Energy 55 (2013) 392e402

Contents lists available at SciVerse ScienceDirect

Energy
journal homepage: www.elsevier.com/locate/energy

Model-based optimization for vapor compression refrigeration cycle

Lei Zhao, Wenjian Cai*, Xudong Ding, Weichung Chang
School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore 639798, Singapore

a r t i c l e i n f o

a b s t r a c t

Article history:
Received 26 September 2012
Received in revised form
25 February 2013
Accepted 26 February 2013
Available online 26 April 2013

This paper presents a model-based optimization strategy for vapor compression refrigeration cycle.
Through analyzing each component characteristics and interactions within the cycle, the optimization
problem is formulated as minimizing the total operating cost of the energy consuming devices subject to
the constraints of mechanical limitations, component interactions, environment conditions and cooling
load demands. A MGA (modied genetic algorithm) together with a solution strategy for a group of
nonlinear equations is proposed to obtain optimal set point under different operating conditions.
Simulation studies are conducted to compare the proposed method with traditional oneoff control
strategy to evaluate its performance. Experiment results of a real practical system are also presented to
demonstrate its feasibility.
2013 Elsevier Ltd. All rights reserved.

Keywords:
Vapor compression refrigeration cycle
Hybrid components models
Global optimization
Modied genetic algorithm
System simulation and testing

1. Introduction
Air conditioning and refrigeration systems, as common residential and industry devices, are widely used to transfer heat between different locations [1]. They control space temperatures for
comfort of human life, food storage and transportation, etc., and
consume a lot of energy for both developing and developed
countries. For instance, statistical data shows air conditioner and
refrigerator account for 28% of home energy consumption in US [2].
For hot and humid tropical country such as Singapore, this ratio can
even rise to over 50% [3]. Among all types of cooling systems,
electricity-based vapor compression cooling systems are still
dominant in the current market. Since the vapor compression
refrigeration cycle is the heart of and consumes most of the energy
in any cooling system, the effort to reduce the energy consumption
through system control and optimization in vapor compression
cooling system is of practical signicance due to both energy
shortage and global warming concerns.
All optimization techniques of vapor compression refrigeration
cycle involve components and system modeling. Even though there
are some discussions on the modeling of compressors, expansion
valves and overall systems, the main works in this area is still
focused on the development of heat exchanger models. So far,
several models have developed including: distributed model, NTU

* Corresponding author. Tel.: 65 6790 6862; fax: 65 6793 3318.

E-mail address: [email protected] (W. Cai).
0360-5442/$ e see front matter 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.energy.2013.02.071

(Number of Transfer Units) model, black box model and hybrid

model. In the distributed model, the heat exchanger is divided into
consecutive segments, the inlet of each segment equals to the
outlet of its former adjacent segment. The nal static state is
determined by recursive calculation. Jia et al. built a distributed
model for dry expansion evaporators to analyze their steady and
dynamic characteristics [4]. To accurately design plate-n heat
exchanger, Zhang et al. developed a 3D distributed parameter
model [5]. In 3-NTU model, the static states of heat exchanger are
determined by a dimensional parameter called the heat transfer
effectiveness 3 . Ren presented an accurate model of counterow
cooling tower with modied 3-NTU method [6]. Arsenyeva et al.
used 3-NTU model of plate and frame heat exchanger in design
procedure to improve heat recovery [7]. The black box model uses a
predetermined function format as its heat transfer function, the
coefcients are determined by curve tting or identication techniques according to experiment data. Ertunc and Hosoz proposed
an adaptive neuro-fuzzy inference system to predict the heat
rejection rate and the temperatures of refrigerant and leaving air of
evaporative condensers. Simulation results showed that the prediction errors are within 5% [8]. Kim et al. derived a polynomial
equation with seven variables to predict the effectiveness of a pilot
dry coil indirect evaporative cooler based on experiment data [9].
The hybrid model lumps the physical information as constants,
calculate the model coefcients through least-square methods
based on experiment data, therefore it takes advantages of both
physical and empirical models, which makes it suitable for online
optimization with high accuracy. Wang et al. derived the hybrid

L. Zhao et al. / Energy 55 (2013) 392e402

Nomenclature
Av
c
f
F
K
H
Hc,fg
Hc,r,i
Hc,r,o
He,fg
He,r,i
He,r,o
_ c;air
m
_ c;air;nom
m
_ e;air
m
_ e;air;nom
m
_r
m
_ r;max
m
_ r;min
m
n
Pc
Pc,max
Pc,min
Pe
Pe,max
Pe,min
Q_ c
Q_ e
Q_ com
Q_
req

Tc,air,i
Tc,max

opening percentage of electronic expansion valve

constants of hybrid model
function
frequency of fans (Hz)
penalty
enthalpy (kJ/kg)
enthalpy difference of gas and liquid saturated
refrigerant in condenser (kJ/kg)
condenser inlet refrigerant enthalpy (kJ/kg)
condenser outlet refrigerant enthalpy (kJ/kg)
enthalpy difference of gas and liquid saturated
refrigerant in evaporator (kJ/kg)
evaporator inlet refrigerant enthalpy (kJ/kg)
evaporator outlet refrigerant enthalpy (kJ/kg)
air ow rate of condenser (kg/s)
nominal air ow rate of condenser (kg/s)
air ow rate of evaporator (kg/s)
nominal air ow rate of evaporator (kg/s)
refrigerant mass ow rate (kg/s)
maximal refrigerant mass ow rate (kg/s)
minimal refrigerant mass ow rate (kg/s)
polytropic exponent
condenser refrigerant saturated pressure (bar)
maximal condenser saturated pressure allowed (bar)
minimal condenser saturated pressure allowed (bar)
evaporator refrigerant saturated pressure (bar)
maximal evaporator saturated pressure allowed (bar)
minimal evaporator saturated pressure allowed (bar)
condenser energy exchange rate (kJ/kg)
evaporator energy exchanger rate (kJ/kg)
mechanical work of compressor (kJ/kg)
required cooling load (kJ/kg)
condenser inlet air temperature ( C)
maximal refrigerant saturation temperature in
condenser ( C)

