Você está na página 1de 10

Automation in Construction 16 (2007) 713 722

www.elsevier.com/locate/autcon

Modelling temperature in intelligent buildings by means of


autoregressive models
G.J. Ros-Moreno , M. Trejo-Perea, R. Castaeda-Miranda,
V.M. Hernndez-Guzmn, G. Herrera-Ruiz
Faculty of Engineering, Universidad Autnoma de Quertaro, Cerro de las Campanas s/n, CP 76010, Quertaro, Qro., Mxico
Accepted 20 November 2006

Abstract
One of the main problems of the intelligent buildings is to give comfort to its occupants and to increase the user's performance at a low cost.
The excessive demand of electric energy due to heating, ventilating, and air-conditioning (HVAC) systems require temperature forecast and
control to make maximum reduction of the electrical energy. The objective of this paper is to investigate in what extent linear autoregressive
models with external input (ARX) and autoregressive moving average models with external input (ARMAX), could be used in order to predict the
interior air temperature of a building. In particular, the obtained results in the classrooms of the Universidad Autnoma de Quertaro, U.A.Q.,
Mxico, are shown. Outside air temperature, global solar radiation flux, outside air relative humidity and air velocity were used as the input
variables. The obtained results showed that the ARX models give a better prediction of the temperature than the ARMAX models, obtaining the
best results with the ARX (2,3,0) with a coefficient of determination of 0.9457 and ARX (2,2,1) with a coefficient of determination of 0.9056.
2006 Elsevier B.V. All rights reserved.
Keywords: Intelligent buildings; Comfort; Autoregressive models

1. Introduction
An important goal of the intelligent buildings is to improve
users' comfort and security with a global reduction in the
consumption of energy [1]. The inside air quality and thermal
comfort should be assured by means of an appropriate heating
and ventilation of the spaces. The consumption of energy by
heating, ventilating and air-conditioning (HVAC) equipment in
industrial and commercial buildings constitutes 50% of the
world energy consumption [2]. The high energetic consumption
in the HVAC systems is related to the use of inefficient control
operation sequences and to failures in the system. Experimental
researches and theoretical studies have demonstrated that a
potential saving is possible by improving the handling of energy
establishing the correct or optimal operation of the HVAC
systems in commercial buildings. The energetic consumption
Corresponding author. Fax: +52 442 192 12 00+1+6064+#.
E-mail address: riosg@uaq.mx (G.J. Ros-Moreno).
0926-5805/$ - see front matter 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.autcon.2006.11.003

can be reduced between 20 and 30% and simultaneously solve


some comfort problems [3].
Engineers and architects have influenced the energy use of
the new buildings through the design of the envelope, the
HVAC systems selection and the operation sequences specification; however, once the building has been finished, the
consumption of energy is decided mainly by its control,
maintenance and by its occupants.
An important factor affecting the efficiency of the HVAC
systems is that most of them are set to operate at design thermal
loads while actual thermal loads, which affect the system, are
time-varying. Therefore, control schemes that consider timevarying loads, could operate more efficiently and keep better
comfort conditions than conventional control schemes. The
adaptable control techniques are necessary in HVAC applications because the characteristics of the process change
continuously by effect of the variations in the climate and
occupation of the building. An adaptable control system
depends on a dynamic model that describes the parameters

