=Paper=
{{Paper
|id=Vol-3118/p13
|storemode=property
|title=Prediction of Refrigeration System Performance Using
Artificial Neural Networks
|pdfUrl=https://ceur-ws.org/Vol-3118/p13.pdf
|volume=Vol-3118
|authors=Iyad Lafta Majid,Ahmed A. M. Saleh,Alaa Abdulhady Jaber
|dblpUrl=https://dblp.org/rec/conf/icyrime/MajidSJ21
}}
==Prediction of Refrigeration System Performance Using
Artificial Neural Networks==
https://ceur-ws.org/Vol-3118/p13.pdf
Prediction of Refrigeration System Performance Using
Artificial Neural Networks
Iyad Lafta Majid1 , Ahmed A. M. Saleh1 and Alaa Abdulhady Jaber1
1
University of Technology, Mechanical Engineering Department, Baghdad, Iraq
Abstract
In this research, a review of the previously conducted simulation approaches for estimating the performance of the vapor
pressure refrigeration systems has been performed. It was found that some researchers have followed the mathematical
approaches, which are based on the principles of thermodynamics, to measure the performance, while many others investigated
the use of artificial intelligence methods. The artificial neural network (ANN) is one of the most widely used methods in
this research field. It showed that the ANN could predict the effect of almost all parameters that significantly impact the
performance of compressed vapor refrigeration systems. Also, ANN was efficient and rapid in reducing time and costs, and
most of the obtained results were close to the experimental results. However, most of the accomplished research considered
the effect of no more than three to four parameters simultaneously. Thus, it is recommended to investigate the concurrent
influence of more and different parameters.
Keywords
artificial neural networks, refrigeration systems, energy efficiency,
1. Introduction 2. Application of ANN for
Refrigeration system production has been significantly Refrigeration Systems
rising in recent decades, has become increasingly vital Performance Predication
in people’s everyday lives. As a result, improving the
refrigeration system design process’s efficiency and prod- Artificial intelligence (AI), machine learning (ML) and
uct performance is critical. One of the most useful tools many other statistical approaches have widely being
for achieving this goal is a computer simulation. The used for different prediction and automation applications
working circumstances and configuration parameters of [9, 10, 11, 12, 13, 14, 15, 16, 17, 18]. However, in terms
the product are supplied first, then the performance is of cooling and refrigeration systems, Prabha, et al. [19]
anticipated, and finally, the configuration parameters of used mathematical models to investigate the impact of
the product are evaluated based on the performance pre- refrigeration system characteristics such as evaporating
diction. If the anticipated performance does not match temperature, condensing temperature, and the mass of
the requirement, the configuration settings should be the refrigerant charge utilized on the system’s perfor-
tweaked, and the simulation should be run again with mance. The researchers constructed such mathematical
the tweaked structural parameters. The process of chang- models to conduct the needed tests utilizing three vari-
ing the parameters and simulating with those changes ables (evaporating temperature, condensing temperature,
will be continued until a set of the best-suited settings and mass of the refrigerant charge) and two levels facto-
is found [1, 2, 3, 4, 5, 6, 7, 8]. This paper describes some rial method refrigerating effect and compressor power.
of the modeling approaches for vapor compression re- The impact of various system variables and their substan-
frigeration systems research that have been published in tial interaction effects on answers were estimated using
various journals or conferences. The findings of these MINITAB software, based on these mathematical models
studies will be deliberated, and the essential elements established for predicting the values of replies. For the
influencing the cooling system performance and stability refrigerants R290/R600, R290/R600a, and LPG, the perfor-
are evaluated. The results of simulation approaches will mance was determined. The influence of system factors
be examined and discussed, and the conclusions made by on performance may be explained using mathematical
the researchers are discussed. Finally, recommendations models. However, the following was concluded:
for future work are provided.
• The models created for a refrigeration system’s
ICYRIME 2021 @ International Conference of Yearly Reports on performance parameters were simple first-order
Informatics Mathematics and Engineering, online, July 9, 2021 quadratic equations correlating the system’s per-
" ayaden88@yahoo.com (I. L. Majid); aamsaleh60@yahoo.com formance parameters. These created models may
(A. A. M. Saleh); alaa.a.jaber@uotechnology.edu.iq (A. A. Jaber)
0000-0001-5709-195X (A. A. Jaber)
be used to forecast system performance based on
© 2021 Copyright for this paper by its authors. Use permitted under Creative any collection of system factors.
Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
Workshop
Proceedings
http://ceur-ws.org
ISSN 1613-0073
CEUR Workshop Proceedings (CEUR-WS.org) • The evaporating temperature has a greater impact
90
Iyad Lafta Majid et al. CEUR Workshop Proceedings 90–98
on refrigerating capacity than refrigerant mass
or condensing temperature.
• The power consumed by the compressor increases
as the evaporating temperature increases.
• Changes in refrigerant mass and condensing tem-
perature have comparable effects on compressor
power but are less substantial.
The vapor compression refrigeration system (VCRS)
components (compressor, condenser, capillary tube, and
evaporator) were tested for irreversibility employing R134
a/LPG refrigerant as a replacement to R134a [20]. Various
experiments were conducted for different temperatures
of evaporator and condenser under restricted settings
to achieve this goal. Under identical experimental set- Figure 1: Figure 1 ANN results for COP [22]
tings, irreversibility in the components of VCRS using
R134a/LPG (liquefied petroleum gas) was lower than ir-
reversibility in the components of VCRS using R134a.
