Energy Conversion and Management 47 (2006) 19271941

www.elsevier.com/locate/enconman

Parameter estimation of internal thermal mass of building

dynamic models using genetic algorithm
Shengwei Wang *, Xinhua Xu
Department of Building Services Engineering, The Hong Kong Polytechnic University Kowloon, Hong Kong, PR China
Received 8 March 2005; accepted 11 September 2005
Available online 27 October 2005

Abstract
Building thermal transfer models are essential to predict transient cooling or heating requirements for performance
monitoring, diagnosis and control strategy analysis. Detailed physical models are time consuming and often not cost eective. Black box models require a signicant amount of training data and may not always reect the physical behaviors. In
this study, a building is described using a simplied thermal network model. For the building envelope, the model parameters can be determined using easily available physical details. For building internal mass having thermal capacitance,
including components such as furniture, partitions etc., it is very dicult to obtain detailed physical properties. To overcome this problem, this paper proposes to present the building internal mass with a thermal network structure of lumped
thermal mass and estimate the lumped parameters using operation data. A genetic algorithm estimator is developed to estimate the lumped internal thermal parameters of the building thermal network model using the operation data collected
from site monitoring. The simplied dynamic model of building internal mass is validated in dierent weather conditions.
2005 Elsevier Ltd. All rights reserved.
Keywords: Lumped thermal parameter; Building internal mass; Thermal network model; Simplied model; Genetic algorithm; Dynamic
thermal performance

1. Introduction
For diagnosis purposes of a whole building [6,16], or for thermal mass control strategies [1,12,18], or even
for energy saving by system retrotting [4,5,27], a reference model of the building is essential for load prediction or cost saving estimation. At the building level as a whole process, many researchers have developed different reference models that can be categorized into physical models and data driven models.
Physical modeling, also called forward modeling, begins with a description of the building system or component of interest and denes the building being modeled according to its physical description. Most simulation models are based on rst principles, such as EnergyPlus [7], DOE-2 [20]. However, a large number of
parameters are needed as inputs for the simulation model. The process of collecting a physical description
*

Corresponding author. Tel.: +852 2766 5858; fax: +852 2774 6146.
E-mail address: beswwang@polyu.edu.hk (S. Wang).

0196-8904/$ - see front matter 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.enconman.2005.09.011

1928

S. Wang, X. Xu / Energy Conversion and Management 47 (2006) 19271941

Nomenclature
A
C
df
f
J
Q
R
T
t

area (m2)
thermal capacitance (J/(m2K))
dierence between two maximum tness values
tness function
objective function
cooling/heating load or heat (kW)
thermal resistance (m2K/W)
air temperature (C or K)
time (s)

Greek symbol
ef
threshold value
Subscripts
act
actual cooling load
conv
convective heat
ei
associated with external wall at ith orientation
est
estimated
fr
fresh air
im
associated with building internal mass
in
inside, indoor air
la
latent heat
r
associated with radiation
rf
associated with roof
rtn
return air
sol
associated with solar air temperature

is time consuming and often is not cost eective, or may even be impossible for some cases (the thermal properties of furniture, partitions etc.). In the modeling process, such indenite properties are usually assumed.
Unsuitable assumptions can make the model deviate from reality, thus decreasing the condence in models.
Data driven models, also called inverse models, include steady [26,28] and dynamic ones [10,17,24,30],
which are capable of capturing dynamics such as mass dynamics to some extents and are better suited to handle inter-correlated forcing functions or independent parameters. Applying regression techniques to the above
can lead to models. It is generally necessary to acquire data over a long period of time with widely varying
conditions in order to train the models for accurate predictions under all conditions. Furthermore, it is not
known how well the models would perform in predicting building energy use if there were a major change
in the control strategies employed, such as would occur when going from a night set up control to a precooling control strategy, because those parameters do not respect the proper physics or the parameters cannot
represent the physical properties [2].
Therefore, a kind of simplied models, which can represent the physical properties of the building system
are preferred for diagnosis, optimal control etc. Braun and Chaturvedi [2] developed an inverse gray box thermal network model for transient building load prediction. In the approach, a second order transfer function
was established from an assumed 3R2C (three resistances, two capacitances) thermal network model to predict
building load. All the parameters of the 3R2C models for the external walls, roof, internal walls etc., whose
values are assumed in certain ranges, need to be identied by a non-linear regression algorithm to minimize
errors between the predictions of the transfer functions and the measured operation data. Liao and Dexter [22]
developed a method to establish a simplied second order physical model to simulate the dynamic behavior of
existing multi-zone heating systems of a residential building. In their method, the total resistance and capac-

