Building and Environment 51 (2012) 379e387

Contents lists available at SciVerse ScienceDirect

Building and Environment

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

Optimal temperature control of intermittently heated buildings using Model

Predictive Control: Part I e Building modeling
Ion Hazyuk a, b, *, Christian Ghiaus a, David Penhouet c
a
INSA-Lyon, CETHIL, UMR5008, F-69621 Villeurbanne, France
b
Technical University of Cluj-Napoca, 15 Constantin Daicoviciu Str., 400020 Cluj-Napoca, Romania
c
CSTB (Centre Scientifique et Technique du Bâtiment), 84 Avenue Jean Jaurès, 77421 Marne-la-Vallée, France

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

Article history: An important performance in building thermal control is to ensure thermal comfort with minimal energy
Received 19 July 2011 consumption. Model Predictive Control (MPC) is considered to be one of the most suited solutions for
Received in revised form this due to its ability to use occupancy schedule and weather forecasts for optimal temperature control.
8 November 2011
MPC relies on a dynamical model of the building, which is the main difficulty of applying it. Therefore,
Accepted 17 November 2011
this paper treats the problems related to building modeling and model parameters identification. A
robust model of the building is obtained in two stages: firstly physical knowledge is used to determine
Keywords:
the structure of a low-order model, then least squares identification method is applied to find the
Lumped capacity model
System identification
numerical values of the model parameters. In order to perform the identification usually there are
Least squares identification method required input/output data records having variations which generally are not accepted in inhabited
Building parameter estimation buildings because of imposed comfort conditions. Also inhabited buildings contain unmeasured
disturbance sources which may degrade the identified model quality. Therefore this paper proposes to
use detailed building models, implemented in dedicated simulation tools, to generate the required input/
output data records instead of measuring them on real buildings. This allows us to apply desired input
signals and to eliminate disturbance sources. Additionally, the paper presents a method to identify the
nonlinearity existing in building thermal behavior, which permits to represent the building by separated
linear and nonlinear blocks. This model representation, used along with the linearization method
proposed in Part II, permits to design the temperature controller without resorting to the nonlinear
system theory.
Ó 2011 Elsevier Ltd. All rights reserved.

1. Introduction its job by maintaining a set-point temperature but without paying

much attention to the energy consumption and the wear and tear of
The general subject of the present work is optimal temperature the elements. Therefore, buildings waste large amounts of energy
control in intermittently heated buildings by using Model Predic- due to poor control performance and have a large potential for
tive Control (MPC). The topic is presented in two papers, one savings by improving it.
focusing on building modeling and the other on MPC. Usually, controllers in Building Energy Management Systems
Thermal control in buildings has received much less attention (BEMS) have comfort as the single objective (commonly indicated
from control engineering communities than other application fields as the set-point temperature); reducing energy consumption and
like aerospace, petro-chemical, electronic or automotive. One of the wear and tear of system elements (e.g. control valves and electrical
reasons is that in these application fields an error in the control motors) is less addressed in the design of temperature controllers
algorithm could cause easy detectable effects. On the contrary, in [1e3]. A promising solution for BEMS is Model Predictive Control
thermal control of buildings the effects of poor control cannot be (MPC). MPC has several features that make it suitable for the
easily noticed. A simple relay or thermostatic valve on radiator does problems encountered in intermittently heated buildings. First, it is
able to use the occupancy schedule and weather forecasts for
optimal temperature control. Second, MPC optimizes not only the
comfort but also an energy criterion. As heating systems generally
* Corresponding author. Bât. Sadi CARNOT, 9 rue de la Physique, 69621 Villeur-
banne Cedex, France. Tel.: þ33 4 72 43 74 84. consume energy to provide thermal comfort, MPC makes a trade-
E-mail address: [email protected] (I. Hazyuk). off between energy savings and thermal comfort. Lastly, MPC is

0360-1323/$ e see front matter Ó 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.buildenv.2011.11.009
380 I. Hazyuk et al. / Building and Environment 51 (2012) 379e387