The second law of efficiency and total irreversibility of matrics arr cooling capacity, power consumption, and
the refrigeration system were predicted using artificial COP. Cross-validation was used to validate each model,
neural network (ANN) models. The absolute fraction resulting in minimum relative errors of 0:15 for cool-
of variance in the range of 0.980–0.994 and 0.951–0.977, ing capacity and coefficient of performance and 0.05 for
root-mean-square error in the range of 0.1636–0.2387 power consumption. Computer simulations were con-
and 0.2501–0.4542, and mean absolute percentage error ducted based on the required validation results to create
in the range of 0.159–0.572 and 0.308–0.931 percent, re- 3D colored figures, as shown in Figure 1. After examining
spectively, were anticipated using the ANN and ANFIS these 3D color surfaces, it was determined that R450A
(adaptive neuro-fuzzy inference system) models. The re- had a little lower cooling capacity than R134a, with a
sults reveal that the ANN model outperforms the ANFIS 10% drop in cooling capacity calculated. Similar findings
model in terms of statistical prediction. were found in power usage, with R450A using around
OUYANG and KANG [21] developed a model for fore- 10% less electricity than the other two refrigerants. R134a
casting the COP of a supermarket refrigeration system. and R513A, conversely, were shown to have extremely
For this purpose, ANN models were created utilizing on- comparable energy characteristics. In terms of COP, it
site testing data. The BP (Back Propagation) and RBF was determined that all three refrigerants behaved in a
(Radial Basis Function) neural networks were trained, relatively comparable manner. After using ANNs and 3D
and the BP network model was optimized using the ge- surface color to analyze the data, it was determined that
netic algorithm (GA). The results showed that both the BP R450A and R513A are suitable refrigerants to substitute
and RBF neural networks could estimate the refrigeration R134a in medium evaporation temperature applications
system’s COP, and the prediction results are extremely in the short term.
close to the real data. The BP and the BP-based GA mod- Instead of CFCs (R12, R22, and R502), HFC- and HC-
els’ mean relative error (MRE) is 1.82 percent and 0.62 based refrigerants and their blends were studied by Ar-
percent, respectively, and their R2 is 0.9518 and 0.9889, caklioǧlu [23]. The COP values of vapor-compression
demonstrating that the BP network can be efficiently op- refrigeration systems were obtained through the ANN
timized using the genetic algorithm. However, the RBF with different refrigerants and their abovementioned mix-
model outperformed the other two techniques. It has tures. In a vapor compression refrigeration system with
the lowest training time and the highest prediction ac- a liquid/suction line heat exchanger, ternary and quartet
curacy with a mean relative error of 0.21 percent and an mixtures of various ratios were calculated to train the
R2 of 0.9996. To determine the input variables of ANN network. The input layer consisted of refrigerant mix-
models, all variables from operating data are analyzed, ing ratios and evaporator temperature, while the output
and it was found that COP is related to all variables. In layer outputs the COP value. The outcomes demonstrate
another research, the ANN was employed to model a that the absolute proportion of variance (R2) values were
micro-cooling system [22]. The primary goal of this re- about 0.9999, and the root means square error (RMSR)
search was to compare the energy efficiency of three values are less than 0.002.
refrigerants: R134a, R450A, and R513A. The ANN was An ANN was used to analyze the COP for a compres-
used to estimate three common energy metrics as a func- sion vapor system using R1234yf [24]. A laboratory test
tion of evaporating and condensing temperatures. These was created to evaluate many parameters at the refriger-
91
Iyad Lafta Majid et al. CEUR Workshop Proceedings 90–98
ation system’s input and output. The temperature, com- 38071 system samples, encompassing transient and sta-
pressor rotation speed, and volumetric flow in the sec- tionary states, all sensors’ signals and those given by mea-
ondary fluids were the input variables. The behavior of suring equipment were utilized. A variable speed vapor
the refrigeration system was modeled using the ANN. compression system was monitored, and numerous vari-
To test the effect of these parameters on the COP, a uni- ables were measured and stored to prepare the training
formly distributed random variable was applied to one set. These metrics were utilized to create both the train-
of the ANN’s inputs. The refrigeration system was ana- ing and validation sets. The samples were separated into
lyzed, and the optimum performance was observed uti- two groups: 85 percent (32360) were utilized for training,
lizing computer simulations employing artificial neural while the remaining 15% (5711) were used for validation.
networks. A data set comprising 54650 values, count- The findings show that the ANN can accurately envisage
ing input and output variables, was built, resulting in an the actual operative behavior of this type of installation.
