Energy and Buildings 42 (2010) 412–421

Contents lists available at ScienceDirect

Energy and Buildings

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

Adaptive HVAC zone modeling for sustainable buildings

Glenn Platt a, Jiaming Li b,*, Ronxin Li b, Geoff Poulton b, Geoff James a, Josh Wall a
a
CSIRO Energy Technology, Australia
b
CSIRO ICT Centre, Australia

A R T I C L E I N F O A B S T R A C T

Article history: Control of energy flows within a building is critical to achieving optimal performance of heating,
Received 16 April 2009 ventilation and air-conditioning (HVAC) systems. To design optimal HVAC control strategies, a dynamic
Received in revised form 28 September 2009 model of the HVAC system – particularly the building zones that it services – is essential. As analysis of
Accepted 4 October 2009
building energy consumption is facilitated by the accurate prediction of indoor environmental
conditions, techniques that dynamically model HVAC zones are crucial, and as such, is an active area of
Keywords: research. This paper focuses on real-time HVAC zone model fitting and prediction techniques based on
Genetic algorithms
physical principles, as well as the use of genetic algorithms for optimization. The proposed approach is
HVAC control
validated by comparing real-time HVAC zone model fitting and prediction against the corresponding
HVAC zone modeling
Kalman filtering experimental measurements. In addition, comparison with prediction results using an algorithm based
on feedback-delayed Kalman filters has demonstrated the superiority of the proposed approach in terms
of prediction accuracy.
ß 2009 Elsevier B.V. All rights reserved.

1. Introduction optimum performance. For the energy-efficient design of buildings

and their associated control systems, it is necessary to obtain
Heating, ventilation and air-conditioning (HVAC) systems play approximate dynamic mathematical models for system compo-
a dominant role in regulating the indoor climate, providing people nents. These models are used for a wide range of tasks including
with a comfortable and safe work environment. In most buildings operating strategy planning, cost analysis and comfort evaluation.
the performance of the HVAC system can influence energy The dynamic model can be especially useful for control strategies
consumption as well as indoor air quality. The amount of energy that require knowledge of the dynamic characteristics of HVAC
used by HVAC systems is a significant concern that impacts on systems.
issues from national policy to personal desires of cost and comfort, The system component that we will focus on in this paper is the
and as such, HVAC modeling is a very active research topic [1,2]. zone. A HVAC zone is a group of adjacent offices and/or spaces
A HVAC system comprises a variety of continuous and discrete serviced by a common air-handling unit (AHU) or air-terminal
control components interacting with the building via sensors and device. Since the external and internal environment of the zone
actuators. Control of the energy flows in the building and its always changes, any model of the zone has to emulate the dynamic
environment is the key to achieving optimum performance from thermal processes within the zone as well as the interaction with
an HVAC system. Such control is secured through monitoring and the environment. Modeling such a time-variant system is a
altering the cooling/heating sources to maintain the desired challenging task.
thermal and air quality conditions in a space, while external and Traditionally, zone models are mainly based on static or short-
internal conditions change over time. The building with its term modeling of the HVAC zone. Four approaches have been
technical equipment and its environment consists of several reported in recent literature. The first uses lumped capacitance in
coupled processes which, in general, cannot be influenced an analogue electric circuit to represent thermal elements of a
independently. This complicates the design of high-quality control building [1,3,4]. A second approach is a building thermal model
algorithms and requires model-based design methodologies for based on energy-and-mass balance [5], where every component of
the HVAC system is represented by energy-and-mass balance
equations. The third approach is based on machine learning, in
Abbreviations: AHU, air-handling unit; BMCS, building management & control which artificial neural networks [6] or support vector machines [7]
system; OA, outside air; RA, return air; SA, supply air; CHWV, chilled water valve; model non-linear processes, such as utility loads or an individual
HWV, hot water valve; SP, set-point; OAF, outside air fan; SAF, supply air fan; RAF,
return air fan.
building’s energy consumption.
* Corresponding author. Tel.: +61 2 9372 4707; fax: +61 2 9372 4161. A fourth approach is evident more recently, as researchers have
E-mail address: [email protected] (J. Li). started exploring the response of HVAC systems to dynamic

0378-7788/$ – see front matter ß 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.enbuild.2009.10.009
 G. Platt et al. / Energy and Buildings 42 (2010) 412–421 413

