Building and Environment 125 (2017) 77e87

Contents lists available at ScienceDirect

Building and Environment

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

Application of artificial neural networks for determining energy-

efficient operating set-points of the VRF cooling system
Min Hee Chung a, Young Kwon Yang a, Kwang Ho Lee b, Je Hyeon Lee c, Jin Woo Moon a, *
a
School of Architecture and Building Science, Chung-Ang University, Seoul, South Korea
b
Department of Architectural Engineering, Hanbat National University, Daejeon, South Korea
c
Department of Digital Appliance R&D Team, Samsung Electronics, Suwon, South Korea

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

Article history: The aim of this study was to develop an Artificial Neural Network (ANN) model that can predict the
Received 23 January 2017 amount of cooling energy consumption for the different settings of the variable refrigerant flow (VRF)
Received in revised form cooling system's control variables. Matrix laboratory (MATLAB) and its neural network toolbox were used
21 August 2017
for the ANN model development and test performance. For the model training and performance eval-
Accepted 22 August 2017
Available online 26 August 2017
uation, data sets were collected through the field measurement. Four steps were conducted in the
development process: initial model development, input variable selection, model optimization, and
performance evaluation. In the initial model development and input variable selection process, seven
Keywords:
Artificial neural network
input variables were selected as input neurons: TEMPOUT, HUMIDOUT, TEMPIN, LOADCOOL, TEMPSA,
Predictive controls TEMPCOND, and PRESCOND. In addition, the initial model was optimized to have 2 hidden layers, 15 hidden
Refrigeration evaporation temperature set- neurons in each hidden layer, a learning rate of 0.3, and a momentum of 0.3. The optimized model
point demonstrated its prediction accuracy within the recommended level, thus proved its potential for
Supply air temperature set-point application in the control algorithm for creating a comfortable indoor thermal environment in an energy-
Condenser fluid temperature set-point efficient manner.
Condenser fluid pressure set-point © 2017 Published by Elsevier Ltd.

1. Introduction The building energy consumption including for space thermal

conditioning is related to global climate change. The consequences
Despite the delay in economic growth and in the publication of of climate change such as global warming will becoming more
government regulations for improving building energy efficiency, intense in the future, contributing to extreme climatic variances [2].
the total energy consumption by commercial and public buildings Global warming will lead to increased cooling energy consumption
in Korea is still increasing [1]. In addition, the proportion of total and peak demand in summer in most regions [3e7]. Therefore, the
building energy consumption comprised by these types of build- rate of cooling energy in total building energy consumption is ex-
ings is still increasing. Between the commercial and public sector, pected to accelerate in the future. Thus, cooling control strategies
space and water heating is the largest component of energy con- will become increasingly more important for improving energy
sumption, which accounts for 34.5% of the total energy consump- efficiency in buildings.
tion followed by cooling, cooking, lighting, and the use of In addition, the use of air-conditioning equipment has increased
appliances and other electronic plug loads of 24.2%, 19.2%, 11.9%, due to the occupants' demand for more comfortable indoor thermal
and 10.3%, respectively. In particular, the share of cooling and conditions. In particular, an increasing number of centralized air
lighting in 2013 increased by 2.4% and 1.8% more than that in 2010 conditioning systems have been applied in mid-to high-rise
respectively, while the share of space and water heating in 2013 buildings due to their higher operating efficiency and greater load-
decreased to 1.9% less than in 2010 [1]. management potential for the designated working conditions.
Furthermore, the centralized systems can be maintained more
conveniently by serving multiple spaces from one central location.
The variable refrigerant flow (VRF) systems, which is one type of
* Corresponding author.
E-mail addresses: [email protected] (M.H. Chung), [email protected] the centralized air conditioning systems, have been increasingly
(Y.K. Yang), [email protected] (K.H. Lee), [email protected] (J.H. Lee), applied for buildings such as commercial buildings, office rooms,
[email protected] (J.W. Moon).

http://dx.doi.org/10.1016/j.buildenv.2017.08.044
0360-1323/© 2017 Published by Elsevier Ltd.
78 M.H. Chung et al. / Building and Environment 125 (2017) 77e87

Nomenclature X inlet air wet-bulb temperature entering the DX coil in

the AHU,
C
ANN artificial neural network Y condenser fluid supply temperature,
C
EX AHU direct expansion air handling unit PLR part load ratio, %
SA supply air CMH flow rate of condenser fluid, m3/h
RA returned air CTEP electric power of fan in the cooling tower, kWh
EA exhausted air TTFE electric power of constant-speed fan in the cooling
OA outdoor air tower, kW
SF supply fan deltim time step in the simulation process
RF return fan compCoef cooling tower fan power correction coefficient
MB mixing box FF flow fraction, dimensionless
OU outdoor unit fpl fraction of full load power, dimensionless
CT cooling tower PCT electric power of the pumps in the cooling tower, kW
TEMPSA air handling unit supply air temperature set-point,
C gwlpm flow rate of condenser fluid, liter/minute
TEMPCOND condenser fluid temperature set-point,
C PLF fraction of full load power, W
PRESCONDcondenser fluid pressure set-point, kg/cm2 FF flow fraction, Dimensionless
POUTUNT electric power of the outdoor unit, kW m current air mass flow, kg/s
LOADCOOL cooling load of DX coil, kWh mdesign design (maximum)air flow, kg/s
LOADCT load of cooling tower, kWh fpl fraction of full load power, dimensionless
ENERGYOUTUNT energy used by the outdoor units, kWh Qtot fan power, W
TEMPCTS condenser water temperature supplied from cooling DP fan design pressure increase, pascals
tower,
C etot fan total efficiency
TEMPCTR condenser water return temperature entering cooling rair air density at standard conditions, kg/m3
tower,
C TEMPOUT average outdoor dry-bulb temperature for last 1 h,
C
FRCT condenser water volume flow rate, m3/h HUMIDOUT average outdoor relative humidity for last 1 h, %
cp specific heat for water, kcal/kg$
C TEMPIN average indoor dry-bulb temperature for last 1 h,
C
r density of water, kg/m3 SOLAR average solar radiation for last 1 h, w/m2
POUTUNT electric power of the outdoor units, kW ENERGYTOT predicted total amount of energy consumption for
QOUTUNT reference capacity of the outdoor unit, kW next 1 h, kWh
COPREF reference coefficient of performance, 4.787 W/W ENERGYCT energy used by the cooling towers, kWh
COPREF reference coefficient of performance, 4.787 W/W ENERGYPUMP energy used by the pumps, kWh
CAPFT cooling capacity ratio according to the entering ENERGYFAN energy used by the fans in the AHU, kWh
condenser fluid and inlet wet-bulb air temperatures, NIN number of input neurons
dimensionless NHN number of hidden neurons
EIRFT electric power ratio according to the entering NON number of output neurons
condenser fluid and inlet wet-bulb air temperatures, NHL number of hidden layers
dimensionless Nd number of training data sets
EIRFPLR electric power ratio according to the part load ratio, MO momentum
dimensionless LR learning rate
EIRFLPM electric power ratio according to the amount of Pi ANN predicted value
condenser fluid, dimensionless Si measured value

