Applied Energy 88 (2011) 198–211

Contents lists available at ScienceDirect

Applied Energy
journal homepage: www.elsevier.com/locate/apenergy

Supervisory and optimal control of central chiller plants using simplified

adaptive models and genetic algorithm
Zhenjun Ma, Shengwei Wang *
Department of Building Services Engineering, The Hong Kong Polytechnic University, Hong Kong

a r t i c l e i n f o a b s t r a c t

Article history: This paper presents a model-based supervisory and optimal control strategy for central chiller plants to
Received 18 May 2010 enhance their energy efficiency and control performance. The optimal strategy is formulated using sim-
Received in revised form 16 July 2010 plified models of major components and the genetic algorithm (GA). The simplified models are used as
Accepted 30 July 2010
the performance predictors to estimate the system energy performance and response to the changes of
Available online 24 August 2010
control settings and working conditions. Since the accuracy of the models has significant impacts on
the overall prediction results, the models used are linear in the parameters and the recursive least
Keywords:
squares (RLS) estimation technique with exponential forgetting is used to identify and update the model
Simplified models
Parameter estimation
parameters online. That is to ensure that the linear models can provide reliable and accurate estimates
Genetic algorithm when working condition changes. The GA, as a global optimization tool, is used to solve the optimization
Optimal control problem and search for globally optimal control settings. The performance of this strategy is tested and
Central chiller plant evaluated in a simulated virtual system representing the actual central chiller plant in a super high-rise
Energy saving building under various working conditions. The results showed that this strategy can save about 0.73–
2.55% daily energy of the system studied, as compared to a reference strategy using conventional settings.
Ó 2010 Elsevier Ltd. All rights reserved.

1. Introduction by Yao et al. [6], in which empirical relationships among controlled

variables, uncontrolled variables and system performance were
Central chiller plants, as major functions of building air-condi- established to determine the optimal control settings. Chow et al.
tioning systems, often contribute a significant proportion to the to- [7] used artificial neural networks (ANNs) and genetic algorithms
tal energy consumed in buildings [1,2]. Well monitored and (GAs) to design an optimal control strategy with aims to achieve
controlled central chiller plants have great potentials in improving energy efficient operation of central chiller plants. Braun et al.
the operational reliability and achieving energy conservation. [8,9] devoted considerable efforts on developing optimal and near
Supervisory and optimal control, which addresses energy and cost optimal control strategies for central chiller plants with and with-
efficient control and operation of air-conditioning systems while out using significant thermal energy storage systems. The major
providing desired indoor thermal comfort and healthy environ- characteristics involved in the studies are that the control variable
ment under dynamic working conditions [3], is one of the main interactions in the system were considered seriously and the con-
achievable approaches to reducing the running costs of central trol settings were optimized systematically based on the trade-offs
chiller plants and providing robust operational performance. among energy (or cost) consumptions of their associated
During the last two decades, many efforts have been made on components.
developing and applying supervisory and optimal control strate- The results obtained from above studies showed that substan-
gies for building central chiller plants [2,4–13]. For instance, a tial amount of energy or costs in central chiller plants can be saved
model-based supervisory control strategy for a direct-fired LiBr when supervisory and optimal control strategies were used. Even a
absorption chiller system was developed by Koeppel et al. [4], in small increase in the overall operating efficiency can result in sig-
which the simulated annealing as a global optimization tool was nificant energy or cost savings. Other benefits can also be achieved
used to solve the optimization problem. Ahn and Mitchell [5] pre- including improved control robustness, enhanced thermal comfort
sented an optimal supervisory control strategy for a chiller plant. of occupants, and reduced maintenance costs, etc. The results also
The optimal strategy was formulated using a quadratic regression showed that supervisory and optimal control strategies are more
equation. The optimal control of a large chiller plant was studied energy efficient and cost effective than local control strategies.
It is worthy noticing that most existing supervisory and optimal
* Corresponding author. Tel.: +852 27665858; fax: +852 27746146. control strategies for central chiller plants were developed using a
E-mail address: beswwang@polyu.edu.hk (S. Wang). model-based approach [4–13], where different types of models

0306-2619/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.apenergy.2010.07.036
 Z. Ma, S. Wang / Applied Energy 88 (2011) 198–211 199

Nomenclature

Symbols h parameter of the model to be determined

a0–a2 coefficients e heat transfer effectiveness
b0–b2 coefficients x the capacity flow rate ratio
C1–C9 coefficients
D0 coefficient Subscripts
C mass flow rate capacity, kW/K a air
cp,a the air specific heat, kJ/(kg K) cd condenser
cp,af the fictitious air specific heat, kJ/(kg K) ch chiller
f fitness function com compressor
h enthalpy, kJ/kg ct cooling tower
hb the saturated air enthalpy at the temperature Tb, kJ/kg CTA CTA tower
INT the integral function CTB CTB tower
J cost function des design condition
M flow rate, kg/s ev evaporator
m, n coefficients fic fictitious
N number hx heat exchanger
NTU number of transfer units in inlet
Q heat transfer rate, kW no near optimal
T temperature, °C p primary side of heat exchangers
Tb the equivalent coil surface temperature, °C pu pump
UA overall heat transfer coefficient, kW/K s secondary side of heat exchangers
V loss function set set-point
W power consumption, kW tot total
x regression variable w water
y observed variable wb wet-bulb

Greek symbols
a parameter
k, c, b coefficients

were used in the control systems to estimate system energy perfor- building thermal systems. The U-D factorization algorithm, which
mance and response to the changes of control settings while the operates upon a factored form of the covariance matrix, was used
time-dependent characteristics of the parameters in models were to enhance the computation efficiency and numerical robustness.
not seriously considered. When models are used in control sys- The user-adaptable comfort control for HVAC systems was de-
tems, their prediction accuracy becomes essential. For HVAC online scribed by Federspiel and Asada [19], in which the controller
control applications, since the working condition of the models has learns to predict the actual thermal sensation of the specific occu-
no noticeable changes in finite time step, the models (parameters pant by tuning parameters of the model of the occupant’s thermal
of models) are only required to be accurate in a limited working sensation and the model parameters were estimated using the
range. Therefore, the models can be reasonable simple and online RLS method. In the optimal control strategy for VAV air-condi-
learning and estimation approaches can be used to identify and up- tioning systems developed by Wang and Jin [15], the adaptive fi-
date the model parameters online to ensure the model accuracy nite-time incremental models of major components were used for
when working condition changes. The models using online learn- performance prediction and the model parameters were updated
ing and estimation approaches for parameters estimations are of- by using the RLS estimation with exponential forgetting. The ma-
ten called self-tuning models. Since new operation data are jor focus of this study was paid on the air-side system while the
continuously used to estimate and update the model parameters, performance of the central chiller plant was predicted by using a
the prediction uncertainty of models can be reduced greatly. very simple approach. The results from above studies demon-
In the control field, the importance of using self-tuning models strated that model-based optimal control strategies using online
in optimal control has been addressed by Åström and Wittenmark parameter estimation techniques taking into account the system
[14]. In the HVAC field, the concept of using self-tuning models dynamics can provide more reliable control and better energy
(online adaptive models) in optimal control was addressed by performance.
Wang and Jin [15] and Nassif et al. [16]. A number of model- Considering that there are relatively few and inadequate studies
based optimal control strategies with online parameter estima- related to supervisory and optimal control of central chiller plants
tion techniques were also developed [15,17–19]. Farris and using adaptive approaches, this paper therefore aims at developing
McDonald [17] presented an algorithm for applying advanced an online adaptive optimal control strategy for central chiller
control concepts in HVAC systems. In order to utilize the optimal plants. The optimal control strategy is formulated using a model-
control approach and obtain closed-loop control equations, a lin- based approach, in which simplified models are used as the
earized system representation was used and the sequential least performance predictors and the model parameters are continu-
squares were developed for the parameter estimation. Real-time ously updated online by using the RLS estimation with exponential
predictive supervisory operation of building thermal systems forgetting. The optimization problem is solved by using a GA. The
with thermal mass was studied by Chen [18], in which the recur- performance of this strategy is tested and evaluated in a simulated
sive least-squares (RLS) techniques with U-D factorization algo- virtual environment representing the actual central chiller plant in
rithm were used for generating the z-transfer function of simple a super high-rise building.
200 Z. Ma, S. Wang / Applied Energy 88 (2011) 198–211

