Building and Environment 133 (2018) 237–245

Contents lists available at ScienceDirect

Building and Environment

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

Design and construction of a non-linear model predictive controller for T

building's cooling system
Arash Erfani, Abbas Rajabi-Ghahnaviyeh, Mehrdad Boroushaki∗
Department of Energy Engineering, Sharif University of Technology, P.O.Box 14565-114, Azadi Ave, Tehran, Iran

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

Keywords: This research aims to optimize a multi-zone Air Handling Unit's (AHU) energy consumption by using a Non-
Model predictive control linear Model Predictive Control (NMPC) approach. In this paper, Genetic Algorithm (GA) and Non-linear au-
Building energy saving toregressive network with exogenous inputs (NARX) have been utilized to design NMPC for a multi-zone AHU.
NARX modeling The NMPC problem could be divided into two main sections: internal model and the optimizer. NARX serves as
Building management system
the controller's internal model to predict the building's thermal dynamics. GA is then used to solve the NMPC
HVAC control
Multi-zone building
problem and find the optimal value of the control signals at each time step. The proposed NMPC jointly mini-
mizes energy consumption of the AHU and the deviation from the set-point temperature. Finally, the designed
controller was implemented and applied to the mentioned AHU. Also, a data acquisition system has been fab-
ricated to secure training and test data for NARX. Utilizing NARX for modeling system's dynamics resulted in a
highly accurate model with an accuracy of 97.71%. The empirical results of the proposed NMPC showed sig-
nificant reduction in gas and electricity consumption of the AHU. NMPC yielded a 55.1% and 43.7% reduction in
electricity and gas consumption of the AHU respectively.

1. Introduction
Considering the increasing importance of energy conservation in
recent decades, various methods and technologies have been developed
to improve different energy systems. Building section is responsible for
almost 40% of the global final energy use and 40% of greenhouse gas
emissions in the United States of America [1,2]. Taking into account the
huge contribution of Heating, Ventilation and Air Conditioning (HVAC)
systems in a building's energy use, optimizing these systems could have
a substantial effect on a building's energy consumption. HVAC systems
account for 50% of the building sector's energy consumption [3].
Hence, optimizing these systems is both economically and en-
vironmentally beneficial. HVAC systems are employed to provide
thermal comfort for the occupants in buildings.
To control HVAC systems, many control strategies have been pro-
posed. The most common controllers applied to HVAC systems are PID
(Proportional-Integral-Derivative), PI and on/off controllers. These
controllers use current thermal parameters to generate control signals.
But, since buildings have a very slow thermal dynamic, a large amount
of energy is wasted in the control process. It has been shown that model
predictive control is the best approach towards controlling building's
HVAC system [4]. MPC utilizes a model of the plant to predict the fu-
ture outputs or states of the system, and then it generates a control
vector that minimizes a certain objective function over a control hor-
izon [5]. Model Predictive Controller (MPC) is suitable for controlling
thermal behavior of a building, because it can handle plants with slow
dynamics, time-variant disturbances and non-linear constraints [6].
MPC is comprised of two integral parts: an internal model to predict the
future outputs and an optimizer to generate the optimal control signals.
Due to the decreasing cost of applying data-driven methods to different
complex plants, there has been a significant increase in utilizing in-
telligent and optimal control methods for Building Management Sys-
tems (BMS). Among the intelligent modeling methods, Artificial Neural
Networks have gained a lot of attention since they can learn the thermal
dynamics of a building pretty well and they are capable of efficiently
predicting the future thermal behaviors of a building. They have been
widely used for HVAC control [7–12] and energy consumption esti-
mation applications [13–19]. HVAC systems have high non-linearity
and delay which makes the control process sophisticated. An Artificial
Neural Network (ANN) has been used as the internal model of the
controller in order to predict the future outputs of the Air Handling Unit
(AHU). Since ANNs have a non-linear nature, the problem becomes a
Non-linear Model Predictive Control (NMPC) one. The most important
task of NMPC is finding the optimal control inputs to apply to the ac-
tuators. Previous researches have demonstrated that Genetic Algorithm
(GA) is the most effective tool for optimizing the performance of HVAC

∗
Corresponding author.
E-mail address: boroushaki@sharif.edu (M. Boroushaki).

https://doi.org/10.1016/j.buildenv.2018.02.022
Received 23 November 2017; Received in revised form 14 February 2018; Accepted 14 February 2018
Available online 19 February 2018
0360-1323/ © 2018 Elsevier Ltd. All rights reserved.
A. Erfani et al. Building and Environment 133 (2018) 237–245

Nomenclature Tmin Minimum allowed temperature (°C)

Tmax Maximum allowed temperature (°C)
HVAC Heating, ventilation and air conditioning systems NARX Non-linear autoregressive network with exogenous inputs
AHU Air handling unit ANN Artificial neural networks
CAV Constant air volume GA Genetic algorithm
VAV Variable air volume BMS Building management system
MPC Model predictive control PCB Printed circuit board
NMPC Non-linear model predictive control REA Robust evolutionary algorithm
PMV Predicted mean vote MREP Mean of relative error percentage
k Time step
Variables Hc Control horizon
Hp Prediction horizon
ṁ ret,z Return air from zth zone (m3/s) z Indices for zones
ṁ dis,z Discharge air to the zth zone (m3/s) p Output delay numbers
Pz Damper position of zth zone (%) q Input delay numbers
Pfa Fresh air damper position (%) rz zth zone reference trajectory
Ta Ambient temperature (°C) →
X Controller input vector
ṁ c Chilled water flow rate (%) →
Y Plant's output
Sahu AHU status →
C Control signals
CCin,w Cooling coil input water (kg/s)
Ptest Number of test patterns
CCout,w Cooling coil output water (kg/s)
Ptrain Number of training patterns
Tz zth zone temperature (°C)
Ts Set-point temperature
T̂z zth zone predicted temperature (°C)