able to handle implicitly the constraints in Multi-Input Multi-  Simulation tools facilitate the construction of the detailed
Output (MIMO) systems. Theoretically, MPC gives the best results model via simple user interface, which is more convenient than
as compared to other control algorithms [4] and recently has been constructing it ourselves.
successfully applied on real occupied buildings [5,6].  It is possible to set particular initial condition, usually required
The application of MPC requires a dynamical model of the by identification methods, what is difficult to achieve in real
system. Although system modeling and identification have world.
a solid theoretical base, their application to real buildings is  It is possible to carry out specific simulations in order to isolate
complicated. Generally for control, and especially for MPC, linear some nonlinear behavior, for the aim of identifying this
low-order models are preferred. However, a typical model of characteristic.
a building, obtained by spatial discretization of heat transfer
equations, can present some nonlinearity and include several The specificity of the proposed approach is to use physical
hundreds of states. Model-size reduction techniques can reduce knowledge for building modeling and identification. An analysis of
this number to less than ten [7,8]. Nevertheless, they require an heat transfer phenomena shows that they are governed by
initial detailed model of the building, which is difficult to obtain, nonlinear laws [23]. Thus, by using physical knowledge, the
and are not effective for nonlinear models. An alternative is to nonlinear characteristic of the building thermal behavior is iden-
obtain a low-order model of the building using experimental tified in order to be used for model linearization in Part II of this
black-box identification methods [9]. However, nonlinear paper.
systems also pose problems for these methods because usually The low-order model estimated in this paper corresponds to
they obtain models which are accurate only around the operating a typical detached house having a 100 m2 living surface. Its detailed
point where they were identified. Moreover, the identified model is implemented in Simbad toolbox [22], under Matlab e
parameters tend to lose physical insight and sometimes their Simulink environment. This well-tested model is used by French
values contradict physical phenomena; for example, complex Scientific and Technical Center for Building (CSTB) to evaluate by
conjugated poles can be obtained though a physical analysis simulation or emulation the performances of real controllers. In
shows that they do not exist in the thermal model of a building case of emulation, the operations are performed in real time and
[10]. In order to avoid this, models with lumped parameters are the communication between the emulated building system and the
used. Here the structure of the low-order model is found directly, controller prototype is done by electrical signals through input/
by applying basic physical principles, and the values of the output ports, just like in real practice.
parameters are estimated by using experimental identification
techniques [11e13]. During the identification, the range of the 2. Dynamic modeling of a building
values of the parameters can be bounded in order to keep their
physical insight [14]. In addition, a class of nonlinear systems can 2.1. State-space modeling
be represented by using basic physical knowledge along with
these models [15]. However, experimental identification of the In order to formulate the reduced model of the building, usual
model parameters requests input/output data from the real simplifying assumptions as time invariant parameters, uniformly
building. To obtain a good quality model, the input signals must distributed properties, etc. are considered. Models derived from
have a large spectrum to excite the system modes and the output physical relations are naturally represented in state-space by a set
signal must have enough variations to observe them [16]. The of first order differential equations. Moreover, the used MPC algo-
best candidate for input signals is a pseudo-random binary rithm also requires the model of the system in state-space
sequence (PRBS). The problem is that in inhabited buildings the representation.
choice of the input signals (i.e. internal heat flux) is very limited. Low-order building models used for control purpose are most
We cannot move beyond the accepted thermal comfort norms for often derived from linear network representations with lumped
a long time and the cost of testing an uninhabited building is parameters [11,13,24e30]. The idea is to apply the principle of
prohibitive. Another problem is that inhabited buildings contain analogy between two different physical domains that can be
plenty of unconsidered heat sources influencing the indoor described by the same mathematical equations. Thus a linear
temperature (i.e. free gains from occupants, electrical appliances, electrical circuit represents the building and the state-space
etc), which degrade the quality of the identified model. equations are obtained by solving that circuit. Here, the tempera-
On the other hand, on the market there are dedicated software ture is equivalent to voltage, the heat flux e to current, the heat
tools to simulate building thermal behavior, such as CODYBA [17], transmission resistance is represented by electrical resistance and
TRNSYS [18], Comfie [19], ESP-R [20], EnergyPlus [21] or Simbad the thermal capacity e by electrical capacity. The equivalent circuit
[22]. They use detailed models based on first principles and of the building is obtained by assembling models of the walls,
constitutive laws of thermodynamics. These models aggregate windows, internal mass, etc. In the case of single-zone buildings,
physical components of a building in a system of algebraic and interior walls are being part of the internal thermal mass while
differential equations. The obtained models tend to be of a very exterior walls are forming the building envelope.
high order and therefore they are not appropriated for the devel- Building envelope is usually represented by 2ReC [11,25,26] or
opment of control algorithms, such as MPC. The advantage is that in 3Re2C [12,31e33] networks. The roof and the floor also can be seen
simulation there are no constraints on input signal choice neither as some kind of wall, and therefore they are modeled by the same
on the disturbance sources. Often, the models used by the network structure. The internal thermal mass is usually repre-
mentioned simulation tools are experimentally validated, so they sented by a single capacity, although some authors modeled it by
can replace the real building for parameter identification. There- 2Re2C network [30] or even omitted it [25]. Windows do not
fore, this paper proposes that data records, required for parameter accumulate thermal energy and therefore they are represented as
identification of low-order models, be generate by simulating simple resistances. Ventilation and infiltrations are also modeled by
a detailed model of the building instead of measuring them on the a resistance.
real building. Besides the possibility of applying PRBS inputs and In the proposed approach a single-zone building is considered,
eliminating disturbance sources, the advantages of the proposed whose equivalent electrical circuit is shown in Fig. 1. The considered
method are: passive components are the envelope, the floor, the windows, the
 I. Hazyuk et al. / Building and Environment 51 (2012) 379e387 381

Fig. 1. Equivalent electrical network representation of a low-order thermal model of a building.