Zone 4 From Office Floors (79-98)

Zones 3&4
Zone 1 To Office Floors (79-98)

(S-B)
SCHWP-78-01 to 03
From Podium and Basement Zone 3
From Office Floors (43-77)
To Podium and Basement

HX-78 HX-78 To Office Floors (43-77)

HX-78
(S-B)

(S-B)
SCHWP-06-11 to 13

(S-B)
SCHWP-42-04 to 06 SCHWP-42-01 to 03

Zone 2 From Office Floors (7-41)

HX-06 HX-06
To Office Floors (7-41)

HX-42 HX-42 HX-42 HX-42 HX-42 HX-42 HX-42

(S-B)
SCHWP-06-03 to 05
(S-B)

(S-B)
SCHWP-06-06 to 10
SCHWP-06-01 to 02

PCHWP-06-01 PCHWP-06-02 PCHWP-06-03 PCHWP-06-04 PCHWP-06-05 PCHWP-06-06

EVAPORAROR EVAPORATOR EVAPORATOR EVAPORATOR EVAPORATOR EVAPORATOR

CONDENSER CONDENSER CONDENSER CONDENSER CONDENSER CONDENSER

CDWP-06-01 CDWP-06-02 CDWP-06-03 CDWP-06-04 CDWP-06-05 CDWP-06-06

COOLING COOLING COOLING COOLING COOLING COOLING COOLING COOLING COOLING COOLING COOLING
TOWER 1 TOWER 2 TOWER 3 TOWER 4 TOWER 5 TOWER 6 TOWER 7 TOWER 8 TOWER 9 TOWER 10 TOWER 11

CTA towers CTB towers

Fig. 1. The schematics of the central chiller plant.

2. Description of the central chiller plant and some water is delivered to the second stage heat exchangers
(HX-78) located on the 78th floor. The water system after the sec-
The central chiller plant concerned in this study is a complex ond stage heat exchangers is the conventional primary–secondary
system in a super high-rise building in Hong Kong. Fig. 1 presents system. All pumps in the secondary water system are equipped
the schematics of this system, in which six identical centrifugal with variable frequency drivers except that the primary chilled
chillers with the capacity of 7230 kW each at the design condition water pumps dedicated to the heat exchangers in Zones 3 and 4
are used to supply the cooling energy for buildings. Each chiller is are constant speed pumps. A more detailed description of this sys-
interlocked with a constant condenser water pump and a constant tem, one can found in Ref. [20].
primary chilled water pump. Taking into consideration the winter
plume abatement purposes, two different types of evaporating
3. Formulation of supervisory and optimal control strategy
cooling towers (named CTA tower and CTB tower, respectively)
are used in this building. Each of the CTA towers (total of six) with-
The supervisory and optimal control strategy for central chiller
out heating coils has a heat rejection capacity of 5234 kW. Each of
plants presented in this paper is developed based on the actual sys-
the CTB towers (total of five) equipped with heating coils has a heat
tem described in Section 2. Since Zone 1 of that building covers all
rejection capacity of 4061 kW at the design condition. All cooling
possible control issues in the secondary chilled water system, the
towers are an in-house type and equipped with variable speed ax-
optimal control strategy presented in the following only includes
ial fans.
Zone 1 while other zones are not considered in order to reduce
The secondary chilled water system is divided into four zones to
the complexity of the control system.
avoid the chilled water pipelines and terminal units from suffering
extremely high pressure. Only Zone 2 is supplied with the second-
ary chilled water directly. For the other three zones, the heat 3.1. Outline of the supervisory and optimal control strategy
exchangers are used to transfer the cooling energy from the chillers
to high zones to avoid the high water static pressure. Zone 1 is sup- The optimal control strategy developed is to optimize the oper-
plied with the secondary chilled water through the heat exchang- ation of the central chiller plants using a systematic approach by
ers (HX-06) located on the sixth floor. Zones 3 and 4 are supplied considering the system level and subsystem level characteristics
with the secondary chilled water through the first stage heat and interactions among the entire system. Fig. 2 illustrates the
exchangers (HX-42) located on the 42nd floor. Some of the chilled overall optimization process of the optimal control strategy. It
water after the first stage heat exchangers is delivered to Zone 3 mainly consists of a rule-based supervisor, a GA optimizer, a cost
 Z. Ma, S. Wang / Applied Energy 88 (2011) 198–211 201

Operation data Model parameter Updated model parameters

estimators

Operation Necessary Trial

Real process data inputs settings Settings
of the central Rule-based GA Cost function Performance
chiller plant supervisor Optimizer estimator predictors
Control Optimal Cost Cost
settings settings Search
ranges
Operating
constraints
Operation data

Fig. 2. Overall optimization process of the optimal control strategy.