output set of values for each input set. The data were The following were predicted: COP, compressor power
arbitrarily divided into two sets: one for training and consumption, cooling capacity, the water temperature at
the other for verification. 70% of the measurements were the condenser outlet, and water-glycol temperature at
utilized to create the training set, while the remaining the evaporator outlet. Ertunc and Hosoz [27] developed
30% were used for validation. The following conclusions an experimental R134a vapor-compression refrigeration
were drawn from the simulation results: •The energy per- unit to test an ANN model. K-type thermocouples were
formance is substantially influenced by the temperature used for all temperature measurements. The intake and
variation of the coolant condenser liquid. • The volumet- exit of each component were fitted with refrigerant ther-
ric flow rates of the coolant condenser liquid have only a mocouples. The airstream entering and leaving the evap-
little impact on energy efficiency (COP). • Variations in orative condenser was measured for both dry and wet
the coolant condenser liquid temperature significantly bulb temperatures. At the compressor’s input and out-
impact the installation’s energy performance variability. flow, refrigerant pressures were recorded. Variable-area
Kamar, et al. [25] investigated the use of the ANN flow meters were used to measure the refrigerant and wa-
to forecast the cooling capacity, compressor power in- ter mass flow rates. In the experimental work, 60 distinct
put, and coefficient of performance in a conventional steady-state test runs were performed to collect training
air-conditioning system for a passenger car. The evap- data and evaluate the proposed ANN. The evaporator
orator, condenser, compressor, and expansion valve are load, air mass flow rate, water mass flow rate, air dry
the four primary components of the system. Tempera- bulb, and wet bulb temperatures at the condenser intake
ture sensors were utilized to measure the cooling fluid, are all inputs to the ANN. The condenser heat rejection
air inside and outside temperatures on the evaporator rate, refrigerant mass flow rate, compressor power ab-
and the condenser. Also, a flow meter, a compressor’s sorbed by the refrigerant, electric power spent by the
speed meter, and a pressure gauge were used to collect compressor motor, and coefficient of performance are
the experimental data. The compressor speed, air tem- the ANN’s outputs. Various system performance charac-
perature at the evaporator inlet, air temperature at the teristics were computed from thermal analysis equations
condenser inlet, and air velocity at the evaporator inlet based on the experimental results and utilized to create
were all varied at steady-state conditions in the experi- and test the ANN model. The available data set from
mental setup. The correlation between the ANN model’s the experimental work was divided into training and
anticipated outputs and the experimental data has a good validation sets to create the ANN for the experimental
agreement in forecasting the system performance. refrigeration system. The training set was assigned to
In another research, ANNs were utilized to build and 70% of the data, while the remaining 30% was used for
validate a variable speed vapor compression device [26]. network testing and validation. For performance evalu-
The experimental test bench is made up of one vapor ation, the projected output parameters were compared
compression circuit and two secondary fluids circuits. to the experimental ones. The RMSE values for the pre-
The R134a working fluid is used in the vapor compres- dicted parameters were quite low when compared to the
sion circuit, which is a single-stage compression system. experimental ranges. The significance of what was dis-
Compressor rotation speed, volumetric flow rates, and covered in this study is that a refrigeration system with
secondary fluid temperatures are the model’s input pa- an evaporating condenser, which is perhaps the most
rameters. The coefficient of performance, compressor difficult to predict using traditional methods, can be mod-
power consumption, cooling capacity, the water temper- eled using ANNs with excellent accuracy. This assists
ature at the condenser outlet, and water-glycol temper- application engineers and makers of these systems in
ature at the evaporator outlet are the model’s output quickly determining their performance without the need
parameters. Sensors in the experimental facility measure for extensive testing.
pressure, temperature, volumetric flow rate, mass flow For illustrating mass flow rate through straight and
rate, compressor speed, and energy usage. To collect helical coil adiabatic capillary tubes in a vapor compres-
92
Iyad Lafta Majid et al. CEUR Workshop Proceedings 90–98
sion refrigeration system, an experimental investigation frequency can save substantial energy. The thermody-
was done with R134a and LPG refrigerant mixture [28]. namic analysis of refrigeration systems can be simplified
Various studies were carried out under steady-state set- using this methodology, and the predicted values were
tings, varying the length of the capillary tube, the in- extremely similar to the actual values.
ner diameter, the coil diameter, and the degree of sub- An experimental configuration of a single-door house-
cooling. The system’s primary components are as follows: hold refrigerator working with R134a and a total capac-
compressor, condenser, expansion valve, evaporator, and ity of 175L was utilized to predict the performance of a
other accessories. The results showed that the mass flow domestic refrigeration system employing R436A as an
rate through helical coil capillary tubes was 5-16% lower alternate refrigerant to R134a [30]. This method can be
than straight capillary tubes. Dimensionless correlation used to calculate the cooling effect, power consumption,
and Artificial Neural Network (ANN) models were con- and performance coefficient of a domestic refrigerator.
