Energy 42 (2012) 241e250

Contents lists available at SciVerse ScienceDirect

Energy
journal homepage: www.elsevier.com/locate/energy

Modeling and optimization of HVAC systems using a dynamic neural network

Andrew Kusiak*, Guanglin Xu
Department of Mechanical and Industrial Engineering, 3131 Seamans Center, The University of Iowa, Iowa City, IA 52242-1527, USA

a r t i c l e i n f o a b s t r a c t

Article history: The energy consumption of a heating, ventilating and air conditioning (HVAC) system is optimized by
Received 24 November 2011 using a data-driven approach. Predictive models with controllable and uncontrollable input and output
Received in revised form variables utilize the concept of a dynamic neural network. The minimization of the energy consumed
22 February 2012
while maintaining indoor room temperature at an acceptable level is accomplished with a bi-objective
Accepted 26 March 2012
optimization. The model is solved with three variants of the multi-objective particle swarm optimiza-
Available online 27 April 2012
tion algorithm. The optimization model and the multi-objective algorithm have been implemented in an
existing HVAC system. The test results performed in the existing environment demonstrate signicant
Keywords:
HVAC
improvement of the system. Compared to the traditional control strategy, the proposed model saved up
Data-driven modeling to 30% of energy.
Dynamic neutral network
2012 Elsevier Ltd. All rights reserved.
Multi-objective particle swarm optimization
Non-linear model

1. Introduction
The past decades have seen signicant growth in the energy
consumed by heating, ventilating, and air conditioning (HVAC)
systems [1]. Design of energy-saving control strategies for HVAC
systems has gained attention. An effective HVAC system control
strategy adjusts controllable variables for effective energy use. A
schematic diagram of a typical HVAC system is shown in Fig. 1. The
system includes two parts, an air handling unit (AHU) and variable
air volume (VAV) terminals.
Numerous papers have reported contributions to the optimization
of HVAC systems with the aim to improve energy efciency. Some
studies have focused on development of analytical models using
physics. Teodosiu et al. [2] proposed an analytical model to assess
thermal comfort by considering indoor air moisture and its transport
by the airow. Nassif et al. [3e5] discussed a supervisory control
strategy to optimize set points of controllers used in a multi-zone
HVAC system. Gebreslassie et al. [6] discussed design of an environ-
mentally conscious absorption cooling system. A multi-objective
model minimizing cost and environmental impact was developed.
The analytical models of HVAC systems are usually computa-
tionally complex and involve assumptions that may not hold in
practice. Unlike physics-based models, neural networks depend
entirely on experimental data which make them application
relevant and easy to compute [7]. Kusiak and Li [8, 9] and Kusiak
et al. [10] developed neural network models minimizing energy
consumption of an air handling unit and variable-air-volume box.
Spindler and Norford [11] presented a system-identication
framework for linear thermal models accommodating the unique
features of mixed-mode buildings. Ben-Nakhi and Mahmoud [12]
applied a general regression neural network to optimize air
conditioning setback scheduling in public buildings.
As HVAC systems exhibit dynamic and non-linear characteris-
tics, a non-linear autoregressive with external input (NARX) model
can capture its properties and states. Many studies have focused on
applying neural networks to identify and control dynamic systems
[13e16]. This paper proposes an optimization model derived by
a dynamic neural network based on the concept of a non-linear
autoregressive with external input (NARX). This model is then
optimized by three variants of the multi-objective particle swarm
optimization (MOPSO) algorithm. The optimization model and the
MOPSO algorithm have been implemented at an industrial test
laboratory setting. The results from the comparative experiment
are reported.
2. Solution methodology
2.1. Non-linear autoregressive with external input

The approach known as the non-linear autoregressive with

* Corresponding author. external input (NARX) is an important modeling of discrete non-
E-mail address: andrew-kusiak@uiowa.edu (A. Kusiak). linear systems. The NARX model is expressed in Eqs. (1)e(4).

0360-5442/$ e see front matter
2012 Elsevier Ltd. All rights reserved.
doi:10.1016/j.energy.2012.03.063
242 A. Kusiak, G. Xu / Energy 42 (2012) 241e250

prediction accuracy of the model, ^xt d with x(t), x(t  d), ., and
Notation x(t  nd) are considered as inputs. Therefore, in the modied
model, Eq. (5) is used in place of Eq. (2).
t current time