ventilation and infiltrations and the internal thermal mass. The roof space equations of the circuit from Fig. 1 the superposition
is not considered because it is assumed that the temperature in the theorem for electrical circuits is applied. This theorem states that
attic space is close to the outdoor air temperature. Thus, the ceiling, the response in any branch of a linear circuit having more than one
delimiting the thermal zone from the attic is being part of the independent source equals to the algebraic sum of the responses
envelope. It is supposed that the ground temperature is constant at given by each independent source acting alone, while all other
a depth of nearly three meters [23] and therefore this layer of independent sources are replaced by their internal impedance. The
ground is being part of the floor. The envelope is represented by internal impedance of an ideal voltage source is zero and that of an
a 2ReC network where its capacity is lumped in Cw and the wall ideal current source is infinite; therefore, all voltage sources are
insulation is represented by two halves of its conductive resistance replaced by short circuits and all current sources are replaced by
Rw/2. The resistances Rco and Rciw represent the convective resis- open circuits.
tances between the envelope and the outdoor and indoor air, By applying the superposition theorem on the circuit shown in
respectively. It is assumed that the varying wind velocity has no Fig. 1, four separated single-input single-output (SISO) models are
influence on the heat exchange coefficient of the envelope surfaces, obtained, each one corresponding to an input/output pair.
so these two resistances are considered constant. The same However, all these models can be embedded in a single multi-input
considerations are made for the floor; its capacity is lumped in Cf single-output model (MISO). Considering the character of each
and the insulation is represented by two halves of its conductive source (command or disturbance), the following MISO model is
resistance Rf /2. The convective resistance between the floor and obtained:
the indoor air is represented by Rcif. The thermal capacity of the
internal mass is lumped in Ca. The heat losses due to ventilation and x_ ¼ Ax þ B1 u þ B2 w
(1)
infiltration are modeled by the resistance Rv. y ¼ Cx þ D1 u þ D2 w
The active elements of the building thermal circuit are outdoor where:
air and ground temperatures, solar radiation falling on the building
envelope and internal heat flux. The internal heat flux is the sum of x ¼ ½ qw qz qf T e is the state vector: qw is the wall temper-
all the internal free gains (from building occupants, electrical ature, qz is the zone temperature and qf is the floor temperature;
appliances, solar radiation through windows) and contributions y ¼ qz e the output of the system;
from heating terminals i.e. radiators. The outdoor air and ground w ¼ ½ qo qg Fs T e the measurable inputs of the system: qo is
temperatures are modeled by ideal voltage sources qo and qg the outdoor air temperature, qg is the ground temperature and
respectively. The solar radiation and internal heat gains are repre- Vs is the solar radiation on the walls;
sented by ideal current sources, Vs and Vg respectively. u ¼ Vg e the command (i.e. the total internal heat flux, which
The output of the model is the indoor temperature. This comes mainly from the radiators, but it also includes internal
temperature is influenced by four different inputs: outdoor air and gains from occupants, solar radiation through the windows, etc.);
ground temperatures, solar radiation and internal sources. Note
that the internal sources are separated in free gains and heating
terminals. The heating terminal is a controllable source, repre- 2 3
senting the command of the system while the free gains are R þ R2 1
6 1 0 7
uncontrollable and considered to be unmeasurable, so they are 6 R1 R2 Cw R2 Cw 7
6 7
unmeasurable disturbances. Outdoor temperature, qo, ground 6 1 Rv þ R2 þ R3 1 7
A ¼ 6
6 R 2 Ca  7 e state matrix;
7
temperature, qg, and solar radiation, Vs, are also uncontrollable 6 Rv R2 R3 Ca R 3 Ca 7
6 1 2R3 þ Rf 7
sources but they can be measured, so they are measurable distur- 4 0  5
bances. For the operating temperature range of the building, the R 3 Cf R3 Rf Cf
model is considered to be linear. Thus, in order to find the state-
382 I. Hazyuk et al. / Building and Environment 51 (2012) 379e387

2 3 frequency of the ground temperature transfer function is smaller

1 Rco
2 36 R1 Cw 0
0 6 R1 Cw 7
7
than the cutoff frequency of the outdoor temperature transfer
6 1 7 6 1 7 function. Therefore, the dynamic gain of the ground temperature is
B1 ¼ 4 5; B2 ¼ 6
6 Rv Ca 0 0 7 7 e input matrices; much smaller than that of the outdoor temperature. All this factors
Ca 6 7
0 4 2 5 indicate that ground temperature influences the dynamic behavior
0 0
Rf Cf of the indoor temperature at least sixteen times less than outdoor
temperature does. Therefore, the heat flux toward the ground can
be ignored for control purpose, considering it as a disturbance
C ¼ ½0 1 0 ; D1 ¼ 0;
which will be compensated by the controller. Eliminating the
D2 ¼ ½ 0 0 0  e the output and feed-through matrices ground source and floor inertia from the thermal model of the
respectively; building, the building model is of second order with only three
inputs. The equivalent linear circuit from Fig. 1 is transformed into
with: R1 ¼ Rco þ Rw/2, R2 ¼ Rciw þ Rw/2, R3 ¼ Rcif þ Rf/2 the representation from Fig. 2 whose state-space model is:

x_ ¼ A x þ B1 u þ B2 w
(4)
2.2. Model analysis y ¼ C x þ D1 u þ D2 w
with:
Once the model of the system is obtained, it can be represented
in several forms, each one with its advantages. For example, the x ¼ ½ qw qz T e state vector;
continuous state-space representation is a natural way to represent y ¼ qz, e output of the system;
system models derived from physical knowledge. However, for u ¼ Vg, w ¼ ½ qo Fs T e inputs of the system;
model analysis it is more convenient to have it in transfer function
2 3
representation. The transition from state-space to transfer function Rco þ Rw þ Rciw 1
representation is made by: 6 ðRco þ Rw =2ÞðRciw þ Rw =2ÞCw ðRciw þ Rw =2ÞCw 7
6 7
A ¼6 7
4 1 Rciw þ Rw =2 þ Rv 5
HðsÞ ¼ CðsI  AÞ1 B þ D (2) ðRciw þ Rw =2ÞCa

