Applied Energy 162 (2016) 666674

Contents lists available at ScienceDirect

Applied Energy
journal homepage: www.elsevier.com/locate/apenergy

Simulation-based optimization of an integrated daylighting and HVAC

system using the design of experiments method
Wonuk Kim, Yongseok Jeon, Yongchan Kim
Department of Mechanical Engineering, Korea University, Anam-Dong, Sungbuk-Gu, Seoul 136-713, Republic of Korea

h i g h l i g h t s

The IDHVAC system is optimized using the integrated meta-model and GA.

The design of experiments method is applied to train the integrated meta-model.

The GA-optimized models are compared with the reference model.

The GA-optimized IDHVAC model shows the best performance among them.

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

Article history: The use of daylight in buildings to save energy while providing satisfactory environmental comfort has
Received 2 May 2015 increased. Integration of the daylighting and thermal energy systems is necessary for environmental
Received in revised form 2 October 2015 comfort and energy efficiency. In this study, an integrated meta-model for a daylighting, heating, venti-
Accepted 24 October 2015
lating, and air conditioning (IDHVAC) system was developed to predict building energy performance by
Available online 11 November 2015
artificial lighting regression models and artificial neural network (ANN) models, with a database that was
generated using the EnergyPlus model. The design of experiments (DOE) method was applied to generate
Keywords:
the database that was used to train robust ANN models without overfitting problems. The IDHVAC sys-
Integrated energy system modelling
Daylighting
tem was optimized using the integrated meta-model and genetic algorithm (GA), to minimize total
Genetic algorithm (GA) energy consumption while satisfying both thermal and visual comfort for occupants. During three
Artificial neural network (ANN) months in the winter, the GA-optimized IDHVAC model showed, on average, 13.7% energy savings against
Design of experiments (DOE) the conventional model.
Energy efficiency 2015 Elsevier Ltd. All rights reserved.

1. Introduction an effective and integrated optimization algorithm for these tech-

Nowadays there is strong demand for environmental comfort in
the workspace for health and productivity. In providing that, heat-
ing, ventilating, and air-conditioning (HVAC) and lighting systems
use approximately 5060% of a buildings energy consumption to
maintain thermal and visual comfort for occupants [1]. In addition,
the use of solar heat and daylight in buildings, to save energy, has
increased [2]. However, when daylighting and HVAC systems are
separately controlled, energy efficiency and environmental com-
fort are often in conflict with each other [3]. Therefore, it is neces-
sary to adopt an integrated control process for the two systems
while satisfying environmental comfort. Since optimized control
to achieve energy conservation and environmental comfort in
buildings is a multi-dimensional and complex nonlinear problem,
Corresponding author. Tel.: +82 2 3290 3366; fax: +82 2 921 5439.
E-mail address: yongckim@korea.ac.kr (Y. Kim).
nologies also needs to be developed under various operating and
geometrical conditions.
Optimization algorithms can be divided into two main cate-
gories: conventional gradient-based methods and gradient-free
direct methods. Since building systems are very often nonlinear,
the gradient-based optimization methods are inapplicable to most
building studies. In contrast, the gradient-free methods are more
suitable to building studies [4] due to the use of stochastic
approaches. Stochastic population-based algorithms, such as the
genetic algorithm (GA), particle swarm optimization, the hybrid
algorithm, and the evolutionary algorithm, are the most frequently
used methods in building performance optimization. The most
well-known and widely accepted method is GA [5]. The GA shows
excellent optimization performance in multi-input, multi-output
non-linear systems [6]. The GA has been used to optimize setpoints
of several devices in HVAC systems at a given time [7,8].

http://dx.doi.org/10.1016/j.apenergy.2015.10.153
0306-2619/ 2015 Elsevier Ltd. All rights reserved.
 W. Kim et al. / Applied Energy 162 (2016) 666674 667

Nomenclature

AHU air handling unit Greek symbols

ANN artificial neural network a azimuth angle ()
CAV constant air volume / slat angle ()
CSTR constraint c ratio (01)
Cv (RMSE) coefficient of variation of the root mean squared error
DOE design of experiments Subscripts
FCU fan coil unit AHU air handling unit
GA genetic algorithm art artificial
HVAC heating, ventilating, and air conditioning day daylight
I illuminance (lux) DOE design of experiments
IDHVAC integrated daylighting and HVAC E equipment
MBE mean bias error FCU fan coil unit
OA outdoor air in indoor
P energy consumption (kW) L lighting
R2 coefficient of determination OA outdoor air
RF relative flow rate based on the maximum value out outdoor
Sch on/off schedule (0 or 1) ref reference
T temperature (C) SP setpoint
tot total
t time
t-1 previous time step
U overall heat transfer coefficient (W m2 K1) w water