S. Wang, X. Xu / Energy Conversion and Management 47 (2006) 19271941

1929

itance of the envelopes and internal mass contribute to the parameters of a simplied second order model, and
the parameters are commissioned with the monitoring operation data. With the simplied model developed, a
soft sensor was built to estimate the average room air temperature for an inferential control scheme for optimizing the boiler controls in multi-zone heating systems [21].
However, some information of building systems, such as building envelopes, is easily available. We can
establish detailed physical models or identify the parameters of them using available property data. Other
information of building systems, such as internal structures, partitions, carpet, furniture etc., is dicult to
obtain. Nevertheless, we can view all the heat storage materials inside the building as a lumped thermal mass,
namely building internal mass, which are physically meaningful. Consequently, half of the building model is
represented by detailed physical parameters and another half is described by identied parameters using operation data. 3R2C models are usually utilized to simulate building envelopes. The nodal placement of 3R2C
models can be determined with simple congurations [29] or an optimized conguration [2] with the detailed
physical description. Building internal mass is represented by a 2R2C (two resistances, two capacitances)
model whose resistances and capacitances are assumed to be constant. The node placement of the building
internal mass is identied with the 3R2C models of the building envelope and the operation data. Searching
for the best values of the model parameters is a non-linear optimization process. A genetic algorithm [23] can
quickly nd a suciently good solution (i.e. near optimal solution) and can also be utilized to search for optimal parameters of the internal mass model to minimize the errors between measured values and prediction.
The parameters of the building internal mass model (2R2C) are identied for a real commercial oce building
with the operation data in a short period time. The model was also validated with dierent operation conditions. The results show that the simplied model with lumped building internal mass has good stability to predict the thermal performance for dierent conditions because it not only describes the behavior or
performance of the building internal mass but also represents the building internal mass with physically meaningful lumped parameters.
2. Thermal network model of buildings
For retrotting analysis, performance monitoring and diagnosis, control strategy development and control
applications of HVAC systems at the building level, a virtual concept is proposed to describe the building system as a whole. A schematic virtual system for a typical building is shown in Fig. 1. The major components of
the system are the building envelopes, internal mass, cooling/heating sources and a virtual air handling unit
(AHU). The virtual AHU can accomplish the functions of maintaining air temperature and humidity within
the building and satisfying the indoor air quality with optimized energy consumption during working hours,

Infiltration
Exhaust
Air Damper

Flow
Station

Temp. & Humidity Sensor

Fresh Air
Fan

Cooling Heating Supply

Fan
Coil
Coil
Mix Air
Plenum

C
C

H
From Boiler

From Chiller

Filter

Exfiltration

Fig. 1. Schematics of a building system.

Internal mass

Return Fan

Re-circulation
Air Damper

Outdoor Air
Damper

Building envelopes

1930

S. Wang, X. Xu / Energy Conversion and Management 47 (2006) 19271941

representing all the installed AHUs in the building. In order to predict overall energy requirements conveniently, a simplied physical model is needed to represent the whole building, including occupant activities,
lighting etc., relative to cooling/heating loads. To arrive at a clear, physically interpretable model, the fundamental assumption is necessary that the whole indoor circulated air volume is at a uniform temperature. This
assumption is fairly common, being employed by most simulation models. It also corresponds with the usual
air temperature measurement in building zones where it is customary to place a few temperature sensors at
representative positions in rooms and determine the average value.
The cooling requirement for a building can be separated into latent and sensible contributions. Latent heat
gains, which become cooling load directly, are associated with the addition of water vapor to the air, which
must ultimately be removed by the air handling system. In a building, latent heat gains primarily consist of
occupant respiration, inltration (exltration) of moisture, induced fresh air and aesthetic sources such as
plants and fountains. Among these, the latent gains from occupants and induced fresh air are easy to estimate
and measure.
The sensible heat gain of buildings is due to the heat transfer from warm surfaces within the buildings.
The sources of sensible heat gains can be categorized as internal and external to the building. Typical internal
sources are occupants, lights, equipments (computers, copy machines etc.), internal structures and furniture
etc., which are warned by all sorts of radiation. External sources of heat gains consist of solar radiation, which
is transmitted through windows and absorbed on external walls and energy conduction through external walls
and windows due to the dierence between the ambient and indoor space temperatures.
The simplied second order physical models of the building system are the compounding of both physical
modeling and data driven modeling. They begin with physical descriptions of certain components if their
detailed physical characteristics are available and assumptions of the physical structures of other components
if their detailed physical characteristics cannot be available. The parameters of the assumed physical models
are identied with known physical descriptions and monitored operation data to provide the most accurate
representation of the building thermal dynamic characteristic on the basis of the assumed models and the
available data. The models not only describe the behavior or performance of the system but also explain
the system physically. Therefore, the models can predict reliably the performance of the global building
system.

Tsol, e ( t)
i:

the number of walls at

different orientations
rf: roof
t: time
e: external wall
in: indoor air
im: internal mass
r: radiation
conv: convective
fr: fresh air
la: latent heat gain
act: actual cooling load

Rrf ,1

Crf , 2
Trf , 2 (t)

2R2C model
of internal mass

Rrf , 3

Qla

C rf , 4

Q fr
Qconv

Trf , 4 (t)

Tsol , ei ( t )

Rei ,1

Rrf ,5

Qr ,1

Qr , 2

Tei , 2 (t ) Rei , 3 Tei , 4 (t) Rei ,5

Tim, 2 (t )
Tim,1 (t)

C ei , 2

Tin (t ) Rim ,2

C ei , 4

Qact

Cim, 2

Rim,1

Cin
Fig. 2. Schematics of the thermal network model of building.