systems [20–24]. Therefore, GA has been used as the optimizer in this
study.
Zhao et al. thoroughly investigated different methods for modeling
building's thermal dynamics. It was shown that data-driven methods are
most useful for on-line control applications in this literature. In addi-
tion, intelligent methods yield more accurate results than statistical
methods which makes them more reasonable for control applications
[25]. Afram and Janabi-Sharifi divided control strategies for HVAC
systems into four categories. Then a rigorous comparison was made
between these methods. Various experimental and simulation results of
different control methods were brought up and investigated based on
the established performance criteria in the literature. It was shown that
MPC proves more useful in HVAC system control applications [4].
Huang simulated zone temperature and damper position in a
Variable Air Volume (VAV) unit and controlled them using MPC. The
MPC results showed better transient response compared to a PI con-
troller [26]. Avci et al. developed an algorithm for determining the
temperature set-point regarding variable electricity price in California
and then tried to track the generated set-point by applying MPC [2].
Morosan et al. presented simulation of zone temperature regulation by
using decentralized, distributed and centralized MPCs [27]. Privara
et al. compared performance of MPC to a well-tuned controller with a
weather compensator in a big university [28]. Oldewurtel et al. took
into account the weather predictions and the uncertainty of the system.
Then, results of a Stochastic Model Predictive Controller (SMPC), non-
stochastic predictive controller or in other terms Certainty Equivalence
(CE) Rule-Based Controller (RBC) and Performance Bound Controller
(PB) were presented and compared with one another. The results
showed that the best performance is obtained when the SMPC is applied
[1]. Ferreira et al. designed and implemented MPC to maintain thermal
comfort quantified as Predicted Mean Vote (PMV) index while mini-
mizing energy consumption. They used ANN to model building's
thermal behavior [29]. Liang et al. designed MPC for controlling a
multi-zone VAV AHU and the simulation results were compared to the
installed PID controller. It was shown that MPC outperforms the PID
controller [6]. Siroky et al. computed set-point temperature for supply
water and room temperature periodically and weather forecasts were
downloaded and served as the MPC inputs. The results showed more
than 15% energy saving potential for a large building's heating system
[30]. Rehrl et al. compared both experimental and simulation results of
two control methods: MPC and exact linearization. They concluded that
in general MPC tracks the reference trajectory better than the other
approach [31]. Fong et al. optimized the operation of a simulated HVAC
system by means of a Robust Evolutionary Algorithm (REA) by de-
termining the optimal set-point of chilled water supply temperature
chiller and set-point of supply air temperature of AHU [32].
This paper addresses the problem of NMPC design and im-
plementation for a multi-zone Constant Air Volume (CAV) air handling
unit. The plant's outputs are the zones' temperature which are to track
set-point temperature, while minimizing the energy consumption. The
control signals in this work are damper positions related to each zone,
chilled water's flow rate and the AHU status (on/off). NMPC problem
could be divided into two main sections: internal model and the opti-
mizer [33]. Artificial neural network serves as the controller's internal
model, and its role is to predict the future temperature of each zone.
The optimal value for each of the control signals is calculated by means
of a genetic algorithm (which is the optimizer in this study) at each step
and then these signals are applied to the AHU through actuators. The
proposed NMPC strategy aims to minimize the electricity and gas en-
ergy carriers' consumption of the aforementioned AHU and the devia-
tion from the set-point temperature of each zone.
The paper is organized as follows. In section 2 the AHU system
along with the data logging system are described. Section 3, elaborates
on the ANN model utilized for predicting the zones' temperature. In
section 4, model predictive control method is succinctly described. In
section 5, designing the proposed NMPC for the AHU is rigorously ex-
plained and then experimental results of applying the NMPC to the
plant are presented and discussed in section 6. Section 7 concludes the
paper.
2. AHU system
2.1. AHU system structure
Case study in this research is an AHU with 3 zones. It is categorized
as a CAV air handling unit, which means that the volume of its input air
is constant while the temperature of this air is variable. A simple
schematic of the aforementioned AHU is shown in Fig. 1. First, the
returned air from the zones enters the mixing chamber and there it is
mixed with the fresh air (if the fresh air damper position is open). In the

238
A. Erfani et al. Building and Environment 133 (2018) 237–245

3. NARX model

A brief introduction to Artificial Neural Networks (ANN) and

especially NARX is given. As for the next part of this section, NARX
model of the building's cooling system is presented at length.

