Building and Environment 103 (2016) 54e69

Contents lists available at ScienceDirect

Building and Environment

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

Analysis of occupants' behavior related to the use of windows in

German households
Davide Calì a, *, Rune Korsholm Andersen b, Dirk Müller a, Bjarne W. Olesen b
a
RWTH Aachen University, E.ON Energy Research Center, Institute for Energy Efficient Buildings and Indoor Climate, Mathieustr. 10, 52074, Aachen,
Germany
b
International Centre for Indoor Environment and Energy, Department of Civil Engineering, Technical University of Denmark, Nils Koppels All

e 402, 2800,
Kgs. Lyngby, Denmark

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

Article history: Real energy performances of buildings depend not only on deterministic aspects, such as building
Received 21 December 2015 physics and HVAC systems, but also on stochastic aspects such as weather and occupants' behavior.
Received in revised form Typically, occupant behavior is not adequately considered when calculating the expected performance.
22 March 2016
As a result, field test studies all over Europe have shown discrepancies between real and expected energy
Accepted 23 March 2016
Available online 1 April 2016
performance of buildings. In order to bridge this gap, stochastic occupants' behavior models could be
embedded into building energy performance simulation software. In order to make such models, there is
a need for a better understanding of occupants' behavior and in particular the reasons of their adjust-
Keywords:
Logistic regression
ments of building controls such as window opening, heating set points, etc. The purpose of this paper
Natural ventilation was to analyze window opening behavior in residential buildings, investigate which drivers lead occu-
Buildings' energy performance pants to interact with windows and how these actions can be modeled. A method to analyze the
Building energy performance simulation probability of a state change of the windows, based on logistic regression, was applied to monitored data
Case study (measured each minute) from two refurbished demonstration buildings. The weather and the five rooms
Field test of the 60 apartments located in the buildings were monitored in terms of air quality and thermal
environment (presence of occupants was not monitored) during four years.
The most common driver to open a window was the time of the day, followed by the carbon dioxide
concentration. The most common driver to close a window was the daily average outdoor temperature,
followed by the time of the day.
© 2016 Elsevier Ltd. All rights reserved.

1. Introduction The real energy performances of buildings depend in general on

deterministic aspects, like building physics and the HVAC systems,
During the last decades, a big effort has been made to build new as well as stochastic (probabilistic) aspects, such as weather and
energy efficient buildings and make energy retrofits to the existing occupants' behavior. While a lot of effort has been spent on
building stock. However, new or retrofitted buildings do not always deterministic modeling of the buildings as a whole, as well as on
perform as predicted: field test studies all over Europe [1e14] show the generation of standard weather files (known as test reference
higher final energy consumption than predicted. The reasons for years) for many regions in the world, only in the last decade the
this discrepancy can be grouped into engineering system issues importance of the dynamic interaction between occupants and
(e.g. components not performing as expected or wrongly installed, buildings has been recognized and researched. A reason for this can
wrong installation of the insulation, bad insulation of thermal be connected to the fact that the relative impact of the occupants'
bridges, tightness, etc.), weather data as well as occupants' behavior on the building energy performance seems to increase
behavior issues (differences in modeled and real behavior when building standards require lower energy use [15]. Including
regarding the interaction with windows and engineering systems). the stochastic aspects of occupants' behavior into building energy
performance simulation software, could hence lead to better pre-
dictions of the buildings' final energy consumption, especially for
new low energy buildings or energy retrofitted buildings.
* Corresponding author.
In order to improve the energy performance forecasts, there is a
E-mail address: [email protected] (D. Calì).

