Proceedings of BS2015:

14th Conference of International Building Performance Simulation Association, Hyderabad, India, Dec. 7-9, 2015.

A NEW METHOD FOR THE OPTIMAL CHILLER SEQUENCING CONTROL

Sen Huang1, Wangda Zuo1*, Michael D. Sohn2

1
Department of Civil, Architectural and Environmental Engineering, University of Miami,
Coral Gable, FL 33146, U.S.
2
Energy Analysis and Environmental Impacts Division, Lawrence Berkeley National Laboratory,
Berkeley, CA 94720, U.S.
*Corresponding Author: [email protected]

(CSC) to define the conditions under which the

ABSTRACT chillers should be brought online or offline according
Cooling Load based chiller sequencing Control to the cooling load, which is represented by some
(CLC) significantly affects the energy performance indicators. Depending on the methods to indicate the
of multiple-chiller plants. The conventional CLC cooling load, the CSC can be categorized as: return
method has two limitations: first, it cannot guarantee chilled water temperature based CSC, bypass flow
the optimal load distribution; second, it may result in based CSC, direct power based CSC, and Cooling
an inappropriate number of operating chillers. Load based CSC (CLC) (Honeywell 1997). Among
Previous research tended to address the two them, the CLC is considered to be the best because
limitations separately. In this paper, we proposed a other methods employ the indirect indicators of the
new CLC method that overcame the two limitations cooling load, which may not be proportional to the
at the same time. The optimization objective is to cooling load (Sun, et al. 2013). In the CLC, the
minimize the total energy consumption of the chiller cooling load is calculated using the chilled water
plant including chillers, cooling towers and pumps. flow rate and the difference between the chilled
The independent variables are the thresholds for water supply temperature and return temperature (Li,
chiller staging and the condenser water set point. We et al. 2014). Then the calculated cooling load would
implemented this method in a Model Predictive be combined with a state machine (Kent, et al. 1991)
Control (MPC) framework so that the optimization to determine when and which chiller should be
can be continuously performed according to the brought online or offline. For instance, Figure 1
predicted cooling load and wet bulb temperature. To shows the CLC for the chiller plant with three
compare the performance, we also implemented two identical chillers. The transition between states
existing CLC optimization methods (the optimal load indicates adding or reducing the number of operating
distribution method and the cooling capacity based chillers. When one or more chillers is operating,
critical points reset method) in the same MPC another chiller should not be brought online/off
framework. Simulation results showed that the unless the measured load is larger/smaller than a
proposed CLC method could provide about 5.6% certain Critical Point (CP). The CP is determined as
annual energy saving for the studied chiller plant follows
compared to the conventional CLC. The performance ∑ (1)
,
of our method is also better than the other two
existing CLC optimization methods. where is the CP to bring the 1 chiller
online, is the cooling capacity of the chiller,
INTRODUCTION and is the safety factor (e.g., 90%). To avoid chiller
In the United States, commercial building cooling short circling, a waiting time and a dead band
equipment consumed around 2.64 quadrillion BTU are usually employed.
(77.4 GWh) primary energy in 2010, which
accounted for about 2.7% of the nation’s total
primary energy usage (U.S. Department of Energy).
On Off
Westphalen, et al. (2001) reported that chiller plants
represented about 35% of the commercial building
cooling energy consumption. Due to their significant Load > CP 1 Load < CP 1- CP db
energy consumption, optimal control of the chiller (Waiting period=t wait) (Waiting period=t wait)

plants is of great interests to this nation.

Among various configurations of chiller plants, a Load > CP 2 Load < CP 2- CP db
(Waiting period=t wait) (Waiting period=t wait)
multiple-chiller plant is one of the most widely used
types. For the multiple-chiller plant, it is
recommended to operate chillers sequentially than Figure 1 The State machine diagram for CLC
simultaneously (ASHRAE 2011). To operate chillers There are two directions in current research to
in sequence, we need a Chiller Sequencing Control enhance the CLC. One is to optimize the load

- 316 -
Proceedings of BS2015:
14th Conference of International Building Performance Simulation Association, Hyderabad, India, Dec. 7-9, 2015.