In building applications, simulations can be useful in several
aspects for improving the design and integration of building con-
trol systems [9]. However, it is difficult to apply building simula-
tion models to a population-based optimized control algorithm,
such as GA, in real time because most models do not allow access
to state variables [10]. One very efficient solution is to use a meta-
model to imitate the behavior of the building model, and then to
optimize individual control variables [4]. The artificial neural net-
work (ANN) has been adopted for the meta-model because it has
tremendous advantages in deriving and extracting accurate pat-
terns from complicated data in the modeling of building systems
[7,1113]. Although the ANN can improve the accuracy of the
model for an integrated system by using more hidden nodes, the
network may then forecast out-of-samples very poorly [14]. This
is the so-called overfitting problem. An early stopping method con-
taining three data sets of training, testing, and validation parts [15]
has been popularly used to avoid overfitting in a building system.
However, despite using the early stopping method, it is still diffi-
cult to achieve both good prediction ability and a robust ANN
model at the same time in a complex integrated system with lim-
ited training data. Therefore, in this study, the database was con-
structed using building performance simulations with the design
of experiments (DOE) method to develop a robust meta-model in
the global optimization process.
Even though various control and optimization methods in
building systems have been developed, an integrated building con-
trol considering the simultaneous interactions of systems is still
lacking. This impedes smart-grid initiatives and market penetra-
tion [16]. An optimization of the setpoints in an integrated day-
lighting and HVAC (IDHVAC) system presents a great challenge,
because the IDHVAC system is a nonlinear, multivariable system.
The objective of this study is to optimize the IDHVAC system for
energy effective operation while considering environmental com-
fort. An integrated meta-model was developed to predict building
performance using artificial lighting regression models and ANN
models, with a database generated by the EnergyPlus model [17].
The setpoints of the IDHVAC system were optimized using the
integrated meta-model and GA to minimize the total energy con-
sumption in the heating season.
2. Experiments on the IDHVAC system
Fig. 1 shows a schematic of the IDHVAC system in a building.
The IDHVAC system consisted of an HVAC system and a daylighting
system. The IDHVAC system was installed and monitored in an
office (6.35 m  3.75 m) in a building in Seoul, Korea [18]. Table 1
shows the descriptions of the reference building and office. The
HVAC system in the building consisted of direct gas-fired absorp-
tion chillers/heaters, a constant air volume (CAV) duct system,
and a fan coil unit (FCU). A CAV box was connected to the AHU,
to maintain thermal comfort by controlling the air flow rate to each
thermal zone. The supply air temperature was controlled by mod-
ulating the water flow rates in the heating and cooling coils,
respectively. The FCU was installed beneath the exterior window,
to manage the thermal load through the window. In the experi-
ments, we measured the temperatures and flow rates of the supply
and return air, and the water in the FCU.
The daylighting system consisted of a motorized venetian blind
(Somfy LS40 motor), and a non-dimmable and two dimmable
96 W fluorescent lights. Generally, a venetian blind can be con-
trolled by the slat angle and height. However, in this study, the
motorized venetian blind was controlled by the slat angle at the
full height, in order to investigate the thermal and visual perfor-
mance of the IDHVAC system. The blind slat angle was controlled
to accept as much radiation as possible, while preventing direct
daylight from entering the room [18]. As shown in Fig. 1, the illu-
minance in the room was measured at two points by photometers
with an accuracy of 5%. The room temperature was measured by a
thermocouple with an accuracy of 0.2 C. The power consumption
of the fluorescent lights was measured using a power analyzer
with an accuracy of 0.1%.
3. Development of the system optimization model
As shown in Fig. 2, an integrated meta-model was developed to
optimize the control variables, using a GA in real time. The inte-
grated meta-model consisted of ANN models and artificial lighting
regression models. The ANN models for the room temperature,
668 W. Kim et al. / Applied Energy 162 (2016) 666674

Constant
Reheat
air volume
coil
box
Valve

Dimming Exhaust Dimming Diffuser Typical

fluorescent grille fluorescent fluorescent

Window
Venetian
blind

Iin,1 Iin,2 Troom

FCU

0.75m
2.00m 1.50m

Fig. 1. A schematic of the IDHVAC system.