ðRciw þ Rw =2ÞRv Ca
where: e state matrix;
2 3
H(s) e is the model in transfer function representation; " # 1 Rco
s e is the complex variable; 0 6 ðR þ R =2ÞC ðRco þ Rw =2ÞCw 7
6 co w w 7
I e is an identity matrix of the same size as the state matrix A. B1 ¼ 1 ; B 2 ¼ 6 7
Ca 4 1 5
0
Rv Ca
As the model (1) has four inputs and one output, four transfer
functions are obtained by applying the transformation (2), each one e input matrices;
connecting an input to the output. The building is represented by
C ¼ ½ 0 1 ; D1 ¼ 0; D2 ¼ ½ 0 0 , e output and feed-through
the superposition of these four transfer functions which show how
matrices respectively;
each input influences the indoor temperature. A transfer function
may be expressed by:
3. Model inputs characterization
YðsÞ
hHðsÞ ¼ K GðsÞ (3)
UðsÞ The identification of model parameters given by equation (4)
needs data records of the inputs and the output of the system.
which shows that the model H(s) can be described by its static Since solar radiation on the building surfaces cannot be directly
(steady-state) gain, K, and its dynamic gain, G(s), which depend on measured and internal heat flux cannot be directly controlled in the
the input frequency. case of water based heating systems, additional relations must be
The most important frequency of the input signal is the one established for these two sources.
corresponding to the period of 24 , i.e. the daily variations of the The second input of the building model, Vs, is the amount of
outdoor temperature qo and solar radiation Vs. As the indoor set- solar radiation falling on the building envelope. This energy flux
point temperature has two levels that are periodically alternated, needs to be obtained from the diffuse and direct radiation available
i.e. day and night set-point, the interior heat flux Vg resulting from in meteorological data. Section 3.1 reviews the relations for heat
control will also have the same period. Among the modeling
assumptions, the ground temperature qg is considered to be
constant. However, if indoor temperature is taken as reference
level, the ground temperature also varies with the same 24 h
period.
A comparison between the static gain of the outdoor tempera-
ture and that of the ground temperature shows that the former is
about eight times larger than the later. This relation reflects the
ratio between thermal conductance of the walls and that of the
floor. In a 24 h period, the variation of the ground temperature as
compared to the indoor temperature is about five degrees. In the
same period, the variation of the outdoor temperature as compared
to the indoor one is at least ten degrees. In addition, the floor
inertia, which in our case includes about three meters of soil, is Fig. 2. Equivalent electrical network representation of a low-order thermal model of
much higher than the envelope inertia. This means that the cutoff a building after the omission of the ground source.
 I. Hazyuk et al. / Building and Environment 51 (2012) 379e387 383

flux Vs calculation from diffuse and direct radiation and gives the f e geographical latitude of the location where the building is,
form of the solar radiation input for the model. positive for north hemisphere;
The third input of the building model is the heat flow from water g e azimuth angle of the surface (angle between the normal to
radiators, Vg, which can be calculated by using water inlet (qin) and the surface and meridian), zero for south facing, negative for
outlet (qout) temperatures: west facing and positive for east facing;
_ w cw ðqin  qout Þ u ¼ 15(t  12) e solar hour angle at the moment t;
Fg ¼ m (5)
b e the angle between the tilted surface and the horizontal
with: plane.