function estimator, performance predictors (i.e., performance mod- defined in Eq. (1) is used to enhance the GA computation efficiency.
els) and model parameter estimators as well as system operating It is determined based on the consideration of the actual heat
constraints. The control settings optimized include the condenser transfer performance of the heat exchangers.
supply water temperature set-point (Tw,cd,set), chiller (chilled) sup-
0:2 6 T w;hx;set  T w;ch;set 6 2:0 ð1Þ
ply water temperature set-point (Tw,ch,set) and heat exchanger (HX-
06 in Zone 1) supply water temperature set-point (Tw,hx,set). Here, where T is the temperature, and subscripts w, hx, ch and set indicate
the heat exchanger supply water temperature refers to the supply water, heat exchanger, chiller and set-point, respectively.
water temperature at the secondary side of heat exchangers. As shown in Fig. 4, the operating number of the heat exchangers
In this strategy, the performance models are used to predict the and the operating number of the variable speed pumps after heat
system energy performance and response to the changes of control exchangers (also called pumps in the secondary side of heat
settings and working conditions. The models used include a simpli- exchangers) in Zone 1 are optimized simultaneously during the
fied chiller model, a heat exchanger model, a fictitious global AHU model prediction process in each GA trial computation. This is be-
coil model, a cooling tower model and fan/pump models. The mod- cause they are controlled based on the water flow rate in the sec-
el parameter estimators adopting the RLS estimation technique ondary side of heat exchangers in this study. The control rules used
with exponential forgetting [14] are used to update the parameters are as follows: if the water flow rate in the secondary side of heat
required by these models using online measurements. Since GA as exchangers predicted by the fictitious global AHU coil model is less
a global optimization tool can provide sufficiently good solutions than 80% of the design water flow rates of the heat exchanger and
with random initialization and has been widely used in the HVAC pump, one heat exchanger and one pump are put into operation.
field [7,12,15,21], a GA optimizer is used to solve the optimization Otherwise, two heat exchangers and two pumps are put into oper-
problem and seek the most energy efficient control settings on the ation (the similar rules can be used for systems with multiple heat
basis of the cost function estimator and performance predictors. exchangers and pumps). In this study, the operating numbers of
The operating constraints give the upper and lower limits of the chillers and cooling towers, and the operating speeds of variable
control settings to be optimized. The rule-based supervisor is used speed pumps were not optimized by the optimal control strategy.
to provide the actual control settings for the real process based on They are considered to be optimized using local control strategies.
the compromise of the control stability and energy savings accord-
ing to the rules presented in Fig. 3 and Section 3.3. 3.2. Definitions of the cost function, GA fitness function and operating
Fig. 4 illustrates the detailed prediction process under a GA trail constraints
computation and the interconnections among the performance
models. In each GA trail computation, the trail settings are firstly The subsystems (i.e., chiller system, heat rejection system and
sent to the cost function estimator. The cost function estimator will chilled water distribution system) in central chiller plants are
check whether the settings satisfy the relationship described in Eq. interacted with each other, and the fact is that the reduction of
(1). A ‘penalty’ cost, i.e., total energy consumption of 15,000 kW h, the power consumption of one subsystem would result in the in-
which is significantly higher than the values in normal operation, crease of the power consumption of the other subsystem with re-
will be given to the GA optimizer directly without commanding spect to the changes of certain control variables [11–13]. Therefore,
any performance model if the relationship is not satisfied. Other- the systematic optimization is aimed to search for the most energy
wise, the cost function estimator will command each performance efficient control settings that minimize the energy consumption of
model to predict the related cost with proper sequence as illus- the entire system by considering the system level and subsystem
trated in Fig. 4. It is worthwhile to point out that the relationship level characteristics and interactions. The objective function for

Control settings A Yes Update the

Cost A
(identified by the GA) control settings
Cost B − Cost A
≥ε
Cost B
Control settings B No Keep control
Cost B
(used in last time interval) settings unchanged

Fig. 3. Working principles of the rule-based supervisor.

202 Z. Ma, S. Wang / Applied Energy 88 (2011) 198–211

Mw,s Sequence of Npu,s Model of pumps Wpu,s

pumps after HXs after HXs
Ta,in,Ta,o,set, Mw,s,
Ma,Tw,in Global AHU Tw,s,in, Heat exchanger Mw,p Model of pumps Wpu,p
coil model Tw,p,in, model before HXs
Tw,hx,set
Nhx
Tw,cd,set,
Tw,ch,set, Mw,s Sequence of Return to the
Tw,hx,set W
Cost function heat exchanger cost function tot
estimator Tw,ct,in,
Twb,in, estimator
Mw CTA tower Ma CTA tower WCTA,tot
model fan model
Tw,ct,in,
Tw,ch,set,
Chiller model Twb,in,
Tw,cd,set,Qev Mw CTB tower Ma CTB tower WCTB,tot
model fan model
Wch,tot

Fig. 4. Illustration of the model prediction process under a GA trail computation.

the system under investigation (only Zone 1 is considered in this 3.3. Description of the GA optimizer and rule-based supervisor
study) can therefore be mathematically expressed as in Eq. (2).
Since the GA used in this study [22] intends to search for maximum Carroll’s Genetic Algorithm driver [22] is modified and used as
values while the optimization problem is to seek the minimum en- the optimizer. This driver initializes a random sample of individu-
ergy consumption, the GA fitness function is therefore defined as in als with different variables to be optimized using the genetic algo-
Eq. (3). rithm approach, i.e., evolution via survival of the fittest. The
selection scheme used is tournament selection with a shuffling
JðT w;set Þ ¼ min W tot ¼ minðW ch;tot þ W ct;tot þ W pu;tot Þ ð2Þ
technique for choosing random pairs for mating. The routine in-
cludes binary coding for the individuals, jump mutation, creep
1 mutation, and the option for single-point or uniform crossover.
f ðT w;set Þ ¼ ð3Þ
JðT w;set Þ Niching (sharing) and an option for the number of children per pair
of parents have been added.
where J is the cost function, f is the fitness function, W is the power
The GA optimizer starts with a trial with a group of random
consumption, and subscripts ct, pu and tot represent cooling tower,
settings within their allowed ranges (defined in Section 3.2) at
pump and total, respectively.
its first generation. Each set-point trail (values of the three tem-
To ensure the central chiller plant to operate properly, a set of
perature set-points) is given to the cost function estimator to
system operating constraints are considered. The input frequencies
predict the system response and the overall cost (i.e., fitness).
of variable speed pumps and cooling tower fans are constrained
According to the fitness of parents and its rules for crossover
between 20 Hz and 50 Hz. To avoid low chiller (chilled) supply
and mutation, the GA optimizer generates the next generation
water temperature set-point causing the problems of ice in evapo-
population, and the system response is then predicted, and the
rators and low efficiency of chillers, and high chiller supply water
fitness is re-evaluated. The criterion to stop the GA optimizer
temperature set-point resulting in the problems of the humidity
is based on the comparison of the best fitness values of two con-
control and inadequate cooling energy provided, the chiller supply
secutive runs. The GA estimator stops when the relative differ-
water temperature set-point is constrained between 5.0 °C and
ence between the two maximum fitness values reaches a
9.5 °C. Taking into account the actual heat transfer performance
threshold value. A GA run is terminated if the number of the cur-
of heat exchangers, the heat exchanger supply water temperature
rent generation equals to a predefined maximum number. Since
set-point is constrained between 5.5 °C and 10.0 °C. To avoid low
the parameters of the GA driver are essential for the convergence
condenser supply water temperature set-point causing low pres-
speed, the parameters used are carefully selected and deter-
sure problems in chillers, a low limit of 18.0 °C is constrained as
mined on the basis of Carroll’s recommendation and simulation
well. In order to avoid finding local optimal control settings and
tests.
to enhance the search efficiency of the GA optimizer, a near opti-
As presented earlier, the rule-based supervisor is used to pro-
mal strategy [6], as illustrated in Eq. (4), is used to determine a
vide the actual control settings for the real process. For each gi-
near optimal condenser supply water temperature set-point. The
ven condition, the GA optimizer will identify a set of optimal
actual search range of the condenser supply water temperature
control settings (i.e., Control settings A in Fig. 3). The resulted
set-point is defined based on this near optimal temperature set-
energy consumption (i.e., Cost A) by using this set of settings
point by using Eq. (5).
is also predicted. In the meantime, the control system will esti-
T w;cd;no ¼ a0 þ a1 T wb þ a2 T 2wb ð4Þ mate the system energy consumption (i.e., Cost B) when contin-
uously using the last settings (i.e., Control settings B). When the
maxð18  C;T w;cd;no  DTÞ 6 T w;cd;set 6 maxð18  C;T w;cd;no þ DTÞ ð5Þ difference (i.e., energy saving) between the two sets of energy
consumption is significant (i.e., e in Fig. 3 is greater than 1.0%),
where a0–a2 are coefficients, and subscripts cd, wb and no indicate the control settings identified by the GA optimizer will be used
condenser, wet-bulb and near optimal, respectively. to update the last settings. Otherwise, the control settings re-
In this study, the coefficients in Eq. (4) were not updated online main unchanged. It is worthy noticing that the ‘cost’ related to
since this near optimal strategy was only used to determine the the use of the last settings is predicted using the same cost func-
search center for the optimal control strategy. The coefficients in tion estimator and performance predictors as that used by the
Eq. (4) used are as follows: a0 = 15.3958; a1 = 0.1476; a2 = 0.0197. GA optimizer.
 Z. Ma, S. Wang / Applied Energy 88 (2011) 198–211 203