structed to forecast the mass flow rate, which was found Seven thermocouple sensors were employed inside the
to be in good agreement with the experimental results, freezer, refrigerator cabin, evaporator, compressor, con-
with absolute fractions of variance of 0.961. The results denser inlets, and outlets. The investigation started by
indicated that the ANN model performed statistically charging R436A mass into the system and calculating
better because ANN model predictions were closer to cooling capacity, compressor effort, and COP for various
experimental values than the dimensionless correlation capillary tube lengths. However, the same capillary tube
model. length was used throughout the R134a tests. Continu-
Based on using artificial neural networks and limited ous tests were conducted throughout the conditions de-
data sets, a study was conducted to estimate the thermo- scribed above, with the evaporator temperature reaching
dynamic performance of an experimental refrigeration -15°C. The pull-down properties and performance factors
system driven by a variable speed compressor [29]. A such as cooling capacity, power consumption, and system
semi-hermetic compressor, an evaporator, a condenser, performance can be determined first. For reference, the
and an externally equalized thermostatic expansion valve ambient temperature should be kept at about 29°C when
make up the experimental variable speed refrigeration changing capillary tubes and refrigerant weights. After
system. The evaporator and condenser were finned tube establishing steady-state conditions, total experimental
heat exchangers that are air-cooled. The evaporator was values were collected. With the experimental data, the
housed in a specially built cold room with electric heaters ANN forecasts for compressor power provide an average
to simulate the refrigeration demand. Temperature and inaccuracy of 2.51% . These results show that, despite
pressure measurements were taken from specific points the wide range of operating conditions, the ANN accu-
of the experimental system to evaluate the system per- rately forecasts the power absorbed by the refrigerant in
formance by modulating the compressor capacity with the compressor. Compared to the experimental COP, the
an inverter. A flow meter built for refrigerant R404a was ANN forecasts are the average error for these predictions
used to measure the mass flow rate of the refrigerant. A is 1.23. The COP forecasts of the ANN were as accurate as
tiny humidity measurement equipment was also used to those of the other performance metrics predictions. The
measure air humidity at the intake and outflow of the results showed that in a household refrigeration system,
condenser channel. Temperatures were taken at 12 places the hydrocarbon refrigerant mixture R436A performs
throughout the system, the pressure was taken at seven better than R134a. For a few input values, the ANN pro-
places, and the refrigerant mass flow rate was measured vides a good response with a significant amount of error.
after the condenser. All of the measurement equipment is As a result, the R436A could be a more energy-efficient,
wired into a data logger with 20 channels for data collec- ozone-friendly, safe, and long-term replacement fluid for
tion. A computer was also attached to the data recorder. R134a in a system. All of the hardware in a residential
All measurements were taken every 5 seconds, and the refrigerator remains the same, except the length of the
data was recorded on a computer using a data logger. capillary tubes, and there was no need to change the
ANNs were employed to study the performance of the lubricating oil when using R436A as the refrigerant.
variable speed refrigeration system, which was the major Yilmaz and Atik [31] established an experimental va-
goal of this research. Compressor frequency, cooling load, por compressor refrigeration setup to explore using R134A
condenser and evaporator temperatures, and condenser as a refrigerant in a vapor compression refrigeration sys-
and evaporator pressures are all input parameters. The tem (Figure 2). The data received from the test results
output parameters were the compressor power consump- were then utilized to simulate the system performance
tion, refrigerant mass flow rate, and experimental and Based on the outcomes of the trials and the accompany-
theoretical COP values. Instead of conducting numerous ing computed coefficient of performance, an Artificial
studies, it was discovered that using a neural network Neural Network was created. In this case, a hermetically
approach was more efficient. This research showed that sealed compressor was used. The power consumed by the
using neural networks to determine the best compressor compressor was measured with an energy meter. After
93
Iyad Lafta Majid et al. CEUR Workshop Proceedings 90–98
and 38 data patterns for Chiller A and B, respectively.
Chiller A had a coefficient of variation of less than 1.5
percent, while Chiller B had a coefficient of variation of
3.9 percent, precisely forecasting the COP.
The dynamics of a vapor compression cycle were also
modeled using artificial neural networks [33]. A semi-
hermetic reciprocating compressor, an air-cooled finned-
tube condenser, three electronic expansion valves, and
three evaporators make up the vapor compression cy-
cle system (one air-cooled finned-tube evaporator and
two electronic evaporators). The input air temperature
of the condenser is controlled by one air duct heater
to simulate outdoor conditions, while the inlet air tem-
perature of the evaporator is maintained at 25 C by the
HVAC system. R134a is the working fluid in the sys-
Figure 2: Figure 2 The developed experimental set-up by [31] tem. The compressor, condenser fan, and evaporator
fan all have inverters to modify their respective frequen-
cies. To adjust the condenser’s incoming air tempera-
the compressor, an air-cooled condenser with a water- ture, a heater was mounted in front of the condenser.
cooled evaporator was installed, with air cooling and a These components are connected in a closed-loop so that
thermostatic expansion valve. The parameters were the the working fluid can be circulated constantly through-
refrigerant temperatures entering and leaving the com- out the system. Compressor rotation speed, evaporator
pressor, condenser, and evaporator, the air temperatures fan frequency, condenser fan frequency, expansion valve
entering and leaving the condenser, and the water tem- opening percentage, outdoor temperature, and indoor
peratures entering and leaving the evaporator inlet and temperature are input parameters, condensing pressure,
outlet pressures evaporator and condenser. The ANN evaporating pressure, subcool, superheat, and system
model has excellent statistical performance as measured power consumption is examples of output parameters.
by the correlation coefficient (R) and the MSE. With a co- The artificial neural networks model may achieve the
efficient of correlation higher than 0.988 and a maximum minimum modeling error and significantly robustness
percentage of error of less than 5% , the outputs pre- against input disturbances and system uncertainties. The
dicted by the ANN model match with experimental data. testing and comparing results using experimental data
The results show that the ANN model can be used suc- have further proven the neural model’s remarkable per-
cessfully to estimate the performance of a very accurate formance.