714

G.J. Ros-Moreno et al. / Automation in Construction 16 (2007) 713722

which draw the behavior of a system; the more accurate the


pattern is to the real characteristics of the process, the more
efficient control actions can be accomplished.
The parameters' adjustment of a model used to represent a
system is called System Identification. The parameters' estimation of the continuous systems in the time is a relevant topic that
has several applications which range from control and signal
processing to astrophysics and economy [4]. This is due to the
fact that the majority of the systems or physical phenomenons
are from continuous nature in time. With the arriving of the
digital computers, research for the control and identification of
these continuous systems and processes in time has been
concentrated in the discretized models with samples of inputs
and outputs of the fundamental systems. A discrete model of the
continuous climatic system in time of a building can be obtained
in several ways; one of them is by an autoregressive relationship
between the discrete output y(t) and the input u(t). Two
structures for autoregressive models commonly used in the
estimation of systems are the autoregressive models with
external input (ARX) and autoregressive moving average
model with external input (ARMAX).
The methods of identification are often classified as grey-box
(semi-physical) or black-box models (non-physical). The greybox models have a structure that is based partly on physical,
chemical or biological laws (like deterministic models) and
partly on empiricism. In the black-box models previous
knowledge of the system is not needed, this could be an
advantage if the information of the system dynamics is limited,
but it implies the problem of selecting an appropriate structure
for the pattern. Another advantage is the possibility of obtaining
a vague model with a relatively small group of measurements.
The pattern can be improved by introducing new data.
In literature, numerous works have tried to respond to the
demand of thermal comfort and energy saving. The development of a dynamic model is a fundamental part of the proposals.
The work presented by Loveday [5] points toward the utility
and the potential of using the models of temperature to handle
resources in intelligent buildings. Some studies related to the
control methods of the HVAC systems based on models of the
system are presented in [611]. Another way in which the
models can be used is by applying them in different strategies of
detecting, locating and predicting the presence of defects which
cause incorrect operations. In [3] an autoregressive algorithm
was used for the development of a detection dynamic model and
a failure diagnosis of a HVAC system. Recent studies focused
on the comfort evaluation in classrooms can be found in Krger
[12] who evaluated classrooms in Brazil, highlighting four

Fig. 1. Blocks diagram of the pattern of black box used to predict the
temperature of the inside air in classrooms.

Fig. 2. Typical school building viewed from the south.

different study areas of comfort: thermal, luminic, acoustic and


ergonomic. The results obtained in different evaluations showed
a strong interdependence among the first three areas. For
evaluating thermal comfort, he considered solar orientation and
building materials on walls and roofs. Hwang [13] presented the
methodology of the ASHRAE standard 55 for the study of
thermal comfort in classrooms in Taiwan. In order to obtain the
data, surveys and field experiments in 26 air-conditioned
classrooms and 10 ventilated in a natural way were applied. The
results demonstrated that air temperature, air movement and
average radiant temperature have a significant influence, but
humidity does not make any statistical significance. Hanqing
[14] carried out a study based on Fanger's thermal comfort
concept [15]. He proposed several concepts by using computing
results obtained from a large eddy simulation (LES). The author
established three indexes of the thermal comfort based on time
averaged parameters, different from Fanger's concept, pointing
out the differences of each one and giving their definition. The
study showed that there are differences between the heat and
energy dissipation model of the human body at the walking state
and the model of quiet seating postures. However, he mentions
that this does not influence thermal comfort concepts, because
several heat and energy dissipation models can be applied to the
calculation equations in order to estimate the three indexes of
the thermal comfort.
The purpose of this work is to determine the appropriate
structure of the (ARX and ARMAX) models to predict the
temperature in the interior of a building (classrooms). It also
pretends to study the external climate variables that should be
included in these models to achieve the best estimation. The
dynamic model of autoregressive structure can become a
practical alternative for the implementation of an adaptable
controller that will allow to increase the building occupants'
comfort and at the same time, to reduce the electric energy
consumption of the HVAC systems.
This paper is organized as follows: in Section 2 the theoretical considerations for the identification of the autoregressive
systems are presented. Section 3 describes the materials and

G.J. Ros-Moreno et al. / Automation in Construction 16 (2007) 713722

715

Fig. 3. Typical plan view (dimensions in meters).

methods for the field measurements and also presents the


accuracy measurements to explain the variation of the
temperature data. In Section 4 a discussion of the obtained
results of the selected models ARX and ARMAX is carried out.
Finally, the conclusions and the future works are presented in
Section 5.
2. Theoretical considerations
ASHRAE STANDARD, claims that to determine a comfort
zone it is necessary to take into account the operative
temperature for given values of humidity, air speed, metabolic
rate, and clothing insulation. The comfort zone is defined in
terms of a range of operative temperatures that provide
acceptable thermal environmental conditions, or in terms of
the combination of air temperature and mean radiant temperature that people find thermally acceptable [16]. The excess of
heat, coming from the atmosphere or generated by their own
metabolism, should be eliminated in order to maintain a
constant temperature in the body and to assure thermal comfort.
By sure, the sensation of thermal comfort is given by the

Fig. 4. Distribution of the sensors in the interior of the classroom 19. Superior
view, plane XY.