d time increment (30 min in this paper) ^t d f2 yt; yt  d; .; yt  n3 d; ^
y xt d;
 (5)
x the actual value of the system state xt; .; xt  n4 d; uy t; uy t  d; .; uy t  n5 d
y the actual value of the system output
^
xt d the predicted value of the system state
^t d
y the predicted value of the system output 2.2. Problem formulation
ux the vector of state parameters
uy the vector used of system output From the designers perspective, the HVAC system should satisfy
ETotal the total energy consumed by the HVAC system the peak load of a building. However, usually the system delivers
EElectricity the energy consumed in the form of electricity a partial load. It may operate at its full capacity for a limited time only.
EGas the energy consumed in the form of natural gas In this paper, the HVAC system consumed energy includes electricity
ESF the energy consumed by the supply fan and natural gas. The electricity is used to power a supply fan (ESF),
ERF the energy consumed by the return fan return fan (ERF), outside air injection fan (EOAF), chillers (ECHL), and
EOAF the energy consumed by the outside air fan pumps (EPump). Natural gas is used by the variable air volume boxes.
ECHL the energy consumed by the chiller Therefore, the total HVAC energy is as expressed in Eq. (6).
EPump the energy consumed by the pump
LPGPM water ow rate of the heating pump ETotal EElectricity EGas (6)
LSWT supply water temperature
where ETotal is the total energy consumption of the HVAC system,
LRWT return water temperature
EElectricity is the consumed electricity, and EGas is the energy
consumed in the natural gas form.
The electricity consumed EElectricity is expressed in Eq. (7).
^
xt d f1 xt; xt  d; .; xt  n4 d;
(1) EElectricity ESF ERF EOAF ECHL EPump (7)
ux t; ux t  d; .; ux t  n5 d
The energy consumed from the natural gas is shown in Eq. (8).


^t d f2 yt; yt  d; .; yt  n1 d;
y
 (2) QGas 500*LPGPM*LSWT  LRWT*T (8)
xt; .; xt  n2 d; uy t; uy t  d; .; uy t  n3 d
where 500 is the conversion factor(BTU min/gallon F
h) for pure
water used in the pump, while LPGPM is the water ow rate of
xt d ^
xt d ex t d (3)
a loop pump (gallon/min), LSWT is the supply water temperature
( F), and LRWT is the return water temperature ( F), T is the time
^t d ey t d
yt d y (4) span (h).
where x and y are the system outputs, ux the input of x, uy the input Since EElectricity unit is Watt-hour while QGas is measured in
of y, d the time delay, ^
xt d is the predicted value x at time t d, British thermal unit (BTU), Eq. (9) provides the BTU to Watt-hour
and y ^t d is the predicted value y at time t d. Parameters conversion in order to evaluate the total energy consumed by the
x(t d) and y(t d) are the actual system outputs, and ex(t d) and HVAC system in a uniform way.
ey(t d) are the corresponding errors at t d.
EGas 0:293*QGas (9)
In the NARX model, the output x at time t, t  d, ., t  nd are the
inputs to y ^t d. However, sometimes the value at time t d, where EGas is the energy consumption from natural gas evaluated in
^
xt d, signicantly impacts the output y(t d). To increase the Watt-hour (Wh).

Damper

Exhaust air Return air Exhaust grile

Damper

VAV terminal
Cooling coil Heating coil
Outside air
Mixed air Supply air
Diffuser
Damper

Valve Valve V-4 Valve

V-1 V-3

Fig. 1. Schematic diagram of a HVAC system.

 A. Kusiak, G. Xu / Energy 42 (2012) 241e250 243

To minimize the total energy consumption, a function should be Table 2

rst established between ETotal, the indoor temperature, and other Candidate parameters selected based on domain knowledge.

input parameters of the HVAC system (see Eq. (10)). Parameter type Parameter Description Unit
name


^t d f2 yt; yt  d; .; yt  n1 d; ^
y xt d; Optimized input SAT-SPT AHU supply air 
F ( C)
 (10)
xt; .; xt  n2 d; uy t; uy t  d; .; uy t  n3 d parameter temperature set point
SASP-SPT Supply air duct static in. WG (kPa)
A function between the indoor temperature and its previous pressure set point
Input parameter CHWC-VLV Chilled water coil valve % Open
values and the input parameters of the HVAC system is established position
(see Eq. (11)). SA-HUMD Supply air humidity % RH

MA-TEMP Mixed air temperature F ( C)
^
xt d f1 xt; xt  d; .; xt  n4 d; ux t; CHWC-EWT Chilled water coil entering 
F ( C)
(11) water temperature
ux t  d; .; ux t  n5 d SA-CFM Supply air fan speed CFM
RA-CFM Return air fan speed CFM

RM-TEMP Room temperature F ( C)