3.4. Description of the performance models and model parameter be calculated based on the inlet and outlet water temperatures
learning tool and the water flow rates at both sides of heat exchangers. Using
the heat transfer coefficients calculated at the current and former
In model-based control and optimization, performance modes two sampling instants, the RLS estimation is then used to estimate
are used to predict the system response and energy performance and update the model parameters.
as well as environmental quality. All models are connected with
Q ¼ e  C min  ðT w;s;in  T w;p;in Þ ð10Þ
the power consumption or operating cost directly. Online mea-
surements collected from the building automation systems (BASs)
1  expðNTUð1  xÞÞ
can be used to tune the model parameters to make them represent e¼ ð11Þ
1  x expðNTUð1  xÞÞ
the actual system accurately. To ensure robust and reliable control
performance, a series of semi-physical models are used in this
UA
study. The details of the models used are presented as follows. NTU ¼ ð12Þ
C min
3.4.1. The fictitious global AHU coil model
UA ¼ b0 þ b1 Mw;s þ b2 M 2w;s ð13Þ
A fictitious global AHU coil is assumed to represent all AHU coils
in the entire zone (Zone 1) of the central chiller plant under study. where C is the capacity flow rate, x is the ratio of the minimum
Given the air inlet and outlet temperatures, humidity, water inlet capacity flow rate to the maximum capacity flow rate, and sub-
temperature and air flow rate, the global AHU coil model is used scripts s and p represent secondary side of heat exchangers and pri-
to predict the required chilled water flow rate in the entire zone mary side of heat exchangers, respectively.
and the outlet water temperature from the AHU coil. The coil mod-
el used is a semi-physical model previously developed by Wang 3.4.3. Chiller model
and Jin [15]. In this model, the total heat transfer rates on the water Given the condenser inlet water temperature, evaporator cool-
side and air side are computed by using Eqs. (6) and (7), respec- ing energy and its outlet water temperature set-point, the chiller
tively. The heat transfer coefficients at the water side and air side model can predict the required power consumption of the com-
are assumed to be related with the water flow rate and air flow rate pressor. The model used in this study is a simplified physical model
only and computed by using Eqs. (8) and (9), respectively. To iden- developed previously [13]. In this model, a fictitious refrigeration
tify the model parameters of cw, ca, bw and ba, both heat transfer cycle, as shown in Fig. 5, is assumed to simplify the complicated
coefficients in the water side and air side need to be calculated thermodynamic processes in the refrigeration systems. The overall
based on the inlet and outlet air and water states of the coil. Using heat transfer coefficients of the evaporator and condenser (UAev,
the heat transfer coefficients calculated at the current and former UAcd) are represented empirically as in Eqs. (14) and (15), respec-
sampling instants, the RLS estimation is used to estimate and up- tively [24]. The actual chiller power consumption (Wcom) is com-
date the parameters of the coil model. puted based on a fictitious power consumption (Wfic), as in Eq.
(16). The fictitious power consumption is computed based on the
Q tot ¼ UAw ðT b  T w;in Þ ð6Þ
fictitious refrigeration cycle. There are nine parameters (C1–C9)
that need to be identified in this model. To identify the parameters
Q tot ¼ UAa ðha;in  hb Þ ð7Þ
of C1–C6, the overall heat transfer coefficients of the evaporator and
condenser need to be calculated based on the evaporator cooling
UAw ¼ cw ðMw Þbw ð8Þ energy and measured compressor power consumption together
with the calculated evaporator and condenser logarithm mean
UAa ¼ ca ðMa Þba ð9Þ temperature differences. To identify the parameters of C7–C9, the
fictitious power consumption (Wfic) needs to be calculated based
where UA is the heat transfer coefficient, M is the flow rate, Q is the on the condensing temperature and evaporating temperature by
heat transfer rate, Tb is the equivalent coil surface temperature, hb is using the fictitious refrigeration cycle. Using the heat transfer coef-
the saturated air enthalpy at the temperature Tb, c and b are the ficients and fictitious power consumptions calculated at the cur-
model parameters to be identified, and subscripts in and a represent rent and former two sampling instants, the RLS estimation
inlet and air, respectively. technique is used to estimate and update the parameters of the
chiller model.
3.4.2. Heat exchanger model
When the global AHU coil model predicted the chilled water
flow rate and chilled water return temperature (i.e., the outlet
water temperature from the AHU coil) at the secondary side of heat
exchangers, the heat exchanger model is then used to predict the
Pressure
required chilled water flow rate and outlet water temperature at Qcd = Qev + Wcom
the primary side of heat exchangers since the chilled water inlet
3 3’ 2’ 2
water temperature in the primary side of heat exchangers is Pcd
known and the outlet water temperature at the secondary side of
heat exchangers is the set-point. In this study, the performance
of the water-to-water heat exchanger with counter flow is mod- Wcom
eled using classical e-NTU method. The actual heat transfer in the
Pev
heat exchanger is computed using Eq. (10). The heat transfer effec-
4 4’ 1’ 1
tiveness (e) and the overall number of transfer units (NTU) are sim-
ulated using Eqs. (11) and Eq. (12), respectively. The overall heat Qev
transfer coefficient is computed by Eq. (13), which is considered
Enthalpy
as a function of the water flow rate at the secondary side of heat
exchangers. There are three parameters (b0–b2) in this model. To Fig. 5. Illustration of the fictitious refrigeration cycle assumed (actual cycle: 1–2–
identify these parameters, the heat transfer coefficient needs to 3–4; fictitious cycle: 10 –20 –30 –40 ).
204 Z. Ma, S. Wang / Applied Energy 88 (2011) 198–211

0:745 1 but the data that is n time units old is weighted by an. The method
C 1 M 0:8
w;eV þ C 2 Q eV þ C3 ¼ ð14Þ
UAeV is therefore called exponential forgetting.