3.1. NARX structure

Since the buildings thermal dynamic is highly nonlinear and in-

volves many uncertainties of parameters, simple physical models
cannot capture the real dynamics of the system. Since ANNs are non-
linear in nature, they can learn sophisticated mappings between inputs
and outputs, so ANN is chosen as the model in this work [34]. The first
thing to determine is the structure of the ANN model. A very important
point to consider for choosing the ANN structure is that the outputs of
the system which are the zones' temperature in this case, are highly
dependent to the previous temperatures.
NARX is a type of Recurrent Neural Network, which is used to
model nonlinear dynamic systems. This network is based on a Multi-
Layer Perceptron (MLP). Now consider a single–input single-output
Fig. 1. General schematic of the system including the studied AHU. system. The model has a single input that is applied to a tapped-delay-
line memory of p units. It has a single output that is applied to a tapped-
delay-line memory of q units. The content of these two tapped-delay-
next level, mixing chamber's output passes through the chilled water's
line memories are used to feed the input layer of the MLP.
coil. It should be noted that only a part of air is cooled at this stage and
It has been proven that NARX have a better performance in learning
some of it is just circulated in the zones. The ratio of air cooled by the
long-term dependencies than the other structures of ANN [35]. Taking
chilled water coil is determined by the damper positions of the zones.
this fact into account, NARX structure was selected for the identifica-
After passing the air through the cooling coil chamber, the air is dis-
tion of the AHU system. Output in NARX structure is computed as
tributed and discharged to the zones by means of blowers and this cycle
follows.
is repeated, while the AHU is functioning. The cooling water is heated
after the hot air passes above the cooling coil. So the cooling water y (k+1) = F [x(k),… x(k-p+1), …, y(k),…y(k-q+1)] (1)
flows in the cycle of the absorption chiller to lose the extra heat as to be
y (k+1) is the future output of the system and x(k-i) and y(k-j) are the
able to chill the passing air again. It should also be mentioned that the
network's inputs and outputs in previous time steps respectively. p and
chilled water supply temperature is constant while the cooling system is
q are the number of input and output delays respectively. ANNs are
functional.
composed of a great number of elements called neurons, which are
connected to each other through weights. To model a plant, usually a
multi-layered perceptron (MLP) structure is used. It consists of an input
2.2. Data logging system
layer, an output layer and one or two hidden layers. Input data is in-
serted to the structure by the means of input nodes and after computing
AHU provides thermal comfort for four classes located at first floor
the output of every neuron in every layer, the final output is generated.
(each two adjacent rooms form a thermal zone) north of an educational
Each neuron's output is calculated by Eq. (2).
building, and the third zone is located at second floor of the mentioned
building above these four classes and it's used as the computer site.
⎛ ⎞
To collect data needed for training the neural network, a Printed
⎜∑ ij j
Oi = fi w o + θi
⎟
Circuit Board (PCB) was designed and constructed which has two tasks. ⎝ j ⎠ (2)
The first task is reading the sensors measured variables and sending it to
a web address through a SIM900 module. The second task is applying
the control signals to the AHU or in other words, carrying out the duties
of the actuators. Fig. 2 shows the electronic board designed for data
gathering.
A ST32F microprocessor is used to process received signals from the
sensors and transmit it through a SIM900 board. It should be mentioned
that microprocessor saves the sensors' measured values 6 times in a
single time step (17 min) and these values are averaged for each time
step to reduce noise and disturbance of data.
To collect temperature data of the 3 zones, three thermometers were
first calibrated and then placed in each zone. These thermometers are
equipped with a SIM900 module to send data to a predefined web
address. Data gathered for training the Non-linear autoregressive net-
work with exogenous inputs (NARX) model is saved in the mentioned
web address and then used for training and test purposes. As for the
actuating role, another web address is used that the module reads the
value written on this address each 17 min and applies those values
(which are the control signals) to the AHU.
Fig. 2. Electronic board used for Data logging.

239
A. Erfani et al. Building and Environment 133 (2018) 237–245

Where oi is the output of unit i, fi is the nonlinear activation function
that determines the neuron's output, wij is the weight that connects unit
j to i, and θi is the bias for unit i. In ANN literature, learning is inter-
preted as the process of tuning the weights between neurons. The most
common algorithm for determining the weights is called the back-
propagation algorithm which minimizes the learning error by a gra-
dient decent method [36,37].
3.2. NARX model of the building's cooling system
The inputs of the NARX model of the AHU system have been defined
as follows:
• Chilled water flow rate (m• ) c
• Damper position of the zones (P , P , P )
1 2 3
step. General idea of the MPC is shown in Fig. 4. In this section we
consider a single-input single-output system, in which y(k) is the output
value at kth time step and s(k) is the set-point trajectory that system's
output (y(k)) is supposed to track. These two trajectories are familiar in
the control system's context. However there's another trajectory known
as reference trajectory (r(k + i|k)) which is defined as the best course
for the current output (y(k)) to follow and get to the set-point trajec-
tory. Reference trajectory plays an important role in the MPC design. It
is commonly assumed that r closes in on the set-point trajectory with a
exponential equation. This process is carried out by a time constant
which determines the speed of error reduction.
Now assume that tracking error at current time step (er(k)) is given
by Eq. (4):
er(k) = s(k) – y(k) (4)

• Air handling unit's status (S ) ahu

Then reference trajectory is chosen so that the error in ith step
• Ambient temperature (T ) a
ahead is as below:
• Fresh air damper position (P ) fa