OA-TEMP Outside air temperature F ( C)
3. HVAC predictive model OA-HUMD Outside air humidity % RH
OA-CO2 Outside air CO2 PPM
concentration
3.1. Data description
IR-RADIA Infrared radiation B/HFt2
SOL-HORZ Solar normal ux B/HFt2
The data set used in this research has been collected from an SOL-BEAM Solar beam B/HFt2
experiment conducted at the Energy Resource Station (ERS) oper- BAR-PRES Barometric pressure mBar
ated by the Iowa Energy Center. The ERS is an energy laboratory for WIND-VEL Outside wind velocity MPH
WIND-DIR Outside wind direction Deg
testing and demonstration of commercial HVAC systems. It has two System state RM-HUMD Room humidity % RH
test areas, A and B, equipped with identical devices and four RM-TEMP Room temperature 
F ( C)
thermal zones. Each thermal zone has a variable air volume (VAV) System output ENERGY Energy consumed by the Wh
box connected to the corresponding air handling unit (AHU) to HVAC system
maintain the thermal comfort of the zone. Weather data has been
collected by sensors installed around the building. The control of
the HVAC system involves two set points, the AHU static pressure algorithm for ranking the importance of parameters for prediction.
set point (SASP-SPT) and the supply air temperature set point (SAT- According to [18,19], the boosting tree algorithm has demonstrated
SPT) that were adjusted in the data collection experiment to reect good performance in parameter selection and therefore it is used in
a range of the HVAC system states. In particular the SP-SPT set point this paper for selecting parameters. Furthermore, since the HVAC
was varied from 1.2 in. WG (0.3 kPa) to 1.8 in. WG (0.45 kPa) at system is dynamic, the system state and energy consumption are
0.2 in. WG (0.05 kPa) increments, while the SAT-SPT varied from signicantly inuenced by past values and the system parameters.
50  F (10  C) to 65  F (18.33  C) at 1  F (0.556  C) increment. The Correlation analysis has been used to determine the relationship
data on more than 300 parameters was recorded at 1-min sampling between the system state and energy consumption with their past
intervals. The experiment was conducted between July 31 to August values at different time steps. Figs. 2 and 3 illustrate the relation-
15, 2010, and September 21 to October 7, 2010. The data used in this ships between energy consumption and indoor room temperature
research includes three subsets. Data set 2 of 2065 instances was with their previous values. Fig. 4 shows the relationship between
used to train the predictive models, while data set 3 and data set 4, energy consumption and indoor room temperature at different
each with 442 instances, were used to test and validate the models. time steps. Correlation analysis has determined the relationship
These data sets are summarized in Table 1. Note that the data in between the system state and the energy consumption with the
Table 1 is expressed in 30-min intervals derived from the original 1- system inputs at different time steps. Fig. 5 shows the relationship
min data. between the energy consumption and the supply fan speed at
different time steps. By combining the correlation analysis with the
3.2. Parameter selection boosting tree algorithm, the parameters, listed in Tables 3 and 4, at
different states have been selected to build the energy consumption
Parameter selection is performed to eliminate parameters of less model and the indoor room temperature model.
importance to the model to enhance comprehensibility, scalability,
and often the accuracy of the resulting models [17]. Table 2 includes 3.3. Model construction
21 parameters that were selected as the candidates for building the
models discussed later in the paper. Boosting tree is a learning In this section, a multi-layer perception (MLP) ensemble algo-
rithm is used to build the predictive models for energy consump-
Table 1 tion and indoor room temperature as it outperforms other
Description of data sets.
algorithms including the chi-squared automatic interaction
Data Data set type Time period Number of detector (CHAID), classication and regression tree (C&RT) algo-
set no. instances rithm, support vector machine (SVM), multi-layer perception
1 Entire data set 07/31e08/15/2010 & 2949 (MLP), boosting tree, random forest, and multivariate adaptive
09/21e10/07/2010 regression spline (MARSpline) algorithms [20]. The models con-
2 Training data set Randomly selected from 2065
data set 1
structed by the MLP ensemble are expressed as follows:
3 Test data set Randomly selected from 442
data set 1
4 Validation data set Randomly selected from 442 ^
xt d f1 xt; xt  d; xt  2d; yt; vx1 t; vx2 t;
data set 1 (12)
vx3 t; vx4 t; vx5 t; vx6 t; vx7 t; c1 t d; c2 t d
244 A. Kusiak, G. Xu / Energy 42 (2012) 241e250

Fig. 2. The correlation coefcient between energy consumption and its previous
values at different time steps. (Note that d is the time step. It is 30 min in this research.)

