Applied Energy 306 (2022) 117975

Contents lists available at ScienceDirect

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

Customized data center cooling system operating at significant outdoor

temperature fluctuations
Mateusz Borkowski, Adam Krzysztof Piłat ∗
AGH University of Science and Technology, Faculty of Electrical Engineering, Automatics, Computer Science and Biomedical Engineering, Department of Automatic
Control and Robotics, Al. Mickiewicza 30, 30-059 Krakow, Poland

ARTICLE INFO ABSTRACT

Keywords: This elaboration presents the configuration of the cooling system of the POLCOM Data Center which utilizes
Data center commercially available components, proposes steering strategy and analyzes their functionality. The designed
Chiller architecture of cooling and dedicated control system is presented to demonstrate the novelty and customization
Freecooling
with respect to the demanding outdoor temperatures existing in the climate of the Małopolska Province. The
COP
cooling system control and operation is illustrated by a comprehensive analysis of compressor and freecooling
Compressor mode
Experimental data
modes. The discussion on embedded functionality of chillers is undertaken. According to features and
limitations of the complex solution, the cooling system reached the annual average coefficient of performance
of 8.63 in 2015 (increase of 110% compared to the year 2014), operating 65% of the time during the year in
compressor mode and 35% in freecooling mode. This coefficient in compressor mode amounted to 4.39 while
in freecooling mode totaled 16.50. It was proved that in the real case under consideration they generated
losses in electricity consumption amounting to 557MWh per year. The real-time experimental data collected
from the commercial Data Center installation are used to present a unique operation of a such complex system.

1. Introduction cooling system is the most energy-intensive component of maintenance

equipment used in Data Centers [14]. The amount of electricity it con-
The currently observed development and popularization of new sumes ranges between 30% and 50% of the total energy consumption
technologies in the IT area entail an increase in the demand for com- of Data Center [6,12,15]. Optimization of a cooling system’s energy
puting power and hard drive space [1–3]. consumption will, therefore, allow for relatively large savings in the
This brings about an increase in demand for, inter alia colocation total energy consumption of server room [16,17]. It is well known
services and cloud computing. The buildings designed to provide this that the Data Center infrastructure and operating modes should be
type of service are professional Data Centers. These are specific pro- designed in an optimal way [18–20]. The optimization of the total
duction facilities in terms of the type of raw materials and consumables workload allocated to servers could be considered during IT installation
used to conduct their production process. This work extends the knowl- as mentioned in [21]. One of the ways to reduce energy consumption by
edge [4–6] with Data Center cooling system located in Poland. In most the cooling system is using natural cooling methods [19,22,23]. They
cases, the only relevant raw material consumed in carrying out the
use the natural cooling resources included in outdoor air or water, for
production process is electricity. It powers IT equipment that provides
example. They can be divided into direct and indirect methods [9,24].
computing and storage assets. Some part of the electricity supplied is
Direct methods are the most effective in terms of energy savings but
consumed to generate computing power, it is almost entirely discharged
are characterized by low reliability and require large expenditures on
by IT hardware as waste in the form of heat energy. This heat energy
maintaining full performance [25]. The main reason is the penetration
is then transported outside the building and dispersed into the outdoor
of pollutants from the outdoor into the cooling system as well as the
air by means of a cooling system [7]. The cooling system also consumes
server room building. This problem is minimized by filtration systems,
electricity for the purposes of heat transfer and dispersion outside the
building. Thus, along with an increase in demand for services provided but those require additional maintenance. Indirect methods are more
by the Data Center, an increase in the amount of electricity it consumes reliable, as they separate the outdoor environment from the cooling
is observed [8–10]. The Data Center building can draw about 50 to system and the building’s interior. However, they are less effective in
100 times more electricity than standard office buildings [11–13]. The terms of energy savings. In general, the freecooling method should take

∗ Corresponding author.
E-mail addresses: [email protected] (M. Borkowski), [email protected] (A.K. Piłat).

https://doi.org/10.1016/j.apenergy.2021.117975
Received 10 June 2021; Received in revised form 25 September 2021; Accepted 27 September 2021
Available online 20 October 2021
0306-2619/© 2021 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
M. Borkowski and A.K. Piłat Applied Energy 306 (2022) 117975

into account air temperature and humidity criteria (see [26]). Study on
airflow in Data Center [27] found that a significant amount of cooling
energy could be saved by operating the chiller system efficiently. It
could be done when the control system has open architecture. In prac-
tice, the choice of the cooling method means striking a balance between
financial outlays, reliability, energy efficiency and service complexity.
In many cases, it is an individual choice dictated by the specificity of
a given building and even the company’s financial standing. A direct
liquid cooling method [28] can also be used, but its crash has a direct
impact on the server and any repair of the server requires intervention
in the cooling system. An thought-provoking recent literature survey on
data center cooling systems technology, power consumption modeling
and control strategy optimization is presented in [29]. In the literature
there is missing information devoted to detailed analysis of real cooling Fig. 1. Outdoor temperature in 2014.
system composed of different components, and methods to steer their
operating modes with respect to requests and external conditions. It is
due to the protected data of Data Center.
freecooling, with a nominal cooling capacity 2 × 129 kW and 2 reserve
chillers with a nominal cooling capacity 2 × 110 kW, which in total
2. Subject of interest
gives 478 kW) would not ensure coverage, with a continued upward
trend, of the projected demand for cooling (with the required N+1
The POLCOM Data Center building is located in Poland, in Skawina,
redundancy). Based on the analysis of historical outdoor temperature
near Krakow. Its geographical location means that the building is in
data, it was found that there was a sufficiently low outdoor temperature
a transitional climate. This climate is defined as a transition between
throughout the year, for a significant proportion of the time, potentially
sea and continental climate. It is characterized by significant rainfall
conducive to the operation of the chillers in freecooling mode. The
throughout the year, large annual fluctuations in air temperature and
frequent weather changes. This is because dry air masses from the historical data also showed a high fluctuation of outdoor temperature
Eurasian continent and humid air from the Atlantic clash over Poland. (both within one day and throughout the year). The significant fluctu-
As a result, the annual temperature amplitude is about 55 ◦ C, and ations in temperature changes and the limited measurement resolution
the daily amplitude is 25 ◦ C. The minimum temperatures in winter in 2014 (amounting to 30 min) provide uncertainty in knowing the
are around −20 ◦ C, and the maximum temperatures in summer are in dynamics of temperature changes in the future. Nevertheless, the ob-
excess of 35 ◦ C. Such a wide range of outdoor air temperature places served changes in temperature over the course of a day indicate a
high demands on the server room’s cooling system. problem with the control of chillers. On the basis of collected outdoor
The building has six modules with an area of about 300 m2 each, temperatures measurements, unsatisfactory performance of chillers,
which are broken down into six independent modules for IT equipment. limited data recorded in parts at 30 min intervals, and with requested
In line with the assumptions of the server room design, IT equipment in increase of cooling system capacity a decision about new chillers in-
each of the modules can be supplied with electrical power of up to 600 stallation and control system modification was established. Till 2014
kW. Therefore, the total power consumed by IT equipment installed in the cooling system had been at the development phase. The chillers
the server room can reach 3.6 MW. The supply of IT modules is secured working in compressor mode only were used in the cooling stage. It
by redundant UPSs on each supply track. The supply of air conditioning was observed that request for cold was increasing. On the basis of the
cabinets is also maintained by UPS. As standard, the IT modules can recorded temperature measurements, the analysis was conducted and
hold eight cubes built from racks. Each cube consists of 16 racks. In it was noted that the temperature changes throughout the year in a
total, each IT modules can hold 128 racks. non-linear manner, the maximum temperatures occurred in the summer
Separation of hot and cold zones was used. The cold air zone is period and reaching 34.0 ◦ C, the minimum temperatures of about -
under the raised floor and inside the rack cubes. The warm air zone is 16.5 ◦ C were recorded in the winter period. During the 24-h period,
above the raised floor, outside the rack cubes. The Computer Room Air the temperature variation can reach up to 21.7 ◦ C and the average 24-h
Handlers (CRAHs) draw warm air from the top of the server module temperature amplitude reaches 10.1 ◦ C. The outside temperature data
and, after cooling, blow it underneath the raised floor. Chilled water are presented in Fig. 1.
from chillers is transported via a pipe to CRAH in the IT modules. The
During the year 2014, an increase in the demand for cooling capac-
cooling pipe is redundant, i.e. CRAH is supplied by two independent
ity was observed. This trend was approximated using linear regression.
pipes. Each CRAH has two independent heat exchangers.
The cooling capacity demand amounted to 336 kW and 402 kW at the
The cost of operating a cooling system is a decisive factor for the
beginning and the end of the year respectively. Thus, demand increased
profitability of server room operations; therefore, optimizing its oper-
by 20% over the course of the year. Based on this approximation, a
ating parameters is crucial. A competitive advantage could be obtained
further increase in cooling power demand to 469 kW at the end of
by extending the functionality of the components by modifying control
2015 was estimated. Therefore, in order to increase the capacity of the
algorithms or their parameters [30].
The architecture of the cooling system in the POLCOM Data Centre cooling system with redundancy, its modification is required. Further
allows for the installation of various types of chillers in order to meet efficiency gains are possible by changing the components of the cooling
the demand for the supply of cooling regardless of the temperature system and gaining at least partial control over their operation.
conditions outside the building. The following questions are raised and finding the answers should
show the operating conditions of the system and allow making future
2.1. Motivation decisions related to the configuration of cooling systems. What is the
variability of outside temperature at the location of the Data Center?
On the basis of historical data from the year 2014, it was deter- How can the cooling system architecture be modified to achieve better
mined that the demand of the server room for cooling capacity was performance and controllability? How does the COP coefficient evolve
characterized by a constant upward trend. The original configuration throughout the year? How can the freecooling operation time of a
of the chillers (2 units of compressor chillers without the possibility of cooling system be maximized?