Cim ,1

S. Wang, X. Xu / Energy Conversion and Management 47 (2006) 19271941

1931

Fig. 2 depicts a schematic representation of an electrical analog for the thermal network building model.
Three main categories are considered: (1) external walls, (2) roof(s) and (3) building internal mass (ground
oor is incorporated into building internal mass). External walls should be considered, respectively, according
to their orientations because the dynamic models of the external walls at dierent orientations have dierent
forcing functions due to the changing position of the sun. External walls and the ceiling/roof are considered in
the 3R2C models. The building internal mass includes oors, partitions, crawl space in ceiling, internal walls,
furniture etc. It absorbs radiant heat through the windows and that from occupants, lighting, machines etc.
and then releases the heat gradually to the air space. Building internal mass is viewed as a 2R2C model, which
consists of two resistances and two capacitances, as shown in Fig. 2. All resistances and capacitances are
assumed to be time invariant. The windows have negligible energy storage and are represented as pure resistances. The eect of varying wind velocity on external wall convection coecients is not considered.
The whole building energy balance can be represented with the following dierential equations:
dT rf;2 t T sol;rf t  T rf;2 t T rf;2 t  T rf;4 t


dt
Rrf;1
Rrf;3
dT rf;4 t T rf;2 t  T rf;4 t T rf;4 t  T in t


C rf;4
dt
Rrf;3
Rrf;5
dT ei;2 t T sol;ei t  T ei;2 t T ei;2 t  T ei;4 t


C ei;2
dt
Rei;1
Rei ; 3
dT ei;4 t T ei;2 t  T ei;4 t T ei;4 t  T in t

C ei ; 4

dt
Rei;3
Rei;5
dT im;1 t
T im;1 t  T im;2 t
Qr;1 
C im;1
dt
Rim;1
dT im;2 t
T im;1 t  T im;2 t T im;2 t  T in t
Qr;2

C im;2
dt
Rim;1
Rim;2
n
X
T ei;4 t  T in t
T rf;4 t  T in t T im;2 t  T in t
Qest

R
Rrf;5
Rim;2
ei;5
i1

C rf;2

 C in

dT in t
Qconv Qfr Qla
dt

Qest Qact

1
2
3
4
5
6

7
8

where C and R are resistance and capacitance, T is temperature, subscripts rf, im, ei and in, indicate roof,
internal mass, the ith external wall and inside, respectively. Qr,1 and Qr,2 absorbed by nodes Cim,1 and
Cim,2, respectively, are the radiation, which includes the radiation from solar radiation through the windows
and from occupants, lights etc. Qconv is the convective heat from windows surfaces, occupants, lights etc. Qfr is
the heat transfer because of fresh air induction, inltration (exltration). Qla is the latent heat gain from occupants etc. Qest andQact are the estimated cooling load with the model and the actual cooling load, which can be
measured from the central plant.
The model parameters of the building envelope (i.e. external walls and roofs) can be determined with simple
congurations or optimized congurations using the property data of the envelope. The model parameters
Cim,1, Rim,1, Cim,2 and Rim,2 of the building internal mass are identied using the operation data, while the
model parameters of the building envelope are obtained in advance. A genetic algorithm (GA) based method
is used for parameter identication as illustrated in the next section.
3. Parameter identication of building internal mass with GA
Finding the best values of the parameters of the 2R2C model of the building internal thermal mass is a nonlinear optimization process. House and Smith [15] employed sequential quadratic programming to compute
the optimal values. Nizet et al. [25] used a conjugate gradient method to develop an optimal control method.
Both optimization methods, as well as other conventional optimization methods have to start from initial