Fig. 4. The correlation coefcient between energy consumption and room tempera-

ture at different time steps.
^t d f2 yt; yt  d; yt  2d; ^
y xt d; xt; xt  d;
vy1 t; vy2 t; vy3 t; vy4 t; vy5 t; vy5 t  d; vy5 t  2d; (13)  
 ~ 
vy6 t; vy6 t  d; vy6 t  2d; c1 t d; c2 t d y  y
APE (16)
where y ^t d is the energy consumption of the HVAC system (an y
objective function); c1(t d) and c2(t d) represent the supply air
Pn
temperature set point and the supply air duct static pressure set i1 APEi
MAPE (17)
point (decision variables). N
In order to obtain the concrete form of f1($) in Eq. (12) and f2($)
s
in Eq. (13), data set 1 (2949 data instances) was divided into three Pn 2
parts: training data set (2065 data instances), test data set (442 data i 1 AEi  MAE
Std AE (18)
instances), and validation data set (442 data instances) as shown in N1
Table 1. In the next sub-section, the performance of the predictive s
models will be validated. Pn 2
i 1 APEi  MAPE
Std APE (19)
3.4. Model validation N1
~ is the predicted
where AE in Eq. (14) represents the absolute error, y
To evaluate the performance of the predictive models built by value obtained from predictive models, y is the actual observed
the MLP ensemble algorithm, four metrics were used: the mean value, and N is the number of data instances that are used for
absolute error (MAE) (Eq. (15)), the standard deviation of absolute training, test, or validation. The data in Table 5 shows the perfor-
error (Std_AE) (Eq. (18)), the mean absolute percentage error mance of the predictive models built by the MPL ensemble algo-
(MAPE) (Eq. (17)) and the standard deviation of absolute rithm. Fig. 6 compares the predicted and observed values of the
percentage error (Std_APE) (Eq. (19)) [21]. In Eq. (15) AE represents energy consumption of HVAC systems on 442 30-min data
the absolute error of a single item, while in Eq. (16) represents the instances drawn from data set 3. The values obtained from the
absolute percentage error of a single item. energy consumption predictive model in Fig. 7 follow closely to the
  observed values. Fig. 7 compares the predicted and observed values
~ 
AE y  y (14) of the room temperature on the same data instances with the
energy consumption generated by the predictive model. The
Pn detailed measurement of the performance for the two models is
i1 AEi
MAE (15) summarized in Table 5. As shown in Table 5, the accuracy for the
N
energy consumption predictive model on validation data set is

Fig. 3. The correlation coefcient between room temperature and its previous values Fig. 5. The correlation coefcient between energy consumption and supply air fan
at different time steps. speed at different time steps.
 A. Kusiak, G. Xu / Energy 42 (2012) 241e250 245

Table 3 Table 5
The parameters selected for building the energy consumption model at time t d. Prediction accuracy of the MLP ensemble algorithm of AHU energy and temperature.

Parameter Parameter name Description Objective Data set MAE MAPE Std_AE Std_MAPE
y(t) Energy Energy consumption at time t Energy consumption Training 153.975 0.064 143.086 0.166
consumption Test 179.836 0.067 217.791 0.083
y(t  d) Energy Energy consumption at time t  d Validation 213.398 0.083 311.036 0.200
consumption Room temperature Training 0.269 0.004 0.263 0.004
y(t  2d) Energy Energy consumption at time t  2d Test 0.287 0.004 0.285 0.004
consumption Validation 0.308 0.004 0.312 0.004
^
xt d RM-TEMP Estimated room temperature at time t d
x(t) RM-TEMP Room temperature at time t
x(t  d) RM-TEMP Room temperature at time t  d
vy1(t) OA-HUMD Outside air humidity at time t
vy2(t) OA-TEMP Outside air temperature at time t
acceptable range in search of the optimal supply air temperature
vy3(t) SA-HUMD Supply air humidity at time t and the supply air duct static pressure set points at the time t d.
vy4(t) SOL-HORZ Solar normal ux at time t The observed room temperature can be calibrated by the sensors
vy5(t) RA-CFM Return air fan speed at time t originally installed in the system while the observed energy
vy5(t  d) RA-CFM Return air fan speed at time t  d
consumption can be computed from Eqs. (6)e(10). The constraints
vy5(t  2d) RA-CFM Return air fan speed at time t  2d
vy6(t) SA-CFM Supply air fan speed at time t in the model are identied by setting the lower and upper bounds
vy6(t  d) SA-CFM Supply air fan speed at time t  d of the control parameters and assigning an acceptable range for
vy6(t  2d) SA-CFM Supply air fan speed at time t  2d room temperature. These settings are restricted within the
c1(t d) SAT-SPT Supply air temperature set point at time t d following limits:
c2(t d) SASP-SPT Supply air duct static pressure set point
at time t d
 The supply air temperature set point varies from 50  F (10  C) to
64  F (17.7  C).
91.7%. For the temperature predictive model, the accuracy on the  The supply air duct static pressure set point varies between
validation data set is 99.6%. Therefore, the two predictive models 1.2 in. WG (0.3 kPa) to 1.8 in. WG (0.45 kPa).
are used to construct the overall optimization model.  The room temperature is maintained between 70.5  F
(21.39  C) and 71.5  F (21.94  C).
4. Model optimization
These three limits are determined by the design of the HVAC
4.1. Model formulation system and preferences of the occupants. Therefore, the optimi-
zation model is formulated as follows:
To minimize the energy consumption of the HVAC system, the
objective function ^
xt d f1 $ in Eq. (12) is minimized while
maintaining the value of y ^t d f2 $ in Eq. (13) within an