climatic conditions, the heat production of the human


metabolism and the transfer of heat with the environment.
In a closed atmosphere there are six main elements that
determine the perception of the environmental quality from the
thermal point of view: Temperature of the air (T), solar radiation
(Ra), wind speed (Vw), relative humidity (Rh), average of the
instantaneous air velocity over an interval of time (Va) [13,16]
and occupation (Oc).
The identification of the system is simplified by omitting
the Oc and Va variables to calculate the model of the classroom building. On one hand, the difficulty of being able to
accurately measure the dissipation of energy of the human
body of the Oc individuals inside the interior of the classrooms is presented. On the other, due to the fact that the
behavior of the classroom becomes highly non-linear during
HVAC ventilation operation inclusion of the Va effect in the
models is discarded.
Based on this, the proposed mathematical model to predict
the temperature inside of the intelligent building of classrooms in this research is based on the analysis of the following
input variables: outside air temperature (To), global solar
radiation flux (Ra), wind speed (Vw), outside air relative
humidity (Rho); being the inside air temperature the output
variable (Ti) (Fig. 1).
Although the climate characteristics are continuous variables, they are measured and registered in discrete form. In the
discrete domain, the dynamic system of the building can be
modeled by means of linear autoregressive relations between
the discrete output y(t) and the discrete input u(t), such as the
autoregressive models with external input (ARX) and autoregressive moving average models with external input
(ARMAX). Assuming a single inputsingle output system,
the following expression can be used to describe this
relationship [4]:
yt a1 yt1 : : : ana ytna b1 ut1
: : : bnb utnb et c1 et1
: : : cn etnc
c

716

G.J. Ros-Moreno et al. / Automation in Construction 16 (2007) 713722

For a multivariable system [4] in which the number of inputs


is given by nu and the number of outputs by ny, A(z) and B(z)
are ny by ny and nu by nu matrices, respectively, which
elements are the polynomials in the shift operator z m (with m
as any natural number). The entries aij (z) and bij (z) of the
matrices A(z) and B(z), respectively, can then be expressed as:
aij z dij a1ij z1 : : : anaij znaij

10

and,
bij z b1ij znkij : : : bnkij znkij nbij 1

11

Fig. 5. Distribution of the sensors inside of the classroom 19. Lateral right view.

where: y = output signal, u = input signal, t = discrete time,


ana = model parameters, bnb = model parameters, cnc = model
parameters, e = error, na = number of poles (positive integer),
nb = number of zeros increased by one (positive integer) and
nc = order of the measured error (positive integer).
The determination of all or some values is obtained by
means of the estimation procedures, that is, the coefficients
enter as parameters to be determined. will be named as the
vector that has all the parameters to estimate. The description
of the model is the following:
yt; Gz; ut Hz; et:

Where ij represents Kronecker symbol.


According to this, it is clear that the ARX structure has a
system which can be defined by means of the number of the
poles na the number of zeros nb 1 and the time delay nk. The
definition of the ARMAX structure additionally requires the
order of the measured error nc to be known. The matrices A(z),
B(z) and the vector C(z) are determined by means of off-line
parameter identification methods.

Because the white noise term enters as a direct error to the


difference equations, the model of Eq. (1) is also known as
the model or structure of error equation. In this case the
parameters of adjustment will be:
a1

a2 : : :

ana

b1

b2 : : :

bnb T

For ARMAX, the Eq. (1) is often represented as:


Azyt Bzutnk Czet

And for ARX as:


Azyt Bzutnk et

Where the matrices A(z) and B(z) and the vector C(z) are
given by:
Az : 1 a1 z1 : : : ana zna

Bz : 1 b1 z1 : : : bnb znb

Cz : 1 c1 z1 : : : cnc znc

and z 1 is the backward shift operator


z1 ut ut1:

Fig. 6. Original data versus results of the (a) autoregressive moving average
model with external input and the (b) autoregressive model with external input
and the difference between both models for the first period of sampling; ,__,
measured;......, estimated.