and dependable vapor compression refrigeration system. Hosoz, et al. [34] The operation of a vapor-compression
Swider, et al. [32] have applied the neural networks were refrigeration system using R134a as the working fluid
to compress vapor in two hermetic vapor-compression and a counter-flow cooling tower was estimated using
liquid chillers, a single-circuited single-screw (Chiller artificial neural networks. The model was then used to
A) and a twin-circuited twin-screw (Chiller B) (Chiller predict numerous performance variables of the refrig-
B). The chilled water outlet temperature is the most im- eration system, including the evaporating temperature,
portant factor in determining the cooling capability of compressor power, coefficient of performance, and the
each chiller. In order to anticipate chiller performance, temperature of the water stream leaving the tower. The
the neural network used the chilled water outlet tem- refrigeration system consists of a reciprocating compres-
perature from the evaporator, the cooling water inlet sor, a water-cooled condenser coupled to the cooling
temperature from the condenser, and the evaporator ca- tower, a thermostatic expansion valve, and an electrically
pacity as input parameters for both chillers. The neural heated evaporator. The system was charged with 600
network chiller models have statistical findings for both g of R134a. For all temperature measurements, K-type
the chiller’s COP and the electrical work. In the lab, the thermocouples were employed. The dry and wet bulb
mass flow rates, inlet and outlet temperatures of chilled temperatures of the air stream were measured at the cool-
and cooling water, and compressor input were all mea- ing tower’s entrance and output. The evaporating and
sured. The cooling water mass flow rate changed during condensing pressures were monitored using Bourdon
various combined chiller operations. This is done for 450 tube gauges. The refrigerant and water mass flow rates
of the 500 measured data patterns in Chiller A and 342 were measured using variable-area flow meters. To build
of the 380 observed data patterns in Chiller B. Only the an artificial neural networks model for the experimental
final validation of the model requires the remaining 50 refrigeration system, the available data set, which con-
sisted of 64 input vectors and their corresponding output
94
Iyad Lafta Majid et al. CEUR Workshop Proceedings 90–98
vectors from the experimental work, was separated into to unity. The largest differences between ANN forecasts
training and test sets. The training set was randomly and experimental observations are 8.03 percent, 1.68 per-
assigned to 75% of the data set, while the remaining 25% cent, and 11.85 percent, respectively, for cooling capacity,
evaluated the network’s performance. Evaporator load, compression work, and COP of the system. It suggests
dry bulb temperature and relative humidity of the air that a properly configured ANN could be a useful tool
stream entering the tower, air mass flow rate, and water for predicting the performance of automobile air con-
mass flow are the five input factors that determine the ditioning systems. This saves time and money in the
refrigeration system’s outputs. The refrigerant mass flow simulation by avoiding the complexity of a first principle-
rate, compressor, condenser heat rejection, coefficient based simulation. Hosoz, et al. [35] used artificial neu-
of performance, evaporating temperature, compressor ral networks to model several mobile air conditioning
discharge temperature, water temperature at the cool- (MAC) system performance metrics. Instead, soft com-
ing tower outlet, and water mass flow rate refrigeration puting techniques, such as the ANNs, can be employed to
system with the cooling tower are all output parame- simulate MAC systems and estimate their performance
ters. Artificial neural networks could adequately depict under various operating scenarios. The performance of
refrigeration systems with cooling towers, according to a MAC system using the alternative refrigerant R1234yf
the findings. This novel technique requires a modest was modeled using the ANN technique. The created
number of experiments rather than extensive experimen- ANN model’s predictions were then compared to exper-
tal study or dealing with a large mathematical model. imental results using statistical performance measures.
Datta, et al. [? ] used the ANN to forecast the thermal A five-cylinder swash plate compressor, a parallel-flow
performance of a vehicle air conditioning system was ex- micro-channel condenser, a laminated type evaporator, a
amined. A finned tube condenser and evaporator, as well receiver/filter/drier, and a thermostatic expansion valve
as a swashplate fixed displacement compressor (driven by were designed for the proposed ANN model from the orig-
the engine) and a thermostatic expansion valve, make up inal components of an R134a MAC system of a compact
the primary refrigerating unit. During start-up and stop, car (TXV). A data collection system was often used to col-
the compressor’s speed fluctuates in lockstep with the lect the measured variables, which were then recorded on
engine’s. A three-phase motor with variable frequency a computer. The compressor speed, intake temperatures
drive was employed as the compressor’s primary mover of the evaporator and condenser air streams, and relative
to explore the influence of speed fluctuation. R134a, the humidity of the air at the evaporator inlet were used as
same refrigerant used in vehicle air conditioning systems, input parameters for the proposed ANN model. The cool-
is also used in the test rig. Separate ducting has been built ing capacity, power absorbed by the refrigerant in the
at the input and outlet plenums of the evaporator and compressor, condenser heat rejection rate, coefficient of
condenser to guide and measure the air streams’ flow rate performance, conditioned air temperature, compressor
and temperature. A variac-controlled electrical heater discharge temperature, refrigerant mass flow rate, and
is installed downstream of the evaporator duct to create pressure ratio across the compressor were the output pa-
variable heat loads similar to that of a driving car. The rameters, on the other hand. The generated ANN model
compressor and blower speeds, as well as the refriger- gives quite accurate predictions, with correlation coeffi-
ant temperature and pressure at various locations, the cients in the range of 0.9159–0.9962 and mean relative er-
refrigerant mass flow rate, the air dry-bulb temperature rors in 2.24–7.46 percent. The findings suggested that an
and relative humidity at the evaporator and condenser ANN technique can be utilized to predict the performance
inlet and outlet, and the airflow rate through both the of R1234yf MAC systems. The ANN model produced very
evaporator and the condenser are all measured. All of accurate predictions for the performance characteristics
the system’s input parameters are the refrigerant charge, of the MAC system. These findings showed that MAC
compressor speed, and blower speed. The output param- systems could be effectively represented using an ANN
eters are the cooling capacity, compression work, and technique rather than comprehensive experiments or
COP. The experimental outcomes are used to build the complex mathematical modeling.