^t d
miny
c1 td;c2 td
subject to :
^
xt d f1 xt; xt
 d; xt  2d; yt; vx1 t; vx2 t; vx3 t; vx4 t; vx5 t; vx6 t; vx7 t; c1 t d; c2 t d
^t d f2 yt; yt  d; yt  2d; ^
y xt d; xt; xt  d; vy1 t; vy2 t; vy3 t; vy4 t; vy5 t; (20)
vy5 t  d; vy5 t  2d; vy6 t; vy6 t  d; vy6 t  2d; c1 t d; c2 t d
50  c1 t d  64
1:2  c2 t d  1:8
70:5  ^ xt d  71:5

where x*(t d) is the predicted value of indoor room temperature

Table 4
The parameters selected for building the indoor room temperature model at time by applying the original supply air temperature set point c1(t d)
t d. and the supply air duct static pressure set point c2(t d). In
minimizing the energy consumption at time stamp t d, the room
Parameter Point name Description
temperature is maintained within a pre-set range. The constrained
x(t) RM-TEMP Room temperature at time t
x(t  d) RM-TEMP Room temperature at time t  d
model (20) is transformed to a bi-objective optimization model
x(t  2d) RM-TEMP Room temperature at time t  2d with the objective functions (21) and (22).
y(t) Energy consumption Energy consumption at time t
vx1(t) RA-CFM Return air fan speed at time t ^t d
Obj1 y (21)
vx2(t) SA-CFM Supply air fan speed at time t
vx3(t) MA-TEMP Mixed air temperature at time t
vx4(t) SA-HUMD Supply air humidity at time t
vx5(t) OA-TEMP Outside air temperature at time t Obj2 maxf0; 70:5  ^
xt dg maxf0; ^
xt d  71:5g (22)
vx6(t) OA-CO2 Outside air CO2 concentration at time t
vx7(t) CHWC-VLV Chilled water coil valve position at time t
c1(t d) SAT-SPT Supply air temperature set point at time t d Then the bi-objective optimization model is presented as
c2(t d) SASP-SPT Supply air duct static pressure set point follows:
at time t d
246 A. Kusiak, G. Xu / Energy 42 (2012) 241e250

minObj1; Obj2
c1 td;c2 td
subject to :
^
xt d f1 xt; xt
 d; xt  2d; yt; vx1 t; vx2 t; vx3 t; vx4 t; vx5 t; vx6 t; vx7 t; c1 t d; c2 t d
^t d f2 yt; yt  d; yt  2d; ^ xt d; xt; xt  d; vy1 t; vy2 t; vy3 t; vy4 (23)
y  t; vy5 t;
vy5 t  d; vy5 t  2d; vy6 t; vy6 t  d; vy6 t  2d; c1 t d; c2 t d
50  c1 t d  64
1:2  c2 t d  1:8

Note that when Obj2 equals 0, all constraints of the model in Eq. vi )vi U0; 41 $pi  xi U0; 42 $pg  xi
(23) are satised. (24)
xi )xi vi
Step 5: If the stop criterion is satised, pg is the nal solution and
4.2. Multi-objective particle swarm optimization algorithm
gbest is the nal optimal tness.
Note that U(0,a) represents the uniform distribution in [0,a]; and
Since the optimization model derived from the MLP ensemble
vi should be within the range [vmax, vmax].
algorithm is complex and the gradients cannot be computed, it is
In order to get CIWPSO, CPSO, and DIWPSO, the concrete
difcult to solve it with traditional algorithms. Particle swarm
modications to the original PSO are listed in the following:
optimization (PSO) inspired by natural bird ocks is a suitable
optimization algorithm [22]. The PSO algorithm is easy to imple-
1) For constant inertia weight particle swarm optimization Eq.
ment and there are few parameters to adjust. In recent years, many
(24) is transformed into Eq. (25).
modications have been made to the original algorithm [23].
Among those modications, constant inertia weight particle swarm
optimization (CIWPSO), constricted particle swarm optimization vi )uvi U0; 41 $pi  xi U0; 42 $pg  xi
(25)
(CPSO), and decreasing inertia weight particle swarm optimization xi )xi vi
(DIWPSO) have good performance in most cases. Hence, these three
PSO variants were applied in this research. The PSO in its original where u is the inertia weight that can reduce the importance of
form [23] is presented next. vmax to satisfy the requirement of controlling the scope of the
The PSO Algorithm [23] search. For constant inertial particle swarm optimization u is
a constant.
Step 1: Initialize a group of particles with random positions
xi RD and velocities vi RD in the search space; 2) For constricted particle swarm optimization Eq. (24) is
expressed as Eq. (26).
Perform the next steps until the pre-set requirements are
satised.  
vi )c vi U0; 41 $pi  xi U0; 42 $pg  xi (26)
xi )xi vi
Step 2: For each particle, compute tness for each particle by
using function f($). where c is the constriction coefcient that can control the
Step 3: Compare each particle tness with its past best value, convergence of the particle and prevent explosion of the particles
pbesti. If current value is better than pbesti, then using current velocity. Usually this constriction coefcient is set to 0.7298.
value instead of pbesti and update pi RD with current location
xi; compare all of the particles pbesti and nd the best one 3) For decreasing inertia weight particle swarm optimization Eq.
assigned as gbest and set its current location as pg RD. (24) is expressed as Eq. (27).
Step 4: Update the particles velocities and positions based on
the following Eq. (24):