http://dx.doi.org/10.1016/j.buildenv.2016.03.024
0360-1323/© 2016 Elsevier Ltd. All rights reserved.
 D. Calì et al. / Building and Environment 103 (2016) 54e69 55

need to focus on occupants' behavior, defined as the set of actions, reasons leading to a reaction in the building occupant and sug-
executed by occupants, modifying the physical conditions of the built gesting him or her to act (they namely “drive” the occupant to an
environment. Occupants also influence the indoor environment action)”. From the analyzed literature review, Fabi et al. conclude
passively through their presence (e.g. through the production of that:
heat, water vapor, carbon dioxide, and volatile organic com-
pounds); thus, a holistic model of the occupants' behavior should 1. Window opening behavior has a strong impact on both the in-
include: door environmental quality and the energy performances of
buildings,
1. Periods of presence and absence of occupants 2. There is not a commonly agreed upon approach among scien-
2. Occupants' interaction with, among others, thermostats/valves, tists about the driving forces of occupants' window opening and
windows and sun-blinds (as schematized in Fig. 1) closing behavior,
3. Most of the existing studies analyze the position of windows
A consistent stochastic model of occupants' behavior in resi- instead of the transition from state “Open” to state “Closed” and
dential buildings based on a large database of observed data has not vice versa, and this might be problematic since the status of the
been developed yet, mostly due to lack of data. In fact, in studies windows influences the indoor environment,
where the presence of occupants was monitored (e.g. through the 4. More efforts should be made to better understand and model
use of time-use survey) and modeled [17,18], no information about occupants' behavior.
occupants' behavior was recorded. Vice versa, in studies focusing
on the evaluation and modeling of occupants' behavior, based on One of the first attempts to mathematically model the occu-
monitoring data, reliable information about real occupancy profiles pants' behavior related to natural ventilation was made by Fritsch
is missing. In this context, in 2014, the International Energy Agency et al. [26]. They use a Markov Chain process and their model is
started the “IEA-EBC Annex 66 e Definition and Simulation of based on the measured window opening angles recorded each half
Occupant Behavior in Buildings” [19]: this aims to bridge the gap hour for four windows in four office rooms. Each window is
between expected and observed energy performance of buildings, modeled with transition probability matrices with six states
by modeling and integrating the occupants' behavior in buildings' respectively (the opening angles are grouped into five opening
simulation software. classes plus a class for the closing status) and adjusted for four
The architecture of a holistic occupants' behavior model could different outdoor temperature ranges. The model considers the
consist of a core for the generation of presence patterns (e.g. based windows to be closed at night and on weekends, and the generation
on Markov chain as in Refs. [17,20,21]), and, on top of this, of a of the window state profiles starts at the beginning of each working
variable number of add ons to reproduce human activities, (e.g. day (and ends therefore at the end of the day). Fritsch et al. [26]
related to settings of the engineering system, change of state of argue that the state of windows during night has, except for two
windows, use of appliances, etc.). Many models have already been cases, always been closed over the entire observed heating period.
developed for reproducing occupants' activities: A plenary litera- However, this consideration is not valid for residential buildings, as
ture review on occupants' behavior models, up to 2013, is offered in shown in many field test studies [25], [27].
Refs. [22]; a more recent review of state of the art models of oc- With their work “Interactions with window openings by office
cupants' behavior is offered in Yan et al. [23]. occupants” [28], Haldi et al. set a milestone in the modeling of
In the European Union (EU28), space heating [24], which is occupants' behavior related to window operation in office build-
influenced by the ventilation rate, accounts for more than two ings. For seven years, the authors monitored temperature, window
thirds of residential buildings' energy consumption. As such, the position (open/closed, no distinction between opening angles), and
energy consumption in residential buildings is under heavy influ- presence of occupants (with passive infra-red sensors) of 14 offices
ence of the occupants' window opening behavior and any attempt in a building, and the outdoor environmental conditions through a
to realistically model indoor environment and energy consumption weather station located 8 km away from the building. They
in residential buildings should include realistic models of the modeled the behavior following several approaches based on lo-
window operation. In order to construct such models, basic gistic probability distributions, Markov chains, continuous-time
knowledge of the reasons for opening and closing windows, is random processes, and hybrid combinations of those models.
needed. They concluded that “[s]upported by rigorous cross-validation,
In 2012 Fabi et al. [25] analyzed more than 70 papers to inves- [they] have demonstrated the superiority of a discrete-time Mar-
tigate why occupants opened and closed windows. In addition to kov process approach and its strong added value compared with
the literature review, the authors define the concept of drivers: existing models”. Further, they wrote: “[f]actors related to indoor
“Factors influencing occupant behavior, both external and indi- air quality (e.g. CO2 or pollutant concentration) should also be
vidual, that could be named with the general term “Drivers”, are the treated”.
A further contribution to the occupants' modeling for natural
ventilation is done by Andersen et al. within several works
[27,29,30]. In the last paper, the authors introduce a model for the
Actions
transition of the state of windows (open to close and vice-versa
close to open) based on observations collected from 15 dwellings
Occupant HVAC
Heat located in Denmark. Contrarily to previous window operation (or
state) models introduced by 2013 and only based on thermal con-
CO2 Lighting
ditions, the authors also monitored indoor carbon dioxide con-
VOC Window centration as an indicator for indoor air quality. The models e one
for the opening and one for the closing of the windows e are based
H2O Appliances on the results of logistic regression analysis. The conclusion of the
manuscript is that indoor CO2 concentration and outdoor temper-
Fig. 1. General schema of occupants' impact on buildings (previously introduced in ature are the two single most important variables (drivers) deter-
Ref. [16]). mining the window opening and closing probability respectively.
56 D. Calì et al. / Building and Environment 103 (2016) 54e69

