Building and Environment 149 (2019) 305–321

Contents lists available at ScienceDirect

Building and Environment

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

Thermal comfort, occupant control behaviour and performance gap – A T

study of office buildings in north-east China using data mining
Cheng Suna,b,∗, Ran Zhanga,b, Steve Sharplesc, Yunsong Hana,b, Hongrui Zhanga,b
a
School of Architecture, Harbin Institute of Technology, Harbin, 150001, China
b
Key Laboratory of Cold Region Urban and Rural Human Settlement Environment Science and Technology, Ministry of Industry and Information Technology, Harbin,
150001, China
c
School of Architecture, University of Liverpool, Leverhulme Building, Liverpool, L69 7ZN, United Kingdom

ARTICLE INFO ABSTRACT

Keywords: Simulation techniques have been increasingly applied to building performance evaluation and building en-
Window-opening behaviour vironmental design. However, uncertain and random factors, such as occupant behaviour, can generate a per-
Office building formance gap between the results from computer simulations and real buildings. This study involved a long-
Cold climate itudinal questionnaire survey conducted for one year, along with a continuous recording of environmental
Cluster analysis
parameters and behaviour state changes, in ten offices located in the severe cold region of north-east China. The
Association rules mining
offices varied from private rooms to open-plan spaces. The thermal comfort experiences of the office workers and
their environmental control behaviours were tracked and analysed during summer and winter seasons. The
interaction of the thermal comfort experiences of the occupants and behaviour changes were analysed, and
window-opening behaviour patterns were defined by applying data mining techniques. The results also gener-
ated window-opening behaviour working profiles to link to building performance simulation software. The aim
was to apply these profiles to further study the discrepancies between simulation and monitored results that arise
from real-world occupant behaviour patterns.

1. Introduction designers with a real perception of building system performance. The

reason for this difference is the large number of input parameters in the
In the process of architectural design and building energy-effi- simulation process and their non-linearity, discreteness, and un-
ciency-evaluation studies, various types of building performance si- certainty, represented by the user behaviour parameters in this study,
mulation techniques have become basic tools for building energy cal- which further increase the complexity of influencing elements [13].
culation, design optimisation, operation management, and building Among these uncertain input parameters, occupant behaviour is a
energy-saving diagnosis [1–5]. Performance-based simulation analysis major factor affecting the thermal comfort and energy efficiency of
methods and evaluation indices are also widely used for the energy- buildings, indicating the importance of establishing a behaviour mode
efficient design of new buildings, energy-saving renovation of existing for modelling and predicting building performance [14–16]. The in-
buildings, energy-efficient technology assessment, and formulation of fluence of occupant behaviour on buildings is greatly influenced by
energy-saving standards [6–9]. The process of architectural design is geographical regions and ethnic cultures [17,18]. This is reflected in
being transformed from result control to process control, and from se- many case studies, especially for residential or office buildings, which
paration of architectural geometric design and evaluation of building tend to have more individual controls [19–21]. Presently, the ex-
performance to the combination of those aspects [9,10]. ploration of the impact of different climates and cultural backgrounds
Although the potential of building performance-based design has on behaviour patterns still needs further development. An accurate
been widely recognised, it is still not possible to provide the best so- model is based on a wide range of data collection. Behaviour-control
lution for designing energy-efficient buildings due to the gap between data collection in recent research is derived from long-time recordings
real results and those expected from architectural design schemes and transverse questionnaire surveys [22,23]. Longitudinal ques-
[11,12]. This discrepancy also fails to give a true feedback on the im- tionnaires, with the characteristics of being time-consuming and la-
pact of a building design on performance; therefore, it does not provide bour-intensive, are relatively rare in occupant-behaviour studies, even

∗
Corresponding author. School of Architecture, Harbin Institute of Technology, Harbin, 150001, China.
E-mail address: suncheng@hit.edu.cn (C. Sun).

https://doi.org/10.1016/j.buildenv.2018.12.036
Received 2 September 2018; Received in revised form 26 November 2018; Accepted 14 December 2018
Available online 15 December 2018
0360-1323/ © 2018 Elsevier Ltd. All rights reserved.
C. Sun et al. Building and Environment 149 (2019) 305–321

Fig. 1. Work flow of the behaviour classi-

fication and building performance simula-
tion optimisation.

though the results, when combined with measured data, can provide window-opening behaviour profiles in selected offices.
more opportunities for exploring changes in behaviour control.
There are four main methods to examine the behaviour mode: 2. Methodology
agent-based modelling, statistical analysis, machine learning, and sto-
chastic modelling [24–29]. Zimmermann [30] first applied the agent- For the extreme Harbin climate of hot summer and cold winter, this
based modelling method to build simulation models for behaviour study established a data set from a long-term survey, with the appli-
control and motivating factors; Haldi and Robinson [31] studied the cation of statistical analyses and data mining techniques, to define
numerical relationship between occupant behaviour and other in- window-opening behaviour and attempted to fix the building perfor-
formation; D'Oca and Hong [32] used data mining to discover occu- mance simulation gap. A longitudinal survey was conducted for a one-
pancy patterns in office spaces; Erickson et al. [33] modelled and es- year period, interviewing for both subjective and objective variables
timated occupancy status and related energy consumption. These relating to occupant thermal comfort and adaptive behaviour. The basic
classic studies, with their different approaches, focused on different characteristics of occupant thermal comfort experiences and behaviour
aspects of behaviour, providing both theoretical support and applica- in the summer and winter were obtained. Logistic regression was ap-
tion guidance for determining patterns of occupant behaviour. plied to analyse the parameters influencing window-opening beha-
Research relating to occupant behaviour in Chinese buildings has viour. Data mining technology combed data, summarised rules, and
only been active in the last few years. For example, Yu [34] conducted a classified categories of these data, obtained from the longitudinal
winter and summer survey amongst elderly occupants to investigate questionnaire survey and field measurements in the summer and winter
their thermal comfort and adaptive behaviour characteristics in a hot seasons. Finally, behaviour profiles were obtained and linked into
summer/cold winter area of China; Song et al. [35] surveyed five office DesignBuilder, and then, the performance simulation was optimised.
rooms located in a cold region of China to identify the influencing Fig. 1 schematically shows the methodological approach.
factors of window-opening behaviour; Xin [36] focused on summer
window-opening behaviour triggers and classification in a hot summer/
2.1. Sample selection
cold winter part of China. Due to China's large regional differences in
climate, research on different climatic regions is imperative. In our
Harbin is a typical city in north-eastern China. It experiences a
previous studies, basic characteristics of the summer occupant beha-
temperate continental monsoon climate with four distinct seasons. The
viour were researched [37], and a comparison of the influencing factors
winter is long and cold, while the summer is short but hot. A district
and predictive models between different modes of occupant behaviour
heating (DH) scheme is widely applied in Harbin, with six months of
in offices were examined [38]. Furthermore, research on the simulation
uninterrupted winter heating. Ten volunteer offices distributed around
optimisation of building performance linked with an occupant beha-
six office buildings in representative districts of Harbin were chosen
viour configuration file is relatively scarce in the literature.
from the samples of the transverse survey, including private offices,
This study focuses on the interaction between thermal comfort and
shared-private offices, and open-plan offices (Fig. 2) [38]. In summer,
occupant behaviour in different-sized offices located in the north-
the background transverse survey revealed that it is uncommon for air
eastern China city of Harbin, which experience a severe cold winter
conditioning (AC) to be used in small-scale offices but was more com-
climate. The study involved a one-year longitudinal questionnaire
monly employed in large open-plan offices. In winter, district heating is
survey and logging of occupant environmental control behaviours in
the most common heating method, but a few buildings still use electric
winter and summer. Window-opening behavioural patterns were
heating (EH). Based on the characteristics of heating and AC systems,
identified using data mining techniques, with an attempt at classifying
typical offices buildings were selected to give a range of different types
the behaviour mode to reflect the characteristics of different behaviour
and sizes.
categories. Next, efforts were made to try and reduce the gap between
All basic building information, including the characteristics of
simulation results and real data by directly linking the behaviour modes
subjects, geometric parameters, and the available environmental
to simulation software, to improve the accuracy of the simulation and
equipment controls, are shown in Table 1. The surveyed offices include
reflect the real mechanism of the impact of occupant behaviour on
four private and shared-private offices, two open-plan offices with 3–10
building simulation in office buildings.
occupants, two open-plan offices with 11–20 occupants, and two open-
This study contributes to findings about thermal comfort and oc-
plan offices with more than 20 occupants. In summer, occupants in
cupant behaviour in different-sized offices with and without air con-
offices D1 and D2 were able to control single-unit AC, and those in
ditioning during the hot summer and cold winter in Harbin regarding
office D1 chose to switch-off the AC when feeling cold. The AC in
the following:
building D was removed in the second week of the summer survey for
equipment replacement. It should be noted that all the offices in this
• Long-term occupant thermal comfort and behaviour characteristics study are in buildings with east- or west-facing main façades, due to the
in private offices, shared-private offices, and open-plan offices;
limitations of the urban layout of the available buildings. However,
• Influencing factors of adaptive behaviour for both summer and according to background research, in Harbin, the main façade in most
winter;
office buildings are oriented east-west, rather than north-south. There
• Defining the window-opening behaviour duration patterns, window- were 80 occupants who completed the questionnaire survey. The
opening behaviour classification, and behaviour profiles in the hot
number of the occupants in each surveyed office was defined as per
summer season and cold winter period via data mining techniques;
ASHRAE Standard 55–2013 [39]. The ratio of male to female was close
• Modifying the building thermal performance gap and verifying the
to 1, similar to the results of transverse surveys.