Fig. 6. Validation of the energy consumption model with 442 data instances. Fig. 7. Validation of the room temperature model with 442 data instances.
 A. Kusiak, G. Xu / Energy 42 (2012) 241e250 247



vt1
i
)ut vti U0; 41 $ pti  xti U0; 42 $ pti  xti
(27)
xt1
i
)xti vt1
i

In Eq. (27), ut is a time function. It is updated based on Eq. (28).

utmax  t
ut umax  umin umin (28)
utmax
To adopt the above three algorithms for solving a multi-
objective optimization model, the following modications are
made according to [24].

Modication 1: Create a set Si to store the non-dominated

solutions for ith particle up to the current time.
Modication 2: Create a set G to store the non-dominated
solutions from all Si at each iteration.
Modication 3: Create an external set E to store the non-
dominated solutions from G at each iteration.
Modication 4: Process to update Si: At each iteration, compare
the current particle solution with the stored solutions. Domi-
nated solutions are removed while non-dominated solutions are
kept.
Modication 5: Process to update G: At each iteration, copy all to
G where dominated solutions are removed.
Modication 6: Process to update external non-dominated set E:
At each iteration, copy solutions from G to E. Remove the
dominated solutions from E.
Modication 7: Process to generate local and global best solu-
tion: For each particle at each iteration, the Euclidean distance
among solutions from the corresponding local non-dominated
set and global non-dominated set are measured. The pair with
minimum distance in the search space is selected as the local
and global best for this particle in under-taking the later velocity
and position updating process.

In contrast to CIWPSO, DIWPSO, and CPSO that were designed

for solving single objective models, multi-objective CIWPSO (MO-
CIWPSO), multi-objective DIWPSO (MO-DIWPSO), and multi-
objective CPSO (MO-CPSO) are aimed at solving multi-objective
models. Fig. 8 shows the ow chart for the multi-objective
Fig. 8. The ow chart of multi-objective particle swarm optimization algorithm.
particle swarm optimization algorithm.

solutions. Scenario 2 implies that energy consumption and room

4.3. Optimization results and analysis
To solve the optimization model represented by (23), the above-
mentioned three multi-objective PSO variants are applied and the
detailed parameters for the algorithms are listed in Table 6.
Data set 4 in Table 1 was used to run the variants of the PSO
algorithm dened in Table 6. The optimal solution is selected from
the nal elite set based on the weighted normalized objective
function (29).
Obj1  Obj1min Obj2  Obj2min
Obj w1 w2 (29)
Obj1max  Obj1min Obj2max  Obj2min
where w1 and w2 are the user-dened weights indicating the
preference of the corresponding objective, and Obj1max and
Obj1min are the maximum and the minimum values of Obj1 in the
nal elite set. Similar notation is used for Obj2max and Obj2min. Note
that w1 w2 1, with w1 and w2 being either constants or func-
tions of other objectives. Table 7 presents two scenarios that
represent different preferences to the objectives. Scenario 1 implies
that energy consumption is the only criterion to be considered
when selecting the single best solution among the non-dominated
temperature are both important criteria when selecting the single
best solution among the non-dominated solutions. In order to
eliminate the bias of the algorithms, each of the three multi-
objective PSO variants was run ten times and the average values
and the corresponding standard deviation were calculated. The
results provided by the three variants of the PSO algorithm are
listed in Table 8. As shown in Table 8, the multi-objective DIWPSO
Table 6
Settings for the three multi-objective PSO variants.
Algorithm Settings
MO-CIWPSO Acceleration coefcients are set to 41 42 2.05.
Inertia weight is set to 0.95. Population size and the
number of iterations are set to 100 and 50, respectively.
MO-DIWPSO Acceleration coefcients are set to 41 42 2.05.
Linearly decreasing inertia weight from 0.9 to 0.4 and the
nal value is reach at the end of the run. Population size
and iteration number are set to 100 and 50, respectively.
MO-CPSO Acceleration coefcients 4 4.1. Constriction factor
c 0.729.
Population size and the number of iterations are set to
100 and 50, respectively.
248 A. Kusiak, G. Xu / Energy 42 (2012) 241e250

Table 7
Two scenarios involving different weight values.