1 1X t
C 4 M 0:8
w;cd þ C 5 ðQ eV þ W com Þ
1=3
þ C6 ¼ ð15Þ Vðh; tÞ ¼ ati ðyðiÞ  xT ðiÞhÞ2 ð21Þ
UAcd 2 i¼1

W com ¼ C 7 þ C 8 W fic þ C 9 W 2fic ð16Þ where V is the loss function, y is the observed variable, h is the
parameter of the model to be determined, x is the regression vari-
able, a is a parameter (0 < a 6 1).
3.4.4. Cooling tower model
Given the inlet water and air temperatures and the water flow
4. Performance tests and analysis of results
rate flowing through the cooling tower, the cooling tower model
can predict the required air flow rate that can control the system
4.1. Test platform
to operate at the desired temperature set-point. In this study, a
simplified model developed by Lebrun et al. was used [23]. In this
A simulated virtual system representing the actual central chil-
model, the cooling tower was treated as an equivalent heat ex-
ler plant under study was used as the test platform of the optimal
changer and modeled using the classical e-NTU method. The simu-
control strategy. This simulated virtual system was constructed
lations of the heat transfer rate and the overall number of transfer
based on a transient simulation program TRNSYS. In this virtual
units are similar to that used in the heat exchanger model pre-
system, a number of local control strategies considering the dy-
sented previously. The heat transfer coefficient of the cooling
namic characteristics of the local process environment were used
tower is simulated using Eq. (17), where cp,af is the fictitious air
to ensure the robust operation and keep track of the control set-
specific heat which is computed by using Eq. (18). Through the log-
tings provided by the optimal control strategy. Since the construc-
arithm transformation, Eq. (17) can be linear in the parameters. To
tion of simulated virtual systems for buildings and air-conditioning
identify the model parameters of Do, m and n, the heat transfer
systems has been well presented in a number of studies [24,25],
coefficient of the cooling tower needs to be calculated based on
the description of the construction of this virtual system is there-
the inlet and outlet air and water states and calculated heat rejec-
fore not provided here for the conciseness of the paper size.
tion capacity. Using the heat transfer coefficients calculated at the
The local control strategies used in the simulated virtual system
current and former two sampling instants, the RLS estimation tech-
are as follows. The chillers were sequenced based on their design
nique is used to identify and update the model parameters.
cooling capacities. A threshold of 10% of the design cooling capac-
 m  n
Mw Ma cp;af ity was used for bringing online and offline of chillers. The operat-
UA ¼ D0   ð17Þ ing number of the cooling towers was controlled based on the
M w;des Ma;des cp;a
operating number of the chillers. For the CTA towers and CTB tow-
ha;out  ha;in ers used in this building, their operating numbers were determined
cp;af ¼ ð18Þ using Eqs. (22) and (23) respectively, where INT is the integral
T wb;out  T wb;in
function and 11 is the total number of towers. For stable control
where cp,a is the air specific heat. and operation, a minimal time interval was introduced in these se-
quence strategies to avoid an alternating fashion of ON/OFF of the
3.4.5. Pump and fan models corresponding components. The temperature controllers were
The power inputs of the cooling tower fan and secondary water used to maintain the condenser supply water temperature at its
pump are modeled to be approximately proportional to their flow set-point by adjusting the operating frequencies of the cooling
rates cubed as in Eq. (19) when the changes of the flow rates are tower fans. The temperature controllers were also used to control
small in a finite step or working range. Since the power consump- the outlet air temperatures of the cooling coils at the desired set-
tion (W) and flow rate (M) are measured, the parameter (k) can be point by modulating the openings of the water control valves.
learnt and estimated directly by using Eq. (20) and updated at each The pressure controller was used to control the operating speeds
sampling instant. of variable speed pumps at the secondary side of heat exchangers
by maintaining the pressure difference of the critical loop at a pre-
W ¼ kM3 ð19Þ determined constant value of 120 kPa. A cascade controller, as
illustrated in Fig. 6, was used to control the operating speeds of
Wk the variable speed pumps at the primary side of heat exchangers
kk ¼ ð20Þ
ðMk Þ3 by maintaining the supply water temperature at the secondary
side of heat exchangers at its set-point. The chiller control system
maintains the evaporator outlet water temperature at the set-point
3.4.6. RLS estimation with exponential forgetting by adjusting the angle of the inlet guide vane of the compressor.
All models presented above are linear in the parameters directly The above temperature set-points were provided by the high-level
or linear in the parameters after through the logarithm transforma- control strategies (such as the optimal control strategy developed
tion. The parameters in the models are considered to be slowly in this study) except that the supply air temperature set-point
time-varying, and be constant within a limited working range. To was maintained at 13.0 °C.
ensure the models to provide reliable estimates when working  
condition changes, the RLS estimation technique with exponential minð2  Nch þ 1; 11Þ
NCTA ¼ INT þ1 ð22Þ
forgetting [14] is used to update the model parameters online. 2
Exponential forgetting is a way to discard old data. It is based on
the assumption that the least squares loss function is replaced by a
 
minð2  Nch þ 1; 11Þ
loss function in which old data is discounted exponentially, as NCTB ¼ INT ð23Þ
2
shown in Eq. (21). The parameter a is called the forgetting factor.
The loss function of Eq. (21) implies that a time-varying weighting where N is the operating number of equipment, and subscripts CTA
of the data is introduced. The most recent data is given unit weight, and CTB indicate CTA tower and CTB tower, respectively.
 Z. Ma, S. Wang / Applied Energy 88 (2011) 198–211 205

Secondary side of HX Primary side of HX

From terminal units

HX

To terminal From primary

units M
T TT M
T system
To primary
HX
system

Temperature Temperature Water flow Water flow

set-point controller set-point controller
Fig. 6. The speed control logic of the variable speed pumps at the primary side of heat exchangers.

4.2. Test conditions what different from that in traditional office zones and office build-
ings. The reason is that Zone 1 of that building mainly serves as
To evaluate the performance of the developed optimal control commercial center involving hotel ballrooms, shopping arcades
strategy (namely Strategy A), two control strategies including and arrival lobbies. In the tests, only Zone 1 of the building was
Strategy A were tested on the test platform described in Section operated while the other zones were not operated. Therefore, the
4.1. The second control strategy (namely Strategy B) is a conven- chillers were sequenced based on the cooling demand of Zone 1
tional control strategy used for performance comparison. In this only.
strategy, the chiller and heat exchanger supply water temperature In order to evaluate the performance of the optimal control
set-points were set to be constant, and the design temperature set- strategy, many tests in various working conditions were per-
points of 5.5 °C and 6.3 °C were used, respectively. The condenser formed. The operational data of the tests in three typical working
supply water temperature set-point was set using the fixed ap- days are selected and presented in this study. The selected three
proach control method. The fixed approach control method intends typical days represent the typical operating conditions of the
to maintain a constant temperature difference (named approach) air-conditioning system in the spring, mild-summer and sunny-
between the condenser supply water temperature set-point and summer, respectively. The variations of the cooling loads in these
ambient air wet-bulb temperature. The condenser supply water typical working days were similar to the variations of the cooling
temperature set-point is therefore derived based on continuously loads presented in Fig. 7. Since the condenser supply water tem-
monitoring of ambient air wet-bulb temperatures. In this study, perature set-points in Strategy B were determined based on the
the design approach of 5.0 °C was used. weather data directly, the dry-bulb temperatures and wet-bulb
To calculate the air flow rates and inlet air states of cooling coils, temperatures in the three typical working days are therefore pre-
a series of assumptions as follows were used. A minimum ratio of sented in Fig. 8. Actually, the weather data and cooling loads of
the fresh air was assumed as 10%. The room design air dry-bulb the second day in Fig. 7 were selected as the operating conditions
temperature was set 23 °C with a 50% relative humidity. The air of the typical spring day.
dry-bulb temperature leaving cooling coil was controlled at 13 °C
with a 95% relative humidity. During the tests, the sampling inter- 4.3. Test and validation of the performance models
val and prediction period used in Strategy A were 60 s and 300 s
respectively, and the simulation time step of the virtual system Based on above test conditions selected, the performances of
simulation was 60 s. major models are validated based on the test platform presented
In the tests, the weather data in Hong Kong was used as the in Section 4.1. Since the modeling approaches of the fan and pump
working conditions. Hong Kong is located at latitude 22°150 north models are very simple, the test and validation results of both
and longitude 114°100 east. The climate in Hong Kong is sub-trop- models are therefore not provided in this study.
ical, tending towards temperate for nearly half the year. In winter, Fig. 9 shows the estimated and ‘measured’ heat transfer coeffi-
the weather is cold and dry. In spring and autumn, the weather is cients (i.e., UA values) at the air side and water side of the global
warm with occasional rain and cooler evenings. In summer, the AHU coil, respectively. The estimated heat transfer coefficients
weather is very hot and humid with occasional showers and thun- were estimated by using the fictitious global AHU coil model while
derstorms, and the temperature often ranges between 26 °C and the ‘measured’ values were calculated based on online measure-
34 °C with humidity of 60–70% during the day time, and even high- ments. It can be observed that the estimated UA values varied clo-
er during night. sely to the variations of the ‘measured’ values. During the high
Since the accuracy of the performance models directly affects cooling demand periods, the estimated UA values at the water side
the performance of the optimal control strategy, the performance were slightly higher than the ‘measured’ values, which demon-
of major component models presented in Section 3.4 is therefore strated that the changes of the estimated values were slower than
validated firstly. The test data in four spring days were selected the changes of the ‘measured’ values. The response of the RLS esti-
to demonstrate their prediction performance. Fig. 7 presents the mator to the change of the system characteristics can be acceler-
weather data and cooling load profiles in the selected four spring ated by reducing the value of the forgetting factor, but a too low
days, where the cooling loads were simulated using EnergyPlus value might cause unstable estimation of the coefficients. From
[26] based on the design data and weather data in the typical year Fig. 9, it can also be found that the heat transfer coefficient at
of Hong Kong. From Fig. 7, it can be observed that the cooling load the water side was much higher than that at the air side in that
profiles (i.e., the cooling loads in Zone 1 of the system) were some- the water is easier to conduct heat than air.
206 Z. Ma, S. Wang / Applied Energy 88 (2011) 198–211

5000 40

4500
Cooling load
35
4000

3500

Cooling load (kW)

Temperature (°C)
30
3000

2500 25

2000
Dry-bulb
20
1500 temperature

1000
15
500
Wet-bulb temperature
0 10
0 6 12 18 24 30 36 42 48 54 60 66 72 78 84 90 96
Time (h)
Fig. 7. Weather data and cooling load profiles of Zone 1 in 4 spring days.

35

30

25
Temperature (°C)

20

15

10 Dry-bulb temp. in sunny-summer day Dry-bulb temp. in mild-summer day

Dry-bulb temp. in spring day Wet-bulb temp. in sunny-summer day
5 Wet-bulb temp. in mild-summer day Wet-bulb temp. in spring day

0
0 2 4 6 8 10 12 14 16 18 20 22 24
Time (h)
Fig. 8. Dry-bulb and wet-bulb temperatures in the three selected typical working days.

1400
'Measured' value in water side Estimated value in water side
1200 'Measured' value in air side Estimated value in air side

1000 Water side

UA (kW/K)

800

600
Air side

400

200

0
0 6 12 18 24 30 36 42 48 54 60 66 72 78 84 90 96
Time (hour)
Fig. 9. Comparison between the estimated and ‘measured’ UA values at both sides of the AHU coil.
 Z. Ma, S. Wang / Applied Energy 88 (2011) 198–211 207

Figs. 10 and 11 show the validation results of the heat exchan- out that, to avoid the effects of strong dynamics and to initialize
ger model and cooling tower model, respectively. It can be ob- the RLS algorithm, the parameters estimated by the RLS estimators
served that UA values in the heat exchanger and cooling towers during the starting phase of the central chiller plant were not rec-
(two types of cooling towers) varied during the day. The trends ommended to be used to update the model parameters.
of their variations were similar to the trends of the variations of
the cooling loads presented in Fig. 7. The estimated UA values were 4.4. Test and validation of the optimal control strategy
all very close to the ‘measured’ values, demonstrating that both
models can provide good performance in prediction. The CTB tower As presented earlier, the values of the control settings directly
with a lower heat rejection capacity has lower heat transfer coeffi- affect the overall energy performance of central chiller plants.
cients than that of the CTA tower with a larger heat rejection The systematic optimization is aimed at seeking the globally opti-
capacity. mal settings that can minimize the overall energy or cost. In this
The comparisons between the model estimated and online section, the performance of the central chiller plant using the set-
‘measured’ instantaneous chiller power consumptions and UA val- tings provided by the two control strategies (Strategy A and Strat-
ues of the chiller evaporator and condenser are presented in Figs. egy B) are evaluated and compared.
12 and 13, respectively. From both figures, it can be found that Figs. 14–16 present the profiles of the optimal temperature set-
the estimated values agreed well with the coincident ‘measured’ points searched by Strategy A for the typical spring day, mild-sum-
values. mer day and sunny-summer day, respectively. These temperature
The above results showed that these performance models com- set-points were optimized systematically based on the trade-offs
bined with the RLS estimation technique with exponential forget- among the power consumptions of chillers, cooing towers and
ting can provide satisfactory performance in prediction and are secondary water pumps. It can be found that the temperature
reliable for online control applications. It is worthwhile to point set-points searched were not constant but varied significantly.

5000
'Measured' value
Estimated value
4000
UA (kW/K)

3000

2000

1000

0
0 6 12 18 24 30 36 42 48 54 60 66 72 78 84 90 96
Time (hour)
Fig. 10. Comparison between the estimated and ‘measured’ UA values of the heat exchanger.

900
'Measured' value of CTA tower
Estimated value of CTA tower
800 'Measured' value of CTB tower
Estimated value of CTB tower
700
CTA tower
600
UA (kW/K)

500

400
CTB tower
300

200

100
0 6 12 18 24 30 36 42 48 54 60 66 72 78 84 90 96
Time (hour)
Fig. 11. Comparison between the estimated and ‘measured’ UA values of the CTA tower and CTB tower.
208 Z. Ma, S. Wang / Applied Energy 88 (2011) 198–211

1200

1000

Chiller power consumption (kW)

800

600

400
Measured value
Estimated value
200

0
0 6 12 18 24 30 36 42 48 54 60 66 72 78 84 90 96
Time (hour)
Fig. 12. Comparison between the estimated and ‘measured’ power consumptions of the chiller.

5000

4000
UA (kW/K)

3000

2000

"Measured' value of evaporator Estimated value of evaporator

1000
"Measured" value of condenser Estimated value of condenser

0
0 6 12 18 24 30 36 42 48 54 60 66 72 78 84 90 96
Time (hour)
Fig. 13. Comparison between the estimated and ‘measured’ UA values of chiller evaporator and condenser.

27
Supply water temperature set-point (°C)

24

21

18 Condenser
15 Chiller
Heat exchanger
12

0
0 2 4 6 8 10 12 14 16 18 20 22 24
Time (h)
Fig. 14. Profiles of the optimal temperature set-points in the typical spring day.
 Z. Ma, S. Wang / Applied Energy 88 (2011) 198–211 209

27

Supply water temperature set-point (°C)

24

21

18
Condenser
15 Chiller
Heat exchanger
12

0
0 2 4 6 8 10 12 14 16 18 20 22 24
Time (h)
Fig. 15. Profiles of the optimal temperature set-points in the typical mild-summer day.

35
Supply water temperature set-point (°C)

30

25

20 Condenser
Chiller
15 Heat exchanger

10

0
0 2 4 6 8 10 12 14 16 18 20 22 24
Time (h)
Fig. 16. Profiles of the optimal temperature set-points in the typical sunny-summer day.

The differences between the chiller supply water temperature set- 628.7 kW h (2.55%), 390.7 kW h (1.35%) and 220.5 kW h (0.73%) to-
point and the heat exchanger supply water temperature set-point tal energy in the typical spring day, mild-summer day and sunny-
varied significantly as well. At light load conditions, a higher chiller summer day, respectively. These savings were achieved through
supply water temperature set-point and a higher heat exchanger applying the optimal control algorithm only and without adding
supply water temperature set-point were searched. The reason is any additional cost.
that the decrease of the power consumption of the chillers was From Table 1, it can also be observed that cooling towers con-
greater than the increase of the power consumption of the second- sumed significant less energy when Strategy A was used, as com-
ary water pumps with the increase of both temperature set-points pared with that of using Strategy B. The savings of cooling
at light load conditions. Therefore, higher temperature set-points towers were resulted by using relatively higher condenser supply
were selected to save more energy. On the contrary, the lower chil- water temperature set-points. The higher condenser supply water
ler and heat exchanger supply water temperature set-points were temperature set-points required less air flow rates of cooling tow-
searched at high load conditions. ers and hence, less energy consumption of cooling tower fans.
Table 1 summarizes the energy consumption of the central chil- However, higher condenser supply water temperature set-points
ler plant (only Zone 1 is considered in this study) in the three se- will require high energy consumption of chillers due to the higher
lected typical days by using the settings provided by the two condensing temperatures in the chiller condenser. In the mean-
different control strategies. Since the constant condenser water time, the chiller performance was affected by the evaporating tem-
pumps and constant primary chilled water pumps consumed rela- perature, which was affected by the chiller supply water
tively constant power consumptions, their energy consumptions temperature set-point. As shown in Table 1, the energy consump-
were therefore not included. Compared with Strategy B using tra- tions of variable speed pumps in the 3 typical days related to the
ditional settings, Strategy A using optimal settings saved about use of Strategy A were much higher than that of using Strategy
210 Z. Ma, S. Wang / Applied Energy 88 (2011) 198–211

Table 1
Energy consumption of the central chiller plant when using different control strategies.

Test case Control strategy Cooling towers Chillers Pumps Total Saving of towers Saving of chillers Saving of pumps Total saving
(kW h) (kW h) (kW h) (kW h) (kW h) (kW h) (kW h) (%)
Spring Strategy B 3836.0 18476.1 2383.3 24695.4 – – – – –
Strategy A 2715.8 18265.3 3085.6 24066.7 1120.2 210.8 –702.3 628.7 2.55
Mild-summer Strategy B 3639.5 21597.3 3602.9 28839.7 – – – – –
Strategy A 3141.7 21421.8 3885.5 28449.0 497.8 175.5 282.6 390.7 1.35
Sunny-summer Strategy B 2591.7 23295.6 4370.6 30257.9 – – – – –
Strategy A 2225.5 23347.5 4464.4 30037.4 366.2 51.9 93.8 220.5 0.73

2100
Cooling tower (Strategy A) Cooling tower (Strategy B)
1800 Chiller (Strategy A) Chiller (Strategy B)
Pump (Strategy A) Pump (Strategy B)
Total (Strategy A) Total (Strategy B)
Power consumption (kW)

1500

1200

900

600

300

0
0 2 4 6 8 10 12 14 16 18 20 22 24
Time (h)
Fig. 17. Profiles of the power consumption in the typical spring day.