306
C. Sun et al. Building and Environment 149 (2019) 305–321

Fig. 2. Site and location of the surveyed offices in Harbin, North-eastern China [38].

2.2. Longitudinal survey questionnaire was conducted for a week among workers in similar of-
fice environments; then, the official questionnaire survey was con-
2.2.1. Panel questionnaire survey ducted. The pilot questionnaires helped to improve the clarity of pre-
The panel questionnaire applied in this study was a survey that used sentation and integrity of the survey content, according to the opinions
the same subjects from the same office environments to track their and feedback from the subjects. Following the pilot survey, a three-part
changes in thermal comfort and adaptive behaviour during the year. questionnaire was conducted, consisting of a start survey, a daily
This panel questionnaire survey was designed so that the results could survey, and a final survey. The questionnaires were sent via WeChat,
be integrated with long-term monitoring behaviour data, so that the the most commonly used social media software in China, at 10:00 a.m.
interactive relationship between naturally ventilated behaviour, occu- in the morning and 3:00 p.m. in the afternoon for two weeks to help the
pant experience, and physical environment parameters could be ob- subjects develop the habit of answering the questions on time. The
tained. purpose of the start questionnaire was to obtain some basic information
Fig. 3 shows the framework of the methodology. A pilot from the subjects, their overall feelings of their office environment, and

Table 1
Basic information for the occupants of the surveyed offices and available facility controls.

Office No. Subjects No. (Surveyed) Male Office type Room size (m2) Orientation Available Control

Summer Winter

A1 1 (1) 0 Private 25.62 Northeast Fan EHc

A2 2 (2) 0 Shared-private 15.47 Southwest Fan EH
B 1 (1) 1 Private 21.74 Northeast Fan DHd
C 2 (2) 0 Shared-private 18.6 Northwest Fan DH
D1 5 (5) 2 Open plan 40.34 West ACa + fan DH
D2 5 (5) 4 Open plan 40.34 West ACa + fan DH
E1 15 (15) 6 Open plan 66.2 Southwest Fan DH
E2 11 (6) 6 Open plan 37.66 Southwest Fan DH
F1 50 (22) 15 Open plan 380 West ACb + fan DH
F2 50 (21) 9 Open plan 380 East ACb + fan DH

Notes:
a. AC is single unit air conditioning;.
b. AC is central air conditioning.
c. EH is electric heating; d. DH is district heating.

307
C. Sun et al. Building and Environment 149 (2019) 305–321

Fig. 3. Framework and content of the three-step panel questionnaire survey [38].

the range of adaptive control behaviours available to them under re- influencing factors of window-opening behaviour in north-east China.
strictive office conditions. The daily survey sought to obtain the occu- Cluster analysis was used to obtain the window-opening duration
pant clothing level, thermal comfort experiences on a seven-point scale patterns of the office occupants via monitoring data over one year. To
[39], and different behavioural status at the time of the questionnaire, form continuous and operational working user profiles, the time of the
involving the most concise and accurate questions. The final survey day was divided into six periods, early morning, morning, noon,
focused on summary questions about the overall thermal comfort ex- afternoon, evening, and night. The average performance of window-
perience during the two-week survey and the satisfaction with the opening behaviour duration in these time periods of summer and winter
questionnaire. was grouped using cluster analysis. The grouping results of weekends
and workdays were separately considered. The summer and winter
2.2.2. Field measurements window-opening duration patterns were then obtained via cluster
The spatial organisation of the office buildings, along with the analysis.
geometric design parameters of the monitored buildings, were mea- Association rules mining was then used to classify the behaviour
sured in detail to facilitate data analysis and simulation modelling, with the results of logistic analysis and cluster mining. Each office was
using an infrared rangefinder (model UT392). Indoor and outdoor classified into its own type, according to the influencing factors of
physical parameters were continuously recorded by a weather station window-opening behaviour and the window-opening duration patterns
(E-Log environmental data logger) and HOBO U12 data loggers (air in summer and winter. Finally, occupant window-opening behaviour
temperature and relative humidity) at 30-min and 15-min intervals, profiles were formed and then linked to the modelling of building
respectively (these intervals are also used by the dynamic building thermal performance, using the dynamic analysis software
energy modelling software DesignBuilder that was used in another part DesignBuilder.
of this study, to be described later). The cluster analysis and association rule mining were employed,
Occupant adaptive control, including the use of fans, AC use in along with the open source data mining program Rapid Miner, to mine
summer, and heating facilities in winter, were recorded by the panel the classification of the window-opening behaviour.
questionnaires at the time the questionnaires were answered, together
with the continuous measurement of window status using the Hobo UX 2.4. Statistical analysis technique
90-001 state/event data loggers. The number and duration of the status
changes were recorded. Due to equipment limitations, the size of Logistic regression analysis is a generalised linear regression ana-
opening could not be recorded, but the windows in the surveyed lysis model, an algorithm used for classification and prediction, which
buildings were all casement windows. This is the most common window characterises the influencing factors of nominal variables and the pre-
type for office buildings in Harbin, and based on the data statistics of dictive probability of the occurrence of events. To solve a problem of
the background transverse survey, they are usually opened fully in most regression or classification, a cost function is established; the optimal
cases. model parameters are iteratively solved by an optimisation method
and, finally, the quality of the model is verified. For binary logistic
2.3. Data analysis regression, when there are only two dependent variables (e.g. happen
or not happen), a regression analysis between conditional probability
Data mining generally refers to the process of searching for hidden P {Y = 1 x } and x is used, substituting the difficult method by at-
information in a large amount of data using algorithms. Fig. 4 shows the tempting to build the relationship between independent and dependent
work flow of the data mining in the classification and characterisation variables directly, which is equivalent to looking at a value in the do-
of the window-opening behaviour in the different sized Harbin offices. main of a continuous function from 0 to 1.
In this research logistic analysis was applied to analyse the influence Equations (1) and (2) describe this relationship of P and x :
factors of window-opening behaviour. The degree of association be-
Logit (P ) = ln P /(1 P) (1)
tween changes in window status and each parameter, including non-
nominal and nominal variables obtained from the panel questionnaires exp( 0 + 1 x1+ ...+ k xk )
and measured datasets, was examined by binary logistic regression. P=
1 + exp( 0 + 1 x1+ ...+ k xk ) (2)
Logic regression analysis results were combined with the measured
distribution characteristics of long-term behaviour to determine the where:

308
C. Sun et al. Building and Environment 149 (2019) 305–321

Fig. 4. Method for window-opening behaviour pattern and working profile definition.

