Journal of Building Engineering 31 (2020) 101339

Contents lists available at ScienceDirect

Journal of Building Engineering

journal homepage: http://www.elsevier.com/locate/jobe

3D pattern identification approach for cooling load profiles in

different buildings
X.J. Luo , Lukumon O. Oyedele *, Olugbenga O. Olugbenga , Anuoluwapo O. Ajayi
Big Data Enterprise and Artificial Intelligence Laboratory, University of the West of England, Frenchay Campus, Bristol, United Kingdom

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

Keywords: Building energy conservation has gained increasing concern owing to its large portion of energy consumption
Pattern identification and great potential of energy saving. In-depth understanding of representative patterns of daily cooling load
Gaussian mixture model clustering profile will facilitate effective building energy system scheduling, fault detection and diagnosis, as well as de­
Cooling load
mand and supply side management. In this study, a novel three-stage approach is proposed for pattern identi­
Data visualization
Energy management
fication of cooling load profiles in different types of buildings. The three stages include data preparation, data
clustering and data visualization. The initial measurement in the building energy management system is con­
ducted at the time step of 15 min. To further explore the characteristics of the building cooling load trend, 1-h
mean pattern, 4-h mean pattern and daily statistical information (i.e. average, minimum and maximum values)
of cooling load are also adopted for data clustering, respectively. To test the generality and robustness of the
proposed approach, one-year historical measurement data collected from the practical chilled water system in
two different buildings are adopted, respectively. The analysis demonstrates that the 3D pattern identification
approach can effectively discover the representative characteristics of the daily cooling load profiles in both
buildings. It is also expected that the proposed 3-stage pattern identification approach is general in adoption and
can be potentially adopted in various types of buildings in different climate zones.

1. Introduction performance using historical measurement data [9,10]. The newly

developed statistical models include multiple linear regression and
As the largest energy-consuming sector, building accounts for over various gaussian process regression models [11,12]; while the newly
one-third of primary energy consumption globally [1]. To improve developed machine learning models include artificial neuron networks,
building energy efficiency, various technologies such as energy demand support vector machines, decision trees and genetic algorithms. On the
prediction [2], system design optimization [3], model predictive control other hand, pattern identification of representative energy load profiles
[4], demand side management [5] as well as energy-efficient prime will further enhance the performance of energy demand prediction,
movers [6,7] have been investigated. system design optimization as well as demand and supply side man­
agement. Clustering-based data-driven models are an unsupervised en­
ergy consumption data analysis approach including the purpose of
1.1. Literature review discovering unlabelled datasets and hidden information, which is widely
used within the studies of building electricity, heating and cooling load
Energy modelling in buildings seeks to quantify energy consumption. profiles.
In recent years, a large number of studies for energy modelling ap­ Pattern identification of electricity consumption profile was mainly
proaches have been proposed for different types of buildings, which can investigated on residential and office buildings. For residential build­
be generally categorised into engineering-based methods and data- ings, Wen et al. [13] proposed a principle component analysis-based
driven methods [8]. Engineering methods adopt the fundamental ther­ k-means clustering algorithm and a shape-based clustering method for
modynamic and physics models to investigate the energy behaviour of pattern recognition of residential electricity consumption. Li et al. [14]
the building. On the other hand, data-driven methods use the statistical proposed a simulated annealing-based k-means clustering approach to
or machine learning methods to investigate building energy

* Corresponding author.
E-mail addresses: [email protected], [email protected] (L.O. Oyedele).

https://doi.org/10.1016/j.jobe.2020.101339
Received 25 December 2019; Received in revised form 13 February 2020; Accepted 6 March 2020
Available online 19 March 2020
2352-7102/© 2020 Elsevier Ltd. All rights reserved.
X.J. Luo et al. Journal of Building Engineering 31 (2020) 101339

Fig. 1. Structure of 3D pattern identification approach.

extract typical electricity consumption patterns of different residential to explore patterns in daily heating load profiles of heat pumps in Danish
buildings. Guo et al. [15] used the k-means clustering method to identify dwellings. Two main clusters were identified: a main cluster with a
the electricity consumption patterns in different households within an relatively constant load profile and a minor cluster with a more distinct
area in China. In these three studies, year-round electricity consumption variation during the daytime. Panagiota et al. [21] proposed a
profiles collected from a large quantity of residential households were clustering-based knowledge discovery approach for evaluating heating
adopted as database. Through various k-means clustering methods, consumption data of the residential buildings from district heating
various residential buildings were grouped into different clusters, while system. It was found that the daily heating load profiles were grouped
the representative electricity consumption patterns were recognized into 5 clusters with regards to their consumption intensity. The clusters
from each cluster. For office buildings, Fan et al. [16] proposed a tem­ were characterized by relatively constant load profiles with a peak in the
poral knowledge discovery methodology for identifying electrical en­ early morning and a peak in the evening, respectively. Lu et al. [22]
ergy consumption of an office building. Several time series data mining adopted the GMM clustering approach to identify the occur time and
techniques were explored and assembled, including the symbolic energy signature of daily heating load patterns in 6 office buildings. Four
aggregate approximation, motif discovery, and temporal association operating patterns of the district heating system were identified,
rule mining. The year-round electricity consumption profile of the including working, on-duty, daytime-nighttime and nighttime-daytime
building was grouped into 4 clusters, while each cluster represented a patterns.
climate (i.e. cold and hot seasons) and day type (i.e. weekdays and Building cooling load can be estimated from various measurement of
weekends). Xiao et al. [17] adopted the entropy-weighted k-means the chilled water system. However, there were quite limited studies
clustering algorithm to identify the typical electrical power consump­ regarding the pattern identification of building cooling loads. Yu et al.
tion patterns in an office building. It was found that electricity con­ [23,24] used the k-means clustering algorithm to evaluate the perfor­
sumption patterns in weekdays were similar to each other, while there mance of the chilled water system. The operating variables including
existed distinctive differences among weekdays, Saturdays and Sundays. load of each operating chiller, the total system load, the quantity of
Li et al. [18] adopted the Gaussian mixture model (GMM) clustering to operating primary chilled water pumps, the quantity of operating
identify the typical daily electricity usage profiles of each individual condenser water pumps, and the quantity of operating cooling towers
building. Through GMM clustering, the year-round daily electricity were adopted as database for clustering analysis. The operating data at
consumption profile was grouped into two or three clusters, which each time step was grouped into 3 clusters, indicating the small, medium
indicate the cold, medium and hot seasons, respectively. In addition, an and heavy loading condition of the chilled water system, respectively.
agglomerative hierarchical clustering to identify electricity usage pro­ An et al. [25] adopted the k-means clustering algorithm to analysis the
files of multiple buildings. As a result, typical patterns and periodical air-conditioning intensity and usage patterns in residential buildings.
variation of daily electricity consumption could be discovered. The daily usage of air conditioning system was grouped into different
Furthermore, Wang et al. [19] adopted the k-means clustering for clusters based on its operating probability. Although these three studies
pattern recognition of water-source heat pump operation using its investigated the chilled water and air conditioning systems, it mainly
electrical power consumption. The one-month profile of pump power focused on the operating parameters of the chilled water system and the
consumption was grouped into 6 different clusters, each of which rep­ air conditioning systems. The variation trend of daily cooling load
resented a typical daily operating schedule of heat pumps. In these four profiles was not investigated.
studies, the electricity consumption profile was grouped mainly ac­
cording to its intensity.
1.2. Research gaps and contribution
Building heating loads are mainly provided by heat pumps and dis­
trict heating system. Carolina et al. [20] adopted the k-means clustering
From the above-discussed comprehensive literature review, three