m_ w e total water mass flow through all radiators; Considering that all surfaces of the envelope are perpendicular
cw e specific heat capacity of the water. to the horizontal plane (the roof is not a part of the envelope but it
completely shades the ceiling so the ceiling is not exposed to solar
The problem with this input is not in measuring it but in controlling radiation), the incidence angle of beam radiation on the walls is:
it. In the obtained building model, the command is considered to be
the heat flow Vg. However, in practice, the command is usually the cosðaT Þ ¼ sinðdÞcosðfÞcosðgÞ þ cosðdÞsinðfÞcosðgÞcosðuÞ
water inlet temperature and not the heat flow. Therefore, in Section þcosðdÞsinðgÞsinðuÞ (9)
3.2 the relation between the inlet water temperature and the cor-
and total solar radiation on a wall is:
responding heat flux is established, allowing the linearization of
the model for the third input.  
rg 1 þ rg
IT ¼ Ib Rb þ þ Id (10)
2 2
3.1. Solar heat flux
Once total solar radiation is calculated for each building side and
One input of the building model is the incident solar radiation on multiplied by the corresponding surface area, the input of the
the surfaces of the building envelope, Vs. Usually, this quantity is not building model is calculated by:
measured directly because only information about diffuse, Id, and
beam, Ib, radiations are available. Therefore, in order to estimate Vs, X
4
Fs ¼ ITi Si (11)
we need to determine the radiation on each surface, multiply it by i¼1
the corresponding surface area and add the results for all sides of the
envelope. Considering an isotropic model of the sky, the incident
solar radiation on a tilted surface, Fig. 3, is calculated by [34]:
3.2. Radiator heat flux
1 þ cosðbÞ 1  cosðbÞ
IT ¼ Ib Rb þ Id þ ðIb þ Id Þ rg (6) The third input of the building model is the heat flux delivered
2 2
by the radiators, Vg. However, for the control of water based
where the ground albedo rg is usually 0.2 and the ratio of beam heating systems, it is inappropriate to use a model with this input
radiation on a tilted surface to that on a horizontal surface is
because the controller cannot act directly on the heat flux. This
calculated by: energy flow depends on several parameters like indoor air
cosðaT Þ temperature, mean radiator temperature, interior walls surface
Rb ¼ (7) temperature and water mass flow through radiators. In practice,
cosðaÞ
the manipulated input is either water inlet temperature or water
The angles a and aT are the incidence angles of the beam radiation flow. Usually, it is avoided to act on water flow because it may
on the horizontal and tilted surfaces respectively, which are induce hydraulic instabilities. Thus, the question that naturally
calculated by: arises is why not to determine a model between the indoor air
temperature and the inlet water temperature. Hereafter is given
cosðaÞ ¼ sinðdÞsinðfÞ þ cosðdÞcosðfÞcosðuÞ
a reason why it is preferable to not do so and how to overcome this
cosðaT Þ ¼ sinðdÞsinðfÞcosðbÞ  sinðdÞcosðfÞsinðbÞcosðgÞ
problem.
þcosðdÞcosðfÞcosðbÞcosðuÞ þ cosðdÞsinðfÞsinðbÞ Thermal energy coming from the radiators is transmitted to
cosðgÞcosðuÞ þ cosðdÞsinðbÞsinðgÞsinðuÞ (8) the indoor environment via convection and radiation. Generally,
the ratio between these two is considered to be 50%, although it
with: may vary depending on many factors. The heat transferred
through convection depends on the difference between the mean
d ¼ 23,45$sin(360(284 þ n)/365) e solar declination in the nth radiator temperature and indoor air temperature while the heat
day of the year; transferred through radiation depends on the difference between
the mean radiator temperature and the temperature of the
surfaces inside the building. However, all the indoor surfaces,
except the interior surfaces of the envelope, are part of the
internal mass. Therefore, for simplicity it is considered that the
radiative heat transfer depends on the difference between radi-
ators’ surface temperature and indoor thermal mass
temperature.
Very often, in models used for control, the radiative heat
transfer is neglected or it is approximated by a linear law [24,29]. If
the difference between the temperatures of the surfaces is not too
large, the radiative heat flux density transferred from a hot surface
Fig. 3. Incident solar radiation on a tilted surface. to a colder one is calculated by:
384 I. Hazyuk et al. / Building and Environment 51 (2012) 379e387

 
4
4r ¼ 3 s Thot 4
 Tcold (12) the entire operating range of the temperature. A simulation based
experiment reveled a nonlinear correlation between the heat flux
where: delivered by radiators and the temperature difference between
inlet water and mean zone temperature, Fig. 4(a). The nonlinearity
3 e surface emissivity; is better illustrated in Fig. 4(b) where total conductance is calcu-
s e StefaneBoltzmann constant; lated by using the measured data from Fig. 4(a):
T e surface temperature in Kelvin degrees.
Fg
Srad hT ¼ (17)
The relation (12) can be written as: qin  qz
   Since in equation (17) the surface area of the radiators, Srad, is
2 2
4r ¼ 3 sðThot þ Tcold Þ Thot þ Tcold Thot  Tcold (13) constant, it is the total heat transfer coefficient, hT, which varies
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} considerable on the operating temperature range, Fig. 4(b).
hr
The heat emitted by a radiator is related to the temperature
The linearization of this relation is based on the following difference qrad  qz and not to qin  qz. However, the controller can
consideration: in buildings, the hot surface is the radiator and its manipulate the inlet water temperature and not the radiator mean
temperature varies between 20 and 60  C. The temperature of the temperature, qrad . Therefore, the temperature difference qin  qz
cold surfaces varies roughly between 15 and 20  C. As in the rela- needs to be used instead of qrad  qz for the calculation of the total
tion (13) temperature is on the absolute scale, T[K] ¼ 273 þ qsurface conductance in equation (17). To express this difference, the radi-
[ C], its usual variations are relatively small compared to its abso- ator surface temperature is approximated by the average temper-
lute value. Thus, these variations do not have a significant relative ature of the inlet and outlet water:
impact on the first two parentheses and therefore they are
approximated by a constant heat transfer coefficient, hr, by taking qin þ qout
qrad ¼ (18)
the mean temperature of the surfaces. 2
Concerning the convective heat transfer, the heat flux density The heat balance equation of the radiator is:
for a vertical surface, taller than 30 cm is described by the following
law [23]: _ w cw ðqin  qout Þ ¼ hT Srad ðqrad  qz Þ
m (19)
   