Table 1
Brief description of the building. EnergyPlus model. In addition, artificial lighting regression models
were developed based on the experimental data [18], in order to
Section Details
estimate the artificial indoor illuminance and lighting energy con-
Location Seoul (latitude 37.58N, longitude 127.02E) sumption. The IDHVAC system was optimized by minimizing the
Building Office building, 7 floors above grade, 1 below grade
total energy consumption in terms of the controllable input vari-
(constructed in 1996)
Total floor area 18,231 m2 ables, while satisfying the constraints for indoor thermal and visual
Plan and height T-shape, floor-to-floor 3.5 m comfort. All calculations were conducted in the MATLAB program
Operating hours Monday to Friday: 0818 using the measured weather data [19] in Seoul, Korea during Jan-
HVAC design Lighting load = 20 W m2, Equipment uary to March 2011.
parameter load = 13.78 W m2
Space design temperature: Cooling = 26 C,
Heating = 20 C
3.1. Database generation using the EnergyPlus model
Window-to-wall North: 23.6%, East: 17.5%, South: 26.9%, West: 17.5%
area ratio
The reference building was simulated using the EnergyPlus
Building envelope
External wall 100 mm brickwork, 79.5 mm polystyrene, 100 mm model based on the data given in Table 1. The windows in the
concrete block, 13 mm gypsum board building were modeled using dimensions typically found in an
(U = 0.35 W m2 K1) office by considering the variation in the window-to-wall area
Internal wall 13 mm plaster board, 50 mm wool quilt, 13 mm plaster ratio for each wall direction (Table 1). The total indoor illuminance
board (U = 0.649 W m2 K1)
Roof 10 mm roof build-up, 144.5 mm glass wool, 200 mm air
was estimated by the summation of the indoor daylight illumi-
gap, 13 mm plaster board (U = 0.259 W m2 K1) nance and indoor artificial illuminance. Based on the split-flux
Window North: 6 mm double LoE tint glass, 13 mm air-space method [17,19], the indoor daylight illuminance was calculated
(U = 1.635 W m2 K1) by the multiplication of the outdoor daylight illuminance and
South, East, West: 3 mm double LoE clear glass, 13 mm
pre-calculated daylight factor. The daylight factor was determined
air-space (U = 1.978 W m2 K1)
by the summation of the direct and reflected daylight incidences.
Office
The indoor artificial illuminance was calculated from evenly dis-
Inner surface Color: cream-white (light reflectance = 60%)
Window Direction: south, Dimension: 3.05 m  1.89 m tributed lighting load. The lighting and equipment loads were
Daylighting and Illuminance: 2 reference points with 500 lux setpoint specified at 20 W m2 and 13.78 W m2, respectively, due to the
light Blind: a motorized venetian (5 mm wide white higher occupancy density of 0.11 person m2 [20]. In addition,
aluminum slats, 180 rotation range) the convective portions of the lighting and equipment loads were
Light dimming method: continuous
assumed to be 41% and 80%, respectively.
The EnergyPlus model has been tested [21], and validated with
total energy consumption, and indoor daylight illuminance were existing programs and experimental data [22,23]. In this study, as
developed to enhance the accuracy of the system model. The shown in Fig. 3, the predicted energy consumption using the Ener-
ANN models were trained using the database generated by the gyPlus model was also compared with the measured data in 2011.
 W. Kim et al. / Applied Energy 162 (2016) 666674 669

2SWLPL]DWLRQSURFHVV

Uncontrollable
'DWDEDVH ,QWHJUDWHGPHWDPRGHO input variables

EnergyPlus
simulation
Meta-models
with DOE (ANN)
Optimization
(GA)

Measured Lighting
data for regression
lighting models Controllable
input variables,
output variables

Fig. 2. Flow chart of the system optimization in the IDHVAC system.

2000 500 Measured

Predicted
Measured
Predicted 400
Energy consumption (GJ)

R2 =0.86
Illuminance (lux)

1600
Cv(RMSE) = 8.1% 300
MSE = -1.8%

1200 200

100
800
0

400 0 1000 2000 3000

Jan Feb Mar Apr May JuneJuly Aug Sep Oct Nov Dec Time (min)
Month
Fig. 4. Validation of the EnergyPlus model for indoor daylight illuminance.
Fig. 3. Validation of the EnergyPlus model for annual energy consumption.
Table 2
Description of variables.
The M&V Guidelines 3.0 [24] recommends that the mean bias error Variable type Variables Description Bound/
(MBE) and coefficient of variation of the root mean squared error, unit
Cv (RMSE), for the monthly energy consumption should be within Controllable input Slat angle (/) Venetian blind slat angle 2084
5% and 15%, respectively. During the same period of time, the variables TSP AHU supply air setpoint 20
EnergyPlus model yielded predictions with an MBE and a Cv temperature 34 C
SchAHU AHU operation status 0 or 1
(RMSE) of 1.8% and 8.1%, respectively, when compared with the
RFw,FCU Relative water flow rate of 01
measured data. Fig. 4 compares the predicted indoor daylight with FCU
the measured data near the exterior window in the IDHVAC system OA-mixing Outdoor-air mixing ratio 0.31
over five days. The indoor daylight illuminance was predicted ratio (c)
using the same weather conditions [19] with the measured data Uncontrollable TOA Outdoor temperature C
over the five days. Generally, the predicted indoor daylight illumi- input variables Azimuth Sun azimuth angle
nance was consistent with the measured data, with a coefficient of angle (a)
Iout Outdoor illuminance lux
determination (R2) of 0.86. Stiff ravine points in the middle of a day
Equipment Energy used in office W
were observed, due to shading from an adjacent building. energy (PE) equipment
Table 2 shows the input and output variables in the IDHVAC Troom(t1) Indoor temperature in the C
system for the simulations. The controllable input variables of previous time step
the IDHVAC system included the blind slat angle, supply air tem- Calculated output Troom Indoor temperature in the C
perature of the CAV box, outdoor-air mixing ratio, operating status variables current time step
Iin Total indoor illuminance lux
(on or off) of the AHU, and relative water flow rate of the FCU. The
Iin,day Indoor daylight illuminance lux
relative flow rate (RF) was defined as the ratio of the actual flow Lighting Lighting energy load W
rate to the maximum flow rate. The BoxBehnken DOE method energy (PL)
[25] was used for the controllable input variables to determine Total energy Total energy load W
the optimum number of simulations to generate the database. (Ptot)
670 W. Kim et al. / Applied Energy 162 (2016) 666674