• Previous temperatures of the zones (T (k) … T z z (k-q+1)) where er( k+ i) = e−iTs/Tref er(k) (5)
q=2
In which Tref is the sampling time and Ts is the time constant that
Other factors such as return and discharge air flow rate, radiation determines the speed of error reduction. So the reference trajectory is
and mixing chamber's output temperature also affect the outputs. But given by Eq. (6).
there were many complications and challenges for setting up suitable
r( k+ i|k) = s( k+ i)− er( k+ i) = s( k+ i)−e−iTs/Tref er(k) (6)
sensors to measure some of the parameters such as return and discharge
air flow rate. Furthermore, many of the variables in this context aren't
In which r(k + i|k) is the value of the reference trajectory at i time
independent and they could be determined if the value of the variables
steps ahead, if system's output is y(k) at kth time step. So the control
in the list are available. Variables such as mixing chamber's output
signals are chosen so that the output trajectory follows the reference
temperatures are placed in this category.
trajectory to reduce the tracking error exponentially [5].
Fig. 3 shows the generic schematic of the optimum designed NARX
As explained earlier, MPC is a method that uses an internal model to
model in this study. One hidden layer of neurons has been considered to
predict the future outputs throughout a time horizon. This time hor-
capture and learn the system's dynamics. The number of hidden layer
izon, for which the future outputs are computed for, is called prediction
neurons and the number of delayed time feedbacks are to be de-
horizon (Hp). Main idea of the MPC is to find the control signal that
termined. The results based on which the optimal structure of the MLP
would lead to the best predicted performance. MPC also utilizes an
has been determined, are presented in section 6. In this study, output is
optimizer to generate the optimal control signals to track a reference
controlled via manipulating the value of the first three variables men-
trajectory or improve performance criteria [33]. In this paper we face a
tioned in the list above. So the variables used to control the AHU are
NMPC problem for which a NARX model (which was described in the
listed in Eq. (3), while two other parameters Pfa and Ta are sample data
preceding section) is used as the internal model of the NMPC and to find
from ambient temperature and fresh air damper position.
the optimal control signals at each time step genetic algorithm has been
→ used. In section 5 the NMPC applied in this study is presented along
C = [ṁ c,P1,P2,P3,Sahu] (3)
with the other important parameters in a NMPC.
As mentioned before, AHU distributes its output air among 3 zones
which are distinguished by their respective indices.

4. Model predictive control method

Model-based predictive control is a rather new control method

which first emerged in chemical processes at late seventies. MPC is not
a certain algorithm. It is just a method and various control strategies
could be used in a MPC scheme. MPC tries to find the most suitable
control signals in order to minimize an objective function subject to
some constraints, using a model of the system. If this model is non-
linear, the MPC problem becomes a non-linear one (NMPC). This pro-
cess of optimizing the objective function is repeated at each time step
and at each time step the optimal control signals are calculated
throughout a time horizon so the objective function is minimized
through this horizon. This time horizon, for which the control signals
are computed for, is called control horizon (Hc). The objective function
could take the form of tracking error, cost of energy, energy con-
sumption, environmental effects and other criteria or a combination of
them.
The most common method in MPC literature is the “receding horizon
method”, in this method only the first element of the control signal
vector is applied to the system (which corresponds to the control signal
Fig. 3. Generic schematic of the optimum NARX structure for modeling the AHU system.
of the current time step(k)) and the process is repeated in the next time

240
A. Erfani et al. Building and Environment 133 (2018) 237–245

Fig. 7. Results of the optimal NARX model (zone 2 results).

Fig. 4. General idea of the model predictive control [5].

Fig. 8. Zone 1 temperature results from NMPC implementation.

Fig. 9. Zone 2 temperature results from NMPC implementation.

Fig. 5. Whole control schematic and connection between NARX model and GA optimizer
where C = [mc, P1, P2, P3, Sahu] and X = [C,Ta,Tz,Ts] and Y = [T1, T2, T3].

Table 1
Results of simulating NARX model with different q values.

Number of feedbacks (q) Training patterns MREP Test pattern MREP

1 98.66 97.47
2 98.65 97.71
3 98.56 97.60

Fig. 10. Zone 3 temperature results from NMPC implementation.

Fig. 6. Error of learning pattern for different number of hidden layer neurons in NARX
Fig. 11. Optimal AHU status generated by the NMPC.
structure.

241
A. Erfani et al. Building and Environment 133 (2018) 237–245

5. Applying NMPC to the AHU system

5.1. Determinations of important NMPC factors

A centralized controller configuration is selected for this work, since

Fig. 12. Optimal cold water flow rate generated by the NMPC.
Fig. 13. Optimal zone 1 damper position generated by the NMPC.
it considers the coupling effect between different outputs (temperature
of different zones). Energy conservation strategies used in this study are
natural thermal storage in building mass and reference temperature
reset during unoccupied hours [4]. Temperature set-points are different
based on the time of day and whether or not that zone is occupied
during that time. During the office hours reference temperature for
zones is set to 25.5 °C which is in the range specified in Ref. [38] and
during the non-office hours it is set to 27 °C.
Three of the crucial design parameter in NMPC, are the value of Hc,
Hp and the time step. First of all, time step should be set. The faster the
system's dynamic the shorter the selected time step should be. In the
HVAC control literature, our system's dynamic is rather slow so a high
sampling rate is not mandatory. Furthermore, choosing a very low
sampling rate (long time step) may cause some problems in the iden-
tification process so that the model cannot capture all the system's
dynamics properly. In this work a 17 min time step is chosen. It is also
common to choose the Hp equal to Hc [5]. As for the control horizon,
the longer the better. But since we are going to implement the designed
NMPC on-line, a compromise between the computing time at each time
step and the efficiency of NMPC is absolutely necessary. Therefore,
Control horizon is set to 7 steps (119 min) in this research, so that the
controller could have enough time to adjust to disturbances and
changes in reference temperatures. Prediction horizon is also set to 7
steps in this context.

5.2. NMPC problem

As stated earlier, NMPC solves an optimization problem to find the

Fig. 14. Optimal zone 2 damper position generated by the NMPC.
best control signal during each time step. Since the plant's model is non-
linear, the optimization problem is most likely a non-convex one.
Solving a non-convex problem is usually a sophisticated task, especially
if one is using analytical methods.
Since we are going to solve an optimization problem, constraints
and objective function must be defined first. The more constraints we
have, the more possibility of facing an infeasibility situation [33].
Considering the on-line implementation of the controller and the many
issues that a high probability of infeasibility may bring with itself, it is
decided to apply as few constraints as possible to the optimization
problem. The only constraint applied to the optimization problem is
regarded to the temperature bounds in the system mentioned in Ref.
[38] (See Eq. (7)).