and hotels that have a wide load variation. The VRF systems are are generally set as constant without considering the energy-
known to provide a more comfortable and stable indoor thermal efficiency. Thus, variable settings that are more energy-efficient
environment as well as to reduce the cooling energy consumption, need to be found and applied.
actively responding to the part load variation [8]. In addition, the The aim of this study is to propose a prediction model that can
VRF system is relatively simple to set up and install by reducing the be applied for an objective system control. The proposed prediction
need for ductwork installation. model was designed to calculate the cooling energy consumption
The VRF system is operated by an outdoor unit with multiple for the different settings of the system control variables. The arti-
indoor units. There are two types of VRF system e water and air ficial neural network (ANN) theory was applied for developing the
types. Fig. 1 shows the water type of VRF system which uses cooled prediction model, in which not only surrounding indoor and out-
water in the outdoor units. The VRF system controls the refrigerant door environmental conditions, but also values of a series of control
flow rate to the multiple indoor units by electronic expansion variables such as air handling unit supply air temperature set-point
valves, whereby each zone is able to operate individually by partial (TEMPSA,
C), condenser fluid temperature set-point (TEMPCOND
cooling load.
C), and condenser fluid pressure set-point (PRES 2
COND kg/cm )
Despite the significant advantages, the more advanced control were included as input variables for calculating the predicted total
method for the VRF system still needs to be addressed. The major energy consumption for the next 1 h (ENERGYTOT, kWh). The total
control variables of the VRF systems are determined by the heu- energy, ENERGYTOT, comprises the energy used by the outdoor units
ristics of the operators. The air handling unit supply air tempera- (ENERGYOUTUNT, kWh), the energy used by the cooling towers
ture set-point (TEMPSA), condenser fluid temperature set-point (ENERGYCT, kWh), the energy used by the pumps (ENERGYPUMP,
(TEMPCOND), and condenser fluid pressure set-point (PRESCOND) kWh), and the energy used by the fans in the air handling unit
 M.H. Chung et al. / Building and Environment 125 (2017) 77e87 79

Fig. 1. Diagram of the VRF cooling system.

(AHU) (ENERGYFAN, kWh). As the prediction model included the different settings of the control variable. The algorithm will
environmental conditions as input variables, the model will be compare the cooling energy for the different settings and will
adaptively applied to the new environment. determine the optimal setting of the control variable. As the
The developed prediction model will be applied in the control optimal values for the control variables are applied, the cooling
algorithm. Fig. 2 shows the flow of the algorithm. The control al- system will create a thermally comfortable indoor environment
gorithm, which will be developed in the next study, will employ the with reduced energy consumption.
ANN model for acquiring the cooling energy consumption for the

2. Factor analysis affecting cooling energy consumption

The energy of the VRF system is principally consumed at the

outdoor units, cooling tower, and AHU fans. The factors deter-
mining the amount of energy consumption are 1) the surrounding
background conditions and 2) the setting of the system variables.
The surrounding background conditions are the outdoor and
indoor environmental factors such as outdoor temperature, hu-
midity, solar radiation, indoor temperature, and internal load that
determine the space cooling load. Among these factors, outdoor
temperature, humidity, solar radiation, and indoor temperature
were considered as input variables of the ANN model, while the
internal load was excluded because it varies according to the
building type and schedule and it is difficult to measure in the
actual buildings in real time.
In addition, the previous cooling load of the building was
selected as one of the input variables of the ANN model in order to
predict the future cooling energy more accurately based on the
assumption that the cooling load would not significantly fluctuate
within the same building. The cooling load was calculated based on
Equations (1) and (2). In the Equations, the condenser water return
temperature of the cooling tower (TEMPCTR) and the condenser
water volume flow rate (FRCT) were related to the cooling load of
the building.