1932

S. Wang, X. Xu / Energy Conversion and Management 47 (2006) 19271941

guesses of the optimal variables and their convergence speed is aected by their initial guesses in most cases. A
genetic algorithm (GA) is a better optimization method, especially when an optimal problem is not perfectly
smooth and uni-modal [23]. It can quickly nd a suciently good solution (i.e. near optimal solution) and can
be applied when a task does not require an absolute optimum. The algorithm has been used to search for
global optimal solutions in air conditioning elds [31,32]. In this study, a GA is utilized to search for optimal
parameters of the 2R2C model of the building internal mass to minimize the errors between measured values
and prediction of the building model. The GA based method is described as follows.
3.1. Objective function of optimization
The simulated prediction of the building model with the dierential equations (1)(7) is used to compare
with the measured cooling load. The optimized parameters are the resistances and capacitances of the
2R2C model of the building internal mass that give the best t with the operation data. The objective function
(J) of such optimization employs the integrated root mean square error dened in Eq. (9).
v
uN
uP
u Qact;k  Qest;k 2
t
J C im;1 ; Rim;1 ; C im;2 ; Rim;2 k1
9
N 1
where Qact is the actual measured cooling/heating load, Qest is the model predicted cooling/heating load and
Cim,1, Rim,1, Cim,2, Rim,2 are the parameters of the 2R2C model. This is a typical non-linear optimization problem. A GA (genetic algorithm) is employed to search for the optimal values as illustrated in Section 3.2.
The measured cooling/heating load is calculated using the return and supply water temperature dierence and
the water ow rate retrieved from the BMS (building management system). To predict the building cooling/
heating load using the building thermal network model, the indoor air temperature and humidity, outdoor
air temperature and humidity, fresh air ow rate, solar radiation, occupancy and internal gains are needed.
The means to collect these data for parameter identication of the internal mass are described in detail in Section 4.
3.2. GA estimator
A GA is an advanced search and optimization technique. It has been developed to imitate the evolutionary
principle of natural genetics. The GA was invented by Holland [14] and further developed in the 1960s and
1970s. Goldberg [11], Davis [8] and Mitchell [23] provided comprehensive overviews and introductions to
GAs. Deb [9] compared the GA search method with traditional methods (the direct exhaustive search method
and the gradient directed search method) for function optimization. One of the main advantages of a GA is
that it is generally robust in nding global optimal solutions, particularly in multi-modal and multi-objective
optimization problems [9]. Extensive research on the theoretical fundamentals and applications of GAs is still
going on, aimed at better computation eciency, improved robustness and so on.
Fig. 3 shows schematically the ow chart of the GA estimator developed for parameter identication of the
2R2C model of building internal mass. It starts with the initial estimates of the individual capacitances and
resistances within assumed ranges. The component with grey background represents the procedures of a
GA run. Multiple runs are allowed. Eq. (10) represents the tness function (f), which is the reciprocal of
the objective function (Eq. (9)).
f f C im;1 ; Rim;1 ; C im;2 ; Rim;2

1
JC im;1 ; Rim;1 ; C im;2 ; Rim;2

10

In the genetic algorithm, the four parameters (Cim,1, Rim,1, Cim,2, Rim,2) constitute the chromosome of an individual, the assumed ranges of these parameters are the search space for these parameters. Initializing the four
parameters produces the initial population to start a GA run.
Termination of a GA run is decided if the number of the current generation is equal to a predened maximum number. At least two runs of the GA process are necessary when running the GA estimator. The

S. Wang, X. Xu / Energy Conversion and Management 47 (2006) 19271941

1933

Fig. 3. Flow chart of the GA estimator for parameter estimation of building internal mass model (2R2C).

criterion to stop the GA estimator is based on the comparison of the best tness values of two consecutive
runs. If the relative dierence between the two maximum tness values (df) is less than a threshold value
(ef, e.g. equals 0.0001), the GA estimator is stopped. A GA driver developed by Carroll [3] is revised and used
in this study.
4. Building system description and data collection
The simplied building dynamic model and the identication method of the lumped thermal parameters of
the building internal mass are validated in a real building. The building system description and simplication
measures, as well as data collection, are briefed below.
4.1. Building system description and simplication
The building in consideration, China Resources Buildings, located at 26th Harbor Road, Hong Kong, was
completed in 1983. The buildings consist of a main building of 50 oors with 180 m height, an attached
building of seven oors with about 28 m height and a basement of three oors. For the main building, the

