Building and Environment 127 (2018) 1–12

Contents lists available at ScienceDirect

Building and Environment

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

A new simplified modeling method for model predictive control in a T

medium-sized commercial building: A case study
Junhua Zhuanga,c,∗, Yimin Chenb,c, Xiangguang Chena
a
School of Chemistry and Chemical Engineering, Beijing Institute of Technology, Beijing, China
b
Department of Civil Architectural and Environmental Engineering, Drexel University, Philadelphia, PA 19104, USA
c
School of Electrical and Information Engineering, Beijing University of Civil Engineering and Architecture, Beijing, China

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

Keywords: Model predictive control (MPC) methods for heating, ventilation and air conditioning (HVAC) systems have been
HVAC studied to improve the control accuracy and reduce energy consumption in recent years. The accuracy of the
Power spectral density model for building thermal dynamics in MPC plays a critical role to accurately control the system. The modeling
Model predictive control method also impacts on the real practice of MPC in buildings due to its cost and scalability. Studies have shown
Feedforward control structure
that an appropriate simplification of modeling procedure has minor impacts on the model accuracy, but in-
creases the modeling efficiency. In this article, variables including weather conditions, occupancy and electricity
are divided into two categories: manipulated variables and random variables. A novel two-step modeling
strategy is proposed for simplifying modeling procedure and increasing model accuracy. Manipulated variables
are used in step response method to develop system model. A low order system is obtained after the model
simplification by observing the response curve. Random variables are used in the power spectral density (PSD)
method for modeling. Transfer function is obtained through calculating the cross-power spectral density (CPSD)
of the system output and input, the PSD of the input, and the ratio of CPSD and PSD. A MPC strategy with
feedforward control structure is proposed to utilize the obtained dynamic characteristics of random variables
and effectively compensate the errors caused by these variables. Field test in a medium-sized commercial
building is implemented to evaluate the MPC strategy. The result shows that a considerable amount of energy
saving is achieved through the proposed MPC.

1. Introduction HVAC optimal control [3]. Khanmirza et al., presented a comparison of

In China, buildings are responsible for 28% of the total energy
consumption [1]. In commercial buildings, heating, ventilation, and air
conditioning (HVAC) system accounts for more than 60% building
electricity consumption to meet the demand of occupants' comfort.
Therefore, a potential benefit can be achieved and environmental im-
pacts can be reduced through energy saving in building HVAC systems.
Compared to HVAC equipment update in existing buildings, the im-
provement of HVAC control strategies is considered as one of the effi-
cient and cost effective ways to increase the building energy efficiency
[2].
Although traditional control methods which include the propor-
tional-integral-derivative (PID) control and the ON/OFF control are still
the most commonly used control methods in most commercial build-
ings, these methods have drawbacks in that system level characteristics
and multiple interactions among different components are not con-
sidered. In recent years, MPC has witnessed a rapid development in the
different control method for a building hybrid heating system, and
proved that maximum cost reduction can be achieved while indoor
temperature requirement can be met by implementing the MPC method
with economical optimization [4]. Salakij et al., developed a MPC
method to incorporate critical building information into control algo-
rithms based on a lower order system model modified from coupled
system model [5,6]. Through simulation and experiment validation,
they demonstrated that the proposed MPC method yielded superior
control performance with lowest energy consumption while main-
taining indoor thermal comfort.
Although there are extensive researches so far on the development
of MPC for increasing energy performance in buildings, there are only a
few real practice examples have been reported from literature. Huang
et al., implemented MPC in an airport terminal building. The simulation
and field experiment demonstrate that energy saving can be achieved
without losing thermal comfort after using the proposed hybrid model
predictive control [7]. Ruano et al., introduced the implementation of

∗
Corresponding author. School of Chemistry and Chemical Engineering, Beijing Institute of Technology, Beijing, China.
E-mail address: zhuangjh@bucea.edu.cn (J. Zhuang).