Fig. 2. Thermal control algorithm. LOADCOOL ¼ LOADCT  ENERGYOUTUNT (1)

80 M.H. Chung et al. / Building and Environment 125 (2017) 77e87

LOADCT ¼ (TEMPCTS  TEMPCTR)  FRCT  cp  r (2) measurement data, indicating that required electric energy input to
meet the load is computed based on the data from field measure-
where LOADCOOL: cooling load of DX coil; LOADCT: load of cooling ment. Therefore, COP of the outdoor units dynamically change
tower; ENERGYOUTUNT: energy used by the outdoor units; TEMPCTS: under various conditions, based on the field measurement data.
is condenser water temperature supplied from cooling tower; Inlet air set-bulb temperature (X) entering the DX coil was
TEMPCTR: condenser water return temperature entering cooling excluded from the ANN model input variables because R2 of X-inlet
tower; FRCT: condenser water volume flow rate; cp: specific heat for air wet-bulb temperature entering the DX coil showed very weak
water; and r: density of water. relationship. It may be due to the fact that inlet wet-bulb temper-
The settings of the system variables are related to the outdoor ature is directly affected by the return air temperature and hu-
unit, the cooling tower fan, the pumps for circulating condenser midity from rooms, which are maintained constant during the
fluids, and the fans in the AHU. Equations (3)e(6) present the system operation. Therefore, there was no noticeable fluctuation of
electric energy consumption by the outdoor units [9,10], cooling X-inlet air wet-bulb temperature entering the DX coil during the
tower fan [11], pumps [12], and fans in the AHU [13,14], system operation. While, entering condenser fluid temperature (Y)
respectively. was selected as an input variable represented by TEMPCOND.
Coefficients in Equation (3)-1e(3)-4 are derived from both The part load ratio (PLR) is related to current cooling load of DX
technical data book and data from field measurement. Similarly, coil and current cooling capacity. However, it is not the variable that
coefficients in Equations (5) and (6)-1 are derived from field mea- can be directly measured. On the other hand, LOADCOOL can be
surement data. It should be noted that those coefficients that can directly measured by multiplying enthalpy difference between
only be applied to specific system used in the test building. The field inlet and outlet of DX coil and is can be used to represent PLR.
measurements for deriving the Coefficients were performed from Therefore, LOADCOOL is selected as an input variable instead of PLR.
June 1 to October 31, 2016 in the office building located in Seoul,
South Korea (Detailed description is given in Section 3.2.1). During CTEP ¼ TTFE  deltim/3600  compCoef (4)
the measurement, AHU discharge air temperature, the flow rate of
the condenser fluid, the condenser fluid temperature, evaporative where CTEP: current electric power of fan in the cooling tower;
temperature of refrigerant, outdoor temperature and humidity ra- TTFE: design electric power of constant-speed fan in the cooling
tio, DX coil cooling load, average indoor air temperature and solar tower (37 kW); deltim: duration of each time step; compCoef:
radiation varied to acquire the training data to be used in ANN cooling tower fan power correction coefficient.
model development. The duration of each time step (deltim) is not the input variable
of ANN model. It is simply the duration of each time-step which is
POUTUNT ¼ QOUTUNT  1/ 5 min in this study. It is used for the unit conversion between power
COPREF  CAPFT  EIRFT  EIRFPLR  EIRFLPM (3) (kW) and energy (kWh).

CAPFT ¼ 1.20869792780842 þ 0.154136903832201X PCT ¼ 0.00000000001gwlpm3 þ

 0.00228343808433177X2, when X < 35
C 0.0000008gwlpm2  0.0017gwlpm  5.5587 (5)

¼ 1.14607075335381 þ 0.149815443534369X where PCT: electric power of the pump; gwlpm: amount of
 0.00182406636707112X2  0.000020998788531433Y2 condenser fluid, which is selected as input variable represented by
 0.000594965675057201XY, When X  35
C (3)-1 PRESCOND.

EIRFT ¼ 1.5590425513611 þ 0.137130646264128X FF ¼ m/mdesign (6)

 0.00248694481270943X2 þ 0.00370594588069732Y
þ 0.000376606609216727Y2 þ 0.000182414794827632XY (3)-2 fpl ¼ 0.0023  0.684FF  1.8832FF2 þ 2.2FF3 (6)-1

EIRFPLR ¼ 0.907453819075204  3.98933921506731PRL Qtot ¼ fpl mdesignDP/(etotrair) (6)-2

þ 8.13631116983235PRLR2  4.05427749810873PLR3, when
PLR  0.5 where FF: flow fraction; m: current air mass flow in kg/s; mdesign:
design(maximum)air flow in kg/s; fpl: fraction of full load power;
¼ 1.5507PLR e 3.2029PLR2 þ 13.808PLR3  20.02PLR4, when Qtot: fan power in watts; DP: fan design pressure increase in Pas-
PLR < 0.5 (3)-3 cals; etot: fan total efficiency; rair: air density at standard conditions
in kg/m3.
EIRFLPM ¼ 1.05566666666667 The current air mass flow (m) directly affects the fan energy as
 0.0000352499999999999CMH (3)-4 shown in Equation. (6) and the fan airflow is directly affected by the
TEMPSA (AHU discharge air temperature). TEMPSA is one of the key
where POUTUNT: electric power of the outdoor unit; QOUTUNT: control variables and thus TEMPSA is selected instead of current
reference capacity of the outdoor unit; COPREF: reference coefficient airflow rate.
of performance; CAPFT: cooling capacity ratio according to the From the factor analysis using these Equations, the discharge air
entering condenser fluid and inlet wet-bulb air temperatures; temperature from the AHU, the flow rate of the condenser fluid, and
EIRFT: electric input ratio according to the entering condenser fluid the condenser fluid temperature were found to be the important
and inlet wet-bulb air temperatures; EIRFPLR: electric input ratio determinants for calculating the electric energy consumption of the
according to the part load ratio; EIRFLPM: electric input ratio ac- cooling system, in addition to the outdoor temperature and hu-
cording to the amount of condenser fluid; X: inlet air wet-bulb midity ratio, DX coil cooling load, average indoor air temperature,
temperature entering the DX coil in the AHU; Y: entering and solar radiation. Therefore, these values were selected as input
condenser fluid temperature; PLR: part load ratio; CMH: flow rate variables of the ANN model.
of condenser fluid. As stated earlier, coefficients in Equation (3)- Among those determinants, cooling tower electricity con-
1e(3)-4 are derived from both technical data book and field sumption is directly affected by outdoor temperature, humidity
 M.H. Chung et al. / Building and Environment 125 (2017) 77e87 81