a1

a2

a3

First period of sampling


ARMAX
4,1,3,0 1.88 1.84 1.32
4,1,2,0 1.83 1.76 1.25
ARX
1,3,0
0.98 0
0
3,3,0
1.28 0.26
0.03
2,3,0
1.29 0.30
0

a4

a5

b11

b12

b13

b21

2.36 10 3 0
2.50 10 3 0

0.39 0
0.37 0

5.87 10 4 0
6.33 10 4 0

0
0

0
0
0

7.14 10 4 3.67 10 3
7.33 10 4 2.98 103
7.54 10 4 2.90 10 3

3.93 10 3 0.02
3.38 10 3 0.02
3.35 10 3
0.02

0
0
0

Second period of sampling


ARMAX
5,3,1,1 2.14 1.34 0.09 0.12 0.02
5,3,2,1 2.42 1.96 0.50 0.08 0.05
ARX
2,2,1
1.50 0.51
0
0
0
3,3,1
1.41 0.30
0.11
0
0
3,2,1
1.41 0.30
0.11
0
0

5.56 10 4 6.17 10 4 9.17 10 5


5.31 10 4 6.63 10 4 1.57 10 4
0
0
0

6.71 10 4
6.20 10 4
6.28 10 4

0.02
0.02

b22

b23

b31

0
0

0.03
0.03

b32

b33

0
0

0
0

b41

0.02
0.03

b42

b43

0
0

0
0

c1

c2

0.48 0.77 0.02


0.43 0.74
0

0.02
0.03
0.03

3.23 10 4 0.51 0.28 0.21 0.10


0.01
0.47
0.44
0.02 0.02
0.01
0.47
0.44 0.02
0.03

0.12
0.22
0.15
0.17
0.15 0.18

0.04
0.04

0.02
0.02

0.03 0.01 0.73 0


0.08
0.01 1.01 0.22

4.57 10 4 1.56 10 4
1.15 10 3 0
3.47 10 4 2.33 10 3 7.93 10 3 7.13 10 3
4.37 10 4 1.79 10 3 2.89 10 4 0

0.11 0.17
0.10 0.16

0.06 0.05
0.06 0.06

0.32 0.32
0
0.01
0.29 0.24 0.04
3.07 10 3
0.32 0.31
0
4.73 10 3

0.02
0
0.05 0.04
0.02
0

c3

0
0
0

0
0
0

0
0
0

0
0
0

0
0
0

0
0
0
0
0

G.J. Ros-Moreno et al. / Automation in Construction 16 (2007) 713722

Table 1
Model structures and typical parameter identification results for the period models giving the inside air temperature (Ti) as a function of outside air temperature (To), global solar radiation flux (Ra), wind speed (Vw),
outside air relative humidity (Rho), and perturbation (e) in the discrete time (t) domain, using the z-transform operator (z 1); a1 to a5, b11 to b43, y, c1 to c3 are regressive coefficients as defined in the model structure
2
3
b11 z1 b12 z2 b13 z3
6 b21 z1 b22 z2 b23 z3 7
To t Ra t Vw t Rho t
1 c1 z1 c2 z2 c3 z3
7
Ti t
6
et
1
2
3 5
4
1
2
3
4
5
1
b31 z b32 z b33 z
1 a1 z a2 z a3 z a4 z a5 z
1 a1 z a2 z2 a3 z3 a4 z4 a5 z5
b41 z1 b42 z2 b43 z3

ARMAX: autoregressive moving average model with external input; ARX: autoregressive model with external input.