model is fitted in soft real-time (every 5 min) it can easily cope

Nomenclature
with the changes of the zone’s dynamics and give accurate internal
temperature prediction.
Two outlet water temperature (8C) Lastly, we present results using an alternative technique, the
Twi inlet water temperature (8C) Kalman filter, to perform the model fitting. The Kalman filter [14] is
To external (outside) temperature (8C) an efficient recursive filter that predicts the state of a dynamic
Tz zone temperature (8C) system from a series of incomplete and noisy measurements. The
Tsp supply air temperature set-point (8C) Kalman filter is traditionally viewed as a prediction-correction
Fsp supply air fan speed set-point (% of maximum filtering algorithm and is a classic method used for dynamic
speed) systems. Recently, various parameter estimation techniques based
on extended Kalman filters have been applied to thermal modeling
Tsa supply air temperature (8C)
and model refinement (e.g. [15,16]). A series of experimental
Fsa supply air fan speed (% of maximum speed)
results using a modified Kalman filter with delayed feedback is
fa supply air flow rate (m3/s) presented for comparison with our own technique. The results
Ca specific heat of air (kJ/(kg 8C)) show that our model is capable of giving more accurate prediction
ra density of the air (kg/m3) for the indoor temperature of a dynamic zone compared to the
Vz volume of the zone (m3) Kalman filter technique.
Uw thermal conductivity of the wall (W/mK) This paper is organized as follows. Section 2 introduces the
Aw heat transfer area of the wall, i.e. area of each wall HVAC system. Section 3 presents a mathematical zone model
(m2) based on physical principles. Section 4 introduces soft real-time
Gw wall thickness (m) zone model fitting and prediction based on genetic algorithms.
Section 5 gives a series of experimental results to quantify the
Ro resistor of outside wall representing its function of
performance of model fitting and prediction. For comparison
transferring heat between outside and inside
purposes, a feedback-delayed Kalman filtering method is pre-
Io current of outside air representing its heat flow
sented in Section 6. Finally, a conclusion is drawn in Section 7.
Ia current of supply air representing its heat flow
Ra resistor of supply air representing the air convec-
2. HVAC system description
tion
Iz current of zone air representing its heat flow Many HVAC systems use a central plant to provide hot water
Cz overall thermal capacity (kJ/C) of the zone (e.g., with temperature of 32 8C) for heating purposes or cold water
(e.g., with temperature of 6 8C) for cooling purposes. Fig. 1 is a
schematic diagram of a typical HVAC system with multiple zones
for cooling purpose.
environments [8–12]. These models, however, tend to be rather The water path is presented by black arrows, which shows that
cumbersome, and do not adequately adapt to the dynamics of the cold water with temperature Two flows from the central plant along
zone environment in real time. In addition, many existing models the pipework and into the cooling coils. The chilled water valves
do not take into account long-term dynamic properties of the (CHWV) are used to control the volume of water flowing through
building, due to changes accumulated over a long time period. the coils. A cooling coil exchanges heat energy between air from
We investigate a real-time model fitting and prediction algorithm the mixing box and water from the central plant, with the output
using only few parameters, that is well suited to short-term water flowing back to the central plant with a slightly higher
predictions and also adapts to long-term dynamics. In Australia, temperature Twi.
the electricity market cycle is half-hourly [13], and our proposed The air path is represented by grey arrows. The outside air can
algorithm is very suitable for applications on this time scale. be drawn in by an outside air fan (OAF) and outside air damper into
The algorithm is applied to a dynamic zone model of a HVAC the mixing box. A mix of outside air and return air then flows
system, intended to allow the implementation of control strategies through the air filter and passes over the cooling coil. A supply air
to reduce energy consumption and improve the quality of the fan (SAF) then forces the air through an insulated supply duct into
indoor environment. The paper has two major contributions. One the zone area as supply air (SA) for the zone. The supply air gains
is the formation of a mathematical model of the zone based on part of its heat from the zone to replace the heat that is leaking
physical principles and circuit theory—the model we propose is the through the walls, roof, etc. The supply air then passes through the
first one that combines energy/mass balance and basic circuit
theory. The second contribution is the real-time model fitting and
prediction algorithm for a time-variant zone, taking advantage of
genetic algorithms (GA). The model and algorithm are then
validated using measurements of a real HVAC system.
Throughout the paper the HVAC zone is represented as a six-
faced box. The zone model is firstly deduced from an energy-and-
mass balance and then represented using electric circuit theory.
Electric circuit elements are used to represent functions of the
building elements, e.g., a wall is modelled as a resistor to represent
its function of transferring heat between outside and inside, and an
entire zone is modelled as a capacitor to represent its heat storage
function. This kind of zone model is very simple and generic,
applicable to a zone with any internal structure and material.
Following development of the zone model structure, a methodol-
ogy of soft real-time zone model fitting and prediction is given, e.g.,
zone model fitting and prediction every 5 or 30 min. Since the zone Fig. 1. Schematic diagram of a typical HVAC system for cooling purpose.
414 G. Platt et al. / Energy and Buildings 42 (2010) 412–421