Table 3
Training, validation, and testing results of the neural network models.

Neural network Number of Number of Number of hidden R2/Cv (RMSE)

model inputs hidden layers layer neurons
Training Validation Testing
ANNDOE
Iin,1 4 3 19,31,30 0.9895/24.7503 0.9895/25.7696 0.9895/25.7615
Iin,2 0.9883/9.7844 0.9876/9.9989 0.9878/9.9048
Troom 11 3 18,25,27 0.9981/0.2023 0.9980/0.2052 0.9981/0.2051
Ptot 11 3 25,24,15 0.9997/34.4756 0.9997/38.1864 0.9997/37.0733
ANNref
Iin,1 4 3 28,21,23 0.9912/17.3700 0.9836/23.9546 0.9850/23.8023
Iin,2 31,30,29 0.9964/5.7308 0.9896/9.8130 0.9910/9.3895
Troom 11 3 7,9,8 0.9931/0.3098 0.9925/0.3239 0.9916/0.3431
Ptot 11 3 13,19,13 1/11.9432 0.9998/24.9041 0.9996/34.7099

7.2 Predicted value by ANNref 2500

6.8 Predicted value by ANNDOE Reference
GA-optimized HVAC
Ptot (kW)

Value of EnergyPlus
6.4 2000 GA-optimized IDHVAC
6.0

Total energy (MJ)

5.6
1500

19.2

19.0 1000
Troom (oC)

18.8

18.6 500

18.4
200 0
Iin,1 (lux)

Jan Feb Mar

150
Month
100
50 Fig. 6. Comparison of total energy consumption for the optimized models.
0
-50
25 110000
250
Outdoor temperature 100000
Iin,2 (lux)

200 20

Outdoor illuminance (lux)

Outdoor illuminance
Outdoor temperature (oC)

90000
150 15
80000
100
10 70000
50
5 60000
0
50000
0
20 40 60 80 40000
Blind slat angle (O) -5 30000
-10 20000
Fig. 5. Results of ANN models, dependence on the blind slat angle.
10000
-15
0
The database was generated by conducting 28 simulations from 1-2 days: Jan., 3-6 days: Feb., 7-8 days: Mar.
-20
January to March, in time steps of ten minutes. 1 2 3 4 5 6 7 8
Days
3.2. Development of the integrated meta-model
Fig. 7. Variations of outdoor temperature and illuminance over eight days.

The integrated meta-model consisted of the artificial lighting

regression models for the indoor artificial illuminance (Iart) and
lighting energy consumption (PL), and the ANN models for the
room temperature (Troom), total energy consumption (Ptot), and
indoor daylight illuminance (Iin,day). The artificial lighting regres-
sion models for the indoor artificial illuminance (Iart) and lighting
energy consumption (PL), which were developed based on the mea-
sured data in terms of dimming signals, were adopted from Xu
et al. [18].
The variables included in the ANN models for room tempera-
ture and total energy consumption were determined by the load
calculation method [26]. The indoor energy balance can be esti-
mated in terms of convection from the zone surfaces, infiltration
and ventilation into the zone, the convective part of internal loads,
and net energy from the supply air. Based on the indoor energy bal-
ance, the selected contributing variables for estimating the room
temperature and the total energy consumption were the blind slat
angle, the outdoor-air mixing ratio, convective parts of artificial
lighting and equipment, supply air temperature, operating status
of the AHU, relative water flow rate of the FCU, outdoor tempera-
ture, and the indoor temperature from the previous time step. In
 W. Kim et al. / Applied Energy 162 (2016) 666674 671

addition, the outdoor illuminance and solar azimuth angle were 30 3600
included to consider the effects of light and heat gain from the Troom, Reference
Troom, GA-optimized HVAC
3200
sun. Therefore, the considered variables in the ANN models for

Indoor illuminance (lux)