4c ¼ 1:78 Tsurf  Tair 1=4 Tsurf  Tair (14) The mean radiator temperature can also be obtained by eliminating
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} the outlet temperature qout from the relations (18) and (19):
hc
_ w cw qin þ hT Srad qz
2m
qrad ¼ (20)
Here, the first parenthesis is also considered to be approxi- 2m_ w cw þ hT Srad
mately constant and equal to the convective heat transfer coeffi-
By extracting the zone temperature from both sides of the equation
cient, hc. By making the hypothesis that radiative and convective
(20), it is shown that qrad  qz can be replaced by qin  qz in the
heat transfer coefficients are constant, the total heat flux density is
calculation of the total conductance:
obtained by summing the relations (13) and (14):
2m_ w cw
4T ¼ 4c þ 4r ¼ ðhc þ hr Þ ðThot  Tcold Þ (15) qrad  qz ¼ ðq  qz Þ (21)
|fflfflfflfflfflffl{zfflfflfflfflfflffl} _ w cw þ hT Srad in
2m
hT
The correlation between the total heat transfer coefficient and
The total heat flux delivered by radiators is obtained by multiplying the temperature difference qin  qz is estimated by making a curve
the total heat flux density, 4T, by the radiators surface area Srad: fitting on the characteristic from Fig. 4(b). The best fitting was
obtained for an exponential correlation:
Fg ¼ Srad 4T (16)
0:2544
Srad hT ¼ 36:85ðqin  qz Þ (22)
Since the temperature difference between radiators and thermal
mass varies usually between zero and forty to fifty degrees, the Thus, the relation between the heat flux delivered by the radi-
total heat transfer coefficient, hT, cannot be considered constant on ators and the inlet water temperature is given by:

Fig. 4. (a) Nonlinear correlation between the temperature difference and the corresponding heat flux and (b) the variation of the total conductance depending on the temperature
difference.
 I. Hazyuk et al. / Building and Environment 51 (2012) 379e387 385

2  
Fg ¼ Srad hT ðqin  qz Þ (23) qz z1 3 2
b11 z1 þ b12 z2
3
 
6 qo z1 7 6 1 þ a z1 þ a z2 7
6 7 6 7
 6 q z1  7
where SradhT is given by the equation (22). This relation will be used 1 2
 6 7
in the second part of the paper for the linearization of the building 1 6 z 7 6 b21 z1 þ b22 z2 7
H z h6 7 ¼ 6 7 (24)
model from the relation (5). 6 Fs z1 7 6 1 þ a1 z1 þ a2 z2 7
6   7 6 7
4 qz z1 5 4 b31 z þ b32 z
1 2 5
 
Fg z1 1 þ a1 z1 þ a2 z2
4. Model parameter identification
This representation is obtained by time discretization of the
Identifying the parameters of a real building is difficult because continuous transfer function model of the building. The continuous
the inputs cannot be varied and the disturbances are very impor- transfer function is obtained by applying the transformation (2) on
tant. An alternative is to identify the parameters of the model given the state-space model from (4). This representation of the model
by equation (5) from input/output data records obtained by simu- permits to have a linear formulation of the identification problem,
lating a detailed model of a reference building. The considered which guarantees the optimality of the solution. Therefore, in this
reference building is a typical detached house, having a living paper the least squares identification method is adopted.
surface of 100 m2 and a volume of 252 m3 (Fig. 5). This is one of the Here, the robustness of the model is given by the fact that
reference buildings in France, which is used by CSTB for perfor- detailed white-box model is projected on a reduced structure
mance evaluation. Its detailed “white-box” model is implemented which was obtained by using physical knowledge. This low-order
in Simbad toolbox and was experimentally validated by CSTB on model is an approximation of the building thermal behavior and
real buildings. therefore it must be accepted that there will always be dissimilar-
ities between the model and the building response. Trying to
4.1. Identification method choice reduce this error by varying model parameters and its structure, we
risk over fitting: a good similarity is obtained for the input signals
Given the abundance of system identification methods, an used during the identification, but when the inputs are changed
appropriated one must be chosen for our application. Grapho- there may be unacceptable differences between the responses. On
analytical methods for impulse or step response are unsuitable the contrary, keeping the model structure as it results from physical
because the building model contains zeros [27] and these analysis will avoid this problem because physical phenomena will
methods are not able to identify them. Therefore, parametrical not change when inputs are varying within the operating range.
identification methods must be employed. Basic algorithms for
4.2. Identification process
the identification of optimal values of the parameters are iterative
min-search and least squares methods. The first ones are used
For the identification of a transfer function model, the system
mainly for situations where the values of physical parameters are
must be in stationary initial conditions, before applying the excita-
investigated. For example this is the case of building energy
tion. In our case, the system states are the zone and wall tempera-
performance assessments [10]. However, in this case the model is
tures, and they were set to zero. Secondly, adequate excitation
represented as a nonlinear correlation between its parameters, so
signals must be chosen, whose role is to excite the system modes. For
it is difficult to guarantee the optimality of the solution. To
outdoor temperature and solar radiation were used statistical
increase the chance of finding the global optimum, the initial
weather records offered by the simulation program. On the contrary,
values need to be close to the optimal solution or constraints need
the imposed internal heat flux was obtained by switching the inlet
to be included in the algorithm in order to bound the physical
water temperature between 20 and 60  C according to a pseudo-
values of the parameters. In the case of the control however,
random binary sequence (PRBS). The building was simulated for
information about physical parameters is not mandatory; a robust
a period of four months using weather data records for Decem-
model is enough in order to predict the building behavior. The
bereMarch from Lyon, France, and a sampling time of one minute.
least squares identification method estimates the parameters of
The mean temperature of the building, which is the model output,
the discrete transfer function representation of the system. In the
was computed as a weighted average of the temperatures from each
case of our low-order building model, the parameters
room. Then, this data set was divided in two halves in order to use
a1 ; a2 ; b11 ; /b32 from the following model representation are
different records for parameters fitting and model validation. The fit
identified:
between the low-order model with identified parameters and the
measured data [35]:
 
normðy^  yÞ
fit ¼ 1 $100% (25)
normðy  meanðyÞÞ
is a measure of the quality of the model; a higher number means
a better model.
The identified transfer function model of the building using
simulation data for sixty days is:
2  
qz z1 3 2 2:284,103 z1  2:283,103 z2 3
 