room and into the return air inlets. Some return air is exhausted Min. Control Point), with a Tz of 21.7 8C is defined as the
through an exhaust air damper. The remaining return air passes Intermediate Control Point. Lastly, a Tz of 27 8C is defined as the
through the re-circulation air damper into the mixing box and fan speed maximum control point (Fan Speed Max. Control Point).
mixes with the outside air. The outside air dampers and exhaust air In a typical HVAC system, control variables related to temperature
dampers control the amount of air flow amount from the outside can often oscillate around their set-point, with fan speed related
air (OA) and return air (RA) streams respectively. variables usually maintained very close to their set-point. Air flow
rate fa is largely proportional to supply air fan speed Fsa.
2.1. Air-handling unit
3. HVAC zone modeling
The air-handling unit (AHU) is a device used to condition and
circulate air as part of a HVAC system. Usually, the AHU is a large In this section, we focus on an HVAC zone modeling technique
metal box containing fans, heating and/or cooling coils, filter racks based on physical principles. The assumption for the model is that
or chambers, sound attenuators, and dampers. Building indoor the supply air flowing into the zone is evenly distributed
temperature is maintained at set-point (SP) values by the AHU, throughout the whole zone; adjacent zones have the same
which usually connects to ductwork that distributes conditioned temperature; and we ignore the thermal storage of all walls.
air to the zone. Firstly, the model is deduced from an energy-and-mass balance.
Fig. 2 shows a typical operation schematic for each AHU. The Then the model is explained using basic electrical circuit theory.
AHU draws air from the zone with temperature Tz(t), which Each zone can be modeled as a box with 6 faces as shown in
determines supply air temperature set-points (expected supply air Fig. 4, in which face ACGE is the wall connecting outside with
temperature, Tsp) and supply air fan speed set-points (expected temperature To; EFHG is the ground floor with temperature TG;
supply air fan speed, Fsp). The AHU forces the air from the mixing ABDC is ceiling with temperature TC; CDHG, BDHF and ABFE are the
box through cooling/heating coils with flow rate fa, which is largely walls connecting with adjacent zones that have the same
proportional to the fan speed Fsa, resulting in supply air with temperature TA.
temperature Tsa, and then discharges the cooled/heated air back The overall heat flow in the zone comes from air convection and
into the zone. A certain amount of fresh air may be introduced from wall conduction, as expressed in the following equation:
the outside air inlet so that fresh air in the zone may be gradually
introduced. dT z f r C a ðT sa  T z Þ þ U w Aw ðT o  T z Þ
Cz ¼ a a (1)
Fig. 3 shows an example of proportional control relationship dt Gw
between the supply air temperature set-point (Tsp) and supply air where Cz = CaraVz is the overall thermal capacity (kJ/C) of the zone;
fan speed set-point (Fsp) versus zone temperature (Tz) for a Ca is the specific heat of air (kJ/(kg 8C)); ra is the density of the air
particular HVAC system, in which the supply air temperature set- (kg/m3); Vz is volume of the zone; fa is the air flow rate (m3/s); Tsa is
point varies between 18 8C and 32 8C, with the supply air fan speed supply air temperature (8C); Tz is the zone temperature (8C); To is
set-point varying between 40% and 100%. The supply air external temperature (8C); the thermal conductivity of the wall is
temperature set-point (Tsp) is the expected supply air temperature, denoted by Uw (W/mK); Aw is heat transfer area (m2), i.e. area of
which is a function of zone temperature, and has inflexion points at each wall; and Gw is wall thickness (m).
the zone temperatures of 19 8C and 21.7 8C. The supply air fan Gw
To simplify this, letting Ro ¼ Uw Aw be a resistor representing the
speed set-point (Fsp) is the expected supply air fan speed, which is thermal properties of the wall, Eq. (1) becomes:
also a function of zone temperature, and has inflexion points at the
zone temperatures of 21.7 8C and 27 8C. A Tz, equal to 19 8C in this dT z f r C a ðT sa  T z Þ þ ðT o  T z Þ
Cz ¼ a a (2)
instance, is defined as the supply air minimum control point (SA dt Ro

Eq. (2) can be explained using RC circuit theory as follows. Heat

flow generated from a temperature difference, is analogous to
electrical current generated from a voltage difference. Therefore,
temperature can be modeled as voltage and heat flow can be
modeled as current, i.e.,
dH
Ia ¼ ¼ f a ra C a ðT sa  T z Þ:
dt
Fig. 5 is a simple analogue electric circuit used to represent
functions of the building elements, in which the outside wall is

Fig. 2. Operation schematic for a typical AHU system.

Fig. 3. Supply air temperature and fan speed set-points vs. zone temperature. Fig. 4. Zone thermal interfaces.
 G. Platt et al. / Energy and Buildings 42 (2010) 412–421 415

where D1 = faraCa, fa is the air flow rate, which is proportional to the

fan speed Fsa; supply air temperature Tsa usually oscillates around
its set-point Tsp, which is a function of zone temperature as shown
in Fig. 3; ra is the density of the air with value 1.204 (kg/m3) at
20 8C; Ca is specific heat of air with value 1.005 (kJ/kg 8C).
Parameter D2 is added to take into account factors such as leakage
of the zone, measurement error, dynamics of the zone, etc.