ratio, and condenser fluid temperature set-point, and is indirectly temperature, indoor occupancy rate, and opening conditions. These
affected by the DX coil load. The electricity consumption of AHU fan models could respond to any external and internal environment
is indirectly affected by the AHU discharge air temperature, which changes and maintain indoor thermal comfort to control the
has a direct impact on the supply air flow through the AHU fans. heating or cooling systems.
Finally, the outdoor unit electricity is directly affected by the flow The ANN-based model can also reduce the modeling size or the
rate of the condenser fluid, condenser fluid temperature, DX coil number of monitoring sensors needed to control the comfort of an
cooling load, and the indoor air temperature. indoor environment. Song et al. [22] proposed a velocity propaga-
tion method based on a dynamic compact zonal model for a data
3. Development of ANN model center and compared the results from a VPM zonal model and
computational fluid dynamics (CFD) simulation in a full-scale data
3.1. Previous ANN applications for building thermal controls center. The results showed that the air flow and temperature dis-
tributions were in good agreement with those obtained using the
The ANN model is based on the way in which the human brain zonal method and the CFD simulation results. Using the zonal
performs computations. ANNs are empirical models that learn from model, it is possible to provide effective thermal control with
experience and generalized data for prediction, pattern recogni- limited monitoring of zonal temperature and air flow. Castilla et al.
tion, function approximation, optimization, and association [15]. [23] stated that the ANN-based model for approximated thermal
Various researchers have proposed many different ANN models. comfort evaluation has economic benefits by reducing the network
One of the most popular and widely used is the multi-layer feed- sensor size for the real-time control of HVAC systems.
forward network [16]. This model typically organizes data in mul- ANN-based models are data driven and do not require a limiting
tiple layers of modes, in which inputs enter the first layer and the assumption about the form of the modeling algorithm. Hence, they
last layer produces outputs. The middle layers, called hidden layers, predict accurate data for controlling a complex thermal environ-
separate the input layer from the output layer. Each node in one ment with efficient monitoring sensors in real time.
layer is connected to every node on the next layer. Each connection
is associated with a numeric number called a weight. The weights
3.2. Process of ANN model development
of each connection, the best number of hidden layers, and the
number of nodes in a hidden layer are determined by many trials
Fig. 3 shows the four major steps for developing the ANN model
using different samples in the training process. The performance of
in this study. The first step was conducted to develop an initial
the model is checked on a validation set. The optimal model is
model. The ANN model was organized with an initial composition
selected with the smallest validation error. The test set for finally
and learning methods. The second step was performed to deter-
checking the overall performance is used to estimate the accuracy
mine the meaningful input variables. The relationship between the
of the selected model. These processes have a significant impact on
input and output variables was statistically analyzed to find the
the accuracy of the prediction.
input variables that have a strong relationship with the output
Numerous successful practical applications of ANNs have been
variables. Input variables with a strong relationship with the output
found in diverse fields. The application of ANNs for building ther-
variables were chosen as input neurons of the model.
mal engineering has become increasingly more common in recent
The third step was performed for optimizing the ANN model.
years. Various authors have applied ANNs to control heating,
Model parameters such as the number of hidden layers (NHL), the
ventilation, and air-conditioning (HVAC) systems to improve in-
number of hidden neurons (NHN), the learning rate (LR), and mo-
door thermal comfort and to reduce energy consumption. Ferreira
mentum (MO) were optimized to produce accurate outputs. Opti-
et al. [17] achieved energy saving and constant thermal comfort
mization was performed in a coupled fashion, so that a series of
from an existing HVAC system control experiment using a model
NHL and NHN were tested together, followed by a series of LR and
based predictive control methodology. The energy savings were
MO together. The ANN model was revised to employ parameter
over 50%. Moon and Jung used an ANN-based model to improve the
values by which the most accurate outputs were produced. The
thermal comfort and energy efficiency of accommodation buildings
final step was conducted to evaluate the prediction performance of
[18]. Two ANN model based algorithms were proposed to control
the optimized model. The applicability of the proposed model was
the set point temperature for external climate and to determine the
setback value and period for an unoccupied period. These proposed
algorithms showed better performance in terms of energy effi-
ciency and thermal comfort than the conventional control
algorithm.
The ANN-based model is not sequential or necessarily deter-
ministic as a building control method. In other words, it can
respond to unexpected changes in real time but cannot execute
programmed instructions. Li et al. [19] developed the ANN-based
control strategy for a direct expansion air conditioning system
that simultaneously considers indoor air temperature and humid-
ity. A command following and disturbance rejection test were
conducted to examine the performance of the proposed controller.
It has been reported that, in a direct expansion air conditioning
system, the ANN-base controller can be used to control the indoor
air temperature and humidity by changing the compressor speed
and supply fan speed. Morel et al. [20] proposed a heating
controller for predicting the external climate conditions, including
the air temperature, solar radiation, and indoor temperature. Moon
et al. [21] used an ANN-based model for optimum control of a
double skin façade depending on the external climate, indoor air Fig. 3. Development process of the ANN model.
82 M.H. Chung et al. / Building and Environment 125 (2017) 77e87

demonstrated by this performance evaluation. A detailed expla- Table 1

nation of each step is as follows. Normalized range by input.