Scenario Weights Description

1 w1 1, w2 0 No AQI constraints
(
1 if y2 70:5; 71:5
2 w1 ; Consider room temperature as
0:5 otherwise
( constraint
0 if y2 70:5; 71:5
w2
0:5 otherwise

Table 8
Performance of three multi-objective PSO variants.

Algorithm Energy savings in Energy savings in

Scenario 1 Scenario 2
MO-CIWPSO 22.77%  0.12% 21.83%  0.13% Fig. 11. Comparison between observed and optimized room temperature in Scenario 1.
MO-DIWPSO 23.00%  0.06% 22.04%  0.13%
MO-CPSO 22.86%  0.13% 21.89%  0.12%

outperforms other two multi-objective PSO variants, and therefore,

the DIWPSO is used in the following analysis. Figs. 9 and 10
compare the observed and the optimized values on data set 4 for
Scenario 1 and 2. In most cases, the energy consumption of the
optimized process is less than the observed one, which means that
the proposed model saves energy. Figs. 11 and 12 illustrate the
room temperature for Scenario 1 and 2. In most cases, the indoor
room temperature can be kept in the range from 68  F (20  C) to
72  F (22.22  C) (desired room temperature range). Since Scenario 2

Fig. 12. Comparison of the observed and optimized room temperature in Scenario 2.

considers the room temperature as a constraint when selecting the

single best solution among the non-dominated solutions, the
number of points violating the desired room temperature range is
less than in Scenario 1. Figs. 13 and 14 present the supply air
temperature set point and the supply air duct static pressure set
point. Based on these set points, the energy savings measured by
the difference between the observed and the optimized energy
consumption shown are shown in Figs. 9 and 10.
Fig. 9. Comparison between observed and optimized energy consumption in Scenario 1. The actual implementation and validation of the DIWPSO
algorithm for Scenario 2 is discussed in the next section.