D'Oca and Hong [31] recently evaluated occupants' behavior certain event to happen (e.g. a window changes its state), and, by
related to natural ventilation in offices using logistic regression definition, PðxÞ ε ½0; 1 c x. Eq. (1) can be rewritten as in Eq. (2).
analysis and two different data-mining approaches, cluster analysis
and association rules mining. They state that indoor air tempera- 1
p¼ (1)
ture, outdoor air temperature, time of the day (office arriving time 1 þ eðaþbxÞ
and early morning) and occupancy presence are the top drivers for
the window opening operation; indoor air temperature, time of the  
p
day (office leaving time and evening), occupancy presence and ln ¼ a þ bx (2)
1p
outdoor air temperature are the top drivers for the window closing Where:
operation.
From the literature review it is evident, that most analyses 1. P(x) (or simply p) is the probability function,
focused on office buildings, and that there is a lack of field test 2. a is the intercept,
studies related to residential buildings. Furthermore, several ap- 3. b is a coefficient,
proaches were used to analyze and model occupants' behavior: 4. x is the explanatory variable.
logistic regression analysis and mixed effect models seem to be
proper instruments for evaluating the drivers, while Markov chains Eq. (2) describes the probability of an event depending on one
are particularly indicated to model the time dependency of pro- explanatory variable, and is therefore used for simple linear
cesses (e.g. opening and closing cycles). regression analysis. For regression analysis with n explanatory
In many of the mentioned papers [21,24e28] there is agreement variables, the probability function p can be expressed as in Eq. (3).
that factors related to indoor air quality could be a driver for oc-
 