717

718

G.J. Ros-Moreno et al. / Automation in Construction 16 (2007) 713722

3. Materials and methods


3.1. Field measurements
The procedure was carried out in the graduate building of the
Engineering school at the Quertaro State University; the
building is shown in Fig. 2. This construction is equipped with
a saving and analysis digital system of the consumption and
quality of electrical energy called MONITO UAQ [17], which
allows a continuous monitoring of the power consumption and
electrical energy parameters in order to execute control actions
that prevent failures and enable energy savings. The structure
of the building has five classrooms, a meeting room and a
control room. The rooms chosen for this study were considered
to be identical.
A schematic plan view of one of these units shows the
relative position of the rooms, doors, windows and overhangs,
in addition to the overall dimensions (Fig. 3). Four classrooms
have an area of 48 m2; one classroom has an area of 40 m2, and
the meeting room an area of 56 m2. A control room, where the
recording and monitoring of the variables were done, has an
area of 16 m2. The height of all of them was 3 m.
The building has a 12 cm thick concrete layer on top. The
concrete layer was armed and has a resistance of 300 kg/cm2.
Lateral walls were built with 14 cm wide 30 cm long cored
bricks. With the purpose of providing a better illumination, the
design included two 4 m long 2 m high windows on the north
side and one on the south side. The walls dividing the rooms are
made of concrete blocks lined up with mortar with finished
interiors and exteriors. It has a 0.5 mm and 0.7 cm thick vinyl
painting. The top is covered by water proof Protexa layer and
red sand.
To measure the climate variables inside classroom 19, twelve
sensors of temperature were installed and distributed in the
interior of the room. Sensors were numbered from 0 to 11. Nine
sensors were distributed in plane XY, placed horizontally at a
distance of 1.5 m on the X axis, and placed on the Y axis at a
distance of 2 m (Fig. 4). Three sensors were placed in the superior
part of Z axis at a distance from floor to roof of 2.9 m (Fig. 5).
At the outside part of the building a resistance temperature
detector (RTD) of the Omega Engineering Inc. was placed in
order to measure temperature in C. A capacitive sensor was
used to measure the relative humidity (%). Wind speed (m/s)
was measured by an anemometer, and to measure the global
solar radiation (W/m2) a pyranometer sensor was utilized. These
components were installed at the southwest part of the
building at 4.5 m from the floor. All sensors were connected to
the TUNA SCCII v4 system, developed in the Biotronics
Laboratory of the Quertaro State University. This system was
configured to sample data every 5 min.
In order to be able to obtain the coefficients for the
mathematical model several measures were done to predict
temperature in the classrooms of the intelligent building. This
research is based on the analysis of the input variables, which
are: To, Ra, Vw, Rho, being the inside air temperature the output
variable (Ti). The measures of the foregoing variables were done
by sampling every 5 min during a period of 36 days.

According to Ljung [4] once the data have been recorded, the
first two thirds or so of the data record were used in order to
determine the model coefficients and the remaining data for
validation.
Even when the measures were done during a continuous
period of time, the data of the first 24 days was used to get the
coefficients of the model. Once these coefficients were
calculated, a prediction of the inside temperature for the next
12 days was done by comparing the predicted temperature of
the model and the real temperature taken in the classroom.
These relation is shown in Fig. 6(a and b).
The approximated temperate of the room was taken from the
average temperature of the 12 sensors for each sample. The
average value calculated corresponds to the inside temperature,
which was compared to the output data of models ARX and
ARMAX. Thus, from only one calculated model, the inside
temperature was calculated.
3.2. Measures of accuracy
The coefficient of determination (R2) is the measure of
correlation between the observations and predictions [18].

Fig. 7. Original data versus results of the (a) autoregressive moving average
model with external input and the (b) autoregressive model with external input
and the difference between both models for the first period of sampling; ,__,
measured;......, estimated.

G.J. Ros-Moreno et al. / Automation in Construction 16 (2007) 713722

Some measures of variances are the percent standard error of the


prediction (% SEP) [19], the coefficient of efficiency (E) [20]
and the average relative variance (ARV) [21]. These estimators
are not biased by the variation range of its elements. These are
used to determine how the model is able to explain the total
variance of the data. The percent standard error of the prediction
is defined as:
v
u N
uX
u
yk k 2
u
t
100 k1
%SEP
12
N
yk
Where yk is the observed output k of the pattern; k is the
estimate output for the pattern; N is the total number of the
generalization patterns and yk is the mean value of the observed
outputs of the prediction set.
The coefficient of the efficiency (E) and the average relative
variance (ARV) are expressed by:
E

Sobs S
S
; ARV
Sobs
Sobs

Sobs

N
X

yk
yk 2 ; S

k1

13

N
X

yk k 2

14

k1

Where Sobs is the measure of variability of the observed


values from their means and S is the measure of the association
between the predicted and observed values. For a perfect match,
R2 and E should be close to 1.0 and the values of %SEP and
(ARV) close to 0.
4. Result and discussion
4.1. Selection of the ARX and ARMAX models
Table 1 shows the coefficients of the selected models for the
first and second period of sampling. It can be observed that all
the selected models included the four variables of the external
climate. It was found that in most cases, the values of the outside

Table 2
Results of the validation of the ARMAX and ARX models corresponding to the
first sampling period
Model

Validation
R2

% SEP

ARV

ARMAX
4,1,3,0
4,1,2,0

0.9082
0.9107

0.9998
0.9997

0.0062
0.7320

2.2344 10 4
3.0476 10 4

ARX
1,3,0
3,3,0
2,3,0

0.9434
0.9455
0.9457

0.9996
0.9986
0.9993

0.85
1.54
1.22

4.1690 10 4
1.4102 10 3
7.4428 10 4

ARMAX: autoregressive moving average model with external input;


ARX: autoregressive model with external input.