training, testing, and validation data sets. The following In [36], the ANN approach was used to predict various
metrics were randomly chosen: training takes up 70% of performance parameters of a cascade vapor compression
the budget (42), while testing and validation take up the refrigeration system using R134a in lower and higher-
remaining 30% . The system’s performance can be accu- temperature refrigeration circuits. The suggested ANN
rately predicted by the ANN. RMSE, MRE, and EI were was trained and tested using steady-state test runs of
0.48-0.74 percent, 5.00-6.50 percent, and 0.80-2.01 percent, an experimental cascade refrigeration system. The ANN
respectively, in the performance measuring parameters. was used to forecast the evaporation temperature in the
When compared to the experimental results’ ranges, the lower circuit, compressor power for each circuit, COP
MSE values are incredibly low. The correlation coeffi- for the lower circuit, and COP for the overall cascade
cient of all performance parameters is extremely close system was developed using the backpropagation algo-
95
Iyad Lafta Majid et al. CEUR Workshop Proceedings 90–98
rithm. The correlation coefficient, mean relative error, and compressor power consumption were 1.87 percent,
and root means square error were used to evaluate the 2.71 percent, 1.79 percent, and 1.64 percent, respectively.
ANN predictions’ performance. The ANN forecasts for For refrigerant mass flow rate, condenser heat rejection,
the cascade refrigeration system usually performed sta- refrigeration capacity, compressor power consumption,
tistically well, with correlation coefficients ranging from root mean square errors were 0.0133 kg h1, 0.0141 kW,
0.953 to 0.996 and MREs ranging from 0.2 to 6.0 percent 0.0140 kW, and 0.0106 kW, respectively. The refrigerant
and extremely low RMSE values compared to the exper- mass flow rate had a correlation coefficient of 0.9983, the
imental data’ ranges. The ANN was utilized outside of condenser heat rejection had a correlation coefficient of
the experimental range to estimate system performance, 0.9980, the refrigeration capacity had a correlation coeffi-
and satisfactory prediction curves were obtained. This cient of 0.9975, and the compressor power consumption
research indicates that the ANN approach may be used had a correlation coefficient of 0.9979.
to model cascade vapor compression refrigeration sys-
tems instead of traditional modeling techniques. As a
result, instead of undertaking an intensive experimental 3. Conclusion
investigation or dealing with a complicated mathematical
This paper has conducted a thorough investigation of the
model, the performance parameters of these systems can
literature and published research from various sources,
be simply identified by doing only a minimum number of
including research papers, conference papers, and ear-
test runs. Tian, et al. [37] presented research on utilizing
lier investigations conducted to estimate refrigeration
an artificial neural network (ANN) to predict the ther-
systems’ performance. The ANN technique was widely
mal performance of a parallel flow (PF) condenser using
applied to determine the performance of an air condi-
R134a as the working fluid. The condenser was divided
tioning system; however, most of the studies looked at
into three sections: refrigerant side, tube side, and air
the impact of some parameters that affect the system per-
side. The following assumptions were made to make the
formance. The considered parameters were about three
research easier: Under steady conditions, all parameters
or four. Thus, it is recommended to consider more vari-
are constant, and the refrigerant and airflow are one-
ables that could affect the refrigeration systems. These
dimensional. Heat conduction along the axial direction
parameters include, for example, the system’s cooling
and radiation heat transfer is not considered; the system
fluid before and after each component, the air tempera-
comprises a refrigerant, air cooling, and heating loop
ture in contact with the evaporator and condenser, the
to ensure that the system automatically responds to the
refrigerant flow rate, the compressor rotation speed, and
presetting parameters. The four key components of the
the system’s high and low pressure. Also, the quantity
refrigerant loop section are the compressor, PF condenser,
of chilled airflow to the condenser and the ambient and
electric expansion valve, and evaporator. These four com-
external temperatures affecting the condenser should be
ponents were thought to be operating in a steady-state
simulated using a heat source to adjust the amount of
mode. The vacuum pump ran for two hours before circu-
heat passed on the condenser to examine its influence.