Tmin < Tz (k) < Tmax (7)

Tmin = 22 °C, Tmax = 28 °C.

There are also the restrictions of the damper positions changing
rates, which in this case they are negligible compared to the time step
and no constraint concerning them is added to the optimization pro-
blem.
Fig. 15. Optimal zone 3 damper position generated by the NMPC. Before the objective function and its solution are elaborated, it is
integral to describe the control signals. It should be noted that since a
Table 2 continuous GA is much more time consuming, a discrete GA is used in
Average temperature of the zones during NMPC and no-controller situation during the this work. So that the decision variables are defined as below:
office hours.

⎧ ṁ c = 5k k=0,1,2,…20
zone NMPC mean temperature Previous system mean temperature
Sahu = {1,2}
⎨
Zone 1 25.758 23.457 ⎩ P1 = 10k , P2 = 10k , P3 = 10k k=0,1,2,…10 (8)
Zone 2 25.763 23.643
Zone 3 25.67 23.313
• =0, 5, 10 …,
Where chilled water flow rate is divided into 20 steps (m c
100) which zero means that the coil is empty and 100 represents the full
capacity of the cooling coil. Sahu indicated the AHU status. The corre-
sponding values for on and off modes are 2 and 1 respectively. The

242
A. Erfani et al.
position of dampers is divided into 10 equal steps (Pz = 0,10 …,100).
Zero means that the air that is going to be blown into the corresponding
zone, passes through the cooling coil chamber whole, while 100 means
that none of air volume is chilled by the cooling water.
In order to set the objective function, performance criteria and en-
ergy saving terms are considered. Energy carriers which are consumed
by the AHU system are gas, which is used to chill the water coming out
of the cooling coil and electricity, which is used to derive the blowers,
motor and fans and water pump. Performance criteria is defined as the
square of deviation from reference trajectory for each zone (Eq. (10)).
So the objective function is given by Equations (9)-(13).
HC
min ⎛∑ N1 + N2−N3 + N4⎞
⎝ i=1 ⎠ (9)
3
N1 = ∑ w1 [T̂z ( k+ i|k)−rz ( k+ i|k)]2
z= 1 (10)
• ( k+ i−1)
N2 = w2 m c (11)
N3 = w3 (P1 ( k+ i−1|k) + P2 ( k+ i−1|k) + P3 ( k+ i−1|k)) (12)
N4 = q.l.(Sahu ( k+ i−1)−1) (13)
In Eq. (9), Hc is the control horizon, T̂z ( k+ i|k) is the model's pre-
diction of the future zone temperatures for i time steps later. rz is the
reference trajectory which is defined in Eq. (6) in which the value for
Ts/Tref is set to 0.3333. Hence, N2 states the chilled water flow rate; N3
represents the air flow rate that passes next to the cooling coil. The less
air passing from the cooling coil chamber, the less the chilled water is
heated and the less gas needed to chill it. AHU's electricity consumption
is denoted here by N4. q expresses the electricity consumption of the
studied AHU per hour (KW) and l is the time step (17 min). w1,w2 and
w3 are the weights that specify the importance of each term in the
objective function. The value of these weights has been determined via
a trial and error during the simulation process so that there is a rational
trade-off between the terms (Eq. (14)).
w1 = 200, w2 = 1.8, w3 = 0.3 (14)
GA has been applied to find the optimal solution of Eq. (9) at each
time step using GA MATLAB function. Seven steps ahead are considered
as the control horizon and at each step there are five control signals
(m• ,P ,P ,P ,S ) to be determined. Therefore, there are 35 decision
c 1 2 3 ahu
variables on aggregate that should be determined at each time step. In
order to solve the optimization problem at each time step, GA uses the
NARX model of the system (Fig. 5) to find the optimum control signals
throughout the control horizon.
At each time step, temperature sensors measure zone temperature Tz
and ambient temperatures Ta and flow meters measure water's flow rate
m• , these measured values along with the damper positions P , AHU
c z
status Sahu and set-point temperature Ts serve as the input for the NMPC
→
( X ). These values are sent to a pre-defined web address so that the
NMPC could read them. At this point NMPC starts the search to find the
optimal control signals. First, GA generates the initial population ran-
domly and then sends the value for the control signals to the NARX C(k
+1|k) and then NARX computes the future outputs throughout the Hc
based on the C vector. At this point, GA has everything to calculate the
value of the objective function (see Eq. (9)). After that the objective
function for each chromosome is calculated, they are sorted based on
the fitness function value and then the best chromosomes are used for
reproduction of the next generation. The same process is repeated for
the next generation until the convergence criterion is met and the op-
timal control signals are found. But in this case, we cannot wait until
the best answer is found by GA because the controller is to be im-
plemented on-line and there is a crucial time limit on the on-line
computation time for finding the optimal control signals, since the
control signals (C(k)) are ideally to be applied to the AHU precisely at
243
Building and Environment 133 (2018) 237–245
the beginning of the kth time step. Therefore, a time limit of 2 min is
considered for GA (which is ignorable compared to the time step) so
that after 2 min GA stops and the best answer is immediately applied to
the AHU as the control signal of the current step C(k). Values of the
control signal is determined so that the zones' temperature in the next
step Tz(k+1) would track the set-point temperature Ts as good as
possible while minimizing the energy consumption (see Eq. (9)). In the
next time step the whole process is repeated from measuring the zones'
→