Input Normalized range

TEMPOUT 10e40
C
3.2.1. Initial model development (1st step)
HUMIDOUT 0e100%
The initial structure of the ANN model consists of three layers: TEMPIN 15e30
C
the input layer, the hidden layer, and the output layer, as shown in SOLAR 0e1000w/m2
Fig. 4. The inputs were selected as variables relevant to the amount LOADCOOL 0e3000 kWh
of energy consumption, as mentioned in Section 2. The initial input TEMPSA 8e20
C
TEMPCOND 15e40
C
neurons comprise eight variables: (1) average outdoor dry-bulb
PRESCOND 0.25e1.25 kg/cm2
temperature for last 1 h (TEMPOUT,
C); (2) average outdoor rela-
tive humidity for last 1 h (HUMIDOUT, %), (3) average indoor dry-
bulb temperature for last 1 h (TEMPIN,
C), (4) average solar radi-
The Levenberg-Marquart algorithm was used as the learning
ation for last 1 h(SOLAR, w/m2), (5) cooling load for last 1 h
method. The initial LR and MO used 0.6 and 0.2, respectively, which
(LOADCOOL, kWh), (6) air handling unit supply air temperature set-
concurred with those obtained in a previous study [26]. The ANN
point (TEMPSA,
C), (7) condenser fluid temperature set-point
model was developed using the Neural Network toolbox in MAT-
(TEMPCOND,
C), and (8) condenser fluid pressure set-point (PRE-
LAB. A total of 200 training data sets were applied in this model
SCOND, kg/cm2). The last variable, PRESCOND, is related to the amount
based on Equation (8).
of condenser fluid. All inputs and outputs in the initial model were
normalized to the measured data range as listed in Table 1.
Nd ¼ (NHN  (NIN þ NON)/2)2 [27] (8)
In the initial model, one hidden layer was used, employing 17
hidden neurons based on Equation (7), which suggests the normal
¼ (17-(8 þ 1)/2)2
number of hidden neurons [24,25]. NHN is the number of hidden
neurons and NIN is the number of input neurons.
¼ 158 sets
NHN ¼ 2NIN þ 1 (7)
The sliding-window method was used for the data management
of the training data sets; advanced data elements expire at a time
The output neuron represents the predicted total energy con-
based on the runoff time series. The sliding-window method can
sumption for the next 1 h (ENERGYTOT, kWh) in a cooling system.
prevent a multiplicative overhead of data [28] and maintain more
The ENERGYTOT refers tothe sum of energy used by the outdoor
flexible data queries without any loss in accuracy [29].
units (ENERGYOUTUNT, kWh), the energy used by the cooling towers
All data sets were obtained from an office building as the test
(ENERGYCT, kWh), the energy used by the pumps (ENERGYPUMP,
building, as shown in Fig. 5. The test building is located in Seoul,
kWh), and the energy used by the fans in the AHU (ENERGYFAN,
South Korea and has a gross floor area of 22,660 m2. This 10-story
kWh).
office building was constructed in 2015. The monitoring period
The sigmoid function for hidden neurons and the pure-linear
was from June 1 to October 31, 2016. The VRF system was used as a
function for the output neuron were chosen as transfer functions.
cooling system. One CT, 11 AHUs, and 27 OUs composed the VRF
system for covering the whole building.
The 11 monitoring zones were selected among low, middle, and
high floors. Monitored variables are listed in Table 2. The data
measured at 1-min intervals are averaged for 1 h. In particular, the
environmental data such as outdoor dry-bulb temperature and
outdoor relative humidity was measured at the weather station
located in the test building site and data for solar radiation was
collected on the roof of the test building. The indoor temperature
data was collected from the 11 sensors installed in the entrance of
the return duct of each AHU. Measured from the actual site and
building, data was supported to represent the actual environmental
conditions of the building.

3.3. Input variable selection (2nd step)

The initial ANN model employed a series of input variables

based on the mathematical analysis in the previous step. In the
second step, the statistical analysis was conducted to determine the
meaningful input variables. Among the eight input variables, five
input variables (TEMPOUT, HUMIDOUT, TEMPIN, SOLAR, and LOAD-
COOL), which were not the system operating variables, were evalu-
ated to determine their relationship with the output variable. The
coefficient of determination (R2) between each input variable and
output variable was calculated for identifying input variables hav-
ing a strong relationship with the output variable. The initial ANN
model was revised to employ these input variables as the final input
neurons. Three-hundred and ninety-six data sets were prepared for
Fig. 4. Initial ANN model. the analysis.
 M.H. Chung et al. / Building and Environment 125 (2017) 77e87 83

Fig. 5. Typical floor plan of the test building.

Table 2
Measurement variables and the unit of measuring time.

Variables The unit of measuring

time

Outdoor dry-bulb temperature 1 min

Outdoor relative humidity 1 min
Indoor dry-bulb temperature 1 min
Solar radiation 1 min
Condenser fluid supply temperature of cooling 1 min
tower
Condenser fluid volume flow rates 1 min
Air handling unit supply air temperature set-point 1 h
Condenser fluid temperature set-point 1 h
Condenser fluid pressure set-point 1 h
Energy used by the outdoor units 1 h
Energy used by the cooling towers 1 h
Energy used by the pumps 1 h
Energy used by the fans in the AHU 1 h

3.4. Model optimization and performance evaluation (3rd and 4th

steps)

To produce more accurate and stable prediction results, the

structure and learning method of the ANN model with the selected
input variables were optimized in the third step. A parametrical Fig. 6. Parametrically tested components and values for ANN optimization.