25
Troom, GA-optimized IDHVAC

Room temperature (oC)

room temperature and total power consumption were given by: 2800

20 2400
T room f /; cOA ; P L ; P E ; T SP ; SchAHU ; RF w;FCU ; T OA ; T roomt1 ; Iout ; a 1
Iin,1, Ref.&GA-optimized HVAC 2000
15 Iin,2, Ref.&GA-optimized HVAC
Ptot f /; cOA ; PL ; PE ; T SP ; SchAHU ; RF w;FCU ; T OA ; T roomt1 ; Iout ; a 2 Iin,1, GA-optimized IDHVAC 1600
Iin,2, GA-optimized IDHVAC
The variables included in the ANN model for the indoor daylight 10 1200
illuminance were determined by the physics of daylight and the 800
parameters pertaining to the window. Park and Athienitis [27] 5
showed that the indoor daylight illuminance was a function of 400
1-2 days: Jan., 3-6 days: Feb., 7-8 days: Mar.
the blind slat angle and outdoor illuminance. Moreover, the solar 0 0
azimuth angle and outdoor temperature were considered to reflect 1 2 3 4 5 6 7 8
the effects of exterior shading and long-wave radiation, respec- Days
tively. Therefore, the considered variables in the ANN model for
the indoor daylight illuminance were given by: Fig. 9. Variations of indoor temperature and illuminance over eight days.

Iin;day f Iout ; T OA ; a; / 3
300
The ANN models were trained by applying the LevenbergMar-
Ref. & GA-optimized HVAC
quardt algorithm [28]. The number of hidden layers in the ANN 280
GA-optimized IDHVAC
models was fixed at three, and the number of neurons in each layer
260
was optimized by the GA. Moreover, sigmoid transfer functions
were used for the input and hidden layers and a linear transfer Lighting energy (W) 240
function for the output layer. Training, validation, and testing data-
sets were randomly separated by ratios of 7:1.5:1.5 for the Energy- 220
Plus results, to prevent overfitting. Based on the early stopping
200
method, the ANN training stopped when the error increased for
the validation data [29]. 180
In addition, the ANN models trained by 28 simulation results
using the DOE method (ANN models of DOE, ANNDOE) were com- 160
pared with them trained by simulation results validated using 140
the measured data (ANN models of reference, ANNref) according 1-2 days: Jan., 3-6 days: Feb., 7-8 days: Mar.
to the blind slat angle with other fixed input variables. Table 3 120
shows the training, validation, and testing results of all the ANN 1 2 3 4 5 6 7 8
models. Based on the coefficient of determination (R2) and Cv Days
(RMSE), all the ANN models yielded satisfactory results for each
Fig. 10. Variation of lighting energy over eight days.
performance. However, as shown in Fig. 5, the ANNDOE models
showed continuous and reasonable changes according to the blind
slat angle, but the ANNref models showed overfitting. Therefore, the 3.3. Development of the system optimization model
ANNref models were inadequate at optimizing setpoints of the IDH-
VAC system. The ANNDOE models were appropriate to estimate the As given in Eq. (4), the total energy consumption of the IDHVAC
total energy consumption and indoor temperature from the con- system was minimized by optimizing the controllable input vari-
trollable and uncontrollable input variables and lighting energy. ables using the GA. The object function was set as the minimization
of the total power consumption. In addition, as shown in Table 2,
the IDHVAC system had five controllable input variables: the blind
120
slat angle, supply air setpoint temperature, AHU on/off status, rel-
110 Ref. & GA-optimized HVAC ative water flow rate of the FCU, and outdoor-air mixing ratio.
100 GA-optimized IDHVAC
Blind slat angle (degree)