4. Dynamic zone model fitting using a genetic algorithm

Having derived a basic model for the zone, our aim is to find
Fig. 5. Simple RC model of an HVAC zone’s thermal performance. optimal parameters for the zone model that best fit with the
measured data. We obtain the solutions using a genetic algorithm
of a kind that has previously achieved successful model fitting [2],
modeled as a resistor (Ro) to represent its function of transferring multi-agent coordination [17–22] and been used in similar
heat between outside and inside. The zone is modeled as a applications related to electricity markets [23].
capacitor (Cz) to represent its function of both ‘charging’ through
heat input from supply air and ‘discharging’ through heat. 4.1. Genetic algorithm
In the figure, the zone’s average temperature is Tz. It has two
sources, one is supply air source with node voltage (temperature) A genetic algorithm (GA) [8,9,24] is a search method inspired by
Tsa; another is outside air source with node voltage (temperature) natural selection and survival of the fittest in the biological world.
To. The current (heat flow) is Ia ¼ dH dt
¼ f a ra C a ðT sa  T z Þ, where A GA is a type of ‘‘evolutionary algorithm’’. By using a GA approach,
(Tsa  Tz) is voltage (temperature) difference in the zone, Ra ¼ excellent results can be achieved through the evolution of the
1= f a ra C a is the variable zone air resistance. The wall has resistance system. A colony of rule sets can be evolved for a number of
Ro. The voltage (temperature difference) across Ro is (To  Tz). generations, improving the performance of the colony. Techniques
Therefore current (heat flow) flowing through Ro is Io = (To  Tz)/Ro. of fitness determination, selection, cross-over, reproduction, and
The voltage (temperature) difference on capacitor Cz is DTz, and mutation are applied to the rules and their chromosomal
thus the current (heat flow) flowing through the capacitor will be representation. Essentially, the parameters we are seeking for
the zone model are ‘‘evolved’’ over successive generations.
dT z The first generation of parameters (initial population) can be
Iz ¼ C z (3)
dt either random or predetermined values. At each subsequent
We know Iz = Ia + Io, which is represented by Eq. (2). generation of parameters, every individual of the population must
Until now we have deduced the zone model from both energy/ be evaluated to be able to distinguish between good and bad
mass balance and basic circuit theory point of view. The zone is a individuals. This is done by mapping the objective function to a
dynamic environment that changes over time. The dynamics of this ‘‘fitness function’’, which is a non-negative well-behaved measure of
environment are partly attributable to events such as people relative fitness of the parameters. The better the fitness of a given
walking in or out, doors and windows being opened or closed. To rule, the more likely it is to be selected. After the two parent rules are
realize adaptability to the dynamic environment and simplify the selected, each is represented by a ‘‘chromosomal’’ string and then
analysis, the zone model shown in Fig. 5 can be represented as a combined to determine the chromosomes of the two resulting rules
time-variant system, represented by a time-variant RC circuit (offspring). These chromosomes are subjected to potential mutation,
(variant RoCz, or variant Ro). and are then converted back to their equivalent rule representation.
Suppose the air supplied into the zone is evenly distributed The selected parents are then replaced in the colony by the offspring.
across the whole zone; outside air with temperature To transfers This mechanism of natural selection is expected to eventually result
the air into the zone through the outside wall, i.e., ACGE in Fig. 4; in a population with a higher performance.
the ceiling, ground and corridor in adjacent zones have the same
temperature as the zone under study, i.e., Tz = TA in Fig. 4. Such a 4.2. Genetic algorithm design for our system
model ignores the thermal storage of all walls.
The discrete expression of (2) is 4.2.1. Fitness function of optimization
The simulated prediction of the zone temperature with the
C z ðT z ðnÞ  T z ðn  1ÞÞ model expressed as Eq. (5) is used to compare with the measured
¼ f a ra C a ðT sa ðn  1Þ  T z ðnÞÞ
Dt zone temperature. The optimized parameters are D1, D2 and Ro of
ðT o ðn  1Þ  T z ðnÞÞ the model that give the best fit with the operational data. The
þ (4)
Ro objective function, O, of such an optimization employs mean
square fitting errors defined in Eq. (6). Eq. (7) represents the fitness
f a ra C a DtT sa ðn  1Þ þ ðDt=Ro ÞT o ðn  1Þ þ C z T z ðn  1Þ function f, which is the reciprocal of the objective function.
T z ðnÞ ¼
C z þ f a ra C a Dt þ ðDt=Ro Þ
1X N
0 2
To take into account factors such as the leakage of the zone and OðD1 ; D2 ; Ro Þ ¼ ðT ðiÞ  T z ðiÞÞ (6)
N i¼1 z
measurement error, a parameter D2 can be added to the above
equation. Therefore the zone temperature can be expressed as where Tz and T 0z are the fitted and measured zone temperature
Eq. (5): respectively, and N is the number of the sample data.

1
f ¼ f ðD1 ; D2 ; Ro Þ ¼ (7)
D1 DtT sa ðn  1Þ þ ðDt=Ro ÞT o ðn  1Þ þ C z T z ðn  1Þ OðD1 ; D2 ; Ro Þ
T z ðnÞ ¼ þ D2
C z þ D1 Dt þ ðDt=Ro Þ
A genetic algorithm (GA) is employed to search for the optimal
(5) values as illustrated in Section 4.2.2.
416 G. Platt et al. / Energy and Buildings 42 (2010) 412–421

Table 1
Parameters for zone modeling.

Name Definition Parameter availability

D1 Time varying coefficient in Eq. (5) Learning parameter

D2 Time varying coefficient in Eq. (5) Learning parameter
Ro Time varying resistor of the wall Learning parameter
connecting outside in Eq. (5)
Tz Average zone temperature Measured parameter
To Outside temperature Measured parameter
Tsa Supply air temperature Measured parameter