model of cooling coil based on the ow characteristics and

nonlinear least-square method [10]. Considering the refrigerant
phase change in heat exchanger, Ding et al. developed the hybrid
model for evaporator with three parameters [11], they further
applied the hybrid method to analyze condenser that contains gas,
liquid and two phase refrigerant, derived the condenser hybrid
model including four parameters [12].
Aiming at ES (energy saving) vapor compression refrigeration
cycles, many research works have been done on improving the
system COP (coefcient of performance). Stoecker claimed to achieve higher COP, both superheat and subcool should be decreased to
near zero, this setting will increase the efciency of heat exchangers [13]. Jensen and Skogestad proposed to add the refrigerant volume (active charge) in the system as an additional control
variable, they also discussed the effect of subcooling on system
efciency [14]. To realize self-optimizing control, they discussed
the selection of the controlled variable to improve the system efciency and performance of control method [15]. In Ref. [16], they
focused on the available degrees of freedom since it is the basis of
optimal operation and plant wide control. Larsen et al. proposed a
gradient method to nd the suboptimal solution for condensing
pressure, while keeping the superheat and evaporating pressure
constant [17]. Selbas et al. optimized the components of subcooled

393

Tc,min

minimum refrigerant saturation temperature in

condenser ( C)
Tc,r,i
condenser inlet refrigerant temperature ( C)
Tc,sc
condenser subcool temperature ( C)
Tc,r,sat
condenser refrigerant saturated temperature ( C)
Tc,r,i
condenser refrigerant inlet temperature ( C)
Tc,r,o
condenser refrigerant outlet temperature ( C)
Te,r,sat
refrigerant saturated temperature in evaporator ( C)
Te,max
maximal refrigerant saturation temperature in
evaporator ( C)
Te,min
minimal refrigerant saturation temperature in
evaporator ( C)
Te,sh
evaporator superheat temperature ( C)
Va
volume at bottom dead center (m3)
Vd
volume when suction valve open (m3)
_
W c;fan
condenser fan power (kW)
_
W
c;fan;nom condenser fan power when air ow rate is mc,air,nom
(kW)
_ com
electricity power consumption of compressor (kW)
W
_
W
evaporator fan power (kW)
c;fan
_
W
e;fan;nom evaporator fan power when air ow rate is me,air,nom
(kW)
_
total power (kW)
W
total
h
enthalpy delivery efciency
u
compressor rotation speed (r/s)
r
inlet refrigerant density (kg/m3)
Subscripts
air
feature of air
c
condenser
com
compressor
e
evaporator
ev
expansion valve
i
inlet
o
outlet
m
mass ow rate
q
heat exchange
h
enthalpy delivery efciency

and superheated vapor compression refrigeration cycle separately

based on exergy analysis [18]. The entransy dissipation-based
optimization technique was introduced to nd the optimal design
and to minimize the operation cost of evaporative cooling system
by Yuan and Chen [19].
Since system states of vapor compression refrigeration cycle are
severely interacted, it is highly possible that solutions of classical
methods are trapped in local minimum instead of global minimum.
On the other hand, the newly developed articial intelligence
computation techniques; have been successfully used in optimization of air conditioning and HVAC (heating, ventilation, and air
conditioning) systems. Fong et al. employed differential evolution
to optimize solar thermal refrigeration system [20]. Kusiak et al.
proposed to optimize the commercial HVAC system with evolutionary computation algorithm [21]. They also used multi-objective
particle swarm algorithm to nd the optimal solution that can
balance energy consumption and thermal comfort [22]. Kusiak and
Xu then used dynamic neural network to construct black box
models for system performance prediction with controllable and
uncontrollable inputs and outputs [23]. Simulation studies based
on experiment conducted at the ERS (Energy Resource Station)
showed that system level optimization that they proposed can
improve overall system operating performance signicantly.

394

L. Zhao et al. / Energy 55 (2013) 392e402

Furthermore, Hovgaard et al. developed a novel economic model

predictive control scheme to reduce the operating cost of supermarket refrigeration system [24].
In this paper, we propose a modied genetic optimization
technique to tackle the optimization problem of vapor compression
refrigeration cycle. The objective function is formulated based on
mathematical models of the major components. The thermodynamic characteristics of the heat exchangers are predicted with the
hybrid models, which are simple yet accurate with some physical
signicance. The power consumptions of the power consuming
equipments are predicted also using the hybrid models. A MGA
(modied genetic algorithm) is developed to search the optimal
settings. Simulation and experimental results on a lab scale pilot
plant demonstrate that a signicant operating cost can be saved by
the proposed method.
2. Working principle and component models
The vapor compression refrigeration cycle is used to remove
heat from a space of lower temperature (cold reservoir) to an
environment of high temperature (hot reservoir). As shown schematically in Fig. 1, it basically consists of four components e
Evaporator, Compressor, Condenser, and Expansion Valve which are
connected in a closed loop so that the refrigerant is continuously
circulated. A peh chart reecting the variation of the refrigerant
states is shown as in Fig. 2 which is briey explained as below.
1. Refrigerant starts from state 1, inlet of compressor, in gas phase
at low pressure and low temperature. As the refrigerant exits