90 minPtot
80 where;
70 20 6 / 6 84
60 20 6 T SP 6 34 4
50 SchAHU 0 or 1
40 0 6 RF w;FCU 6 1
30
0:3 6 cOA 6 1
20 In each time step, the controllable input variables were opti-
10 mized to satisfy the objective function with the fixed uncontrol-
1-2 days: Jan., 3-6 days: Feb., 7-8 days: Mar.
lable input variables, while satisfying the constraints for indoor
0
1 2 3 4 5 6 7 8 thermal and visual comfort. As given in Eq. (5), the constraints
Days for indoor thermal and visual comfort were considered in terms
of the room temperature and total indoor illuminance, respec-
Fig. 8. Variation of blind slat angle over eight days. tively. The constraints of the room temperature and total indoor
672
125
Reference
GA-optimized HVAC
GA-optimized IDHVAC
100
Total energy (MJ)
75
50
25
0
1 2 3 4 5 6
Days
Fig. 11. Variation of total energy consumption over eight days.
illuminance were normalized by the difference in each value
between the upper and lower bounds. The constraints could be sat-
isfied when both terms in Eq. (5) were equal to zero. Generally, the
indoor temperature from 20 C to 23 C [30] and the illuminance of
500 lux [31] are recommended for indoor thermal and visual
comfort, respectively, in the heating condition. The lower and
upper bounds for the variables were set as follows: supply air
temperature from 20 C to 34 C, total indoor illuminance from
500 to 550 lux, indoor temperature from 20 C to 23 C, direct day-
light blocked for anti-glare and outdoor-air mixing ratio higher
than 0.3. In addition, the direct daylight was blocked by controlling
the blind slat angle based on the predicted solar altitude angle
[18,27].
CSTR max0; 20  T room =23  20 fmax0; 500  Iin;1
max0; Iin;1  550 max0; 500  Iin;2
max0; Iin;2  550g=550  500
4. Results and discussion
System optimization results using the GA-optimized HVAC and
GA-optimized IDHVAC models were compared with those using
the reference model. All three models used dimming fluorescent
lights. Xu et al. [18] reported that the dimming fluorescent light
saved about 11% of lighting energy. The reference model was set
based on the results of the previous experiments [18]. In the refer-
ence model, the blind slat angle was optimized to maximize day-
light within the range of no-direct daylight to prevent glare at a
given time, while the HVAC system was optimized using a conven-
tional on/off control with the fixed setpoints. In the GA-optimized
HVAC model, the blind slat angle was optimized using the same
method as the reference model, while the setpoints in the HVAC
system were optimized using the GA to minimize the objective
function [7,8]. Therefore, the reference and GA-optimized HVAC
models were unable to consider both visual and thermal comfort.
In the GA-optimized IDHVAC model, both the blind slat angle
and the setpoints in the HVAC system were optimized using the
GA to minimize the objective function.
Fig. 6 shows the total energy consumption of the IDHVAC sys-
tem in the building located in Seoul, Korea. During three months
of the winter of 2011, the average total energy consumptions of
the GA-optimized HVAC and GA-optimized IDHVAC models were
11.7% and 13.7% lower, respectively, than that of the reference
model. Therefore, the GA-optimized IDHVAC model showed the
W. Kim et al. / Applied Energy 162 (2016) 666674
highest energy savings among them because of the simultaneous
optimizations of the HVAC system and venetian blind. The opti-
mization method of the HVAC system was the major difference
between the GA-optimized HVAC model and the reference model,
while the optimization method of the venetian blind was the major
1-2 days : Jan. difference between the GA-optimized IDHVAC model and the GA-
3-6 days : Feb. optimized HVAC model. Generally, the optimization of the HVAC
7-8 days : Mar.
system showed more energy saving potential than the optimiza-
tion of the venetian blind. In addition, as the outdoor temperature
increased, according to the change in month from January to
March, the energy saving potential of the GA-optimized IDHVAC
model against the GA-optimized HVAC model decreased, due to
the decreased effects of the venetian blind optimization.
The effects of the input parameters on the system performance
were typically analyzed for eight days with various outdoor tem-
peratures and illuminances. Fig. 7 shows the outdoor temperature
7 8
and outdoor illuminance for the eight days in the winter of 2011.
The eight days consisted of two days (12) in January, four days
(36) in February, and two days (78) in March. The outdoor tem-
perature and illuminance varied with large fluctuations, and their
maximum values were observed in the middle of the day. There-
fore, due to the variations in heating load and daylight with time,
the IDHVAC system was optimized in order to cover the heating
load, and to satisfy visual comfort at a given time.
Fig. 8 shows the optimized blind slat angles for the optimization
models for the eight days. In the reference and GA-optimized HVAC
models, the optimized blind slat angle at a given time strongly
depended on the solar altitude angle due to the no-direct daylight
control regardless of the outdoor illuminance. In the GA-optimized
IDHVAC model, the optimized blind slat angle at a given time was
determined to minimize the total energy consumption, with con-
sideration of the setpoints in the HVAC system. Therefore, even
though the general trends of the optimized blind slat angles for
all optimization models were consistent, the GA-optimized IDH-
VAC model showed lower blind slat angles than the reference
and GA-optimized HVAC models, to satisfy visual comfort and min-
imize the heat loss through the window. In all optimization mod-
5 els, several peak blind slat angles were observed during the eight
days, because of the fully opened blind not having any direct day-
light before sunset. In addition, the optimized blind slat angle sig-
nificantly varied in the morning of the last two-days (78), because
of the change in the time of sunrise.
Fig. 9 presents the room temperature and indoor illuminance of
the optimization models. The room temperatures were generally
maintained within thermal constraints with all models, except at
the beginning of HVAC operation. The room temperature of the last
three-days (68) increased because of the increased outdoor tem-
perature. The illuminance on the 1st point near the window with
the reference and GA-optimized HVAC models was very high on
days having high outdoor illuminance because the blind slat angles
of these models were only controlled to block direct daylight in
stepped movements. On the contrary, the illuminance of the GA-
optimized IDHVAC model was successfully maintained within
visual constraints due to the optimization of the blind slat angle
in every time step.
Fig. 10 shows the lighting energy of the optimization models
over the eight days. The general trends of the lighting energy for
these models were consistent. Unusually, the lighting energy trend
showed a ravine on the morning of the eighth day, because light-
ing energy decreased in the morning with higher outdoor illumi-
nance and a higher blind slat angle in the first step, due to the
solar altitude angle. The indoor daylight illuminance depended
on the outdoor illuminance, solar altitude angle, and blind slat
angle [32]. Therefore, in the middle of a day, the lighting energy
was reduced, because the outdoor illuminance was relatively high.
However, in the middle of the day, the GA-optimized IDHVAC
 W. Kim et al. / Applied Energy 162 (2016) 666674 673