Fig. 13. Recommended supply air temperature set point compared to the observed
Fig. 10. Comparison between observed and optimized energy consumption in Scenario 2. value in Scenario 1.
 A. Kusiak, G. Xu / Energy 42 (2012) 241e250
Fig. 14. Recommended supply air duct static pressure set point compared to the
observed value in Scenario 1.
Fig. 15. Energy consumption of AHU-A and AHU-B at the rst stage.
5. Model implementation and result analysis
To validate the energy savings produced by the model presented
in this paper, an experiment was designed at the Energy Resource
Station (ERS) of the Iowa Energy Center from May 15 to 22, 2011. In
the experiment, AHU-A and AHU-B were operated simultaneously.
The AHU-A and AHU-B serve four identical thermal areas. The only
difference between the two systems was that the AHU-A was
controlled by the proposed approach while the AHU-B was
controlled by the traditional approach. The experiment included
Fig. 16. Indoor room temperature of AHU-A and AHU-B at the rst stage.
249
Fig. 17. Comparison of energy consumption of system A and B for the same set points
at the second stage.
two stages. The rst stage from May 15 to 20, 2011 aimed at col-
lecting data from the two systems while the second stage from May
21 to 22, 2011 was to establish the energy consumption bias
between the two systems controlled by the traditional approach. At
the rst stage, AHU-A was controlled by the optimization model
while AHU-B operated with xed set points: the supply air
temperature set at 55  F (12.78  C) and the supply air duct static
pressure set at 1.4 in. WG (0.35 kPa). Since these two systems use
identical devices and they serve identical test areas, the difference
in energy consumption points to the effectiveness of the proposed
optimization approach. As shown in Fig. 15, the energy consump-
tion for AHU-A and AHU-B are 568.88 and 781.16 kWh, respectively.
AHU-A therefore consumed less energy than AHU-B, thus
producing an energy savings of 27.18%. Fig. 16 illustrates the room
temperature for AHU-A and AHU-B. Although the range of indoor
room temperature of AHU-A is larger than for AHU-B, it usually falls
in the normal range of 68  F (20  C) to 72  F (22.22  C). Only for
a limited time does it exceed the present constraints. Thus, the
indoor room temperature is considered to be at an acceptable level.
In the second stage, the two systems were operated with identical
xed set points: the supply air temperature was set at 55  F
(12.78  C) and the supply air duct static pressure was set at
1.4 in. WG (0.35 kPa). As shown in Fig. 17, the energy consumption
of AHU-A and AHU-B are 277.22 and 266.49 kWh, respectively.
AHU-A therefore consumes 3.87% more energy than AHU-B for the
same control settings. Considering for the bias between the two
systems, the energy savings provided by the optimization model
are adjusted to 29.99% Fig. 18 shows the adjusted energy compar-
ison after the adjustment.
Fig. 18. Comparison of the bias e adjusted energy consumption of system A and B.
250 A. Kusiak, G. Xu / Energy 42 (2012) 241e250
6. Conclusion
In this research, the energy consumption of a heating ventilation
and air conditioning (HVAC) system was optimized with a dynamic
neural network. A model was built with data mining algorithms to
optimize the controllable set points of the HVAC system in order to
provide energy savings while maintaining the room temperature
within an acceptable range. Three modied multi-objective particle
swarm optimization (MOPSO) algorithms were applied to solve the
optimization model. The computational results indicated that the
multi-objective decreasing inertia weight particle swarm optimi-
zation (MO-DIWPSO) outperformed two other variants of the same
algorithm. The MO-DIWPSO algorithm was implemented on the
actual HVAC system. The experiments demonstrated that the
optimization model saved 29.99% in energy consumption. Future
research will focus on improving the accuracy of the model and
approaches to improve handling of user preferences.
Acknowledgement
This research has been supported by the Iowa Energy Center,
Grant No. 08-01.
References
[1] Prez-Lombard L, Ortiz J, Pout C. A review on building energy consumption
information. Energy and Buildings 2008;40(3):394e8.
[2] Teodosiu C, Hohota R, Rusaouen G, Woloszyn M. Numerical prediction of
indoor air humidity and its effect on indoor environment. Building and
Environment 2003;38(5):655e64.
[3] Nassif N, Kajl S, Sabourin R. Evolutionary algorithms for multi-objective
optimization in HVAC system control strategy. Fuzzy Information 2004;1:
51e6.
[4] Nassif N, Kajl S, Sabourin R. Optimization of HVAC control system strategy
using two objective genetic algorithm. HVAC & Research 2005;11(3):459e86.
[5] Nassif N, Moujaes S. A cost-effective operating strategy to reduce energy
consumption in a HVAC system. International Journal of Energy Research
2008;Vol. 32:543e58.
[6] Gebreslassie BH, Guillen-Gosalbez G, Jimenez L, Boer D. Design of environ-
mentally conscious absorption cooling systems via multi-objective optimiza-
tion and life cycle assessment. Applied Energy 2009;86(9):1712e22.
[7] Lu T, Viljanen M. Prediction of indoor temperature and relative humidity using
neural network models: model comparison. Neural Computing & Application
2009;18(4):345e57.
[8] Kusiak A, Li M. Reheat optimization of the variable-air-volume box. Energy
2010;35(5):1997e2005.
[9] Kusiak A, Li M. Cooling output optimization of an air handling unit. Applied
Energy 2010;87(3):901e9.
[10] Kusiak A, Tang F, Xu G. Multi-objective optimization of HVAC system with an
evolutionary computation algorithm. Energy 2011;36(5):2440e9.
[11] Spindler HC, Norford LK. Naturally ventilated and mixed-mode buildings e
part I: thermal modeling. Building and Environment 2009;44(4):736e49.
[12] Ben-Nakhi AE, Mahmoud MA. Energy conservation in buildings through
efcient A/C control using neural networks. Applied Energy 2002;73(1):5e23.
[13] Narendra KS, Parthasarathy K. Identication and control of dynamical systems
using neural network. IEEE Transactions on Neural Network 1990;1(1):4e27.
[14] Levin AU, Narendra KS. Control of nonlinear dynamical systems using neural
network: controllability and stabilization. IEEE Transactions on Neural
Network 1993;4(2):192e206.
[15] Levin AU, Narendra KS. Control of nonlinear dynamical systems using neural
network e part II: observability, identication, and control. IEEE Transactions
on Neural Network 1996;7(1):30e42.
[16] Maria J, Menezes Jr P, Barreto GA. Long-term time series prediction with the
NARX network: an empirical evaluation. Neurocomputing 2008;71(16e18):
3335e43.
[17] Wang J. Data mining: opportunities and challenges. North Sydney: Idea
Group; 2003.
[18] Friedman J. Stochastic gradient boosting. Stanford University Statistics
Department; 1999.
[19] Hastie T, Tibshirani R, Firedman JH. The elements of statistical learning. New
York: Springer; 2001.
[20] Kusiak A, Li M, Tang F. Modeling and optimization of HVAC energy
consumption. Applied Energy 2010;87(10):3092e102.
[21] Casella G, Berger R. Statistical inference. Pacic Grove, CA: Duxbury Press;
1990.
[22] Kennedy J, Eberhart RC. Particle swarm optimization. In: Proceedings of
international conference on neural networks. Perth, Australia: IEEE Press;
1995. p. 1942e8.
[23] Poli R, Kennedy J, Blackwell T. Particle swarm optimization: an overview.
Swarm Intelligence 2007;1(1):33e57.
[24] Abido MA. Two-level of non-dominated solutions approach to multi-objective
particle swarm particle swarm optimization. In: Genetic and evolutionary
computation conference. 2007. p. 726e33.