the compressor at state 2, it is still in gas phase but with high

pressure (saturation pressure) and high temperature. The superheated vapor refrigerant enters the condenser at state 2,
where the heat from the working uid is transferred to secondary uid as the secondary uid has relatively lower temperature. Inside the condenser, the refrigerant is in two phases
(liquid mixed with gas). When the refrigerant exits the
condenser after transferring the heat to the secondary uid, the
temperature is below the condensation temperature to ensure
the fully liquidated refrigerant.
2. The refrigerant in liquid phase enters the expansion valve at the
state 3 where the pressure of the uid is decreased through the
throttling effect of the expansion valve. The sudden decrease in
the pressure of the refrigerant causes it to partially evaporate
which in turn reduces the saturation temperature. Thus, when
the refrigerant exits the valve it is in low pressure, low temperature and two phases (liquid mixed with gas).
3. At state 4, the refrigerant enters the evaporator at low temperature and low pressure condition. Since the refrigerant
temperature is lower than the temperature of cold reservoir,
the refrigerant completely evaporates at a constant temperature with the help of evaporator fan to increase the heat
transfer rate. At the exit of the evaporator, the temperature is
slightly above the evaporating temperature for the safe operation of compressor.
The mathematical models of the four components which realize
the transition of the refrigerant between the states and are
instrumental in the model-based optimization are given below.

Fig. 1. Vapor compression refrigeration cycle.

L. Zhao et al. / Energy 55 (2013) 392e402

395

2.3. Compressor
In a compressor, the mass ow rate and the refrigerant energy
change during compression stroke can be expressed [25,26]:

 ccom;m;3 
Pc
u
Pe

(5)

 n1

Pc n
_r
 1 um
Pe

(6)


_r
m

ccom;m;1  ccom;m;2

and

Q_ com

Fig. 2. peh chart of vapor compression cycle.

2.1. Evaporator
Based on mass and energy balance, a simple hybrid model to
describe the heat transfer properties in evaporator is given as [11]





_ r ce;1 m
_ cr e;3 Te;air;i  Te;r;sat
He;g  He;r;i m
Q_ e
!ce;3
_r
m
1 ce;2
_ e;air
m

(1)

where ccom,m,1, ccom,m,2 and ccom,m,3 are constants determined by

curve tting, Pc, Pe, u, n, Va, Vd, and Q_ com are condensing temperature, evaporating pressure, compressor rotation speed, polytropic
exponent, volume at bottom dead center, volume when suction
valve open and mechanical work input to refrigerant by compressor
respectively. Since the parameters n, Va and Vd are constants for a
given compressor and at a specied working environment, Eq. (6)
can be further simplied to a hybrid model form:

_ r Pe
Q_ com ccom;q;1 um


 ccom;q;2
Pc
1
Pe

(7)

where ccom;q;1 n=n  1Va  Vd and ccom;q;2 n  1=n.

where ce,1, ce,2 and ce,3 are constants obtained by tting experiment
_ r, m
_ e;air , Te,air,i, Te,r,sat and Q_ e are the enthalpy of
data, He,g, He,r,i, m
saturated gas phase refrigerant in evaporator, refrigerant enthalpy
of evaporator inlet, mass ow rates of refrigerant, air outside
evaporator, temperature of inlet air and saturated refrigerant of
evaporator, heat exchanging rate of evaporator, respectively (see
Appendix B for the calculations of He,g, He,r,i and Te,r,sat).
In addition, an energy balance equation, i.e., energy absorbed by
the refrigerant is equal to energy reduced in the cold reservoir, i.e.


_ r He;r;o  He;r;i
Q_ e m

n
Va  Vd Pe
n1

(2)

will also be used in later development, where He,r,o is refrigerant

enthalpy at evaporator outlet (see Appendix B for the calculation of
He,r,o).

2.4. Expansion valve

The mass ow rate of expansion valve is determined by valve
opening percentage, pressure difference and inlet refrigerant density. Its mass ow rate is given by [25]

p
_ r cev;1 cev;2 Av
rPc  Pe
m

(8)

where cev,1 and cev,2 are constants, Av and r are opening percentage
of electronic expansion valve and density of inlet refrigerant
respectively (see Appendix B for the calculation of r). Since the
expansion process is isenthalpic, it is assumed that refrigerant
enthalpy is constant which implies Qev 0.
2.5. Energy balance of overall cycle

2.2. Condenser
Similar to the evaporator, condenser can also be represented by
a hybrid model [12]:





_ cr c;4 Tc;r;sat  Tc;air;i cc;2 m
_ r Tc;r;i  Tc;r;sat Hc;fg m
_r
cc;1 m
_
Qc
!cc;4
_r
m
1 cc;3
_ c;air
m
(3)
where cc,1, cc,2, cc,3 and cc,4 are constants calculated by tting
experiment data, Hc,fg is enthalpy difference between saturated
_ c;air is the mass ow
liquid and gas phase refrigerant in condenser, m
rate of air outside condenser. Tc,r,sat, Tc,r,i, Tc,air,i are temperature of
saturated refrigerant, inlet refrigerant and inlet air of condenser, Q_ c
is the heat transferring rate of condenser (see Appendix B for the
calculations of Hc,fg and Tc,r,sat).
The corresponding energy balance equation is


_ r Hc;r;i  Hc;r;o
Q_ c m

(4)

where Hc,r,i and Hc,r,o are enthalpy of inlet and outlet refrigerant (see
Appendix B for the calculations of Hc,r,i, and Hc,r,o).

Finally, the corresponding energy balance equation for the

overall system is given as

Q_ c Q_ com Q_ e

(9)

3. Optimization problem formulation

The objective of global optimization for vapor compression
refrigeration cycle is to satisfy cooling load requirement of cold
reservoir with minimal energy consumption. Mathematically, it can
be formulated as

_
_
_
_
Min W
total W com W c;fan W e;fan
_
_
Subject to Q e Q req

(10)

_
_
_
_
_
where W
total ; W com ; W c;fan ; W e;fan and Q req are total power consumption, power consumption of compressor, condenser fan power
consumption, evaporator fan power consumption and required
cooling load respectively. The power consumption models of the
compressor, condenser fan and evaporator fan are formulated according to their working principles.