719

Table 3
Results of the validation of the ARMAX and ARX models corresponding to the
second sampling period
Model

Validation
R2

% SEP

ARV

ARMAX
5,3,1,1
5,3,2,1

0.8861
0.8906

0.8525
0.8640

16.97
15.34

0.1475
0.1360

ARX
2,2,1
3,3,1
3,2,1

0.9056
0.9093
0.9096

0.9325
0.9363
0.9391

10.81
10.50
10.27

0.0675
0.0637
0.0609

ARMAX: autoregressive moving average model with external input;


ARX: autoregressive model with external input.

climate data which were older than 20 min did not improve the
model performance. In most cases, outside climate data older
than 15 min were not considered. Consequently, the resulting
models were rather compact. This should be considered as an
advantage since this allows their use in climate control
applications.
The coefficients of the ARMAX and ARX models of Table 1
of the first period with regard to the coefficients of the ARMAX
and ARX models of the second period show a minimum
variation. It is observed that most of the coefficients do not
differ and they present a similarity according to the two periods.
But the coefficients of the ARMAX models of the first period of
sampling with regard to the coefficients of the ARX models of
the first and second periods show a greater variation among
them. In a similar way the coefficients of the ARMAX models
of the second period of sampling with regard to the coefficients
of the ARX models of the second and first period of sampling
show a bigger variation.
In order to assess the accuracy of the ARMAX and ARX
models, the simulated results are compared with the original
data for the inside air temperature. Fig. 6(a and b) show an
example for both first and second period models. The difference
between the measured and the estimated inside air temperatures
is shown in these figures. The measured data are shown versus
the estimated ones for the ARMAX (4,1,2,0) model obtaining
an estimate of 82.22% and for the ARX (2,3,0) an estimate of
the 87.71%. It can be observed that the models which give a
better prediction of the temperature are the ARX models with
regard to the ARMAX models.
The assessment corresponding to the second period of
sampling can be observed in Fig. 7 showing the accuracy of the
ARMAX and ARX models. The simulated results are compared
with the real data for the temperature of the interior air. The
differences between the measured temperature and the estimated one in the interior of the classrooms between both models are
shown in Fig. 7(a and b). The ARMAX (5,3,2,1) model shows
an estimate of 82.27% and the ARX (3,2,1) an estimate of
87.33%. Based on the results of Figs. 6 and 7(a and b) the
models that give a better prediction of the temperature are the
ARX models with regard to the ARMAX models.
Based on the results in Table 1 and on Figs. 6 and 7(a and b)
it is observed that in both periods of sampling, the models

720

G.J. Ros-Moreno et al. / Automation in Construction 16 (2007) 713722

Fig. 8. Scatterplot of the observed versus estimated temperature variation for the
ARMAX (4,1,2,0) model corresponding to the first sampling period.

Fig. 10. Scatterplot of the observed versus estimated temperature variation for
the ARX (2,3,0) model corresponding to the first sampling period.

presented similar behaviors, but the ARX models are more


compact and give a better estimate for the internal temperature
of the air with regard to the ARMAX model. It can be also
appreciated that the best models for the ARMAX of the first and
second period are the models (4,1,2,0) and (5,3,2,1) and that
they have an estimate of 82.22 and 82.27% respectively. The
best models of the first and second period for the ARX are
(2,3,0) and (3,2,1) and they have an estimate of the 87.71 and
87.33% respectively.
The simulated results on the real data of the lineal regression
for the 5 selected models from a total of 21 are shown in Table 2.
It can be observed that the coefficient of determination is close
to 1 (that is 0.9082) and that the percent standard error of the
prediction is close to 0 (that is 0.015), while the coefficient of
efficiency is superior to 0.9992 and the average relative variance
is very close to 0 (that is 0 .00074). When the results of the
autoregression were compared, it could be observed that most of