lating R134a through the system to empty the refrigerant
The system’s performance, the compressor’s volumet-
loop. ANN takes into account the dry temperature, wet
ric efficiency, the compression ratio, and the amount
bulb temperature, and velocity of the incoming air stream,
of deep chilling were projected to replicate most of the
mass flow rate, and the temperature and pressure of the
system’s characteristics. It should be emphasized that
refrigerant entering the condenser. All of the perfor-
ANN outperforms other nonlinear data approaches sig-
mance parameters’ R2 values were very close to unity,
nificantly. Without a prior understanding of the rela-
demonstrating that the ANN model can reliably predict
tionships between input and output variables, ANN can
the performance parameters of the PF condenser. Finally,
conduct nonlinear models. This will benefit designers
Tian, et al. [38] utilized an ANN technique to forecast the
in their decision-making and provide them with more
performance of an electric vehicle air conditioning sys-
flexibility in adjusting design criteria.
tem. Experiments were carried out by altering the scroll
compressor speeds, EEV apertures, and ambient tempera-
tures. 119 experimental data sets were acquired for ANN References
training and testing. Scroll compressor speed and EEV
were used as input variables to the ANN. The condenser [1] M. Hosoz, H. M. Ertunç, H. Bulgurcu, Performance
inlet air temperature, and evaporator inlet air tempera- prediction of a cooling tower using artificial neural
ture were used. The output variables included refrigerant network, Energy Conversion and Management 48
mass flow rate, condenser heat rejection, refrigeration (2007) 1349–1359.
capacity, and compressor energy consumption. In the es- [2] E. Pasqualotto, S. Federici, A. Simonetta,
tablished ANN, mean relative errors for refrigerant mass M. Olivetti Belardinelli, Usability of brain
flow rate, condenser heat rejection, refrigeration capacity, computer interfaces, in: Everyday Technology
96
Iyad Lafta Majid et al. CEUR Workshop Proceedings 90–98
for Independence and Care, IOS Press, 2011, pp. ning based on knowledge base and expertise shar-
481–488. ing, volume 2472, 2019, p. 41 – 47.
[3] A. Simonetta, M. C. Paoletti, Designing digital cir- [15] A. A. Jaber, K. M. Ali, Artificial neural network
cuits in multi-valued logic, International Journal based fault diagnosis of a pulley-belt rotating sys-
on Advanced Science, Engineering and Information tem, Int J Adv Sci Eng Inform Technol 9 (2019)
Technology 8 (2018) 1166–1172. 544–551.
[4] R. Brociek, G. De Magistris, F. Cardia, F. Coppa, [16] G. C. Cardarilli, L. D. Nunzio, R. Fazzolari, D. Gi-
S. Russo, Contagion prevention of covid-19 by ardino, A. Nannarelli, M. Re, S. Spanò, A pseudo-
means of touch detection for retail stores, volume softmax function for hardware-based high speed
3092, 2021, p. 89 – 94. image classification, Scientific Reports 11 (2021).
[5] F. Mazzenga, A. Simonetta, R. Giuliano, M. Vari, doi:10.1038/s41598-021-94691-7.
Applications of smart tagged rfid tapes for local- [17] S. Brusca, G. Capizzi, G. Lo Sciuto, G. Susi, A new
ization services in historical and cultural heritage design methodology to predict wind farm energy
environments, in: 2010 19th IEEE International production by means of a spiking neural network–
Workshops on Enabling Technologies: Infrastruc- based system, International Journal of Numerical
tures for Collaborative Enterprises, IEEE, 2010, pp. Modelling: Electronic Networks, Devices and Fields
186–191. 32 (2019) e2267.
[6] E. Pasqualotto, A. Simonetta, S. Federici, [18] N. Brandizzi, V. Bianco, G. Castro, S. Russo, A. Wa-
M. Olivetti Belardinelli, Usability evaluation jda, Automatic rgb inference based on facial emo-
of bcis (2009). tion recognition, volume 3092, 2021, p. 66 – 74.
[7] G. Capizzi, F. Bonanno, C. Napoli, Hybrid neural [19] G.-l. Ding, Recent developments in simulation
networks architectures for soc and voltage predic- techniques for vapour-compression refrigeration
tion of new generation batteries storage, in: 2011 systems, International Journal of refrigeration 30
International Conference on Clean Electrical Power (2007) 1119–1133.
(ICCEP), IEEE, 2011, pp. 341–344. [20] J. Gill, J. Singh, O. S. Ohunakin, D. S. Adelekan,
[8] G. De Magistris, S. Russo, P. Roma, J. Starczewski, Ann approach for irreversibility analysis of vapor
C. Napoli, An explainable fake news detector based compression refrigeration system using r134a/lpg
on named entity recognition and stance classifica- blend as replacement of r134a, Journal of Thermal
tion applied to covid-19, Information (Switzerland) Analysis and Calorimetry 135 (2019) 2495–2511.