distribution between the operating chillers and the chiller plant equipped with chillers and pumps that
other is to reset the CPs according to the estimation can handle variable chilled water flow rates.
of the chiller capacity. We will discuss the concepts
Cooling Capacity based CPs Reset
and the limitations of both methods in the following
sessions. The conventional CLC method assumes that the
chiller cooling capacity at any operating conditions is
Optimal Load Distribution equal to the chiller’s nominal capacity, which is a
According to the ASHRAE Handbook (2011), the capacity measured at the nominal operating
load distribution for the multiple-chiller plant is to condition. However, the actual cooling capacity of a
operate chillers at as the highest Partial Load Ratio chiller varies with its operating conditions (Sun, et al.
(PLR, the ratio of the cooling load handled by the 2013, Li, et al. 2014). As shown in Figure 3, the
chiller to its nominal cooling capacity) as possible. chiller’s capacity increases up to 110% of its nominal
However, the ASHRAE Handbook also points out capacity when the temperature of the condenser
higher chiller PLR does not necessarily mean better water entering the chiller ( , ) decreases from
operational efficiency. To describe chiller operational 23.89 oC (nominal condition) to 18.89 oC. Therefore,
efficiency, we use a coefficient of performance it is possible that the actual cooling capacity of the
(COP), which is the ratio of the cooling energy the operating chillers in a chiller plant is larger than the
chiller provides to its power consumption. Figure 2 summation of their nominal capacities. That means
shows that the highest COPs may occur at relatively that the chiller plant can meet higher cooling load
low PLRs for the three different chillers. without turning on an additional chiller. Since we
usually have a dedicated primary chilled water pump
and a dedicated condenser water pump to each
chiller, reducing the number of the operating chillers
can save energy for their dedicated pumps (ASHRAE
2011). To address this issue, some researchers
proposed model based cooling capacity estimation
methods to reset the CPs according to the chiller
operating conditions (Sun, et al. 2009, Sun, et al.
2013, Li, et al. 2014). However, as we mentioned
above, adjusting the number of the operating chillers
would affect the efficiency of the chiller plant since it
will also change the PLR of each chiller. Because
these CPs reset methods did not consider the side
effect on the chiller efficiency, they may not lead to
the optimal chiller operation.
Figure 2 The relationship between PLRs and relative
COPs for three chillers in the chiller dataset
provided by EnergyPlus (Crawley, et al. 2001)

In order to achieve higher operational efficiency,

some researchers developed model based
optimization methods to adjust the PLR of each
chiller individually according to a given cooling load
(Chang 2004, Chang, et al. 2005, Chang 2006, Chang
2007, Ardakani, et al. 2008, Chang, et al. 2009, Lee,
et al. 2009, Fan, et al. 2011, Geem 2011, Coelho, et
al. 2013, Chen, et al. 2014, Coelho, et al. 2014).
However, the PLR cannot be directly controlled, so it
is not possible to implement these methods directly
in real-world applications. Some scholars revised the
Figure 3 The relationship between the temperature of
above methods by replacing the PLR with other
controllable parameters, such as the chilled water the condenser water entering the chiller and relative
flow rate through each chiller (Yu, et al. 2007, Yu, et cooling capacity for three chillers calculated in the
al. 2008), the temperature set points of the chilled chiller dataset provided by EnergyPlus
water leaving each chiller (Chang, et al. 2006,
Chang, et al. 2008), and the combination of the To sum up, although the load distribution control and
previous two parameters (Lu, et al. 2011). However, CPs reset may interact with each other, they were
these methods still have some limitations. For only studied separately in previous studies. In
instance, the methods of adjusting the chilled water response to this, we proposed a new method to
flow rate through chillers can only be applied to the consider both of them in the CLC optimization

- 317 -
Proceedings of BS2015:
14th Conference of International Building Performance Simulation Association, Hyderabad, India, Dec. 7-9, 2015.