the models gave good results although the performances of the


ARMAX models were not as good as those of the ARX models.
Table 3 shows the results of the lineal regression for the 5
selected models from a total of 36. It is observed that they had a
coefficient of determination close to 1 (that is 0.8861) and a
percent standard error of the prediction close to 0 (that is
0.1697), while the coefficient of efficiency is superior to 0.8525
and the average relative variance is very close to 0 (that is
0.1475). Comparing the results of autoregression it was
observed that the models gave good results although the
performances of the ARMAX models again were not as good
than those of the ARX models.
Comparing the results of Tables 2 and 3, it is observed that
the structures of the patterns of a bigger order, and in
consequence more complex showed a lower performance.
Finally, these clearly illustrate that the ARX models showed a
better performance than the ARMAX models.

Fig. 9. Scatterplot of the observed versus estimated temperature variation for the
ARMAX (5,3,2,1) model corresponding to the second sampling period.

Fig. 11. Scatterplot of the observed versus estimated temperature variation for
the ARX (3,2,1) model corresponding to the first sampling period.

G.J. Ros-Moreno et al. / Automation in Construction 16 (2007) 713722

The Scatterplot and the linear regression between the


observed data and the estimated ones are shown in Fig. 8.
The characteristics of the temperature variation provided the
lower deviations around the regression line. It is observed that
the best ARMAX model is the (4,1,2,0) where a (R2 = 0.9107)
coefficient of determination is presented, a percent standard
error of prediction of (% SEP = 0.7320), a coefficient of
efficiency of (E = 0.9997), besides average relative variance of
(AVR = 3.0476 10 4).
The linear regression for the ARMAX (5,3,2,1) model in
Fig. 9, presents a (R2 = 0.8906) coefficient of determination, a
standard error of prediction of (% SEP = 15.34), a coefficient of
efficiency of (E = 0.8640), besides average relative variance of
(AVR = 0.1360), showing a lower correlation in comparison
with the ARMAX (4,1,2,0) model.
The ARMAX models present good linear regressions
regarding to the measured data versus the observed ones, but
they do not improve the performances of the model ARX. The
ARX (2,3,0), Fig. 10, has the lowest dispersion around the
models shown in Table 2, presenting the best coefficient of
determination (R2 = 0.9457), one of the lower percent standard
error of prediction (% SEP = 1.22) and the best coefficient of
efficiency of (E = 0.9993) besides an average relative variance
of (AVR = 7.4428 10 4) which is the lowest.
In Fig. 11 a Scatterplot with good results can be appreciated.
It shows a low dispersion of the data, given by the ARX (3,2,1)
model, improving this way any of the performances of the
ARMAX models with a (R2 = 0.9096) coefficient of determination, a percent standard error of prediction of (%
SEP = 10.27), a coefficient of efficiency of (E = 0.9391) and an
average relative variance of (AVR = 0.0609), shown in Table 3.
It is observed that the ARMAX models give good
predictions of the interior temperature of the classrooms, but
these results are improved by the ARX models, due to a good
adjustment, the models should present a coefficient of
determination and a coefficient of efficiency close to 1; the
main characteristic that is observed in the results of the ARX
models according to the data that are given in Tables 2 and 3.
Likewise, they should present a percent standard error of
prediction and the average relative variance close to zero, a
characteristic that the ARX models also present in general; this
explains the capacity of the ARX models to give a better
prediction.
Figs. 811 show the dispersion diagram of the temperature
data. It can be considered that the models that will be included
are the ARX models as identification models of the building
for the prediction of the internal temperature. Besides, they
present coefficients of smaller order and they obtain a lower
dispersion; such case was observed with the model of Fig. 10,
ARX (2,3,0).
5. Conclusion and future work
This work studied how to apply the linear autoregressive
ARX and ARMAX models as an approach to the dynamic
behavior of the temperature of the air inside the classrooms. For
this, measurements of temperature, global solar radiation flux,