zone area with flow rate fa and then exchanges heat with the zone.
The next step zone temperature Tz(t + 1) is a function of the zone
model, supply air temperature and supply air flow rate. The AHU
GA operation is shown in Fig. 6.
Fig. 6. AHU GA operation. Fig. 6 are supply air temperature set-points, fan speed set-
points, current zone temperature and current external tempera-
4.2.2. Optimization ture. The output is the next step zone temperature.
As shown in Fig. 2, the supply air temperature set-points As expressed in Eq. (5), the zone model is determined by several
(expected supply air temperature, Tsp) and supply air fan speed set- parameters. Some of them are known; some are measured; some
points (expected supply fan speed Fsp) are predetermined inputs can be calculated; and some are unknown parameters that need to
based on the current zone temperature (Tz(t)) specified in the be learnt by the genetic algorithm. Table 1 summarizes the
building management and control system (BMCS) logic. The BMCS parameters for zone modeling, where the set of learning
controls supply air fan speed Fsa based on set-points to allow mixed parameters must each be learnt by the GA.
air to pass over the cooling/heating coil with air flow rate fa, Fig. 7 shows schematically the flow chart of the GA learner
resulting in supply air with temperature Tsa. Supply air with developed for fitting the zone model parameters. In the genetic
temperature Tsa flows through the insulated supply duct into the algorithm, the three parameters (D1, D2, Ro) constitute the
chromosome of an individual. Initializing the three parameters
produces an initial population to start the first generation GA
process. Termination of a GA process is decided if the number of
the current generation is equal to a predefined maximum
generation number. At least two generations are necessary when
running the GA optimization.

Fig. 8. Real-time zone model fitting and prediction using GA learner: (a) flow chart
of ½ hour zone model fitting and prediction; (b) time series of model fitting and
Fig. 7. Flow chart of GA learner for fitting the zone model parameters. prediction.
 G. Platt et al. / Energy and Buildings 42 (2010) 412–421 417

4.2.3. Real-time zone model fitting and prediction using GA ture Tz(n + 1); using predicted zone temperature Tz(n + 1) to get
As we have described above, the HVAC zone is a dynamic next step predicted zone temperature Tz(n + 2); and so on.
environment. To cope with its dynamics, a real-time zone model T z ðnÞ ) T z ðn þ 1Þ ) T z ðn þ 2Þ )    ) T z ðn þ MÞ
fitting using GA learner is developed. Fig. 8(a) shows schematically
the flow chart of ½ hour zone model fitting and prediction based on
the GA learner introduced above. 5. Experimental results
Fig. 8(b) shows the time series of model fitting and prediction.
In the figure T 0z represents the measured zone temperature and Tz 5.1. Training data
represents the calculated (fitted and predicted) zone temperatures.
Suppose current time is tn. Training data collected in past time A number of measurements have been conducted to gather
period, Vi ¼ t n  t nN , is used for model fitting using the genetic historical data for the purposes of thermal model fitting within the
algorithm, and generates the best parameters for the zone model, CSIRO Energy Centre Level 3 (East) office wing, in which three
which is called the fitted model. The fitted model is then used to AHUs service three adjacent HVAC zones consisting of 18 offices
predict the future zone temperature, e.g., temperature during time and associated corridors. Each office or space has a dry bulb
period V j ¼ t nþM  t n . The training data used in the process temperature sensor to detect the internal temperature. The zone
includes measured supply air temperature Tsa, external tempera- temperature is the average value across all offices and/or spaces in
ture To and supply air fan speed Fsa. The process is repeated for the zone. Fig. 9 shows the CSIRO energy centre building. Fig. 10
every tk time interval. The details of the process are as follows. gives a floor plan of the AHUs and their associated offices/spaces,
where different zones are represented as different grey color
1. The measured zone temperature over a past time period is shades. AHU-08 solely services office 331, with AHU-09 and AHU-
represented as
½T 0z ðn  NÞ; T 0z ðn  N þ 1Þ; . . . ; T 0z ðn  2Þ; T 0z ðn  1Þ; T 0z ðnÞ

where N is the number of sample data points in the past time

period.
2. The fitted zone temperature based on Eq. (5) is represented as
½T z ðn  NÞ; T z ðn  N þ 1Þ; . . . ; T z ðn  2Þ; T z ðn  1Þ; T z ðnÞ

3. The predicted zone temperature based on Eq. (5) is represented

as
½T z ðnÞ; T z ðn þ 1Þ; T z ðn þ 2Þ; . . . ; T z ðn þ M  1Þ; T z ðn þ MÞ

4. In each generation of GA, each population of learning

parameters gives the calculated zone temperature one by one
using Eq. (5), i.e., using measured zone temperature T 0z ðn  NÞ to
calculate zone temperature Tz(n  N + 1); using fitted (calcu-
Fig. 10. CSIRO Energy Centre Level 3 (East).
lated) zone temperature Tz(n  N + 1) to get another zone
temperature Tz(n  N + 2) and so on.
T 0z ðn  NÞ ) T z ðn  N þ 1Þ )    ) T z ðn  1Þ ) T z ðnÞ

5. Compare the measured zone temperatures with the fitted ones

to calculate fitness value for each population as Eq. (7).
6. ‘‘Parent’’ rules are selected based on a fitness value. After the two
parent rules are selected, each is represented by a ‘‘chromoso-
mal’’ string and are then combined to form two new
chromosomes. These chromosomes are subjected to potential
mutation, and are then converted back to their equivalent rule
representation. The selected parents are then replaced in the
colony by the offspring.
7. Repeat the process steps 4–6 for the number of required
generations and obtain the best population, i.e., the best
learning parameters.
8. Based on the best learned parameters for the past time period,
do the prediction for the future time one by one, i.e., using
current zone temperature Tz(n) to get predicted zone tempera-

Fig. 9. The CSIRO Energy Centre, Newcastle NSW, Australia. Fig. 11. Collected data for the AHU between 3rd and 4th of April 2007.
418 G. Platt et al. / Energy and Buildings 42 (2010) 412–421

Fig. 12. External temperature.