temperature to finding the optimal control signal C(k+1) is found. Y is
the AHU output which consists of the zones' temperatures.
→
Y (k) = [T1 (k),T2 (k),T3 (k)] (15)
→ →
X = [ C ,Ta,Tz,Ts] (16)
6. Result and discussion
6.1. NARX model results
Training data was collected from July 13 to July 17, 2017 and test
data was gathered through 2 days: July 30 and July 31, 2017. 527
patterns were collected and saved, which 26.75% or in other terms 141
of these patterns were used for testing the neural network and the re-
maining 386 patterns were used for training the NARX. MATLAB code
for NARX was developed. To evaluate the performance of the NARX
model, Mean of Relative Error Percentage (MREP) is considered as the
accuracy criterion of the results (Eq. (17)).
Ptest
|di−ui | 100
MREP = ∑ di
*
Ptest
i= 1 (17)
di and ui are the desired output and NARX output respectively. Ptest is
the number of test patterns. In order to have the best performance, the
optimal value for the design parameters of the NARX should be ob-
tained. First the number of feedback time steps from the NARX outputs
(Fig. (3)) should be optimized. In order to do so, NARX with different q
values (Eq. (1)) was examined and the results are presented in Table 1.
As it could be concluded from Table 1, the optimal number of
output delays is q = 2. Since it is clear that the zones' temperature are
only a function of current values of the X vector parameters, thus the
best value for the p is one. Now to determine the number of hidden
layer neurons, the number of hidden layer neurons has been swept from
1 to 50 and the result are presented in Fig. 6.
Fig. 6 shows that for more than 30 neurons in the hidden layer, no
significant change in the performance of the model is detected and
model's error has very small fluctuations. Selecting a big value for the
number of hidden layer neurons, leads to a better model but it also
increases the computation time, so a compromise between the two
factors is essential. The number of hidden layer neurons in this work
has been set to 35. Thus the optimum NARX structure would be 13-35-
3. Model's prediction results for the test patterns are presented in Fig. 7.
The optimal model's accuracy is 97.71%.
6.2. NMPC implementation results
Design parameters of the NMPC have been determined in section
3.2. In this section, first the results obtained from on-line im-
plementation of NMPC are represented and then we discuss and eval-
uate these results. NMPC was run on a desktop computer with a dual
core CPU and 2.4 GHz clock frequency. To provide the ambient tem-
perature predictions throughout the controller implementation,
weather predictions presented in Ref. [39] are utilized. NMPC was
tested on August 21 and August 22, 2017 on the AHU. Figs. 8–10 show
the temperature of the three aforementioned zones during the im-
plementation of the NMPC.
Figs. 8–10 indicate that the suggested NMPC was successful in
A. Erfani et al.
following the reference temperature assigned to each zone. There are
three important points that could be derived from these pictures. The
first point is that during the night hours, controller tracked the set-point
temperature by natural ventilation in the building, since the ambient
temperature dropped at night. The second point is that in Figs. 9 and 10
the controller couldn't provide the set-point temperature during night
and afternoon hours because the ambient temperature was below 27 °C
and the cooling system wasn't capable of increasing the room's tem-
perature. The last and probably the most important point in these fig-
ures is that 2 h before the set-point temperature changes, zones' tem-
perature deviation from the set-point increases. This phenomenon
occurred because the system's thermal dynamic is slow which is a
consequence of the strong thermal inertia. Therefore, it takes time for
the room temperature to change. So tracking the future set-point starts
2 h (control horizon) before the change takes place.
The optimal control signals applied to the AHU during the test are
shown in Figs. 11–15. It should be mentioned that in Fig. 12, the flow
rate percentage is directly related to the actual flow rate.
It should be noted that before the NMPC was applied to the system,
the AHU was equipped with no controller. The air handling unit worked
in full operation mode during the weekdays which meant the cooling
coil flow rate was set to 100 and all the damper position were set to
zero and the AHU was constantly operating as well.
Figs. 11–15 imply that most part of the energy saving potential in
the system was related to the non-office hours, during which zones were
cooled, while there was no occupant in them and a part of the energy
saving potential was achieved through the strong thermal inertia of the
building. During the night hours, mostly small values have been as-
signed to the cooling coil flow rate and the NMPC often orders the AHU
to turn off. Also, the air volume passed through the cooling coil
chamber was usually a low percentage of whole, which led to low gas
consumption.
To quantify the energy conservation potential achieved by the
NMPC, energy consumption is compared using two methods. In the first
method, to compare the electricity consumption we have assumed that
the electrical power consumed by the AHU is constant. Based on this
consumption to find out how much electricity consumption has been
decreased, the total time that the AHU was off is calculated according to
the signals applied to the plant. Using this simple method, it was found
out that NMPC yields 55.1% reduction in electricity consumption.
In the second method, to calculate the gas savings achieved by the