2
X.J. Luo et al. Journal of Building Engineering 31 (2020) 101339

� Although data clustering approaches have been widely adopted for

pattern identification of energy load profiles in residential and office
buildings, they mainly focused on electricity and heating loads. In
previous cooling load studies, only the operating parameters of the
chilled water system were investigated, while there is a lack of study
regarding the daily profile of the building cooling loads.
� Owing to its low computational complexity, k-means clustering was
widely adopted in previous research works for pattern identification.
However, the initial clustering centroids were randomly selected,
which might lead to convergence to local minimum points and dif­
ficulty in achieving global optimal solutions. However, GMM clus­
tering can approximate any probability distribution by adjusting the
quantity of mixture models, while there is a lack of study in exploring
the effectiveness of GMM clustering in pattern identification of daily
cooling load profile.
� Furthermore, previous studies mainly focus on the residential and
office buildings. Due to the unique functions, the cooling load profile
in campus and hotel buildings also worth investigation. However,
none of the previous studies considered the energy loads in campus
and hotel buildings.
� Last but not least, most of the previous research works mainly
focused on the methodology of data clustering. However, data
preparation is critical in providing the high-quality database while
data visualization is significant in analyzing the clustering results.

Therefore, the contribution of this study is to propose a robust and

generalized 3D pattern identification approach for daily cooling load
Fig. 2. Procedure of GMM clustering. profiles which can be adopted in different types of buildings. The 3D
indicates the three-stage of the pattern identification approach: data
preparation, data clustering and data visualization, while GMM is
Table 1 adopted for data clustering. To demonstrate the robustness and gener­
Building information. ality of the proposed pattern identification approach, the historical
Building type Campus Hotel measurement data from a campus and a hotel building from the same
Number of floors 37 15
Floor area 63500 160000 Table 2
Location Hong Kong Island Hong Kong Island
Key design parameters of the chilled water system.
Rated cooling capacity of each chiller (kW) 2103
major research gaps were identified: Rated chilled water flow rate of each chiller (L/s) 100.4
Set-point of chilled water leaving temperature (� C) 7
Design chilled water return temperature (� C) 12
� Engineering building energy simulation methods require both high Design COP 6.2
level of technical knowledge and detailed information of building Rated condenser water flow rate of each chiller (L/s) 110.0
materials to improve its modelling accuracy; Set-point of cooling water leaving temperature (� C) 37.37
Design cooling water return temperature (� C) 32.0

Fig. 3. Schematic diagram of the chilled water systems.

3
X.J. Luo et al. Journal of Building Engineering 31 (2020) 101339

Fig. 4. Year-round cooling load of the two buildings.

4
X.J. Luo et al. Journal of Building Engineering 31 (2020) 101339

Fig. 5. BIC values of campus building for different time steps.

hot-tropic area are adopted to test the performance of the proposed initial cooling energy generated by the chiller plant. However, operating
pattern identification approach. It is also expected that the proposed schedule of the chilled water system should be determined by the initial
approach can be generalized and adopted in diverse climate zones. cooling energy demand at the chiller side. Therefore, in this study, the
The remainder of the paper is organized as follows: the next section cooling load profile is considered to be the initial cooling energy
illustrates the structure of the proposed pattern identification approach; generated by the chiller plant. To gain the information about the initial
the third section presents the information regarding the chilled water cooling load profile, historical measurement data of the chilled water
system; the fourth section discusses the identified representative pat­ system should be collected from the building energy management sys­
terns of building cooling loads; the fifth section shows implication for tem, covering total quantity of operating chillers G; leaving temperature
practice and future direction while the last section draws the conclusion. Tl;g;i;j , return temperature Tr;g;i;j and mass flow rate Mg;i;j of each chiller;
as well as the leaving temperature Tl;b;i;j , return temperature Tr;b;i;j and
2. Structure of the pattern identification approach mass flow rate Mb;i;j at the main branch. i stands for the number of the
day, while j represents the time step in each day.
In this study, a three-stage pattern identification approach is pro­ Since the temperature difference between the inlet and outlet of the
posed to process and analyse the building cooling load profile. The circulation water pump is relatively low, the sum of cooling load from
structure of the proposed pattern identification approach is shown in each chiller should be approximate to the cooling load from the main
Fig. 1. The three stages consist of data preparation, data clustering and branch. Therefore, the chilled water system can be estimated using two
data visualization. methods: from each operating chiller Qc,i,j and from main branch Qb,i,j:
g¼G
X �
Qc;i;j ¼ Mg;i;j Cp Tr;g;i;j Tl;g;i;j (1)
2.1. Data preparation g¼1

To provide high-quality data for data clustering and data visualiza­ Qb;i;j ¼ Mb;i;j Cp Tr;b;i;j Tl;b;i;j
�
(2)
tion, data preparation is conducted through three steps: data collection,
data cleaning and data pre-processing. where Cp is the specific heat capacity of chilled water.

2.1.1. Data collection 2.1.2. Data cleaning

Due to heat loss on various pipes and circulation pumps, the actual Missing and outlier values are mainly caused by sensor malfunctions
cooling load provided by the air handling unit would be smaller than the

5
X.J. Luo et al. Journal of Building Engineering 31 (2020) 101339

Fig. 6. BIC values of hotel building for different time steps.

and signal transmission problems in the building management system.

Qc;i;j þ Qb;i;j
The missing and outlier values of the measurement data have a negative Qi;j ¼ (4)
2
impact on data analysis. Hence, the aim of data cleaning is to improve
the quality of raw measurement data by excluding missing and outlier
2.1.3. Data pre-processing
values. In this study, after detecting the missing values, outliers of the
In building energy management system, sensor measurement is
measurement data are identified through the following procedures:
generally recorded at 15 min. From the 15-min measurement data, the
database Q for the year-round cooling load is consolidated as a 365 � 96
� When the chiller is operating, the temperature of the leaving chilled
matrix. Q ¼ ½Q1 ; Q2 ; …Qi ; …; Q365 � , each Qi is a vector representing
water temperature should be lower than the return chilled water
daily cooling load profile of the ith day, and Qi ¼ ½Qi;1 Qi;2 …Qi;j … Qi;96 �.
owing to the heat exchange with the cooling tower. Therefore,
To provide an in-depth investigation on the daily cooling load profile,
chilled water can be delivered to the air handling units for space
the initial 15-min measurement data is averaged at every 1 h and 4 h,
cooling purposes. When space cooling is not needed, chillers and
respectively. In addition, daily brief information (i.e. daily minimum,
cooling towers are turned off, thus the temperature of leaving chilled
maximum and average value) of the cooling load profile is also obtained.
water is equal to that of the returning chilled water. Therefore, at
To make the cooling load relatively flatten, the average value of the
each time step, Qc;i;j and Qb;i;j should not be lower than 0. The
15-min cooling load Qi;j is obtained at every 4-time-steps to obtain the 1-
negative value of Qc;i;j and Qb;i;j might be caused by the faulty sensor
h mean value.
measurement thus should be excluded.
� Due to the heat loss in pipes and circulation pumps, there would be Qi;ðj þ Qi;ðj þ Qi;ðj þ Qi;ðj
(5)
1Þ�4þ1 1Þ�4þ2 1Þ�4þ3 1Þ�4þ4
Xi;j ¼
little difference between Qc;i;j and Qb;i;j . Therefore, the relative error r 4
between Qc;i;j and Qb;i;j should not be higher than 10%: Thus, the daily 1-h mean cooling load profile Xi is a vector consisting
8 �
�
�
� � of 24 elements, and Xi ¼ ½Xi;1 Xi;2 …Xi;j … Xi;24 �. Meanwhile, the database
< 2 Qc;i;j Qb;i;j
>
max Qc;i;j ; Qb;i;j > 0 X for the 1-h mean value is consolidated as a 365 � 24 matrix, and: X ¼
r¼ Qc;i;j þ Qb;i;j (3)
>
: ½X1 ; X2 ; …; X365 �.
0 else Furthermore, to gain briefer understanding of the overall trend of the
After clearing out the missing and outlier values, the cooling load of daily cooling load profile, the average value of the 1-h mean cooling load
the chilled water system Qi;j is calculated as: Xi;j is calculated at every 4 time steps to obtain the 4-h mean value.

6
X.J. Luo et al. Journal of Building Engineering 31 (2020) 101339

Fig. 7. Clustering results of the campus building at 15-min time step.

7
X.J. Luo et al. Journal of Building Engineering 31 (2020) 101339

Fig. 8. Clustering results of the campus building of 1-h mean value.

� P
Xi;ðj þ Xi;ðj þ Xi;ðj þ Xi;ðj φk ðQ�μk ; k Þ Gaussian probability density function of the kth
(6)
1Þ�4þ1 1Þ�4þ2 1Þ�4þ3 1Þ�4þ4
Yi;j ¼ P
4 mixture component, where μk denotes the mean matrix, while k is
Therefore, the daily 4-h mean cooling load profile Y i is a vector the covariance matrix. φk can specify the spread and orientation of
consisting of 6 elements, and Y i ¼ ½Yi;1 Yi;2 … Yi;6 �. After that, the data­ the distribution:
base Y for the 4-h mean value is consolidated as a 365 � 6 matrix, and 2 0 13
0 1 σ2 σ … σ
Y ¼ ½Y 1 ; Y 2 ; …; Y 365 �. 6 μ1 B 1 12 1k
C7
To summarize the daily trend of the cooling load profile, its average, 6B C B σ σ 2 … σ C7
6B μ2 C B 21 2 2k C7
φk 6@ A; B C7
minimum and maximum values in each day is consolidated as the daily 6 … B … C7
4 @ A5
brief information. μk 2
σ k1 σk2 … σ k
Qi;1 þ Qi;2 þ … þ Qi;96
Zi;1 ¼ (7) In this study, each cluster is assumed to be represented by either
96
identical or different mixture models. Namely, for “true” case, all the
�
Zi;2 ¼ max Qi;1 ; Qi;2 ; …; Qi;96 (8) mixture models share the same covariance matrix; while for “false” case,
different covariance matrix are assigned to various mixture models. For
�
Zi;3 ¼ min Qi;1 ; Qi;2 ; …; Qi;96 (9) both cases, both diagonal and full Gaussian distribution are studied.
Given the database Q and pre-set K, a GMM fitting is to estimate the
As a result, the daily brief information Zi is a vector consisting of 3 P
values of parameter sets fαk ; μk ; k gðk ¼ 1; 2; …KÞ to ensure that the
elements, and Zi ¼ ½Zi;1 Zi;2 Zi;3 �. The database Z for the daily brief in­
GMM has the maximum likelihood. The fitting procedure is carried out
formation is consolidated as a 365 � 3 matrix, and Z ¼ ½Z1 ; Z2 ; …; Z365 �. using Expectation Maximization algorithms, which consists of three
steps: initialisation, expectation step and maximization step. The initial
2.2. Data clustering values of parameter sets are randomly selected at first. After that, an
iteration of expectation step and maximization step is conducted to
GMM-based clustering uses the feature of GMM to group multiple improve the estimation of the parameter sets.
observations from the database into different clusters. In order to In the expectation step, each daily cooling load profile Qi at the ith
investigate the representative patterns of the cooling load profile, GMM day is assigned to one of the mixture model k which has the highest
clustering is conducted to group the databases of Q, X, Y and Z into probability γ ik . The posteriori probability γik is calculated based on the
P
different clusters. given values of parameter set fαk ; μk ; k g:
� P�
ωk φk Qi �μk ; k
2.2.1. GMM clustering γ ik ¼ PK P� (12)
The GMM is a probabilistic model, which uses a weighted combi­ k¼1 αk φk ðQi jμk ; k

nation of several Gaussian distributions to represent the database Based on the γ ik obtained from the expectation step, the parameter
[26–28]. The procedure of GMM clustering is illustrated in Fig. 2. In this P
sets fαk ; μk ; k g are updated in the maximization step:
subsection, database Q is illustrated as an example. The procedure of
GMM clustering for X, Y and Z is similar, while the dimension d is 96, 24, φk
ωk ¼ (13)
6 and 3 for Q, X, Y and Z, respectively. N
The d-dimensional GMM Ψ ðQÞ with K mixture components can be
1 X N
described as: μk ¼ γ Q (14)
φk j¼1 ik i
X
K � � X�
�
Ψ ðQÞ ¼ ωk φk Q�μk ; (10)
X 1 X N
(15)
k¼1 k
¼ γ ðQ μk ÞðQi μk ÞT
( ) k
φk j¼1 ik i
� � X� � �X�� d2 1 X1
� � �
φk Q�μk ; ¼ 2π� � exp ðQ μk ÞT ðQ μk Þ (11) P
k k
2 k where φk ¼ Ni¼1 γ ik , N denotes the quantity of vectors in the database,
and N ¼ 365.
where The expectation step and maximization step are iteratively computed
until the updated parameters of all the mixture models do not change
K quantity of mixture models further. Therefore, the likelihood function is determined as:
PK
ωk weight of the kth mixture component, 0 < ωk < 1, and k¼1 ωk ¼
1

8
X.J. Luo et al. Journal of Building Engineering 31 (2020) 101339

Fig. 8. (continued).

9
X.J. Luo et al. Journal of Building Engineering 31 (2020) 101339

Fig. 9. Clustering results of the campus building of 4-h mean value.

( !)
X
N X
K X
L θjQi Þ ¼ ln ωk φk ðQi jμk ; (18) 1
A ¼ Kd þ Kdðd þ 1Þ þ ðK 1Þ (20)
i¼1 k¼1 k 2
Once the GMM has been fitted, the observations belonging to the
same mixture component are considered to be grouped in the same 2.3. Data visualization
cluster.
To investigate the clustering results of year-round cooling profile, the
2.2.2. Performance evaluation criterion calendar view is adopted to illustrate the load distribution of the whole
Bayesian Information Criterion (BIC) is the most widely adopted year. Meanwhile, through plotting the cooling load profile based on the
indicator for statistical model selection. Assuming that the input data is cluster number, the representative and characteristics of each cluster
originally generated according to an unknown GMM (i.e. ψtrue) and the can be identified.
GMM is fitted using the generated input data (i.e. ψcandidate), BIC is
adopted to measure the difference between ψtrue and ψcandidate. There­ 3. Chilled water system under study
fore, the GMM with an optimal K value will result in the lowest BIC. The
optimal K value in the intra-building clustering used in this study was To keep track of the changing cooling load, chilled water systems
determined based on the range of 2–20: must respond by varying chilled water flow rate or chilled water tem­
perature differential. Variable chilled water pumps and bypass valve are
BIC ¼ LðθjQi Þ þ
A
logðNÞ (19) two representative approaches to adjust the cooling load of the water-
2 cooled chilled water system. Variable chilled water pumps can be used
to adjust the flow rate of chilled water while bypass valve is generally
where A is the total quantity of free parameters, and can be calculated
adopted to change the chilled water temperature differential.
as:
To test the performance and demonstrate the generality of the pro­
posed pattern identification approach, it is implemented on two typical

10
X.J. Luo et al. Journal of Building Engineering 31 (2020) 101339

Fig. 10. Daily average, minimum and maximum cooling load of the campus building.

water-cooled chilled water systems installed in a campus building and a 0 during night time. For the campus building, there were approximately
hotel building, respectively. The basic building information is summa­ three types of cooling load during mid-May to mid-October, while two
rized in Table 1. The chilled water system 1 used in the high-rise campus types in other period of the year. For the hotel building, there were about
building consisted of two identical centrifugal water-cooled chillers. The two categories of cooling load from mid-May to mid-October, while only
differential pressure bypass was adopted in chilled water system 1 to one type in other time periods.
allow surplus chilled water passing through when the air handling units
were under part-load operation. The chilled water system 2 adopted in 4.2. Performance evaluation of GMM clustering
the high-rise hotel building was comprised of four identical chillers,
while the maximum operating quantity of chillers was two. Each chiller The Bayesian Information Criterion values introduced in Section
had an associated chilled water pump and a cooling water pump, 2.2.2 were adopted to evaluate the GMM clustering performance. The
respectively. The variable chilled water pumps were equipped in the variation between BIC values and quantity of components in the
chilled water system 2. The same type of water-cooled chillers and Gaussian mixture model K under different cases are summarized in
cooling towers were installed in both chilled water systems. The sche­ Figs. 5 and 6. For each building under four different cases (i.e. shared
matic diagrams of the two chilled water systems are shown in Fig. 3, covariance is true while sigma is diagonal, shared covariance is false
while its key design parameters are summarized in Table 2. while sigma is diagonal, shared covariance is false while sigma is full,
shared covariance is false while sigma is full), the BIC values were
4. Results and analysis evaluated. For both the campus and the hotel buildings, the optimal K
was found when shared covariance was false while sigma was full. For
In this section, the performance of each stage of the proposed 3D the campus building, the smallest BIC value was 4.20 � 105, 1.19 � 105,
pattern identification approach was evaluated, respectively. 2.94 � 104 and 1.24 � 104 when K was 9, 18, 4 and 6 for 15-min, 1-h, 4-
h, and daily cooling load, respectively. For the hotel building, the
4.1. Performance assessment of data preparation smallest BIC value was 3.80 � 105, 1.06 � 105, 2.84 � 104 and 1.45 �
104 when K was 10, 19, 5 and 5 for 15-min, 1-h, 4-h, and daily cooling
To assess the performance of data collection, data cleaning and data load, respectively. Therefore, the similar smallest BIC value and optimal
pre-processing, the year-round cooling load of the two buildings were K was identified for the two cooling load profiles under the same time-
investigated. The 15-min measurement, 1-h mean, 4-h mean, and daily step value.
average cooling load of the two buildings are shown in Fig. 4.
For both the campus and hotel buildings, cooling loads were higher 4.3. Data visualization and analysis
from mid-May till mid-October. From the 15-min measurement, it can be
seen that the variation of the cooling load in the campus building was To investigate the characteristics of the year-round cooling load of
quite large in each day. In other words, the cooling load tended to be the campus and the hotel building, the distribution of the cooling load

11
X.J. Luo et al. Journal of Building Engineering 31 (2020) 101339

Fig. 11. Clustering results of the hotel building at 15-min time step.

12
X.J. Luo et al. Journal of Building Engineering 31 (2020) 101339

Fig. 12. Clustering results of the hotel building of 1-h mean cooling load.

profile in a calendar view was visualised to explore the different load � Clusters C13–C18 represented the mid-season (i.e. April, late
distribution among the whole year while the representative cooling load October and November). There were more days clustered into the
in each cluster were identified through plotting the clustered cooling mid-season compared to that from the 15-min measurement, espe­
load profiles. cially the Sundays in September. During this period, the 1-h mean
value of cooling load was low during night-time while almost con­
4.3.1. Chilled water system in the campus building stant at about 1800 kW during daytime (8–20 h).
The clustering results of the 15-min measurement, 1-h mean value, 4-
h mean value and daily brief are summarized in this section. As shown in Fig. 5(c), the optimal K value was 4 while the 4-h mean
As shown in Fig. 5(a), the optimal K value was 9 the 15-min mea­ value of daily cooling load profile was grouped into 4 different clusters.
surement of daily cooling load profile was grouped into 9 different The cluster distribution at each day is shown in Fig. 9(a), while the daily
clusters. The cluster distribution at each day is shown in Fig. 7(a), while cooling load profile in corresponding clusters are shown in Fig. 9(b).
the grouped daily cooling load profile in different clusters are summa­ Different from the 15-min measurement and 1-h mean value, the whole
rized in Fig. 7(b). It is found that: year cooling load profile was briefly divided into two seasons: cold and
hot seasons. It is found that most of the days in the cold season (i.e.
� Clusters C1–C4 were mainly distributed among the cold season (i.e. January to March, late November and December) were grouped into the
January–March and December), during which the cooling load var­ single cluster C1, during which the 4-h mean value of cooling load was
ied from 0 to 2000 kW. Due to the relatively low cooling load during around 200–1200 kW. Meanwhile, most of the days during the hot
the cold season, the chillers were continuously switched on-and-off, season (i.e. April-early November) are grouped into Clusters C2 and C3,
which thus resulted in the persistent variation of cooling load in each during which the 4-h mean value of cooling load was around 1000 kW at
day. night and 2000 kW at daytime, respectively. Cluster C4 stood for few
� Clusters C5–C7 were generally found in the hot season (i.e. June to days during which there existed a decreasing trend from 2 to 6 h.
early October). During the night time, there existed continuous Moreover, the 4-h mean value of cooling load profiles in each group
variation of cooling load between 0 to 2000 kW. During the daytime were more similar to each other in the respective clusters. It is because
(8–19 h), it jumped to about 4000 kW, and varied between 2000 to that the effects of abrupt variation caused by the continuously switching
4000 kW. on-and-off were minimized through the long time-step average.
� Clusters C8 and C9 represented the mid-season (i.e. April, May, late As shown in Fig. 5(d), the optimal K value was 6 while the sets of
October and November). During the mid-season, there also existed daily average, minimum and maximum cooling load was grouped into 6
continuous variation of cooling load between 0 to 2000 kW during different clusters. The cluster distribution at each day is shown in Fig. 10
night-time, while it was relatively constant between 1000 to 2000 (a), while the daily minimum-average and maximum-average values of
kW during the daytime (8–19 h). cooling load are shown in Fig. 10(b) and (c), respectively. Similar to the
4-h mean value of cooling load, the whole year was also roughly divided
As shown in Fig. 5(b), the optimal K value was 18 thus the 1-h mean into two seasons: cold and hot seasons. It is found that Clusters C1 and
value of daily cooling load profile was grouped into 18 different clusters. C2 represented the relatively cold months (i.e. January–May, November
The cluster distribution at each day is shown in Fig. 8(a), while the daily and December), in which the daily average, minimum and maximum
cooling load profile in corresponding clusters are shown in Fig. 8(b). It is cooling load were in the range of [400, 900], [0, 100], [1700, 2700],
found that: respectively. Clusters C3–C5 stood for the relatively hot months (i.e.
June–October), while the daily average, minimum and maximum
� Clusters C1–C5 were mainly distributed among the cold season (i.e. cooling load were in the range of [1700, 2300], [50, 250], [3500, 4500],
January–March and December), while there were more days respectively. Cluster C6 was spread among both hot and cold seasons.
assigned into the cold season compared to those from 15-min mea­ From Fig. 10(b), the daily minimum value of cooling load was higher
surement. In the cold season, the 1-h mean value of cooling load than that in the cold season while lower than that in the hot season.
varied from 0 to 1200 kW. Due to the relatively low cooling load However, the daily maximum value of cooling load was similar to that in
during the cold season, the chillers were continuously switched on- the cold season.
and-off, which resulted in the continuous waves in each day.
� Clusters C6–C12 were generally found in the hot season (i.e. June to 4.3.2. Chilled water system in the hotel building
early October). Compared to the 15-min measurement, there were The clustering results of the 15-min measurement, 1-h mean value, 4-
fewer days grouped into the hot season. During this period, there also h mean value and daily brief are summarized in this section.
existed continuous variation between 0 to 1200 kW during night- As shown in Fig. 6(a), the optimal K value was 10, therefore, 15-min
time, while the 1-h mean value of cooling load was approximately daily cooling load profile was grouped into 10 different clusters. The
between 2200 to 3800 kW during daytime (8–20 h). cluster distribution at each day is shown in Fig. 11(a), while the daily

13
X.J. Luo et al. Journal of Building Engineering 31 (2020) 101339

Fig. 12. (continued).

14
X.J. Luo et al. Journal of Building Engineering 31 (2020) 101339

Fig. 13. Clustering results of the hotel building at 4-h time step.

cooling load profile in corresponding clusters are shown in Fig. 11(b). It In other words, the cooling load difference between night-time and
is found that: day-time was less than that in the hot season.

� Clusters C1–C4 were mainly distributed among the cold season (i.e. As shown in Fig. 6(b), the optimal K value was 19 thus the 1-h mean
January–March, November and December), during which the cool­ value of daily cooling load profile of the hotel building was grouped into
ing load was around 1000 kW from 7 to 24 h. Compared to that in the 19 different clusters. The cluster distribution at each day is shown in
campus building, the constant cooling load period was longer while Fig. 12(a), while the daily cooling load profile in corresponding clusters
the frequency of switch on-and-off was lower. are shown in Fig. 12(b). It is found that:
� Clusters C5–C8 represented the hot season (i.e. June to early
October). During this period, the cooling load was about 2200 kW � Clusters C1–C7 were mainly distributed among the cold season (i.e.
and 3000 kW during 0–6 h and 6–23 h, respectively. During this January–March, late November and December), during which the
season, there seemed to be less frequency of switching on-and-off cooling load varied from 0 to 1200 kW. Due to the relatively low
chillers. cooling load during the cold season, the chillers were continuously
� Clusters C9 and C10 were generally found in the middle season (i.e. switched on-and-off during some of the days during 0–7 h, which
April, May and late). Similar to that in the hot season, there was less resulted in the waves of 1-h mean value of cooling load. During 7–24
frequency of switching on-and-off in this season. Moreover, in h, the cooling load was relatively constant at 1100 kW.
Cluster C9, the cooling load was about 1600 kW and 2200 kW during � Clusters C8–C15 were generally found in the hot season (i.e. June to
0–7 h and 7–24 h, respectively; in Cluster 10, the cooling load was early October). During this period, the cooling load was about 2000
about 1200 kW and 1800 kW during 0–7 h and 7–24 h, respectively. kW during night, while it increased to 3000 kW from the 9–22 h, and
dropped to 2000 kW after the 23rd h.

15
X.J. Luo et al. Journal of Building Engineering 31 (2020) 101339

Fig. 14. Clustering results of hotel building for daily average cooling load.

Table 3
Summary of the similarity and differences among the two chilled water system.
Building Load profile Clusters Load characteristics
type
Cold season Middle season Hot season

Campus 15-min 9 Jan–Mar, Dec: 0–2000 kW, Apr, May, late Oct, Nov: 0–2000 kW during June-early Oct: 0–2000 kW during night,
building measurement continuously on-and-off night, continuously on-off; 1000–2000 kW continuously on-off; 2000–4000 kW
during daytime (8–20 h). during daytime (8–20 h).
1-h mean 18 Jan–Mar, Dec: 0–1200 kW, Apr, May, late Oct, Nov: 0–1000 kW during June-early Oct: 0–1200 kW during night,
continuously on-and-off night, continuously on-off; 1000–2200 kW continuously on-off; 2200–3800 kW
during daytime (8–20 h). during daytime (8–20 h).
4-h mean 4 Jan–Mar, late Nov, Dec: around N.A. Apr-early Nov: around 1000 kW during
200–1200 kW night while 2000–2800 kW during
daytime (8–20 h)
Daily brief 6 Jan–May, Nov, Dec: N.A. June–Oct:
Average [400 900] Average [1700–2300]
Minimum [0–100] Minimum [50 250]
Maximum [1700–2700] Maximum [3500–4500]
Hotel 15-min 10 Jan–Mar, late Nov, Dec: continuously April, May and late Oct: relatively constant June-early Oct: relatively constant
building measurement on-off during 0–7 h, while relatively around 1600 kW during night (0–7 h), while around 2000 kW during night (0–7 h),
constant at 1000 kW during 7–24 h. around 1800–2000 kW during daytime (7–24 while around 3000 kW during daytime
h). (7–24 h)
1-h mean 19 Jan–Mar, late Nov, Dec: almost April, May, late Oct and early Nov: about June-early Oct: about 2000 kW during
constantly at 1100 kW during 7–24 h, 1500 kW during night (0–7 h), while 2000 kW night (0–7 h), while 3000 kW during
while several ups-and-downs during during daytime (7–24 h). daytime (7–24 h).
0–7 h.
4-h mean 5 Jan–May, late Oct–Dec: around June-early Oct: about 2200 kW during
500–2500 kW. night (0–7 h) while 3000 kW during
daytime (7–24 h).
Daily brief 5 Jan–May, Nov and Dec: Average June–Oct:
[600 2400] Average [2400–3000]
Minimum [0–1800] Minimum [1200–2300]
Maximum [1000–2500] Maximum [3200–4500]

16
X.J. Luo et al. Journal of Building Engineering 31 (2020) 101339

� Clusters C16–C19 represented the mid-season (i.e. April, late From the above analysis, it is seen that the 15-min measurement
October and early November). During this period, the cooling load could provide the detailed trends of the daily cooling load, the 1-h mean
was about 1500 kW during night, while it increased to 2000 kW value was able to illustrate the overall trend of the daily cooling profile,
during daytime (7–24 h). the 4-h mean could show the general trend of the building cooling load,
and the daily brief load profile demonstrated the maximum, minimum
As shown in Fig. 6(c), the optimal K value was 5 while the 4-h mean and average cooling load in different seasons. For both the campus and
value of daily cooling load profile was grouped into 5 different clusters. the hotel building, the 1-h mean value resulted in the most clusters while
Different from the 15-min measurement and 1-h mean value of cooling both the 4-h mean value and daily brief values resulted in the fewest
load, the whole year was briefly divided into two seasons: cold and hot clusters through GMM clustering. By partially avoiding the frequent
seasons. Most of the days in the cold season (i.e. January–May, late changes of cooling load, the 1-h mean value resulted in the most accu­
October–December) were grouped into Clusters C1 and C2, during rate clustering thus could explore the most of characteristics of the daily
which the 4-h mean value of cooling load was around 500–2500 kW. cooling load profile. Moreover, 4-h mean value and daily brief value is
Most of the days during the hot season (i.e. June-early October) were also helpful when simpler and quicker analysis is needed.
grouped into Clusters 3 and 4, during which the cooling load was around It is also discovered that different characteristics of the cooling load
2200 kW and 3000 kW at night-time and daytime, respectively. Cluster can be identified in hot, cold and middle seasons, respectively:
C5 stood for few days during which there existed a decreasing trend at
the end of each day. Most of the days which were grouped into Cluster � During the cold season, there usually existed ups-and-downs of the
C5 were Thursdays. During these days, the cooling load dropped to cooling load during night in the two chilled water system, while it
0 during 21–24 h. This might be caused by the pre-scheduled mainte­ was relatively constant during daytime;
nance of the related system. Moreover, the 4-h mean value of cooling � During the middle and hot seasons, there were still some ups-and-
load in each group were more similar to each other in the respective downs during the night in the campus building, while the cooling
clusters. It was because that the effects of abrupt variation caused by the load was relatively constant during daytime and night in the hotel
continuously switching on-and-off were minimized through the long building;
time-step average (see Fig. 13). � For the campus building, the daytime was around 8–20 h, while the
As shown in Fig. 6(d), since the optimal K value was 5, the sets of daytime was much longer in the hotel building: 7–24 h.
daily average, minimum and maximum cooling load was grouped into 5
different clusters. The cluster distribution at each day is shown in Fig, 14 From the above analysis, it is seen that when the cooling load was
(a), while the daily minimum-average and maximum-average values of relatively low, there existed a lot of ups-and-downs of the cooling load.
cooling load are shown in Fig. 14(b) and (c), respectively. Similar to the In other words, the chillers were working at low part-load-ratio during
4-h averaged cooling load, the whole year was also roughly divided into most of the time, which resulted in low energy conversion efficiency. It
two seasons: cold and hot seasons. It is found that Clusters C1–C4 rep­ is suggested that chilled water storage could be installed to shift the peak
resented the relatively cold months (i.e. January–May, November and cooling demand and help the chillers operate at relatively constant load.
December), in which the daily average, minimum and maximum cooling
load were in the range of [600, 2400], [0, 1800] and [1000, 2500], 5. Implication for practice and future direction
respectively. Cluster C5 stood for the relatively hot months (i.e. June­
–October), in which the daily average, minimum and maximum cooling In this study, a 3D pattern identification approach was proposed for
load were in the range [2400, 3000], [1200, 2300] and [3200, 4500], building cooling loads. Due to the limitation of available historical data,
respectively. Moreover, due to the switching on-and-off of chillers at the only one-year measurement data was adopted as the database in this
21st h, those mentioned Thursdays in C5 from 4-h mean cooling load study. In practical application, several years’ measurement data can be
data were clustered into C1 due to the low minimum and average adopted as the database of the proposed pattern identification approach
cooling load in each day. to further improve its accuracy and effectiveness.
Based on the recorded profile of historical operating data of the
4.4. Summary of the proposed 3D pattern identification approach chilled water system, the proposed pattern identification approach can
be adopted to discover the characteristics and identify the representative
In this study, the 3D pattern identification approach was proposed cooling loads in campus and hotel buildings. Through effective identi­
for the building cooling loads. fication of representative daily cooling load profile.

� At the first stage, the historical measurement data of the chilled � Energy-efficient system management can be implemented: During
water system was collected from the building management system, low cooling load period, chilled water storage or other methods
while data cleaning was conducted to improve the quality of raw should be adopted to avoid frequent switching on-and-off of chillers
measurement by excluding the missing and outlier values. Mean­ hence to improve chiller operating efficiency;
while, measurement data was processed at different time steps (i.e. � Effective demand and supply side management can be adopted: Due
the 1-h mean value, 4-h mean value, daily average, minimum and to the different characteristics of cooling loads in different types of
maximum value) to evaluate the different trends of the daily cooling buildings, the power grid could make personalized intervention
load profile. strategies for users during different periods from the point of view of
� At the second stage, GMM clustering was conducted to group the 15- demand and supply side management;
min measurement, 1-h mean value, 4-h mean value and the daily � Fault detection and diagnosis can be implemented for the energy
brief value of the daily cooling load profile into different clusters. management system: based on the representative daily cooling load,
� At the third stage, data visualization of clustering group in a calendar abnormal energy consumption of the cooling system can be diag­
view and the grouped daily cooling load profile was implemented to nosed through comparing the real-time operating data to the repre­
investigate the characteristics of the cooling load profile around the sentative cooling load profiles.
whole year.
Both the outdoor weather conditions (i.e. outdoor air dry-bulb
Through the proposed pattern identification approach, the similarity temperature, relative humidity, solar radiation, wind speed, etc) and
and differences of the cooling load among the campus and the hotel indoor operating schedules (i.e. occupancy, lighting, etc) have effect on
building were found and summarized in Table 3. the actual cooling load [2]. It is worthwhile to see the relationship

17
X.J. Luo et al. Journal of Building Engineering 31 (2020) 101339

between clustering result of outdoor weather conditions, indoor oper­ pattern identification are highlighted as follows:
ating schedules and the cooling load profile. It also worth investigating
the clustering result of an integrated information among outdoor � Through the proposed pattern identification approach, the year-
weather conditions, indoor operating schedules and cooling load profile. round cooling load profile can be grouped into different clusters,
In future study, the applicability of the proposed pattern identifica­ with each cluster illustrating a distinctive characteristic of daily
tion approach should be tested on other types of building such as resi­ cooling load profile;
dential, office and hospital buildings. Moreover, the performance of � The 15-min measurement, 1-h mean value, 4-h mean value and the
other types of clustering algorithms in feature identification, such as daily brief load profile could provide the detailed trend, overall
connectivity-based clustering, centroid-based clustering and density- trend, general trend and brief information of the daily cooling load,
based clustering should be evaluated. Meanwhile, it is worthwhile to respectively;
see whether this pattern identification approach is also effective on � The 1-h mean value resulted in the most clusters while the 4-h mean
other types of chilled water system (i.e. air-cooled chilled water system) value and daily brief values resulted in the fewest clusters through
and in different climatic regions. It is also expected that the proposed 3D GMM clustering. Therefore, the 1-h mean value resulted in the most
pattern identification can formulate a standard tool to facilitate feature accurate clustering thus could explore the most of features of the
identification of cooling loads in various types of buildings and even daily cooling load profile. On the other hand, 4-h mean value and
district area thus to provide valuable retrofiting suggestion for chilled daily brief value is helpful in quicker analysis.
water system and buildings.
Through comparing the clustering results of the campus and hotel
6. Conclusion building, it is also found out that:

This paper proposes a 3D pattern identification approach for build­ � Compared to that in the hotel building, there are more frequent
ing cooling load using Gaussian mixture model-based clustering. switching on-and-off of chillers in the campus building;
� There are roughly 3 seasons resulted from 15-min measurement and
� At the first stage, various historical measurement data of the chilled 1-h mean value of cooling load profile, while 2 seasons from 4-h
water system is collected from the building energy management mean value and daily brief information (i.e. average, minimum
system; Data cleaning is then conducted to improve the quality of and maximum value of daily cooling load profile);
measurement data by excluding the missing values and detecting � The optimal quantity of components in the Gaussian mixture model
outliers in raw measurement data; Furthermore, data pre-processing for each time-step’s load profile is similar between the campus and
is conducted to obtain the 1-h mean value, 4-h mean value as well as the hotel building, which also demonstrates the robustness and
daily average, minimum and maximum value of cooling load to have generality of the proposed approach.
better understanding of the different trends of daily cooling load.
� At the second stage, data clustering is implemented on the 15-min CRediT authorship contribution statement
measurement, 1-h mean value, 4-h mean value as well as daily
average, minimum and maximum value of cooling load, respectively. X.J. Luo: Conceptualization, Methodology, Software, Writing -
Bayesian Information Criterion (BIC) value is then adopted to select original draft. Lukumon O. Oyedele: Funding acquisition, Project
the optimal quantity of components of Gaussian mixture model in administration. Olugbenga O. Olugbenga: Supervision, Writing - re­
each case. Therefore, each day of the year is grouped into different view & editing. Anuoluwapo O. Ajayi: Data curation.
clusters based on its daily cooling load profile.
� At the third stage, based on the year-round distribution and clustered Acknowledgement
cooling load profiles, data visualization of calendar view and
grouped cooling load profile is conducted to extract the unique The authors would like to acknowledge and express their sincere
feature of each cluster. gratitude to The Department for Business, Energy & Industrial Strategy
through grant project number TEIF-101-7025. Opinions expressed and
The successful implementation of the proposed pattern identification conclusions arrived at are those of the authors and are not to be
approach on two chilled water systems in different types of buildings has attributed to BEIS.
demonstrated its robustness and generality. The main findings from the

Nomenclature

A Total quantity of free parameters in GMM

BIC Bayesian Information Criterion
d Dimension, or quantity of elements in each vector
Cp Specific heat of chilled water (J kg-1 K -1)
G Quantity of operating chillers
K Quantity of components in Gaussian mixture models
L Likehood function
M Mass flow rate (L/s)
N Quantity of vectors in the database
Q 15-min cooling load
Q Vector of database for 15-min cooling load
Q Database of 15-min cooling load
r Relative error
T Temperature (K)
X 1-h mean cooling load

18
X.J. Luo et al. Journal of Building Engineering 31 (2020) 101339

X Vector of database for 1-h mean cooling load

X Database of 1-h mean cooling load
Y 4-h mean cooling load
Y Vector of database for 4-h mean cooling load
Y Database of 4-h mean cooling load
Z Average, minimum or maximum value of daily cooling load
Z Vector of database for daily brief information of cooling load
Z Database of daily brief information of cooling load
ω Weight of mixture component
φ Gaussian probability density function
μ Mean matrix of GMM
P
Covariance matrix of GMM
γ Posteriori probability

Subscripts
b Branch
c Chiller
g Number of operating chiller
i Day of the year
j Time step of the day
k Number of GMM
l Leaving
r Return

Abbreviations
GMM Gaussian mixture model

References [13] L. Wen, K. Zhou, S. Yang, A shape-based clustering method for pattern recognition
of residential electricity consumption, J. Clean. Prod. 212 (2019) 475–488.
[14] Kang L, Fei W, Zhao Z, Zeng M, Hong S, Chun L and Jing L. Analysis on residential
[1] X.J. Luo, K.F. Fong, Y.J. Sun, M.K.H. Leung, Development of clustering-based
electricity consumption behavior using improved k-means based on simulated
sensor fault detection and diagnosis strategy for chilled water system, Energy
annealing algorithm. In Power and Energy Conference at Illinois, IEEE1-6.
Build. 186 (2019) 17–36.
[15] Z. Guo, K. Zhou, X. Zhang, S. Yang, Z. Shao, Data mining based framework for
[2] X.J. Luo, L.O. Oyedele, A.O. Ajayi, C.G. Monyei, O.O. Akinade, L.A. Akanbi,
exploring household electricity consumption patterns: a case study in China
Development of an IoT-based big data platform for day-ahead prediction of
context, J. Clean. Prod. 195 (2018) 773–785.
building heating and cooling demands, Adv. Eng. Inf. 41 (2019) 100926.
[16] C. Fan, F. Xiao, H. Madsen, D. Wang, Temporal knowledge discovery in big BAS
[3] G.Y. Zhu, T.T. Chow, K.F. Fong, C. Lee, X.J. Luo, Design optimisation and
data for building energy management, Energy Build. 109 (2015) 75–89.
performance appraisal of a combined cooling, heating and power system primed
[17] F. Xiao, C. Fan, Data mining in building automation system for improving building
with Maisotsenko combustion turbine cycle, Energy Convers. Manag. 177 (2018)
operational performance, Energy Build. 75 (2014) 109–118.
91–106.
[18] K. Li, Z. Ma, D. Robinson, J. Ma, Identification of typical building daily electricity
[4] X.J. Luo, K.F. Fong, Development of multi-supply-multi-demand control strategy
usage profiles using Gaussian mixture model-based clustering and hierarchical
for combined cooling, heating and power system primed with solid oxide fuel cell-
clustering, Appl. Energy 231 (2018) 331–342.
gas turbine, Energy Convers. Manag. 154 (2017) 538–561.
[19] J. Wang, G. Li, H. Chen, J. Liu, Y. Guo, S. Sun, Y. Hu, Energy consumption
[5] X.J. Luo, K.F. Fong, Development of integrated demand and supply side
prediction for water-source heat pump system using pattern recognition-based
management strategy of multi-energy system for residential building application,
algorithms, Appl. Therm. Eng. 136 (2018) 755–766.
Appl. Energy 242 (2019) 570–587.
[20] C.M.R. Carmo, T.H. Christensen, Cluster analysis of residential heat load profiles
[6] X.J. Luo, K.F. Fong, Development of 2D dynamic model for hydrogen-fed and
and the role of technical and household characteristics, Energy Build. 125 (2016)
methane-fed solid oxide fuel cells, J. Power Sources 328 (2016) 91–104.
171–180.
[7] X.J. Luo, K.F. Fong, Investigation on part-load performances of combined cooling
[21] P. Gianniou, X. Liu, A. Heller, P.S. Nielsen, C. Rode, Clustering-based analysis for
and power system primed by solid oxide fuel cell with different bottoming cycles,
residential district heating data, Energy Convers. Manag. 165 (2018) 840–850.
J. Power Sources 429 (2019) 127–148.
[22] Y. Lu, Z. Tian, P. Peng, J. Niu, W. Li, H. Zhang, GMM clustering for heating load
[8] A. Ortega, M. Manana, Characterization and analysis of energy demand patterns in
patterns in-depth identification and prediction model accuracy improvement of
airports, Energies 10 (1) (2017) 119.
district heating system, Energy Build. 190 (2019) 49–60.
[9] T. Ahmad, H. Chen, Short and medium-term forecasting of cooling and heating
[23] F.W. Yu, K.T. Chan, Using cluster and multivariate analyses to appraise the
load demand in building environment with data-mining based approaches, Energy
operating performance of a chiller system serving an institutional building, Energy
Build. 166 (2018) (2018) 460–476.
Build. 44 (2012) 104–113.
[10] T. Ahmad, H. Chen, A review on machine learning forecasting growth trends and
[24] F.W. Yu, K.T. Chan, Assessment of operating performance of chiller systems using
their real-time applications in different energy systems, Sustain. Cities Soc. 2019
cluster analysis, Int. J. Therm. Sci. 53 (2012) 148–155.
(2019) 102010.
[25] J. An, D. Yan, T. Hong, Clustering and statistical analyses of air-conditioning
[11] T. Ahmad, H. Chen, R. Huang, G. Yabin, J. Wang, J. Shair, H.M. Akram, S.
intensity and use patterns in residential buildings, Energy Build. 174 (2018)
A. Mohsan, M. Kazim, Supervised based machine learning models for short,
214–227.
medium and long-term energy prediction in distinct building environment, Energy
[26] J. Han, J. Pei, M. Kamber, Data Mining: Concepts and Techniques, Elsevier, 2011.
158 (2018) 17–32.
[27] C.M. Bishop, Pattern Recognition and Machine Learning, springer, 2006.
[12] T. Ahmad, H. Chen, Y. Guo, J. Wang, A comprehensive overview on the data driven
[28] J.D. Banfield, A.E. Raftery, Model-based Gaussian and non-Gaussian clustering,
and large scale based approaches for forecasting of building energy demand: a
Biometrics (1993) 803–821.
review, Energy Build. 165 (2018) 301–320.

19