1934

S. Wang, X. Xu / Energy Conversion and Management 47 (2006) 19271941

rst and second oors serve as a shopping center with open corridors at the perimeter zone connecting the
dierent buildings. The third, fourth and fth oors are restaurants. The sixth oor is used for the chiller
plant. The commercial oces are located from the 7th to the 49th oors with 2262 m2 (58 by 39 m) per oor
except that the 15th, 31st and 48th oors are for refuge use. The 50th oor is used for a banquet hall with 6 m
height. For the attached building, the rst and second oors also serve as a shopping center with open corridors at the perimeter zone connecting dierent buildings. The third and fourth oors are used as an exhibition
hall with atrium in the center. The 5th, 6th and 7th oors are used for oces with 1738 m2 (22 by 79 m) per
oor, and the roof is covered with a swimming pool. The basements are used as a garage.
All the buildings are constructed primarily of heavy weight steel concrete with transportation systems and
AHU plants in the cores. The external walls above the 2nd oor are a multi-layer construction consisting of
ve layers of homogeneous materials including 300 mm high density concrete between 13 mm face brick and
about 13 mm plaster with outside and inside air lms. The oors are 150 mm high density concrete with about
a 50 mm refurbishment layer. For the oce oors, the ratio of window to wall is about 25%. For restaurant
oors and exhibition hall, there are almost no external windows. The exterior of the 1st and 2nd oors of the
main building and the attached building are glass curtains and mostly shaded outside.
All the buildings are air conditioned with the chiller water plant on the 6th oor except that the banquet
hall on the 50th oor is supplied with a separate air cooled package unit, and the basements are almost not air
conditioned. The condensing heat is taken away by a sea water cooling system. The plant is equipped with
three centrifugal chillers with each having 3866 kW cooling capacity, two chillers with each having
1055 kW cooling capacity and one chiller with 3164 kW cooling capacity. Therefore, the cooling capacity
is approximately 137 W/m2 (gross area 123,000 m2 not including the 50th oor and basements). The water
system conguration is constant primary and variable secondary water. Most of the air conditioning
terminals are AHUs, which are located in the core areas. Basically, in the oce oors, two CAVs (constant
air volume) service the perimeter zone and two or three VAVs (variable air volume) service the inner
zone.
For convenient modeling, the following simplication measures are assumed: (1). The refuge oors are
open air with window orices. The heat transfer from the refuge oors to the adjacent oors are calculated
through the structure of 150 mm high density concrete with a simplied 3R2 C model; (2). The ceiling of
the 49th oor is considered as adiabatic because the 50th oor is air conditioned with a separate unit; (3).
The ceiling of the 7th oor of the attached building is also considered as adiabatic because the roof is covered
with a swimming pool; (4). The ground oor is merged into the building internal mass without special
considerations.
4.2. Data collection and processing
A site survey was conducted and original design information was collected to build the complete proles of
occupancy and the use of lighting and equipment. The occupancy load and the internal load from lighting and
equipment were estimated according to the rules established in a previous research on Hong Kong buildings
[13] as briefed in the following. The normal occupancy period of the oces, shopping center and restaurant is
from 8:00 am to 6:00 pm, from 10:30 am to 10:00 pm and from 6:30 am to 10:00 pm or later, respectively. The
design densities of occupancy for the three places are 9, 4.5 and 2 m2 per person, respectively. The design
equipment powers for the three places are 25, 30 and 55 W/m2, respectively. The design lighting powers for
the three places are 25, 70 and 35 W/m2, respectively. The normal patterns of occupancy, equipment power
and lighting power are shown in Fig. 4 (occupancy load, light power and equipment power pattern are in fractions of their respective peak values). The internal gain from occupancy, lighting and equipments can be split
into convective and radiative components (occupancy heat gains: latent heat 40%, convective 20% and radiative 40%; lighting heat gains: convective 50% and radiative 50% (mostly uorescent lights); equipment heat
gains: convective 8020% and radiative 2080%).
The oces, shopping centers and restaurants are supplied with xed amounts of fresh air with the ventilation rates of 10, 7 and 7 L/s per person in the occupancy periods. Although the building is tight, mostly with
xed windows, the inltration rate is considered as 0.1 ach (air changes per hour) in the occupied periods and
0.5 ach in the unoccupied periods.

Daily pattern of occupant load (-)

S. Wang, X. Xu / Energy Conversion and Management 47 (2006) 19271941

1.0
Office
Shopping center
Restaurant

0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0

12

16

20

24

16

20

24

16

20

24

Time (h)

(a)

Daily pattern of eqipment load (-)

1935

1.0
Office
Shopping center
Restaurant

0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0

12

Time (h)

(b)

Daily pattern of Lighting load (-)

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0

(c)

Office perimeter zone

Office interior zone
Shopping center
Restaurant

12

Time (h)

Fig. 4. Normal patterns of occupancy load, equipment power load and lighting power load.

The return air temperature of the AHUs and the actual cooling load of the chiller plant were measured and
recorded by the BMS. The outdoor air temperature and humidity were measured by the BMS as well. An average return air temperature was taken as the uniform indoor air temperature, as shown in Eq. (11).
T in

N
X
Ai
T rtn i
A
0
i1

11

1936

S. Wang, X. Xu / Energy Conversion and Management 47 (2006) 19271941

where Tin and Trtn is the uniform indoor air temperature and the return air temperature of the ith AHU. Ai
and A0 are the served area of the ith AHU and the total served area.
Solar radiation cannot be measured directly at the site. Instead, the horizontal global solar radiation was
obtained from the Hong Kong Royal Observatory. It was decomposed into direct normal solar radiation and
diuse solar radiation with the relationship established by Lam and Li [19] or by Yik and Chung et al. [33] for
Hong Kong. The direct normal solar radiation and diuse solar radiation were used to calculate the solar air
temperature on the dierent external wall surfaces and the solar radiation transmitted through the windows.
5. Validation results and analysis
The parameters of the 2R2C model of the building internal mass were identied using the monitored data
of fourteen consecutive days (two weeks) in the summer season. The indoor air temperature can also be estimated with the identied parameters and the recorded cooling load. The operation data of the building were
also collected to assess the accuracy and stability of the building model with the identied parameters of the
building internal mass model to estimate the cooling load and the indoor air temperature in other operation
conditions.
5.1. Identied parameters of 2R2C building internal mass model

Outdoor air temperature (oC)

35.0

4.00
Outdoor air temperature
Horizontal global solar radiation

30.0

3.50
3.00
2.50

25.0

2.00
1.50

20.0

1.00
0.50

15.0

0.00
24 48 72 96 120 144 168 192 216 240 264 288 312 336

Horizontal global solar radiation (MJ/h)