cupants' window operation. There is also agreement that factors p
related to thermal comfort should be included in models of window ln ¼ a þ b0 x0 þ b1 x1 þ … þ bn xn (3)
1p
opening behavior. However, these conclusions are based on data
from a limited number of windows in a limited number of apart- Andersen et al. [27] suggest to include interaction terms into the
ments. The objective of this work was to investigate which drivers probability function. They argue that “the probability might depend
lead occupants to open or close windows, based on one full year of differently on x1 at one level of x2 as compared to another level of
measurements from 300 windows in 60 apartments. x2”. For example, when inferring the probability of opening (or
The drivers analyzed were related to thermal comfort (room air closing) a window, it may happen that the coefficient bi of the xi
temperature, room relative humidity), indoor air quality (carbon explanatory variable at a certain time slot (e.g. morning) differs
dioxide concentration), weather (wind speed, outdoor tempera- from the coefficient bi at a different time slot (e.g. night): in an
ture, outdoor relative humidity) and time of the day. Compared to extreme case, an increase of the temperature might result in an
other studies on modeling the state change of windows, the data increase of the probability of opening a window in the morning,
used within this work had three main advantages: the availability and in a decrease of the probability of opening a window in the
of both data related to thermal comfort and to indoor air quality, evening. Based on their suggestion, in order to include the inter-
the big number of geometrically identical apartments (however action terms gi,j, Eq. (3) can therefore be written as Eq. (4). Only
with different engineering systems) and a ten times higher time interaction terms between continuous and categorical variables
resolution. This high time resolution allowed for a precise study of were used.
the drivers for a window state change, without missing short  
p
opening/closing phases (more on this topic in Section 3.2). ln ¼ a þ b0 x0 þ b1 x1 þ … þ bn xn þ g1;2 x1 x2 þ …
1p
2. Description of the applied methodology þ g1;n x1 xn þ … þ gn1;n xn1 xn
(4)
Logistic regression is a well-established statistical method used
to analyze and model binary dependent variables (such as the state The “forward and backwards” selection of the variables for the
of a window, closed or open, or the change of state of a window). It regression models was executed based on the Akaike information
is used in many scientific fields, from medicine and economics to criterion (AIC), as suggested by Schweiker and Shukuya [33]: This
engineering. As illustrated in the introduction, the method has allowed for a selection of a “best model” containing only the most
been successfully adopted to describe the probability of opening important explanatory variables (variables that have a consistent
and closing events of windows. The main objective of the statistical impact on the probability function). In practice, the process for the
analyses was to understand the drivers leading occupants' to act selection of the best model was implemented using the ‘step’
(change the state of windows), thus providing researchers a back- function with the ‘glm’ function in the statistical language “R”, with
ground about how occupants' make decisions, correlated to ther- n explanatory variables. It is described by the following steps:
mal comfort, indoor air quality, weather and time of the day. In
order to tackle this research question, the logistic regression anal- 1. Each variable was fitted by the regression model (in a single-
ysis with multiple explanatory variables was applied singularly to variable model), and the AIC calculated for each fit;
each monitored window. In this way, for each window, it was 2. The variable with the lowest AIC was selected, and the model
possible to determine which variables had a major influence on the was fitted n1 times with the selected variable and each of the
probability of a change of window state, and which did not. Thus, n1 remaining variables;
the variables were classified depending on the number of times 3. The model based on two variables with the lowest AIC was
they appeared in the models (the logistic regression equations). The selected and the AIC of this model is compared to the best
variables used most frequently in the models were regarded as single-variable model (the single-variable model with the
more important in general for opening/closing actions. lowest AIC); Then:
Logistic regression [32] was used as analysis and modeling a. If the new model (two-variables model) had a consistently
method. Logistic regression is based on the logistic function as lower AIC, the process went further to step 4,
expressed in Eq. (1). P(x) expresses the probability function for a b. Otherwise the single-variable model was chosen;
 D. Calì et al. / Building and Environment 103 (2016) 54e69 57