396

L. Zhao et al. / Energy 55 (2013) 392e402

3. Condensing and evaporating pressure

3.1. Power consumption of compressor

Using a hybrid model to describe the delivery coefcient of
compressor, hcom [26]

Pc;min  Pc  Pc;max

(16a)

hcom ccom;h;1 ccom;h;2 Pc =Pe ccom;h;3

Pe;min  Pe  Pe;max

(16b)

(11)

the power consumption of compressor is given as

_ com
W

where Pc,min, Pc,max, Pe,min, and Pe,max are minimal and maximal
condensing pressure, minimal and maximal evaporating pressure.

Q_ com

(12)

hcom

4. Superheat

where ccom;h;1 ; ccom;h;2 and ccom;h;3 are constants calculated by catalog or experiment data.
3.2. Power consumptions of condenser fan and evaporator fan
The power consumptions of fans are inuenced by two parameters: mass ow rates of uids and the pressure difference
between the inlets and outlets and can be described [27]:

_
_
W
c;fan W c;fan;nom cc;fan;0 cc;fan;1
cc;fan;2

_ c;air
m
_ c;air;nom
m

_ c;air
m
_ c;air;nom
m

!2
cc;fan;3

7 + C  Te;sh  25 + C

(17)

Superheat temperature too low will cause oscillation in the cycle

[28], while the COP of system will decrease steeply if the superheat
temperature is too high [13]. The upper and lower limits depend on
the system congurations.
5. Subcool

!
Tc;sc  0

_ c;air
m
_ c;air;nom
m

!3 !
(13a)

(18)

Subcool only has lower bound to ensure the refrigerant at

expansion valve inlet is at liquid phase for ashing.
6. Compressor frequency (for the testing system)

_
_
W
e;fan W e;fan;nom ce;fan;0 ce;fan;1
ce;fan;2

_ e;air
m
_
me;air;nom

_ e;air
m
_ e;air;nom
m

!2
ce;fan;3

_ e;air
m
_
me;air;nom

30 Hz  F  50 Hz
!3 !
(13b)

(19)

This physical constraint is directly determined by compressor

working range.
7. Mass ow rate of refrigerant

where cc,fan,0 to cc,fan,3 and ce,fan,0 to ce,fan,3 are coefcients deter_

_
_
mined by catalog or experiment data, W
c;fan ; W e;fan ; W c;fan;nom ;
_
_
_
W e;fan;nom ; mc;air;nom and me;air;nom are measured power consumptions of condenser fan, measured power consumption of evaporator fan, nominal power consumption of condenser fan, nominal
power consumption of evaporator fan, nominal mass ow rate of
condenser fan and mass ow rate of evaporator fan respectively.
In the optimization problem, there are two types of constraints,
i.e., interaction between the components represented by Eqs. (1)e
(5) and (7)e(9), and physical limitations of the components and
uids, including:
1. Air mass ow rates of condenser and evaporator fans

_r m
_ r;max
_ r;min  m
m

(20)

_ r;min and m
_ r;max are minimal and maximal refrigerant
where m
mass ow rate. The upper bound of mass ow rate of refrigerant is
determined by expansion valve and compressor capacities, the
lower bound is determined by ow meter working range.
8. EEV opening percentage

0 < Av  1

(21)

To reduce the variables and simplify the solution procedures, we

notice that

_ c;air  m
_ c;air;max
_ c;air;min  m
m

(14a)

_ e;air  m
_ e;air;max
_ e;air;min  m
m

(14b)

_ c;air;min ; m
_ c;air;max ; m
_ e;air;min and m
_ e;air;max are minimal and
where m
maximal condenser air mass ow rates, minimal and maximal
evaporator air mass ow rates respectively.
2. Condensing and evaporating temperatures

Tc;air;i  Tc;r;o  Tc;min  Tc;r;sat  Tc;max  Tc;r;i

(15a)

Te;r;i  Te;min  Te;r;sat  Te;max  Te;air;i

(15b)

where Tc,min, Tc,max, Te,min and Te,max are minimal and maximal
condensing temperatures, minimal and maximal evaporating
temperatures.

 Pc, Hc,fg and Tc,r,sat

 Pe, He,fg and Te,r,sat
 Hc,r,o and He,r,i
have same physical properties, therefore, we can use Pc; Pe; and
Hc,r,o to replace Tc,r,sat and Hc,fg; Te,r,sat and He,fg; He,r,i, in the system
objective, constraints and power consumption equations. All variables used in the optimization problem, according to their rules,
can be further classied into three groups:
 Uncontrollable variables (Q_ e , Tc,air,i, Te,air,i): They are determined
by user demand or environmental conditions, these variables
are xed during each calculation step. Among these three
variables, the outdoor air temperature, Tc,air,i, is the most critical
data, the other two uncontrollable variables, Te,air,i and Q_ e , are
also affected by Tc,air,i.

L. Zhao et al. / Energy 55 (2013) 392e402

 Independent variables (Pc, Pe, u, Tc,r,o): Independent variables

can be selected randomly within the constraint limits to meet
the optimization objective. The obvious choices for indepen_ c;air ; m
_ e;air as they are directly
dent variables are u; Av ; m
controlled by compressor, EEV (electronic expansion valve) and
fan frequencies. However, Pe, Pc, u and Tc,r,o are assigned to be
independent variables in this work in order to reduce the
complexity of solving process as well as to increase solution
accuracy.
_ r; T ; m
_
_
 Dependent variables (m
; Te;r;o ; m
; Q_ ; Q_
c;r;i

c;air

e;air

com

and Av): These values can be resolved through the interaction

and physical constraint equations for the given independent
variables. These variables can then be used to calculate the
power consumption equations.