For the concerned building, the basic structure of the oor layer is 150 mm high density concrete. Because
the internal mass includes the basic structure of the oor layer, the partitions, the internal walls, furniture etc.,
the searching scopes of the parameters of the building internal mass with the GA estimator should be much
larger than the capacitance and resistance of the basic structure of the oor layer. The searching scope of the
two resistances of the building internal mass was assumed as the resistance of the indoor air lm plus three
times the resistance of the oor basic structure. The searching scope of the two capacitances of the building
internal mass was assumed as three times the capacitance of the oor basic structure.
The operation data of the building for fourteen consecutive days (two weeks) in the typical summer season
were used to identify the parameters of the 2R2C building internal mass model. The outdoor air temperature
and the horizontal global solar radiation for the parameter identication case are shown in Fig. 5. Most of the
days were sunny and cloudy. Some days were in sunny periods with a few showers.
The air conditioning systems in the oces were shut down except from 8:00 am to 18:00 pm of oce hours.
The air conditioning systems in the restaurants and shopping centers were shut down a little later. Almost each
oce oor has fan coil units to supply the cooling load of communication or computation rooms for nonoce hours, such as night time and holidays. The cooling load in these hours was only a very small part
of the total building cooling load in normal oce hours. The temperatures measured in non-oce hours cannot represent the indoor air temperature because the sensors installed in the air chamber (in the core of the
building) and the air in the oces is stagnant. Therefore, only the operation data during oce hours were used

Time (h)

Fig. 5. Outdoor air temperature and horizontal global solar radiation (parameter identication case).

S. Wang, X. Xu / Energy Conversion and Management 47 (2006) 19271941

1937

to identify the parameters of the building internal mass model. It is worth noting that the operation data in
non-oce hours are preferably used in the regression ttings for parameter identication when reliable measurements in that period are available.
The identied parameters using the GA estimator are: Cim,1 = 648,729 J/(m2K), Cim,2 = 73,793 J/(m2K),
Rim,1 = 0.299 m2K/W, Rim,2 = 0.0282 m2K/W.
Fig. 6 presents a comparison between the model predicted cooling load with the identied parameters of the
building internal mass model and the actual measured cooling load. It shows that the model predicted cooling
load followed the dynamic prole of the actual measured cooling load well. The average error (average of
absolute deviation) was 7.8% for the data points during oce hours from 8:00 am to 18:00 pm. If all the data
points, including those in oce hours and non-oce hours, were used for model validation, the error was
slightly higher, 18.5%, because the model tended to over predict the cooling load for non-oce hours. It
can be explained that the measured uniform indoor air temperature, as Eq. (11), is higher than the actual
temperature in non-oce hours, which is described as follows.
Fig. 7 presents the model predicted uniform indoor air temperature with the identied parameters of the
building internal mass model and the measured cooling load in the period used for the parameter identication. It shows that the identied model predicted well the trends of the indoor air temperature. At the beginning of oce hours, the indoor air was cooled by the air handling systems. In contrast, when the air handling
systems were shut down in non-oce hours, the indoor air was warmed by heat transfer through the building
envelopes and released heat from the building internal mass, which absorbed amounts of radiant heat from
solar radiation, occupancy etc. The comparison shows that the model predicted indoor air temperature in
the oce hours agreed well with the measured uniform indoor air temperature, which was measured
using the measured return air temperatures as shown in Eq. (11). In non-oce hours, the model predicted

15000

Cooling Load (Kw)

Actual measured cooling load

Model predicted cooling load

12000

9000

6000

3000

0
0

48

96

144

192

240

288

336

Time (h)

Fig. 6. Actual measured cooling load vs model predicted cooling load (parameter identication case).

30.0

Indoor air temperature (oC)

"Measured" uniform indoor air temperature

Model predicted indoor air temperature

27.0

24.0

21.0

18.0

15.0

48

96

144

192

240

288

336

Time (h)

Fig. 7. Measured uniform indoor air temperature vs model predicted indoor air temperature (parameter identication case).

1938

S. Wang, X. Xu / Energy Conversion and Management 47 (2006) 19271941

indoor air temperature was higher than the measured uniform indoor air temperature. It is because, in oce
hours, the air was circulating, and the measured return air temperature in the return air chamber could well
represent the indoor air temperature. However, in non-oce hours, the measured return air temperature was
obviously lower than the real indoor air temperatures in the summer season because the air was stagnant and
the measured return air temperature in the return air chamber (in core area) was not aected (i.e. heated) signicantly by the outside weather condition. Therefore, the model predicted indoor air temperature could represent the uniform indoor air temperature more accurately compared with the measured uniform indoor air
temperature in non-oce hours.
5.2. Validation of building model
To validate the building model and the parameter identication method, the building model was used to
predict the cooling load with the measured uniform indoor air temperature in two other operation periods.
One was also in the summer season, lasting for two weeks, and the other was in the winter season, lasting for
one week. The same building model was also used to predict the uniform indoor air temperature with the given
cooling load. The internal heat gains were kept unchanged in the two cases.
In the summer case, the weather condition was similar to that used for the training process as shown in
Fig. 5. Fig. 8 presents the model predicted cooling load prole using the building model and the actual measured cooling load prole. It shows that the predicted cooling load could follow the trends of actual energy
consumption well. The average error between the predicted and measured cooling loads was 9.7% when the
data points in non-oce hours were excluded. The average error was 22.0% if all the data points (both in
non-oce hours and oce hours) were included. It shows that the building model can predict the cooling load
with acceptable accuracy for practical application.
The predicted uniform indoor air temperatures in the oce hours were close to the measured indoor air
temperatures based on the measured return air temperatures as shown in Fig. 9. The model predicted temperature well matched the actual indoor air temperature and its dynamic trend. Fig. 9 also displayed the measured indoor air temperatures in the non-oce hours (at night and weekend days), which were much lower
than the predicted uniform indoor air temperatures. The situation was similar to that stated in Section 5.1.
The building model was also validated with a week long operation data in a winter case in Hong Kong.
Fig. 10 shows the outdoor air temperature prole and horizontal global solar radiation prole. The outdoor
air temperature was much low than that in the summer season. Most of the days were sunny, and the intensities of the solar radiation are a little lower than that used for the training process. Fig. 11 presents the actual
measured cooling load and model predicted cooling load. The model predicted cooling load agreed with the
actual measured cooling load in oce hours with a corresponding average error of 12.0%. The total error,
including all the data points, was higher, at 20.1%, because there was signicant deviation in the 7th day
(Sunday).

15000

Cooling Load (Kw)

Actual measured cooling load

Model predicted cooling load

12000

9000

6000

3000

48

96

144

192

240

288

336

Time (h)

Fig. 8. Actual measured cooling load vs model predicted cooling load (validationsummer case).

S. Wang, X. Xu / Energy Conversion and Management 47 (2006) 19271941

1939

Indoor air temperature (oC)

30.0
"Measured" uniform indoor air temperature
Model predicted indoor air temperature

27.0

24.0

21.0

18.0

15.0
0

48

96

144

192

240

288

336

Time (h)

Outdoor air temperature (oC)

25.0

3.00
Outdoor air temperature
Horizontal global solar radiation

2.50
20.0

2.00
1.50

15.0

1.00
0.50

10.0

24

48

72

96

120

144

0.00
168

Horizontal global solar radiation (MJ/h)

Fig. 9. Measured uniform indoor air temperature vs model predicted indoor air temperature (validationsummer case).

Time (h)

Fig. 10. Outdoor air temperature and horizontal global solar radiation (validationwinter case).

10000

Cooling Load (Kw)

Actual measured cooling load

Model predicted cooling load

7500

5000

2500

24

48

72

96

120

144

168

Time (h)

Fig. 11. Actual measured cooling load vs model predicted cooling load (validationwinter case).

6. Conclusion
Simplied models, which can represent the physical properties of a building system are desired for diagnosis, control strategy analysis etc. The simplied thermal network model describes all the heat storage materials
inside the building, including the oors, internal walls etc., as a simplied lumped thermal network model with
2R2C structure. The procedure to estimate the lumped thermal parameters of the building internal mass
model (2R2C) with the genetic algorithm presented in the study can be eectively used to identify the parameters using short term operation data.