4. The yet excluded n2 variables were used to fit the model language “R” using the packages “Companion to Applied Regres-
together with the two variables of the “two variables model” sion” (car) and the package “Mixed GAM Computation Vehicle with
with the lowest AIC, in a “three variables model” (this is the so- GCV/AIC/REML Smoothness Estimation” (mgcv), within the pro-
called “forward selection”). Further, from each of the three- gramming environment R Studio.
variables models, three two-variables models, obtained by
dropping each of the variables recursively, were fitted (this is 3. Description of the demonstration buildings and of the
the so-called “backward selection”). Then: monitoring system
a. If none of the three-variable or “new generated” two-variable
models had a consistently lower AIC than the two-variables Measurements were conducted in three refurbished buildings
model with the lowest AIC from step 3, the model with the (Fig. 3) located in southern Germany. The buildings were built at
lowest AIC from step 3 was chosen, the end of the nineteen-fifties, with relatively poor materials and
b. Otherwise, the process went further with the same criteria, unpretentious engineering systems. The three buildings, located
up to n-variables models. next to each other, are geometrically identical. Each building has 3
entrances, and each entrance provides access to 10 apartments over
Further, a k-fold cross validation (with K ¼ 10) was executed on 5 floors. The buildings have a total of 90 apartments. The apart-
the top of the described selection process: Therefor, each data ments have the same floor plan. They consist of a kitchen, a bath-
sample was partitioned into ten sub-samples. Nine sub-samples room, three rooms and a corridor. In this paper, the following
were used for the training of the models, while one sub-sample nomenclature was used: “B” for building, e.g. “B2” refers to the
was used to test the model. This was done by using the measured group of the 30 apartments in building 2, E for entrance, e.g. B2E1
input variables of the 10th subsample (the one which was not used indicated the 10 apartments located in building 2 and accessible
for the training) as input to the model and comparing the model through entrance 1. The buildings are 52 m wide and 10 m long.
output with the monitored window position. The operation was Before retrofit, the gables had 4 cm insulation material and the
executed 10 times: each subsample was used once as a test sub- windows were double glazed; the roof, the floor to the cellar and
sample, while the remaining nine sub-samples were used as the façade were not insulated. The apartments were heated
training samples. The analysis and modeling activity was executed through gas stoves installed in each living room. Domestic hot
for the window state change “opening” and the window state water was produced through gas flow heaters installed in each
change “closing” separately. By windows with more panels, p ¼ 0 apartment.
corresponds to both panels closed, while p ¼ 1 corresponds to at
least one panel open. In order to infer the probability of opening 3.1. The retrofit of the buildings
and closing (the “state change probability”), the data sample was
partitioned into two sub-samples: sub-sample “window closed”, to Between 2008 and 2010 the buildings were retrofitted. The
infer the probability that a window will be opened and sub-sample retrofit layouts were designed together with the municipal society
“window open”, to infer the probability that a window will be owning the buildings. Various engineering system components and
closed. The complete modeling process is graphically described in building insulation materials as well as new windows (Fig. 4) were
the flux diagram of Fig. 2. The analyses were executed in the selected and combined, to generate seven different retrofit layouts:
in this way, the retrofit layouts can be compared, and useful
knowledge about optimal retrofitting can be gained. Building 1 and
building 2 are connected to a district heating network, while
Partitioning the sample into 2 sub-samples building 3 is heated through different types of heat pumps (HP).
„window open“ and Depending on the entrance, radiators (Rad), ceiling heating (CH),
„window closed“ floor heating (FH) and ventilation heating (VH) were installed to
deliver the heating energy to the indoor heated spaces. Standard
water heaters and low temperature peripheral domestic hot water
Partitioning the sub-sample into 10 sub- (DHW) heaters were installed. In the peripheral solution the DHW
samples for k-fold cross validation is generated in each apartment through so called fresh water heat
exchanger stations (FWHX). The seven retrofit layouts are sche-
matically described in Table 1. More information about the build-
Start K-Fold Cross validation, K=1 ings and the retrofit layout can be found in Refs. [34e38].
It should be noted that there are five windows per apartment;
two of them (in kitchen and bathroom) have one opening panel.
The two windows in the bedrooms and in the children rooms have
Regression analysis on the sub-sample, three opening panels: for those windows, the panels on the left and
choice of best-model through AIC, K=K+1 on the right can be opened independently and were monitored; the
store chosen model as „K-Best-Model“ central panel can only be open (but not tilted) when the right panel
of the window is also completely open. The window (balcony-door)
in the living room has three panels: the one on the left can be open
No
K=10 ? be open irrespective of the state of the other two panels and was
monitored, the panel in the middle can only be open (but not tilted)
Yes only if the left panel is open, and was not monitored. The panel on
the right is fixed.
Comparison of the 10 K-Best-Models and
choice of the K-Best-Model with best fitting 3.2. The monitoring system and the observed data

Fig. 2. Flux diagram of the process aimed to generate and obtain the best fitting To evaluate the energy performances of the refurbished build-
model. ings and the occupants' behavior, a comprehensive high time
58 D. Calì et al. / Building and Environment 103 (2016) 54e69

Fig. 3. One building façade before retrofit (a). One back façade after the retrofit (c). Floor space of the apartments, for one entrance with south-gable (b). Relative position and
orientation of the buildings (d).