397

a constant which can be specied by user, the rule of thumb of the

selection is 3 times of the rated compressor power. By calculating
Eq. (21), the tness value of each individual chromosome is converted into predetermined interval to control the evolution speed
in an acceptable range.
4.3. Evolution
In evolution, four major evolutionary operations: selection,
crossover, mutation and reinsertion are performed. The elitist
roulette wheel selection [29], single point crossover and single bit
mutation operators are employed. The crossover and mutation
points are selected randomly in each generation.
4.4. Termination

4. Solution procedures
To solve the constrained nonlinear optimization problem, a
modied genetic algorithm (MGA) combined with a solution
strategy for a group of nonlinear equations is devised for higher
efciency and less computation complexity. This algorithm is
divided into four parts: encoding, construction of tness function,
evolution and termination [29], which are briey discussed below.

The genetic algorithm terminates if one of the following criterion is satised:

1. Predetermined maximal generation number is reached.

4.1. Encoding

start

Each independent variable in constraint Eqs. (15a), (16a), (16b)

and (19) is converted into a binary string. All of these strings are
combined together to form a chromosome. Gray coding is preferred
for overcoming the hidden representative bias in binary representation. The lengths of the binary strings are determined by the
control precision of the corresponding variables: the more precise
set point control requires the longer binary string. For instance, if
the compressor controllers accuracy is 102, then 10 bits is enough
for u to describe the compressor setting. After encoding, a random
of population of chromosomes, is generated such that, when
decoded to real values, the variable values lie within their limitations. The whole generation should possibly try to include the
entire range of values.

determine component model parameters

measure air temperature and calculate cooling load

reduce variables by practical consideration

initialize parameters of modified genetic algorithm

gen>max gen?
Y

4.2. Constructing tness function

After setting the random variables, eight dependent variables

are solved according to interaction constraint equations based on
the coded independent variables (see Appendix B). Since the
_
objective of this optimization problem is to minimize W
total , the
tness function should be constructed to give the maximum value
_
to the smallest W
total . In addition, the tness function should reect
whether the chromosome violates the constraints. Based on these
considerations, the tness function is constructed as follows:

fitness

1
_
_
_ com W
W
e;fan W c;fan Ktotal

gen=gen+1

calculate dependent variables

calcuate fitness

evolution and store the best one

(21)
termination criterion satisfied?

where the denition of total system penalty function Ktotal is as

following:

Ktotal

8
>
>
>
<

8
>
P
>
>
Ki
: cp
i1

if
if

8
P
i1
8
P

Y
find indiviual with largest fitness among all generations

Ki 0

i1

(22)

calculate power consumption of optimal set point

Ki s0
end

where Ki, i 1,.,8 are the individual state penalty functions (see
Appendices A and B for the detailed denition and calculation), cp is

Fig. 3. Procedure owchart.

398

L. Zhao et al. / Energy 55 (2013) 392e402

Table 1
Physical limits of variables.

1.5

Variables

Lower bound

Upper bound

Fc,fan
Fe,fan
Tc,r,sat
Te,r,sat
Pc
Pe
Te,sh
Tc,sc
Fcom
_r
m

15 Hz
15 Hz
Tc,air,i
12  C
8 bar
2 bar
5 C
0 C
30 Hz
0.008 kg/s

35 Hz
30 Hz
Tc,r,i
Te,air,i
15 bar
5 bar
25  C
NA
50 Hz
0.03 kg/s

MGA
on-off control

1.45

Energy consumption(kW)

1.4
1.35
1.3
1.25
1.2
1.15

33

1.1
1.05

32

Temperature(C)

12

14
Time(h)

16

18

20

Fig. 6. Energy consumptions for working hours.

31

30

4.5. Remark

29

28

27

10

12

16

20

24
28
Time(h)

32

36

40

44

48

Fig. 4. Measured temperatures for two consecutive days.

2. The optimal tness of group does not change much over several
steps, the bound for determining whether its value changes too
much is predetermined according to experience.
3. All chromosomes converge to nearly the same tness groups.
Compared with the traditional genetic algorithm, there is a
difference of the proposed algorithm, particularly modied for such
kind of problems, which is discussed in the following remark.

A main drawback of traditional genetic algorithm when applied

to vapor compression refrigeration cycle is that when Q_ e is
approaching to system cooling capacity, it is highly possible that all
individuals in last generation do not satisfy the physical constraints, even though several individuals in previous generation do,
the good chromosome might lost after evolution. The proposed
algorithm is to store the individuals that satisfy the physical constraints with maximum tness value in each generation step when
Q_ e is large. As a result, only chromosome with maximal tness is
stored for each generation and to compare with the nal generation, this modication increases the possibility of nding a feasible
solution.
The owchart of MGA for optimization of vapor compressor
cycle is illustrated in Fig. 3 which can be summarized as follows:
Step 1: Determine the parameters of the component models by
catalog or experimental data.
Step 2: Measure Te,air,i and Tc,air,i, and calculate Q_ e .

25
4.6
4.4

20
Percentage of time interval(%)

4.2

Cooling load(kW)

4
3.8
3.6
3.4
3.2

15

10

3
2.8
2.6

12

16

20

24
28
Time(h)

32

36

40

Fig. 5. Assumed cooling load proles in simulation.

44

48

0
29.5

30

30.5

31
31.5
Temperature(C)

32

Fig. 7. Bar chart of one day temperature prole.

32.5

33

L. Zhao et al. / Energy 55 (2013) 392e402

399

3.5
Energy consumptions:MGA
Energy savings:MGA

Energy consumption(kWh)

2.5

1.5

0.5

0
29.5

30

30.5

31
31.5
Temperature(C)

32

32.5

33

Fig. 8. System energy consumptions and savings.

