buildings

Article
The Hourly Energy Consumption Prediction by KNN for
Buildings in Community Buildings
Goopyo Hong 1 , Gyeong-Seok Choi 2 , Ji-Young Eum 2 , Han Sol Lee 2 and Daeung Danny Kim 3, *

1 Department of Architectural Engineering, Kangwon National University, Samcheok 25913, Korea

2 Department of Building Energy Research, Korea Institute of Civil Engineering and Building Technology,
Goyang 10223, Korea
3 Architectural Engineering Department, King Fahd University of Petroleum and Minerals (KFUPM),
Dhahran 31261, Saudi Arabia
* Correspondence: [email protected]

Abstract: With the development of metering technologies, data mining techniques such as machine
learning have been increasingly used for the prediction of building energy consumption. Among
various machine learning methods, the KNN algorithm was implemented to predict the hourly
energy consumption of community buildings composed of several different types of buildings. Based
on the input data set, 10 similar hourly energy patterns for each season in the historic data sets
were chosen, and these 10 energy consumption patterns were averaged. The prediction results were
analyzed quantitatively and qualitatively. The prediction results for the summer and fall were close
to the energy consumption data, while the results for the spring and winter were higher than the
energy consumption data. For accuracy, a similar trend was observed. The values of CVRMSE for the
summer and fall were within the acceptable range of ASHRAE guidelines 14, while higher values of
CVRMSE for the spring and winter were observed. In sum, the total values of CVRMSE were within
the acceptable range.

Citation: Hong, G.; Choi, G.-S.; Eum, Keywords: hourly energy consumption; KNN algorithm; energy pattern; community buildings
J.-Y.; Lee, H.S.; Kim, D.D. The Hourly
Energy Consumption Prediction by
KNN for Buildings in Community
Buildings. Buildings 2022, 12, 1636. 1. Introduction
https://doi.org/10.3390/
With notable concern for CO2 emission increases, energy consumption by buildings
buildings12101636
has been significantly increasing [1–3]. According to the “Energy Statistics Handbook in
Academic Editor: Benedetto Nastasi 2020” provided by the Korea Energy Agency, a large amount of energy is accounted for
Received: 8 September 2022
in building sectors [4]. Thus, much attention has been paid to a reduction of energy con-
Accepted: 7 October 2022
sumption by buildings in South Korea (hereafter Korea). There have been many attempts
Published: 9 October 2022
to reduce or optimize building energy consumption. Focusing on passive or active design
strategies, many studies have performed investigations to improve energy efficiency in
Publisher’s Note: MDPI stays neutral
buildings by enhancing the thermal performance of building envelopes, replacing them
with regard to jurisdictional claims in
with advanced mechanical systems or installing renewable energy systems [5–13]. While
published maps and institutional affil-
these design strategies have played an important role in building energy management,
iations.
the effects on the reduction of building energy consumption vary from 3% to 10% of the
total building energy consumption [14,15]. In addition, a better energy performance can be
expected only when these design strategies are implemented in the early building design
Copyright: © 2022 by the authors.
stage.
Licensee MDPI, Basel, Switzerland. With the rapid development of information and communication technologies, building
This article is an open access article energy management and prediction becomes a fundamental key for improving energy
distributed under the terms and efficiency as well as for reducing building energy consumption [16]. By implementing
conditions of the Creative Commons metering technologies, the specific information on building operation and energy con-
Attribution (CC BY) license (https:// sumption data are more available for analyses of understanding energy use behaviors in
creativecommons.org/licenses/by/ buildings [17]. The large amount of data (i.e., hourly or sub-hourly energy performance
4.0/). data, etc.) collected by some systems such as building automation systems and building

Buildings 2022, 12, 1636. https://doi.org/10.3390/buildings12101636 https://www.mdpi.com/journal/buildings

Buildings 2022, 12, 1636 2 of 15

energy management systems can be used to predict their dynamic interactions affecting
building energy consumption [18,19]. For example, Jairo et al. have analyzed the inter-
relationship between the electrical loads of each space in relation to the total building
energy consumption for a full week of operation from the data obtained by the energy
management systems [20]. Similarly, Alam and Devjani have implemented the building
management systems to obtain the data at 5 min interval for one year to analyze the energy
patterns of educational buildings [21]. Thus, it is inevitable to understand a huge number
of data sets of energy consumption for developing effective energy saving strategies.
Traditionally, statistical methods have been employed to handle sets of data for iden-
tifying the patterns of building energy consumption. However, it is difficult to handle a
huge amount of data. According to the study by Fan et al., data mining techniques are
effective to clarify the massive data sets [22]. Therefore, the present study exists to predict
the energy consumption of buildings by utilizing the data mining technique. Three years
(2018–2020) of energy consumption in community buildings were used to train the mining
technique such as the KNN (K-nearest neighbor) algorithm. By recognizing data patterns
through the calculation of distances between test and training data sets, the KNN method is
increasingly used for the analysis of large historical data sets [23]. The community buildings
were composed of offices, a gymnasium, exhibition hall, etc. The energy consumption data
of the community buildings were collected. By implementing the KNN algorithm, the
seasonal energy consumption patterns were identified by the KNN method. The clustered
energy consumption patterns were averaged, and then, these averaged patterns were
verified with the collected energy consumption data qualitatively and quantitatively. By
using the suggested method in the present study, the energy consumption of buildings in
a community can be forecasted faster than other machine learning techniques. Currently,
much attention has been paid to energy-sharing strategies in community buildings [24–26].
Therefore, the prediction of energy consumption by the KNN algorithm can be helpful to
develop effective energy-sharing strategies for community buildings.

2. Literature Review
The use of machine learning algorithms has been increasingly implemented to analyze
data as well as to continuously learn judgments or predictions for the future [27]. In the
case of building energy predictions, steady-state approaches such as averaged monthly
or hourly calculation methods have been generally used instead of dynamic approaches
due to high computational cost. In the comparison of the energy demand of buildings by
energy simulations, Fantozzi et al. mentioned that dynamic approaches can provide more
accurate prediction than those obtained by the average monthly calculation method [28].
The accuracy of the hourly model was also pointed out by the study of Ballarini et al. [29].
Considering this point, it is necessary to implement machine learning techniques for
achieving the accuracy of the energy consumption prediction by handling a huge amount
of data sets.
For the present study, the energy demand patterns were recognized by the machine
learning methods. To identify energy usage patterns, clustering techniques have been
commonly used, which are generally unsupervised learning techniques for recognizing
inherent data patterns [30]. Typically, these clustering techniques can be categorized as
several methods: partition, hierarchical, and density-based [31]. As one of the partitioning
clustering algorithms, the K-means algorithm has been widely used in several studies for
the energy pattern identification due to its high efficiency [30]. However, the use of K-
means algorithm had difficultly handling massive time series of similar electric load profiles,
and the calculation can be computationally expensive [32]. Considering these difficulties,
several researchers proposed supervised data mining-based approaches. According to
the study of Xiong and Yao, common supervised classification algorithms are Artificial
Neural Networks (ANN), Support Vector Machine (SVM), K-nearest neighbor (KNN), and
so on [27].
Buildings 2022, 12, 1636 3 of 15

