Available online at www.sciencedirect.

com

ScienceDirect
Availableonline
Available onlineatatwww.sciencedirect.com
www.sciencedirect.com
Energy Procedia 00 (2017) 000–000

ScienceDirect
ScienceDirect
www.elsevier.com/locate/procedia

Energy
EnergyProcedia
Procedia140 (2017) 000–000
00 (2017) 447–457
www.elsevier.com/locate/procedia

AiCARR 50th International Congress; Beyond NZEB Buildings, 10-11 May 2017, Matera, Italy

Modelling and control of a free cooling system for Data Centers

The 15th International Symposium on District Heating and Cooling
Alessandro Beghia, Luca Cecchinatob, Giuseppe Dalla Manac, Michele Lionelloa,
Assessing the feasibility of using
Mirco Rampazzo a, the heat
*, Enrico Sistib demand-outdoor
temperature function
Department for a long-term
of Information Engineering,
a
University of Padova,district
via Gradenigo heat demand
6/B, I-35131 Padova, Italy forecast
b
Everbyte s.r.l., via Ugo Foscolo 3, I-35131 Padova, Italy
c
Vertiv, via
a Leonardo da Vincia 16/18, I-35028 Piove di Sacco, Italy
I. Andrića,b,c*, A. Pina , P. Ferrão , J. Fournier ., B. Lacarrièrec, O. Le Correc
b

a
IN+ Center for Innovation, Technology and Policy Research - Instituto Superior Técnico, Av. Rovisco Pais 1, 1049-001 Lisbon, Portugal
b
Abstract Veolia Recherche & Innovation, 291 Avenue Dreyfous Daniel, 78520 Limay, France
c
Département Systèmes Énergétiques et Environnement - IMT Atlantique, 4 rue Alfred Kastler, 44300 Nantes, France
Data centers are facilities hosting a large number of servers dedicated to data storage and management. In recent years, their power
consumption has increased significantly due to the power density of the IT equipment. In particular, cooling represents
approximately
Abstract one third of the total electricity consumption, therefore efficiently cooling data centers has become a challenging
problem and it represents an opportunity to reduce both IT energy costs and emissions environmental impact. The efficiency of
computers room air
District heating conditioning
networks (CRAC) systems
are commonly addressed can inbe the
increased using
literature as both
one advanced
of the most control techniques
effective solutionsand for
newdecreasing
free cooling the
technologies, such as the indirect adiabatic cooling (IAC), that is the humidification of air
greenhouse gas emissions from the building sector. These systems require high investments which are returned through the under adiabatic conditions. Water
heat
sprinkled
sales. Dueby spray
to thenozzles
changed humidifies
climate and cools down
conditions and the air taken
building from thepolicies,
renovation outside, which then cools
heat demand in down the computers
the future room
could decrease,
air by means of a crossflow heat
prolonging the investment return period. exchanger. In this way, the process air temperature is economically reduced and the cooling
process
The mainis effective
scope of even when isthe
this paper tooutside temperature
assess the feasibilityisofwarmer
using the than thatdemand
heat desired–inoutdoor
the computers
temperatureroom. Besidefor
function theheat
traditional
demand
approach,
forecast. The district of Alvalade, located in Lisbon (Portugal), was used as a case study. The district is consistedcontrol
that improves energy efficiency of CRAC systems through advanced hardware design, nowadays advanced of 665
systems
buildingsoffer thevary
that opportunity
in both to improve both
construction efficiency
period and performance
and typology. by mostly
Three weather acting on
scenarios software
(low, medium,components.
high) andInthreeparticular,
district
a renovation
model-based paradigm
scenarios candeveloped
were result very(shallow,
useful inintermediate,
the design ofdeep). the controller.
To estimate Thistheapproach involvesheat
error, obtained three main steps:
demand valuesplant
were
modelling,
compared controller
with results design,
from aand simulations.
dynamic In this paper,
heat demand model,First-Principle Data-Driven
previously developed (FPDD) techniques
and validated have been considered
by the authors.
inThe
the results
modellingshowedphase, in when
that order to obtain
only a model
weather changeas simple as possible
is considered, thebut accurate
margin enough.
of error Allbetheacceptable
could main components
for someofapplications
the plant,
such
(the as fans,inspray
error annual nozzles,
demand heat exchanger,
was lower than and20%the for
computers
all weatherroomscenarios
have beenconsidered).
taken into account
However, andafter
theyintroducing
have been calibrated
renovation
exploiting
scenarios,realthedata.
errorThe
valuedynamics
increasedof the
up tocomputers room variables
59.5% (depending on the (e.g. temperature)
weather are slower
and renovation than those
scenarios of the components
combination of
considered).
the cooling system, due to higher thermal inertias of the computers room. Therefore, fans, heat
The value of slope coefficient increased on average within the range of 3.8% up to 8% per decade, that corresponds to the exchanger, and spray nozzles are
described
decrease by in static models,ofwhereas
the number heatingthe computers
hours of 22-139hroomduring
is described by a LTI
the heating dynamic
season model. Once
(depending on theobtained a model
combination of of the plant,
weather and
a renovation
simulation scenarios
environment based onOn
considered). Matlab/Simulink
the other hand,isfunction
designedintercept
accordingly. The developed
increased for 7.8-12.7% control system (depending
per decade is hierarchical: a
on the
supervisor determines the best combination of CRAC water and process air flows which minimizes
coupled scenarios). The values suggested could be used to modify the function parameters for the scenarios considered, and the total power consumption,
while satisfying
improve the cooling
the accuracy demand.
of heat demand This system energy management problem is formulated as a non-linear optimization problem,
estimations.

© 2017 The Authors. Published by Elsevier Ltd.

Peer-review under responsibility of the Scientific Committee of The 15th International Symposium on District Heating and
* Corresponding author. Tel.: +39 049.827.7768; fax: +39 049.827.7799.
Cooling.
E-mail address: rampazzom@dei.unipd.it

Keywords: Heat demand; Forecast; Climate change

1876-6102 © 2017 The Authors. Published by Elsevier Ltd.
Peer-review under responsibility of the scientific committee of the AiCARR 50th International Congress; Beyond NZEB Buildings.

1876-6102 © 2017 The Authors. Published by Elsevier Ltd.

Peer-review under responsibility of the Scientific Committee of The 15th International Symposium on District Heating and Cooling.
1876-6102 © 2017 The Authors. Published by Elsevier Ltd.
Peer-review under responsibility of the scientific committee of the AiCARR 50th International Congress; Beyond NZEB Buildings
10.1016/j.egypro.2017.11.156
448 Alessandro Beghi et al. / Energy Procedia 140 (2017) 447–457
Alessandro Beghi et al. / Energy Procedia 00 (2017) 000–000

subject to internal air condition requirements and system operating constraints. The optimization problem is repeatedly solved at
each supervision period by using a population based stochastic optimization technique (Particle Swarm Optimization). Results of
simulations show that the proposed control system is effective and minimizes the input electric power while satisfying both the
data center thermal load and system operating constraints.
© 2017 The Authors. Published by Elsevier Ltd.
Peer-review under responsibility of the scientific committee of the AiCARR 50th International Congress; Beyond NZEB
Buildings.

Keywords: Cooling Device; Energy Efficiency; HVAC

1. Introduction

Data centers are facilities hosting a large number of servers dedicated to massive computation and storage. They
can be seen as a composition of information technology (IT) systems, which provides services to the end users, and a
support infrastructure, which supplies power and cooling. Power consumption in data centers has increased
significantly in the past few years. In particular, costs for air conditioning systems represent more than one third of
the total energy consumption, therefore efficiently cooling data centers has become a challenging problem. New free
cooling technologies, coupled with advance control systems, allow the increase of the efficiency of computers room
air conditioning (CRAC) systems, [1]. Among the available free cooling technologies, the indirect adiabatic cooling
(IAC) grants low energy consumption while meeting air conditioning needs, [2]. Nowadays, advanced control systems
give the chance to improve both system efficiency and performance by mostly acting on software components.
Modelling and control of data centers is discussed in various works in the latest literature, for example in [3] a control-
oriented data centers model is depicted, including the coupling in the dynamics between computational and cooling
resources. In [4], a control strategy that provides the best trade-off between energy consumption and cooling needs
satisfaction is presented. In [5], a MPC approach is used to obtain a cooling control system for data centers based on
the use of indirect fresh air.
In this paper, a model-based approach is developed for the design of an efficient control strategy for CRAC systems.
This approach involves three main steps: system modelling, analysing and developing of the controller, and simulating
both the plant and the controller. In particular, a model for the IAC system has been obtained by resorting to First-
Principle-Data-Driven technique, [6]. Moreover, a simulation environment based on Matlab/Simulink is designed
accordingly. Then, a hierarchical control system for CRAC optimal operation has been developed. Specifically, the
problem is formulated as a non-linear constrained optimization problem and it is solved by a population based
stochastic optimization technique, the PSO (Particle Swarm Optimization). Simulations results show that the proposed
control strategy minimizes the input total power while satisfying the operational constraints.

2. Indirect Adiabatic Cooling

In this paper, an IAC system is considered. It exploits the outside air to cool an internal environment, avoiding
external and internal airstreams to directly mix. Cool air taken from the outside is forced through a heat exchanger
(HX) and then immediately exhausted, whereas internal air is drawn from the room and circulated through the other
side of the HX before being re-inserted into the room. In order to economically reduce even more the temperature of
the air used for cooling, the outside air is humidified before it enters the heat recovery unit. This reduces the use of
traditional expansion air conditioning systems, reducing the cooling costs. This also makes the cooling system
effective even during periods when the outside air temperature is warmer than the desired internal room condition. In
particular, a system of spray humidifiers delivers moisture to the air using water. The fact that the internal air never
mixes with the outside air reduces the possibility of internal air contamination due to external pollutants. The IAC
considered system is depicted in Fig. 1. Two distinct air passages can be distinguished, called respectively primary
(or supply) and secondary (or process). Main components of the cooling system are fans, spray nozzles, HX, and
computers room.
 Alessandro Beghi et al. / Energy Procedia 140 (2017) 447–457 449
Alessandro Beghi et al. / Energy Procedia 00 (2017) 000–000

Fig. 1 - Computers room indirect adiabatic cooling system.

3. Modelling of the CRAC system

The first step in the modelling of a real system is collecting and systematic treatment of available knowledge.
Depending on the availability of the a priori knowledge, a first-principle or a data-driven model, or a combination of
them, can be applied. First-principle (FP) models describe in depth the physics of the system, allowing high capacity
of performance prediction, but usually it is difficult to obtain the specific relationships and values of certain
parameters, even if the physical principia are known. Instead, data-driven (DD) models exploit experimental data
acquired from the plant to obtain a suitable description of the system. Lowering time in the modelling phase is
guaranteed by this kind of models, but meaningful information is required in order to cover all the system operation
range. To obtain a CRAC model as simple as possible but accurate enough, First-Principle Data-Driven (FPDD)
models, which are a combination of FP and DD ones, are here considered. They take advantages of FP and DD
approaches and they are suitable for simulation and model-based control design purposes. In the next subsections, as
an example, models of spray nozzles, HX, and computers room are described. Further details of those models and a
description of that of the fans can be found in [7]. It is worth highlighting that the dynamics of the variables of interest
(e.g. temperatures) of the computers room are slower than those of spray nozzles, HX, and fans. Therefore, the
computers room is described by a dynamic LTI model, whereas spray nozzles, HX and fans are described by static
models.

3.1. Spray Nozzles

In the following, the lumped static model obtained for the spray nozzles is described. Adiabatic humidification
process is characterized by negligible variations of wet bulb temperature and enthalpy between inlet and outlet
conditions. A ‘wet’ effectiveness of the process is defined as function of the specific humidity x:

xb  xa
 wet  . (1)
xc  xa
450 Alessandro Beghi et al. / Energy Procedia 140 (2017) 447–457
Alessandro Beghi et al. / Energy Procedia 00 (2017) 000–000

The specific humidity xc is referred to saturation condition, whereas xa and xb are the specific humidity of the air at
the inlet and outlet sections respectively, Fig. 2.
In fair approximation, the humidifier ‘thermal’ effectiveness is calculated as function of inlet and outlet dry air
temperatures as follows:

Tb  Ta
 hum  . (2)
Tc  Ta

Fig. 2 - Adiabatic humidification process.

The humidifier can be considered as a heat exchanger working with a fictitious moisture fluid, which is characterized
by a fictitious specific heat, [8, 9]. Furthermore, the water can be considered as a quasi-isothermal fluid. The
humidifier outlet air temperature Ta,ex,hum is calculated as function of the inlet air temperature Ta,su,hum and the wet bulb
temperature Twb,su,hum by using the ε-NTU method (Effectiveness-Number of Transfer Unit), [10]:

Ta ,ex,hum  Ta , su,hum  (Twb, su,hum  Ta , su,hum ) hum , (3)

with

 hum  1  e  NTU , (4)

AU
NTU   . (5)
Cmin

The evaporative source is assumed as having an infinite capacity flow rate in the calculation of the ‘thermal’
effectiveness in Eq. (4). In Eq. (5) the term Ċmin is the minimal thermal capacity flow rate.
The fictitious overall heat transfer coefficient AU is calculated as function of both air and water flow rates 𝑚𝑚̇𝑎𝑎 and
𝑚𝑚̇𝑤𝑤 as follows:
 Alessandro Beghi et al. / Energy Procedia 00 (2017) 000–000
Alessandro Beghi et al. / Energy Procedia 140 (2017) 447–457 451

pa pw
 m   m 
AU  AU n  a   w  . (6)
 m a ,n   m w,n 

In Eq. (6), 𝐴𝐴𝑈𝑈𝑛𝑛 , 𝑚𝑚̇𝑎𝑎,𝑛𝑛 , and 𝑚𝑚̇𝑤𝑤,𝑛𝑛 are respectively the overall heat transfer coefficient and the air and water flow rates
at nominal conditions. Moreover, the parameters 𝑝𝑝𝑎𝑎 and 𝑝𝑝𝑤𝑤 are calibrated on experimental data.
Finally, the humidifier outlet air specific humidify can be calculated as follows:

Ta ,ex,hum  Ta ,su,hum
xa ,ex,hum  xa , su,hum  xwb,ex,hum  xa , su,hum  . (7)
Twb,ex,hum  Ta ,su,hum

3.2. Heat exchanger

The crossflow heat exchanger HX receives the air taken from the computers room (primary or supply air) and cools
it with the external air (secondary or exhaust air), which is treated, if need be, by the adiabatic humidifier. The heat
exchanged between primary and secondary airflows can be only sensible, when the specific humidity of both airflows
remains constant, or both sensible and latent, when condensation occurs. In fact, when the temperature of the
secondary air is lower than the dew point temperature of the supply air, this one condenses, creating a film of water
on the supply channels surface of the HX. For these reasons, the overall model of the HX is divided in two sub-models.
The ‘Dry sub-model’ considers only sensible heat exchange phenomena and it is a grey box model calibrated on
experimental data. On the other hand, the ‘Wet sub-model’ describes both sensible and latent heat exchanges.

3.2.1. Dry sub-model

The Fanning friction factor f and the Nusselt number Nu can be described using the correlations suggested by Abu-
Khader and Polley in [11].
The Fanning friction factor f is given by:

1
 16 3 3
n2 3
f     n1 Re  ,   (8)
 Re  

and the Nusselt number Nu is calculated as follows:

Nu  Nu  n10
turb  Nu lam
n10 n10
.
(9)

Laminar and turbulent components of the Nusselt number Nuturb and Nulam are given by:

Nuturb  n3 f Re n4 Pr n5 , (10)
 Alessandro Beghi et al. / Energy Procedia 00 (2017) 000–000
452 Alessandro Beghi et al. / Energy Procedia 140 (2017) 447–457



Nulam  n63  n73  n8 Gz n9  n7  ,
3 3


 (11)

where Re is the Reynolds number, Pr is the Prandtl number, and Gz is the Graetz number.
Once calibrated parameters n1, n2, …, n10 on experimental data, convective heat transfer coefficients in dry conditions
for both supply and exhaust air αdry,S and αdry,E can be calculated as follows:

i Nu
 dry  . (12)
dh

Then, the overall heat transfer coefficient is:

1
 Aplates Aplates s 
K dry     plates  , (13)
A  m 
 S dry, S dry, S AE dry, E  dry, E

where Aplates and splates are respectively the total area and the thickness of the plates in the HX, while AS and AE are
respectively the supply and exhaust side exchange surface areas.
Once obtained the overall heat transfer coefficient, P-NTU method is used to calculate outputs for dry channels
outflow, that are temperatures, humidities and heat flows.

3.2.2. Wet sub-model

Condensation process occurs in supply air channels when the wall surfaces temperature is lower than the dew point
temperature of the supply air. The process is described, for example, by Threlkeld et al. correlation, [12].
The convective heat transfer coefficient of supply air in wet-fin condition can be calculated as function of that in dry-
fin condition as follows:

bw
 wet , S   dry, S , (14)
c p, S ,i

where bw is the slope of the temperature-specific enthalpy curve of the saturated air at the wall mean temperature tw:

hsat
bw  . (15)
t tw

The Wet sub-model outputs are obtained using the P-NTU method.

3.2.3. Condensation process evaluation

In order to evaluate the condensation process in supply channels, we exploit a fictitious dimensionless coordinate
ξ that represents the portion of the channels where the condensation occurs and which is defined as function of the
inlet and outlet sections wall temperature tw,A and tw,B.
 Alessandro Beghi et al. / Energy Procedia 140 (2017) 447–457 453
Alessandro Beghi et al. / Energy Procedia 00 (2017) 000–000

Depending on the value of ξ, weights ωdry and ωwet are accordingly introduced. If ξ ≥ 1 then there is no
condensation and ωdry = 1, ωwet = 0, whereas, if ξ ≤ 1 then the supply air channels are completely wet and ωdry =
0, ωwet = 1. If 0 ≤ ξ ≤ 1 then there is a mixed condition and ωdry = 1 - ωwet, ωwet = 1 – ξ.
Once obtained the weights, sensible, latent and total heat flow rates of the overall HX model are obtained by
combining those of the two sub-models as follows:

qsens  dryqdry  wet qwet , sens , (16)

qlat  wet qwet ,lat , (17)

qtot  qsens  qlat . (18)

3.3. Computers room

The lumped-parameter thermal dynamic model of the computers room is derived by means of the electrical-thermal
analogy. In particular, a RC (Resistance-Capacity) model has been employed, where electrical resistances represent
thermal resistances between adjacent nodes, whereas electrical capacities represent nodes thermal capacities, Fig. 3.
Air temperatures and humidities of the room and working temperatures of electrical devices are the most important

Fig. 3 - The computers room RC model.

variables. Internal and external thermal loads (e.g. Qload,int, Toutdoor) are treated as disturbances. Temperatures and
humidities of the primary air are inputs and outputs of the model.
The dynamic of the most important variables (e.g. rack temperature Track, room temperature Troom, wall temperatures
Twall,in and Twall,out, and the computers room humidity xroom,out) are described by the following equations:

Track 1  Track  Troom 

 Qload,int  , (19)
 crack  R1 

Troom 1  Troom  Track Troom  Twall,in 

 m air hroom,in  hroom,out    , (20)
 croom  R1 R2 
454 Alessandro Beghi Beghi
Alessandro et al. / Energy Procedia
et al. / Energy 00 (2017)
Procedia 140000–000
(2017) 447–457

Twall,in 2  Twall,in  Troom Twall,in  Twall,out 

   , (21)
 cwall  R2 R3 

Twall,out 2  Twall,out  Twall,in Twall,out  Toutdoor 

   , (22)
 cwall  R3 R4 

xroom,out m air
 xroom,in  xroom,out , (23)
 M air,room

where R1, R2, R3, R4 are the thermal resistances, whereas Cwall, Crack, Croom are the thermal capacities. Mair,room
represents the total mass of air in the computers room and 𝑚𝑚̇𝑎𝑎𝑎𝑎𝑎𝑎 is the room air flow rate.

4. Control of CRAC systems

The optimization of the CRAC system operation guarantees significant energy savings. The optimal operation is
obtained by choosing the best combination of the modes of operation that minimises the input power consumption
while satisfying cooling load demand and constraints (e.g. computers room air temperature restrictions). It is worth
noticing that, the cooling needs can be achieved with various combinations of CRAC modes of operation. In Fig. 4,
the iso-temperature computers room curves are depicted as function of air process flow rate and supply water flow
rate.
In this paper, we employ a hierarchical control system. Local control loops (e.g. standard regulators) are used to
regulate the water-flow sprinkled by spray nozzles and the air-flow rate by process fan, which characterize the mode
of operation of the CRAC system. A supervisory control loop has to determine the set-points for the local loops
minimizing the power consumption, while satisfying operational constraints.
For this purpose, at each supervision period (e.g. 15 min), the local set-points are obtained by solving a non-linear
constrained optimization problem, which can be formulated on the given supervision time interval as follows:

arg min J x, u, d  , subject to: (24a)

u

, (24b)

hx, u, d  0 , (24c)

where J is the objective function to minimize, e.g. the total input electric power. In Eq. (24b), Σ represents the overall
CRAC model, which includes fans, spray nozzles, HX, and the computers room. States (i.e. air temperatures and
humidities), inputs (i.e. air and water flow rates) and disturbances (i.e. cooling load) of the system are the vectors x,
u, d, while y are the output variables (i.e. input power consumption, air temperatures and humidities). Finally, Eq.
(24c) represents the constraints on the CRAC modes of operation.
To solve the optimization problem (24) we have employed an ad-hoc PSO (Particle Swarm Optimization) algorithm,
which intensifies its search ability in specific promising regions, [13]. PSO is a derivative-free, stochastic and
population-based algorithm, inspired by the motion of bird flocks or fish schooling. It is often used to optimize
 Alessandro Beghi et al. / Energy Procedia 140 (2017) 447–457 455
Alessandro Beghi et al. / Energy Procedia 00 (2017) 000–000

functions in rather unfriendly non-convex, non-continuous search spaces, [14]. Each particle of the swarm represents
a candidate solution and it moves with its own velocity in the multidimensional search space, determines its own
position and calculates its fitness using an objective function. The interaction between particles grants global and
collective search capabilities, allowing the particles to move towards the global extremum.

Fig. 4 - The iso-temperature computers room curves

In the following, an example of application of the control strategy is reported. The aim is to determine the best
combination of CRAC water and air flow that minimizes the total electric power consumption while ensuring fixed
computers room air temperature conditions. The air-flow of the computers room is kept constant at its nominal value
and initial CRAC nominal condition are depicted in Table 1.

Table 1. CRAC nominal conditions.

Cooling load 𝑄𝑄𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙,𝑖𝑖𝑖𝑖𝑖𝑖 75 𝑘𝑘𝑘𝑘
External temperature 𝑇𝑇𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 25 °𝐶𝐶
Process inlet temperature 𝑇𝑇𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝,𝑖𝑖𝑖𝑖 25 °𝐶𝐶
−1
Process inlet humidity 𝑥𝑥𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝,𝑖𝑖𝑖𝑖 8 ∙ 10−3 𝑘𝑘𝑔𝑔𝑤𝑤𝑤𝑤 𝑘𝑘𝑔𝑔𝑑𝑑𝑑𝑑
Process outlet temperature 𝑇𝑇𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝,𝑜𝑜𝑜𝑜𝑜𝑜 25 °𝐶𝐶
−1
Process outlet humidity 𝑥𝑥𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝,𝑜𝑜𝑜𝑜𝑜𝑜 8 ∙ 10−3 𝑘𝑘𝑔𝑔𝑤𝑤𝑤𝑤 𝑘𝑘𝑔𝑔𝑑𝑑𝑑𝑑
Process air flow-rate 𝑚𝑚̇𝑎𝑎 15 ∙ 103 𝑚𝑚3 ℎ−1
Room inlet temperature 𝑇𝑇𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟,𝑖𝑖𝑖𝑖 25 °𝐶𝐶
−3 −1
Room inlet humidity 𝑥𝑥𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟,𝑖𝑖𝑖𝑖 8 ∙ 10 𝑘𝑘𝑔𝑔𝑤𝑤𝑤𝑤 𝑘𝑘𝑔𝑔𝑑𝑑𝑑𝑑
Room outlet temperature 𝑇𝑇𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟,𝑜𝑜𝑜𝑜𝑜𝑜 25 °𝐶𝐶
−3 −1
Room outlet humidity 𝑥𝑥𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟,𝑜𝑜𝑜𝑜𝑜𝑜 8 ∙ 10 𝑘𝑘𝑔𝑔𝑤𝑤𝑤𝑤 𝑘𝑘𝑔𝑔𝑑𝑑𝑑𝑑
3
Room air flow-rate 𝑚𝑚̇𝑎𝑎𝑎𝑎𝑎𝑎 39 ∙ 10 𝑚𝑚3 ℎ−1
Supply water 𝑚𝑚̇𝑤𝑤 100 𝑠𝑠 −1

The solution of the optimization problem (24) is obtained for various values of the constraint on the computers room
air temperature: in Fig. 5 the green stars represent the optimal combinations of water and air flows. It is worth noticing
456 Alessandro Beghi et al. / Energy Procedia 140 (2017) 447–457
Alessandro Beghi et al. / Energy Procedia 00 (2017) 000–000

that, for a fixed thermal load, as the desired computers room temperature decreases, then it is more convenient, in
terms of energy savings, increase the water-flow rather than increase the secondary air-flow.
In Fig. 6 the green stars represent the energy consumption that corresponds to the best combination of the
parameters for various values of the constraint on the computers room temperature.

Fig. 5 - Optimal combinations of modes of operation.

Fig. 6 - Energy consumption corresponding to the best combination of the modes of operation.
 Alessandro Beghi et al. / Energy Procedia 140 (2017) 447–457 457
Alessandro Beghi et al. / Energy Procedia 00 (2017) 000–000

Conclusions

In this paper, an efficient control strategy for IAC systems for data centers has been designed using a model-based
optimization approach. In particular, the main components of the CRAC system, such as spray nozzles, heat exchanger
and computers room, have been modelled through FPDD techniques, granting simplicity and reliability. A
Matlab/Simulink simulation environment has been developed accordingly. The operation of the cooling system is
optimized by minimizing the energy consumption while satisfying the constraint on the cooling demand. The
corresponding nonlinear and constrained optimization problem has been successfully solved by using a PSO
algorithm. The effectiveness of the proposed control architecture is confirmed by simulation results.

References

[1] Van der Geer J, Hanraads JAJ, Lupton RA. The art of writing a scientific article. J Sci Commun 2000; 163:51-9.
[2] Strunk Jr W, White EB. The elements of style. 3rd Ed. New York: Macmillan; 1979.
[3] Mettam GR, Adams LB. How to prepare an electronic version of your article. In: Jones BS, Smith RZ, editors. Introduction to the electronic
age. New York: E-Publishing Inc; 1999. p. 281-304.
[1] Shao S., Tian C., Xu H., Zhang H., Zou H. 2014. Free cooling of data centers: A re-view, Renewable and Sustainable Energy Reviews, 35,
171-182.
[2] Gallaher D., Weaver R., Weerts B. A. 2012. Green data center cooling: Achieving 90% reduction: Airside economization and unique indirect
evaporative cooling. In Green Technologies Conference, 1-6.
[3] Garone E., Krogh B. H., Parolini L., Sinopoli B. 2010. A hierarchical approach to energy management in data centers. 49th IEEE Conference
on Decision and Control (CDC), 1065-1070.
[4] Krogh B. H., Parolini L., Sinopoli B., Tolia N. 2010. A cyber-physical systems approach to energy management in data centers. In
Proceedings of the 1st ACM/IEEE In-ternational Conference on Cyber-Physical Systems, 168-177.
[5] Endo H., Fukuda H., Kasajima T., Kodama H., Kondo M., Ogawa M., Sugimoto T. 2015. Development of a cooling control system for data
centers utilizing indirect fresh air based on Model Predictive Control. In Ultra Modern Telecommunications and Control Systems and
Workshops (ICUMT), 7th International Congress on, 132-137.
[6] Czop P., Kost G., Sławik D., Wszołek G. 2011. Formulation and identification of first-principle data-driven models. Journal of Achievements
in materials and manufac-turing Engineering, 44(2), 179-186.
[7] Sisti E. 2014. Modelling of an Evaporative Cooling System for Energy Saving Applica-tions in Data Centers. Master’s thesis, University of
Padova, Italy (in Italian).
[8] Lebrun J., Silva C. A., Trebilcock F., Winandy E. 2004. Simplified models for direct and indirect contact cooling towers and evaporative
condensers. Building Services En-gineering Research and Technology, 25(1), 25-31.
[9] Cuevas C., Fonseca N. 2010. Experimental and Theoretical Study of Adiabatic Humidi-fication in HVAC&R Applications. Ingeniare. Revista
chilena de ingeniería, 18(2), 243-252.
[10] Bergman T. L., DeWitt D.P., Incropera F.P., Lavine A.S. 2006. Fundamentals of Heat and Mass Transfer, 6th edition, US: published by John
Wiley & Sons.
[11] Abu-Khader M. M., Polley G. T. 2005. Interpreting and applying experimental data for plate-fin surfaces: problems with power law
correlation, Heat transfer engineer-ing, 26(9), 15-21.
[12] Kuehn T. H., Ramsey J. W., Threlkeld J. L. 1998. Thermal environmental engineer-ing (Vol. 188). Upper Saddle River, NJ: Prentice Hall.
[13] Beghi A., Cecchinato L., Cosi G., Rampazzo M. 2012. A pso-based algorithm for opti-mal multiple chiller systems operation. Applied
Thermal Engineering, 32, 31 – 40.
[14] Eberhart R., Kennedy J. 1995. “Particle swarm optimization,” in Neural Networks. IEEE International Conference on, 4, 1942–1948.