10 servicing the northern and southern zones respectively. In

addition AHU-09 also services the corridor.
The data we used in this work was collected between the 3rd
and 4th of April 2007, as shown in Fig. 11. Each day, the AHU runs
between 7:00 am and 5:00 pm, with a constant supply air set-
point of 18 8C. The data sample rate was 5 min. Our experiments
were carried out using Matlab on a laptop with 2 GB RAM and an
Intel Core 2 Duo T7600 CPU running at 2.33 GHz.

5.2. Real-time zone modeling and temperature prediction

An HVAC zone is a dynamic environment with people walking

in and out, doors opening and closing, etc. To cope with such an
environment, a dynamic model is necessary. To achieve this we
have developed a soft real-time model fitting and prediction
algorithm. Every V time interval, the past Vi period data is used for
zone model fitting with the GA, then the fitted zone model is used Fig. 13. Supply air fan speed (a) and temperature (b).
to predict future Vj period zone temperature as described in
Section 4.2.3.
set-point Tsp, and supply air fan speed Fsa is maintained very close
5.2.1. Supply air and external air prediction to its set-point Fsp.
As shown in Fig. 6, the future zone states are a function of the
zone model, external temperature and supply air. The predicted 5.2.2. Model fitting and zone temperature prediction
supply and external air properties are essential to model and
predict zone temperature. 5.2.2.1. Using predicted external temperature. In our experiments,
In our experiments, external temperature in some future period external temperature in some future time, e.g., ½ h, 1 h or 2 h, is
of time is assumed to be the same as current external temperature. assumed to be the same as the current one and used for model
This assumption is reasonable for short-term prediction. Fig. 12 fitting and zone temperature prediction. Experiment results are
shows the external temperature T̂o for ½ hour prediction, i.e., next shown in Table 2 and Fig. 14.
½ hour external temperature is very close to the current one. Table 2 gives fitting and prediction accuracy measurement
The supply air properties include temperature and flow rate using the mean square error (MSE). Fig. 14 shows the zone model
that are proportional to fan speed. Both expected supply air fan fitting and zone temperature prediction results for different time
speed and temperature can be calculated through the supply air strategies. The past ½ hour’s data is used to fit the model, and then
set-point function as shown in Fig. 3. Suppose we have current predict the next ½ hour zone temperature, and this process repeats
zone temperature Tz, then based on the function shown in Fig. 3 we every ½ hour as shown in Fig. 14(a).
will know our expected supply air fan speed Fsp and temperature
Tsp. If we assume the BMCS is able to obtain the expected supply air Table 2
perfectly, Fsp and Tsp can be used as predicted supply air fan speed Parameters for zone modeling.
F̂sa and temperature T̂sa , i.e., F̂sa ¼ F sp and F̂sa ¼ F sp . Applying T̂o , F̂sa Time strategy Fitting error (8C) Prediction error (8C)
and T̂sa to the zone model, the next step zone temperature will be
½ hour fitting, prediction 0.0008 0.0034
predicted T̂z , which will be used to get next step Fsp and Tsp, then F̂sa and ½ hour repetition
and T̂sa afterwards, and so on. Fig. 13 shows the predicted supply 1 hour fitting, prediction 0.0013 0.0099
air fan speed and temperature based on the set-point function and and 1 hour repetition
compared with the measured one. As we have mentioned before, 1 hour fitting, 2 hour prediction 0.0031 0.0287
and 2 hour repetition
the real supply air temperature Tsa generally oscillates around its
 G. Platt et al. / Energy and Buildings 42 (2010) 412–421 419

Fig. 15. Zone model fitting and temperature prediction with precise external
temperature; (a) ½ hour fitting, ½ hour prediction and ½ hour repetition; (b) 1 hour
fitting, 2 hour prediction and 2 hour repetition.

5.2.2.3. Model fitting and zone temperature prediction for variable

supply air set-point data. Supply air set-point function as illu-
strated in Fig. 3 has constant value at the two ends and variable
value in the middle. The experiments above work on constant
supply air temperature set-point, e.g. 18 8C, where it was found our
Fig. 14. Zone model fitting and temperature prediction for different fitting and
proposed model fitting and prediction algorithm works very well.
prediction strategies; (a) ½ hour fitting, ½ hour prediction and ½ hour repetition; To investigate the algorithm’s adaptability to variable supply air
(b) 1 hour fitting, 1 hour prediction and 1 hour repetition; (c) 1 hour fitting, 2 hour temperature set-points, further experiments were conducted.
prediction and 2 hour repetition. The data for these experiments was collected between the 11th
and 13th of March 2008 as shown in Fig. 16. Each day, the AHU runs
between 4:00 am and 5:00 pm, with the supply air set-point varied
From the experimental results, we can see that the model fitting between 18 8C and 20 8C. Data sample rate is 5 min. Fig. 17 shows
is very good for different fitting times, and the prediction accuracy is the results of different fitting and prediction strategies for such
strongly dependant on the prediction period. The shorter the variable set-point data. From the figure, we can see that our
prediction time, the more accurate the prediction for the indoor algorithm can properly adapt to variable supply air set-points.
temperature that is achieved. This difference is due to the dynamics
of zone environment and external temperature variation. 6. A Kalman filtering method with delayed feedback