Among these algorithms, the ANN algorithm is the most typical regression and
commonly used technique in which the concept of this algorithm resembles the human
brain by using interconnected processors, but this method requires a large number of
parameters and massive training data sets [33]. In the case of the SVM algorithm, it is widely
applied for pattern classification, which computes the linear regression function [34,35].
However, it requires a large number of historical data sets and design inputs [36]. As one of
the supervised machine learning algorithms, the KNN can be also used to recognize data
patterns. Contrary to other machine learning algorithms, the KNN only defines the number
of the nearest neighbors not requiring other parameters [27]. Because of these convenient
issues, the KNN algorithm can be increasingly applied for pattern recognition. Several
studies revealed the high accuracy and computational efficiency of the data classification
by the KNN algorithm [37,38].
Generally, the KNN algorithm calculates the distance between test and training sam-
ples and then returns k closest samples by using the linear search method to find the exact
k nearest neighbors [37]. This means that the KNN algorithm can map the relationship
between independent and dependent variable spaces. The computational complexity
of the KNN algorithm is proportional to the size of the training data set for each test
sample [39,40]. The sample distance can be calculated by using the Minkowski distance
equation [36]:
n 1/p
d = ( ∑ | xi − yi | p ) (1)
i =1

where xi and yi are the coordinates of the sample points in multidimensional space. d is the
absolute distance (the Manhattan distance), when p is 1. When p is 2, d is the linear distance
(the Euclidean distance). If the sample distribution is unbalanced, the KNN algorithm
weighs points regarding the inverse of the distance to minimize the impact of the distance
between the test and training samples [36]. For example, closer neighbors of a query point
have a more significant impact on the result than farther neighbors [41]. Since the region of
the k-neighborhood is determined by the values of k, the classification performance can be
easily affected by outliers of the smaller or larger value of k [42]. In addition, the k data
points closest to the test point can be obtained by calculating the distances [27]. Since the
number k is the number of data points needed, the clustering model can be oversensitive
to sample points near the test point, if the number k is too small. If it is too large, a poor
clustering result can be produced [27]. The best number k can be determined by cross
validation by considering its accuracy.

3. Methodology
In the present study, a prediction strategy based on the classification of hourly energy
consumption patterns was proposed by using the KNN algorithm for several buildings
in the A community buildings, where these buildings are located in Gyeonggi-do, South
Korea (Figure 1). While most previous studies have focused on single building function
such as residential or commercial buildings, this study focuses on energy consumption
patterns of community buildings composed of different building types. The A community
buildings include six different buildings such as office, auditorium, gymnasium, and so on.
Three years (2018–2020) of hourly electricity consumption data obtained from i-SMART
of Power planner in KEPCO were utilized [43]. These data were classified seasonally. For
each season, the maximum daily energy consumption data of the A community buildings
were identified. In addition, the daily weather data were recognized. By implementing
the KNN algorithm, the energy consumption patterns for each season were identified and
validated with the historic energy consumption data.
Buildings 2022, 12, 1636 4 of 15
Buildings 2022, 12, x FOR PEER REVIEW 4 of 15

Figure 1. The buildings in the A community.

Figure 1. The buildings in the A community.
3.1. Data
3.1. Data Preparation
Preparation
This study analyzed the hourly electricity data for three years (2018–2020) of the A
This study analyzed the hourly electricity data for three years (2018–2020) of the A
community. Among these data, the hourly energy consumption data for two years (2018
community. Among these data, the hourly energy consumption data for two years (2018
and 2019) were used as training data sets, while the data for the year of 2020 were used as
and 2019) were used as training data sets, while the data for the year of 2020 were used
test data sets. In addition, three years of TMY weather data for the location of the A com-
as
munitydata
test sets. In
buildings addition,
were used as three
input years of TMY
data. Table weather
1 shows the data for the
summary of location
the KNNof the A
community
method. The description of buildings in the community is presented in Table 2. Theof
buildings were used as input data. Table 1 shows the summary the KNN
en-
ergy consumption of each building is presented in Table 3. As the area of the building energy
method. The description of buildings in the community is presented in Table 2. The
increased, high
consumption ofenergy consumption
each building was observed.
is presented in TableIn the case
3. As theofarea
the community
of the buildingcenter,
increased,
the energy
high energyconsumption
consumption waswashigher than that
observed. In of
thethe auditorium,
case while the area
of the community center,of the
the energy
community center
consumption waswas smaller
higher thanthanthatthat of the
of the auditorium.while
auditorium, This can
thebe caused
area of thebycommunity
the
longer was
center building occupancy
smaller schedule
than that of theofauditorium.
the community center.
This can Inbeaddition,
caused by thethe
totallonger
energybuilding
consumption
occupancy in 2019 of
schedule wastheslightly
community more increased
center. In than that in
addition, the2018 (about
total energy 2%), while
consumption in
about a 5% decrease in the total energy consumption was observed
2019 was slightly more increased than that in 2018 (about 2%), while about a 5% compared with thatdecrease
in 2020. This can be caused by limited personal and social activities by COVID-19.
in the total energy consumption was observed compared with that in 2020. This can be
caused
Table 1. by limited
Summary of personal and social activities by COVID-19.
the KNN algorithm.

Characteristics
Table 1. Summary of the KNN algorithm.
Type of building Office, auditorium, gymnasium, accommodation, community center, exhibition
Time interval Hour, day, year Characteristics
Metric RMSE and CVRMSE
Type of buildingElectrical
Office, auditorium,(Mainly
consumption gymnasium,
heatingaccommodation,
and cooling) community center, exhibition
Input
Time interval Weather: Ambient temperature,
Hour, day, year solar radiation, humidity ratio, wind speed

Metric RMSE and CVRMSE

Table 2. Building description.
Electrical consumption (Mainly heating and cooling)
Input Building Floors Gross Floor Area (m2)
Weather: Ambient temperature, solar radiation, humidity ratio, wind speed
Office 3 floors and 1 basement floor 3917
Auditorium 2 floors and 1 basement floor 1809
Gymnasium 2 floors 1357
Table 2. Building description.
Accommodation 2 floors and 1 basement floor 2743
Community center 4 floors 1485
Building Floors Gross Floor Area (m2 )
Exhibition 2 floors and 1 basement floor 1103
Office 3 floors and 1 basement floor 3917
Auditorium 2 floors and 1 basement floor 1809
Gymnasium 2 floors 1357
Accommodation 2 floors and 1 basement floor 2743
Community center 4 floors 1485
Exhibition 2 floors and 1 basement floor 1103
Buildings 2022, 12, x FOR PEER REVIEW 5 of 15
Buildings 2022,
Buildings 12, 1636
2022, 12, x FOR PEER REVIEW 55of
of 15
15

Table 3. Energy consumption of the A community buildings.