Step 3: Initialize the parameters of MGA: population size,

maximum number of generations, precisions of variables, generation gap and probabilities of crossover and
mutation.
Step 4: Perform the evolutionary operations.
Step 5: Repeat Step 5 and store chromosome according to Remark
until termination criterion is satised.
Step 6: Find the chromosome with largest tness value among
all generations and calculate corresponding power
consumption.
After calculating the solution, control system will set the frequencies of compressor, evaporator fan, condenser fan and expansion valve opening according to the solution, resulting in all
operating states equal or close to their corresponding optimal value.

Fig. 10. Picture of the lab scale vapor compression refrigeration system.

Considering the large inuences of evaporating and condensing

temperatures on operating states and system energy consumption,
the fans will be used to slightly adjust them to decrease the error
caused by model inaccuracy. In practice, the precisions of the state
variables in optimization algorithm are usually higher than that of
fans and compressors controllers, the closest values that controllers
can achieve will be chosen as the set point. This approach can
minimize the difference between desired and practical output.

Fig. 9. Schematic of the vapor compression refrigeration system.

400

L. Zhao et al. / Energy 55 (2013) 392e402

schemes have the same trends, but MGA is more sensitive to the
change of cooling load. The largest and smallest energy saving of
MGA compared with oneoff control scheme is 1.44% and 14.78%
respectively.
To analyze the energy consumptions at different temperature
conditions, the measured temperatures are regrouped according to
the temperature range. The percentage of temperature in the
measured day from 8am to 19pm is shown as in Fig. 7.
To estimate the energy saving effect of MGA, the total energy
consumptions for both MGA and oneoff control schemes of the
temperature prole in Fig. 7 are calculated and presented in Fig. 8,
which shows that the average energy saving of 8.52% is achieved.

1.5
MGA test
on-off test
MGA sim
on-off sim

1.45

Energy consumption(kW)

1.4
1.35
1.3
1.25
1.2
1.15

6. Experiment

1.1
1.05

10

12

14
Time(h)

16

18

20

Fig. 11. Energy consumptions.

5. Simulation
The simulation studies are based on the models of a lab scale
pilot plant. The coefcients of models are determined through
catalog tting 200 groups of experiment data which are obtained
by different compressor, condenser fan and evaporator fan frequencies. All constraints also follow the system physical limits listed as in Table 1.
To simulate the energy saving effect in practical situations,
environmental temperatures of two consecutive days are measured
as shown in Fig. 4.
Assuming the set point of indoor temperature is 23  C and the
highest cooling load is equal to the maximum cooling capacity of
the pilot plant, the measured temperature data can then be proportionally converted to the cooling load shown as in Fig. 5.
The simulation results for energy consumption of the proposed
MGA and traditional oneoff control scheme from 8am to 19pm are
illustrated in Fig. 6. It can be seen that energy consumptions of MGA
are lower than that of oneoff control scheme for all cooling loads,
but the difference is smaller when the cooling load is near its
maximum cooling capacity. The energy consumptions of the two

To verify the effect of the proposed MGA, experimental tests are

conducted on a lab scale multi-evaporator vapor compression cycle.
The schematic and photo of the experimental system are illustrated
in Figs. 9 and 10, respectively.
The test bench consists of a single-stage vapor compression
formed by basic components: variable speed compressor, n-tube
condenser, thermostatic expansion valve and isolated n-tube
evaporator. The heat exchangers work with secondary uids
loops, which represents and simulates environmental and thermal
load.
The experiment system includes a semi-hermetic reciprocating
compressor, an air-cooled nned-tube condenser, three electronic
expansion valves and three evaporators (one air-cooled nnedtube evaporator and two electronic evaporators). One air duct
heater controls the inlet air temperature of condenser for simulating outdoor condition, and the inlet air temperature of evaporator is constantly kept as 25  C by HVAC system. The working uid
used for the system is R134a. The compressor, the condenser fan
and the evaporator fan are equipped with inverters to adjust their
corresponding frequencies. An air duct heater is installed in front of
the condenser to control the temperature of condenser inlet air. In
addition, several temperature and pressure sensors are installed for
detecting state variables. To measure the mass ow rate, a ow
meter with maximum 4% full scale error is used. The measurement
range of the pressure transducers and the temperature transmitters
are 0e16 bar and 40  Ce200  C, with their maximum full scale

Fig. 12. System energy consumptions and savings.

L. Zhao et al. / Energy 55 (2013) 392e402

error are within 0.3% and 0.3  C respectively. The positions of

these sensors are as following (refer to Fig. 9):
 refrigerant temperature and pressure at compressor outlet (T1
and P1);
 temperature and pressure at condenser inlet (T2 and P2);
 temperature and pressure at condenser outlet (T3 and P3);
 temperature and pressure at EEV inlet (T5 and P5);
 temperature and pressure at evaporator outlet (T6 and P6);
_ r ).
 mass ow rate after receiver (m
 humidities at evaporator inlet and outlet (R.H.1, R.H.2).
To simplify the test and calculation, the average temperatures
and cooling load demands of each hour from 8am to 19pm are used
during testing. The energy consumptions of both MGA and oneoff
control schemes for results of both simulation and testing are
compared in Fig. 11. It shows that EC (energy consumptions) are
higher in the experiment results than that of simulation for nearly
all testing groups, the average and largest difference between
experiment and simulation results are 3.43% and 10.47% respectively, which might be caused by the modeling errors. However, the
results are accurate enough for the MGA to be implemented in
practical applications. Moreover, the difference between experimental results of MGA and oneoff control scheme veries the energy saving effect of proposed method. The smallest and largest
differences between these two groups of data are 0.03% and 13.14%
respectively with the average overall energy saving (ES) of 8.45% as
shown in Fig. 12.
7. Conclusion
In this paper, a model-based optimization strategy for vapor
compression refrigeration cycle was developed, which was formulated as minimizing the total operating cost of the energy
consuming devices subject to the constraints of mechanical limitations, component interactions, environment conditions and cooling
load demands. A modied genetic algorithm together with a solution strategy for a group of nonlinear equations was proposed to
effectively obtain the optimal set point under different operating
conditions. The experimental results showed that the set points
calculated by MGA can reduce energy consumption compared with
traditional oneoff control, the overall energy saving for a typical day
during 8ame19pm is 8.45%. Furthermore, the average difference
between simulated and experimental results for energy consumptions is 3.43%, which demonstrated the effectiveness of the proposed method. An interesting topic along this direction is to study
the quantitative effect of each independent variable to the overall
system performance, therefore to reduce the search space, which is
currently under study and the ndings will be reported later.
Acknowledgements
The work was funded by National Research Foundation of
Singapore: NRF2008EWT-CERP002-010. The other project partners
are also acknowledged.