60000 decrease in the blind slat angle [33]. However, when the outdoor
(a) 40.0% illuminance was low, such as on the second and fifth days, no
35.0%
energy savings were observed by optimizing the venetian blind.
Outdoor illuminance (lux)

50000 Energy savings (%) 30.0%

25.0% Therefore, supervisory control for the IDHVAC system is required
20.0%
15.0%
with respect to the outdoor illuminance and outdoor temperature.
40000 10.0%
5.0%
0% 5. Conclusions
30000
In this study, an optimization model for the IDHVAC system was
developed using the GA, ANN, and DOE methods. An integrated
20000
meta-model for the IDHVAC system was developed to predict
building performance using artificial lighting regression models
10000 and ANN models, with a database generated using the EnergyPlus
model. The ANN models with DOE showed reasonable results with-
-8 -6 -4 -2 0 2 4 6 8 10 12
out overfitting problems. The controllable input variables in the
Outdoor temperature (oC)
IDHVAC system were optimized using the integrated meta-model
and GA to minimize total energy consumption, while satisfying
60000
the constraints for indoor thermal and visual comfort. During three
(b) 11.5%
months of the winter of 2011, the optimization of the HVAC system
10.0%
Outdoor illuminance (lux)

50000 Energy savings (%) 8.5% showed more energy saving potential than the optimization of the
7.0%
5.5%
venetian blind. The effects of the optimization in the blind slat
4.0% angle became more dominant at a higher outdoor illuminance
40000 2.5%
1.0% and relatively lower outdoor temperature. During the three
0% months, the GA-optimized HVAC and the GA-optimized IDHVAC
30000 models showed on average 11.7% and 13.7% lower total energy
consumption than the reference model, respectively. The proposed
20000
optimization model may provide an effective control solution for
the IDHVAC system. In addition, it can be applied to different con-
trol strategies for integrated non-linear systems.
10000