P is the probability Rj is the average distance inside group j found by averaging the
P /(1 P ) is the odds ratio distance between each cluster data point and the cluster centre, and
Mij is the distance between the centre of each group.
The changes in behaviour status often correspond to the categorical
variables, such as window open and closed. In this study, binary logistic According to Equation (3), a smaller DBI value indicates better
regression was applied to define the relationship between the related performance for the cluster algorithm result. Groups with low DBI in-
variables and window-opening probabilities. The significance of the dicators represent clusters of low internal distances (i.e., high cluster
variables, based on a likelihood ratio test, using a 5% significance level, similarity) while high DBI indicators represent clusters of high internal
was tested to estimate the regression coefficients. distance (i.e., low cluster similarity).

2.5. Data mining techniques 2.5.2. Association rules mining

The purpose of applying association rules mining is to find the re-
2.5.1. Cluster analysis lationship between variables in large data sets and reveal the implicitly
Cluster analysis is processed to classify similar objects into different related features in the data [41]. The general form of association rules
groups or subsets by statistical classification, so that all the member mining can be presented by Equation (5).
objects in the same subset have similar attributes.
X Y (5)
In this study, the window-opening duration modes in the observed
offices were analysed using the K-means clustering approach [40]. This where:
method involves a vector quantisation of clusters and is the most
commonly used algorithm for basic clustering. For a data set D, K- X is the preceding item of the rule,
means clustering initially distributes the n data points in D into k Y is the latter item of the rule,
random clusters. Each cluster is associated with a centroid (centre X and Y can be a project or an item set from the data set.
point), and the distance from each data point to all k centroids is cal-
culated. A data point is then assigned to the cluster whose centroid is Although many association rules relationships can be identified via
closest to it so that similar data points can be gathered together. It is an the method, only a few of the relationships may be valid. There are two
iterative method, and the next iteration calculates the new centroids of values for evaluating the validity of the mining results: Confidence and
these new clusters by calculating the average of the distances between Support.
the points and the centroid. This iteration continues until convergence Confidence is the measurement of the accuracy of the association
is achieved. rules. It describes the probability of item Y containing item X , and
The similarity between clusters is usually evaluated via the distance reflects the possibility of Y appearing under the condition of X . If the
between groups, and the distance is obtained through the measured confidence level is high, the possibility of the emergence of X is high,
Euclidean distance (Equation (3)). reflecting the conditional probability of Y under a given X . Its formula
can be described as Equation (5)
d (a , b) = d (b, a) = (b1 a1)2 + (b2 a2)2 + (bn an ) 2 (3)
T (X Y )
where, CX Y =
T (X ) (6)
a = (a1, a2 , an ) , where:
b = (b1, b2 , bn) ,
T (X ) represents the number of transactions that contain the project
and a and b are two points in Euclidean space. X,
The performance of clustering was evaluated using the Davies- T (X Y ) means the number of transactions that contain both the
Bouldin Index (DBI) index. The DBI index refers to the ratio between project X and the project Y .
the average distance in the group and between the groups (Equation
(4)) Support measures the universality of the association rules and re-
n
Ri + Rj presents the probability of the concurrent occurrence of project X and
1
E= maxi j project Y , and the formula is
n i=1
Mij (4)
T (X Y)
where: SX Y =
T (7)

n is the group number, where:

309
C. Sun et al. Building and Environment 149 (2019) 305–321

T represents the total number of transactions outdoor temperature cooled down, the occupants of offices with 3–10
persons (offices D1 and D2) felt very hot. Occupants felt the same level
Confidence and Support can only measure the validity of the results of warmth in offices of 10–20 people during all the survey runs, in-
of association rules, but they fail to measure whether the results are dicating that the average value of the thermal sensation vote changed
practical. Therefore, the index of lift is applied to measure whether the little when the outdoor temperature varied between 26.9 °C and
appearance of X can motivate Y . The index of lift, shown in Equation 30.2 °C. The thermal sensation vote of office occupants using AC
(7), is the ratio of Confidence to later Support, and the greater the value equipment was neutral because of the stable indoor environment.
is, the better are the results. In winter, most of the mean thermal sensation vote results were in
the level between cool and neutral, except those from the open-plan
CX Y T (X Y ) T (Y )
LX Y = = / offices with 3–10 occupants in which the value was between cool and
SY T (X ) T (8)
cold (Fig. 6). According to the line chart of the mean thermal sensation
In this study, the frequent pattern growth algorithm (FP-Growth vote, divided by office size, all occupants' experiences in the winter
algorithm) was applied to define the classifications of window-opening were between cool and warm.
behaviour with the results of influencing factors and the duration From the thermal satisfaction vote results in Fig. 7, only occupants
modes of window-opening. in offices with three to ten persons experienced low satisfaction below
‘dissatisfied’ in most cases in summer. The average level of thermal
satisfaction corresponded to the level of thermal sensation in offices
3. Results and discussion
with 3–20 and > 20 occupants. Private and shared-private offices had
a neutral assessment about the indoor environment, despite the hot
The results were analysed to define the window-opening behaviour
experience of the thermal sensation, which may be because the occu-
type for modifying the simulation gap. The main outcomes are sum-
pants working in the more independent office environments had more
marised as follows:
control over their behaviour. In winter, the occupants in offices with
3–10 persons also presented a low level of satisfaction, while others in
3.1. Thermal comfort characteristics the range were ‘a little dissatisfied’ to ‘a little satisfied’, which is con-
sistent with the mean thermal sensation voting (Fig. 8).
During the summer season in Harbin, the indoor temperature Tin
was maintained at around 30 °C in the surveyed natural ventilation
buildings, and around 27.5 °C in the AC offices. The Chinese evaluation 3.2. Behaviour control characteristics
standards for indoor thermal environments in civil buildings (GB/T
50785-2012) [42] limit the range of Tin to between 18 °C and 28 °C, One day time was divided into six intervals to obtain the formation
which means that the indoor temperature of all naturally ventilated of a continuous window-opening behaviour profile, containing early
offices were in the uncomfortable range. In winter, the indoor tem- morning time, morning time, noon time, afternoon time, evening time,
perature of all surveyed offices was in the comfortable temperature and night time (Table 2).
range. Table 2 presents the average value of window-opening duration in
During the summer survey, the outdoor temperatures were gen- each surveyed office on workdays and weekends in summer and winter,
erally lower for the second half of the questionnaire, with the average with the value of the variance measuring the dispersion of the recording
value dropping from 30.2 °C to 26.9 °C. In winter, the change was up- data. Around 50% of the surveyed offices from small-to large-scale
wards, from −17.2 °C to −13.5 °C. Fig. 5 to Fig. 8 shows the scatter exhibited a window-opening time of no closures across the entire
plots of average thermal sensation and thermal satisfaction votes for summer typical season in July during day and night, which means the
each surveyed office building, and the line chart shows the mean value occupants never closed the window during this period. It is worth
of votes for occupants from offices of different sizes and layouts. noting that most of these buildings also showed the extremely opposite
The summer thermal sensation votes show that all kinds of office performance of having totally closed windows during the days and
buildings with the same size had a certain degree of consistency nights on weekdays and weekends in December and January.
(Fig. 5). Occupant sensations in offices with 1–2 persons were very hot In summer, four rooms (A2, D1, D2, D2', and F2) kept their windows
in the first week and neutral in the second week due to the decline in in an open state during the work time, with most of the occupants
outdoor temperature. After the removal of AC equipment, although the keeping to a routine of opening the window when they arrived and

Fig. 5. Scatter plots and line chart of average thermal sensation vote of occupants in different-sized offices in summer.