6 qo z1 7 6 7
6   7 6 1  1:991z1 þ 0:9907z2 7
  6 q z1 7 6 9:033,107 z1  9:024,107 z2 7
1 6 z 7 6
7 ¼ 6 7
H z h6 7
6 Fs z1 7 6 1  1:991z1 þ 0:9907z2 7
6  7 4 1:589,105 z1  1:589,105 z2 5
4 qz z1 5
 
Fg z1 1  1:991z1 þ 0:9907z2
(26)
Fig. 5. The blueprint of the reference building.
386 I. Hazyuk et al. / Building and Environment 51 (2012) 379e387

Fig. 6. Comparison between the mean temperature of the building obtained by simulating the detailed white-box model and the low-order model; (top) during parameter
identification, (bottom) during model validation.

The fit between the temperatures calculated by the detailed and Thus, the low-order building model structure was obtained by
low-order model for the data used for parameter identification solving its equivalent electrical circuit representation and the
(Fig. 6 on top) is 96%. The same comparison, but this time for the parameter values were identified from data obtained by simu-
last two months of simulation, with different excitation signals lating its detailed model in a dedicated simulation toolbox [22].
(Fig. 6 on bottom), the obtained fit is 93%. It can be noticed that the The building was characterized by a single-zone lumped capacity
fit for data used for model validation is not as good as for data used model for which the state-space representation was calculated.
for parameter identification. Nevertheless, in both cases the ob- Initially, four inputs were considered (outdoor and ground
tained low-order model reproduces very well the dynamic of the temperatures, solar radiation and internal heat flux) but the model
mean temperature of the building. These results show that a second analysis has shown that only three of them have a predominant
order model is well fitted to describe the building thermal behavior, impact on the model dynamics; therefore, the ground temperature
at least for control purpose. was omitted. The output of the model was defined as the mean
temperature of the rooms, averaged by their surfaces. As in
simulation there were no disturbance inputs that could be an
5. Conclusions
additional source of incertitude, it was possible to use the standard
least squares method for parameter identification. The building
The paper presents a possible way to obtain a low-order model
was excited for a four month period using statistical weather
of the building thermal behavior, which is usually required by
records for solar radiation and outdoor temperature, and the heat
model-based temperature controllers.
flux resulted by switching the radiators inlet water temperature
Modeling of building thermal behavior for control purpose is
between 20 and 60  C, according to a PRBS. The simulation results
particularly difficult because of the large number of states (modes).
were divided in two parts; the first part of the sequence was used
Therefore, they are often represented by lumped parameters, which
for parameter identification and the second one for model vali-
permit to obtain directly a low-order model. In order to identify
dation. Model validation has shown that a second order model can
correctly the parameters, the building must be excited by inputs
reproduce the building thermal behavior with a good fidelity (the
having a large spectrum of frequencies and the output must have
fit between the results is larger than 90%).
enough variations to observe the building dynamics. The problem is
As in the control process the real manipulated input is the inlet
that in inhabited buildings it is difficult to have such conditions
water temperature and not the heat flux delivered by the radiators,
because of imposed thermal comfort norms, prohibitive costs and
it was shown that the relation between these two is nonlinear.
technological constraints. Also inhabited buildings contain plenty
Therefore this correlation was identified in order to be used for
of heat sources that influence the indoor temperature, which may
model linearization in the second part of the paper.
degrade the identified model quality.
The proposed approach is to generate the required input/output
data records by simulating a detailed model of the building instead Acknowledgments
of measuring them on the real building. This approach, which is
directly applicable to new buildings, allows us to applying This work has been completed in the project “Intelligent
a pseudo-random binary signal (PRBS) for internal heat flux and to Buildings and Rational Management of Renewable Energy” sup-
eliminate the sources of disturbances, but also to set particular ported financially by the French National Agency for Research. The
initial condition, required by the identification method, and to carry doctoral work of the author Ion Hazyuk was supported by the
out specific simulations for the identification of the nonlinear French National Agency for Research and by the Romanian Ministry
characteristic, present in building thermal behavior. for Education and Research.
 I. Hazyuk et al. / Building and Environment 51 (2012) 379e387 387