1940

S. Wang, X. Xu / Energy Conversion and Management 47 (2006) 19271941

The parameters of the internal thermal network model of a commercial oce building were identied using
operation data of two weeks in the summer season. The results show that the trained model can predict the
cooling load and average indoor air temperature in oce hours with good accuracy. The entire building thermal network model, consisting of building envelope models and internal thermal model, can also be extrapolated well to other operating conditions. The predicted cooling load in oce hours in other operating
conditions using the building thermal network model shows the average error was about 10% compared to
the actual measured cooling load in the other two test conditions. The model can also predict average indoor
air temperature in oce hours with good accuracy.
In summary, the model has good robustness to predict the thermal performance under dierent operation
conditions by capturing the dynamic characteristics of the building system correctly. The good robustness of
the model is due to the parameters of the building internal mass model (2R2C) and the model structure, which
not only describes the behavior or performance of the building internal mass but also represents the building
internal mass physically. The model can be used to predict building thermal performance for practical applications with acceptable accuracy and good reliability.
Acknowledgements
The research presented in this paper was nancially supported by a research grant of the Research Grants
Council (RGC) of the Hong Kong SAR and a research grant of The Hong Kong Polytechnic University.
References
[1] Ben-Nakhi A, Mahmoud M. Cooling load prediction for buildings using general regression neural networks. Energy Conversion
Manage 2004;45(1314):212741.
[2] Braun J, Chaturvedi N. An inverse gray-box model for transient building load prediction. HVAC & R Res 2002;8(1):7399.
[3] Carroll DL. FORTRAN genetic algorithm (GA) driver; 2001. Version 1.7a. Available from: http://cuaerospace.com/carroll/ga.html,
http://cuaerospace.com/carroll/gatips.html.
[4] Chen Z, Nelson R, et al. Comparison of methods for predicting monthly post-retrot energy use in buildings. ASHRAE Trans
2003;109(2):44959.
[5] Chirarattananon S, Taweekun J. A technical review of energy conservation programs for commercial and government buildings in
Thailand. Energy Conversion Manage 2003;44(5):74564.
[6] Claridge D, Liu M, et al. Whole building diagnostics, 1999. Available from: <http://poet.lbl.gov/diagworkshop/proceedings>.
[7] Crawley, Drury B. Energyplus: energy simulation program. ASHRAE J 2000;42(4):4956.
[8] Davis L. Handbook of genetic algorithms. New York: Van Nostrand Reinhold; 1991.
[9] Deb K. Genetic algorithms for function optimisation. In: Ferrera F, Verdegay JL, editors. Genetic algorithms and soft
computing. Heidelberg: Physica-Verlag; 1996. p. 329.
[10] Dhar A, Reddy T, et al. A Fourier series model to predict hourly heating and cooling energy use in commercial buildings with
outdoor temperature as the only weather variable. J Solar Energy Eng 1999;121:4753.
[11] Goldberg D. Genetic algorithms in search, optimisation, and machine learning. New York: Addison-Wesley; 1989.
[12] Hasnain S. Review on sustainable thermal energy storage technologies, Part II: cool thermal storage. Energy Conversion Manage
1998;39(11):113953.
[13] HK-BEAM Society. An environmental assessment for new building developments, HK-BEAM Version 2003. HK-BEAM Society,
Kowloon, Hong Kong; 2003.
[14] Holland J. Adaptation in natural and articial systems. Cambridge, MA: MIT Press; 1992.
[15] House J, Smith T. Optimal control of a thermal system. ASHRAE Trans 1991;97(2):9911001.
[16] House J, Kelly G. (NIST). An overview of building diagnostic; 1999. Available from: <http://poet.lbl.gov/diag-workshop/
proceedings>.
[17] Kalogirou S, Neocleous C, et al. Heating load estimation using articial neural networks. In: Proc. CLIMA 2000 Conf., Brussels
(Belgium); 1997.
[18] Keeney K, Braun J. Application of building precooling to reduce peak cooling requirements. ASHRAE Trans 1997;103(1):8306.
[19] Lam C, Li HW. Correlation between global solar radiation and its direct and diuse components. Building Env 1996;31(6):52735.
[20] Lawrence Berkeley Laboratory. DOE-2 engineering manual version 2.1C. Berkeley, CA, Lawrence Berkeley Laboratory; 1982.
[21] Liao Z, Dexter A. An inferential control scheme for optimizing the operation of boilers in multi-zone heating systems. Building
Services Eng Res Technol 2003;24(4):24556.
[22] Liao Z, Dexter A. A simplied physical model for estimating the average air temperature in multi-zone heating systems. Building Env
2004;39:101322.
[23] Mitchell M. An introduction to genetic algorithm. The MIT Press; 1997.

S. Wang, X. Xu / Energy Conversion and Management 47 (2006) 19271941

1941

[24] Minoru K, Charles D, et al. Hourly thermal load prediction for the next 24 h by ARIMA, EWMA, LR and an articial neural
network. ASHRAE Trans 1995;101(1):186200.
[25] Nizet J, Lecomte L, Litt F. Optimal control applied to air conditioning in building. ASHRAE Trans 1984;90(1B):587600.
[26] Reddy T, Claridge D. Using synthetic data to evaluate multiple regression and principal component analyses for statistical modeling
of daily building energy consumption. Energy Buildings 1994;21:3544.
[27] Reddy T, Kissock J, et al. Uncertainty in baseline regression modeling and in determination of retrot savings. J Solar Energy Eng
1998;120:18592.
[28] Ruch D, Chen L, et al. A change-point principal component analysis (CP/PCA) method for predicting energy usage in commercial
buildings: the PCA model. J Solar Energy Eng 1993;115:7784.
[29] Seem J, Klein S, et al. Transfer functions for ecient calculation of multi-dimensional transient heat transfer. J Heat Transfer
1989;111:512.
[30] Subbarao K, Burch J, et al. How to accurately measure the load coecient of a residential building. Sol Eng 1990 Twelfth Annu Int
Sol Energ Conf; 1990. p. 41925.
[31] Wang SW, Jin XQ. Model-based optimal control of VAV air-conditioning system using genetic algorithm. Building Env
2000;35(6):47187.
[32] Wang SW, Wang JB. Robust sensor fault diagnosis and validation in HVAC systems. Trans Inst Meas Control 2002;24(3):23162.
[33] Yik WH, Chung TM, Chan KT. A method to estimate diuse radiation in Hong Kong and its accuracy. Trans Hong Kong Inst Eng
1995;2:239.