http://dx.doi.org/10.1016/j.buildenv.2017.10.022
Received 2 July 2017; Received in revised form 27 September 2017; Accepted 16 October 2017
Available online 18 October 2017
0360-1323/ © 2017 Elsevier Ltd. All rights reserved.
J. Zhuang et al.
an intelligent model based predictive control (IMBPC) system which
consists of the software and hardware in a real university campus
building [8]. Through the experiment for almost one month, HVAC
electrical saving can be reached and thermal comfort is also maintained
by using the IMBPC.
MPC achieves higher control quality compared with PID or ON/OFF
control due to its three features including: predictive model, rolling
optimization and feedback correction [9,10]. The procedure of mod-
eling is a critical prerequisite for a successful implementation of MPC in
buildings [11], because a control process will react to the prediction of
input changes or other disturbances [12]. Modeling can be divided into
physical-based modeling (or white box modeling) method and data-
driven based modeling (or black box modeling) method. Recently, data-
driven modeling methods have been widely utilized to model HVAC
systems for their advantages on cost-effectiveness and flexibility [13].
Artificial neural network (ANN) is one of widely employed data-driven
method in building energy prediction (modeling) due to the ability to
deal with nonlinear, multivariable modeling problems. Afram et al.,
gave a good review on artificial neural network (ANN) based MPC
system design [14]. A dynamic temperature setpoints profiles of the
zone air and buffer tank can be generated through MPC to save oper-
ating cost. Huang et al., developed an ANN model-based system iden-
tification method to model multi-zone buildings [15]. The model in-
cludes different energy inputs to increase model accuracy for MPC.
Ferreira et al., proposed a discrete model-based control method which
include three components as predictive models based on ANN, the cost
function aiming to minimize energy consumption and maintain thermal
comfort and a discrete branch and bound approach for optimization
[16].
As for physical-based method, it can provide detailed building in-
formation based on the analysis of building material features. So some
researchers proposed that rigorous model or physical model should be
implemented [17–19], but it is seldom employed in real engineering
practice for the high costs and expert involvement. Furthermore some
key parameters such as the heat transfer coefficient and specific heat
capacity of building envelope cannot be easily obtained, especially in
large size commercial buildings, and when there exists high couples
among different components in a system. Fortunately, some studies
show that a simplified model can achieve acceptable accuracy in
comparison to a detailed physical model for control [20–23]. Sourbron
et al., established a second-order model and a fourth-order model,
which included additional solar and internal gains in the identification
data set to improve performance as contrast to the second-order one, for
concrete core activation (CCA) [20]. The result indicated that both
second-order and forth-order model can achieve a satisfactory control
performance. Bălan et al., demonstrated a simple solution with a
second-order model for thermal modeling of a house which included
experimental identification of the model's parameters. Based the model,
they successfully used a predictive control algorithm to control the
thermal system in a house [21]. Karlsson et al., established a simplified
two-node lumped model for a single room [22]. Prívara et al., identified
a real building model by using subspace methods and successfully ap-
plied it into real control [23].
Similar to these low-order models stated above, resistance and ca-
pacitance (RC) network models or inverse models are widely used for
building energy estimation [24,25]. RC network modeling possesses a
meaningful and transparent configuration, making the developed
model more understandable and reliable to use [7]. Since the models
established by the RC method or some simplified methods are all linear
models, it is actually already using a priori knowledge – the system can
be approximated as a linear system, when using these methods. As for
large sized buildings, state-space modeling approach based on RC
method can also be applied [26], although sometime this approach are
applied in a single zone building model for details [25]. In practice, a
large sized building includes many zones, rather than two zones as in
Ref. [26]. Many zones will produce a high dimension matrix of the
2
Building and Environment 127 (2018) 1–12
state-space which results in great difficulties in practical application
[43]. For the purpose of chiller plant control, it is not necessary to
model all zones in the building. Moreover, the structures and para-
meters of a RC model obtained from one building can hardly be em-
ployed in other building. This makes the scalability of a RC model to be
lower in the real practice [11]. Therefore, some simplified methods
have been proposed to cope with this issue. For example, Braun et al.,
modeled a whole large sized building as a single zone instead of de-
veloping a multi-zone RC model [25]. The result indicated that there
was a close match between the predicted cooling load and the actual
cooling load. Sourbron et al., also demonstrated that the second-order
model can achieve equal control performance as the forth-order one
[20]. Richard et al. proved that it is feasible to simplify high order
systems to low order ones in engineering applications [12,27,28].
The system parameters identification (SPID) is another challenge
after the determination of system structure. Various SPID methods have
been proposed to optimize model parameters. O'Dwyer et al., developed
a system identification method from data in which significant un-
measured disturbances are present [29]. In this method, high-order
simulation models for control strategy analysis and low-order zone
models for optimization were developed separately for MPC. Karlsson
et al., identified the parameters by using step-response method and
pulse-response method [22]. It is effective to model the indoor tem-
perature influenced by the supply heat flux, since the supply heat flux
can be a step change or an approximate ideal impulse function. How-
ever, these methods that are called experimental methods can only be
applied to variables that can be manipulated (e.g. the supply heat flux,
the supply chilled water temperature), and can't applied to random
ones (e.g. the outdoor temperature, the solar radiation). By using si-
mulation, Fraisse et al., studied the internal surface temperature re-
sponse after a step change in external or internal temperature [30]. Yao
et al., also studied the dynamic response of air temperature in different
zones when subject to a step change of ambient temperature, occupant
number and etc. [31]. However, some weather variables such as out-
door temperature, solar radiation, cannot be step changed in real
practice.
Data-driven methods for model parameters identification can be
applied after system structures are determined (e.g. a simplified lower-
order models). These data-driven methods are better at handing
random variables such as weather prediction data. Various methods
including least squares method [32], regression method [33], Kalman
filter [23], ensemble Kalman filter [34], ARMAX model identification
and subspace identification [35] have been employed to estimate the
model parameters. Models developed through random variables can be
obtained by these methods. Some research works treat manipulated
variables such as temperatures and flow rates of supply chilled water, as
random variables [13]. For example, Li et al., pointed out that model
accuracy and calculation efficacy cannot be guaranteed through the
model identification methods which lack of active building excitation
[13]. This is because database data does not contain plenty of useful
information about systems, in other words, data rich and information
poor [36]. Fig. 1 shows a summary of models and parameter identifi-
cation methods mentioned above.
Although a number of successful cases of modeling can be found,
some issues still need to be solved to apply MPC in the real practice.
These include:
• Models inevitably have errors. The system control error cannot be
eliminated successfully even with high-order models because it is
difficult to determine the model parameters.
• The presence of external uncertain disturbance such as weather
prediction has significant impacts on the control quality.
Concerning the above problems, this article aims to develop a
strategy of MPC with feedforward control structure, which has both the
MPC feedback structure and the feedforward structure [37,38]. As
J. Zhuang et al.
Fig. 1. The summary of models and parameter identification.
mentioned above, the inputs of the controlled system can usually be
divided into two categories, one is manipulated variables, and the other
is uncontrollable variables, which are often random variables. The
purpose of control is to continuously adjust manipulated inputs to
overcome the influence of uncontrollable inputs, and make the output
has the desired dynamic characteristics. For irregular disturbances such
as measurement errors, this adjustment can only be made if it is re-
flected to the detectable variables such as outputs, and therefore the
feedback method must be used. However, for those random variables
(e.g. weather condition) that can be predicted, it is unreasonable if they
can be corrected by feedback as disturbances. The feedback control
only works when the error has impact on the output, which results in a
certain hysteresis. Furthermore, the feedback control does not take full
advantage of the known information of this part of the inputs. Thus, it is
more reasonable to implement a feedforward control which can take
advantage of the known dynamic characteristics of this part of the in-
puts, and can effectively compensate the errors caused by these inputs
[39].
This paper is as follows: Section 2 introduces the workflow of this
study including building and HVAC description. The bypass valve
control model is presented in Section 3, which use step response
method for manipulated variables and PSD method for random vari-
ables. Section 4 introduces the proposed MPC strategy with feedforward
control structure. In Section 5, a case study is used to demonstrate the
effectiveness of the proposed modeling and control strategy. Conclusion
and a future research plan are given in Section 6.
2. System description
2.1. Description of working flow
Various inputs, such as occupancy, electricity, outdoor temperature
etc., to the HVAC system have tremendous impacts on the system be-
havior. Various inputs should be analyzed respectively due to their
different roles and impacts. And they can be divided into some cate-
gories: constants, ignored variables, manipulated variables and random
variables.
In this study, two types of variable are pre-defined such that ma-
nipulated variables and random variables to develop a simplified con-
trol model. The step response method is used to model the system under
3
Building and Environment 127 (2018) 1–12
Fig. 2. The working flow of this study.
manipulated variables, where the system model is the bypass valve
control model. PSD method is applied to model the system under
random variables that are outdoor temperature and solar radiation. In
applying the PSD method, some parameters may be adjusted to make
the model more matched with the real system. These models are ex-
pressed in the form of transfer functions, and then applied in a MPC
method with feedforward control structure to the system. The whole
flow of the study is shown in Fig. 2.
2.2. Building description
The test building is a medium-sized commercial building of a
shopping mall in Sichuan province, China, located at 29°35′ north la-
titude, 103°43′ east longitude. The building has 31.4 m in height and
41,000 m2 with six-floor on the ground and two-floor basement. From
outside layer to inside layer, the exposed walls are all made of granite,
thermal insulation materials, gravel concrete, hollow brick, gravel
concrete, and wall tile or lime. The exposed windows are all double-
glazing windows. Fig. 3 shows the external shape of the test building.
Fig. 4 shows the building view in Google SketchUp™, in which the di-
rection of the building can be clearly seen.
2.3. HVAC system description
The chiller plant has three centrifugal chillers with variable speed
drivers for six floors above the ground and two screw chillers for two
floors underground. In this study, only the six floors part above the
ground will be studied, and the part is served by 28 air handling units
(AHUs) and 22 fan coils (FCs). Each AHU has a controller that controls a
valve to change chilled water flow according to return air temperature
by using the PID control method. The return air temperature is set to
26 °C. Each FC controls chilled water flow by using ON/OFF control
method according to the zone temperature directly. This study does not
focus on the control of zone temperatures, for the control performance
J. Zhuang et al. Building and Environment 127 (2018) 1–12

Fig. 3. The test building view.

of AHUs and FCs, by and large, is perfect, which means that zone
temperatures are controlled at about 26 °C. Although, the flow through
a single FC is at an ON/OFF status which can be simplified as a discrete
value, continuous value can also be obtained for the flow through the
whole system due to the flow through 28 AHUs, that is continuous
value, is much larger than 22 FCs and 22 FCs run at different time. The
cooling rated capacity of each of the three centrifugal chillers is
2286 kW, and the consumption power of each is 437 kW. The chillers
are parallel connected. Every day, the HVAC system works from 9:20 to
21:00. Usually there are two chillers running in July and August, and
the rest of a chiller is for redundancy and maintenance in most time,
unless in the very hot weather. The plant is shown in Fig. 5. Primary
pump system is used for chiller plant which includes constant speed Fig. 5. The composition of the plant.
pumps. A bypass valve is used to balance the cooling capacity of the
plant to the internal cooling load. The bypass valve is adjusted ac-
cording to the chilled water differential pressure which is set at
0.12 MPa. The system was designed to be operated as 7 °C for the
supply chilled water and 5 °C for the temperature difference between
supply and return water.

Fig. 4. The test building view in Google SketchUp™.

4
J. Zhuang et al.
3. Modeling procedure
3.1. Determination of model inputs and outputs
Model input and output variable selection is crucial to the control
system. Variable selection must be consistent with the process re-
quirement. The accuracy of system identification and model scalability
should also be considered. Most of researchers treat zone temperatures
or required amount of cooling or heat as controlled variables that is
system outputs.
This study focuses on how to generate and deliver cooling capacity
more efficiently. There is no potential for energy conservation in the
chilled water delivery on the whole because the chilled water pumps
are constant speed pumps. So the goal of saving energy in the primary
cooling system becomes a goal to save chiller energy consumption. The
load is determined by weather, occupant and electricity, the cooling
water system is not adjusted in this study, and the chilled water flow is
also approximately constant. In order to improve the efficiency of the
chillers, there is only one way that is to increase the supply chilled
water temperature.
The indoor temperature Ti is influenced by heat gain from outdoor
environment Q̇out , the solar radiation Q̇sol , the internal gains by occu-
pancy and electricity Q̇int , and the cooling capacity from the chiller
plant Q̇cool , as shown in Fig. 6. Controllers serving the AHUs and FCs
control the chilled water flow according to the indoor temperature, and
then the bypass valve opening level is controlled by another controller
that serves the bypass valve for maintaining the chilled water differ-
ential pressure. These two types of controllers are considered to be
perfect in this study. Under the condition that the supply chilled water
temperature remains approximately unchanged, the cooling capacity
changes in different bypass valve opening level. A closed-loop control
system is demonstrated in Fig. 6. Before the study in this paper was
implemented, the bypass valve opening level was controlled to meet the
building cooling load at a given supply chilled water temperature
(7 °C). But when the supply chilled water temperature increases to
improve the chiller efficiency and save energy, the bypass valve
opening level decreases correspondingly. The control objective is
therefore set to decrease the bypass valve opening level as small as
possible to meet the requirement of internal cooling load. So the bypass
valve opening level is used as the model output in this study.
As for input variables, they can be divided into two categories. One
category is manipulated variable that is the supply chilled water tem-
perature in this study. Another category are random variables that af-
fect the load, such as occupancy, electricity, outdoor temperature, solar
radiation, etc., which are treated as input variables during modeling,
and are treated as interference in control. Ref.21 points out that too
many inputs cannot improve the accuracy, but increase the difficulty of
identification. In this study, two weather inputs – outdoor temperature
and solar radiation, which are usually considered to have the greatest
impact on indoor temperature, are selected. The occupancy schedule is
relative fixed, and the occupancy density is about 0.03 Person/m2 in the
building, except for about 20% increase of occupancy between 19:00 to
21:00 each day. So in this study, the occupancy was regarded as a
Fig. 6. Simplified scheme of the system.
5
Building and Environment 127 (2018) 1–12
constant, and was not an input variable. Such a simplification does not
have much effect on the control. Similarly, the electricity power density
is also an approximately constant (value is 30 W/m2), not a variable,
since the electricity power schedule is also relative fixed.
3.2. Transfer function
A single-input single-output linear system can be described by a
transfer function G(s) as Eq. (1) [31].
Y (s ) L (y (t ))
G (s ) = =
U (s ) L (u (t )) (1)
Where s is the complex frequency variable that can be represented as a
a + bj form, Y(S) and U(S) are the Laplace transforms of the output y(t)
and the input u(t), respectively, and L (•) denotes the Laplace transform.
The Laplace transform is a mathematical tool used to solve linear or-
dinary differential equations.
A linear model has the properties of superposition and scalability,
which imply that the result of multiple inputs is the sum of their in-
dividual effects. In this study, the bypass valve control model includes
three inputs and one output, and their relationship can be described as
Eq. (2)
Y (s ) = Gsup (s ) Usup (s ) + Gout (s ) Uout (s ) + Gsol (s ) Usol (s ) (2)
where Usup(s), Uout(s) and Usol(s) denote the supply chilled water tem-
perature, outdoor temperature and solar radiation, respectively, Y(s) is
the system output – the bypass valve opening level, and Gsup(s), Gout(s),
Gsol(s) are three transfer functions of one output corresponding to three
inputs respectively. Eq. (2) can also be written in vector form as follows:
⎡Usup (s ) ⎤
Y (s ) = [Gsup (s ) Gout (s ) Gsol (s )] ⎢ Uout (s ) ⎥ = G(s)U(s)
⎢ ⎥
⎣ Usol (s ) ⎦ (3)
And the objective of modeling in this study is to find the correct transfer
functions.
As mentioned above, two of the three inputs are random variables
that will be modeled by using PSD method, and the remaining one is
manipulated variable that will modeled by step response analysis
method.
3.3. Model order
As mentioned above, the system is a three-inputs one-output system,
and the system model G(S) includes three transfer functions Gsup(s),
Gout(s) and Gsol(s) respectively. The order of each transfer function is
determined by the tradeoff between accuracy and complexity of mod-
eling. The increase of order leads to the improvement of model accu-
racy, but increase of the complexity in modeling and control. The order
of the system model is the sum of the orders of the three transfer
functions, and then the model order is also the tradeoff between ac-
curacy and complexity.
Resistance and capacitance (RC) method has been widely used to
model building thermal response. Fig. 7 shows a simple two-node
model for a hypothetical building external wall. The purpose of the
model is to reveal how outdoor temperature To affects indoor tem-
perature Ti, therefore, in this study, the model with indoor and outdoor
temperature is a part of the whole system model. A simple model shown
in Fig. 7(a) is used to approximate the relationship [23]. The zone node
has a thermal capacitance equal to C1, and R1 represents the thermal
resistance between outdoor and indoor. It is obvious that the model is a
first order model as shown in Eq. (4).
dTi
R1 C1 + Ti = To
dt (4)
A more complex model (see Fig. 7(b)) is also used to illustrate the
J. Zhuang et al.
relationship between the outdoor and indoor temperature [23]. The
thermal resistance of the external wall is divided into R1 and R2, and the
thermal capacitance of the wall C2 is added as an additional model
node. The model is described by state space equations as shown in Eq.
(5) and (6).
dTi 1 Fig. 9. The block diagram of the model.
C1 = (Tow − Ti )
dt R1 (5)
dTow 1 1
C2 = (To − Tow ) + (Ti − Tow )
dt R2 R1 (6)
Then the relationship between To and Ti can be rewritten as a second
order differential equation as shown in Eq. (7). In other words, the
model is a second order model.
d 2Ti dT
R1 R2 C1 C2 + (R1 C1 + R2 C2 − R2 C1) i + Ti = To
dt 2 dt (7)
It can be seen, the order of model has been determined at the be-
ginning of modeling by using RC method. However, any model is an
approximation to the modeled object, and at the beginning of modeling,
the information about the order has not yet been obtained. The order of
the model cannot be determined unless the response characteristics of
the model can be obtained after the experiment or data analysis are
implemented. At the same time, the tradeoff between accuracy and
complexity should also be considered when developing the model.
Therefore, at the beginning of modeling, the order of the model should
not be determined, and should be expressed as a “RC network”, as
shown in Fig. 7(c).
3.4. Modeling under manipulated variables
The manipulated variable is the supply chilled water temperature. If
the variable changes, the bypass valve opening level must change. As
shown in the dashed box of Fig. 8, if the load is constant, given a rise in
the supply chilled water temperature, the AHUs' valve opening level
will become larger, and the FCs' valve will open for longer time. It
causes the chilled water differential pressure to become smaller. In
order to maintain the differential pressure at the set value of 0.12 MPa,
the controller will adjust the bypass valve to make the opening level
smaller.
This paper studies the relationship between the supply chilled water
temperature and the bypass valve opening level. Taking into account
the short time in experiment, the load changes can be ignored, and the
block diagram of the model is shown in Fig. 9. The block in the figure
represents the HVAC system that contains the closed loop control
6
Building and Environment 127 (2018) 1–12
Fig. 7. RC model.
system as shown in the dashed box in Fig. 8. In other words, the HVAC
system with existing control system is the modeling object in this study.
Step response method can be used to determine the system transfer
function by injecting step signal and observing the system response
output. If the system is approximated as a first order system, it can be
expressed as a first order differential equation as Eq. (8)
dy (t ) 1 K
+ y (t ) = usup (t − τ )
dt T T (8)
And the corresponding transfer function is defines as:
Ke−τs
Gsup (s ) =
Ts + 1 (9)
where K is the gain, T is the time constant, and τ is the pure lag time.
These three parameters can be determined by the step response method.
As shown in Eq. (3), the transfer function Gsup(s) can be obtained by
keeping another two variables as zeroes. Of course, the zero means no
deviation from equilibrium. Thus, to minimize the impact of the load
during the experiment, it is appropriate to select a cloudy or rainy day
to do the experiment. Therefore August 1, 2016 was selected because it
was a rainy day. The difference between the highest and lowest tem-
perature of the day was 4 °C. The solar radiation was small and had
little change during the daytime. The influence from these two factors
were neglected as the fluctuation of the bypass valve opening level
caused by these two factors was small. Test result is presented as
Fig. 10. The step excitation was generated by setting the supply chilled
water temperature set point.
The transfer function can be obtained graphically [40]. The curve of
the bypass valve opening level is smoothed by least square method
firstly as in Fig. 11. By observing, the output curve is S-shaped, so the
system is a high order system. It is appropriate in engineering that the
transfer function is approximated as a first order model. The gain K can
be calculated directly from the steady-state values of the input and
output, and τ and T can be determined by plotting as shown in Fig. 11.
A tangent line is drawn for response curve, which intersects the two
steady state line. The time between the two intersections is time con-
stant T. Pure lag time can be determined by the time between the first
Fig. 8. The control block diagram of the system.
J. Zhuang et al.
Fig. 11. The transfer function determination by plotting.
intersection and the starting point of the step response. In the three step
responses in Fig. 10, it is untrusted at 9:20, for the initial value of 0%
did not reflect the operating condition of the HVAC system, but in-
dicates that it is turned off. The model parameters estimated by two
experiments at 14:00 and 17:30 are similar, and the average of two
identifications is taken as the model parameters. Finally, the transfer
function is:
7.03e−14.9s
Gsup (s ) =
84.25s + 1
3.5. Modeling for random variables by PSD method
It is obvious that outdoor temperature and solar radiation can not be
step changed, and the step response method is not applicable for these
two random variables. For random variables, a stochastic process
modeling method should be used. In this study, power spectral density
(PSD) is chosen for modeling, since it has the ability to suppress irre-
levant signals, and to improve the modeling accuracy.
Firstly, the cross power spectral density (CPSD) Syu (jω) between the
output and the input signal, and the PSD Suu (jω) of the input signal are
acquired, and they are the Fourier transform of the corresponding
correlation function as Eqs. (10) and (11). In this study, the output
signal y is the bypass valve opening level, and the input signal u is
outdoor temperature or solar radiation.
Syu (jω) = F (Ryu (τ ))
Building and Environment 127 (2018) 1–12
Fig. 10. The step response test result.
Suu (jω) = F (Ruu (τ )) (11)
where F (•) denotes the Fourier transform, Ruu (τ ) is the auto-correlation
function of the input signal, and Ryu (τ ) is the cross-correlation function
between the output and the input signals. In the calculation of the two
correlation function, some irrelevant variables such as measurement
error, noise, etc. are suppressed, because the so-call irrelevant is that
the cross-correlation function between two variables is zero. In other
words, the Ruu (τ ) and Ryu (τ ) do not include the information of irrele-
vant variables such as measurement error and noise.
Then, Eq. (12) is employed to obtain the transfer function. The
principle is that when a random signal passes through a linear system,
the ratio between the CPSD Syu (jω) and the PSD Suu (jω) is just the
system transfer function. When the complex frequency is defined as
s = jω, the transfer function in the Laplace transform G(s) must be
equal to the one in the Fourier transform. G(jω) [45]. It is interesting to
find that the formula of the transfer function for random variables is
similar to Eq. (1) for deterministic signals.
Syu (jω)
G (jω) =
Suu (jω) (12)
Unlike manipulated variables, modeling with random variables re-
quires a significant number of data samples. Data quality is as im-
portant as the data quantity when implementing a data-driven based
method. The data for the transition process is not desirable, for ex-
ample, the data obtained during the HVAC system startup in the
morning and shutdown in the evening, since from Eq. (2), modeling
with random variables requires that the manipulated variable does not
change, and the transition process means the manipulated variable
changed. Data of August 20–22, 2016 were selected. The sampling in-
terval was 30 s, and the data was obtained from 10:00 to 21:00 when
the HVAC system is under operation. During the experiment, the supply
chilled water temperature was maintained at the equilibrium value
7 °C.
In this study, the output is bypass valve opening level and the inputs
are outdoor temperature and solar radiation. As given in Eq. (12), the
CPSD of output and input, and PSD of input should be estimated, re-
spectively. Then the quotient of the CPSD and the PSD will also be
calculated. The results of the quotient for different inputs are presented
in terms of amplitude and phase at different frequencies in Figs. 12 and
13. A discrete filter is applied to fit the frequency response data by
using least squares method [44]. And the result will be a form where
both the numerator and the denominator are polynomials, which is
usually the form of the transfer function. Considering the tradeoff be-
(10) tween simplification and accuracy, second-order model were chosen to
7
J. Zhuang et al. Building and Environment 127 (2018) 1–12

Fig. 12. Transfer function amplitude and phase for outdoor

temperature.

obtain the transfer functions given as: inputs are determined, they are transformed to discrete form for sub-
sequent use. Secondly, these transfer functions are converted into the
2.125s + 1.268
Gout (s ) = form of state equations. Then the form of the cost function for MPC can
354.2s 2 + 42.31s + 1
be determined. The objective of this study is to minimize the opening
0.002352s + 0.01451 level of the bypass valve. Therefore, the following cost function can be
Gsol (s ) = obtained.
213.2s 2 + 25.42s + 1
N

4. Control system design

J (k ) = ∑ [w (k + i) − y (k + i k )]2
i=0 (13)

Firstly, as the three transfer functions which are related to the three subject to:

Fig. 13. Transfer function amplitude and phase for solar radiation.

8
J. Zhuang et al.
xk + i + 1 k = Axk + i k + Buk + i k + Ddk + i k , ∀ i = 0, …, N − 1 (14)
yk + i k = Cxk + i k (15)
u min ≤ uk + i k ≤ u max (16)
where N is the prediction horizon, the symbol uk + i k denotes the vari-
able u at time k+i predicted at time k, u is a vector of the control inputs
within the prediction horizon N, that is the supply chilled water tem-
perature as a scalar in this study, dk + i k are the random variables, out-
door temperature and solar radiation where the predicted value can be
obtained from the local weather forecast station and it represents the
feedforward control structure, yk + i k is the predicted output at time k,
which is obtained by iteratively solving Eqs. (14) and (15) using the
control inputs and random variables, u min and u max are set to 7 °C and
16 °C respectively.
As discussed in section 3.1, the smaller the bypass valve opening
level is, the higher the supply chilled water temperature is, which
means higher chiller efficiency. However, it means that the cooling
capacity is less than the requirement of the building cooling load when
the bypass valve is completely closed (i.e., the opening level is 0%). So
the opening level set point w (k + i) is set to be 10% to avoid the fully
closed bypass valve.
The output variable is the controlled variable, also known as the
control objective. When HVAC system energy saving is studied in this
paper, chiller profile should be incorporated into system modeling if
electricity consumption is chosen as the control variable. This would
increase the model complexity and therefore electricity consumption is
not adopted as the control output variable in this research. The flow
rate of supply chilled water is approximately constant due to the con-
stant speed pumps. Therefore, energy saving can be realized by in-
creasing the supply water temperature while maintaining the zone
comfort. There is no need to obtain the quantitative relationship be-
tween the supply chilled water and chiller energy consumption because
such relationship is not needed in the optimization calculation.
In Eq. (13), the supply water temperature is not used as a controlled
variable. However, in this study, the bypass valve opening level is se-
lected as the controlled variable as shown in Fig. 14. Here the HVAC
system with the existing control system is demonstrated in the dash box
in Fig. 8 and is illustrated in Fig. 9. Although the bypass valve opening
level is a manipulate variable in the original system as shown in Fig. 8,
it is used as a controlled variable as shown in Fig. 14 for the following
reason. If the supply water temperature is selected as the controlled
variable, the control structure in Fig. 14 has to be changed in Eq. (14) as
a constraint condition. Because in order to ensure the indoor comfort,
the supply water temperature cannot increase without limit, in other
words, the bypass valve opening level must not be closed. Solving this
constraint is same as the handling the opening level when it is used as
the controlled variable, therefore, can also be solved through Eqs.
(13)–(16).
The prediction horizon, N is normally selected as large as possible in
the premise of ensuring the computation ability. In this study, the
HVAC system is turned off in every night. Thus, the prediction horizon
at each sample time is variable. In each sampling time, the optimal
sequence {y∗ (k + i)}iN=−01 is obtained by using Eqs. (13)–(16). The first
9
Building and Environment 127 (2018) 1–12
one of the sequence y*(k) is implemented on the process. As new in-
formation becomes available at the next sampling time, the above
process will be repeated. Then a new sequence is generated, and the
first value of the new sequence is used to control.
The cost function in Eq. (13) minimizes a sum of the deviations of
the bypass valve opening level from the set point, which means the
control target will be achieved. And the control structure of the system
is shown in Fig. 15. In the figure, random variables are outdoor tem-
perature and solar radiation, and they directly act on the prediction
model and play the role of the feedforward. The prediction model,
which is composed of the three transfer functions described above,
predicts the future value of bypass valve opening level based on the two
random variables and the supply chilled water temperature. The pre-
dicted supply chilled water temperature enters the controller as a
feedback value, and then the controller, according to the optimization
control algorithm, adjusts the supply chilled water temperature to
control the bypass valve opening level to the setpoint.
5. Results and discussions
5.1. Model validation
The performance of the model under random variables is evaluated
by the normalized root mean squared error (NRMSE) fitness value:
n
1 ∑1 (y (k ) − yˆ (k ))2
NRMSE =
ymax − ymin n (17)
where y is the real output value and ŷ is the calculated output value. If
the matching between the calculated value and the real value is too
low, the parameters of the PSD method should be changed, and then the
PSD method is re-applied to modeling until a satisfactory matching
degree is obtained. The parameters include samples of overlap from
section to section and the selected window function. After sensitivity
test, parameters for PSD were set as follows: the window is 10 that
means a Hamming window of equal length is used, and the overlap is
the default value 50%.
The bypass valve opening level calculated by the weather data of
August 23, 2016 is compared to the real opening level, as shown in
Fig. 16. NRMSE is 0.127 meaning that the model has a fitness of 87.3%.
Fig. 17 demonstrates the discrepancy between the calculated value Yp
and the real value Yr, all the blue point are clustered around the line:
Yr = Yp. From these two figures, it can be seen that the model accuracy
is satisfied. The comparison does not include the transition process
because the transition process after chillers re-start is not modeled in
this study.
5.2. Control simulation
Before incorporating the MPC strategy in an online experiment, it
was tested under simulation environment. The objective of the simu-
lation study was to examine the closed loop performance. Usually,
when MPC is applied online, the control object outputs are measured by
the sensors and then fed back to the controller. For simulation purpose,
Fig. 14. The control structure of this study.
J. Zhuang et al.
Fig. 16. Comparison of the calculated and real value.
Fig. 17. Bypass valve opening level prediction error.
the controlled object in Fig. 15 is replaced by the model same as the
prediction model. Although the prediction model is different from the
real system, MPC can still overcome the small model mismatch, since
MPC algorithm includes a feedback alignment.
The weather prediction only have the maximum and minimum
outdoor temperatures for the next day, and the solar radiation has no
prediction. Outdoor temperature forecast can be replaced by sinusoidal
curve [41]. Solar radiation prediction can be obtained by Eq. (18) [42]:
10
Building and Environment 127 (2018) 1–12
Fig. 15. The block diagram of the MPC with feedforward
structure.
CC CC ⎞2
Ih = ⎧I0⋅sinh⋅⎡C0 + C1⋅ + C2⋅⎛ + C3⋅(θn − θn − 3) + C4 ϕ ⎤ − C5 ⎫
⎨ ⎢
⎣ 10 ⎝ 10 ⎠ ⎥
⎦ ⎬
⎩ ⎭
/k (18)
Where Ih is total horizontal radiation, I0 is solar constant and the value
is 1353 W/m2, h is solar elevation angel, CC is cloud cover and the
value is 0,1, …10, θn and θn − 3 are the temperatures at some moment
and 3 h ago, ϕ is relative humidity, and C0…C5 and k are constants
given in Table 1. Fig. 18 demonstrates the simulation result from test
data obtained in August 24, 2016. Since the transition process after
chillers re-start is not studied, the control is implemented when system
reaches a steady state.
5.3. Control experiment
After simulation, the data in same day is also used in real test.
Fig. 18 (a) and (b) show the bypass valve opening level and the supply
chilled water temperature in the simulation and the actual experiment
respectively. The results are very similar. It shows that MPC can over-
come the model mismatch caused by simplification in modeling pro-
cess. MPC algorithm includes three parts: prediction model, rolling
optimization and feedback correction, and the latter two ensure that the
control system has a feedback structure, which can tolerate a certain
degree of model mismatch.
The electricity consumption in August 23, 2016 and August 24,
2016 were used to evaluate the energy saving potential by increasing
supply chilled water temperature. The traditional control method was
implemented in August 23, which from 9:00 to 21:00, the average
outdoor temperature and solar radiation, and the electricity consump-
tion of the two days are listed in Table 2. It can be seen that two days of
climate is almost the same. And it can obtain a 16% chiller energy
saving that the control method for increasing supply chilled water
temperature is applied, compared with the traditional method that
maintain the temperature at 7 °C. In fact, if the outdoor temperature is
lower than the two experiment days, the energy saving potential is
greater due to the higher supply water temperature that can meet the
lower load. Increasing the supply chilled water temperature may affect
the dehumidification effect. And the indoor relative humidity is also
listed in Table 2. It can be seen that the humidity is slightly higher by
Table 1
Constant value used in Eq. (18) [42].
C0 C1 C2 C3 C4 C5 k
0.3645 0.48 −0.6335 0.0495 −0.0011 40.566 0.9900
J. Zhuang et al. Building and Environment 127 (2018) 1–12

Fig. 18. Comparison results of MPC strategy between simulation

and experiment.

Table 2 and PSD method are used respectively for system modeling under these
Comparison results of the electricity consumption in two days. two categories of variables. Step response method is more applicable in
the real practice due to its simplicity and easy to implement. PSD
Date Avg. outdoor Avg. solar Electricity Indoor relative
temperature radiation consumption humidity method belongs to data driven methods, and its accuracy depends on
the richness of the data. Given the proposed method, higher model
Aug. 23 35.4 °C 1145 W/m2 7337 kWh 49.3%–53.5% accuracy is achieved and modeling complex is lowered through being
Aug. 24 35.1 °C 1089 W/m2 6161 kWh 62.7%–68.7%
applied in a medium sized commercial building in China.
The test bypass valve in chiller plant was controlled through the
proposed MPC strategy. Energy saving can be achieved by the proposed
using the control method of this study, due to increased supply chilled
MPC strategy.
water temperature and dehumidification capacity decreased. Although,
We expect to further improve the accuracy of this bypass valve
humidity control is critical to maintain thermal comfort, the range of
control model by including more random variables such as occupancy
relative humidity in this building is set to 55%–70% which can be
and electricity rate. It is worth studying that how much of these vari-
maintained even the chilled water supply temperature fluctuates in
ables that will be added to improve the accuracy of the model, and then
certain scope. Therefore, in this research, the humidity control was not
improve the control performance.
considered.

6. Conclusion and future work
Buildings are usually approximated as linear systems, and there is a
tradeoff between the model accuracy and the control performance. If a
control strategy can well compensate model error, the model can be
simplified to a low order model, vice versa. In the real practice, the
error caused by the approximation of a high-order system to a low-
order system is allowable, and the approximation can greatly simplify
the analysis of modeling. This study proposes a new modeling method
for a medium-sized commercial building. A MPC strategy with feed-
forward control structure is developed to increase the supply chilled
water temperature.
In the proposed modeling framework, the inputs are divided into
two categories as manipulated variable and random variable to analyze
their effects on the system by different methods. Step response method
Acknowledgements
The research work presented in this paper is financially supported
by Beijing Municipal Education Commission science and technology
program [Grant No. SQKM201510016002].
References
[1] C.B. Zhou, S.J. Sheng, Study on integrated design of passive energy saving in hot
summer and warm winter region residential building, J. FUZHOU Univ. Nat. Sci.
Ed. 39 (5) (2011) 760–764 (in Chinese).
[2] Z.F. Xue, Y. Jiang, Study on the air conditioning energy consumption guideline in
commercial buildings, HV AC 35 (1) (2005) 37–41 (in Chinese).
[3] A. Afram, F. Janabi-Sharifi, Theory and applications of HVAC control systems – a
review of model predictive control (MPC), Build. Environ. 72 (1) (2014) 343–355.
[4] E. Khanmirza, A. Esmaeilzadeh, A.H.D. Markazi, Predictive control of a building
hybrid heating system for energy cost reduction, Appl. Soft Comput. 46 (2016)
407–423.

11
J. Zhuang et al.
[5] S. Salakij, N. Yu, S. Paolucci, P. Antsaklis, Model-Based Predictive Control for
building energy management. I: energy modeling and optimal control, Energy
Build. 133 (2016) 345–358.
[6] N. Yu, S. Salakij, R. Chavez, S. Paolucci, M. Sen, P. Antsaklis, Model-based pre-
dictive control for building energy management: Part II–Experimental validations,
Energy Build. 146 (2017) 19–26.
[7] H. Huang, L. Chen, E. Hu, A new model predictive control scheme for energy and
cost savings in commercial buildings: an airport terminal building case study, Build.
Environ. 89 (2015) 203–216.
[8] A.E. Ruano, S. Pesteh, S. Silva, H. Duarte, G. Mestre, P.M. Ferreira, H.R. Khosravani,
R. Horta, The IMBPC HVAC system: a complete MBPC solution for existing HVAC
systems, Energy Build. 120 (2016) 145–158.
[9] C.R. Cutler, B.L. Ramaker, Dynamic matrix control – a computer control algorithm,
Proceedings of the 1980 Joint Automatic Control Conference, Aug. 13-15, 1980 San
Francisco.
[10] S.J. Qin, T.A. Badgwell, A survey of industrial model predictive control technology,
Control Eng. Pract. 11 (7) (2003) 733–764.
[11] X. Li, J. Wen, Building energy consumption on-line forecasting using physics based
system identification, Energy Build. 82 (2014) 1–12.
[12] B.C. Kuo, F. Golnaraghi, Automatic Control Systems, eighth ed., John Wiley & Sons,
Inc., 2003.
[13] A. Afram, F. Janabi-Sharifi, Review of modeling methods for HVAC systems, Appl.
Therm. Eng. 67 (1–2) (2014) 507–519.
[14] A. Afram, F. Janabi-Sharifi, A.S. Fung, K. Raahemifar, Artificial neural network
(ANN) based model predictive control (MPC) and optimization of HVAC systems: a
state of the art review and case study of a residential HVAC system, Energy Build.
141 (2017) 96–113.
[15] H. Huang, L. Chen, E. Hu, A neural network-based multi-zone modelling approach
for predictive control system design in commercial buildings, Energy Build. 97
(2015) 86–97.
[16] P. Ferreira, A. Ruano, S. Silva, E. Conceicao, Neural networks based predictive
control for thermal comfort and energy savings in public buildings, Energy Build.
55 (2012) 238–251.
[17] J. Sun, A. Reddy, Optimal control of building HVAC & R systems using complete
simulation-based sequential quadratic programming (CSB-SQP), Build. Environ. 40
(5) (2005) 657–669.
[18] B. Tashtoush, M. Molhim, M. Al-Rousan, Dynamic model of an HVAC system for
control analysis, Energy 30 (10) (2005) 1729–1745.
[19] F. Tahersima, J. Stoustrup, H. Rasmussen, P.G. Nielsen, Thermal analysis of an
HVAC system with TRV controlled hydronic radiator, Automation Sci. Eng. 19 (11)
(2010) 756–761.
[20] M. Sourbron, C. Verhelst, L. Helsen, Building models for model predictive control of
office buildings with concrete core activation, J. Build. Perform. Simul. 6 (3) (2012)
1–24.
[21] R. Bălan, J. Cooper, K.-M. Chao, S. Stan, R. Donca, Parameter identification and
model based predictive control of temperature inside a house, Energy Build. 43
(2–3) (2011) 748–758.
[22] H. Karlsson, C.-E. Hagentoft, Application of model based predictive control for
water-based floor heating in low energy residential buildings, Build. Environ. 46 (3)
(2011) 556–569.
[23] S. Prívara, J. Široký, L. Ferkl, J. Cigler, Model predictive control of a building
heating system: the first experience, Energy Build. 43 (2–3) (2011) 564–572.
12
Building and Environment 127 (2018) 1–12
[24] Y. Ma, A. Kelman, A. Daly, F. Borrelli, Predictive control for energy efficient
buildings with thermal storage: modeling, stimulation, and experiments, IEEE
Control Syst. 32 (1) (2012) 44–64.
[25] J.E. Braun, N. Chaturvedi, An inverse gray-box model for transient building load
prediction, HVAC & R Res. 8 (1) (2002) 73–99.
[26] J. Hu, P. Karava, A state-space modeling approach and multi-level optimization
algorithm for predictive control of multi-zone buildings with mixed-mode cooling,
Build. Environ. 80 (7) (2014) 259–273.
[27] R.C. Dorf, R.H. Bishop, Modern Control Systems, eleventh ed., Pearson Prentice
Hall, 2008.
[28] D.E. Seborg, T.F. Edgar, D.A. Mellichamp, F.J. Doyle III, Process Dynamics and
Control, third ed., John Wiley & Sons, Inc, 2011.
[29] E. O'Dwyer, L. De Tommasi, K. Kouramas, M. Cychowski, G. Lightbody, Modelling
and disturbance estimation for model predictive control in building heating sys-
tems, Energy Build. 130 (2016) 532–545.
[30] G. Fraisse, C. Viardot, O. Lafabrie, G. Achard, Development of a simplified and
accurate building model based on electrical analogy, Energy Build. 34 (10) (2002)
1017–1031.
[31] Y. Yao, K. Yang, M. Huang, L. Wang, A state-space model for dynamic response of
indoor air temperature and humidity, Build. Environ. 64 (2013) 26–37.
[32] P.R. Armstrong, S.B. Leeb, L.K. Norford, Control with building mass – Part I:
thermal response model, ASHRAE Trans. (2006) 112.
[33] N.T. Gayeski, P.R. Armstrong, L.K. Norford, Predictive pre-cooling of thermo-active
building systems with low-lift chillers, HVAC & R Res. 18 (5) (2012) 858–873.
[34] B. Huchuk, C.A. Cruickshank, W. O'Brien, H.B. Gunay, Recursive Thermal Building
Model Training Using Ensemble Kalman Filters, Ottowa, Canada, eSim, (2014).
[35] L. Ferkl, J. Široky, Ceiling radiant cooling: comparison of ARMAX and subspace
identification modelling methods, Build. Environ. 45 (1) (2010) 205–212.
[36] P. Defabianis, A. Cimma, D. Tessaris, R. Lala, L. Basano, P. Appendino,
E. Caravaggio, Integration of fault detection and diagnosis with energy manage-
ment standard ISO 50001 and operations and maintenance of HVAC systems, Rehva
World Congress and International Conference on Iaqvec Clima, 2013.
[37] D.S. Carrasco, G.C. Goodwin, Feedforward model predictive control, Annu. Rev.
Control 35 (2) (2011) 199–206.
[38] J.A. Rossiter, G. Valencia-Palomo, Feed forward design in MPC, Control Conference,
2009.
[39] E.J. Davison, The feedforward control of linear multivariable time-invariant sys-
tems, Automatica 9 (5) (1973) 561–573.
[40] B.W. Bequette, Process Dynamics: Modeling, Analysis and Simulation, Prentice
Hall, 1998.
[41] P.C. Tabares-Velasco, C. Christensen, M. Bianchi, Verification and validation of
EnergyPlus phase change material model for opaque wall assemblies, Build.
Environ. 54 (7) (2012) 186–196.
[42] Q.Y. Zhang, H.X. Yang, Typical Meteorological Database Handbook for Buildings,
China Architecture & Building Press, 2012 (in Chinese).
[43] Y. Ma, Model Predictive Control for Energy Efficient Buildings, PhD Dissertations
University of California, Berkeley, 2012.
[44] L. He, Z. Jin, The population predicting based on the curve fitting least square
method, International Conference on Automation, Mechanical Control and
Computational Engineering, 2015.
[45] A.V. Oppenheim, A.S. Willsky, I.T. Young, Signals and Systems, Prentice-Hall Single
Processing Series Prentice-Hall, Englewood Cliffs, 1983.