air tem-
perature. 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 sample
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 minimiseminimize 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.
The trained artificial neural network model of the sample case is
shown in Fig. 9. This shows 7-7-6-6 network structure with 7 hidden
layers yielded the good model for the accurate estimation of outputs,
with minimum MSE during training because it has higher stability and
faster convergence rate. This trained network can be used for simulat-
ing the system outputs for the inputs that have been introduced before.
4. Performance indices
356
D.B. Jani et al. Renewable and Sustainable Energy Reviews 80 (2017) 352–366
Q cc mp (h5 − h 4)
COP = =
Qr mr (h8 − h7) (10)
where mp and mr denotes the mass flow rate of process air as well as
regeneration air streams, respectively.
Similarly, the coefficient of performance (COP) for the solid
desiccant – vapor compression hybrid cooling system can be given as
follows [37]
Q cc
COPT =
Et (11)
where Qcc is the cooling capacity which is defined as following [38]
Q cc = mp (h1 − h5) (12)
where Et is the total energy consumption of the system expressed by
[39]
E t = Q r + Wc + Woth (13)
Woth shows energy consumption of other equipments are fans, desiccant
wheel and heat wheel motor. While Qr is the heat supply to regenera-
tion air stream by a reactivation heat source is given as [40]
Q r = mr (h8 − h7) (14)
COPw and COPH can also be defined by [41]
Q cc Q
COPw = and COPH = cc
W Qr (15)
Fig. 8. Architecture of proposed ANN network (BP) for rotary dehumidifier [30].
From last equation we obtain:
Table 1 1 1 1
Training parameters used for the ANN model [30]. = +
COPT COPW COPH (16)
Sr. no. Training parameter Type/value
Similarly, the coefficient of performance (COP) for the solar assisted
1 Training function TRAINLM solid desiccant – vapor compression hybrid air-conditioning can be
2 Performance function MSE given as follows [42–57]
3 Transfer function TANSIG
4 epochs 1000 Q DEC
COPDEC =
5 min_gradmin _grad 0.000001 Q HC (17)
6 mu 0.001
7 mu_dec 0.1 where solar cooling coefficient of performance (COPDEC) is the ratio
8 mu_inc 10 between useful cooling output of the desiccant cooling cycle (QDEC) and
9 max_timemax _time inf the regeneration heat delivered by the solar heating coil (QHC). So,
10 goal 0.000000000000001
11 max_failmax _fail 50 Q DEC = mp (h3 −h2) and Q HC = mr (h8 −h7) (18)
while coefficient of performance (COP) of vapor compression
where Twb is the wet bulb temperature of moist air. systems (COPVCR) is expressed as follows [58–70]
The rates of moisture added to air by the evaporative coolers in the QL
process and return lines are obtained by following [35] COPVCR =
Wi (19)
mw1 = mp (w4 − w3) and mw2 = mr (w6 − w5) (9) where QL is cooling effect in kW and Wi is electric power consumption
for VCR compressor in kW.
where m w1 and m w2 are the rates of moisture added to air in the
evaporative coolers in supply and return lines respectively. 5. Important literature survey on applications of ANN
The coefficient of performance (COP) for the solid desiccant
evaporative cooling system can be expressed by [36] The earliest effort was made to apply artificial neural network to
357
D.B. Jani et al. Renewable and Sustainable Energy Reviews 80 (2017) 352–366
optimize energy consumption in HVAC system by Curtiss et al. [71]. with a maximum possible error of ± 5% and a correlation coefficient up
The network was able to estimate energy use better than the traditional to 0.989 and 0.999 during dehumidification and regeneration operat-
regression techniques, and the energy management system was able to ing modes, respectively. The reported results revealed that the desic-
maintain thermal comfort and saves more energy as compared to a cant bed can be modelled using ANNs with a high degree of accuracy.
fixed set-point or basic temperature reset algorithm. Later on, Kah Uçkan et al. [78] proposed mathematical equations derived from
et al. [72] proposed a modular network using a multilayer perceptron Artificial Neural Networks (ANNs) for the prediction of dry bulb
for controlling the temperature of the air-conditioning system used in a temperature and the specific humidity at the dehumidifier exit. The
public transport. The input parameters used for training the network neural network model consists of five inputs and two output neurons
were the temperature, number of passengers and time of day. A model that define the outlet conditions (DBT and humidity ratio) of the rotary
was also used for decision making to set the system to an optimum dehumidifier. The results predicted by the ANNs model are compared
state of operation that maintains thermal comfort inside conditioned with the actual measured data by using input variables. It is found that
space. The authors claimed that the system offered fast learning with the mean absolute percentage errors (MAPE) for DBT and humidity
good accuracy. Datta et al. [73] presented a similar application where a ratio are found to be 0.80% and 1.56%, respectively and the correlation
number of networks were compared, with the objective of identifying coefficient (R) values are obtained to be approximately 0.986 for both
the important inputs to the network which will facilitate on-line output variables while the root mean square errors (RMSE) are found
prediction and thereby implement refrigeration and HVAC system as 0.54% and 0.18% for DBT and humidity ratio, respectively. Zhu and
diagnostics, process control, optimization and energy management in Chen [79] developed ANN model of marine rotary desiccant air-
retail stores. The network showed a better performance as compared to conditioning by using multivariate linear regression analysis method.
conventional multiple regression techniques. Kajl et al. [74] proposed The results show that the square of multiple coefficient correlation (R2)
an artificial neural network model which can fill the gap between of the prediction models reached 0.999 after the seventh corrected and
simplified and detailed prediction methods for building energy con- the relative error much less than 1%. Neural network is used to analyze
sumption. The neural network training and testing data set were based the effect of significant operating parameters on moisture adsorption
on simulation results of numerous office buildings carried out with the rate in desiccant wheel by Rafique [80] because in analytical model it is
DOE-2 software program. It is found that the proposed model retains difficult to predict relationship between inlet and outlet parameters.
all the merits of the simple steady state methods and additionally, it Gawad et al. [81] proposed artificial neural network (ANN) and
gives certain advantages of the detailed dynamic methods, as for adaptive neuro-fuzzy inference system (ANFIS) artificial intelligent
example, the interaction between the envelope and the HVAC systems techniques to estimate the critical characteristics of the rotary desiccant
of the building. Kalogirou [5] carried out an important study to review dehumidifier. The parameters used in the modeling are like as flow
various applications of artificial neural network in energy systems used velocity, shape of cross-section of flow path and matrix material.
in building cooling. Errors reported in models discussed in review were Moreover, both these artificial intelligence techniques use experimen-
well a within acceptable limit, which suggests that artificial neural tally measured results for training and validation of model.
networks can readily be used for modeling in other fields of energy Zendehboudi [82] proposed hybrid model based on least squares
production and use. support vector machine (LSSVM) and genetic algorithm (GA) to
Artificial neural network modeling of rotary desiccant dehumidifier estimate DBT and humidity ratio of process air as well as regeneration
was carried out first by Cejudo [75]. The neural network model consists air at the dehumidifier exit, dehumidifier effectiveness, moisture
of a four-input–four-output network that calculates the outlet condi- removal capacity (MRC), and sensible energy ratio (SER) for desiccant
tions from inlet ones. Real data are used to validate the ANN model and materials as silica gel and molecular sieve by considering different
to train the neural network. A small discrepancy between calculated process/regeneration section (P/R) area ratios. The capability of the
and measured values was found mainly due to the fact that the model was evaluated through three different statistical error tests. The
dehumidification is assumed to be adiabatic in theoretical model as obtained results show that, integration of LSSVM and GA is very
well as the heat losses in the ducts and the dehumidifier were not accurate technique for estimating the performance of desiccant wheel,
considered in the ANN model, but in the experimental facility these having a mean average error (MAE) is less than 0.23, determination
losses can occur. The second attempt was made by Parmar and coefficient (R2) is greater than 0.994, and mean squared error (MSE) is
Hindoliya [76] to model desiccant dehumidifier with help of ANN. A less than 0.072. Jani et al. [30] developed an artificial neural network
neural network model was proposed using a neural network toolbox of (ANN) model to predict the performance of a desiccant wheel for
MATLAB with feed forward back propagation method. The model was different process air inlet conditions. DBT, humidity ratio and flow rate
validated against experimental data sets. A number of training algo- of the process as well as regeneration air streams of dehumidifier and
rithms with feed forward back propagation method were used for the regeneration temperatures are used as inputs to the model. The outputs
modeling of rotary dehumidifier to identify a training algorithm with of the model are outlet DBT and humidity ratio of process as well as
least mean square error (MSE). The performance of all training regeneration air streams of dehumidifier, the moisture removal rate
algorithms was analyzed and training algorithm trainlm (Levenberg– (MRR) and the effectiveness of the dehumidifier. Performance estima-
Marquardt back propagation) is found most suitable for the estimation tions through ANN are compared with the experimentally measured
of outputs which have least mean square error. The proposed model data and a close agreement is observed between them.
was predicted the specific humidity and temperature at the dehumidi- Summary of dehumidifier parameters which are studied by pre-
fier exit well within the range of experimentally measured values. vious researchers have been tabulated in Table 2.
Characteristics of a dehumidifier made up of silica-gel were predicted Ertunc and Hosoz [83] developed an ANN model to predict the
using artificial neural networks (ANNs) over a wide range of operating performance of evaporative condenser cooling system. The ANN model
conditions by Mahmoud et al. [77]. Two different Multi-layer was used for estimating various performance parameters of the system,
Perceptron feed forward Neural Networks (MLPNN) were developed namely the condenser heat rejection rate, refrigerant mass flow rate,
to predict dry bulb temperature and relative humidity of process air compressor power, electric power input to the compressor motor and
stream at the dehumidifier exit during dehumidification and reactiva- the system coefficient of performance (COP). The ANN estimations are
tion operating modes. Training and testing process of the proposed in good match with the experimental values with higher correlation
ANN model was based on experimentally measured data. Effects of coefficients and very low root mean square errors. The obtained results
desiccant mass on air mass flow rate, shelves span and air dry bulb show that such systems can alternatively be modelled using ANNs
temperature were predicted. Results showed that the ANN model within a high degree of accuracy. Yigit and Ertunc [84] predicted dry
predicts data that are very close to the available experimental data bulb temperature and humidity ratio at exit of a sensible heat
358
D.B. Jani et al. Renewable and Sustainable Energy Reviews 80 (2017) 352–366
359
D.B. Jani et al. Renewable and Sustainable Energy Reviews 80 (2017) 352–366
Fig. 10. Influence of different input variables on exit conditions of process and reactivation air streams for standardised regression coefficient values [88].
360
D.B. Jani et al. Renewable and Sustainable Energy Reviews 80 (2017) 352–366
Fig. 11. The regression plot between ANN estimations and the experimental data base for (a) Training, (b) Validation, (c) Test, and (d) Combined set [29].
Table 5
The network performance with various neuron numbers in the hidden layer [94].
361
D.B. Jani et al. Renewable and Sustainable Energy Reviews 80 (2017) 352–366
362
D.B. Jani et al. Renewable and Sustainable Energy Reviews 80 (2017) 352–366
Fig. 16. Comparison of ANN model training results of (a) Log-Sig and (b) Tan-Sig transfer functions [99].
363
D.B. Jani et al. Renewable and Sustainable Energy Reviews 80 (2017) 352–366
Table 7
Summary of ANN applications for different space cooling systems.
364
D.B. Jani et al. Renewable and Sustainable Energy Reviews 80 (2017) 352–366
365
D.B. Jani et al. Renewable and Sustainable Energy Reviews 80 (2017) 352–366
desiccant air-conditioning by adding a dynamic correction factor. Appl Therm Eng [100] Wang H, Cai W, Wang Y. Modeling of a hybrid ejector air conditioning system
2015;78:101–9. using artificial neural networks. Energy Convers Manag 2016;127:11–24.
[80] Rafique MM. A statistical analysis of desiccant dehumidifier for air conditioning [101] Sohani A, Zabihigivi M, Moradi MH, Sayyaadi H, Balyani HH. A comprehensive
application. Int J Hybrid Inform Technol 2015;8:245–56. performance investigation of cellulose evaporative cooling pad systems using
[81] Gawad AFA, Radhwi MN, Wafiah AM, Softah GJ. Computational and artificial predictive approaches. Appl Therm Eng 2017;110:1589–608.
intelligence study of the parameters affecting the performance of heat recovery [102] Mohanraj M, Jayaraj S, Muraleedharan C. Applications of artificial neural
wheels. Am J Energy Eng 2015;3:79–94. networks for refrigeration, air-conditioning and heat pump systems—a review.
[82] Zendehboudi A. Implementation of GA-LSSVM modelling approach for estimating Renew Sustain Energy Rev 2012;16:1340–58.
the performance of solid desiccant wheels. Energy Convers Manag [103] Mohanraj M, Jayaraj S, Muraleedharan C. Applications of artificial neural
2016;127:245–55. networks for thermal analysis of heat exchangers – a review. Int J Therm Sci
[83] Ertunc HM, Hosoz M. Artificial neural network analysis of a refrigeration system 2015;90:150–72.
with an evaporative condenser. Appl Therm Eng 2006;26:627–35. [104] Sahlot M, Riffat SB. Desiccant cooling systems: a review. Int J Low-Carbon
[84] Yigit KS, Ertunc HM. Prediction of the air temperature and humidity at the outlet Technol 2016;11:489–505.
of a cooling coil using neural networks. Int Commun Heat Mass Transf [105] Saidur R, Masjuki HH, Jamiludhin MY. A new method to investigate the energy
2006;33:898–907. performance of a household refrigerator-freezer. Int Energy J 2006;7:9–15.
[85] Ertunc HM, Hosoz M. Comparative analysis of an evaporative condenser using [106] Hosoz MH, Ertunc M. Modeling of a cascade refrigeration system using artificial
artificial neural network and adaptive neuro-fuzzy inference system. Int J Refrig neural networks. Int J Energy Res 2006;30:1200–15.
2008;31:1426–36. [107] Bacthler H, Browne MW, Bansal PK, Kecman V. Neural networks—a new
[86] Hosoz M, Ertunc HM, Ozguc AF. Modelling of a direct evaporative air cooler using approach to model vapour compression heat pumps. Int J Energy Res
artificial neural network. Int J Energy Res 2008;32:83–9. 2001;25:591–9.
[87] Kiran TR, Rajput SPS. An effectiveness model for an indirect evaporative cooling [108] Swider DJ, Browne MW, Bansal PK, Kecman V. Modeling of vapour compression
(IEC) system: comparison of artificial neural networks (ANN), adaptive neuro- liquid chillers with neural networks. Appl Therm Eng 2001;21:311–29.
fuzzy inference system (ANFIS) and fuzzy inference system (FIS) approach. Appl [109] Wei X, Li N, Peng J, Cheng J, Su L, Hu J. Analysis of the effect of the CaCl2 mass
Soft Comput 2011;11:3525–33. fraction on the efficiency of a heat pump integrated heat-source tower using an
[88] Koronaki IP, Rogdakis E, Kakatsiou T. Thermodynamic analysis of an open cycle artificial neural network model. Sustainability 2016;8:1–14.
solid desiccant cooling system using artificial neural network. Energy Convers [110] Zheng X, Ge TS, Wang RZ. Recent progress on desiccant materials for solid
Manag 2012;60:152–60. desiccant cooling systems. Energy 2014;1:1–15.
[89] Sohani A, Sayyaadi H, Balyani HH, Hoseinpoori S. A novel approach using [111] Yang J, Chan KT, Dai T, Yu FW, Chen L. Hybrid artificial neural network – genetic
predictive models for performance analysis of desiccant enhanced evaporative algorithm technique for condensing temperature control of air – cooled chillers.
cooling systems; 2016. 〈http://doi.org/10.1016/j.applthermaleng.2016.06.121〉. Procedia Eng 2015;121:706–13.
[90] Hosoz M, Ertunc HM. Artificial neural network analysis of an automobile air [112] Deb C, Eang LS, Yang J, Santamouris M. Forecasting diurnal cooling energy load
conditioning system. Energy Convers Manag 2006;47:1574–87. for institutional buildings using artificial neural network. Energy Build
[91] Hosoz M, Direk M. Performance evaluation of an integrated automotive air 2015;1:1–14.
conditioning and heat pump system. Energy Convers Manag 2006;47:545–59. [113] Lee SJ, Kim HJ, Dong HW, Jeong JW Energy saving assessment of a desiccant –
[92] Yilmaz S, Atik K. Modeling of a mechanical cooling system with variable cooling enhanced evaporative cooling system in variable air volume applications. Appl
capacity by using artificial neural network. Appl Therm Eng 2007;27:2308–13. Therm Eng. 〈http://dx.doi.org/10.1016/j.applthermaleng.2017.02.007〉.
[93] Gandhidasan P, Mohandes MA. Predictions of vapor pressures of aqueous [114] Guo J, Lin S, Bilbao JI, White SD, Sproul AB. A review of photovoltaic thermal
desiccants for cooling applications by using artificial neural networks. Appl Therm (PV/T) heat utilisation with low temperature desiccant cooling and dehumidifi-
Eng 2008;28:126–35. cation. Renew Sustain Energy Rev 2017;67:1–14.
[94] Atik K, Aktas A, Deniz E. Performance parameters estimation of MAC by using [115] Rafael PM, Alexis PF, Carlos RB, Jesus APA. Comparison of linear regression and
artificial neural network. Expert Syst Appl 2010;37:5436–42. artificial neural networks models to predict heating and cooling energy demand,
[95] Li N, Xia L, Shiming D, Xu X, Chan MY. Dynamic modeling and control of a direct energy consumption and CO2 emissions. Energy 2017;118:24–36.
expansion air conditioning system. Appl Energy 2011;91:290–300. [116] Erdemir D, Ayata T. Prediction of temperature decreasing on a green roof by using
[96] Hernandez JA, Rivera W, Colorado D, Moreno-Quintanar G. Optimal COP artificial neural network. Appl Therm Eng 2017. http://dx.doi.org/10.1016/
prediction of a solar intermittent refrigeration system for ice production by means j.applthermaleng.2016.10.145.
of direct and inverse artificial neural networks. Sol Energy 2012;86:1108–17. [117] Alaidroos A, Krarti M. Numerical modeling of ventilated wall cavities with spray
[97] Kamar HM, Ahmad R, Kamsah NB, Mustafa AFM. Artificial neural networks for evaporative cooling system. Energy Build 2016;130:350–65.
automotive air-conditioning systems performance prediction. Appl Therm Energy [118] Speerforck A, Schmitz G. Experimental investigation of a ground-coupled
2013;50:63–70. desiccant assisted air conditioning system. Appl Energy 2016;181:575–85.
[98] Mora JE, Carlos PT, Gonzalez FN, Ocampo JD. Thermodynamic properties of [119] Yin C, Rosendahl L, Luo Z. Methods to improve prediction performance of ANN
refrigerants using artificial neural networks. Int J Refrig 2014;46:9–16. models. Simul Model Pract Theory 2003;11:211–22.
[99] Tian Z, Qian CH, Gu B, Yang L, Liu F. Electric vehicle air conditioning system
performance prediction based on artificial neural network. Appl Therm Eng
2015;89:101–14.
366