proposed NMPC, the amount of thermal energy lost by the chilled water
after the heat exchanging process with air is measured by means of two
thermometers which measured the cooling water temperature before
and after the heat exchange with air, and a flow meter to measure the
chilled water's flow rate. Efficiency of the absorption chiller was as-
sumed to be constant throughout the test. It was found out that the
proposed NMPC yields 43.7% gas savings for the AHU.
To prove the efficiency of the proposed controller, NMPC was
compared with simulation results of a simple on/off controller. In this
scenario, AHU unit was turned on 3 h before the start of the office hours
and was turned off at the end of the office hours. The simulation results
showed 36.6% and 26.3% reduction in electricity and gas consumption
respectively. By comparing these results to the NMPC, it is concluded
that the proposed NMPC outperforms a scheduled controller. NMPC
leads to 18.5% and 17.4% more saving in electricity and gas energy
carriers respectively in comparison to the on/off controller.
In Table 2 the average zones' temperature during the office hours
are presented for the NMPC and the previous system. As it is presented
in Table 2, NMPC was able to tackle the problem of achieving energy
savings in a building's AHU system while maintaining the thermal
comfort of the occupants.
7. Conclusion
In this paper a practical non-linear model predictive controller for
244
Building and Environment 133 (2018) 237–245
controlling a multi-zone air handing unit is presented. The proposed
NMPC aims to minimize gas and electricity consumption of the AHU
while maintaining the thermal comfort of the building's occupants. The
control signals manipulated to alter the zone temperature in this study
are cooling coil's flow rate, damper position of each zone and AHU
status (on/off). To predict the future outputs of the plant (which is the
temperature of each zone) a recurrent neural network NARX model has
been developed in MATLAB environment. To generate the optimal
value of the control signals in each time step, genetic algorithm has
been utilized.
To investigate and demonstrate the effectiveness of the proposed
control strategy, designed NMPC was implemented and applied to the
studied AHU. To store and send data, a web-based approach has been
used in this work. The experimental results of the NMPC demonstrate
the effectiveness of this approach toward controlling a building's
thermal behavior. NMPC tracked the set-point temperature very well
and achieved significant savings in gas and electricity consumption si-
multaneously. The results indicate a 55.1% and 43.7% reduction in
electricity and gas consumption respectively, compared to the previous
setup.
A future extension of this work is considering the effect of occu-
pants' presence in the NMPC framework. Also considering the effect of
other factors such as relative humidity in the definition of thermal
comfort could be a new challenge in this subject. Another issue that
could be addressed in the future is considering the importance of oc-
cupant hours in controller design, such as weighting different hours of a
day in the objective function. It also should be noted that designing a
time-varying model and integrating it into the NMPC framework could
be extremely instrumental for the NMPC and would improve its ro-
bustness against various uncertainties.
This is our first step to implement the nonlinear controller in AHU
systems. One could peruse this job with more contemporary controllers
such as: fuzzy, intelligent, and sliding model … controllers. In addition,
to improve the NARX model performance, capturing more weather
variability is recommended. This point can simply be over by extending
the data logging and train periods, which leads to improve the results of
the MPC controller.
References
[1] F. Oldewurtel, A. Parisio, C.N. Jones, M. Morari, D. Gyalistras, M. Gwerder,
V. Stauch, B. Lehmann, K. Wirth, Energy efficient building climate control using
Stochastic Model Predictive Control and weather predictions, Am. Control Conf.
(ACC) 2010 (2010) 5100–5105, http://dx.doi.org/10.1109/ACC.2010.5530680.
[2] M. Avci, M. Erkoc, A. Rahmani, S. Asfour, Model predictive HVAC load control in
buildings using real-time electricity pricing, Energy Build. 60 (2013) 199–209,
http://dx.doi.org/10.1016/j.enbuild.2013.01.008.
[3] L. Pérez-Lombard, J. Ortiz, C. Pout, A review on buildings energy consumption
information, Energy Build. 40 (2008) 394–398, http://dx.doi.org/10.1016/j.
enbuild.2007.03.007.
[4] A. Afram, F. Janabi-Sharifi, Theory and applications of HVAC control systems - a
review of model predictive control (MPC), Build. Environ. 72 (2014) 343–355,
http://dx.doi.org/10.1016/j.buildenv.2013.11.016.
[5] J.M. Maciejowski, Predictive Control with Constraints, first ed., Prentice hall, 2002,
https://www.amazon.com/Predictive-Control-Constraints-Jan-Maciejowski/dp/
0201398230.
[6] W. Liang, R. Quinte, X. Jia, J.Q. Sun, MPC control for improving energy efficiency
of a building air handler for multi-zone VAVs, Build. Environ. 92 (2015) 256–268,
http://dx.doi.org/10.1016/j.buildenv.2015.04.033.
[7] A.I. Dounis, C. Caraiscos, Advanced control systems engineering for energy and
comfort management in a building environment-A review, Renew. Sustain. Energy
Rev. 13 (2009) 1246–1261, http://dx.doi.org/10.1016/j.rser.2008.09.015.
[8] A.E. Ben-Nakhi, M.A. Mahmoud, Cooling load prediction for buildings using general
regression neural networks, Energy Convers. Manag. 45 (2004) 2127–2141, http://
dx.doi.org/10.1016/j.enconman.2003.10.009.
[9] C. Chang, J. Zhao, N. Zhu, Energy saving effect prediction and post evaluation of