optimization process was conducted in a coupled fashion. The

prediction performances of models with a series of hidden layers
After optimization, the prediction performance of the optimized
(NHL) and hidden neurons (NHN) were tested together and a series
ANN model was evaluated in the fourth step. Prediction perfor-
of learning rates (LR) and momentum (MO) were evaluated
mance was evaluated based on three terms: coefficient of deter-
together. When the first two parameters (i.e., NHL and NHN) were
mination (R2), coefficient of variation root mean square error
tested to find the optimal numbers, the other two parameters (i.e.,
(CVRMSE), and mean bias error (MBE) between the Pi and Si values.
LR and MO) were fixed at the initial values (0.6 learning rate and 0.2
One hundred new data sets were used for the performance tests.
moment). Once the optimal values of the first two parameters were
determined, the optimal values were applied and the next two
parameters were tested. Fig. 6 summarizes the parametrical values 4. Result analysis
that were used to optimize each component. The statistical corre-
lation (R2) between the predicted (Pi) and simulated (Si) values was 4.1. Input variable selection
calculated for each parametrical value, and the values that pro-
duced the smallest R2 values were determined to be the optimal The coefficients of determination (R2) between the five input
parameter values. Two-hundred new data sets were collected for variables (TEMPOUT, HUMIDOUT, TEMPIN, SOLAR, and LOADCOOL) and
the optimization, using the same method as that explained in the one output variable (ENERGYTOT) are shown in Fig. 7. Among the
Section 3.2. five input variables, three variables (TEMPOUT, TEMPIN, and
84 M.H. Chung et al. / Building and Environment 125 (2017) 77e87

Fig. 7. Coefficients of determination (R2) between the five input variables and one output variable.
 M.H. Chung et al. / Building and Environment 125 (2017) 77e87 85

LOADCOOL) demonstrated a relatively closer relationship of 0.4429, 4.3. Performance evaluation

0.3489, and 0.2856, respectively. On the other hand, the other two
variables (HUMIDOUT and SOLAR) showed very weak results of Prediction performance of the optimized ANN model was
0.0575 and 0.0849, respectively. evaluated using three terms:R2, CVRMSE, and MBE. Fig. 9 shows the
The R2 values of the three variables (TEMPOUT, TEMPIN, and coefficient of determination (R2) between the Pi and Si from the
LOADCOOL) implied a weak correlation with the output variable, optimized ANN model. The R2 was 0.8136. In addition, the CVRMSE
even though they presented relatively higher R2. When the data and MBE between Pi and Si were calculated as 11.28% and 1.18%,
sets were collected from the test building, the exterior and interior respectively. According to the ASHRAE Guideline14-Measurement
thermal conditions such as TEMPOUT, HUMIDOUT, TEMPIN, SOLAR, of energy and demand savings [30], the prediction accuracy of
and LOADCOOL were simultaneously changed. Thus, the correlation the prediction model can be verified with a CVRMSE value under
between each input and output seemed to be weak, as expected. If 30% and MBE value under 10%. Thus, prediction accuracy of the
only one variable (e.g., TEMPOUT) had been changed while other ANN model in this study is verified.
variables were fixed, the correlation between TEMPOUT and ENER- The analysis results regarding R2, CVRMSE, and MBE between Pi
GYTOT would be greater than the result in this study. and Si indicate that the ANN model properly predicts the cooling
Table 3 summarizes the CVRMSE (%) between the Pi from the energy consumption for the different combinations of the system
test building and Si from the initial ANN model according to the operating mode as well as the given exterior and interior thermal
different input variable applications. Four ANN models with conditions. Based on prediction accuracy, the model demonstrated
different combinations of the two input variables (HUMIDOUT and a potential for application in the control algorithm for suggesting
SOLAR), which presented lower R2, were evaluated. The lowest the optimal system operating mode of the cooling system.
CVRMSE value (15.49%) was deduced when the HUMIDOUT was
included in the input variables and SOLAR was excluded from the 5. Conclusion
input variables.
Even though the solar radiation is deeply related to the amount In this study, an ANN based predictive model was developed for
of cooling energy, it could be an obstructive factor to accurately and calculating the cooling energy consumption for the different set-
stably predict the amount of cooling energy in several reasons. tings of the cooling system's control variables. Four steps were
First, the number of measuring point for the solar radiation is conducted in the development process: initial model development,
limited. Normally horizontal (or vertical) solar radiation is input variable selection, model optimization, and performance
measured on a specific spot such as a roof, but the cooling load or evaluation. The summary of findings is a follows.
energy is differentiated by the orientation of the building zone.
Thus, the amount of solar radiation may not be a direct determinant (1) Initial model development: the initial model was designed to
of the amount of energy consumption in actual building. Second, have nine input neurons in the input layer, including five
the solar radiation sometimes continues to fluctuating by the sky exterior and interior thermal conditions (TEMPOUT, TEMPIN,
conditions such as in partly cloudy conditions. The instantaneous or LOADCOOL, HUMIDOUT, and SOLAR) and three system oper-
even average solar radiation could be a reason to make the model ating variables (TEMPSA, TEMPCOND, and PRESCOND). In addi-
prediction unstable. Third, the effect of solar radiation to the tion, NHL was 1, NHN was 17, LR was 0.6, and MO was 0.2.
amount of cooling energy consumption could be different due to (2) Input variable selection: in the correlation analysis between
the different thermal mechanisms in the building. For example, the the input variables and output variable, the coefficient of
storing and discharging ratio of the solar energy is different for determination (R2) was relatively higher by the TEMPOUT,
buildings. A specific model, which is suitably tuned to one building, TEMPIN, and LOADCOOL than by the HUMIDOUT and SOLAR. In
may work improperly to the other building. addition, the ANN model employing the HUMIDOUT, while
From the comparison of R2 and CVRMSE, the four input vari-
ables (TEMPOUT, TEMPIN, and LOADCOOL, and HUMIDOUT) were
included in a series of input neurons. Therefore, eight input vari- Table 3
CVRMSE (%) according to the input (SOLAR and HUMIDOUT) selection.
ables including four system operating variables (TEMPREF, TEMPSA,
TEMPCOND, and PRESCOND) comprised the input layer. Case TEMPOUT TEMPIN LOADCOOL SOLAR HUMIDOUT CVRMSE(%)