B. The high energy consumption of pumps was caused by using formulated using simplified linear self-tuning models and the ge-
higher chiller and heat exchanger supply water temperature set- netic algorithm. The performances of the self-tuning models and
points. Although higher chiller and heat exchanger supply water the optimal control strategy were tested and evaluated on a simu-
temperature set-points consumed more energy of pumps, but they lated virtual system representing the actual central chiller plant in
helped to save the energy consumption of the chillers. Therefore, a super high-rise building. The results showed that the simplified
the actual energy consumption of chillers was affected by both linear models combined with the recursive least squares (RLS) esti-
the condenser and chiller supply water temperature set-points. mators can provide satisfactory accuracy and reliability in
In comparison with using Strategy B, the chillers consumed prediction.
210.8 kW h and 175.5 kW h less energy in the typical spring and The evaluation of the optimal control strategy showed that this
mild-summer days respectively, but they consumed 51.9 kW h strategy is capable of optimizing the overall system performance.
more energy in the typical sunny-summer day when Strategy A Compared with a reference strategy using traditional settings, this
was used. optimal strategy can save about 0.73–2.55% daily energy of the sys-
Fig. 17 illustrates more details of the energy consumptions of tem studied. This part of energy savings was achieved through
the cooling towers, chillers, variable speed pumps and their total optimizing the three temperature set-points only. The total energy
energy consumptions in the typical spring day by using the two saving percentage is relatively small when recognizing the model-
different control strategies. Similar to many central chiller plants, ing mismatch and the performance benefits in practice might likely
chillers are the major energy consumers in this central chiller be hard to measure. However, it could be expected that this opti-
plant. The power consumptions of cooling towers and variable mal strategy can help to save more energy in practice when it is
speed pumps were relatively small, particularly at light load used together with the near optimal/optimal local control strate-
conditions. gies for heat exchangers and pumps used in this study. The reasons
Based on the above results, it can be found that Strategy A are that the reference strategy used in this study has already some-
(developed optimal strategy) considering system-level characteris- what optimized the set-points/operation modes and the control
tics and interactions among the overall central chiller plant is more strategies used in practice might be much simpler than the refer-
energy efficient and cost effective, as compared with the control ence strategy used.
strategy using traditional settings. Although the strategy presented in this study was developed
based on the actual configuration of Zone 1 of the building, the
methodology used can be easily extended to any other zones of
5. Conclusions the building or any other primary–secondary chilled water sys-
tems. It is worthy noticing that, when all zones in the building
This paper presented an optimal control strategy for central are considered, the optimal chilled water supply temperature
chiller plants for energy efficiency. The optimal strategy was set-point searched will be somewhat different from the values
 Z. Ma, S. Wang / Applied Energy 88 (2011) 198–211 211

searched in this study in which only Zone 1 is into consideration, [10] Yu FW, Chan KT. Optimization of water-cooled chiller system with load-based
speed control. Appl Energy 2008;85(10):931–50.
due to the interactions among different zones.
[11] Ma ZJ, Wang SW. An optimal control strategy for complex building central
chilled water systems for practical and real-time applications. Build Environ
Acknowledgement 2009;44(6):1188–98.
[12] Lu L, Cai WJ, Soh YC, Xie LH, Li SJ. HVAC system optimization – condenser
water loop. Energy Convers Manage 2004;45(4):613–30.
The research work presented in this paper is financially sup- [13] Ma ZJ, Wang SW, Xiao F. Online performance evaluation of alternative control
ported by a grant of the Hong Kong Research Grants Council strategies for building cooling water systems prior to in-situ implementation.
(5308/08E) and a grant from The Hong Kong Polytechnic University Appl Energy 2009;86(5):712–21.
[14] Åström KJ, Wittenmark B. Adaptive control. 2nd ed. New York: Addison-
as well as the support from Sun Hung Kai Real Properties Limited. Wesley Publishing Company, Inc.; 1995.
The authors would like to thank Mr. W.K. Pau, Sun Hung Kai Real [15] Wang SW, Jin XQ. Model-based optimal control of VAV air-conditioning
Properties Limited for his essential support to this research work. system using genetic algorithm. Building Environ 2000;35(6):471–87.
[16] Nassif N, Moujaes S, Zaheer-uddin M. Self-tuning dynamic models of HVAC
system components. Energy Build 2008;40(9):1709–20.
References [17] Farris DR, McDonald TE. Adaptive optimal control – an algorithm for direct
digital control. ASHRAE Trans 1980;86(1):880–93.
[1] Kusiak A, Li MY. Cooling output optimization of an air handling unit. Appl [18] Chen TY. Real-time predictive supervisory operation of building thermal
Energy 2010;87(3):901–9. systems with thermal mass. Energy Build 2001;33(2):141–50.
[2] Chang YC, Chan TS, Lee WS. Economic dispatch of chiller plant by gradient [19] Federspiel CC, Asada H. User-adaptable comfort control for HVAC systems.
method for saving energy. Appl Energy 2010;87(4):1096–101. Trans ASME, J Dyn Syst, Meas, Control 1994;116(3):474–86.
[3] Wang SW, Ma ZJ. Supervisory and optimal control of building HVAC systems: a [20] Ma ZJ, Wang SW, Pau WK. Secondary loop chilled water in super high-rise.
review. HVAC&R Res 2008;14(1):3–32. ASHRAE J 2008;50(5):42–52.
[4] Koeppel EA, Klein SA, Mitchell JW, Flake BA. Optimal supervisory control of an [21] Fong KF, Yuen SY, Chow CK, Leung SW. Energy management and design of
absorption chiller system. HVAC&R Res 1995;1(4):325–42. centralized air-conditioning systems through the non-revisiting strategy for
[5] Ahn BC, Mitchell JM. Optimal control development for chilled water plants heuristic optimization methods. Appl Energy 2010;87(11):3494–506.
using a quadratic representation. Energy Build 2001;33(4):371–8. [22] Carroll DL. FORTRAN genetic algorithm (GA) driver. Version 1.7a; 2001.
[6] Yao Y, Lian ZW, Hou ZJ, Zhou XJ. Optimal operation of a large cooling system <http://cuaerospace.com/carroll/ga.html>.
based on empirical model. Appl Therm Eng 2004;24(16):2303–21. [23] Lebrun J, Silva CA, Trebilcock F, Winandy E. Simplified models for direct and
[7] Chow TT, Zhang GQ, Lin Z, Song CL. Global optimization of absorption chiller indirect contact cooling towers and evaporative condensers. Build Serv Eng
system by genetic algorithm and neural network. Energy Build Res Technol 2004;25(1):25–31.
2002;34(1):103–9. [24] Wang SW. Dynamic simulation of building central chilling system and
[8] Braun JE, Klein SA, Beckman WA, Mitchell JW. Methodologies for optimal evaluation of EMCS on-Line control strategies. Build Environ 1998;33(1):1–20.
control to chilled water systems without storage. ASHRAE Trans [25] Husaunndee A, Visier JC. Building HVAC system in control engineering – a
1989;95(1):652–62. modeling approach in a widespread graphical environment. ASHRAE Trans
[9] Braun JE. A near-optimal control strategy for cool storage systems with 1999;105(1):319–29.
dynamic electric rates. HVAC&R Res 2007;13(4):557–80. [26] EnergyPlus; 2005. <http://www.eere.energy.gov/buildings/energyplus/>.