310
C. Sun et al. Building and Environment 149 (2019) 305–321

Fig. 6. Scatter plots and line chart of average thermal sensation vote of occupants in different-sized offices in winter.

Fig. 7. Scatter plots and line chart of average thermal satisfaction vote of occupants in different-sized offices in summer.

Fig. 8. Scatter plots and line chart of average thermal satisfaction vote of occupants in different-sized offices in winter.

closing the window when they left. In winter, among these surveyed In Harbin, it is a very common phenomenon that people work
offices, D1, D2, and F2 rooms also showed a short duration of opening overtime on weekends. In summer, the windows in the offices with day-
windows at the time of people's arrival or their lunch break. night window-opening behaviour were open on weekends throughout

311
C. Sun et al. Building and Environment 149 (2019) 305–321

Table 2
Window-opening duration of weekdays and weekends in each surveyed office in summer and winter with average and variance value.

Summer results

Window-opening duration on work days (hours) (average value (variance)) Window-opening duration on weekends (hours) (average value (variance))

6 a.m.-9 am 9am-12am 12am-3pm 3pm-6pm 6pm-9pm 9pm-12pm 6–9am 9am-12am 12am-3pm 3pm-6pm 6pm-9pm 9pm-12pm

A1 3 (0) 3 (0) 3 (0) 3 (0) 3 (0) 6 (0) 3 (0) 3 (0) 3 (0) 3 (0) 3 (0) 6 (0)
B 3 (0) 3 (0) 3 (0) 3 (0) 3 (0) 6 (0) 3 (0) 3 (0) 3 (0) 3 (0) 3 (0) 6 (0)
A2 0.45 (0.2) 2 (0.1) 2 (0.6) 0.25 (0.1) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0)
C 3 (0) 3 (0) 3 (0) 3 (0) 3 (0) 6 (0) 3 (0) 3 (0) 3 (0) 3 (0) 3 (0) 6 (0)
D1 1.25 (0.1) 3 (0.1) 2.45 (0.2) 2 (0.2) 0.25 (0) 0 (0) 0 (0) 1 (0.9) 1 (0.2) 0.5 (0.1) 0 (0) 0 (0)
D2 1 (0.3) 1 (0.4) 0.25 (0.2) 0.25 (0.2) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0)
D2′ 1 (0.4) 2 (0.2) 2 (0.2) 1.25 (0.1) 0 (0) 0 (0) 0.5 (1) 2 (0.2) 2 (0.4) 0.75 (0.8) 0 (0) 0 (0)
E1 3 (0) 3 (0) 3 (0) 3 (0) 3 (0) 6 (0) 3 (0) 3 (0) 3 (0) 3 (0) 3 (0) 6 (0)
E2 1.5 (1.2) 2.5 (1.4) 2.5 (0.4) 2.5 (0.4) 2.25 (1) 5 (2) 2.5 (1) 3 (0.1) 3 (0.1) 3 (0.6) 2.5 (1) 6 (1)
F1 3 (0) 3 (0) 3 (0) 3 (0) 3 (0) 6 (0) 3 (0) 3 (0) 3 (0) 3 (0) 3 (0) 6 (0)
F2 1.25 (0.2) 3 (0.1) 2.45 (0.3) 2 (0.1) 0.25 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0)

Winter results

Window-opening duration on work days (hours) (average value (variance)) Window-opening duration on weekends (hours) (average value (variance))

6am-9am 9am-12am 12am-3pm 3pm-6pm 6pm-9pm 9pm-12pm 6–9am 9am-12am 12am-3pm 3pm-6pm 6pm-9pm 9pm-12pm

A1 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0)
B 0.02 (0) 0.05 (0) 0.1 (0.1) 0.1 (0.1) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0)
A2 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0)
C 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0)
D1 0.25 (0) 0 (0) 0.05 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0)
D2 0.1 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0)
E1 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0)
E2 0.04 (0) 0.07 (0) 0.2 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0)
F1 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0)
F2 0 (0) 0.16 (0) 0.04 (0) 0.02 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0)

Notes:
a. A1 and B are private offices, A2 and C are shared-private offices, D1, D2 (3–10 persons, AC was closed or removed for equipment update), D2’ (with AC on in
summer), E1 and E2 (11–20 persons) are open-plan offices, F1 and F2 (> 20 persons) are open-plan offices (with AC in summer).
b. The offices on bold were those open all the time in summer and closed all the time in winter.

the summer. The other offices of the ‘routine type’ presented a greater throughout the year, temperature and relative humidity indoors and
dispersion of window-opening behaviours, which may be due to in- outdoors, as well as the corresponding thermal sensation, temperature
creased randomness of the overtime work period on the weekends. In preference, and humidity feelings, were related. Meanwhile, the spe-
winter, the window-opening behaviour of occupants of all the surveyed cific statistics from the summer and winter analysis showed that the
office buildings was remarkably consistent, that is, no window-opening window-opening behaviour in the surveyed office rooms did not show
at all during the winter season, which may be due to the fact that the high correlation to the temperature or humidity change in an individual
cold winter reduced the overtime hours or the overtime on weekends in season.
winter was not too long. Due to the characteristics of window-opening In summer, the occupant window status varied with outdoor tem-
behaviour, most surveyed office rooms showed a correlation with habit perature only in offices F1 and F2. Correspondingly, the thermal sen-
in winter and summer. sation feeling influenced the window-opening behaviour of occupants
in F1 and F2, and temperance preference in F2. The window-opening
3.3. Behaviour influencing factors changes were also influenced by indoor relative humidity and humidity
feelings for the occupants in F1 and F2. Occupants in F1 also thought
The correlation between potential influencing factors and window- the air movement and overall satisfaction were the reasons for their
opening behaviour was analysed in summer, winter, and two quarters behavioural changes towards the window.
to assess the main factors affecting behaviour in a single season and In winter, there were only two surveyed offices presenting corre-
across different seasons (Table 3). The correlation of physical para- lation of environmental physical parameters and thermal sensation
meters, consisting of indoor and outdoor temperature and relative hu- evaluation vote. There was significant correlation for the temperature
midity, with the window state was determined using logistic analysis. preference, air movement, and overall satisfaction with the window
The occupants' experience of the thermal environment were also in- status in office B but no physical factors, and in addition to the tem-
cluded in the discussion for further understanding of the interaction perature preference, users of office F2 were affected by outdoor tem-
between the occupants' thermal comfort experience and the window- perature, indoor humidity, and the corresponding thermal sensation
opening behaviour control. Nominal variables, e.g. season and and humidity feeling.
morning/afternoon, were tested for correlation with the window- In summary, for all surveyed buildings, the window status change
opening behaviour. had the strongest correlation with the seasons (Table 4). Changes in
According to the basic features of window-opening of office build- physical parameters within a certain threshold in summer and winter
ings in different scales, there are significant differences in the opening only affected the window-opening behaviour of users in a few build-
duration in winter and summer, and this was also verified by the result ings. In Section 3.2, the statistical results of window-opening duration
of correlation analysis. From the perspective of data analysis showed that some of the occupants exhibited the habit of opening all

312
C. Sun et al. Building and Environment 149 (2019) 305–321

Table 3
Influencing factors of window-opening behaviour in summer, winter, and across the entire year.