1 O O O O O 21.35
4.2. Model optimization 2 O O O O X 21.84
3 O O O X X 15.93
4 O O O X O 15.49
Table 4 summarizes the CVRMSE (%) results between the Pi and
Si for the different combinations of NHL and NHN. The CVRMSE for
a different NHL (1 … 5) and NHN (10 … 20) ranged between 12.31
Table 4
and 24.64. The case for the model employing 2 NHL and 15 NHN
CVRMSE (%) between the Pi and Si for various NHL and NHN.
presented the least CVRMSE. Therefore, the optimal composition
regarding NHL and NHN was determined to be 2 and 15, CVRMSE (%) NHL
respectively. 1 2 3 4 5
Table 5 summarizes the CVRMSE results for the different LRs (0.1 NHN 10 14.89 14.13 18.16 15.34 24.64
… 1.0) and MOs (0.1 … 1.0). The values for NHL and NHN were fixed 11 16.78 13.67 14.39 14.08 17.66
as 2 and 15, as determined in the previous step. The least RMSE was 12 13.24 17.55 14.57 13.48 24.43
presented by a model with 0.3LR and 0.3MO with 11.28% CVRMSE. 13 13.09 14.16 12.49 15.44 20.03
14 14.44 14.34 13.93 15.63 13.12
Thus, the optimized ANN model employed these values, 0.3 and 0.3,
15 12.72 12.31 12.83 18.91 13.08
for the learning rate and momentum, respectively. 16 21.64 15.06 14.47 12.61 13.76
Fig. 8 shows the ANN model after input variable selection and 17 15.58 14.38 16.22 12.69 14.00
optimization. The final model had seven input neurons: TEMPOUT, 18 13.19 13.36 13.59 15.75 12.95
HUMIDOUT, TEMPIN, LOADCOOL, TEMPSA, TEMPCOND, and PRESCOND, 19 16.24 13.93 15.96 13.72 13.66
20 14.20 13.85 14.63 14.36 14.32
with NHL of 2; NHN of 15; LR of 0.3; and MO of 0.3.
86 M.H. Chung et al. / Building and Environment 125 (2017) 77e87

Table 5
CVRMSE (%) between the Pi and Si for various numbers of LR and MO.

CVRMSE (%) LR

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

MO 0.1 15.90 17.58 16.41 14.20 13.84 16.00 15.25 15.25 12.94 12.32
0.2 12.79 14.36 15.63 12.27 15.18 12.31 12.85 13.36 12.91 13.30
0.3 15.78 13.95 11.28 14.16 12.79 13.71 18.32 15.86 14.54 15.08
0.4 12.94 13.58 15.09 14.81 16.33 13.04 14.38 14.15 17.00 15.40
0.5 13.30 12.58 12.93 13.82 13.14 16.97 16.57 11.99 17.40 12.21
0.6 13.88 12.51 11.67 13.57 13.10 15.06 15.42 13.03 15.22 13.36
0.7 13.09 11.71 14.81 15.42 14.28 12.98 12.88 13.29 14.93 13.72
0.8 12.93 14.57 13.69 15.62 13.31 12.57 13.20 13.05 15.33 14.98
0.9 16.08 15.53 14.11 15.39 13.89 14.36 13.97 13.48 13.04 13.31
1.0 13.66 13.53 11.99 12.68 13.07 11.38 16.34 15.27 12.30 14.30

Fig. 9. Coefficient of determination (R2) between Pi and Si from the optimized ANN
model.

prediction accuracy of the optimized ANN was demonstrated

under the recommended level.

From the development process employed in this study, the

optimized ANN model showed prediction accuracy and potential
for application to the control algorithm for determining the most
Fig. 8. Optimized ANN model.
energy-efficient cooling system operation. Further study is needed
to develop the control algorithm embedded in the ANN model
excluding SOLAR, demonstrated the most accurate predic- developed in this study and to test its thermal and energy perfor-
tion performance with the least CVRMSE (%) between the Pi mance. In particular, the algorithm needs to have supporting
from the test building and Si from the initial ANN model. functions to exclude the exceptional prediction errors of the ANN
Thus, the initial model was organized to employ eight input model for the stable system operation.
variables in the input layer: four exterior and interior ther- In addition, the proposed model in this study excluded the
mal conditions (TEMPOUT, TEMPIN, LOADCOOL, and HUMID- amount of internal load from the input variables because the
OUT) and three system operating variables (TEMPSA, proper method to estimate the real-time amount of internal load
TEMPCOND, and PRESCOND). from the actual building has not been introduced yet. Once a
(3) Model optimization: in the prediction performance analysis trustful method to calculate the heat gain from always varying
of the ANN model with different combinations of NHL and occupants, use of lighting fixtures, and equipments is developed,
NHN as well as LR and MO, the least CVRMSE (%) was ac- further study is required to consider the amount of internal load as
quired when the NHL was 2, the NHN was 15, the LR was 0.3, one of the input variable of the prediction model.
and the MO was 0.3. Thus, the model was optimized to Lastly, since the proposed model does not directly mean to be
employ these values for the structure and learning method. applied to the all types of VRF system, new models are required for
(4) Performance evaluation: through the prediction perfor- the different types of VRF systems. For example, a model and an
mance evaluation using the R2, CVRMSE, and MBE, the algorithm working independently for each thermal zone and AHU
needs to be investigated. Moreover, models for the air type VRF
systems are necessary to be investigated as well.
 M.H. Chung et al. / Building and Environment 125 (2017) 77e87 87