2
M. Borkowski and A.K. Piłat Applied Energy 306 (2022) 117975

3. Theory outdoor temperature drops below 𝑇1 , the management system starts

the second chiller to use the maximum available cooling capacity in
3.1. Efficiency freecooling mode. After 𝛥𝑡1 , it checks if both chillers are working
in freecooling mode. If that is not the case, the system recognizes
Since this paper concerns the analysis of data obtained from a real that freecooling at 𝑇1 is impossible and turns off one of the chillers.
installation, this knowledge is most valuable from the standpoint of Another attempt to switch the system into freecooling mode is made
the analysis of environmental conditions, the operation of individual at 𝑇2 and 𝑇3 after 𝛥𝑡2 at 𝛥𝑡3 in sequence. It is not possible to directly
components and the control system. One of the major indicators show- control the switching point of a single chiller between operating modes,
ing the efficiency of a cooling system is its 𝐶𝑂𝑃 . Its expansion [31] is which means that it is not possible to determine a specific switching
shown in Eq. (1). temperature. Chiller’s built-in controller is working on the basis of its
𝐨𝐮𝐭𝐩𝐮𝐭 𝐜𝐨𝐨𝐥𝐢𝐧𝐠 𝐜𝐚𝐩𝐚𝐜𝐢𝐭𝐲 [𝐤𝐖𝐡] (𝐜𝐨𝐨𝐥𝐢𝐧𝐠) own decision strategy and does not allow remote management. Time
𝐶𝑂𝑃 = (1) constants and hysteresis are introduced to guarantee the resistance of
𝐞𝐥𝐞𝐜𝐭𝐫𝐢𝐜𝐚𝐥 𝐩𝐨𝐰𝐞𝐫 𝐢𝐧𝐩𝐮𝐭 [𝐤𝐖𝐡] (𝐞𝐥𝐞𝐜𝐭𝐫𝐢𝐜𝐚𝐥)
the control system to outdoor temperature fluctuations.
3.2. Analysis Freecooling is the most desirable mode of operation in terms of
energy efficiency. As the outdoor temperature increases, however, the
The determination of the COP coefficient was possible thanks to cooling capacity of the chillers decreases, until any of the chillers
the available actual process data: the cooling capacity of the cooling decides to switch on the compressor. Switching the compressor on by
system, the electric power of installed chillers. The analysis of the any of the chillers sends information to the BMS system about the
data acquired used basic statistical methods to determine the minimum need to switch the cooling system to compressor mode. This is done
values 𝑥𝑚𝑖𝑛 , maximum values 𝑥𝑚𝑎𝑥 , mean 𝑥̄ and standard deviation by switching off one of the chillers. A single working chiller continues
𝜎. Linear trends were approximated using regression graded by the to work in compressor mode.
quality factor 𝑅2 . The quality of approximation by analytical function
was measured by mean squared error (MSE). The analyzed temperature 4. Design and development of cooling system modernization
data and operating modes of the cooling system were characterized by
Gaussian distributions in single- and multi-modal data in the general When designing the cooling system modernization of the POLCOM
form (2) and (3). Data Center, two assumptions were made. The system should be energy
The values of the coefficients were obtained through global para- efficient and, due to the favorable temperatures in the autumn–winter–
metric optimization using genetic algorithm configured to 50,000 gen- spring period, use the option of freecooling mode. It should also have
erations and function tolerance set to 10−12 under the 𝑀𝑆𝐸 quality a low maintenance requirement. With regard to the energy efficiency
indicator. The obtained analytical form of distribution presents the criterion, the decision was made to chose the external chillers with the
peak values that are useful for further studies on temperature and freecooling function as the primary source of cooling for the server
chillers operation prediction. room. Alternatively, the use of cooling towers was considered, but
( ) abandoned due to the higher maintenance requirements. Given the
1 −(𝑥 − 𝜇)2
𝑓𝜇,𝜎 (𝑥) = √ 𝑒𝑥𝑝 (2) limited time between the design phase and commissioning of the cool-
𝜎 2𝜋 2𝜎 2
ing system, the decision was made to use off-the-shelf cooling system
∑𝑛
components, which enabled the system to be commissioned within the
𝑓𝑛 (𝑥) = 𝑤𝑖 𝑓𝜇𝑖 ,𝜎𝑖 (𝑥) (3)
𝑖=1 required timeframe.
The installation of two new, high-efficiency chillers was realized
3.3. Custom control system and two low-efficiency reserve chillers were removed from the cooling
system architecture.
A single chiller can work in two modes. The first is the mode To sum up, the Data Center, after modernization, has four chilled
of operation based on screw compressors circulating freon refriger- water chillers: new external chillers – 2 pcs, old internal chillers – 2
ant (compressor operation). The second mode involves bypassing the pcs. The simplified topology of the cooling system is presented in the
compressors; the production of cooling power is done by discharging diagram in Fig. 2. Blue lines mark the pipe transporting cold chilled
heat from chilled water to the air using water/air exchangers and fans water from cold sources to the server chambers. The red line marks the
generating air flow through these exchangers (freecooling mode). A pipe transporting hot water, which, after heating in server chambers,
pair of chillers (system) can, in turn, work in two basic modes of returns to the chillers in order to cool down.
operation:
4.1. Requirements for the modernized cooling system
• freecooling — where one or both chillers work in freecooling
mode,
Based on the historical data from the year 2014 detailed in Sec-
• compressor — where one or both chillers work in compressor
tion 2.1 and the business requirements, the indicators that the cooling
mode.
system should have after the modernization were determined.
The master building monitoring system (BMS) communicates with The COP coefficient for the year 2014 for the cooling system
each device, but its operation is limited to collecting information about amounted to 4.12, with 3.48 in May to the end of September and 5.03
the operating parameters of individual units in order to present them for the rest of the year. The minimum value of COP in May–October
to the building’s technical staff. The only exception is the BMS system’s was 1.57, and 1.85 for the rest of the year. The maximum COP during
decision to start or stop chiller units (see Fig. 2). Therefore, the master the year was 10.04, while in the period May to the end of September
system indirectly attempts to obtain the appropriate operating mode 8.89. It was assumed that the COP of the modernized cooling system
of a single chiller (compressor or freecooling mode). This is done by should be 50% higher than that of the original cooling system for the
splitting in real time the demand for cooling power into one or two year 2014 (when it amounted to 4.12), so in the year 2015 it should
chillers at a given moment. The goal is to maximize working time in have not been less than 6.18.
freecooling mode (work with full cooling capacity on both chillers in An analysis of the temperature distribution in the year 2014 was
freecooling mode). performed, obtaining the parameters shown in Table 4 for the math-
Let us study a custom control system operation (see Fig. 2). Let ematical model (3). The dominant temperature ranges −1÷1 ◦ C and
us assume a single chiller works during compressor operation. If the 9÷12 ◦ C, the asymmetry of the distribution with the predominance of