Nominal variables Non-Nominal variables

Season Tin Tout RHin RHout Clo

S W H S W H S W H S W H S W H

A1 ✓ × × ✓ × × ✓ × × ✓ × × × × × ✓
B ✓ × × × × × × × × × × × × × × ×
A2 ✓ × × ✓ × × ✓ × × ✓ × × ✓ × × ✓
C ✓ × × × × × × × × × × × × – × ×
D1 ✓ × × ✓ × × ✓ × × ✓ × × ✓ × × ×
D2 ✓ × × ✓ × × ✓ × × ✓ × × ✓ × × ×
E1 ✓ × × ✓ × × ✓ × × × × × ✓ × × ✓
E2 ✓ × × ✓ × × ✓ × × ✓ × × × × × ✓
F1 ✓ × × ✓ ✓ × ✓ ✓ × ✓ × × × × × ✓
F2 ✓ × × ✓ ✓ ✓ ✓ ✓ ✓ ✓ × × × × × ✓

Interval variables

Thermal sensation feeling Temperature preference Humidity feeling Air movement Overall satisfaction

Office No. S W H S W H S W H S W H S W H
A1 × × ✓ × × ✓ × × ✓ × × × × × ✓
B × × ✓ × ✓ ✓ × × × × ✓ × × ✓ ✓
A2 × × ✓ × × ✓ × × × × × × × × ×
C – × ✓ – × ✓ – × ✓ – × × – × ✓
D1 × × ✓ × × ✓ × × × × × × × × ✓
D2 × × ✓ × × ✓ × × × × × × × × ✓
E1 × × ✓ × × ✓ × × ✓ × × ✓ × × ✓
E2 × × ✓ × × ✓ × × ✓ × × × × × ✓
F1 ✓ × × × × × ✓ × ✓ ✓ × ✓ ✓ × ×
F2 ✓ ✓ ✓ ✓ ✓ ✓ ✓ × ✓ × × × × × ×
Tested but not relevant parameters Smell; Outdoor noise; Outdoor Air Quality;
Time: (Morning or afternoon); fan use; AC use

Notes:
√ means p < 0.05, the correlation is significant; × means p > 0.05, the correlation is not significant.
- means no data was collected because of holidays and other reasons.
S is summer; W is winter; and H is the results considering winter and summer.
Tin is indoor temperature; Tout is outdoor temperature; RHout is outdoor relative humidity; Rhin is indoor relative humidity; Clo is thermaI resistance of clothing;
No. is number.

Table 4 the behaviour.

Statistics of window-opening behaviour influencing factors across the whole
year.
3.4. Window opening duration patterns
Influencing factors Offices No.
With the application of cluster analysis, three window-opening
A1 B A2 C D1 D2 E1 E2 F1 F2
duration patterns in summer and four patterns in winter were obtained
Season ✓ via the data of occupant performance of window-opening behaviours in
Thermal comfort × √a × × × × × × √b √c offices from private offices to open-plan offices in north-east China.
Habit ✓ × ✓

Notes: 3.4.1. Summer window-opening duration patterns

a.√ means the factor affect window-opening behaviour in winter. In the typical summer month of July, according to the statistics in
b.√ means the factor affect window-opening behaviour in summer. Table 2, some of the surveyed offices exhibited a window open time of
c.√ means the factor affect window-opening behaviour in summer and winter. the entire duration of the month, while others showed a close con-
nection between the working hours and window open time. The cluster
results (Fig. 9) agree with these basic statistics of the window-opening
windows in summer and closing all windows in winter. Some other behaviour, and three types are defined.
offices kept the routine of opening the window during work time in The first type was where windows were open the entire time on
summer and winter. Office E2 was the only office with random window- weekdays, which consisted of private offices A1 and B, shared-private
opening behaviour. Combining with the analysis of the basic features of office C, and open-plan offices E1 and F1. The offices belonging to this
the window-opening behaviour in the previous section, it can be type were also those with the continuous window-opening behaviour
speculated that the main factors influencing behaviour included season, during the weekends. Type two was the “working time opening routine”
habit, and thermal comfort experiences, and this will be further tested example, which was also consistent with the statistics of Table 2,
in the next section as the premise input for association rules mining, via comprising shared-private office A2 and open-plan offices D2 (with and
classifying the categories of behaviour. Other variables, including without single unit AC) and F2 (with central AC). For this type, the
smell, outdoor noise, outdoor air quality, time (morning or afternoon), occupants used the window during the work time period. This pattern
and other behaviour were also tested but showed no correlation with of behaviour similarly occurred during the weekends. After cluster

313
C. Sun et al. Building and Environment 149 (2019) 305–321

Fig. 9. Cluster-mining results of summer window-opening behaviour duration for the ten surveyed offices.

314
C. Sun et al. Building and Environment 149 (2019) 305–321

Fig. 10. Cluster-mining results of winter window-opening behaviour duration for the ten surveyed offices.

315
C. Sun et al. Building and Environment 149 (2019) 305–321

mining, E2 belonged to a separate category, and it also had the phe- 3.5. Window-opening behaviour classification and profiles
nomenon of open window during the day and night, but its duration
was shorter than the duration of type 1 in each period of time. It can be Association rules mining was applied for classifying the window-
seen from Table 2 that the duration of the E2 open window in each part opening behaviour together with the summer and winter opening
of the day has great discreteness. The duration of occupants working in duration and influencing factors. To get significant results from the
E2 had a larger randomness by the monitoring results from the space association rules mining, the values of Support, Confidence, and lift were
occupancy sensor HOBO UX90-006x, which may lead to this discrete- set at the minimum thresholds of 30%, 80%, and 1, respectively. The
ness. criteria were prescribed for each rule mined, which is that at least 30%
of the data contained the premise and conclusion, in which the prob-
ability that a premise led to a conclusion was greater than 80%.
3.4.2. Winter window-opening duration patterns Simultaneously, all the results mined were positively correlated with
Data from the typical two winter months, December and January, lift > 1. Finally, five types of window-opening modes for summer and
were included in the data mining. Four types of window-opening be- winter seasons were obtained based on mining of the monitoring da-
haviour duration were defined (Fig. 10). Table 2 also reveals that, on tabase for the entire year. The types and modes are summarised in
the weekends, all windows of the monitored offices were in the state of Fig. 11.
being totally closed during these two months. Four of these types of window-opening behaviour modes were all-
Type 1 was windows opened for a short time of 1 min or less during season and habit-motivated, except Type 5. The seasons are the most
the early morning and morning time, which included five situations in influential factors of the occupant window-opening behaviour in all
which offices had entirely closed windows, except office D2. The other types of office buildings. With the great changes in the physical data of
three types of window-opening behaviour in winter involved a rela- temperature and relative humidity in different seasons, the behaviour
tively longer open window time of about 15 min, but each of the open of window-opening had an extremely large impact. Simultaneously,
times were concentrated at different time periods: type two was in the these types of behavioural modes were also significantly driven by
early morning, type three was in the morning time, and type four was in behavioural habits, in which occupant behaviour during each of the
the noon time. These three types of windows also had a very short time seasons presented a stable window-opening duration.
of open window, 1–3 min during other periods of the day. Six of the surveyed offices showed an agreement between the

Fig. 11. Classification results of the window-opening behaviour with the duration patterns and influencing factors as premise conditions in summer and winter.