13 (2022). doi:10.3390/info13030137. [21] W.-l. OUYANG, R.-q. KANG, Ann methods for
[9] R. Giuliano, G. C. Cardarilli, C. Cesarini, L. Di Nun- cop prediction of supermarket refrigeration sys-
zio, F. Fallucchi, R. Fazzolari, F. Mazzenga, M. Re, tem, DEStech Transactions on Computer Science
A. Vizzarri, Indoor localization system based on and Engineering (2019).
bluetooth low energy for museum applications, [22] J. Belman-Flores, A. Mota-Babiloni, S. Ledesma,
Electronics 9 (2020) 1055. P. Makhnatch, Using anns to approach to the en-
[10] F. Fallucchi, M. Coladangelo, R. Giuliano, E. William ergy performance for a small refrigeration system
De Luca, Predicting employee attrition using ma- working with r134a and two alternative lower gwp
chine learning techniques, Computers 9 (2020) 86. mixtures, Applied Thermal Engineering 127 (2017)
[11] F. Bonanno, G. Capizzi, S. Coco, C. Napoli, A. Lau- 996–1004.
dani, G. Sciuto, Optimal thicknesses determi- [23] E. Arcaklioğlu, Performance comparison of cfcs
nation in a multilayer structure to improve the with their substitutes using artificial neural net-
spp efficiency for photovoltaic devices by an hy- work, International Journal of Energy Research 28
brid fem - cascade neural network based approach, (2004) 1113–1125.
2014, p. 355 – 362. doi:10.1109/SPEEDAM.2014. [24] S. Ledesma, J. M. Belman-Flores, Analysis of cop
6872103. stability in a refrigeration system using artificial
[12] F. Fallucchi, H. Steffen, E. W. De Luca, Creating neural networks, in: 2016 International Joint Con-
cmdi-profiles for textbook resources, in: Research ference on Neural Networks (IJCNN), IEEE, 2016,
Conference on Metadata and Semantics Research, pp. 558–565.
Springer, 2018, pp. 302–314. [25] H. M. Kamar, R. Ahmad, N. Kamsah, A. F. M.
[13] A. A. Jaber, A. A. Saleh, H. F. M. Ali, Prediction of Mustafa, Artificial neural networks for automotive
hourly cooling energy consumption of educational air-conditioning systems performance prediction,
buildings using artificial neural network, Space Applied Thermal Engineering 50 (2013) 63–70.
10137 (2019) m3. [26] J. M. Belman-Flores, S. Ledesma, M. Garcia, J. Ruiz,
[14] S. Russo, C. Napoli, A comprehensive solution for J. L. Rodríguez-Muñoz, Analysis of a variable speed
psychological treatment and therapeutic path plan- vapor compression system using artificial neural
97
Iyad Lafta Majid et al. CEUR Workshop Proceedings 90–98
networks, Expert systems with applications 40 vehicle air conditioning system performance pre-
(2013) 4362–4369. diction based on artificial neural network, Applied
[27] H. M. Ertunc, M. Hosoz, Artificial neural network Thermal Engineering 89 (2015) 101–114.
analysis of a refrigeration system with an evapora-
tive condenser, Applied Thermal Engineering 26
(2006) 627–635.
[28] J. Gill, J. Singh, Use of artificial neural network
approach for depicting mass flow rate of r134a/lpg
refrigerant through straight and helical coiled adia-
batic capillary tubes of vapor compression refriger-
ation system, International Journal of Refrigeration
86 (2018) 228–238.
[29] Ö. Kizilkan, Thermodynamic analysis of variable
speed refrigeration system using artificial neural
networks, Expert Systems with Applications 38
(2011) 11686–11692.
[30] M. V. Reddy, C. C. Sekhar, D. R. Reddy, Performance
prediction of domestic refrigeration system using
r436a refrigerant as alternative refrigerant to r134a
with ann, International Journal of Innovative Re-
search in Science, Engineering and Technology 6
(2017) 19008–19017.
[31] S. Yilmaz, K. Atik, Modeling of a mechanical cool-
ing system with variable cooling capacity by using
artificial neural network, Applied Thermal Engi-
neering 27 (2007) 2308–2313.
[32] D. Swider, M. Browne, P. Bansal, V. Kecman, Mod-
elling of vapour-compression liquid chillers with
neural networks, Applied thermal engineering 21
(2001) 311–329.
[33] L. Zhao, W.-J. Cai, Z.-H. Man, Neural modeling of
vapor compression refrigeration cycle with extreme
learning machine, Neurocomputing 128 (2014) 242–
248.
[34] M. Hosoz, H. M. Ertunc, H. Bulgurcu, An adaptive
neuro-fuzzy inference system model for predicting
the performance of a refrigeration system with a
cooling tower, Expert Systems with Applications
38 (2011) 14148–14155.
[35] M. Hosoz, M. S. Kaplan Kaplan, M. C. Aral, H. M.
Ertunc, Artificial neural network modelling of a
mobile air conditioning system using refrigerant
r1234yf, in: 5th International Symposium on Inno-
vative Technologies in Engineering and Science 29-
30 September 2017 (ISITES2017 Baku-Azerbaijan),
2017.
[36] M. Hosoz, H. Ertunc, Modelling of a cascade re-
frigeration system using artificial neural network,
International Journal of Energy Research 30 (2006)
1200–1215.
[37] Z. Tian, B. Gu, L. Yang, F. Liu, Performance predic-
tion for a parallel flow condenser based on artificial
neural network, Applied Thermal Engineering 63
(2014) 459–467.
[38] Z. Tian, C. Qian, B. Gu, L. Yang, F. Liu, Electric
98