-8 -6 -4 -2 0 2 4 6 8 10 12
Outdoor temperature (oC)
Fig. 12. Energy savings of the optimization models: (a) the GA-optimized HVAC
model against the reference model, and (b) the GA-optimized IDHVAC model
against the GA-optimized HVAC model.
model showed slightly increased lighting energy over the GA-
optimized HVAC model, because it had a lower blind slat angle
than the GA-optimized HVAC model. Consequently, the decreased
heat loss through the window with the lower blind slat angle and
the increased lighting energy led to a decrease in the heating load
of the room.
Fig. 11 shows the total energy consumption for each of the opti-
mization models over the eight days. As the daily average outdoor
temperature and illuminance increased, the total energy consump-
tion decreased. However, in the reference model the total energy
consumption on the fourth day was lower than on the fifth day
despite similar outdoor temperatures, because the fourth day had
a lower heating load due to higher solar radiation. Fig. 12 shows
the contour map of energy savings according to the outdoor illumi-
nance and daily average outdoor temperature. As shown in Fig. 12
(a), energy savings of the GA-optimized HVAC model against the
reference model were observed on most days, due to the dominant
effects of the optimization of the HVAC system. Energy savings
were relatively small on the days when outdoor temperatures
and illuminance were high. However, as shown in Figs. 11 and 12
(b), energy savings of the GA-optimized IDHVAC model against the
GA-optimized HVAC model were observed only on the third and
fourth days because of the dominant effects of the optimization
in the blind slat angle at a higher outdoor illuminance and rela-
tively lower outdoor temperature. On the third and fourth days,
the indoor daylight illuminance was sufficient due to the higher
outdoor illuminance. And the heat loss through the window
decreased, even at the lower outdoor temperature, with the
Acknowledgements
This work was supported by the Human Resources Develop-
ment Program (No. 20144010200770) and the Energy Technology
Development Program (No. 20142010102660) of the Korea Insti-
tute of Energy Technology Evaluation and Planning (KETEP),
through a grant funded by the Korea Government Ministry of
Trade, Industry and Energy.
References
[1] Prez-Lombard L, Ortiz J, Pout C. A review on buildings energy consumption
information. Energy Build 2008;40:3948.
[2] Kristl Z, Koir M, Trobec Lah M, Krainer A. Fuzzy control system for thermal
and visual comfort in building. Renew Energy 2008;33:694702.
[3] Dounis A, Caraiscos C. Advanced control systems engineering for energy and
comfort management in a building environmenta review. Renew Sustain
Energy Rev 2009;13:124661.
[4] Magnier L, Haghighat F. Multiobjective optimization of building design using
TRNSYS simulations, genetic algorithm, and artificial neural network. Build
Environ 2010;45:73946.
[5] Nguyen A-T, Reiter S, Rigo P. A review on simulation-based optimization
methods applied to building performance analysis. Appl Energy
2014;113:104358.
[6] Zolpakar NA, Mohd-Ghazali N, Ahmad R. Simultaneous optimization of four
parameters in the stack unit of a thermoacoustic refrigerator. Int J Air-
Conditioning Refrig 2014;22:1450011.
[7] Kusiak A, Tang F, Xu G. Multi-objective optimization of HVAC system with an
evolutionary computation algorithm. Energy 2011;36:24409.
[8] Kusiak A, Xu G. Modeling and optimization of HVAC systems using a dynamic
neural network. Energy 2012;42:24150.
[9] Treado S, Delgoshaei P, Windham A. Simulation-based approaches for building
control system design and integration. In: Proc build simulations 2011 12th
conf int build perform simulations assoc Sydney; 1416 November 2011. p.
254451.
[10] Hoes P, Loonen R, Trcka M, Hensen J. Performance prediction of advanced
building controls in the design phase using ESP-r, BCVTB and MATLAB.
Loughborough, UK: BSO12 (IBPSA-England); 2012.
[11] Wei X, Kusiak A, Li M, Tang F, Zeng Y. Multi-objective optimization of the
HVAC (heating, ventilation, and air conditioning) system performance. Energy
2015;83:294306.
[12] Wong SL, Wan KKW, Lam TNT. Artificial neural networks for energy analysis of
office buildings with daylighting. Appl Energy 2010;87:5517.
674 W. Kim et al. / Applied Energy 162 (2016) 666674
[13] Tang F, Kusiak A, Wei X. Modeling and short-term prediction of HVAC system
with a clustering algorithm. Energy Build 2014;82:31021.
[14] Zhang GP, Patuwo BE, Hu MY. A simulation study of artificial neural networks
for nonlinear time-series forecasting. Comput Oper Res 2001;28:38196.
[15] Sarle WS. Stopped training and other remedies for overfitting. In: Proc. 27th
symp. interface comput. sci. stat.; 1995. p. 35260.
[16] Shaikh PH, Nor NBM, Nallagownden P, Elamvazuthi I, Ibrahim T. A review on
optimized control systems for building energy and comfort management of
smart sustainable buildings. Renew Sustain Energy Rev 2014;34:40929.
[17] ENERGYPLUS. EnergyPlus engineering reference. Lawrence Berkeley National
Laboratory; 2013.
[18] Xu L, Kim WU, Kim Y. Energy performance of dimmable fluorescent lamps in
the integrated daylighting control system in cold climate. In: Proc. 6th Asian
conf. refrig. air cond., Xian, China: ACRA; 2012.
[19] Yun G, Kim K. An empirical validation of lighting energy consumption using
the integrated simulation method. Energy Build 2012.
[20] ANSI/ASHRAE. ASHRAE standard 62.1-2004. Ventilation for Acceptable Indoor
Air Quality; 2004.
[21] Henninger R, Witte M. EnergyPlus testing with building thermal envelope and
fabric load tests from ANSI/ASHRAE standard 140-2011. Washington, DC
20585-0121, The United State of America; 2013.
[22] Loutzenhiser PG, Maxwell GM, Manz H. An empirical validation of the
daylighting algorithms and associated interactions in building energy
simulation programs using various shading devices and windows. Energy
2007;32:185570.
[23] Li DHW, Lam TNT, Wong SL, Tsang EKW. Lighting and cooling energy
consumption in an open-plan office using solar film coating. Energy
2008;33:128897.
[24] M&V Guidelines. Measurement and verification for federal energy projects.
version 3. US Department of Energy; 2008.
[25] Myers RH, Anderson-Cook CM. Response surface methodology: process and
product optimization using designed experiments. John Wiley & Sons; 2009.
[26] Spitler JD. Load calculation applications manual. Atlanta, GA: ASHRAE; 2009.
[27] Park K-W, Athienitis AK. Workplane illuminance prediction method for
daylighting control systems. Sol Energy 2003;75:27784.
[28] Guillemin A, Morel N. An innovative lighting controller integrated in a self-
adaptive building control system. Energy Build 2001;33:47787.
[29] Piotrowski AP, Napiorkowski JJ. A comparison of methods to avoid overfitting
in neural networks training in the case of catchment runoff modelling. J Hydrol
2013;476:97111.
[30] SAREK. SAREK handbook. 3rd ed. Seoul, Korea: The Society of Air-Conditioning
and Refrigerating Engineers of Korea; 2011.
[31] IESNA. The IESNA lighting handbook: reference & application. 9th ed. New
York, New York, USA: Illuminating Engineering Society of North America;
2000.
[32] Park K-W, Athienitis AK. Development and testing of an integrated daylighting
control system. ASHRAE Trans 2005;111.
[33] Shahid H, Naylor D. Energy performance assessment of a window with a
horizontal Venetian blind. Energy Build 2005;37:83643.