References [17] Noël J, Roux JJ, Schneider PS. CODYBA, a design tool for buildings performance
simulation. In: Proceedings of the 7th International IBPSA Conference
“Building Simulation 2001”. Rio de Janeiro, Brazil; 2001. p. 67e74.
[1] Dounis AI, Caraiscos C. Advanced control systems engineering for energy and
[18] Klein SA, Beckman WA, Mitchell JW, Duffie JA, Duffie NA, Freeman TL, et al.
comfort management in a building environment e a review. Renew Sust
TRNSYS 16 e mathematical reference. Solar Energy Laboratory, University of
Energ Rev 2009;13:1246e61.
Wisconsin-Madison; 2004.
[2] Liao Z, Swainson M, Dexter AL. On the control of heating systems in the UK.
[19] Peuportier B, Sommereux IB. COMFIE: passive solar design tool for multizone
Build Environ 2005;40:343e51.
buildings e user’s manual. École des Mines de Paris: Centre d’énergétique;
[3] Peeters L, Van der Veken J, Hens H, Helsen L, D’Haeseleer W. Control of heating
1994.
systems in residential buildings: current practice. Energ Buildings 2008;40:
[20] Clarke JA. Energy simulation in building design. 2nd ed. Oxford: Butterworth-
1446e55.
Heinemann; 2001.
[4] Kummert M, André P, Nicolas J. Optimised thermal zone controller for inte-
[21] EnergyPlus. EnergyPlus engineering reference: the reference to EnergyPlus
gration within a Building Energy Management System. In: Proceedings of
calculations. Lawrence Berkeley National Laboratory; 2009.
CLIMA 2000 conference; 1997. Brussels, Belgium.
[22] Husaunndee A, Lahrech R, Vaezi-Nejad H, Visier JC. SIMBAD: a simulation
[5] Kolokotsa D, Pouliezos A, Stavrakakis G, Lazos C. Predictive control techniques
toolbox for the design and test of HVAC control systems. 5th international
for energy and indoor environmental quality management in buildings. Build
IBPSA conference. Prague, Czech Republic; 1997. p. 269e76.
Environ 2009;44:1850e63.
[23] Incropera FP, DeWitt DP, Bergman TL, Lavine AS. Fundamentals of heat and
[6] Prívara S, Siroký J, Ferkl L, Cigler J. Model predictive control of a building
mass transfer. 6th ed. John Wiley & Sons; 2006.
heating system: the first experience. Energ Buildings 2011;43:564e72.
[24] Achterbosch GGJ, de Jong PPG, Krist-Spit CE, van der Meulen SF, Verberne J.
[7] Palomo Del Barrio E, Lefebvre G, Behar P, Bailly N. Using model size reduction
The development of a comvenient thermal dynamic building model. Energ
techniques for thermal control applications in buildings. Energ Buildings
Buildings 1985;8:183e96.
2000;33:1e14.
[25] Bénard C, Guerrier B, Rosset-Louërat M-M. Optimal building energy
[8] Sempey A, Inard C, Ghiaus C, Allery C. Fast simulation of temperature distri-
management: part I e modeling. J Sol Energ Eng 1992;114:2e12.
bution in air conditioned rooms by using proper orthogonal decomposition.
[26] García-Sanz M. A reduced model of central heating systems as a realistic
Build Environ 2009;44:280e9.
scenario for analyzing control strategies. Appl Math Model 1997;21:535e45.
[9] Ríos-Moreno GJ, Trejo-Perea M, Castañeda-Miranda R, Hernández-
[27] Ghiaus C, Hazyuk I. Calculation of optimal thermal load of intermittently
Guzmán VM, Herrera-Ruiz G. Modelling temperature in intelligent build-
heated buildings. Energ Buildings 2010;42:1248e58.
ings by means of autoregressive models. Automat Construct 2007;16:
[28] Madsen H, Holst J. Estimation of continuous-time models for the heat
713e22.
dynamics of a building. Energ Buildings 1995;22:67e79.
[10] Mejri O, Palomo Del Barrio E, Ghrab-Morcos N. Energy performance assess-
[29] Liao Z, Dexter AL. A simplified physical model for estimating the average air
ment of occupied buildings using model identification techniques. Energ
temperature in multi-zone heating systems. Build Environ 2004;39:1013e22.
Buildings 2011;43:285e99.
[30] Wang S, Xu X. Simplified building model for transient thermal performance
[11] Coley DA, Penman JM. Simplified thermal response modelling in building
estimation using GA-based parameter identification. Int J Therm Sci 2006;45:
energy management. Paper III: demonstration of a working controller. Build
419e32.
Environ 1996;31:93e7.
[31] Gouda MM, Danaher S, Underwood CP. Building thermal model reduction
[12] Jiménez MJ, Madsen H, Andersen KK. Identification of the main thermal
using nonlinear constrained optimization. Build Environ 2002;37:1255e65.
characteristics of building components using MATLAB. Build Environ 2008;43:
[32] Kummert M, André P, Nicolas J. Development of simplified models for solar
170e80.
buildings optimal control. In: Proceedings of ISES EuroSun 96 Congress.
[13] Kummert M, André P, Nicolas J. Optimal heating control in a passive solar
Freiburg, Germany; 1996. p. 1055e61.
commercial building. Solar Energy 2001;69:103e16.
[33] Xu X, Wang S. Optimal simplified thermal models of building envelope based
[14] Chen TY, Athienitis AK. Investigation of practical issues in building thermal
on frequency domain regression using genetic algorithm. Energ Buildings
parameter estimation. Build Environ 2003;38:1027e38.
2007;39:525e36.
[15] Ghiaus C, Chicinas A, Inard C. Grey-box identification of air-handling unit
[34] Duffie JA, Beckman WA. Solar engineering of thermal processes. 3rd ed.
elements. Control Eng Pract 2007;15:421e33.
Hoboken, NJ: John Wiley & Sons; 2006.
[16] Söderström T, Stoica P. System identification. Prentice Hall International;
[35] Ljung L. System identification toolbox 7 user’s guide. MathWorks; 2007.
2001.