simultaneously. To demonstrate the usage of the variables and assumed they were constant during the
proposed method, we built a Model Predictive period , Δ :
Control (MPC) framework and implemented the , , , ∈ , Δ , (4)
proposed CLC method in the framework. We also
implemented the optimal load distribution and the , ∈ , Δ
cooling capacity based CPs reset methods in the (5)
MPC framework and evaluated the performance of The objective function is to minimize the total energy
the three methods in a case study. consumption under the constraints of physical plants.
The optimization problem can be defined as:
METHODOLOGY min | , ∈ , Δ (6)
To implement the proposed CLC, we built a MPC
subject to:
framework to find the optimal CPs and the optimal
, , , , , , (7)
condenser water set point according to the operating
condition (cooling load and wet bulb temperature).
To enable the MPC, it is indispensable to have a , (8)
model that can realistically represent both the
physical and the control system of the chiller plant.
1 , (9)
However, conventional building modelling tools,
such as EnergyPlus, are not suitable for this purpose where , , and , , are the low and the high
since they tend to highly idealize the control process bounds for , and are the
,
(Piette, et al. 2012). Thus, we selected Modelica that
low and the high bounds for , is the
is an equation-based object-orient modelling
language for dynamic systems (Modelica Association high bound for . There are also other
2000). constrains such as that the temperature of the
temperature of the chilled water leaving
Optimization Formulation chillers, ,lea should be equal to , and
Here we consider a chiller plant with chillers and these constrains were considered in the system
cooling towers. Each chiller has a dedicated model.
constant speed chilled water pump and a dedicated
Model Predictive Control
constant speed condenser water pump. The towers
have variable cooling tower fans controlled by the Figure 4 shows the configuration of our MPC.
same set point for the temperature of the condenser and for a future period (termed as prediction
water leaving the tower, , . Assuming that the horizon) are given by prediction models and the
set points for the temperature of the chilled water prediction horizon would be divided into steps
leaving the chillers, , , are constant, the total (termed as control horizon). For the control horizon
power of chillers, pumps, and cooling towers, , at starting from , , and are used as
time can be described as follows: input variables to perform the optimization, then the
generated optimum , and would be
∑ ∑
, , ,
(2) used to obtain ∆ which would be used in the
, , ,.., , , , ,
optimization at next control horizon.
where , , , , are the powers of
the chiller, the dedicated chilled water pump and
CASE STUDY
condenser water pump for the chiller, and the To compare the performance of proposed CLC
cooling tower, respectively. is the cooling load, method with the optimal load distribution method
is the wet bulb temperature, and is the and the cooling capacity based CPs reset method, the
state vector of the system, including the operating following case study was performed.
status of chillers (On/Off) as well as temperatures of Case Description
chillers and cooling towers.
As shown in Figure 5, we studied a chiller plant with
Then the energy consumption of the chiller plant for three identical chillers and three identical cooling
a period from to ∆ is towers. Each chiller has one dedicated chilled water
∆ pump, one dedicated condenser water pump and one
| . (3) dedicated cooling tower. The model of the chiller is
York_YK2771kW, which has the nominal cooling
The operating status and PLR of each chiller are
capacity ( ) as 2771 kW. The corresponding
modulated for energy saving by
chiller performance curves from the chiller dataset
adjusting ,.., . In addition, the , is
provided by EnergyPlus are adopted in this study.
controlled by changing , so that the cooling For the cooling tower, the nominal fan power is
capacity of the chillers can also be regulated. We 37.285 kW (50 HP). The fan power is assumed to be
used , and as the independent proportional to the cubic of the fan speed ratio. The
nominal wet bulb temperature and approach

- 318 -
Proceedings of BS2015:
14th Conference of International Building Performance Simulation Association, Hyderabad, India, Dec. 7-9, 2015.

Figure 4 MPC configuration

temperature is 23.89 C (75 F) and 0.89 oC (1.6 oF),
o o
models and shown as an icon in the system model.
respectively. The chilled and condenser water pumps The chillers and cooling towers were modelled using
are constant speed pumps and their powers are 34 ElectricEIR and YorkCalc models, both from
kW and 47 kW, respectively. Modelica Buildings library (Wetter, et al. 2014). We
used Modelica_StateGraph2 library (Otter, et al.
2009) to model the CLC prescribed in Figure 6 and
the implementation is shown as Figure 8.