air-conditioning system in public buildings, Energy Build. 43 (2011) 3243–3249,
http://dx.doi.org/10.1016/j.enbuild.2011.08.025.
[10] S.W. Wang, Y.M. Chen, Fault-tolerant control for outdoor ventilation air flow rate in
buildings based on neural network, Build. Environ. 37 (2002) 691–704, http://dx.
doi.org/10.1016/S0360-1323(01)00076-2 Pii S0360-1323(01)00076-2\r.
[11] A.E. Ben-nakhi, M. a Mahmoud, Energy conservation in buildings through efficient
A/C control using neural networks, Appl. Energy 73 (2002) 5–23.
A. Erfani et al.
[12] S. Soyguder, Intelligent system based on wavelet decomposition and neural network
for predicting of fan speed for energy saving in HVAC system, Energy Build. 43
(2011) 814–822, http://dx.doi.org/10.1016/j.enbuild.2010.12.001.
[13] S. Kalogirou, Artificial neural networks for the prediction of the energy consump-
tion of a passive solar building, Energy 25 (2000) 479–491, http://dx.doi.org/10.
1016/S0360-5442(99)00086-9.
[14] T. Olofsson, S. Andersson, R. Östin, A method for predicting the annual building
heating demand based on limited performance data, Energy Build. 28 (1998)
101–108, http://dx.doi.org/10.1016/S0378-7788(98)00004-8.
[15] J.F. Kreider, D.E. Claridge, P. Curtiss, R. Dodier, J.S. Haberl, M. Krarti, Building
energy use prediction and system identification using recurrent neural networks, J.
Sol. Energy Eng. 117 (1995) 161, http://dx.doi.org/10.1115/1.2847757.
[16] J. Yang, H. Rivard, R. Zmeureanu, On-line building energy prediction using adap-
tive artificial neural networks, Energy Build. 37 (2005) 1250–1259, http://dx.doi.
org/10.1016/j.enbuild.2005.02.005.
[17] P.A. González, J.M. Zamarreño, Prediction of hourly energy consumption in
buildings based on a feedback artificial neural network, Energy Build. 37 (2005)
595–601, http://dx.doi.org/10.1016/j.enbuild.2004.09.006.
[18] R. Yokoyama, T. Wakui, R. Satake, Prediction of energy demands using neural
network with model identification by global optimization, Energy Convers. Manag.
50 (2009) 319–327, http://dx.doi.org/10.1016/j.enconman.2008.09.017.
[19] B.B. Ekici, U.T. Aksoy, Prediction of building energy consumption by using artificial
neural networks, Adv. Eng. Software 40 (2009) 356–362, http://dx.doi.org/10.
1016/j.advengsoft.2008.05.003.
[20] L. Giraud-Moreau, P. Lafon, A Comparison of evolutionary algorithms for me-
chanical design components, Eng. Optim. 34 (2002) 307–322, http://dx.doi.org/
10.1080/03052150211750.
[21] Y. Asiedu, R. Besant, P. Gu, HVAC duct system design using genetic algorithms,
HVAC R Res. 6 (2000) 149–173, http://dx.doi.org/10.1080/10789669.2000.
10391255.
[22] J.A. Wright, HVAC optimisation studies: sizing by genetic algorithm, Build. Serv.
Eng. Technol. 17 (1996) 7–14 http://bse.sagepub.com/content/17/1/7.abstract.
[23] L. Lu, W. Cai, Y.C. Soh, L. Xie, S. Li, HVAC system optimization - condenser water
loop, Energy Convers. Manag. 45 (2004) 613–630, http://dx.doi.org/10.1016/
S0196-8904(03)00181-X.
[24] W. Huang, H.N. Lam, Using genetic algorithms to optimize controller parameters
for HVAC systems, Energy Build. 26 (1997) 277–282, http://dx.doi.org/10.1016/
245
Building and Environment 133 (2018) 237–245
S0378-7788(97)00008-X.
[25] H.X. Zhao, F. Magoules, A review on the prediction of building energy consumption,
Renew. Sustain. Energy Rev. 16 (2012) 3586–3592, http://dx.doi.org/10.1016/j.
rser.2012.02.049.
[26] G. Huang, Model predictive control of VAV zone thermal systems concerning bi-
linearity and gain nonlinearity, Contr. Eng. Pract. 19 (2011) 700–710, http://dx.
doi.org/10.1016/j.conengprac.2011.03.005.
[27] P.-D. Moroşan, R. Bourdais, D. Dumur, J. Buisson, Building temperature regulation
using a distributed model predictive control, Energy Build. 42 (2010) 1445–1452,
http://dx.doi.org/10.1016/j.enbuild.2010.03.014.
[28] S. Privara, J. Siroky, L. Ferkl, J. Cigler, Model predictive control of a building
heating system: the first experience, Energy Build. 43 (2011) 564–572, http://dx.
doi.org/10.1016/j.enbuild.2010.10.022.
[29] P.M. Ferreira, A.E. Ruano, S. Silva, E.Z.E. Conceição, Neural networks based pre-
dictive control for thermal comfort and energy savings in public buildings, Energy
Build. 55 (2012) 238–251, http://dx.doi.org/10.1016/j.enbuild.2012.08.002.
[30] J. Siroky, F. Oldewurtel, J. Cigler, S. Privara, Experimental analysis of model pre-
dictive control for an energy efficient building heating system, Appl. Energy 88
(2011) 3079–3087, http://dx.doi.org/10.1016/j.apenergy.2011.03.009.
[31] J. Rehrl, M. Horn, Temperature control for HVAC systems based on exact linear-
ization and model predictive control, 2011 IEEE Int. Conf. Control Appl. CCA 2011
(2011) 1119–1124, http://dx.doi.org/10.1109/CCA.2011.6044437.
[32] K.F. Fong, V.I. Hanby, T.T. Chow, System optimization for HVAC energy manage-
ment using the robust evolutionary algorithm, Appl. Therm. Eng. 29 (2009)
2327–2334, http://dx.doi.org/10.1016/j.applthermaleng.2008.11.019.
[33] F. Allgower, A. Zheng, Nonlinear Model Predictive Control, springer, 2000.
[34] R. Beale, T. Jackson, A. Hilger, Neural Computing : an Introduction, (1990).
[35] T. Lin, B. Horne, P. Tino, C.L. Giles, Learning long-term dependencies in NARX
recurrent neural networks, IEEE Trans. Neural Network. 7 (1996) 1329–1351.
[36] M. Boroushaki, M.B. Ghofrani, C. Lucas, Identification of a nuclear reactor core (
VVER ) using recurrent neural networks, Ann. Nucl. Energy 29 (2002) 1225–1240.
[37] S. Haykin, Neural Networks and Learning Machines, third, Pearson, 2009.
[38] Thermal Environmental Conditions for Human Occupancy, Cond. Am. Soc. Heat.
Refrig. Air, (2003).
[39] Iran weather data, available at: https://www.accuweather.com/en/ir/iran-
weather.