721

wind speed and outside air relative humidity were carried out as
input variables to the system, and different structures of ARX
and ARMAX models were intended. By means of programming, the acting indexes were calculated for each one of the
structures; selecting those models with better prediction of the
real conditions of interior temperature. Based on the selected
models, the influence of the input variables is discussed in the
accuracy of the models. The best prediction results were
obtained by the structures of ARX (2,3,0) model with a
coefficient of determination of 0.9457 and the ARX (3,2,1) with
a coefficient of determination of 0.9096.
The information presented in this work will serve as a
complement for a strategy of intelligent control in HVAC
equipments, where, based on the model of the controlled
building will determine the best control sequence in order to get
the desired reference with a minimum expense of energy.
References
[1] J.K.W. Wong, H. Li, S.W. Wang, Intelligent building research: a review,
Autom. Constr. 14 (2005) 143159.
[2] M.S. Imbabi, Computer validation of scale model test for building energy
simulation, Int. J. Energy Res. 14 (1990) 727736.
[3] H. Yoshida, S. Kumar, RARX algorithm based model development and
application to real time data for on-line fault detection in VAV AHU units,
IBPSA Building Simulation, vol. 99, 1999, pp. 161168.
[4] L. Ljung, System Identification, Theory for the User, Prentice-Hall, Upper
Saddle River, 1999.
[5] D.L. Loveday, G.S. Virk, J.Y.M. Cheung, D. Azzi, Intelligence in buildings:
the potential of advanced modelling, Autom. Constr. 6 (1997) 447461.
[6] R.M.C. De Keyser, Application of extended prediction self-adaptive
control, in: K. Warwick (Ed.), Implementation of Self-Tuning Controllers,
Peter Peregrinus, London, 1989, pp. 195219.
[7] A.L. Dexter, P. Haves, A robust self-tuning predictive controller for HVAC
applications, ASHRAE Trans. 95 (2) (1989).
[8] A.L. Dexter, P. Haves, Improved energy management of building
system environment though self-tuning control of HVAC plant, SERC
Final Report, Science and Engineering Research Council, Swindon,
1990.
[9] J.M. Dion, L. Dugard, A. Franco, N.M. Tri, D. Rey, MIMO adaptive
constrained predictive control case study: an environmental test chamber,
Automatica 27 (4) (1991) 611626.
[10] W.J. Graham, A.L. Dexter, Practical experience of self-tuning temperature control in an office building, Proc. Int. Symp. Resent Advances in
Control and Operation of Building HVAC Systems, Trondheim, 1984,
pp. 124133.
[11] A.R. Guesalaga, H.W. Kropholler, Improved temperature and humidity
control using Hx synthesis, Proc. IEE D137 (6) (1990) 374380.
[12] E.L. Krger, P.H.T. Zannin, Acoustic, thermal and luminous comfort in
classrooms, Build. Environ. 39 (2004) 10551063.
[13] R.L. Hwang, T.P. Lin, N.J. Kuo, Field experiments on thermal comfort in
campus classrooms in Taiwan, Energy Build. 1 (38) (2006) 5362.
[14] W. Hanqing, H. Chunhua, L. Zhiqiang, T. Guangfa, L. Yingyun, W.
Zhiyong, Dynamic evaluation of thermal comfort environment of airconditioned buildings, Build. Environ. 14 (11) (2006) 15221529.
[15] P.O. Fanger, A.K. Melikov, et al., Turbulence and draft, ASHRAE J. 31 (4)
(1989) 1823.
[16] ASHRAE, ASHRAE Standard 55: Thermal Environmental Conditions for
Human Occupancy, American Society of Heating, Refrigerating and Airconditioning Engineers, 2004.
[17] M. Trejo Perea, J.G. Ros Moreno, E. Rivas, V. Rauch, Savings and
analysis of the consumption and quality of the energy, 1er International
Congress of Engineering, Universidad Autnoma de Quertaro, Mxico,
2005, pp. 253261.

722

G.J. Ros-Moreno et al. / Automation in Construction 16 (2007) 713722

[18] J. Neter, M.H.C.J. Nachtsheim, W. Wasserman, Applied Linear Statistical


Models, 4th ed., Irwin, 1996.
[19] S. Ventura, M. Silvia, D. Prez-Bendito, C. Hervs, Artificial neural
networks for estimation of kinetic analytical parameters, Anal. Chem. 67
(1995) 15211525.

[20] P.K. Kitanidis, R.L. Bras, Real time forecasting with a conceptual hydrological
model. 2. Applications and results, Water Resour. Res. 16 (1980) 10341044.
[21] R. Grio, Neural networks for univariate time series forecasting and their
application to water demand prediction, Neural Netw. World (1992)
437450.

Você também pode gostar