Table 3.
Table 3. Energy
Energy consumption
consumption of
of the
the A
A community
community buildings.
buildings.
Total
Community Total
Office Auditorium Gymnasium Accommodation Community Exhibition Energy
Total
Year Office Auditorium Gymnasium Accommodation Center
Community Exhibition Energy
Year (kWh)
Office (kWh)
Auditorium (kWh)
Gymnasium (kWh)
Accommodation
Center (kWh)
Exhibition Consumption
Energy
Year
(kWh) (kWh) (kWh) (kWh) (kWh)
Center
(kWh) Consumption
(kWh) (kWh) (kWh) (kWh)
(kWh)
(kWh)
(kWh) (kWh)
Consumption
2018 1,216,539 891,488 772,653 1,188,113 897,536 413,645 (kWh)
(kWh)
5,379,974
2018
2019
2018 1,216,539
1,236,482
1,216,539 891,488
906,103
891,488 772,653
785,319
772,653 1,188,113
1,207,590
1,188,113 897,536
912,250
897,536 413,645
420,426
413,645 5,379,974
5,468,170
5,379,974
2019
2020
2019 1,236,482
1,183,300
1,236,482 906,103
867,130
906,103 785,319
751,542
785,319 1,207,590
1,207,590
1,207,590 912,250
873,013
912,250 420,426
402,343
420,426 5,468,170
5,232,980
5,468,170
2020
2020 1,183,300
1,183,300 867,130
867,130 751,542
751,542 1,207,590
1,207,590 873,013
873,013 402,343
402,343 5,232,980
5,232,980
As shown in Figure 2, the data collection began from 01 January 2018 to 31 December
2020.As
AsIn shown
general,in
shown
Figure
inthe
2, theenergy
building
Figure
data collection begancan
consumption
2, the data collection
from
began from be01 Januaryby2018
affected
01 January 2018someto 31 December
to considerable
31 December
2020. Insuch
factors general, the building
as building designenergy consumption
variables, occupants’ can be affected
behaviors, by some
building considerable
systems opera-
2020. In general, the building energy consumption can be affected by some considerable
factors
tion and such
theirasefficiency,
building design
and variables,
climatic occupants’
conditions [21]. behaviors,
For this building
study, the systems
hourly opera-
electrical
factors such as building design variables, occupants’ behaviors, building systems operation
tion and
consumptiontheir efficiency,
data, mainly and climatic
heating and conditions
cooling,[21].[21].
were For For this
collected study,
for each the hourly
building, electrical
and their efficiency, and climatic conditions this study, the hourly and then,
electrical
consumption
these data were data, mainlytoheating
summed reduceand cooling, were
computational collected
errors for as
as well each building,
toeach
promote and then,
computa-
consumption data, mainly heating and cooling, were collected for building, and
these data
tional were summed
efficiency. As a to reduce
result, 26,280 computational
records were errors as and
available, wellthese
as to were
promote
brokencomputa-
down
then, these data were summed to reduce computational errors as well as to promote
tional efficiency.
seasonally
computational (Figure As3).a Table
result,
efficiency. As4a26,280
result,records
presents were available,
the seasonal
26,280 records energy and these
consumption
were available, andweredata
these broken
of the
were down
com-
broken
seasonally
munity (Figure
buildings 3).
for Table
three 4 presents
years. Amongthe seasonal
the energy energy consumption
consumption
down seasonally (Figure 3). Table 4 presents the seasonal energy consumption data of data data
for of
threethe com-
years,
the
munity
the
community buildings
biggest energy
buildingsfor three
consumptionyears. Among
was Among
for three years. the
observed energy
theinenergy consumption
2019, and data
this trenddata
consumption for
was for three
alsothree
shownyears,
in
years,
the
mostbiggest
seasons
the biggest energy consumption
for winter. was
exceptconsumption
energy was observed
observed in in 2019,
2019, and
and this
this trend
trend was
was also shown in
also shown in
most
most seasons except for
seasons except for winter.
winter.

Figure 2. Energy consumption of the A community buildings for three consecutive years.
Figure 2. Energy
Figure 2. Energy consumption of the
consumption of the A
A community
community buildings
buildings for
for three
three consecutive
consecutive years.
years.

(a) Spring.
(a) Spring.
Figure 3. Cont.
 Buildings
Buildings 2022,
2022, 12, 163612, x FOR PEER REVIEW 6 of 15 6 of 15

(b) Summer.

(c) Fall.

(d) Winter.
Figure 3. Seasonal hourly energy consumption data of the A community buildings for three years.
Figure 3. Seasonal hourly energy consumption data of the A community buildings for three years.
Table 4. Seasonal energy consumption and the hourly average of the A community buildings for
Table 4. Seasonal energy consumption and the hourly average of the A community buildings for
three years.
three years.
Season Year Total Energy Consumption (kWh) Hourly Average (kWh)
Season Year 2018 1,150,590
Total Energy Consumption (kWh) 521Average (kWh)
Hourly
Spring
2018 2019 1,249,676
1,150,590 566
521
2020 1,149,997 520
Spring 2019 1,249,676 566
2018 1,166,838 528
Summer 2020 2019 1,149,997
1,176,430 532 520
2018 2020 1,098,648
1,166,838 497 528
Summer 2019 2018 1,176,219
1,176,430 538 532
Fall 2019 1,197,061 556
2020 1,098,648 497
2020 1,208,927 553
2018 2018 1,176,219
2,222,270 1028
538
Fall Winter 2019 2019 2,099,024
1,197,061 961 556
2020 2020 2,088,060
1,208,927 1009 553
2018 2018 5,752,180
2,222,270 656 1028
Total 2019 5,862,931 673
Winter 2019 2,099,024 961
2020 5,545,632 631
2020 2,088,060 1009
2018 5,752,180 656
Total 2019 5,862,931 673
2020 5,545,632 631
Buildings 2022, 12,
Buildings 2022, 12, x1636
FOR PEER REVIEW 7 7of
of 15
15
Buildings 2022, 12, x FOR PEER REVIEW 7 of 15

Based
Based onon the
the hourly
hourly average
average energy
energy consumption
consumption for for each
each season,
season, four
four dates
dates were
were
Based on the hourly average energy consumption for each season, four dates were
selected
selected for
for each
each season.
season. As
As shown
shown inin Figure
Figure 4,
4,hourly
hourly energy
energy consumption
consumption on on the
the selected
selected
selected for each season. As shown in Figure 4, hourly energy consumption on the selected
dates
dates for
for three
three years
years are
are presented.
presented. Even
Even though
though there
there was
was aa little
little difference
difference in
in energy
energy
dates for three years are presented. Even though there was a little difference in energy
consumption,
consumption, similar
similar patterns
patterns in
in each
each season
season for
for three
three years
years are
are observed
observed visually,
visually, except
except
consumption, similar patterns in each season for three years are observed visually, except
for
for the
the spring
spring season.
season. Moreover,
Moreover, thethe daily
daily weather
weather data
data (outdoor
(outdoor temperature)
temperature) andand the
the
for the spring season. Moreover, the daily weather data (outdoor temperature) and the
cooling and heating energy consumption in each season for three years are presented in
cooling and heating energy consumption in each season for three years are presented in
cooling and heating energy consumption in each season for three years are presented in
Figure
Figure 5.
5. As
As presented,
presented, the
the energy
energy consumption
consumption patterns
patterns for
for heating
heating andand cooling
cooling were
were
Figure 5. As presented, the energy consumption patterns for heating and cooling were
different,
different, but
but similar
similar patterns
patterns within the heating
within the heating and
and cooling
cooling were
were identified.
identified.
different, but similar patterns within the heating and cooling were identified.

(a) Spring. (b) Summer.

(a) Spring. (b) Summer.

(c) Fall. (d) Winter.

(c) Fall. (d) Winter.
Figure 4. Daily energy consumption for each season of the A community buildings for three years.
Figure 4.
Figure 4. Daily
Daily energy
energy consumption for each
consumption for each season
season of
of the
the A
A community
community buildings
buildings for
for three
three years.
years.
Spring/Fall Summer Winter
Spring/Fall Summer Winter
30
30
25
25
°C°C

10
10
0
0
- 10
- 10
1 4 8 12 16 20 24
1 4 8 12
Hour 16 20 24
Hour
(a). The outdoor temperature distribution
(a). TheSummer
Spring/Fall
outdoor temperature distribution Spring/Fall Winter
Spring/Fall Summer Spring/Fall Winter
2500 2000
2500 2000
consumption

consumption
consumption

consumption

2000 1500
2000 1500
1500
(kWh)

(kWh)

1500
(kWh)

1000
(kWh)

1000 1000
Energy

1000
Energy

500
Energy

Energy

500 500
500
0 0
01 4 8 12 16 20 24 01 4 8 12 16 20 24
1 4 8 12
Hour 16 20 24 1 4 8 12
Hour 16 20 24
Hour Hour
(b). Cooling energy consumption. (c). Heating energy consumption.
(b). Cooling energy consumption. (c). Heating energy consumption.
Figure 5. Daily weather data and cooling and heating energy consumption for three years.
Figure 5.
Figure 5. Daily
Daily weather
weather data
data and
and cooling and heating
cooling and heating energy
energy consumption
consumption for
for three
three years.
years.
 Buildings 2022, 12, 1636 8 of 15

3.2. The KNN Algorithm

Generally, the proper number of k can be selected after some experiments by using
an optimized tool [44]. In the present study, the distances between the patterns in the
Buildings 2022, 12, x FOR PEER REVIEW training data sets were computed to find the optimal number of 8 ofk.15Among the nearest
patterns, the optimal number of k nearest patterns was extracted from the historical data
sets. In addition, the weighed averaging values of each pattern were created for forecasting
3.2. Thefuture
KNN Algorithm
values [45]. To assess the forecast model of electricity consumption pattern by the
Generally, the proper number
KNN algorithm, of k can be selected
two performance after some
assessments experiments
were employed:by using
root mean square error
an optimized tool [44]. In the present study, the distances between the patterns
(RMSE) and coefficient of variation of the root mean squared error (CVRMSE). in the While RMSE
trainingindicates
data sets were computed to find the optimal number of k. Among the nearest
the size of the variation, CVRMSE indicates the deviation degree of error of the
patterns, the optimal number of k nearest patterns was extracted from the historical data
training data for the comparison between different data sets [46]. These equations are
sets. In addition, the weighed averaging values of each pattern were created for forecast-
expressed
ing future values [45].as Equations
To assess the(2) and (3)
forecast [46]:of electricity consumption pattern by
model
the KNN algorithm, two performance assessments were semployed: root mean square er-
ror (RMSE) and coefficient of variation of the root mean squared
∑in=1 Yerror (CVRMSE).
ˆ 2 While


i − Yi
RMSE indicates the size of the variation, CVRMSE
RMSEindicates
= the deviation degree of error (2)
n
of the training data for the comparison between different data sets [46]. These equations
are expressed as Equations (2) and (3) [46]: s
2
∑in=1 Yi − Yiˆ ∑in=1 Yi


∑ (𝑌𝑌 − 𝑌𝑌𝑖𝑖 )𝑛𝑛 ^ 2
� 𝑖𝑖=1 𝑖𝑖 =
RMSE =CVRMSE / (2) (3)
𝑛𝑛 n n
where Yi is the ith actual ∑ ˆ is the ith predicted value. n is the total number of
value,
𝑛𝑛
𝑌𝑌𝑖𝑖^ )2 Y∑
(𝑌𝑌𝑖𝑖 −and i 𝑛𝑛𝑖𝑖=1 𝑌𝑌𝑖𝑖 (3)
CVRMSE = � 𝑖𝑖=1 /
data in the KNN algorithm model. 𝑛𝑛 𝑛𝑛

where Yi is the ith actual value, and 𝑌𝑌𝑖𝑖^ is the ith predicted value. n is the total number of
4. Results
data in the KNN algorithm model.
4.1. Forecasting Results
4. Results For the KNN algorithm, the periodicity of the variable was found (p = 24). Thus,
4.1. Forecasting Results distance was employed and calculated by Equation (1). First, the outdoor
the Euclidean
Fortemperature
the KNN algorithm, the periodicity
patterns of the variable
for each season was found (pAs
were classified. = 24).
can Thus, the
be shown in Figure 6, 10
Euclidean distance
similar was employed
outdoor and calculated
temperature patternsby Eq. the
(i.e., 1. First, the outdoor
k number = 10) temperature
for two years (2018 and 2019)
patternswere
for each season were
identified basedclassified.
on testAsdata
can beonshown in Figuredates
the selected 6, 10 similar
for the outdoor
year 2020. In the spring,
temperature patterns (i.e., the k number = 10) for two years (2018 and 2019) were identified
10 similar hourly outdoor temperature patterns for two years were chosen based on the
based on test data on the selected dates for the year 2020. In the spring, 10 similar hourly
daily outdoor temperature on 20 April in 2020 (blue curve). By averaging these chosen 10
outdoor temperature patterns for two years were chosen based on the daily outdoor tem-
peratureoutdoor temperature
on 20 April patterns,
in 2020 (blue theaveraging
curve). By prediction of chosen
these the hourly outdoor
10 outdoor temperature pattern for
temper-
ature patterns,
the springthe prediction
was obtainedof the(red
hourly outdoor
curve). In atemperature
similar way, pattern
similar foroutdoor
the springtemperature patterns
was obtained
on the(red curve).
other In a similar
seasons way, similarfor
were classified outdoor temperature
two years (2018 patterns
and 2019) on the
based on the test data
other seasons
sets on were
theclassified
selectedfor two years
dates of 12(2018 and 2019)
August, based onand
16 October, the test
20 data sets onAs a result, it can be
January.
the selected dates of 12 August, 16 October, and 20 January. As a result,
seen that the predictions for the summer, fall, and winter showed a similar it can be seen that pattern with the
the predictions for the summer, fall, and winter showed a similar pattern with the test
test data on these specific dates. In the case of spring, a somewhat difference of the outdoor
data on these specific dates. In the case of spring, a somewhat difference of the outdoor
temperature
temperature between thebetween theand
prediction prediction andonthe
the test data thetest datadate
specific on wasthe observed.
specific date was observed.

15 27, Oct, 2018

07, Nov,2019
14, Apr, 2019
14 11, Oct, 2018
23, Apr, 2018
08, Apr, 2019
13 26, Oct, 2019
11, Apr, 2018
10, Nov,2019
12 12, Nov,2019
True for 20, Apr, 2020
Estimated Electricity consumption for 20, Apr, 2020
°C

11

10

1 4 7 10 13 16 19 22
Hour
(a) Spring.

Figure 6. Cont.
 Buildings 2022, 12, 1636 9 of 15
Buildings 2022, 12, x FOR PEER REVIEW 9 of 15

17, Jul, 2018

26, Aug,2019
30 13, Jul, 2018
23, Jul, 2019
18, Aug,2019
14, Jul, 2018
16, Aug,2019
18, Aug,2019
28 01, Sep,2018
03, Jul,2018
True for 12, Aug, 2020
Estimated Electricity consumption for 12, Aug, 2020
°C
26

24

22

1 4 7 10 13 16 19 22
Hour

(b) Summer.
06, Nov,2018
16 10, Nov,2018
28, Mar,2018
28, Apr,2019
04, May,2018
15 11, Nov,2019
03, Apr,2018
09, Nov,2018
14 24, Apr,2018
07, Nov,2018
True for 16, Oct, 2020
Estimated Electricity consumption for 16, Oct, 2020
13
°C

12

11

10

9
1 4 7 10 13 16 19 22
Hour

(c) Fall.
6

2
°C

12, Jan,2018
06, Dec,2018
0 20, Feb,2018
11, Jan,2019
22, Jan,2018
20, Jan,2019
24, Feb,2018
-2 19, Feb,2018
14, Jan,2018
21, Feb,2018
True for 20, Jan, 2020
Estimated Electricity consumption for
20, Jan, 2020
1 4 7 10 13 16 19 22
Hour

(d) Winter.

Figure Figure
6. The outdoor temperature
6. The outdoor patterns from
temperature the KNN
patterns fromalgorithm.
the KNN algorithm.

As with Asthe outdoor

with temperature
the outdoor prediction,
temperature the hourly
prediction, theenergy
hourlyconsumption pat-
energy consumption patterns
terns for each season were identified (Figure 7). Based on the test data on 20 April in 2020
for each season were identified (Figure 7). Based on the test data on 20 April in 2020 (Blue
(Blue curve),
curve),1010similar
similarpatterns
patternsin in
2018 and
2018 2019
and were
2019 clustered
were for the
clustered spring
for the season.
spring season. The red
curve presents the average value of the chosen 10 similar energy consumption patterns in
2018 and 2019. For the other seasons, 10 similar patterns based on the test data on specific
dates were classified, and then, these were averaged. As shown in Figure 7, it can be seen
uildings 2022, 12, x FOR PEER REVIEW 10 of 15

Buildings 2022, 12, 1636

The red curve presents the average value of the chosen 10 similar energy consumption 10 of 15
patterns in 2018 and 2019. For the other seasons, 10 similar patterns based on the test data
on specific dates were classified, and then, these were averaged. As shown in Figure 7, it
can bethat
seenthethat
test the
datatest data
on the on thedates
selected selected
in thedates
summerin the
andsummer
fall were and
closefall were
to the close to
predicted
the predicted energy consumptions
energy consumptions by the KNN by the KNN
algorithm. algorithm.
While there wasWhile there was
little difference in little differ-
the energy
ence in the energyduring
consumption consumption during
the afternoon the afternoon
between between
the test data and thethe test datainand
prediction the pre-
the winter,
dictiona similar
in the winter, a similar
pattern was pattern
observed. was
In the observed.
spring, In the
a different spring,
pattern a different
of energy pattern of
consumption
energybetween
consumption between
the test data on thethe test data
selected on the
date and selected date
the prediction and the prediction
was observed. This was also was
observed. This was also observed in the outdoor temperature prediction in the spring.
observed in the outdoor temperature prediction in the spring.

1100 27, Oct, 2018

07, Nov,2019
14, Apr, 2019
1000 11, Oct, 2018
23, Apr, 2018
08, Apr, 2019
Energy consumption(kWh)

900 26, Oct, 2019

11, Apr, 2018
10, Nov,2019
800 12, Nov,2019
True for 20, Apr, 2020
Estimated Electricity consumption for 20, Apr, 2020
700

600

500

400

300

1 4 7 10 13 16 19 22
Hour

(a) Spring.
1200 17, Jul, 2018
26, Aug,2019
13, Jul, 2018
23, Jul, 2019
18, Aug,2019
14, Jul, 2018
1000 16, Aug,2019
Energy consumption(kWh)

18, Aug,2019
01, Sep,2018
03, Jul,2018
True for 12, Aug, 2020
800 Estimated Electricity consumption for 12, Aug, 2020

600

400

1 4 7 10 13 16 19 22
Hour

(b) Summer.

Figure 7. Cont.
Buildings 2022, 12, x2022,
Buildings FOR12,PEER
1636 REVIEW 1111
of of
15 15

1200 06, Nov,2018

10, Nov,2018
28, Mar,2018
28, Apr,2019
04, May,2018
11, Nov,2019
1000 03, Apr,2018

Energy consumption(kWh)
09, Nov,2018
24, Apr,2018
07, Nov,2018
True for 16, Oct, 2020
800 Estimated Electricity consumption for 16, Oct, 2020

600

400

1 4 7 10 13 16 19 22
Hour

(c) Fall.
12, Jan,2018
2000 06, Dec,2018
20, Feb,2018
11, Jan,2019
1800 22, Jan,2018
20, Jan,2019
24, Feb,2018
Energy consumption(kWh)

1600 19, Feb,2018

14, Jan,2018
21, Feb,2018
True for 20, Jan, 2020
1400 Estimated Electricity consumption for 20, Jan, 2020

1200

1000

800

600

1 4 7 10 13 16 19 22
Hour

(d) Winter.

FigureFigure
7. Energy consumption
7. Energy patterns
consumption byby
patterns thethe
KNN
KNNalgorithm.
algorithm.

4.2. Evaluation of Forecasting Accuracy by the KNN Method

4.2. Evaluation of Forecasting Accuracy by the KNN Method
Table 5 summarizes the forecasting accuracy by using the assessment indicators.
Table 5 summarizes
As shown below, the the
RMSE forecasting
and CVRMSEaccuracy byofusing
values four the assessment
seasons and the indicators.
total energyAs
shown below, the RMSE and CVRMSE values of four seasons and the
consumption are presented. The RMSE and CVRMSE of the total energy consumptiontotal energy con-
sumption
were are presented.
139.44 kWh andThe RMSE
24.15%, and CVRMSE
respectively. of the total
According to theenergy
ASHRAEconsumption were
guidelines 14,
139.44the
kWh and 24.15%,
acceptable respectively.
range of the CVRMSE According
value can tobethe ASHRAE as
differentiated guidelines 14, the
hourly, daily, andac-
ceptable range[47].
monthly of the
By CVRMSE
considering value can beacceptable
the hourly differentiated
rangeasofhourly,
30%, thedaily,
CVRMSEand values
monthly
[47]. By considering the hourly acceptable range of 30%, the CVRMSE values of the
of the summer and fall were below 30%, while the values of the spring and winter sum-
were
somewhat higher than or close to 30%. For the total energy consumption forecasting, the
mer and fall were below 30%, while the values of the spring and winter were somewhat
CVRMSE was below 30%.
higher than or close to 30%. For the total energy consumption forecasting, the CVRMSE
was below 30%.

Table 5. Summary of the indicator for RMSE and CVRMSE.

Season Spring Summer Fall Winter Total

RMSE CV-RMSE RMSE CV-RMSE RMSE CV-RMSE RMSE CV-RMSE RMSE CV-RMSE
Metric
(kWh) (%) (kWh) (%) (kWh) (%) (kWh) (%) (kWh) (%)
Mean 382.93 34.29 89.46 12.21 94.32 13.78 211.32 29.74 139.44 24.15
Buildings 2022, 12, 1636 12 of 15

Table 5. Summary of the indicator for RMSE and CVRMSE.

Season Spring Summer Fall Winter Total

Metric RMSE CV-RMSE RMSE CV-RMSE RMSE CV-RMSE RMSE CV-RMSE RMSE CV-RMSE
(kWh) (%) (kWh) (%) (kWh) (%) (kWh) (%) (kWh) (%)
Mean 382.93 34.29 89.46 12.21 94.32 13.78 211.32 29.74 139.44 24.15

5. Discussion
By using the KNN method, the energy consumption patterns of the community
buildings including several building functions for two years were clustered. Based on
the k value of 10, 10 hourly energy consumption profiles for each season were chosen.
In addition, their averaging values (i.e., the prediction) were compared with the energy
consumption data of the community buildings. Similar with the present study, building
electrical demand was predicted by using the KNN algorithm in the study of Gomez-
Omella et al. [45]. With 122 residential and small office buildings, they utilized two different
versions of the KNN algorithm for an effective analysis of a large historical time series of
electricity consumption data sets. While there was a little difference in the prediction results
and computational efficiency between the two KNN algorithms, both the KNN methods
provided good accuracy. In the case of the study by Liu et al., two step clustering approaches
including density-based spatial clustering application and the k-means algorithm were
used to identify daily electricity usage patterns of three office buildings [17]. However,
it was difficult for the density-based clustering application to select the values of radius
and the minimum number of observations, and the fast KNN algorithm was implemented.
As shown in the results of the present study, the CVRMSE values of the prediction were
somewhat higher than that obtained from the previous KNN studies. It can be seen that
the overestimated value of the present study was caused by the different schedules and
functions of the various types of buildings such as office, auditorium, gymnasium, and
so on in the community buildings, while several studies focused on one or two different
building functions. Moreover, the results predicted by the KNN method are presented in
Table 6 by using the Mean Bias Error (MBE) (Equation (4)).

2
∑n Yi − Yiˆ
MBE = i=1 n × 100 (4)
∑i=1 Yi

where Yi is the ith actual value and Yiˆ is the ith predicted value. n is the total number of
data in the KNN algorithm model.

Table 6. Summary of the indicator for the MBE.

Season Spring Summer Fall Winter Total

MBE (%) 12.5 −2.8 4.3 10.7 11.2

Considering the acceptable error of the MBE (±10), the value of the spring was not
acceptable, while the values of the summer and fall were acceptable. In the case of winter,
the MBE value was slightly over the acceptable range. As observed in the CVRMSE values,
spring carried the biggest prediction errors. It can be seen that the errors in the spring were
caused by very irregular weather profiles in each year. This added greater probabilistic
patterns to the prediction by the KNN method as opposed to the KNN predictions with
clear weather characteristics of other seasons. As a result, this led to the overestimated
MBE values in the total energy consumption prediction.
Moreover, the KNN algorithm was used for various purposes. For example, Xiong
and Yao implemented this method to establish a personalized adaptive thermal comfort
environment for occupants’ indoor environmental preferences [27]. In their study, the KNN-
based thermal comfort model proved to have good accuracy after a large amount of data
training. In addition, they calculated the distance for the parameters of the thermal comfort.
Buildings 2022, 12, 1636 13 of 15

Martinez et al. used the KNN method to handle time series patterns seasonally [44]. They
applied different KNN schemes for different seasons. For the chiller system optimization,
Ho and Yu utilized the KNN regression to discover optimal strategies [48]. Using the
k number of 3, the design parameters were selected to identify specific strategies to the
existing chiller system. Most previous studies pointed out the importance of the k number
since the classification performance is quite affected by the k number [36]. For the present
study, the k value was 10. To find the best k value, several k values were investigated.
However, the k numbers above 10 were not investigated due to computational efficiency.
In addition, the authors in the previous studies also stated that different scale combinations
can influence the accuracy of forecasting. Even though hourly energy consumption for
two years was used for the study, this seemed to be insufficient in that the overestimated
CVRMSE values for some seasons were obtained. For further study, it is required to
consider the prediction performance by different distances and different KNN schemes for
better accurate prediction. In addition, it is also necessary to compare the prediction results
with other machine learning techniques with the same data used for the present study.
By training three years of energy consumption data sets with weather parameters, the
KNN method clustered similar patterns of energy consumptions with them in 2020. Even
though it seemed that the predicted results were overestimated, this can be overcome with
more historic data sets. In addition, it can be seen that the data sets were classified where
similar weather characteristics were observed. Thus, it can expected to predict building
energy consumption in more than a year with the data sets of weather information.

6. Conclusions
As with the development of mining techniques, the use of machine learning methods
has been significantly increased. For the present study, the hourly energy consumption data
were predicted by the machine learning technique, especially the KNN algorithm, which
explores the whole training data sets for clustering based on the input test sample. Thus, the
energy consumption pattern classification by using the KNN algorithm can recognize each
pattern and then provide targeted predictions within each pattern. By using three years
of hourly energy consumption data, the hourly energy consumption of the A community
buildings composed of several types of buildings was predicted by the KNN algorithm.
The outcomes of the study were as follows:
(a) For the KNN algorithm, the periodicity of the variable was set at 24. Thus, Euclidean
distance was chosen for clustering the hourly energy consumptions for each season.
(b) By investigating several k numbers, 10 was selected. Using this k number, 10 similar
hourly energy consumption patterns for two years (2018 and 2019) were clustered
based on the test data on seasonal specific dates in 2020. Then, these 10 clustered
hourly energy consumption patterns were averaged for the prediction. As a result,
the averaged energy consumption for the summer and fall were close to the test data,
while the prediction results of the spring and winter were a little higher than the test
data.
(c) The accuracy of the prediction by using the KNN method was assessed by the RMSE
and CVRMSE quantitatively. The CVRMSE values of the summer and fall ranged from
12–13%, which was within the acceptable range provided by ASHRAE guidelines 14.
In the case of spring and winter, the values of CVRMSE were higher than 30% or close.
In sum, the total CVRMSE was about 24%.
Considering the outcome of the present study, the total value of CVRMSE was within
the acceptable range of ASHRAE guidelines 14. However, the CVRMSE values of the
prediction results were still high in that the accuracy seemed to be overestimated. This
can be caused by insufficient historic data sets and design inputs of the various types of
buildings in the A community buildings. Further study will include design inputs of vari-
ous building functions for more accurate prediction by using the KNN method. Moreover,
it is necessary to compare the prediction results with other machine learning techniques
to enhance the accuracy of the KNN method. As shown in the results, the prediction by
Buildings 2022, 12, 1636 14 of 15

using the KNN algorithm was obtained with three years of energy consumption data of the
community buildings composed of various building functions. By improving the accuracy,
the suggested method can be implemented to develop energy sharing strategies for the
community buildings.

Author Contributions: G.H. contributed to the concept and design of the research and drafted the
manuscript. G.-S.C., J.-Y.E. and H.S.L. collected and analyzed the data. D.D.K. wrote the manuscript.
All authors have read and agreed to the published version of the manuscript.
Funding: This work is supported by the Korea Agency for Infrastructure Technology Advancement
(KAIA) grant funded by the Ministry of Land, Infrastructure and Transport (grant RS-2019-KA153277).
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Oh, M.; Jang, K.M.; Kim, Y. Empirical analysis of building energy consumption and urban form in a large city: A case of seoul,
south korea. Energy Build. 2021, 245, 111046. [CrossRef]
2. Lei, L.; Chen, W.; Wu, B.; Chen, C.; Liu, W. A building energy consumption prediction model based on rough set theory and deep
learning algorithms. Energy Build. 2021, 240, 110886. [CrossRef]
3. Pi, Z.X.; Li, X.H.; Ding, Y.M.; Zhao, M.; Liu, Z.X. Demand response scheduling algorithm of the economic energy consumption in
buildings for considering comfortable working time and user target price. Energy Build. 2021, 250, 111252. [CrossRef]
4. Agency, K.E. Energy Statistics Handbook. 2020. Available online: https://www.Energy.Or.Kr/web/kem_home_new/new_main.
Asp (accessed on 7 September 2022).
5. Calero, M.; Alameda-Hernandez, E.; Fernández-Serrano, M.; Ronda, A.; Martín-Lara, M.Á. Energy consumption reduction
proposals for thermal systems in residential buildings. Energy Build. 2018, 175, 121–130. [CrossRef]
6. Golbazi, M.; Aktas, C.B. Energy efficiency of residential buildings in the U.S.: Improvement potential beyond iecc. Build. Environ.
2018, 142, 278–287. [CrossRef]
7. Park, B.; Srubar, W.V.; Krarti, M. Energy performance analysis of variable thermal resistance envelopes in residential buildings.
Energy Build. 2015, 103, 317–325. [CrossRef]
8. Hachem-Vermette, C. Multistory building envelope: Creative design and enhanced performance. Sol. Energy 2018, 159, 710–721.
[CrossRef]
9. Athienitis, A.K.; Barone, G.; Buonomano, A.; Palombo, A. Assessing active and passive effects of façade building integrated
photovoltaics/thermal systems: Dynamic modelling and simulation. Appl. Energy 2018, 209, 355–382. [CrossRef]
10. Zhang, R.; Nie, Y.; Lam, K.P.; Biegler, L.T. Dynamic optimization based integrated operation strategy design for passive cooling
ventilation and active building air conditioning. Energy Build. 2014, 85, 126–135. [CrossRef]
11. Huide, F.; Xuxin, Z.; Lei, M.; Tao, Z.; Qixing, W.; Hongyuan, S. A comparative study on three types of solar utilization technologies
for buildings: Photovoltaic, solar thermal and hybrid photovoltaic/thermal systems. Energy Convers. Manag. 2017, 140, 1–13.
[CrossRef]
12. McLarty, D.; Brouwer, J.; Ainscough, C. Economic analysis of fuel cell installations at commercial buildings including regional
pricing and complementary technologies. Energy Build. 2016, 113, 112–122. [CrossRef]
13. Gautam, K.R.; Andresen, G.B. Performance comparison of building-integrated combined photovoltaic thermal solar collectors
(bipvt) with other building-integrated solar technologies. Sol. Energy 2017, 155, 93–102. [CrossRef]
14. Kim, D.-B.; Kim, D.D.; Kim, T. Energy performance assessment of hvac commissioning using long-term monitoring data: A case
study of the newly built office building in south korea. Energy Build. 2019, 204, 109465. [CrossRef]
15. Suh, H.S.; Kim, D.D. Energy performance assessment towards nearly zero energy community buildings in south korea. Sustain.
Cities Soc. 2019, 44, 488–498. [CrossRef]
16. Capozzoli, A.; Piscitelli, M.S.; Brandi, S.; Grassi, D.; Chicco, G. Automated load pattern learning and anomaly detection for
enhancing energy management in smart buildings. Energy 2018, 157, 336–352. [CrossRef]
17. Liu, X.; Ding, Y.; Tang, H.; Xiao, F. A data mining-based framework for the identification of daily electricity usage patterns and
anomaly detection in building electricity consumption data. Energy Build. 2021, 231, 110601. [CrossRef]
18. Yu, Z.; Fung, B.C.M.; Haghighat, F. Extracting knowledge from building-related data—A data mining framework. Build. Simul.
2013, 6, 207–222. [CrossRef]
19. Ding, Y.; Brattebø, H.; Nord, N. A systematic approach for data analysis and prediction methods for annual energy profiles: An
example for school buildings in norway. Energy Build. 2021, 247, 111160. [CrossRef]
20. Diaz-Acevedo, J.A.; Grisales-Noreña, L.F.; Escobar, A. A method for estimating electricity consumption patterns of buildings to
implement energy management systems. J. Build. Eng. 2019, 25, 100774. [CrossRef]
21. Alam, M.; Devjani, M.R. Analyzing energy consumption patterns of an educational building through data mining. J. Build. Eng.
2021, 44, 103385. [CrossRef]
Buildings 2022, 12, 1636 15 of 15

22. Fan, C.; Xiao, F.; Li, Z.; Wang, J. Unsupervised data analytics in mining big building operational data for energy efficiency
enhancement: A review. Energy Build. 2018, 159, 296–308. [CrossRef]
23. Yoon, Y.R.; Lee, Y.R.; Kim, S.H.; Kim, J.W.; Moon, H.J. A non-intrusive data-driven model for detailed occupants’ activities
classification in residential buildings using environmental and energy usage data. Energy Build. 2022, 256, 111699. [CrossRef]
24. Li, Y.; Hu, S.; Hoare, C.; O’Donnell, J.; García-Castro, R.; Vega-Sánchez, S.; Jiang, X. An information sharing strategy based on
linked data for net zero energy buildings and clusters. Autom. Constr. 2021, 124, 103592. [CrossRef]
25. Henni, S.; Staudt, P.; Weinhardt, C. A sharing economy for residential communities with pv-coupled battery storage: Benefits,
pricing and participant matching. Appl. Energy 2021, 301, 117351. [CrossRef]
26. Herenčić, L.; Kirac, M.; Keko, H.; Kuzle, I.; Rajšl, I. Automated energy sharing in mv and lv distribution grids within an energy
community: A case for croatian city of križevci with a hybrid renewable system. Renew. Energy 2022, 191, 176–194. [CrossRef]
27. Xiong, L.; Yao, Y. Study on an adaptive thermal comfort model with k-nearest-neighbors (knn) algorithm. Build. Environ. 2021,
202, 108026. [CrossRef]
28. Fantozzi, F.; Romeo, C.; Salvadori, G.; Leccese, F.; Gazzarri, F. Simulation of the annual energy demand of buildings through
averaged monthly and hourly calculation methods: A comparative analysis. Build. Simul. Conf. Proc. 2019, 6, 3839–3846.
29. Ballarini, I.; Costantino, A.; Fabrizio, E.; Corrado, V. The dynamic model of en iso 52016-1 for the energy assessment of buildings
compared to simplified and detailed simulation methods. In Proceedings of the Building Simulation 2019, Rome, Italy, 2–4
September 2019; pp. 3847–3854.
30. Rajabi, A.; Eskandari, M.; Ghadi, M.J.; Li, L.; Zhang, J.; Siano, P. A comparative study of clustering techniques for electrical load
pattern segmentation. Renew. Sustain. Energy Rev. 2020, 120, 109628. [CrossRef]
31. Aghabozorgi, S.; Seyed Shirkhorshidi, A.; Ying Wah, T. Time-series clustering—A decade review. Inf. Syst. 2015, 53, 16–38.
[CrossRef]
32. Verleysen, M.; François, D. The curse of dimensionality in data mining and time series prediction. In Computational Intelligence
and Bioinspired Systems; Cabestany, J., Prieto, A., Sandoval, F., Eds.; Springer: Berlin/Heidelberg, Germany, 2005; pp. 758–770.
33. Mohandes, S.R.; Zhang, X.; Mahdiyar, A. A comprehensive review on the application of artificial neural networks in building
energy analysis. Neurocomputing 2019, 340, 55–75. [CrossRef]
34. Ring, M.; Eskofier, B.M. An approximation of the gaussian rbf kernel for efficient classification with svms. Pattern Recognit. Lett.
2016, 84, 107–113. [CrossRef]
35. Santolamazza, A.; Cesarotti, V.; Introna, V. Anomaly detection in energy consumption for condition-based maintenance of
compressed air generation systems: An approach based on artificial neural networks. IFAC-Pap. 2018, 51, 1131–1136. [CrossRef]
36. Han, H.; Zhang, Z.; Cui, X.; Meng, Q. Ensemble learning with member optimization for fault diagnosis of a building energy
system. Energy Build. 2020, 226, 110351. [CrossRef]
37. Deng, Z.; Zhu, X.; Cheng, D.; Zong, M.; Zhang, S. Efficient knn classification algorithm for big data. Neurocomputing 2016, 195,
143–148. [CrossRef]
38. Wang, Z.; Parkinson, T.; Li, P.; Lin, B.; Hong, T. The squeaky wheel: Machine learning for anomaly detection in subjective thermal
comfort votes. Build. Environ. 2019, 151, 219–227. [CrossRef]
39. Xindong, W.; Shichao, Z. Synthesizing high-frequency rules from different data sources. IEEE Trans. Knowl. Data Eng. 2003, 15,
353–367. [CrossRef]
40. Wu, X.; Zhang, C.; Zhang, S. Database classification for multi-database mining. Inf. Syst. 2005, 30, 71–88. [CrossRef]
41. Wylie, T.; Schuh, M.A.; Angryk, R.A. Enabling high-dimensional range queries using knn indexing techniques: Approaches and
empirical results. J. Comb. Optim. 2016, 32, 1107–1132. [CrossRef]
42. Gou, J.; Du, L.; Zhang, Y.; Xiong, T. A new distance-weighted k-nearest neighbor classifier. J. Inf. Comput. Sci. 2012, 9, 1429–1436.
43. I-Smart, Power Planner, Kepco. Available online: https://home.kepco.co.kr/kepco/main.do (accessed on 7 September 2022).
44. Martínez, F.; Frías, M.P.; Pérez-Godoy, M.D.; Rivera, A.J. Dealing with seasonality by narrowing the training set in time series
forecasting with knn. Expert Syst. Appl. 2018, 103, 38–48. [CrossRef]
45. Gómez-Omella, M.; Esnaola-Gonzalez, I.; Ferreiro, S.; Sierra, B. K-nearest patterns for electrical demand forecasting in residential
and small commercial buildings. Energy Build. 2021, 253, 111396. [CrossRef]
46. Dong, Z.; Liu, J.; Liu, B.; Li, K.; Li, X. Hourly energy consumption prediction of an office building based on ensemble learning
and energy consumption pattern classification. Energy Build. 2021, 241, 110929. [CrossRef]
47. American Society of Heating, Refrigerating and Air Conditioning Engineers. Ashrae Guideline 14-2002, Measurement of Energy
and Demand Savings—Measurement of Energy, Demand and Water Savings. 2002. Available online: https://webstore.ansi.org/
Standards/ASHRAE/ashraeguideline142002 (accessed on 7 September 2022).
48. Ho, W.T.; Yu, F.W. Chiller system optimization using k nearest neighbour regression. J. Clean. Prod. 2021, 303, 127050. [CrossRef]