401

Evaporator air mass ow rate penalty:

8
_
_
_
0
< 

 if me;air;min  me;air  me;air;max

_ e;air;min  m
_ e;air;min > m
_ e;air m
_ e;air
_ e;air 
if
m
K2 
m


:
_ e;air;max  m
_ e;air > m
_ e;air m
_ e;air;max
_ e;air 
m
if m
(A2)
Condensing saturated temperature penalty:

K3

0




 Tc;r;i  Tc;r;sat Tc;r;sat 

if Tc;r;o  Tc;r;sat  Tc;r;i

if Tc;r;i < Tc;r;sat
(A3)

Evaporating saturated temperature penalty:

K4

if Te;r;sat  Te;air;i
if Te;r;sat > Te;air;i

0




 Te;r;sat  Te;air;i Te;r;sat 

Superheat penalty:

8
0 
< 

 if Te;r;sat 7  Te;r;o  Te;r;sat 25
if Te;r;o < Te;r;sat 7
K5 
Te;r;o  Te;r;sat Te;r;o 
:
Te;r;o  Te;r;sat Te;r;sat 
if Te;r;o > Te;r;sat 25
(A5)
Subcool penalty:

K6

0 



 Tc;r;o  Tc;r;sat Tc;r;sat 

if Tc;r;o  Tc;r;sat
if Tc;r;o > Tc;r;sat

(A6)

Refrigerant mass ow rate penalty:

8
0
< 

 
 m
_ r;min  m
_r m
_ 
K7
 r 
: 

_ r;max  m
_r m
_ r
m

_rm
_ r;max
_ r;min  m
if m
_ r;min > m
_r
if m
_r>m
_ r;max
if m

(A7)

EEV opening percentage penalty:

8
<
K8

0
1
:
jAv  1=Av j

0  Av  1
Av < 0
Av > 1

(A8)

Appendix B. Calculation of state penalty functions

Following procedures are used to calculate penalty for individual system state variable (dependent variable):
_ r and
Step 1: for the assigned independent variables Pc, Pe and u, m
Av are determined by Eqs. (5) and (8), where r is a linear
function of Pc and Tc,r,o

r fr Pc ; Tc;r;o ar Pc br Tc;r;o cr

(B1)

and the coefcients ar, br, cr can be obtained by curve tting for
give refrigerant.
Step 2: Q_ com and Q_ c are determined by Eqs. (7) and (9)
Step 3: for He,r,o and Hc,r,o determined through


He;r;o fHe;r;o Pe ; Te;r;o aHe;r;o Pe bHe;r;o Te;r;o cHe;r;o

Appendix A. Denition of state penalty functions

(A4)

(B2)

and

Condenser air mass ow rate penalty:

8
_
_
_
0
< 

 if mc;air;min  mc;air  mc;air;max

_ c;air;min  m
_ c;air;min > m
_ c;air m
_ c;air
_ c;air 
if
m
K1 
m


:
_ c;air;max  m
_ c;air > m
_ c;air m
_ c;air;max
_ c;air 
m
if m
(A1)


Hc;r;o fHc;r;o Pc ; Tc;r;o aHc;r;o Pc bHc;r;o Tc;r;o cHc;r;o

(B3)

where the coefcients aHe;r;o , bHe;r;o , cHe;r;o , aHc;r;o , bHc;r;o and cHc;r;o can be
obtained by curve tting for the given refrigerant. Te,r,o can then be
solved by

402

Te;r;o

L. Zhao et al. / Energy 55 (2013) 392e402

He;r;o  cHe;r;o

.
bHe;r;o
 aHe;r;o Pe

(B4)

_ r He;r;i and He,r,i Hc,r,o

where He;r;o Q_ e =m
Step 4: for the Hc,r,i determined through


Hc;r;i fHc;r;i Pc ; Tc;r;i aHc;r;i Pc bHc;r;i Tc;r;i cHc;r;i

(B5)

where the coefcients aHc;r;i , bHc;r;i and cHc;r;i can be obtained by curve
tting for the given refrigerant. Tc,r,i can then be determined by

Tc;r;i

.
Hc;r;i  aHc;r;i Hc;r;i  cHc;r;i
bHc;r;i

(B6)

_ r Hc;r;o
where Hc;r;i Q_ c =m
Step 5: for He,g, Te,r,sat, Hc,fg and Tc,r,sat expressed by

He;g fHe;g Pe aHe;g Pe2 bHe;g Pe cHe;g

(B7)

Te;r;sat fTe;r;sat Pe aTe;r;sat Pe2 bTe;r;sat Pe cTe;r;sat

(B8)