Figure 5 The schematic drawing of the studied

chiller plant

Figure 7 The diagram of the Modelica model in

On Off
the system level for the chilled water plant

Load > CP 1 or T chw,lea > Tchw,set Load < CP 1- CP db

(Waiting period=t wait) (Waiting period=t wait)

Load > CP 2 or T chw,lea > Tchw,set Load < CP 2- CP db

(Waiting period=t wait) (Waiting period=t wait)

Figure 6 The CLC

The operation of chillers is managed by the CLC ( =

90%, = 900s, = 50 ton). We modified the
conventional CLC by adding , as another
indicator so that , would not exceed ,
(shown in Figure 6). In the condenser water loop, a
three-way valve is employed to modulate the flow Figure 8 The implementation of the CLC in
rate through the cooling towers to avoid overcooling. Modelica
The condenser water is considered to be overcooled
if the temperature of the condenser water leaving the We used the historic data of weather and cooling load
cooling tower is less than 12.78 oC (55 oF). measured from a district cooling system in
We used Dymola 2005 FD01 (http://www.3ds.com/) Washington DC as the input variables. This is
as the Modelica simulation platform. Figure 7 shows equivalent to have a perfect prediction model. Figure
the Modelica model in the system level. The sub- 9 shows the on-site measurement of hourly cooling
systems, e.g. chillers, were packaged in sub-system load data in 2012 while Figure 10 shows the hourly

- 319 -
Proceedings of BS2015:
14th Conference of International Building Performance Simulation Association, Hyderabad, India, Dec. 7-9, 2015.

wet bulb temperature on the same year from a dataset intentionally. Therefore, , was fixed as the
called Quality Controlled Local Climatological Data nominal value.
(National Climatic Data Center). A linear Method Opt 2 was designed to evaluate the potential
interpolation was applied to obtain the data between energy saving from the cooling capacity based CPs
the sampling points. reset which aims to reduce the number of operating
chillers. To realize the objective, and could
be larger than the nominal values. We would not
adjust the chiller capacity. Thus, , was fixed as
the nominal value.
Method Opt 3 was our proposed method, which
attempted to consider both energy saving methods
mentioned above. Therefore, the ranges of and
were the combination of those in the Opt 1 and
2. Besides, we also modulated the capacities of the
chiller by adjusting , .
For comparison, we also designed a baseline case to
represent the conventional CLC. For the baseline
case, , and , were all fixed as the
nominal values: ηCC , 2ηCC and 23.89 oC.
Figure 9 The annual hourly cooling load
Result
Figure 11 shows the energy saving of Opt 1, 2 and 3
compared to the baseline. For Opt 1, the energy
savings from chillers was mostly offset by the
increasing energy used of the pumps. Thus, the total
system energy saving for Opt 1 was only around
0.3% although it saved about 4% chiller energy. For
Opt 2, there was almost no energy saving can be
obtained, which means it is difficult to save energy
by reducing the number of operating chillers when
the capacities of the chillers were not regulated. For
Opt 3, the energy saving was more significant than
that in Opt 1 and 2. The annual total energy saving
was around 5.6%. The chiller energy saving ratio in
Opt 3 was 11% while the cooling tower energy
Figure 10 The annual hourly wet bulb temperature
increased by 41%. In addition, the pump energy rose
by 1.7%.
MPC Setting
In this study, we used GenOpt (Wetter 2001) as the
optimization engine and Hooke Jeeves method
(Hooke, et al. 1961) as the optimization algorithm.
The optimization settings for three methods are listed
in the Table 1.

Table 1 Optimization settings for different scenarios

Method ton ton ,

oC

Opt 1 [0,1] [ ,2 ] Fixed as

Opt 2 [1,1.2] [2,2.4] 23.89
Opt 3 [0,1.2] [ , 2. 4 ] [13.89,23.89 ]
(The intervals for CP and , are 0.1 and 1 oC,
respectively)
Figure 11 Annual simulation result
Method Opt 1 was designed to assess the energy
saving from the optimal load distribution only. It did The comparison of simulation results for the three
not consider the possible capacity increase due to the CLC optimization methods shows that if we combine
decreased , . Thus, and should not be the optimal load distribution and the cooling capacity
larger than the nominal values. In addition, the
capacities of the chillers were not regulated

- 320 -
Proceedings of BS2015:
14th Conference of International Building Performance Simulation Association, Hyderabad, India, Dec. 7-9, 2015.