3
M. Borkowski and A.K. Piłat Applied Energy 306 (2022) 117975

Fig. 3. Temperature histogram in 2014.

Fig. 4. Chillers cooling power 𝑃 in compressor (CP) and freecooling (FC) modes. On
the courtesy of manufacturer.

4.2. Refrigeration chillers after modernization

Two external chillers based on two twin-screw compressors (R134a

refrigerant type) named ‘‘Primary external chillers’’ (see Fig. 2) were
chosen as the new cooling units. They have a freecooling function. They
are controlled by their own microprocessor controller. According to the
catalogue card, they are characterized by the nominal operating pa-
rameters: electrical power consumed: 149 kW; cooling power: 442 kW;
COP: 3.12; in the following conditions: water temperature 15/10 ◦ C,
outdoor temperature 30 ◦ C, glycol 20%. In freecooling mode cooling
power is 220 kW at outdoor temperature 1 ◦ C. The characteristics of
cooling power obtained from the manufacturer are presented in Fig. 4.
Internal chillers with an external Freon condenser (R407c refrigerant
type) already installed on site remained as back-up. They have four
screw compressors, working in two independent hydraulic circuits and
are controlled by their own microprocessor controller. According to the
catalogue card, it is characterized by the nominal operating parameters:
electrical power consumed: 44 kW; cooling power: 129 kW; COP: 2.96
Fig. 2. Cooling and control systems diagram. in the following conditions: water temperature 12/7 ◦ C, condensing
temperature 50 ◦ C. Summarizing, the modernization resulted in the
increased performance by 664 kW of cooling capacity, giving 1142 kW
positive temperatures, which affect the operation of the cooling system, in total (growth of 140%).
were noted. Assuming that the temperature histogram would be similar
in the following 2015, it was estimated that with a fixed switching 4.3. Built-in chiller controller
point determined at 6 ◦ C, the system would operate annually in the
freecooling mode 38% and in compressor mode 62% of the time (see The built-in chiller’s control system is decentralized; there is no mas-
Fig. 3). ter control unit for all the cooling system components. Each chiller has
its own built-in controller that manages the parameters of a single unit.
The required capacity of the cooling system after the modernization
Typically, cooling system components are supplied by external man-
should have a minimum of 469 kW (see Section 2.1). To ensure greater
ufacturers. They are characterized by closed hardware and software
accuracy of measurement data, the resolution of recorded process data architecture. In particular, there is no information on the implemented
should be increased from 30 min to 5 min. The control system should be control (operation) algorithm. To some extent, chiller’s controllers pro-
able to resist significant 24-h fluctuations in outside temperature. The vide information about their work in the form of monitoring variables.
requirements described in this paragraph were the basis for designing The limited number of parameters selected by the manufacturer can be
the modernization of the cooling system (new chillers and control changed locally or remotely by a master control system (e.g. setpoint).
strategy). The chiller’s control algorithm implemented by the given controller

4
M. Borkowski and A.K. Piłat Applied Energy 306 (2022) 117975

Fig. 5. Outdoor temperature in 2015. Fig. 6. Temperature histogram in 2015.

Table 1
is unknown. This is to guarantee reliable operation of the device. Statistical values of the outdoor temperature in individual seasons of the year 2015.

However, this is a serious limitation for the master control system, Season Temperature [𝑜 C]

restricting its flexibility. 𝑇𝑚𝑖𝑛 𝑇𝑚𝑎𝑥 𝑇̄ 𝜎

Spring 20.03 ÷ 21.06 −5.9 33.9 11.1 6.9
4.4. Custom control system Summer 21.06 ÷ 23.09 2.3 36.2 18.2 6.4
Fall 23.09 ÷ 22.12 −7.4 23.3 5.5 5.0
Winter 22.12 ÷ 20.03 −13.5 15.1 1.2 4.5
The real-time data acquisition system was configured to store all
Annual −13.5 36.2 9.1 8.6
available measurements and control signals at 5 min intervals, which
solved the problem of data consistency, ensured that time stamps were
recorded, and guaranteed that the entire state of the Data Center was
recorded for later analysis, both in terms of efficiency and the course of Table 4. The temperature below 6 ◦ C occurred 41% of the time during
any failures. The recorded data is an invaluable source of information, 2015 and temperatures above 6 ◦ C occurred 59% of the time. The
especially as it comes from the operation of an actual commercial hypothetical operating time of the cooling system in freecooling mode
cooling system. Based on the collected process data, the constraints should have therefore been 3% higher than the 38% assumed on the
related to the dynamics of chillers switching and the time required to basis of the 2014 temperature distribution and should have amounted
stabilize the operation of the cooling system, the time constants of the to 41%.
control algorithm 𝛥𝑡1 and 𝛥𝑡2 were proposed to be set for a period of
15 min. The temperature hysteresis points of the algorithm were set at
5.1. Analysis of the outdoor temperature in individual months and seasons
𝑇1 = 8 ◦ C, 𝑇2 = 7 ◦ C and 𝑇3 = 6 ◦ C respectively. The differences between
the hysteresis points 𝑇1 , 𝑇2 and 𝑇3 , determined on the temperature
In Table 1 the statistics of outdoor temperature for particular sea-
variability index shown in Table 2, were 1 ◦ C.
sons are presented. A very high gradient in the temperature change is
observed. It makes a very demanding conditions for the cooling system
5. Research on outdoor temperature
and the control algorithm. Summer is the most demanding season of the
year for the cooling system. This is due to the large temperature range
The outdoor temperature sensor was located on the roof of the
as well as the highest maximum temperatures during the year. In turn,
building in a shaded place. This place was chosen because there was
winter is the least demanding, during which temperatures are at the
no influence of outside factors on temperature measurement. It was
lowest level and have the smallest spread compared to other seasons.
characterized by a large airflow, and at the same time there were no de-
vices nearby that generated or emitted heat. The information indicated The temperature distribution in January and September is notewor-
in Section 2.1 showed that for the proper functioning of the cooling thy (see Table 2). The distribution in January is characteristic because
system, temperature measurement is a key source of data. Solving the of the smallest difference between the first and third quartiles com-
login problem resulted in the number of temperature measurements pared to other months. On the other hand, it has a significant amount
reaching 105,120 points per year. Measurement in 5-minute intervals of outliers, both upper and lower. In turn, September is characterized
allowed for a detailed analysis of changes in outside temperature. by the highest number of upper outliers. Typical summer months
This data is presented in Fig. 5, and statistical values are listed (July and August) have no outliers at all. The analysis of temperature
in Table 4. The increased measurement resolution resulted in a more values showed that there are days with similar temperature profiles
accurate temperature distribution. As in the year 2014, the domi- throughout the year.
nant temperature ranges of approx. 0÷3 ◦ C and 8÷11 ◦ C, the distri- It can be observed that in the period June–September the value of
bution asymmetry with predominantly positive temperatures can be the outside temperature is higher than 6 ◦ C. Based on the temperature
distinguished. analysis, it is expected that the freecooling mode will not be used
It is worth emphasizing the high temperature fluctuation even during this period.
within a single month. In turn, the outdoor temperature histogram The variability of outside temperature at the location of Data Center,
(Fig. 6) illustrates the time share of individual temperature values calculated as average value of maximal gradients observed daily at
throughout the year. 15 min intervals, is shown in the last column of Table 2. It was
The fitted curve, which is a combination of four Gaussian distribu- decided to set the temperature hysteresis points 𝑇1 , 𝑇2 and 𝑇3 at 1 ◦ C
tions described by Eq. (3), is marked in red. The values of coefficients (see chapter 4.4) used by control algorithm. One can find that further
for individual distributions are presented in Table 4. The 𝑀𝑆𝐸 of cooling system performance and controllability can be increased by the
the fitted curve is 0.988. Now compare this distribution with the dynamical determination of temperature variability and adaptation of
corresponding operating mode distributions presented in Fig. 10 and in hysteresis width.

5
M. Borkowski and A.K. Piłat Applied Energy 306 (2022) 117975

Table 2
Statistical values of the outdoor temperature in individual months of the year 2015.
Month Temperature [𝑜 C]
𝑇𝑚𝑖𝑛 𝑇𝑚𝑎𝑥 𝑇̄ 𝜎 variability over
15 minutes
January −13.5 13.8 3.8 3.8 1.1
February −9.1 13.4 4.3 4.3 1.1
March −5.9 19.7 4.8 4.8 1.4
April −2.7 24.3 6.3 6.3 1.4
May 0.5 26.6 4.7 4.7 1.2
June 6.0 33.9 5.5 5.5 1.3
July 6.6 34.8 5.9 5.9 1.2
August 6.0 36.2 6.6 6.6 1.3
September 2.3 35.9 5.4 5.4 1.3
October −2.8 23.3 4.9 4.9 1.1
November −7.4 16.4 4.9 4.9 1.1
December −11.6 14.5 4.3 4.3 1.2

6. Research and discussion on cooling system operation

6.1. Verification of custom control system operation

The verification of the control system was conducted on the basis of

data recorded for the period of the year 2015. The failure-free operation
of the cooling system over the mentioned year is a confirmation of
the correctness of the operation of the implemented control algorithm.
For a detailed analysis of the control implementation by the algorithm,
23 days with a similar temperature profile were observed. Two repre-
sentative dates were chosen: 23-Feb-2015 and 8-Oct-2015. Both days
were characterized by moderate temperatures, without extremely low
and high values, and a similar time course. The minimum (0.6 ◦ C,
0.4 ◦ C) and maximum (13.4 ◦ C, 13.7 ◦ C) temperatures, mean values
(5.9 ◦ C, 6.0 ◦ C) and standard deviation (4.0, 4.3) are similar for both
days. Based on the convergence of these parameters, it was determined
that the comparison of the chillers’ operation on both days would be
relevant, since the impact of the outdoor temperature was minimized.
Fig. 7. Day analyzed 23-Feb-2015 and 8-Oct-2015..
In Fig. 7 the red and blue dots mark the moments of transition and
the cooling system exiting the freecooling mode. The legend for these
points includes the time of day and the outdoor temperature, respec-
tively. Taking into account the profile of electric power consumption the control system started the second of the chillers. It also started
by chillers, a point of entering the freecooling mode was set at 6:10 at working in freecooling mode. Both chillers worked in freecooling mode
0.7 ◦ C on 8-Oct-2015, and a point of exit at 10:35 at 7.1 ◦ C. In turn, simultaneously, consuming approx. 45 kW of electric power. At 11:45,
on 23-Feb 2015, the entry point for freecooling mode took place at one of the chillers was switched to compressor work mode. The control
22:55 at 4.9 ◦ C and the exit point at 10:35 at 8.6 ◦ C. Therefore, for system shut down the second chiller. The single chiller worked in
both days the temperature of entering freecooling mode was lower than compressor mode until 22:55. At this time, the only running chiller
the temperature of leaving this mode. The working time in freecooling switched to freecooling mode.
mode on 23-Feb-2015 was 11 h and 10 min while on 8-Oct-2015 it was To achieve better performance, the control system should have
4 h and 25 min. started the second chiller at 18:20, since they would be able to work
It should be noted that chillers should work in freecooling mode for in freecooling mode from then on. A necessary prerequisite, however,
the entire period for which the outdoor temperature was lower than the would be to ensure that the controllers of both chillers switch them
temperature of the freecooling exit point (according to the property that to freecooling mode. Based on historical observations, it is concluded
the lower the outdoor air temperature, the greater the cooling capacity that the chillers’ internal controllers switch them to freecooling mode
of the chiller working in freecooling mode). The maximum possible only when the cooling capacity reserve is sufficiently high. There-
working time of the cooling system in freecooling mode was therefore fore, the control system did not start the second chiller early enough,
16 h and 15 min on 23-Feb-2015 (until 11:45 and from 18:20) and 14 h because it was not certain that both chillers would work in freecool-
and 35 min on 8-Oct-2015 (until 10:35 and from 20:00). ing mode. On 8-Oct-2015, one chiller initially worked in compressor
The cooling capacity changed in a similar way during each day. mode, consuming over 80 kW of electric power. Then at 6:10, the
Performance changes on individual days were small, below 4% of the control system started the second of the chillers. It started working
day’s average and therefore could be omitted. Thus, the average cooling in freecooling mode. Meanwhile, the unloaded first chiller switched
capacity on 23-Feb-2015 was 336 kW and on 8-Oct-2015 – 444 kW. The to freecooling mode. Both chillers then worked in freecooling mode,
increase in average performance was 108 kW (32%). Such increase in consuming approx. 30÷51 kW of electric power. At 10:35, one of the
the cooling capacity of aggregates resulted in a 111% increase in power chillers went into compressor mode. The control system shut down the
consumption. Meanwhile, the average COP fell by 37% (from 8.37 to second chiller. The single chiller worked in compressor mode until the
5.30) (see Table 3). end of the day.
On 23-Feb-2015, one chiller was initially working in freecooling In this case, the control system worked correctly; however, due to
mode, consuming less than 30 kW of electric power. Then at 9:40, the inability to steer the operation mode of individual chillers directly,

6
M. Borkowski and A.K. Piłat Applied Energy 306 (2022) 117975

Table 3
COP profile parameters.
Data 23-Feb-2015 8-Oct-2015
𝐶𝑂𝑃𝑚𝑖𝑛 4.08 3.25
𝐶𝑂𝑃𝑚𝑎𝑥 11.94 13.55
̄
𝐶𝑂𝑃 8.37 5.30
𝜎 2.94 2.59

Fig. 9. Chiller’s COP related to outdoor temperature and cooling capacity.

outdoor temperature, the control system worked as intended and the

cooling system operated in a stable manner. Despite the limitations due
to the closed architecture of the chillers and the lack of direct control, it
was proven that indirect control of the chillers worked well in practice
in a commercial cooling system.
Fig. 8. Demand for cooling power.

6.3. Operating specifics of external chillers

switching on the second chiller and thus the transition to freecooling Historical data for 2015 do not contain information on the operation
mode took place at a much lower outdoor temperature compared to the mode of the cooling system. This had to be estimated based on the
maximum temperature recorded when the system was in freecooling analysis of data on electric power consumption by each of the chillers.
mode. The average daily power consumption increased from 46 kW to For this purpose, chillers were started in conditions of maximum elec-
97 kW, i.e. by 51 kW (111%). tric power consumption in freecooling mode. The amount of electrical
power consumed was measured. The maximum power consumption of
6.2. Cooling system performance a single chiller in freecooling mode was 30 kW.
Based on historical observations, it was assumed that the minimum
The operation of the Data Center is a dynamic process, due to the
power consumption of a single chiller working in compressor mode was
variable demand for cooling power resulting from updated IT config-
clearly higher than the maximum power consumption of one chiller
uration and the variable load related to the use of individual servers.
working in freecooling mode. Based on this, it was determined that the
Over the course of the year, the amount of IT hardware installed in
power consumption of a single chiller of up to 30 kW meant that it
the server room increased. Thus, the demand for cooling capacity to be
worked in freecooling mode, and above 30 kW — that it worked in
realized by chillers increased. The profile of the change in demand for
compressor mode.
cooling power is shown in Fig. 8. At the beginning of the year 2015,
Fig. 10 shows the operating time of the cooling system in a given
the trend of demand for cooling capacity was 425 kW, and at the end
operating mode in relation to the outdoor temperature. There are two
of the year it reached 562 kW. Thus, demand increased by 63% over
distributions corresponding to compressor and freecooling work. Their
the course of the year. Trend was approximated using linear regression
analytical representation (3) was computed and obtained parameters
(Offset of regression was 425, slope of regression was 1.413 × 10−3 and
are listed in Table 4. Peak values corresponds to temperatures of 0.6 ◦ C
𝑅2 = 0.872).
Based on the 2014 data, it was estimated that at the end of 2015 and 11.0 ◦ C. The freecooling mode is characterized by symmetrical
the cooling capacity demand was going to reach 469 kW. In reality it distribution, while compressor mode asymmetrical one due to asymme-
reached 562 kW, 93 kW more than anticipated. The rate of increase try of outdoor temperature profile for warm days. Both distributions
was higher than expected. overlap. The reason is the hysteresis of switching the cooling system
The cooling system’s demand for cooling capacity is closely related between modes, as well as the increasing demand for cooling capacity
to the amount of IT equipment installed by POLCOM customers and throughout the year. One can find that the knowledge of tempera-
presents the actual dynamics of business processes. ture model can be used for prediction of cooling system operation.
The increase in demand for cooling power has the greatest impact Moreover, the distribution parameters show week dynamics in year to
on the chiller’s COP factor at outdoor temperature range below 10 ◦ C in year outdoor temperature fluctuations, but strong correlation of cooling
freecooling mode (see Fig. 9). The reason is the relatively low efficiency system operating modes to temperature confirm the correct operation
of chillers in freecooling mode. Moreover, the demand for cooling of the cooling system.
capacity increases above the freecooling capacity of a single chiller The actual operating time of the cooling system in freecooling mode
during the year. Therefore, with increased demand for cooling power, amounted to 35% and was 3% less than the hypothetical operating time
it is necessary to run both chillers in freecooling work. assumed. Taking into account the different temperature distribution
The control system does not work properly in this case, as it does not between the years 2014 and 2015, the freecooling operating time was
switch chillers in freecooling mode at favorable temperature conditions. even 6% less than the hypothetical operating time of 41% updated on
With a perfectly working control system, it would be possible for the the basis of the 2015 temperature data.
chillers to work longer in freecooling mode and thus conserve electric- It is important to note that it is the chiller’s internal controller, not
ity. The reason for this is that direct control of the operating mode of the control system that determines whether the chiller is taken out
the chillers is not possible. Based on the tests and analyses carried out, of freecooling mode and switched to compressor mode. Direct control
it is concluded that the custom control algorithm works correctly in of the switching is not possible due to the closed architecture of the
a dynamic outdoor environment. Despite the occurring fluctuations in chiller.

7
M. Borkowski and A.K. Piłat Applied Energy 306 (2022) 117975

Fig. 10. Histogram of freecooling/compressor work. Fig. 11. Annual cooling power demand.

Table 4
Gaussian distribution coefficients.
𝑛 𝑇2014 𝑇2015 FC+CP FC CP
4 4 4 2 2
𝑤1 107.676 1947.300 1412.910 2404.520 1292.277
𝜇1 −0.353 0.541 0.851 0.394 8.637
𝜎1 1.586 4.149 2.673 4.219 −4.976
𝑤2 378.253 789.433 2027.480 1300.582
𝜇2 9.453 9.096 8.363 15.149
𝜎2 11.874 4.261 9.850 −9.745
𝑤3 175.608 503.713 696.471
𝜇3 2.427 −22.929 −2.412
𝜎3 1.197 0.521 1.701
𝑤4 16.150 1522.979 448.489
𝜇4 49.999 12.317 20.289
𝜎4 1.628 11.907 10.593 Fig. 12. Annual outdoor temperature.

The reason for the too short, in relation to the expected operation corresponding to the following situations: single chiller operation in
time in freecooling mode, was therefore the lack of possibility to freecooling mode, operation of two chillers in freecooling mode, work
directly control the selection of the cooling mode for the chillers. The in compressor mode. The following data samples were analyzed: electri-
deficiency of the only indirect control available in this configuration, cal power absorbed by chillers, cooling capacity, outdoor temperature
which consists only of sharing the cooling capacity between the chillers, and the chillers’ coefficient of performance (COP).
is apparent. The analyzes carried out later in the paper showed that the Fig. 11 illustrates the efficiency of the cooling system with the
cooling system in freecooling mode is possible at temperatures below separation of the chillers’ operating modes meeting the desired demand
about 7÷11 ◦ C (depending on the system load, compare to Fig. 16). for cooling power (compare with Fig. 8). Operation of one chiller in
freecooling mode was possible until the second half of March. This was
The temperature histogram shows that such temperatures occur more
due to the correlation between low demand for cooling capacity and
often during the year than higher temperatures. The most common
low outdoor temperatures (see Fig. 12). Over time, the demand for
temperatures during the year are temperatures between 0 ◦ C and 2 ◦ C.
cooling capacity increased, which prevented the cooling system from
Therefore, the climatic conditions of the server room are favorable for
working on one chiller in freecooling mode. In the middle of the year
using freecooling as a way of cooling the server room.
(May till the end of September), the cooling system worked mainly in
Should component manufacturers provide full control over the
compressor mode. The increase in demand for cooling capacity was
equipment fitted with such advanced cooling systems? From the man-
relatively small in this period (by 10%, from 400 kW to 440 kW), and
ufacturer’s standpoint, this is obviously risky, but from the point of
there was no significant impact of such a change on the COP value
view of the server room operator, this is desirable. In other words,
(Fig. 14). At the end of the year, when the outdoor temperatures had
the automation system of the device delivered should be equipped
returned to around a few degrees Celsius, it was necessary to work on
with the ability to safely reconfigure parameters through an external both chillers in freecooling mode to ensure the production of cooling
master control system, with an open form of the control algorithm. The capacity at the appropriate level.
chillers are designed as universal devices. However, based on the data
analyzed, it was found that the specifics of the server room’s cooling 6.4.1. Electrical power consumption
system prefer an individual control mode, going beyond the available Fig. 13 shows the total power consumption of the chillers over 2015.
functionality. The analyzes show that the chillers delivered have good The impact of the operating mode on the amount of electrical power
functionality; however, the efficiency of work in the entire cooling consumed is clearly visible. The lowest power consumption is assigned
system could be higher. to the freecooling operation of one chiller.
With reference to the power consumption described in Section 6.3,
6.4. Impact of the change in demand for cooling capacity of chillers on their the power consumption for the two chillers operating in freecooling
operating efficiency — based on the analysis of annual work ranged from 30 kW to 60 kW. The mode with the highest demand for
electric power was the compressor mode, where the power consump-
Apart from outdoor air temperature, the demand for cooling ca- tion reached 250 kW. The compressor mode was also characterized by
pacity is a major factor affecting the operating conditions of chillers. the largest amplitude of changes in electric power consumption (190
To illustrate this relationship (see Figs. 11÷15), an appropriate data kW), where it was 10 kW in freecooling mode on one chiller, and 20
analysis was made. It involved the separation of measuring points kW on both chillers.

8
M. Borkowski and A.K. Piłat Applied Energy 306 (2022) 117975

Table 5
Cooling system work time in individual operating modes.
Operating mode Working time COP
[days] [%] mean [%]
Compressor 236 65 4.39 100%
Freecooling
single chiller 69 19 17.69 403%
Freecooling
both chillers 60 16 15.09 344%

Fig. 13. Total power consumption of chillers.

Fig. 15. Operation modes of the cooling system.

Fig. 14. Annual COP.

6.4.2. Performance coefficient

The evolution of the COP throughout the year is shown by the
two envelopes covering weekly minimums and maximums of the COP
throughout the year. The operating time and efficiency data of the
cooling system in the individual operating modes are shown in Table 5.
Please note that the COP in freecooling mode was 303% and 244%
greater than in compressor mode. The increase of the COP coefficient
over time is visible in freecooling mode with both chillers working (see Fig. 16. Daily maximum outdoor temperatures in freecooling mode.

Fig. 14). This is due to the fact that the chiller consumes part of its
electrical power for ‘‘idle’’ work not directly related to the production of
cooling energy. Mainly it involves pumping glycol through circulation 6.5. Remarks on control
pumps. Regardless of what load the chiller works with, the circulation
pumps must be started and consume the same amount of electrical In order to estimate the maximum achievable working time in
freecooling mode at the temperature conditions observed in 2015, the
power. In the tested chiller, this was about 20 kW. The element that
maximum outdoor temperature was established, at which freecooling
regulates its efficiency in freecooling mode are fans, which consume
mode work took place on each day of the year in which the system went
up to 10 kW of electrical power, i.e. only 33% of the maximum total
into compressor mode (see Fig. 16). Trend was approximated using
power consumed by a chiller working in freecooling mode. Hence, by linear regression (offset of regression was 12.7, slope of regression was
increasing the chillers’ load with thermal power, the power consump- −12.621 × 10−3 and 𝑅2 = 0.638). Based on the analysis of maximum
tion of the fans is only increased. Therefore, increasing the load on the temperatures, it was assumed that the maximum outdoor temperature
chiller in freecooling mode causes a simultaneous increase in its COP at which it is possible to work in freecooling mode is determined by
(the effect of the outdoor temperature should be ignored here). a straight line connecting points at 12.7 ◦ C at the beginning of the
It was also observed that the system achieved the highest effi- year and 8.3 ◦ C on the last day of the year. Many of the maximum
ciency with one chiller working. The reason was that only the circu- temperatures significantly deviate from the values resulting from the
lation pumps in one chiller worked. The pumps of the second chiller designated straight line. Values below the straight line represent losses
were turned off and therefore did not increase the electric power during freecooling operation caused by the cooling system leaving
the freecooling mode when the outdoor temperature is too low. The
consumption.
decrease in the maximum temperature at which the cooling system
Fig. 15 illustrates the scope of the cooling system’s work in in- worked in freecooling mode results from the increase in demand for
dividual modes of operation in relation to the outdoor temperature cooling power (which is presented in Fig. 8).
and the demand for cooling power. The increase in the demand for Starts during freecooling operation translated into losses associated
cooling power causes a decrease in the maximum outdoor temperatures with increased electricity consumption. The average power consump-
at which the cooling system works in freecooling mode. It also limits tion for freecooling operation (single and both units), for the outdoor
the possibility of working on a single chiller in freecooling mode. temperature ranged from 3 ◦ C to 12 ◦ C (temperature range for which

9
M. Borkowski and A.K. Piłat Applied Energy 306 (2022) 117975

amplitudes of up to 25 ◦ C. Knowing the temperature distribution, the

changes in demand, it is possible to predict the operation of the cooling
system in the forthcoming year.
Based on the research, it was determined that the outdoor tem-
peratures at which the system could operate in the freecooling mode
occurred for 188 days, which constituted 52% of the operating time
during the 2015 year. In fact, the freecooling system only covered
129 days. The working time of the cooling system in the freecooling
mode amounted to 129 days, and in the compressor mode 236 days per
year. This corresponded to 35% and 65% of working time, respectively.
During freecooling operation, the system operated on both chillers
for 60 days and on one chiller for 69 days. Failure to fully utilize
the time available during the year for freecooling operation resulted
in a loss in electricity consumption of 92 MWh, representing 12%
of the total energy consumption of the cooling system. The average
Fig. 17. Hysteresis of cooling system operation modes.
COP in compressor mode amounted to 4.39, in freecooling mode with
one chiller running 17.69, and in freecooling mode with two chillers
running 15.09. In freecooling mode it was therefore 303% and 244%
the freecooling mode was exited too early) amounted to 43 kW. Aver- greater than the COP in compressor mode, respectively.
age power consumption for the same outdoor temperature range but
Peak to peak values of the electric power consumed in compressor
for compressor operation was 108 kW. The difference in electric power
mode was 19 times greater than that of the freecooling mode. The
consumption was therefore 65 kW on average. For the time range of
maximum electric power consumed by the cooling system was 250 kW;
59 days, it gave 92 MWh of additional electricity consumed per year.
the minimum power in the freecooling mode was 30 kW.
The annual electricity consumption of the cooling system was 800
The control algorithms built into the factory controllers of chillers
MWh. If the switching algorithm worked perfectly (the outputs from
were pointed to as the cause of the limited system operation time in
the freecooling mode would coincide with the red line in Fig. 16), the
the freecooling mode.
annual electricity consumption would be 708 MWh. So it would be 12%
lower than the actual consumption. The designed cooling system, combined with a dedicated control
This regularity is also visible in Fig. 17, showing the outdoor system, achieved the COP of 8.63 in the year 2015. This coefficient
temperature range for the cooling system’s work in individual operating represents a 110% increase in efficiency compared to the year 2014 (see
modes. It shows the operation of the cooling system in compressor Section 4.1). It generated losses in electricity consumption amounting
mode from an outdoor temperature of −4.2 ◦ C. The data analysis (see to 557 MWh per year. During the period of dominant compressor
Fig. 16) showed that compressor operation at such low temperatures operation (May to the end of September), the COP varied between 1.73
was not necessary, because it would be possible to operate the system and 16.71 giving an average of 4.62 and for the rest of the year between
in the economical freecooling mode. Due to the chiller’s internal control 2.56 and 21.18 respectively with an average of 11.83. As a conclusion
algorithm, the areas of working modes overlapped. from the discussion presented, the solution proposed is to change the
The analysis showed that the time when the temperature throughout approach to the design of cooling systems to have a full control over
the year was lower than the values determined by the straight line in the operation of chillers. This should be provided by manufacturers or
Fig. 16 was 188 days (52%). by developing a dispersed cooling system with individually selected
Therefore, it can be stated that with a fully effective control system components. A master control algorithm should be designed for the
(if the chiller controller allowed for direct control of its operating operation of the cooling system in a custom way, taking into account
modes and thus it would be possible to extend the strategy of superior the time of year, day, outdoor climatic conditions and measurements in
control), the working time in freecooling mode should also amount to the server chambers, with particular emphasis on direct control of the
188 days. So it should be 59 days longer than the time the system really process of switching the system between operating modes. Moreover,
achieved. the extended control system should react to temperature changes on an
The reason the cooling system did not achieve the theoretically pos- ongoing basis, while keeping in mind the time of year so as to include
sible freecooling operating time was the inability to directly set chillers the trend of changes.
operating mode. To achieve better controllability, the cooling system It is believed that the presented real experimental data and con-
architecture should be modified by replacing embedded chillers con- clusions related to commercial installation operation will be useful
trollers with units allowing full control. It will also rise the possibility to for researchers and manufacturers interested in the field of cooling
maximize freecooling operation time. One can consider maximization systems, as well for Data Center designers due to specific climate
freecooling operation through application of adiabatic cooling, keeping conditions.
in mind advantages and disadvantages of this solution. Further study conflictofinterest The authors declare that they have no known
on multi-objective optimization [32] can boost the performance and competing financial interests or personal relationships that could have
efficiency of the cooling system and save the energy. appeared to influence the work reported in this paper.

7. Conclusions
CRediT authorship contribution statement
The POLCOM cooling system was upgraded by the installation of
chillers allowing operation in freecooling and compressor modes. The Mateusz Borkowski: Conceptualization, Methodology, Software,
custom control system designed and implemented to operate in real Validation, Formal analysis, Investigation, Resources, Data curation,
conditions met the design assumptions and correctly controlled the Writing – original draft, Writing – review & editing, Visualization,
operation of chillers. The cooling system, due to the location of the Data Funding acquisition. Adam Krzysztof Piłat: Conceptualization,
Center in specific climatic conditions, required effective operation with Methodology, Formal analysis, Resources, Data curation, Writing –
annual amplitudes of the outdoor temperature about 50 ◦ C and weekly original draft, Writing – review & editing, Visualization.

10
M. Borkowski and A.K. Piłat Applied Energy 306 (2022) 117975

Acknowledgments [15] Hamann H. A measurement-based method for improving data center energy effi-
ciency. In: 2008 IEEE International Conference on Sensor Networks, Ubiquitous,
and Trustworthy Computing. 2008. p. 312–3.
Research was funded by Polish Ministry of Science and Higher
[16] Zhou F, Li C, Zhu W, Zhou J, Ma G, Liu Z. Energy-saving analysis of a case data
Education (MNiSW) Project No. 0014/DW/2018/02 and carried out in center with a pump-driven loop heat pipe system in different climate regions
cooperation of POLCOM and AGH University of Science and Technol- in China. Energy Build 2018;169:295–304. http://dx.doi.org/10.1016/j.enbuild.
ogy; Faculty of Electrical Engineering, Automatics, Computer Science 2018.03.081.
and Biomedical Engineering. [17] Deymi-Dashtebayaz M, Valipour Namanlo S, Arabkoohsar A. Simultaneous use
of air-side and water-side economizers with the air source heat pump in a
data center for cooling and heating production. Appl Therm Eng 2019;161.
References http://dx.doi.org/10.1016/j.applthermaleng.2019.114133.
[18] Cheung H, Wang S. Optimal design of data center cooling systems concerning
[1] Tian H, He Z, Li Z. A combined cooling solution for high heat density data multi-chiller system configuration and component selection for energy-efficient
centers using multi-stage heat pipe loops. Energy Build 2015;94:177–88. http: operation and maximized free-cooling. Renew Energy 2019;143:1717–31. http:
//dx.doi.org/10.1016/j.enbuild.2015.03.002. //dx.doi.org/10.1016/j.renene.2019.05.127.
[2] Zhou F, Wei C, Ma G. Development and analysis of a pump-driven loop heat [19] Cho J, Woo J. Development and experimental study of an independent row-based
pipe unit for cooling a small data center. Appl Therm Eng 2017;124:1169–75. cooling system for improving thermal performance of a data center. Appl Therm
http://dx.doi.org/10.1016/j.applthermaleng.2017.06.108. Eng 2020;169:1–16. http://dx.doi.org/10.1016/j.applthermaleng.2019.114857.
[3] Lyu C, Chen G, Ye S, Liu Y. Enclosed aisle effect on cooling efficiency in [20] Durand-Estebe B, Le Bot C, Mancos JN, Arquis E. Simulation of a temperature
small scale data center. Procedia Eng 2017;205:3789–96, 10th International adaptive control strategy for an IWSE economizer in a data center. Appl Energy
Symposium on Heating, Ventilation and Air Conditioning, ISHVAC2017, 19–22 2014;134:45–56. http://dx.doi.org/10.1016/j.apenergy.2014.07.072.
October 2017, Jinan, China, http://dx.doi.org/10.1016/j.proeng.2017.10.132. [21] Habibi Khalaj A, Scherer T, Siriwardana J, Halgamuge SK. Multi-objective
[4] Cho J, Kim Y. Improving energy efficiency of dedicated cooling system and efficiency enhancement using workload spreading in an operational data center.
its contribution towards meeting an energy-optimized data center. Appl Energy Appl Energy 2015;138:432–44. http://dx.doi.org/10.1016/j.apenergy.2014.10.
2016;165:967–82. http://dx.doi.org/10.1016/j.apenergy.2015.12.099. 083.
[5] Jin C, Bai X, Yang C, Mao W, Xu X. A review of power consumption models of [22] Zhang H, Shao S, Tian C, Zhang K. A review on thermosyphon and its integrated
servers in data centers. Appl Energy 2020;265:1–18. http://dx.doi.org/10.1016/ system with vapor compression for free cooling of data centers. Renew Sustain
j.apenergy.2020.114806. Energy Rev 2018;81:789–98. http://dx.doi.org/10.1016/j.rser.2017.08.011.
[6] Castellazzi L, Maria A, Bertoldi P. Trends in data centre energy consumption [23] Zhang Y, Wei Z, Zhang M. Free cooling technologies for data centers: energy
under the European code of conduct for data centre energy efficiency. 2017, p. saving mechanism and applications. Energy Procedia 2017;143:410–5, Lever-
1–43. http://dx.doi.org/10.2760/358256. aging Energy Technologies and Policy Options for Low Carbon Cities, http:
[7] Capozzoli A, Primiceri G. Cooling systems in data centers: State of art and //dx.doi.org/10.1016/j.egypro.2017.12.703.
emerging technologies. Energy Procedia 2015;83:484–93, Sustainability in En- [24] Ma Y, Ma G, Zhang S, Zhou F. Cooling performance of a pump-driven two phase
ergy and Buildings: Proceedings of the 7th International Conference SEB-15, cooling system for free cooling in data centers. Appl Therm Eng 2016;95:143–9.
http://dx.doi.org/10.1016/j.egypro.2015.12.168. http://dx.doi.org/10.1016/j.applthermaleng.2015.11.002.
[8] Ebrahimi K, Jones GF, Fleischer AS. A review of data center cooling technology, [25] Yang Y, Wang B, Zhou Q. Energy saving analysis of free cooling system in the
operating conditions and the corresponding low-grade waste heat recovery data center. Procedia Eng 2017;205:1815–9, 10th International Symposium on
opportunities. Renew Sustain Energy Rev 2014;31:622–38. http://dx.doi.org/10. Heating, Ventilation and Air Conditioning, ISHVAC2017, 19–22 October 2017,
1016/j.rser.2013.12.007. Jinan, China, http://dx.doi.org/10.1016/j.proeng.2017.10.239.
[9] Yang Y, Wang B, Zhou Q. Energy saving analysis of free cooling system in the [26] Amado E, Schneider P, Bresolin C. Free cooling potential for Brazilian data
data center. Procedia Eng 2017;205:1815–9, 10th International Symposium on centers based on approach point methodology. Int J Refrig 2021;122:171–80.
Heating, Ventilation and Air Conditioning, ISHVAC2017, 19–22 October 2017, http://dx.doi.org/10.1016/j.ijrefrig.2020.11.010.
Jinan, China, http://dx.doi.org/10.1016/j.proeng.2017.10.239. [27] Chu W-X, Wang C-C. A review on airflow management in data centers. Appl
[10] Zhang X, Lindberg T, Xiong N, Vyatkin V, Mousavi A. Cooling energy con- Energy 2019;240:84–119. http://dx.doi.org/10.1016/j.apenergy.2019.02.041.
sumption investigation of data center IT room with vertical placed server. [28] He W, Ding S, Zhang J, Pei C, Zhang Z, Wang Y, Li H. Performance optimization
Energy Procedia 2017;105:2047–52, 8th International Conference on Applied of server water cooling system based on minimum energy consumption analysis.
Energy, ICAE2016, 8–11 October 2016, Beijing, China, http://dx.doi.org/10. Appl Energy 2021;303:1–10. http://dx.doi.org/10.1016/j.apenergy.2021.117620.
1016/j.egypro.2017.03.581. [29] Zhang Q, Meng Z, Hong X, Zhan Y, Liu J, Dong J, et al. A survey on data
[11] Cho K, Chang H, Jung Y, Yoon Y. Economic analysis of data center cooling center cooling systems: Technology, power consumption modeling and control
strategies. Sustainable Cities Soc 2017;31:234–43. http://dx.doi.org/10.1016/j. strategy optimization. J Syst Archit 2021;119:1–17. http://dx.doi.org/10.1016/
scs.2017.03.008. j.sysarc.2021.102253.
[12] Cho J, Yang J, Lee C, Lee J. Development of an energy evaluation and design [30] Canova A, Laudicina G, Lazzeroni P, Perez-Mora N, Repetto M. Exploitation and
tool for dedicated cooling systems of data centers: Sensing data center cooling optimal management of free cooling for industrial refrigeration. Energy Convers
energy efficiency. Energy Build 2015;96:357–72. http://dx.doi.org/10.1016/j. Manage 2019;198:1–16. http://dx.doi.org/10.1016/j.enconman.2019.111815.
enbuild.2015.03.040. [31] Deng J, Wei Q, Qian Y, Zhang H. Does magnetic bearing variable-speed
[13] Carbó A, Oró E, Salom J, Canuto M, Macías M, Guitart J. Experimental and centrifugal chiller perform truly energy efficient in buildings: Field-test and
numerical analysis for potential heat reuse in liquid cooled data centres. Energy simulation results. Appl Energy 2018;229:998–1009. http://dx.doi.org/10.1016/
Convers Manage 2016;112:135–45. http://dx.doi.org/10.1016/j.enconman.2016. j.apenergy.2018.08.062.
01.003. [32] Gupta R, Asgari S, Moazamigoodarzi H, Down DG, Puri IK. Energy, exergy and
[14] Nadjahi C, Louahlia H, Lemasson S. A review of thermal management and computing efficiency based data center workload and cooling management. Appl
innovative cooling strategies for data center. Sustain Comput: Inform Syst Energy 2021;299:1–20. http://dx.doi.org/10.1016/j.apenergy.2021.117050.
2018;19:14–28. http://dx.doi.org/10.1016/j.suscom.2018.05.002.

11