Table 5
Summary of window-opening behaviour classification of office occupants.

Office Type Office number Duration mode Motivation factors

Private office A1 Type 1 AO(S)+AC(W) Season, habits

B Type 4 AO(S)+WO(W) Season, habits, thermal comfort (in winter)
Shared- Private office A2 Type 2 WO(S)+AC(W) Season, habits
C Type 1 AO(S)+AC(W) Season, habits
3-20 Open-plan office D1 Type 3 WO(S + W) Season, habits
D2 Type 2 WO(S)+AC(W) Season, habits
E1 Type 1 AO(S)+AC(W) Season, habits
E2 Type 5 Type 3 + WO(W) Season
> 20 Open-plan office F1 Type 1 AO(S)+AC(W) Season, habits, thermal comfort (in summer)
F2 Type 3 WO(S + W) Season, habits, thermal comfort

Notes: S is summer; W is winter.

316
 Table 6
Window-opening behaviour profile of occupants in different-sized offices on weekdays and weekends during summer and winter season, divided into six periods.
C. Sun et al.

Duration of window-opening behaviour (hours)

Summer

Weekdays Weekends

Type 1: A1, C, E1, F1

6–9am 9–12am 12am-3pm 3–6pm 6–9pm 9pm-6am 6–9am 9–12am 12am-3pm 3–6pm 6–9pm 9pm-6am
3 3 3 3 3 9 3 3 3 3 3 9

Type 2: A2,D2

6–9am 9–12am 12am-3pm 3–6pm 6–9pm 9pm-6am 6–9am 9–12am 12am-3pm 3–6pm 6–9pm 9pm-6am
1 1.75 1.5 1 0 0 1 1 1 0.25 0 0

Type3: D1

6–9am 9–12am 12am-3pm 3–6pm 6–9pm 9pm-6am 6–9am 9–12am 12am-3pm 3–6pm 6–9pm 9pm-6am
1 1.75 1.5 1 0 0 1 1 1 0.25 0 0

Type3: F2

6–9am 9–12am 12am-3pm 3–6pm 6–9pm 9pm-6am 6–9am 9–12am 12am-3pm 3–6pm 6–9pm 9pm-6am
1 1.75 1.5 1 0 0 1 1 1 0.25 0 0

Type 4: B

317
6–9am 9–12am 12am-3pm 3–6pm 6–9pm 9pm-6am 6–9am 9–12am 12am-3pm 3–6pm 6–9pm 9pm-6am
3 3 3 3 3 9 3 3 3 3 3 9

Type 5: E2

6–9am 9–12am 12am-3pm 3–6pm 6–9pm 9pm-6am 6–9am 9–12am 12am-3pm 3–6pm 6–9pm 9pm-6am
1.5 2.5 2.5 2.5 2.25 5 1 1 1 0.25 0 0

Duration of window-opening behaviour (hours)

Summer Winter

Weekdays Weekdays Weekends

Type 1: A1, C, E1, F1

6–9am 6–9am 9–12am 12am-3pm 3–6pm 6–9pm 9pm-6am 6–9am 9–12am 12am-3pm 3–6pm 6–9pm 9pm-6am
3 0 0 0 0 0 0 0 0 0 0 0 0

Type 2: A2,D2

6–9am 6–9am 9–12am 12am-3pm 3–6pm 6–9pm 9pm-6am 6–9am 9–12am 12am-3pm 3–6pm 6–9pm 9pm-6am
1 0 0 0 0 0 0 0 0 0 0 0 0

Type3: D1

6–9am 6–9am 9–12am 12am-3pm 3–6pm 6–9pm 9pm-6am 6–9am 9–12am 12am-3pm 3–6pm 6–9pm 9pm-6am
(continued on next page)
Building and Environment 149 (2019) 305–321
C. Sun et al. Building and Environment 149 (2019) 305–321

summer and winter window-opening behaviour modes, which were all

open in summer season and all closed in winter season (AO, AC) for

9pm-6am

9pm-6am

9pm-6am
Type 1, and work-time open for summer and winter for Type 3. For
Type 3, the occupants opened the window when they arrived, whilst in

0
summer the duration was close to full-time open during the working
hours; in winter the duration was reduced to less than 15 min. This
meant that the occupants of this type of office, who were executing the
work-time window-opening behaviour mode (WO) still maintained this

6–9pm

6–9pm

6–9pm
state in winter, with a significantly reduced duration of window-
0

0
opening.
In Type 1, office F1 was also motivated by thermal comfort ex-
periences, including thermal sensation, humidity feeling, and air
movement in summer. In Type 3, only F2 showed similar window-
3–6pm

3–6pm

3–6pm
opening behaviour factors, consisting of temperature preference and
0

0
humidity feelings. The occupant behaviour of Type 2 exhibited work-
time opening in summer and all-closed in winter with season and habit
as the motivations. Office B in Type 4 exhibited an all-opening mode in
12am-3pm

12am-3pm

12am-3pm
summer and work-time opening in winter during the noon time, with a
short open duration of about 15 min. The behaviour mode changed in
winter due to the increasing need for thermal satisfaction and better air
0

flow. Type 5, office E2, has also been described in the previous dis-
cussion, and it was noted that, because of the greater flexibility in
working hours, its window-opening behaviour showed a strong corre-
9–12am

9–12am

9–12am

lation only with the seasons, and the window-opening duration did not
demonstrate patterns consistent with other category types in the
0

summer. E2 did show the same performance as office B with a short

open duration of 15 min in the noon period in winter.
The categorised types are classified by office scales from private
Weekends

6–9am

6–9am

6–9am

office to open office. Table 5 shows the types of window-opening pat-

terns and motivational factors of occupant window-opening behaviour
0

in different-sized offices. Table 6 shows the results of window-opening

behaviour profile of occupants in different-sized offices on weekdays
and weekends during summer and winter season. For the private office,
9pm-6am

9pm-6am

9pm-6am

offices A1 and B had the same mode of full-time opening all summer,
while there was a different behaviour in winter, with all the windows
0

closed in office A1 and open for a short duration in office B. For the
shared-private office, offices A2 and C showed the same performance of
all windows being closed in winter, while A2 belonged to the all-open
mode, and C exhibited the work-time opening mode in summer. For the
6–9pm

6–9pm

6–9pm

open-plan offices with 3–20 occupants, offices D1, D2, E1, and E2 were
0

not consistent with each other, displaying four different modes. It

should be noted that D2 retained the habit of opening the window at
the time of occupant arrival, but D2 was classified as a fully closed type
due to its very short window-opening time when cluster analysis was
3–6pm

3–6pm

3–6pm

performed. For the open-plan offices of more than 20 occupants, the

0

two surveyed offices also showed inconsistent window-opening dura-

tion patterns, but their window-opening behaviour was affected by
season, habit, and thermal comfort both in summer and by season, and
12am-3pm

12am-3pm

12am-3pm

by habit in winter. The offices with a thermal comfort experience mo-

tivation were all in the common range of indoor temperature and re-
0.25

0.25

lative humidity, with no large difference as in the other surveyed of-

0

fices.
After obtaining the window-opening behaviour classification, be-
Duration of window-opening behaviour (hours)

havioural profiles were finally formed, which considered the results of

9–12am

9–12am

9–12am

the mode classification definitions and the temporal degree of sub-

0.25

division of the DesignBuilder software. Usually, the time step was set as
0

15-min intervals to obtain a suitable simulation speed. The window-

opening durations of less than 15 min cannot be calculated because it is
too short a duration under this setting. The results are listed in Table 4.
Weekdays
Winter

6–9am

6–9am

6–9am
Type 1: A1, C, E1, F1

0.25
Table 6 (continued)

3.6. Building simulation optimisation

0

Four offices from two of the surveyed buildings were selected to

Type 5: E2
Type3: F2
Weekdays

Type 4: B

investigate whether the application of behavioural models, by linking

Summer

6–9am

6–9am

6–9am

the new window-opening behavioural profile into DesignBuilder soft-

1.5
1

ware, could be effective in reducing the differences between simulated

318
C. Sun et al. Building and Environment 149 (2019) 305–321

and real environmental data. The weather file provided by the only the data on July 26 were simulated for comparison with the
DesignBuilder resource platform was not a 2017 weather file (the year template inside DesignBuilder and the behaviour modes of this research
of this study). To overcome this, the outdoor weather station conditions (offices D1 and D2 were natural ventilated rooms due to the AC being
measured in the study were matched with two days in summer and one removed in the second half of July). In winter, the outdoor temperature
day in winter that were very similar to the DesignBuilder weather file was very low. When the DesignBuilder window-opening schedule
data. These days (July 10 and 26 and December 6) were used in the template was used, a full-time window-opening mode during working
simulations. A calculation method for infiltration into DesignBuilder hours, the windows were closed after the indoor temperature was re-
was used by entering the behaviour pattern code into the software. The duced to a certain extent.
calculated method needed to meet the conditions is shown in Equation These four offices respectively belonged to Type 1, Type 2, Type 3,
(8) and Type 2. The original window-opening behaviour mode inside the
software was due to the work time of the occupants. Figs. 12 and 13
AND AND
Tzone _ air Tsetpo int Tzone _ air Toutside _ air the schedule value (9) present a comparison of the percentage difference of indoor tempera-
ture during the simulated work time with behaviour patterns of no
In summer, for an office with AC, for example D1 in the first half of behaviour (no opening), mode template inside DesignBuilder, and the
July, the windows would be closed when the outdoor temperature was real pattern from the data mining results in summer, with the real
higher than the indoor one. Therefore, it was impossible for an office measured data as the baseline. The result of the winter real mode was
with AC to be simulated, as the real scenario involved the occupants very close to the no window-opening behaviour; therefore, only
using the windows and AC together. Therefore, for offices D1 and D2,

Fig. 12. Indoor temperature discrepancy applying no behaviour control, mode inside DesignBuilder, and real window-opening behaviour modes mined for private
office A1 and shared-private office A2, with the measured temperature as the baseline during work time.

Fig. 13. Indoor temperature discrepancy applying no behaviour control, mode inside DesignBuilder, and real window-opening behaviour modes mined for open-plan
offices D1 and D2, with the measured temperature as the baseline during work time.

319
C. Sun et al. Building and Environment 149 (2019) 305–321

comparisons of the simulation results with the real mode and the the open-plan offices. Due to limited volunteer participation of the in-
template are shown. vestigated offices, using offices from just one building to reduce the
In summer, the mode detail of work-time opening type (WO) ob- interference of other variables was difficult. This study considered ten
tained from the cluster analysis was very close to the DesignBuilder offices of different sizes as the research objects and obtained some
mode (Office A2, D1, and D2). The results of all behaviour types with meaningful results. Extensive research of more building types is still
the summer WO pattern were close to those of mode template inside necessary for further discussion.
DesignBuilder. The discrepancy between the real all-opening pattern The behaviour influencing factors shown in this study do not match
(AO, office A1) and the DesignBuilder one was around 2.5% and 0.8 °C. the previous work by Xin [36], which found that in the summer season
All the rooms with the no behaviour control showed a relatively high the environmental variables lose their predictive power of window-
value, while the differences for office D1 and D2 were smaller, which opening probability. This may be due to the distinct seasons with larger
may be because the nearby rooms were all offices with AC. temperature differences in north-east China. Season had a great influ-
In winter, due to the calculation method of DesignBuilder, the ence on window-opening behaviour in this study. In a single season, the
window always changed to closed status when the temperature goes change of window-opening behaviour of most occupants was very in-
down and, in particular, when the temperature was lower than the set active. These inactive occupants did not change the state of windows
point temperature for heating. For all the offices in building A (Fig. 13), with temperature. Some of the results from this study agree with the
with the AC mode in winter, the difference between the template in DB work of Song [35] about the influencing variables, while the findings of
and the real mode result was around 13% with a temperature difference this study showed that season and habit are the major affecting para-
of 4.5 °C. In building D (Fig. 13), the real pattern of D1 was with a 15- meters. The results of this study prove again the necessity of research on
min opening in the early morning, and D2 was the all-closed mode. The occupant behaviour to help revise and refine simulation results from
discrepancy of D1 was 8% and 1.5 °C, and for D2, it was 10% with a building performance software.
temperature difference of 1.8 °C.
Acknowledgements
4. Conclusion and limitation
The authors would like to acknowledge that this paper was finan-
This study applied data mining techniques to obtain the real occu- cially supported by the China National Key R&D Program (Grant No.
pant window-opening behaviour modes during a one-year period that 2016YFC0700200), the National Natural Science Foundation of China
involved longitudinal questionnaire surveys and behaviour state re- (Grant No. 51708149), and the China Postdoctoral Science Foundation
cording of different-sized offices in the severe cold winter climate of (Grant No. 2017M621276).
Harbin. Window-opening duration patterns using cluster analysis and
influencing factors motivating behaviour via logistic analysis were de- References
fined for conditions to further classify the behaviour modes and form
the behavioural profiles that were used in building performance simu- [1] G. Augenbroe, Trends in building simulation, Build. Environ. 37 (2002) 891–902.
lation software. The findings of this study can be concluded as: [2] T. Hong, S.K. Chou, T.Y. Kong, Building simulation: an overview of developments
and information sources, Build. Environ. 35 (2000) 347–361.

• In summer, the thermal sensation score was high, except from the
[3] D.B. Crawley, Contrasting the capabilities of building energy performance simula-
tion programs, Build. Environ. 43 (2008) 661–673.
offices with AC, during a certain range of outdoor temperature [4] J.A. Clarke, Energy Simulation in Building Design, Butterworth Heinemann, 2001.
[5] J.L.M. Hensen, R. Lamberts (Eds.), Building Performance Simulation for Design and
changes, while occupants from private and shared-private offices Operation, Routledge, 2011.
had better satisfaction with the same thermal sensation feeling. In [6] S.-H. Lin, D.J. Gerber, Evolutionary energy performance feedback for design:
winter, most of the occupants experienced a neutral thermal comfort multidisciplinary design optimization and performance boundaries for design de-
cision support, Energy Build. 86 (2014) 416–441.
level in offices of different sizes;
•
[7] C. Sun, Y. Han, Construction of office building environment information system in
Generally, season and habit were the major driving factors of severe cold area, N. Archit. 10 (2015) 4–9.
window-opening behaviour. During the summer or winter runs, the [8] Y.E. Kalay, Performance-based design, Autom. Construct. 4 (1999) 395–409.
[9] Y. Han, H. Yu, C. Sun, Simulation-based multiobjective optimization of timber-glass
behaviour tended to exhibit a stable change that was not influenced
residential buildings in severe cold regions, Sustainability 9 (2017) 2353–2371.
by physical environmental changes. Half of the surveyed office [10] C. Sun, Y. Han, Study on the theory of building digital energy saving design under
buildings showed the mode of window-opening day and night with the guidance of green performance, Architect. J. 11 (2016) 89–93.
no closing during summer, and all-closed in winter for the same [11] E. Mantesi, C.J. Hopfe, M.J. Cook, J. Glass, P. Strachan, The modelling gap:
quantifying the discrepancy in the representation of thermal mass in building si-
occupants. mulation, Build. Environ. 131 (2018) 74–98.
• Three summer patterns and four winter patterns of window-opening [12] P. Wilde, The gap between predicted and measured energy performance of build-
ings: a framework for investigation, Autom. Construct. 41 (2014) 40–69.
duration were obtained via cluster analysis. Five types of modes
[13] C.A. Menezes, A. Cripps, D. Bouchlaghem, R. Buswell, Predicted vs. actual energy
were classified and used to generate window-opening behaviour performance of non-domestic buildings:using post-occupancy evaluation data to
profiles. The types of patterns were not correlated with office size reduce the performance gap, Appl. Energy 97 (2012) 355–364.
but were more related to habit. [14] Da Yan, et al., IEA EBC Annex 66: definition and simulation of occupant behavior in

• For the surveyed office, in the summer simulation, the difference

buildings, Energy Build. 156 (2017) 258–270.
[15] M. Bonte, Impact of occupant's actions on energy building performance and thermal
between the simulated results via working-time open mode (WO) sensation, Energy Build. 76 (2014) 219–227.
and the mode inside DesignBuilder was quite small, and thus, the [16] K.-U. Ahn, C.-S. Park, Correlation between occupants and energy consumption,
Energy Build. 116 (2016) 420–430.
mode profile inside the software could have replaced the WO mode. [17] D&R International Ltd, Buildings Energy Data Book, (2011).
The AO pattern of window-opening behaviour suggested that the [18] The Energy Data and Modeling Center, Handbook of Energy & Economic Statistics
simulation performance gap could be fixed from 2.5% in summer. In in Japan, (2011).
[19] Z. Li, Y. Jiang, W.Q. Fan, Environmental parameters and air conditioning behavior
the winter simulation, the difference between the real-mode calcu-
of residential air conditioning energy consumption impact analysis, Air Cond. 37
lated result and the template of DesignBuilder was more significant, (2007) 67–71.
from 10% to 13%. Due to the very short window-opening length of [20] Japan Institute of Architecture. Japan Residential Energy Consumption Online
the winter WO mode, the correction level for the indoor thermal Database. http://tkkankyo.eng.niigata-.ac.jp/HP/HP/database/japan2/index.htm.
[21] P. Hoes, J. Hensen, M.G.L.C. Loomans, B. de Vries, D. Bourgeois, User behavior in
performance simulation of the two modes is very close. whole building simulation, Energy Build. 41 (2009) 295–302.
[22] M. Jia, R.S. Srinivasan, A.A. Raheemb, From occupancy to occupant behavior: an
Clearly, this study has imperfections with the limitation of the analytical survey of data acquisition technologies, modeling methodologies and
simulation coupling mechanisms for building energy efficiency, Renew. Sustain.
number of buildings surveyed and the number of recorded windows in

320
C. Sun et al. Building and Environment 149 (2019) 305–321

Energy Rev. 68 (2017) 525–540. framework, Energy Build. 88 (2015) 395–408.

[23] R. Li, G. Dane, C. Finck, W. Zeiler, Are building users prepared for energy flexible [33] V.L. Erickson, M.Á. Carreira-Perpiñán, A.E. Cerpa, Observe: occupancy-based
buildings? - a large-scale survey in The Netherlands, Appl. Energy 203 (2017) system for efficient reduction of HVAC energy, 10th International Conference on
623–634. Information Processing in Sensor Networks (IPSN), IEEE, 2011, pp. 258–269.
[24] S.P. Corgnati, E. Fabrizio, D. Raimondo, M. Filippi, Categories of indoor environ- [34] H Yu Y Jiao, T. Wang, Y. An, Y. Yu, Thermal comfort and adaptation of the elderly
mental quality and building energy demand for heating and cooling, Build Simul. 4 in free-running environments in Shanghai, China, Build. Environ. 118 (2017)
(2011) 97–105. 259–272.
[25] HB Gunay, W. O'Brien, I. Beausoleil-Morrison, A critical review of observation [35] S. Pan, Y. Xiong, Y. Han, Z. Zhang, L. Xia, S. Wei, J. Wu, M. Han, A study on
studies, modeling, and simulation of adaptive occupant behaviors in offices, Build. influential factors of occupant window-opening behaviour in an office building in
Environ. 70 (2013) 31–47. China, Build. Environ. 133 (2018) 41–50.
[26] C.J. Andrews, H. Chandra Putra, C. Brennan, Simulation modeling of occupant [36] X. Zhou, T. Liu, X. Shi, X. Jin, Case study of window operating behaviour patterns in
behavior in commercial buildings, Prepared by the Center for Green Building at an open-plan office in the summer, Energy Build. 165 (2018) 15–24.
Rutgers University for the Energy Efficient Buildings Hub, Philadelphia, PA, 2013. [37] R. Zhang, C. Sun, S. Sharples, H.R. Zhang, Thermal comfort and behaviour control
[27] A. Mahdavi, C. Pröglhöf, User Behavior and Energy Performance in Buildings, in mix-mode office buildings in Harbin, China, Proc. Sustainable Design of the Built
Internationalen Energiewirtschaftstagung an der TU Wien (IEWT), Wien, Austria, Environment 2017 Conference, vol. 1, 2017, pp. 318–330 (London).
2009, pp. 1–13. [38] C. Sun, R. Zhang, S. Sharples, Y.S. Han, H.R. Zhang, A longitudinal study of sum-
[28] J. Zhao, B. Lasternas, K.P. Lam, R. Yun, V. Loftness, Occupant behavior and sche- mertime occupant behaviour and thermal comfort in office buildings in northern
dule modeling for building energy simulation through office appliance power China, Build. Environ. 143 (2018) 404–420.
consumption data mining, Energy Build. 82 (2014) 341–355. [39] ANSI/ASHRAE Standard 55-2013, Thermal Environmental Conditions for Human
[29] B. Dong, B. Andrews, Sensor-based occupancy behavioral pattern recognition for Occupancy, (2013), p. 14 www.ashrae.org.
energy and comfort management in intelligent building, Proc. Build. Simul. (2009) [40] S. Shukla, S. Naganna, A review on K-means data clustering approach, Int. J. Inf.
1444–1451. Comput. Technol. 4 (2014) 1847–1860.
[30] G. Zimmerman, Modeling and simulation of individual user behavior for building [41] S. Kotsiantis, D. Kanellopoulos, Association rules mining: a recent overview, GESTS
performance predictions, Proc. 2007 Summer Computer Simulation Conference, International Transactions on Computer Science and Engineering, vol. 32, 2006, pp.
Society for Computer Simulation International, 2007, pp. 913–920. 71–82.
[31] F. Haldi, D. Robinson, On the behaviour and adaptation of office occupant, Build. [42] Evaluation standard for indoor thermal environment in civil buildings, Ministry of
Environ. 43 (2008) 2163–2177. Housing and Urban-rural Development of the People's Republic of China & State
[32] S. D'Oca, T. Hong, Occupancy schedules learning process through a data mining Administration for Market Regulation, (2012).

321