Acknowledgements [15] I.D. Basheer, M. Hajmeer, Artificial neural networks: fundamentals,

computing, design, and application, J. Microbiol. Meth. 43 (2000) 3e31.
[16] G. Zhang, B.E. Patuwo, M.Y. Hu, Forecasting with artificial neural networks:
This research was supported by the Department of Digital the state of the art, Int. J. Forecast. 14 (1) (1998) 35e62.
Appliance R&D Team, Samsung Electronics and by the Basic Science [17] P.M. Ferreira, A.E. Ruano, S. Silva, E.Z.E. Conceica ~o, Neural networks based
Research Program through the National Research Foundation (NRF) predictive control for thermal comfort and energy savings in public buildings,
Energy Build. 55 (2012) 238e251.
of Korea funded by the Ministry of Science, ICT and Future Planning [18] J.W. Moon, S.K. Jung, Development of a thermal control algorithm using
(grant number 2015R1A1A1A05001142). artificial neural network models for improved thermal comfort and energy
efficiency in accommodation buildings, Appl. Therm. Eng. 103 (2016)
1135e1144.
References [19] N. Li, L. Xia, D. Shiming, X. Xu, M.Y. Chanm, Dynamic modeling and control of a
direct expansion air conditioning system using artificial neural network, Appl.
[1] Korea Energy Economics Institute, 2014 Energy Consumption Survey, 2015. Energy 91 (2012) 290e300.
[2] G.A. Meehl, C. Tebaldi, More intense, more frequent, and longer lasting heat [20] N. Morel, M. Bauer, M. El-Khoury, J. Krauss, Neurobat, a predictive and
waves in the 21st Century, Science 305 (2004) 994e997. adaptive heating control system using artificial neural networks, Int. J. Sol.
[3] J.A. Dirks, W.J. Gorrissen, J.H. Hathaway, D.C. Skorski, M.J. Scott, T.C. Pulsipher, Energy 21 (2001) 161e201.
M. Huang, Y. Liu, J.S. Rice, Impacts of climate change on energy consumption [21] J.W. Moon, J.H. Lee, Y. Yoon, S. Kim, Determining optimum control of double
and peak demand in buildings: a detailed regional approach, Energy 79 (2015) skin envelope for indoor thermal environment based on artificial neural
20e32. network, Energy Build. 69 (2014) 175e183.
[4] H. Wang, Q. Chen, Impact of climate change heating and cooling energy use in [22] Z. Song, B.T. Murray, B. Sammakia, A dynamic compact thermal model for data
buildings in the United States, Energy Build. 82 (2014) 428e436. center analysis and control using the zonal method and artificial neural net-
[5] J. Huang, K.R. Gurney, The variation of climate change impact on building works, Appl. Therm. Eng. 62 (2014) 48e57.
energy consumption to building type and spatiotemporal scale, Energy 111 
[23] M. Castilla, J.D. Alvarez, M.G. Ortega, M.R. Arahal, Neural network and poly-
(2016) 137e153. nomial approximated thermal comfort models for HVAC systems, Build. En-
[6] M. Santamouris, Cooling the buildings- past, present and future, Energy Build. viron. 59 (2013) 107e115.
128 (2016) 617e638. [24] D. Datta, S.A. Tassou, D. Marriott, Application of neural networks for the
[7] Y. Petri, K. Caldeira, Impacts of global warming on residential heating and prediction of the energy consumption in a supermarket, in: CLIMA 2000,
cooling degree-days in the United States, Sci. Rep. 5 (2015) 12427. http://dx. 1997, pp. 98e107. Brussels (Belgium).
doi.org/10.1038/srep12427. [25] J. Yang, H. Rivard, R. Zmeureanu, On-line building energy prediction using
[8] D. Zhang, X. Zhang, J. Liu, Experimental study of performance of digital vari- adaptive artificial neural networks, Energy Build. 37 (2005) 1250e1259.
able multiple air conditioning system under part load conditions, Energy [26] J.W. Moon, K. Kim, H. Min, ANN-based prediction and optimization of cooling
Build. 43 (2011) 1175e1178. system in hotel rooms, Energies 8 (2015) 10775e10795.
[9] The U.S. D.O.E, EnergyPlus Engineering Reference V. 8.5; the Reference to [27] S.A. Kalogirou, M. Bojic, Artificial neural networks for the prediction of the
EnergyPlus Calculations, Chapter 15.11.1 Variable Refrigerant Flow Heat Pump energy consumption of a passive solar building, Energy 25 (2001) 479e491.
Model (System Curve Based Model), 2011, p. 1121. [28] M. Datar, A. Gionis, P. Indyk, R. Motwani, Maintaining stream statistics over
[10] Samsung Electorincs Co, DVM S Water Technical Data Book, 2015. sliding windows, Siam J. Comput. 31 (2002) 1794e1813.
[11] Society of Heating Air-Conditioning and Sanitary Engineers of Japan [29] Y. Chabchoub, G. He brail, Sliding Hyper Log Log: estimating cardinality in a
commissioning WG, HVACSIMþ(J) User Manual, 1998. data stream over a sliding window, in: 2010 IEEE International Conference on
[12] The U.S. D.O.E, EnergyPlus Engineering Reference V. 8.5; the Reference to Data Mining Workshops, IEEE, 2010, pp. 1297e1303.
EnergyPlus Calculations, Chapter 15.17.3 Variable Speed Pump, 2011, p. 1257. [30] American Society of Heating, Refrigerating, and Air-conditioning Engineer.
[13] The U.S. D.O.E., EnergyPlus Engineering Reference V. 8.5; the Reference to ASHRAE Guideline14-measurement of Energy and Demand Savings, ASHRAE
EnergyPlus Calculations, Chapter 15.9.2.3 Simulation, 2011, p. 1015. Inc., 2002.
[14] The U.S. D.O.E., EnergyPlus InputOutput Reference V. 8.5; the Reference to
EnergyPlus Calculations, Chapter 1.41.3 Fan:VariableVolume, 2011, p. 1907.