5.2.2.2. Using precise external temperature. In order to see the effect To evaluate the performance of our modeling technique, we
of inaccurate external temperature predictions on the zone model compare it with a Kalman-filter-based modeling technique. The
prediction accuracy, experiments were run using the real external standard Kalman filter is a recursive filter that estimates the state
temperature. The results are shown in Fig. 15 for short time of a dynamic and often noisy system from a series of incomplete
prediction (e.g., ½ hour) and long time prediction (e.g., 2 h). By measurements, which are sometimes further degraded by
comparing corresponding results in Figs. 14 and 15, we can see that measurement noise. The main mechanism in a Kalman filter is
for ½ hour prediction the zone temperature prediction error feedback control: the filter iteratively estimates the process state
caused by inaccurate external temperature predictions is much and obtains feedback in the form of measurements. Thus, each
smaller than for the 2 hour predictions. iteration comprises two distinct phases prediction and correction.
420 G. Platt et al. / Energy and Buildings 42 (2010) 412–421

Fig. 16. Variable supply air data for the AHU between 11th and 13th of March 2008.

In the prediction phase, the next state of the system is predicted

based upon predetermined process equations. This reflects the a
priori understanding of the process. The corresponding estimated
error covariance matrix is also computed in this phase. Both these
estimates are updated when a measurement becomes available at
the next time step. The amount of the correction in this update
phase is heavily dependent on a factor called the Kalman gain,
which is calculated from a model of how the process is measured/
observed, the covariance of the measurement noise, and the
estimated error covariance.
Unfortunately, the standard Kalman filter is not directly
applicable to our zone modeling problem. Our requirement is
for a single phase of continuously sliding-forward prediction that
lasts a relatively long period, without any injection of feedback
measurements at any time during that period. In our application,
measurements are only available at the end of each period.
Therefore, a key ingredient in the Kalman approach is not
accessible.
On the other hand, feedback-based correction is definitely
necessary. This is because for a dynamic process our a priori Fig. 17. Zone model fitting and temperature prediction for different fitting and
knowledge is highly unlikely to be accurate all the time. Therefore, prediction strategies; (a) ½ hour fitting, ½ hour prediction and ½ hour repetition;
(b) 1 hour fitting, 1 hour prediction and 1 hour repetition; (c) 1 hour fitting, 2 hour
in the complete absence of measurement-based updates, the
prediction and 2 hour repetition.
prediction error is likely to be unacceptable.
A possible approach to addressing the dilemma is to take
advantage of historical trends in the corrections or updates. Our described in two phases in a forward-sliding time window, as
experiments are based upon such an attempt, albeit a relatively follows:
straightforward one.
Our system is described by the state vector 1. For the initial N time intervals, the standard Kalman algorithm is
  used. The update amount in each iteration is recorded for future
Tz
xn ¼ use;
F sa
2. Starting from the nth interval, and through to the (n + M)th
Although the temperature in the HVAC zone can be rather interval, the corrections are no longer based on instant feedback,
accurately determined, the measured fan speed is subject to but instead based on the correction applied at the previous
significant noise due to mechanical inaccuracy and latency. In our corresponding time instant;
experiments, the measurement noise is assumed to be 10 times as 3. The above two phases are repeated as the time window slides
high as the process noise. The Kalman procedure we employ can be forward.
 G. Platt et al. / Energy and Buildings 42 (2010) 412–421 421

volatility of wholesale electricity prices and assist constrained

networks during summer and winter demand peaks. The proposed
method is quite suitable for near-term predictions as required by
Australia’s half-hour electricity market cycle. As follow-up work,
we are focussing on heat exchanger modeling in order to model the
whole HVAC system and provide HVAC power usage predictions.
With such predictions, we can properly control energy flows in the
entire building and thus holistically optimize HVAC performance,
as well as managing its power consumption strategically for the
purpose of demand management.