Figure 12 Daily energy saving in Opt 3

Figure 13 Annual distribution of the optimal , , and in Opt 3

based CPs reset, the energy saving from the CLC As shown in Figure 12, the chiller energy
optimization can be significantly increase as shown consumption was saved for the most of time in the
in the results of Opt 3. To understand the result, we studied year, which could be attributed to both the
provide some detail analysis as follows. optimal load distribution and lower , .
Sometimes, the chiller energy consumption may

- 321 -
Proceedings of BS2015:
14th Conference of International Building Performance Simulation Association, Hyderabad, India, Dec. 7-9, 2015.

increase, such as a few hours in September. In those NOMENCLATURE

cases, the number of operating chillers decreased to
= Temperature
reduce the energy use by dedicated pumps, although
= Cooling load
the PLR for each operating chiller raised and the total
chiller energy increased. = Energy consumption
= Power
The cooling tower energy consumption mostly = Partial load ratio
increased because the system cooled the condenser = State vector
water to a lower set point with the aim of reducing = Critical point for chiller staging control
chiller energy. It is also interesting to see that cooling = Safety factor
tower energy consumption reduced sometimes in the = Cooling capacity
summer (May to September). This happened when
= Cooling towers ,chillers and Pumps
the number of operating cooling towers increased and
= Wet bulb temperature
the cooling load handled by each towers decreased.
= Condenser water
The cooling tower energy consumption dropped
= Chilled water
because the variable speed cooling tower fans
= Chiller
worked more effectively at partial speeds rather than
= Cooling tower
the full speed.
= Pump
The pump energy consumption was increased or = Set point
reduced around the year depending on the number of = Instantaneous time
operating chillers. In the summer, the pump energy = Start time for a period
consumption usually increased which indicates that = Lower limit
more chillers were operating in Opt 3 compared with = Higher limit
the baseline. In the rest time, the pump energy = Entering the chiller
consumption was reduced which means the cooling = Leaving the chiller
load was met with less chillers.
= Dead band
Based on Figure 13, we can see that the optimal = Waiting time
, was mostly different from that the constant
set point of 23.89 oC in the baseline. In general, the ACKNOWLEDGEMENT
value of optimal , was high in summer and low This research was supported by the U.S. Department
at the other season. The optimal was larger than of Defense under the ESTCP program.
that in the baseline for most of time, which indicates The authors thank Marco Bonvini, Michael Wetter,
that we could delay the start of the second chiller to Mary Ann Piette, Jessica Granderson, Oren Schetrit,
save energy. In addition, the optimal is Rong Lily Hu and Guanjing Lin for the support
sometimes less than that used in the baseline provided through the research.
(reduced by up to 33%). This means that we could This research also emerged from the Annex 60
also save energy by increasing the running time of project, an international project conducted under the
the second chiller. On the other side, the optimal umbrella of the International Energy Agency (IEA)
was usually small than that in the baseline (dropped within the Energy in Buildings and Communities
by up to 50%), which means we could enhance the (EBC) Programme. Annex 60 will develop and
energy efficiency by making the third chiller operate demonstrate new-generation computational tools for
more frequently. building and community energy systems based on
Modelica, Functional Mockup Interface and BIM
CONCLUSION standards.
In this study, we proposed a new CLC method for
multi-chiller plants by combining the optimal load REFERENCES
distribution method and the cooling capacity based Ardakani, A. J., et al. 2008. A novel approach for
CPs reset method. The new method was implemented optimal chiller loading using particle swarm
using a MPC framework. Our case study showed that optimization. Energy and Buildings 40(12):
the new method could provide a higher energy saving 2177–2187.
for the whole chiller plant compared with the optimal ASHRAE 2011. ASHRAE Handbook HVAC
load distribution method and the cooling capacity Application. Atlanta, GA USA
based CPs reset method alone. Chang, Y. 2007. An innovative approach for demand
In this study, the evaluation of this proposed method side management—optimal chiller loading
was limited to the application in the chiller plant with by simulated annealing. Energy 31(12):
identical chillers. In future study, we could assess the 1883–1896.
performance of this method for chiller plants with Chang, Y. 2004. A novel energy conservation
non-identical chillers. method - optimal chiller loading. Electric
Power Systems Research 69(2-3): 221-226.

- 322 -
Proceedings of BS2015:
14th Conference of International Building Performance Simulation Association, Hyderabad, India, Dec. 7-9, 2015.

Chang, Y. 2006. An outstanding method for saving Li, Z., et al. 2014. Stochastic chiller sequencing
energy-optimal chiller operation. IEEE control. Energy and Buildings 84(2014):
transactions on energy conversion 21(2): 203-213.
527-532. Lu, Y., et al. 2011. Using cooling load forecast as the
Chang, Y., Chen, W. 2008. Optimal chilled water optimal operation scheme for a large multi-
temperture calculation of mutiple chiller chiller system. International Journal of
systems using Hopfield neural network for Refrigeration 34(8): 2050–2062.
saving energy. Energy 34(4): 448-456. Modelica Association 2000. Modelica- A Unified
Chang, Y., et al. 2006. Simulated annealing based Object-Oriented Language for Physical
optimal chiller loading for saving energy. Systems Modeling.
Energy Conversion and Management 47(15- National Climatic Data Center. Quality Controlled
16): 2044-2058. Local Climatological Data. Retrieved May
Chang, Y., et al. 2009. Evolution strategy based 14, 2015, from
optimal chiller loading for saving energy http://www.ncdc.noaa.gov/data-access/land-
Energy Conversion and Management 50(1): based-station-data/land-based-
132–139. datasets/quality-controlled-local-
Chang, Y., et al. 2005. Optimal chiller sequencing by climatological-data-qclcd.
branch and boun method for saving energy. Otter, M., et al. 2009. A New Formalism for
Energy Conversion and Management 46(13- Modeling of Reactive and Hybrid Systems.
14): 2158-2172. In Proc. of the 7th Modelica Conference
Chen, C., et al. 2014. Applying smart models for Como, Italy.
energy saving in optimal chiller loading. Piette, M. A., et al. 2012. Responsive and Intelligent
Energy and Buildings 68 Part A: 364-371. Building Information and Control for Low-
Coelho, L. d. S., et al. 2014. Optimal chiller loading Energy and Optimized Grid Integration In
for energy conservation using a new Proc. of American Council for Energy
differential cuckoo search approach. Energy Efficient Economy summer study on energy
75: 237-243. efficiency in buildings proceedings.
Coelho, L. d. S., Mariani, V. C. 2013. Improved Sun, Y., et al. 2009. Chiller sequencing control with
firefly algorithm approach applied to chiller enhanced robustness for energy efficient
loading for energy conservation. Energy and operation. Energy and Buildings 41(11):
Buildings 59: 273–278. 1246-1255.
Crawley, D. B., et al. 2001. EnergyPlus: creating a Sun, Y., et al. 2013. In situ performance comparison
new-generation building energy simulation and evaluation of three chiller sequencing
program. Energy and Buildings 33(4): 319- control strategies in a super high-rise
331. building. Energy and Buildings 61: 333-343.
Fan, B., et al. 2011. Optimal control strategies for U.S. Department of Energy. Buildings Energy Data
multi-chiller system based on probability Book. Retrieved May 15, 2014, from
density distribution of cooling load ratio. https://catalog.data.gov/dataset/buildings-
Energy and Buildings 43(10): 2813-2821. energy-data-book.
Geem, Z. W. 2011. Solution quality improvement in Westphalen, D., Koszalinski, S. 2001. Energy
chiller loading optimization. Applied Consumption Characteristics of Commercial
Thermal Engineering 31(10): 1848-1851. Building HVAC Systems Volume I:Chillers,
Honeywell 1997. Engineering Manual of Automatic Refrigerant Compressors,and Heating
Control for Commercial Buildings. Systems.
Minneapolis, Minnesota. Wetter, M. 2001. GenOpt-a generic optimization
Hooke, R., Jeeves, T. A. 1961. `` Direct Search'' program. In Proc. of the 7th IBPSA
Solution of Numerical and Statistical Conference, Rio de Janeiro, Brazil.
Problems. Journal of the Association for Wetter, M., et al. 2014. Modelica Buildings library.
Computing Machinery 8(2): 212-229. Journal of Building Performance Simulation
Kent, A., Williams, J. G. 1991. Encyclopedia of 7(4): 253-270.
Computer Science and Technology: Volume Yu, F. W., Chan, K. T. 2008. Improved energy
25 - Supplement 10: Applications of performance of air cooled centrifugal
Artificial Intelligence to Agriculture and chillers with variable chilled water flow.
Natural Resource Management to Energy Conversion and Management 49(6):
Transaction Machine CRC Press. 1595–1611.
Lee, W., Lin, L. 2009. Optimal chiller loading by Yu, F. W., Chan, K. T. 2007. Optimum load sharing
particle swarm algorithm for reducing strategy for multiple-chiller systems serving
energy consumption. Applied Thermal air-conditioned buildings. Building and
Engineering 29(8-9): 1730–1734. Environment 42(4): 1581–1593.

- 323 -