Fig. 4. Pictures of the windows and of the façade after retrofit.

resolution monitoring system was designed and installed (by the the distribution system of both domestic hot water (DHW) and
University of Applied Science Karlsruhe). The monitoring system heating (HE). In building 2 and building 3, the following variables
collected energy flows in the heat generation, the heat storage and were monitored every minute in each room of the 60 apartments:
 D. Calì et al. / Building and Environment 103 (2016) 54e69 59

Table 1
Description of each retrofit layout in each of the 3 entrances in the three buildings: The HVAC system is supplied by either District heating (DH) or heat pump (HP). The
ventilation is exhaust air ventilation (EAV) or ventilation with heat recovery (HR). The apartments are heated through radiators (Rad), floor heating (FH), warm air heating (VH)
or ceiling heating (CH). Domestic hot water is produced through central heat exchanger (HX) or fresh water heat exchanger stations (FWHX).

Insulation Windows U-Val HVAC DHW

2 2 2
B1, 2160 m , 30 apartments 14 cm 0.035 W/(m K) 1.3 W/(m K) DH, EAV, Rad Central HX
B2E1, 720 m2 10 Apartments 16 cm 0.021 W/(m2K) 1.3 W/(m2K) DH, EAV, window frame HR, Rad Apartment FWHX
B2E2, 720 m2 10 Apartments 16 cm 0.021 W/(m2K) 0.8 W/(m2K) DH, EAV, Rad Central HX
B2E3 720 m2 10 Apartments 16 cm 0.021 W/(m2K) 1.3 W/(m2K) DH, EAV, FH Central HX
B3E1, 720 m2 10 Apartments Vacuum: 0.8 W/(m2K) CO2-Probe HP, HR central ventilation, FH Apartment FWHX
4 cm 0.008 W/(m2K)
4 cm 0.021 W/(m2K)
B3E2, 720 m2 10 Apartments Vacuum: 0.8 W/(m2K) CO2-Probe master HP þ slave HP, apart. HR, VH Apartment FWHX
4 cm 0.008 W/(m2K)
4 cm 0.021 W/(m2K)
B3E3, 720 m2 10 Apartments Vacuum: 1.3 W/(m2K) Air HP þ exhaust air HP, CH Apartment FWHX
4 cm 0.008 W/(m2K)
4 cm 0.021 W/(m2K)

1. Air temperature [ C] was 11  C, against the typical value for this region (based on his-
2. Relative humidity [%] torical data and used e.g. for the generation of the energy labels of
3. CO2 concentration [ppm] buildings) equal to 8.9  C. The values related to the indoor condi-
4. Volatile organic compounds (VOC) [-], tions are plotted in Fig. 6(d)e(f): each box shows the data collected
5. Light on the ceiling [Lux], in the 50 rooms (10 apartments) of each building entrance (hence,
6. Infrared/visible light ratio (to recognize the light source) [-], each box represents over 25 million data points, which explains the
7. Window opening sensors (open/closed). high density of the flyers in the figures). In general, building 3 had
higher temperatures and lower relative humidity than building 2.
All variables (except for window position) were monitored The CO2 concentration was lower in B2E3 and B3, than in B2E1 and
through room monitoring units (RMUs) as shown in Fig. 5. Further B2E2; however, it was not possible to find any dependency be-
information about the monitoring system can be found in Ref. [39]. tween the air quality level and the different retrofit solutions.
In building 1, only 7 apartments were equipped with RMUs. The As already mentioned in the introduction, it is important to
evaluation presented in this manuscript was therefore related to B2 monitor the opening cycles of the windows event-based, or within
and B3. The windows' positions were monitored through wired a short time interval. Fig. 7 shows the duration of time with open
reed switches installed in the windows' frames and permanent windows. In colder months, the duration was much shorter than in
magnets installed in the frame of the windows' moveable panels. warmer periods.
The monitoring started in 2010 (The buildings were completely
occupied since spring 2011), the measurements were collected
each 60 s, and the data was stored in HDF5 files. Due to crashes of 4. