Energy and Buildings 92 (2015) 19–28

Contents lists available at ScienceDirect

Energy and Buildings

journal homepage: www.elsevier.com/locate/enbuild

Modeling and identification of the cooling dynamics of a tropical

island hotel
Boris G. Vega Lara a,∗ , Luis M. Castellanos Molina b , José P. Monteagudo Yanes c
a
Department of Physics, Faculty of Engineering, University of Cienfuegos, Cienfuegos, Cuba
b
Department of Mechanical Engineering, Faculty of Engineering, University of Cienfuegos, Cienfuegos, Cuba
c
Center for Energy and Environmental Studies, Faculty of Engineering, University of Cienfuegos, Cienfuegos, Cuba

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

Article history: Minimizing energy consumption of heating, ventilation and air conditioning (HVAC) systems in build-
Received 19 September 2014 ings has experienced an increasing attention recently. Mainly motivated by the exploitation of building
Received in revised form 13 January 2015 automation systems (BAS), as well as simulation tools, innovative practices and methodologies have been
Accepted 20 January 2015
introduced to reduce the costs of energy required for heating and air conditioning of buildings. Outstand-
Available online 30 January 2015
ing has been the establishment of model predictive control (MPC) as a control strategy for the optimal
operation of HVAC systems. The basis for MPC is a dynamic model, which is objectively challenging and
Keywords:
time-consuming to obtain. This paper presents two modeling approaches of the return water temper-
Building modeling
System identification
ature of a central chiller plant based on data of the real operation of a building, weather disturbances,
Central chiller plant and the temperature of a reference thermal zone. It integrates building’s real measurements with a room
Parameter estimation simulator model. In this work, an original method for accurately describing the cooling dynamics of a
case study hotel on a tropical island is proposed.
© 2015 Elsevier B.V. All rights reserved.

1. Introduction with central air conditioning systems, chiller plants are commonly
used to provide cooling energy in the form of chilled water to
Tourism plays an important role in the economical developing of maintain the thermal conditions required for indoor areas. The
tropical islands in the Caribbean Sea. In Cuba, tourism is an industry operation of chillers leads to a huge electricity consumption and
with steady annual growth and a tendency to increase the number a peak demand, that is why any program of energy saving and
of hotels and rooms built. Raising the quality of service, reducing reduction of energy costs in a hotel should focus to reduce the con-
costs and environmental conservation are continuous challenges sumption of the HVAC (heating, ventilation and air conditioning)
in this area. system.
The hotel sector is generally characterized by high, and some- A series of papers about energy efficient operations of com-
times irrational, energy consumption. This is because of the plex chilled water systems under various working conditions are
inherited concept that the hotel’s main function is to give maximum reported in references [1–3]. Ma and Wang [1,2] reported a model-
comfort to its customers to any price. However, there are oppor- based supervisory and optimal control strategy for central chiller
tunities to reduce energy consumption and costs without affecting plants in complex building air-conditioning systems to enhance
the level and quality of services through an effective energy man- their energy efficiency and control performance. The optimal strat-
agement. egy is formulated using simplified models of major components
Concerning the energy costs for Caribbean hotels, electricity and genetic algorithm (GA) [2]. Wang et al. [3] presented an adap-
represents the largest bill, where the air conditioning and the light- tive optimal control strategy for online control of complex chilled
ing systems are the most power demanding. Air conditioning can water systems involving intermediate heat exchangers to enhance
account for about 65% of total electricity consumption, mainly due operation and energy performances. A simulated virtual platform
to the high solar radiation and ambient temperatures. For hotels representing a chilled water system in a super high-rise building
was established to validate and evaluate the above strategies.
An extensive research on the application of a model predictive
∗ Corresponding author at: Department of Physics, Faculty of Engineering, Uni-
control (MPC) of thermal energy storage in building cooling sys-
versity of Cienfuegos, Carretera a Rodas Km 4, Cienfuegos, Cuba. Tel.: +53 43550991.
tems have been published in the literature [4–6]. These works deal
E-mail address: [email protected] (B.G.V. Lara). with buildings, at university campus, equipped with a water tank

http://dx.doi.org/10.1016/j.enbuild.2015.01.036
0378-7788/© 2015 Elsevier B.V. All rights reserved.
20 B.G.V. Lara et al. / Energy and Buildings 92 (2015) 19–28

their total thermal load was modeled as a function of the indoor

Nomenclature set point temperature and the weather disturbances. Subsequently,
a simple model of the chilled water plant is added. Considering
T temperature (◦ C) the whole system, a model predictive controller for respecting the
C thermal capacity (J/◦ C) comfort requirements of the building and minimizing the energy
UA overall heat transfer coefficient (W/◦ C) consumption is designed and evaluated in simulation.
I intensity of solar radiation (W/m2 ) From the aforementioned review, it has clearly demonstrated
Occ occupancy the key importance of modeling buildings and HVAC systems
t time (s) for developing optimal control strategies, and consequently to
k discrete time index improve the energy efficient performance of cooling and heating
ṁ water mass flow rate(kg/s) plants. Building modeling approaches, e.g. first-principle models
cp specific heat of water (J/kg ◦ C) used on simulation tools and statistically based models, have been
n total of thermal zones the current trends that have encouraged the present investigation.
 parameter vector In contrast with previous works, in this research we focus on the
F state equation identification of suitable models for describing the cooling dynam-
G output equation ics (i.e., the heat exchange between the chilled water, thermal zones
and the environment) of a case study hotel on a tropical island. The
Subscripts
contribution of this paper is to present two modeling approaches
room reference thermal zone
of the return water temperature of a central chiller plant based on
extra additional state
(i) data of the real operation of a hotel (i.e., chilled water tempera-
w water
tures and hotel’s occupancy), (ii) weather disturbances (iii) and the
s supply
temperature of a reference thermal zone, previously constructed
r return
and modeled in TRNSYS. The models will be used for future appli-
amb ambient
cations of indoor climate control and energy efficient performance
prim primary circuit
reports.
bypass bypass line
This paper is organized as follows. Section 2 describes the
global global
building modeling approaches. Section 3 introduces the case study
c continuous
building, the central chiller plant and the real operation profile
d discrete
of the hotel. Section 4 recalls the model identification of a refer-
ence thermal zone. In Section 5, two modeling approaches of the
return water temperature are proposed, while Section 6 exposes
used for actively storing cold water produced by a series of chillers. the results. Finally, Section 7 concludes the article and remarks
Simplified models of chillers, cooling towers, tank and buildings future works.
are developed and validated for the purpose of model based con-
trol design. A MPC for the chillers operation is designed in order to
optimally store the thermal energy in the tank by using predictive 2. Modeling and identification for buildings
knowledge of building loads and weather conditions. The precinct
has a significantly enhanced level of instrumentation in order to 2.1. Building modeling techniques
support the development and demonstration of energy-efficient
A comprehensive literature related to building modeling and
technologies and practices.
identification has been published in a number of journals and con-
The MPC presented in reference [7] uses both weather forecast
and thermal model of a real building to inside temperature control. ference proceedings. One approach is to use the first-principle
Subspace methods were applied to identify a multiple input mul- models [13,14], which are largely used on simulation tools (e.g. Trn-
tiple output system (MIMO), that describes the dynamics of the sys, EnergyPlus, ESP-r), but these models are not explicit and cannot
indoor temperature and return water temperature of the heating be used for control directly. The alternative is to use statistically
system. Alternatively to this black-box statistical approach, a RC based models when large measurements data sets are available [8].
modeling of the same building was implemented by Široký et al. Prívara et al. [11,15,16] classified the building modeling tech-
[8]. Both modeling practices captured the thermal capacity of the niques in:
building. The results from real operation on a large university build-
ing proved the supremacy of predictive controller over a well tuned • Subspace methods family (4SID) is a family of algorithms estimat-
weather-compensated control, with the savings of 15−28%. ing a model of a system in a state space form. They work purely
In the other hand, the use of simulation tools (e.g. TRNSYS, Ener- in a statistical manner and belong to the black-box identification
gyPlus, ESP-r) for model identification of buildings has become algorithms. The complete mathematical treatment can be found
an innovative practice nowadays [9–11]. A medium weight office in the literature [17]. It was applied, for example, in references
building with two zones separated by a concrete wall, heating by a [7,9].
thermo-active building system (TABS) was constructed in TRNSYS • Prediction error methods (PEM) are the most commonly used sta-
environment [10]. Using a pseudo-random binary sequence as the tistical identification techniques. Their objective is to minimize
excitation input signal, a linear time invariant (LTI) model of the one-step ahead prediction error by optimizing parameters of pre-
system was identified using grey-box technique. In a co-simulation specified model structure [18]. This approach was employed for
framework [9,11], a real-life example of a large multi-zone office modeling of a room temperature in office buildings [19,20].
building is modeled linking EnergyPlus(EP) and Matlab via Build- • MPC relevant identification (MRI) is an approach minimizing
ing Controls Virtual Test Bed (BCVTB). It was shown that statistically multi-step ahead prediction errors [21]. The horizon for an error
based algorithms were the viable option for modeling such a com- minimization commensurate with a prediction horizon of the
plex structure building. predictive controller.
Archetypal guestrooms of a tropical island hotel were con- • Deterministic semi-physical modeling uses resistance capacitance
structed in TRNSYS by Lara et al. [12]. Applying a RC approach, (RC) network analogue to an electric circuitry to describe the
 B.G.V. Lara et al. / Energy and Buildings 92 (2015) 19–28 21

process dynamics and is often referred to as a grey-box modeling. There is extensive flexibility in choosing various predictor struc-
This technique was used for modeling real buildings [4,8]. tures, i.e., g functions, and this gives a corresponding freedom
• Probabilistic semi-physical modeling (PSPM) utilizes stochastic in defining “good” models in terms of prediction performance
differential equations for the description of the system to be iden- [18].
tified. Then a maximum likelihood estimation (ML) is employed A measure of model quality is the Akaike’s Information Criterion
to obtain unknown parameters. This method naturally enables an (AIC), defined in Eq. (9), where the model is tested on a validation
incorporation of prior information. It was used for modeling the data set. Comparing different models under this criterion, the most
heat dynamics of buildings [14,22]. accurate model has the smallest AIC [18].

The building modeling approach applied in this paper is based 2d

AIC = log VN (, Z N ) + . (9)
on prediction error methods. It is essential to underline that N
this technique will not be used to estimate parameters for a
pre-specified ready-made model structure (e.g. ARX, ARMAX, Box- 3. Case study: tropical island hotel
Jenkins or Output Error), but physically parametrized ones are
treated. For its significance to this work, the main aspects of this 3.1. Building and system description
methodology are briefly discussed in the following.
This case study considers a colonial mansion built in 1869, one of
2.2. Prediction error methods the most charming boutique hotels in Cuba. After a recent period of
restoration, the hotel was rebuilt with 36 standard rooms, 11 junior
From a selected certain model structure M, with particular mod- suite rooms, and 2 suite rooms.
els M() parametrized using the parameter vector  ∈ DM ⊂ Rd , the The central chiller plant of the hotel is a primary-secondary
set of models defined is: chilled water system that consists of: two screw compressor units
    with partial heat recovery (229.9 kW of cooling capacity and R-
M∗ = M   ∈ DM . (1) 134a as refrigerant) to provide cooling energy for the building, and
Each model represents a way of predicting future outputs. The constant speed primary and secondary pumps to distribute chilled
predictor could be a linear or nonlinear filter, which corresponds water across the chiller evaporator and thermal zones, respectively.
to one-step-ahead prediction for a general system described in The primary loop is decoupled with the secondary loop through the
reference [18]. The task with it is to decide upon how to use the bypass line, as shown in Fig. 1.
information contained in the batch of data collected from the sys- A building automation system (BAS) was installed in the hotel to
tem to select a proper value ˆ N of the parameter vector, and hence monitor the chiller plant, the sanitary hot water system, the quality
a proper member M(ˆ N ) in the set M∗ . of pool water, the cold chambers, and to control the starting or
Starting from a data set ZN (see Eq. (2)) collected from the sys- stopping of pumps, lighting systems, air handling units (AHU), etc.
tem, with inputs and outputs recorded over a time interval 1 ≤ t ≤ N, A database for energy management contains historical records of
supply and return chilled water temperatures, sanitary hot water
Z N = {u(1), y(1), . . ., u(N), y(N)}, (2) temperatures, active power of the hotel, etc.
The present study is based on data from the real operation of the
a general predictor of a parametrized model leads to:
hotel, collected by the BAS during periods between April to October
ŷ( t| ) = g(, Z t−1 ). (3) 2013. Time series of supply and return water temperatures from the
chiller, the weather, and the occupancy profile of the hotel for April
A parameter estimation method is summarized as the following 2nd–6th and 8th–16th can be found in Figs. 2 and 3, respectively.
general procedure [18]: The weather data were provided by the Provincial Weather Insti-
tute of Cienfuegos, and the occupancy by the commercial staff of
1. Form the sequence of prediction errors from observed data and the hotel.
the predictor ŷ( t| ), It is important to remark that nearly persistent excitation tests
ε(t, ) = y(t) − ŷ( t| ), t = 1, 2, . . ., N. (4) for the building took place throughout April; the normal operation
of the hotel was stopped due to foreseen shutdowns of chiller com-
Here  is the finite dimensional vector used to parametrize the pressors, and that is evident when both temperatures, the supply
mapping from past inputs and outputs to the space of the model and return water temperatures from the chiller, become practically
outputs. equal during this time. Further, the incidence of outliers caused
2. Possibly filter the prediction errors through a linear filter L(q) (q by signal spikes or by measurement malfunctions, can be seen in
denotes the shift operator, qu(t) = u(t + 1)), the figures. Hence, after extracting the informative data portions
into segments and merge them into one multiexperiment data set,
εF (t, ) = L(q)ε(t, ). (5)
these measurements (i.e., the estimation data) were conveniently
3. Select a scalar valued, positive function l(.). Typically, a quadratic used for input-output statistical identification experiments. Each
norm would be a first candidate, magnitude was recorded with a sampling interval of 5 min.
The interaction of the hotel with the chilled water system is
1 2
l(ε) = ε . (6) divided into two parts. Initially, the temperature of a reference
2
thermal zone (i.e., a critical guestroom) is modeled as a function of
4. Minimize the sum of these norms: weather and the temperature of the water supplied to the actuator
(i.e., a fan coil unit) located inside the room, as discussed in the next
ˆ N = argminVN (, Z N ), (7)
 section. This model will be useful for evaluating the occupants’ ther-
mal comfort in a near future climate control application. Finally, the
where
return water temperature from the building is modeled as a func-
1
N
tion of this reference room temperature, the hotel’s occupancy, the
VN (, Z N ) = l(εF (t, )). (8) ambient temperature and the chilled water supply temperature,
N
t=1 which is a control variable.
22 B.G.V. Lara et al. / Energy and Buildings 92 (2015) 19–28

Fig. 1. Hotel primary-secondary chilled water system.

4. Model identification of a reference thermal zone occupants’ thermal comfort for the worst scenario (i.e., the room
with the highest cooling load) was chosen. It means that, if the
4.1. The zone and HVAC system description reference zone is “comfortable”, consequently the comfort for the
rest of the zones is ensured. In this plan, the selected thermal
Measuring and recording of real temperatures of guestrooms, zone was the corner Suite Room 208, located in the 3rd floor. This
offices or other parts of the hotel, ideal for building mathematical guestroom has one of the biggest space area (66.8 m2 ) for air con-
models using identification techniques, have not been a common ditioning as a single thermal zone, only exceeded by the hotel’s
practice in such facilities. Alternatively, an original solution is pro- restaurant and the summit meeting room, not considered in this
posed: to construct, simulate and model a reference thermal zone study. It has façade orientations to the East (31.28 m2 of front area)
on a simulation tool, in this case TRNSYS. and the South (39.4 m2 of front area), balconies with doors made
In view of the case study hotel has a diversity of room types, of wood (4.05 m2 of area) and (2.84 m2 of area) glass, and conse-
to model each guestroom, office rooms, small shop, etc. might be quently receiving the effects of solar radiation on the external walls,
a complex and tedious task. A reference zone for evaluating the doors and the roof. This reference thermal zone was constructed in

Fig. 2. The data set for April 2nd–6th.

 B.G.V. Lara et al. / Energy and Buildings 92 (2015) 19–28 23

Fig. 3. The data set for April 8th–16th.

TRNSYS environment (component Type56), taking into account the Eq. (10) led to one of the most accurate structure for representing
geometry of the room, solar orientation, materials and dimensions the dynamics of the room temperature,
of walls, windows and doors, internal heat gains due to equipment
and occupants, and the HVAC system used. xroom (k + 1) = Axroom (k) + Buroom (k),
The HVAC system used in the building is a fan coil unit (i.e., (10)
yroom (k) = Cxroom (k) + Duroom (k),
chilled water is circulated through a series of fluid-to-air heat
exchangers to provide cooling energy to the building zones). For
this case, a real prototype of fan coil unit is firstly modeled from where xroom ∈ R2 , uroom ∈ R3 and yroom ∈ R, and A, B, C and D
a manufacturer data-sheet, and latterly implemented in TRNSYS. are matrices of appropriate dimensions. These system matri-
Considering that the fan coil unit is working at its maximum cool- ces are a canonical version of a black-box linear model in free
ing capacity (i.e., highest and constant water flow by the pipes and parametrization, estimated by subspace identification method,
using simulation data from TRNSYS. The manipulated variable and
air flow by the fan), a static model for describing the sensible cool-  T
ing power, as a function of the room temperature and the chilled disturbances are combined in vector uroom = Tw , Tamb , Iglobal .
water temperature, was achieved. This cooling power is put into The output vector is yroom = [Troom ]. The vector of system states is
practice in the environment as a mechanical ventilation. composed as xroom = [Troom , Textra ]T , where the state Textra has no
physical meaning. The sampling period of the model is 1/12 h.
Due to the bypass line, the water leaving the chiller plant (i.e., the
4.2. Modeling the reference thermal zone water with the supply temperature Tw,s ) is mixed with the return
water (i.e., the water with the temperature Tw,r ) from the hotel
The objective of modeling the reference thermal zone is to find before they flow into the fan coil units, as it can be seen in Fig. 1.
out the relationship between the selected room’s temperature with For that reason, the temperature of the chilled water reaching the
the chilled water supply and return temperatures, as well as the fan coil units, Tw , was substituted by a proper expression obtained
weather disturbances. from an energy balance, Eq. (11).
Before going into the modeling details, two simplifying supposi-
tions were considered. (i) The fan coil unit dynamics are faster than (ṁw,prim + ṁw,bypass )cp Tw = ṁw,prim cp Tw,s + ṁw,bypass cp Tw,r . (11)
the one of the room temperature, then the actuator’s static model is
suitable for simulations, (ii) and relative humidity is not controlled Taking into account the nominal water flows of the constant
in the room (60% is fixed in simulations). Meteorological condi- speed pumps in the primary (60 m3 /h) and secondary (80 m3 /h)
tions (i.e., ambient temperature and solar radiation) are simulated circuits, Tw is a linear combination of the measured chilled water
using TRNSYS Type109 with the year weather profile correspond- supply and return temperatures.
ing to Havana, Cuba. Timebase and TRNSYS simulation time-step
equal to 1/12 h for Type56 was chosen, allowing good compromise Tw = 3/4Tw,s + 1/4Tw,r . (12)
between simulation accuracy and execution time.
Once the thermal zone is constructed in TRNSYS environment To conclude this section, the actual temperature of the refer-
and the fan coil unit attached, the system is excited with a pseudo- ence thermal zone can be estimated, using Eqs. (10) and (12), from
random multilevel sequence of the chilled water temperature, Tw . real values of chilled water supply and return temperatures and
In the following and relied on black-box modeling approach, differ- weather disturbances. From now on, this estimation will be consid-
ent model structures (ARX, ARMAX, Box-Jenkins, Output Error and ered as a “measured” temperature of the reference thermal zone,
state space form) were tested. The discrete-time state space model Troom .
24 B.G.V. Lara et al. / Energy and Buildings 92 (2015) 19–28

5. Model identification of the return water temperature parameters the constant terms of the right hand side of Eq. (17), the
continuous-time nonlinear (CNL) model for explaining the return
In this section two approaches for modeling the return water water temperature resumes as:
temperature were considered. At first, a deterministic semi-
physical approximation was regarded. Afterward, a black-box
dT w,r
strategy was implemented. = c,1 (Tw,s − Tw,r ) + c,2 (Tamb − Tw,r )
dt
5.1. Simplifying assumptions + c,3 · Occ (Troom − Tw,r ). (18)

In order to reduce the complexity of the problem, some realistic

Grouping together Eqs. (18) with the continuous-time version
suppositions were taking into account.
of (10) and (12), the full model that represents the interaction of
the reference thermal zone with the chilled water system is a set
• As constant speed pumps were installed, the water mass flow
of first-order differential equations:
rates in primary and secondary circuits are assumed to be con-
stant during the experiments. ẋ(t) = Fc (t, x(t), u(t), c,1 , c,2 , c,3 ),
• As the water circuit is closed, the water thermal capacitance is
assumed to be time invariant. y(t) = Gc (t, x(t)), (19)
• Based on the fact that the same fan coil unit, considering the
x(0) = x0 ,
manufacturer cooling capacity, was installed in most of the
guestrooms for air conditioning, it is assumed that the over-
in which the state, output and input variables are respectively
all heat transfer coefficient (thermal transmittance/conductance)
defined according to:
for each actuator is the same and time invariant.
• The amount of thermal zones for air conditioning is mainly ⎡ ⎤
Tw,r
defined by the number of occupied guestrooms, i.e., the occu- ⎢ ⎥
pancy of the hotel in this instant of time. x = ⎣ Troom ⎦ ,

5.2. Deterministic semi-physical approach  Textra 

Tw,r
y= ,
Initially, a deterministic semi-physical approach is concerned Troom (20)
for modeling the return water temperature. From first principles,
⎡ ⎤
Tw,s
the differential equation describing the heat exchange between the ⎢ ⎥
return water with the supply water, the ambient and the thermal ⎢ Tamb ⎥
⎢
u=⎢ ⎥.
zones for air conditioning is as follows: ⎥
⎣ global ⎦
I
dT w,r
Cw = ṁw cp (Tw,s − Tw,r ) + UAamb (Tamb − Tw,r ) Occ
dt
+ UA1 (Troom,1 − Tw,r ) + · · · + UAn (Troom,n − Tw,r ). (13) Finally, the unknown parameters of this nonlinear grey-box model
should be estimated according to the procedures described in Sec-
tion 2.2, using measured data collected in April, 2013.
As the overall heat transfer coefficient for each fan coil unit was The drawback of this scheme is the nonlinear dynamic behav-
assumed equivalent, then ior of the return water temperature, in this case bi-linear between
UA1 = UA2 = . . . = UAn ≡ UA. (14) inputs and states. An MPC based on this model will result in a non-
convex optimization problem, which can be challenging to solve
Therefore, extracting UA as a common factor, and posterior n, the [23]. Therefore, a linear model version could be an alternative to
terms of the heat exchange between the return water temperature this disadvantage.
and the thermal zones can be transformed to:

UA(Troom,1 + · · · + Troom,n − n · Tw,r ) 5.3. Black-box approach

T 
room,1 + · · · + Troom,n
= n · UA − Tw,r , (15) Inspired in the black-box identification techniques, the follow-
n
ing model was found after a series of trial-and-error experiments.
where n is the total of thermal zones for air conditioning, precisely Regardless from physical principles, the primary interests of the
the occupancy (Occ) of the hotel in this instant of time. Denot- model in Eq. (21) were in fitting the data and being a linear structure
ing T room as the average temperature of the thermal zones, it is oriented to future control applications.
calculated as:
Tw,r (k + 1) = d,1 Tw,r + d,2 Troom + d,3 Tw,s + d,4 Tamb + d,5 Occ .
Troom,1 + · · · + Troom,n
T room = . (16)
n (21)
Subsequently, the lumped capacitance model for the return water Gathering Eqs. (20), (10) and (12), the full discrete-time linear
temperature can be rewritten as: (DL) model that represents the interaction of the reference thermal
dT w,r ṁw cp UAamb zone with the chilled water system is a set of first-order difference
= (Tw,s − Tw,r ) + (Tamb − Tw,r ) equations:
dt Cw Cw
UA x(k + 1) = Fd (k, x(k), u(k), d,1 , d,2 , d,3 , d,4 , d,5 ),
+ · Occ (T room − Tw,r ), (17)
Cw
y(k) = Gd (k, x(k)), (22)
where T room is substituted by the temperature of the reference
x(0) = x0 ,
thermal room, i.e., Troom , modeled in Section 4.2. Combining in
 B.G.V. Lara et al. / Energy and Buildings 92 (2015) 19–28 25

Table 1 Table 2
Parameters of CNL model. Parameters of DL model.

Parameters Value Standard dev. Unit Parameters Value Standard dev. Unit

 c,1 1.17 × 101 6.82 × 10−1 h−1  d,1 3.05 × 10−1 2.09 × 10−2 –
 c,2 1.48 × 100 9.48 × 10−2 h−1  d,2 4.57 × 10−2 6.27 × 10−3 –
 c,3 3.63 × 10−2 2.26 × 10−3 h−1  d,3 5.85 × 10−1 1.75 × 10−2 –
 d,4 5.03 × 10−2 5.28 × 10−3 –
 d,5 1.67 × 10−2 9.22 × 10−4 ◦
C

in which the state, output and input variables are in correspon-

dence with Eq. 20, respectively. In this case, 5 unknown parameters
the cooling dynamics of a case study hotel, using prediction error
should be estimated using the methods discussed in Section 2.2,
method.
and measured data collected in April, 2013.
The estimated parameters for the resultant models are pre-
sented in the following. In Table 1, parameters and their standard
6. Results deviations estimated for the CNL model are listed. In Table 2, idem
for the DL model.
Real measurements were used as input-output data for estimat- A cross-validation process was performed as a method to vali-
ing parameters in black-box and grey-box models that describe date the models. As it was stated before, measurements from

Fig. 4. Validation of models’ responses versus measurements. (a) Return water and room temperatures for April’s validation data set. (b) Return water and room temperatures
for June 28th–July 3rd.
26 B.G.V. Lara et al. / Energy and Buildings 92 (2015) 19–28

Fig. 5. Validation of models’ responses versus measurements. (a) Return water and room temperatures for August–September, experiment I. (b) Return water and room
temperatures for August–September, experiment II.

several days of April was used as estimation data. Validation data values higher than 90%, for the periods of April, June-July, and
corresponded to periods of April (not considered in the estimation August-September. Likewise, there is a wide coincidence between
data), June, July, August, September and October. Comparison of the outputs of both models, shown in the large intersection of each
the simulated outputs of the models with the measured data are response in the figures. A similar behavior between the room tem-
depicted in Figs. 4–6. perature and the chilled water return temperature was observed.
For evaluating the models’ quality, a normalized root mean At the end of September and the beginning of October, Fig. 6, the
square error (NRMSE) fitness value, defined as Eq. (23), was used: models’ quality decreased due to the change of the experimental
    conditions for estimating the parameters; the operating point of the
yk − ŷk  primary and secondary pumps was moved (i.e., they were pumping
NRMSE fit = 1−  
2
100%, (23)
yk − E (yk ) different water mass flow rate regarding the initial settings) with
2 the replacing of the central chiller plant for a new one.
Finally, both models were compared using the Akaike’s Infor-
where E stands for the expected value operator.
mation Criterion, Eq. (9). The CNL model (AICCNL = 2.59) is slightly
As it can be seen in Figs. 4 and 5, the identified models capture
more accurate than the DL one (AICDL = 2.90). This small difference
satisfactorily well the dynamics of the return water temperature,
between both AIC factors can be explained recalling that the CNL
with NRMSE fitness values around 70%, and even better of the
model has only 3 parameters to be estimated, instead of 5 for the
room temperature (in case of Troom , “measured” data versus simu-
DL model, with practically the same NRMSE fitness values.
lated outputs for the models was considered), with NRMSE fitness
 B.G.V. Lara et al. / Energy and Buildings 92 (2015) 19–28 27

Fig. 6. Validation of models’ responses versus measurements. (a) Return water and room temperatures for September 23rd–29th. (b) Return water and room temperatures
for October 1st–9th.

7. Conclusions and future works Apart from the methodology for modeling the return water tem-
perature, the reference thermal zone’s model was simulated with
This work has revealed a novel integration of a building simula- data of the real operation of the case study hotel. This analysis
tion software, for modeling a reference thermal zone, with data of showed that the currently used open-loop strategy for controlling
the real operation of a hotel and weather disturbances, to present the central chiller plant, considering empirical criterions, has some-
two modeling approaches of the return water temperature of a times carried on violations of the occupants’ thermal comfort (i.e.,
central chiller plant. Troom > 24◦ C). Hence, future works will be directed to develop con-
Two appropriate models were found for describing precisely the trol applications for minimizing energy consumption of the HVAC
cooling dynamics of a case study hotel on a tropical island. The first system whereas the comfort requirements of the hotel are satisfied.
proposed model, based on a deterministic semi-physical approach,
conducted to reproduce the response of the return water temper- Acknowledgements
ature with good accuracy. However, this model led to a bi-linear
dynamic behavior between one of the input and the states, which B. Vega deeply thanks the Swiss Federal Commission for Scholar-
will not be suitable for predictive control. The second proposed ships for Foreign Students (FCS) for their financial support provided
scheme, founded on a black-box paradigm, produced comparable to his research stay at the Automatic Control Laboratory at ETH
results, but in a linear discrete-time structure proper for an optimal Zurich, as well as Prof. Dr. Manfred Morari for very valuable discuss-
control application. ions and suggestions to his investigation. Moreover, the authors
28 B.G.V. Lara et al. / Energy and Buildings 92 (2015) 19–28

like to thank to Frauke Oldewurtel and Jiří Cigler for their indis- [10] J. Cigler, S. Prívara, Z. Váňa, E. Žáčeková, L. Ferkl, Optimization of predicted
pensable comments and recommendations to improve the quality mean vote index within model predictive control framework: computationally
tractable solution, Energy Build. 52 (2012) 39–49, http://dx.doi.org/10.1016/
of the paper. j.enbuild.2012.05.022 http://www.sciencedirect.com/science/article/pii:
S0378778812002770
Appendix A. Supplementary data [11] S. Prívara, J. Cigler, Z. Váňa, F. Oldewurtel, C. Sagerschnig, E. Žáčeková,
Building modeling as a crucial part for building predictive control,
Energy Build. 56 (2013) 8–22, http://dx.doi.org/10.1016/j.enbuild.2012.10.024
Supplementary data associated with this article can be http://www.sciencedirect.com/science/article/pii:S0378778812005336
found, in the online version, at http://dx.doi.org/10.1016/j.enbuild. [12] B.G.V. Lara, O.B. Padrón, J. Cigler, F. Oldewurtel, M.A.R. Borroto, Model predictive
control for a tropical island hotel, in: 11th REHVA World Congress-CLIMA, 2013,
2015.01.036. pp. 5917–5927.
[13] H. Madsen, J. Holst, Estimation of continuous-time models for the heat
References dynamics of a building, Energy Build. 22 (1) (1995) 67–79, http://dx.doi.org/
10.1016/0378-7788(94)00904-X, http://www.sciencedirect.com/science/
article/pii:S037877889400904X
[1] Z. Ma, S. Wang, An optimal control strategy for complex building central
[14] K.K. Andersen, H. Madsen, L.H. Hansen, Modelling the heat dynam-
chilled water systems for practical and real-time applications, Build. Environ.
ics of a building using stochastic differential equations, Energy Build.
44 (6) (2009) 1188–1198, http://dx.doi.org/10.1016/j.buildenv.2008.08.011
31 (1) (2000) 13–24, http://dx.doi.org/10.1016/S0378-7788(98)00069-3
http://www.sciencedirect.com/science/article/pii:S0360132308002114
http://www.sciencedirect.com/science/article/pii:S0378778898000693
[2] Z. Ma, S. Wang, Supervisory and optimal control of central chiller plants
[15] S. Prívara, J. Cigler, Z. Váňa, F. Oldewurtel, E. Žáčeková, Use of partial least
using simplified adaptive models and genetic algorithm, Appl. Energy
squares within the control relevant identification for buildings, Control Eng.
88 (1) (2011) 198–211, http://dx.doi.org/10.1016/j.apenergy.2010.07.036
Pract. 21 (1) (2013) 113–121, http://dx.doi.org/10.1016/j.conengprac.2012.09.
http://www.sciencedirect.com/science/article/pii:S0306261910003065
017 http://www.sciencedirect.com/science/article/pii:S0967066112002006
[3] S. Wang, D. Gao, Y. Sun, F. Xiao, An online adaptive optimal control strategy for
[16] S. Prívara, Z. Váňa, E. Žáčeková, J. Cigler, Building modeling: selection of the most
complex building chilled water systems involving intermediate heat exchang-
appropriate model for predictive control, Energy Build. 55 (2012) 341–350,
ers, Appl. Therm. Eng. 50 (1) (2013) 614–628, http://dx.doi.org/10.1016/
http://dx.doi.org/10.1016/j.enbuild.2012.08.040, S0378778812004446.
j.applthermaleng.2012.06.010 http://www.sciencedirect.com/science/article/
[17] P.V. Overschee, B.D. Moor, Subspace Identification for Linear System. Theory-
pii:S1359431112004310
Implementation-Applications, Kluwer Academic Publishers, 1996.
[4] Y. Ma, F. Borrelli, B. Hencey, B. Coffey, S. Bengea, P. Haves, Model predictive con-
[18] L. Ljung, System Identification: Theory for the User, 2nd ed., Upper Saddle River,
trol for the operation of building cooling systems, IEEE Trans. Control Syst. Tech-
NJ, Prentice Hall PTR, 1999.
nol. 20 (3) (2012) 796–803, http://dx.doi.org/10.1109/TCST.2011.2124461.
[19] S. Wu, J.-Q. Sun, A physics-based linear parametric model of room temperature
[5] Y. Ma, F. Borrelli, B. Hencey, B. Coffey, S. Bengea, P. Haves, Model predictive
in office buildings, Build. Environ. 50 (2012) 1–9, http://dx.doi.org/10.1016/
control for the operation of building cooling systems, in: American Control
j.buildenv.2011.10.005 http://www.sciencedirect.com/science/article/pii:
Conference (ACC), 2010, 2010, pp. 5106–5111.
S0360132311003544
[6] Y. Ma, F. Borrelli, B. Hencey, A. Packard, S. Bortoff, Model predictive con-
[20] G. Mustafaraj, J. Chen, G. Lowry, Development of room temperature and relative
trol of thermal energy storage in building cooling systems, in: Decision and
humidity linear parametric models for an open office using bms data, Energy
Control, 2009 held jointly with the 2009 28th Chinese Control Conference,
Build. 42 (3) (2010) 348–356, http://dx.doi.org/10.1016/j.enbuild.2009.10.001
Proceedings of the 48th IEEE Conference on CDC/CCC 2009, 2009, pp. 392–397,
http://www.sciencedirect.com/science/article/pii:S0378778809002394
http://dx.doi.org/10.1109/CDC.2009.5400677.
[21] D. Shook, C. Mohtadi, S. Shah, A control-relevant identification strat-
[7] S. Prívara, J. Široký, L. Ferkl, J. Cigler, Model predictive control of a
egy for GPC, IEEE Trans. Autom. Control 37 (7) (1992) 975–980,
building heating system: the first experience, Energy Build. 43 (2-3)
http://dx.doi.org/10.1109/9.148352.
(2011) 564–572, http://dx.doi.org/10.1016/j.enbuild.2010.10.022 http://www.
[22] P. Bacher, H. Madsen, Identifying suitable models for the heat dynamics of
sciencedirect.com/science/article/pii:S0378778810003749
buildings, Energy Build. 43 (7) (2011) 1511–1522, http://dx.doi.org/10.1016/
[8] J. Široký, F. Oldewurtel, J. Cigler, S. Prívara, Experimental analysis of model pre-
j.enbuild.2011.02.005 http://www.sciencedirect.com/science/article/pii:
dictive control for an energy efficient building heating system, Appl. Energy
S0378778811000491
88 (9) (2011) 3079–3087, http://dx.doi.org/10.1016/j.apenergy.2011.03.009
[23] F. Oldewurtel, A. Parisio, C.N. Jones, D. Gyalistras, M. Gwerder, V.
http://www.sciencedirect.com/science/article/pii:S0306261911001668
Stauch, B. Lehmann, M. Morari, Use of model predictive control and
[9] S. Privara, Z. Vana, D. Gyalistras, J. Cigler, C. Sagerschnig, M. Morari, L. Ferkl,
weather forecasts for energy efficient building climate control, Energy
Modeling and identification of a large multi-zone office building, in: IEEE
Build. 45 (2012) 15–27, http://dx.doi.org/10.1016/j.enbuild.2011.09.022
International Conference on Control Applications (CCA), 2011, pp. 55–60,
http://www.sciencedirect.com/science/article/pii:S0378778811004105
http://dx.doi.org/10.1109/CCA.2011.6044402.