References

[1] G. Hudson, C.P. Underwood, A simple building modeling procedure for MATLAB/
SIMULINK, in: Proceedings of the International Building Performance and Simu-
lation Conference, Kyoto, 1999.
[2] K. Tor, F. Ritter, Using a genetic algorithm to optimize the fit of cognitive models,
in: Proceedings of the International Conference on Cognitive Modeling, 2004, pp.
308–313.
[3] .G.J. Levermore, Building Energy Management Systems, E & FN Spon, London, 1992
[4] G.J. Levermore, Simple model for an optimiser, Building Services Engineering
Research & Technology 9 (3) (1988) 109–116.
[5] B. Tashtoush, M. Molhim, M. Rousan, Dynamic model of an HVAC system for
Fig. 18. Feedback-delayed Kalman prediction of the zone temperature for future
control analysis, Energy 30 (2005) 1729–1745.
time windows of ½ hour (a), 1 h (b) and 2 h (c).
[6] S. Karatasou, M. Santamouris, V. Geros, Modeling and predicting building’s energy
use with artificial neural networks: methods and results, Journal of Energy and
Buildings 38 (2006) 949–958.
Table 3 [7] B. Dong, C. Cao, S. Lee, Applying support vector machines to predict building
Feedback-delayed Kalman prediction of the zone temperature. energy consumption in tropical region, Journal of Energy and Buildings 37 (2005)
545–553.
Time strategy Prediction Error [8] S.W. Wang, X.H. Xu, Simplified building model for transient thermal performance
estimation using GA-based parameter identification, International Journal of
½ hour prediction 0.0043
Thermal Science 45 (4) (2006) 419–432.
1 hour prediction 0.0133
[9] S.W. Wang, X.H. Xu, Parameter estimation of internal thermal mass of building
2 hour prediction 0.0635 dynamic models using genetic algorithm, Energy Conversion and Management 47
(13–14) (2006) 1927–1941.
[10] S.W. Wang, Dynamic simulation of building VAV air-conditioning system and
evaluation of EMCS on-line control strategies, Building and Environment 34 (6)
Several experiments have been conducted using the constant (1999) 681–705.
supply air set-point data as used in Section 5.2.2.1. The results [11] C.-S. Park, G. Augenbroe, T. Messadi, M. Thitisawat, N. Sadegh, Calibration of a
using feedback-delayed Kalman zone temperature prediction for lumped simulation model for double-skin façade systems, Energy and Buildings
36 (2004) 1117–1130.
future time windows of ½ h (a), 1 h (b) and 2 h (c) are presented in [12] N. Djuric, V. Novakovic, F. Frydenlund, Heating system performance estimation
Fig. 18. Note that in this figure the first N data points are based using optimization tool and BEMS data, Energy and Buildings 40 (2008) 1367–
upon instant feedback as per the standard Kalman filter. The 1376.
[13] National Electricity Market Management Company (NEMMCO), Australia, 2007.
subsequent results are computed using the feedback-delayed http://www.nemmco.com.au/, accessed December 2008.
algorithm. [14] G. Welch, G. Bishop, An Introduction to the Kalman Filter, Technique Report
Error analysis of the feedback-delayed Kalman prediction of the Department of Computer Science, University of North Carolina at Chapel Hill,
Chapel Hill, NC 27599-3175, 2006.
zone temperature is tabulated in Table 3. The rows represent
[15] F. Ashley, Emery, Estimating parameters and refining thermal models by using the
predictions over time windows of ½ h, 1 h, 2 h and 3 h. As in extended Kalman filter approach, Journal Heat Transfer 126 (5) (2004) 809–817.
Table 2, the prediction errors are quantified by the MSE. [16] Saeid Habibi, Parameter estimation using a combined variable structure and
By comparing Tables 2 and 3, we can clearly see that our real- Kalman filtering approach, Journal of Dynamic Systems, Measurement, and
Control 130 (September (5)) (2008) 051004 (14 pages).
time HVAC zone modeling and temperature prediction algorithm [17] I.F. MacGill, Optimising decentralised power system operation using dual evolu-
achieves more accurate prediction of zone temperatures compared tionary programming, Ph.D. Thesis, UNSW, Australia, 1998.
to the feedback-delayed Kalman filter method. [18] Y. Guo, J. Li, G. James, Evolutionary optimisation of distributed energy resources,
in: The 18th Australian Joint Conference on Artificial Intelligence (AI’05), Sydney,
Australia, December 5–9, (2005), pp. 1086–1091.
7. Conclusions [19] J. Li, G. Poulton, G. James, Agent-based distributed energy management, in: 20th
Australian Joint Conference on Artificial Intelligence (AI’07), Gold Coast, Queens-
land, Australia, December 2–6, (2007), pp. 569–578.
This paper has introduced a soft real-time HVAC zone modeling [20] R. Li, J. Li, G. Poulton, G. James, Agent-based optimisation systems for electrical
and temperature prediction algorithm using a simple time varying load management, in: Proc. First International Workshop on Optimisation in
zone model with only a few parameters. The algorithm through Multi-Agent Systems (OPTMAS), Estoril, Portugal, (2008), pp. 60–69, in pewaa
May.
real-time learning can adequately adapt to the dynamics of the [21] J. Li, G. Poulton, G. James, Y. Guo, Multiple energy resource agent coordination
HVAC zone environment in an efficient manner, dealing with both based on electricity price, Journal of Distributed Energy Resources 5 (April–June
short-term unexpected changes and relatively long-term accu- (2)) (2009) 103–120.
[22] J. Li, G. James, G. Poulton, Set-points based optimal multi-agent coordination for
mulated changes. Modeling and prediction results from this work
controlling distributed energy loads, in: 3rd IEEE International Conference on
have shown to be promising. The shorter the prediction time, the Self-Adaptive and Self-Organizing Systems, San Francisco, California, September
more accurate the prediction of indoor temperature that can be 14–18, 2009.
achieved. Based on this work, we can model and predict zone [23] F.R. Tully, R.J. Kaye, Unit commitment in competitive electricity markets using
genetic algorithms, Trans. of the Institute of Electrical Engineers of Japan 117-B
status, which is a necessary step for managing and controlling (6) (1997) 815–821.
HVAC systems as distributed energy resources [20] to reduce the [24] J.P. Rennard, Introduction to genetic algorithms, PhD thesis, 2000.