Hc;fg fHc;fg Pc aHc;fg Pc2 bHc;fg Pc cHc;fg

(B9)

and

Tc;r;sat fTc;r;sat Pc aTc;r;sat Pc2 bTc;r;sat Pc cTc;r;sat

(B10)

where the coefcients aHe;g , bHe;g , cHe;g , aTe;r;sat , bTe;r;sat , cTe;r;sat , aHc;fg , bHc;fg ,
cHc;fg , aTc;r;sat , bTc;r;sat , cTc;r;sat can be obtained by curve tting for the
_ e;air and m
_ c;air
given refrigerant and operating conditions. Then m
are determined through Eqs. (1) and (3).
References
[1] Dossat RJ, Horan TJ. Principles of refrigeration. 5th ed. Upper Saddle River, N.J.,
U.S.A: Prentice Hall; 2002.
[2] US household electricity report. Washington D.C, U.S.A: Energy Information
Administration; 2005.
[3] Green building design guide. Singapore: Building & Construction Authority;
2007.
[4] Jia X, Tso CP, Jolly P, Wong YW. Distributed steady and dynamic modelling of
dry-expansion evaporators. International Journal of Refrigeration 1999;22:
126e36.
[5] Zhang L, Yang C, Zhou J. A distributed parameter model and its application in
optimizing the plate-n heat exchanger based on the minimum entropy
generation. International Journal of Thermal Sciences 2010;49:1427e36.
[6] Ren C-Q. Corrections to the simple effectiveness-NTU method for counterow
cooling towers and packed bed liquid desiccanteair contact systems. International Journal of Heat and Mass Transfer 2008;51:237e45.

[7] Arsenyeva OP, Tovazhnyansky LL, Kapustenko PO, Khavin GL. Optimal design
of plate-and-frame heat exchangers for efcient heat recovery in process industries. Energy 2011;36:4588e98.
[8] Metin Ertunc H, Hosoz M. Comparative analysis of an evaporative condenser
using articial neural network and adaptive neuro-fuzzy inference system.
International Journal of Refrigeration 2008;31:1426e36.
[9] Kim M-H, Kim J-H, Choi A-S, Jeong J-W. Experimental study on the heat exchange effectiveness of a dry coil indirect evaporation cooler under various
operating conditions. Energy 2011;36:6479e89.
[10] Wang Y-W, Cai W-J, Soh Y-C. A simplied modeling of cooling coils for control
and optimization of HVAC systems. Energy Conversion and Management
2004;45:2915e30.
[11] Xudong D, Wenjian C, Lei J. Evaporator modeling e a hybrid approach.
Applied Energy 2009;86:81e8.
[12] Xudong D, Wenjian C, Lei J. A hybrid condenser model for real-time applications in performance monitoring, control and optimization. Energy Conversion and Management 2009;50:1513e21.
[13] Stoecker WF. Industrial refrigeration handbook. New York, U.S.A: McGrawHill; 1998.
[14] Jensen JB, Skogestad S. Optimal operation of simple refrigeration cycles: part I:
degrees of freedom and optimality of sub-cooling. Computers & Chemical
Engineering 2007;31:712e21.
[15] Jensen JB, Skogestad S. Optimal operation of simple refrigeration cycles: part
II: selection of controlled variables. Computers & Chemical Engineering
2007;31:1590e601.
[16] Jensen JB, Skogestad S. Steady-state operational degrees of freedom with
application to refrigeration cycles. Industrial and Engineering Chemistry
Research 2009;48:6652e9.
[17] Larsen LFS, Thybo C, Stoustrup J, Rasmussen H, A method for online steady
state energy minimization, with application to refrigeration systems. In: 2004
43rd IEEE Conference on Decision and Control (CDC), December 14, 2004e
December 17, 2004, Nassau, Bahamas, 2004, pp. 4708e4713.
[18] Selbas R, Kzlkan , Sencan A. Thermoeconomic optimization of subcooled
and superheated vapor compression refrigeration cycle. Energy 2006;31:
2108e28.
[19] Yuan F, Chen Q. A global optimization method for evaporative cooling systems
based on the entransy theory. Energy 2012;42:181e91.
[20] Fong KF, Lee CK, Chow CK, Yuen SY. Simulationeoptimization of solarethermal refrigeration systems for ofce use in subtropical Hong Kong. Energy
2011;36:6298e307.
[21] Kusiak A, Tang F, Xu G. Multi-objective optimization of HVAC system with an
evolutionary computation algorithm. Energy 2011;36:2440e9.
[22] Kusiak A, Xu G, Tang F. Optimization of an HVAC system with a strength
multi-objective particle-swarm algorithm. Energy 2011;36:5935e43.
[23] Kusiak A, Xu G. Modeling and optimization of HVAC systems using a dynamic
neural network. Energy 2012;42:241e50.
[24] Hovgaard TG, Larsen LFS, Edlund K, Jrgensen JB. Model predictive control
technologies for efcient and exible power consumption in refrigeration
systems. Energy 2012;44:105e16.
[25] Arora CP. Refrigeration and air conditioning. 2nd ed. New Delhi, India: Tata
McGraw-Hill; 2000.
[26] Ding X, Jia L, Cai W, A hybrid modeling for the real-time control and optimization of compressors. In: 2009 4th IEEE Conference on Industrial Electronics and Applications, ICIEA 2009, May 25, 2009eMay 27, 2009, Xian,
China, 2009, pp. 3256e3261.
[27] Dixon SL. Fluid mechanics and thermodynamics of turbomachinery. 4th ed.
Boston, U.S.A: Butterworth-Heinemann; 1998.
[28] Mithraratne P, Wijeysundera NE. An experimental and numerical study of
hunting in thermostatic-expansion-valve-controlled evaporators. International Journal of Refrigeration 2002;25:992e8.
[29] Goldberg DE. Genetic algorithms in search, optimization, and machine
learning. Reading, Mass, U.S.A: Addison-Wesley Pub. Co.; 1989.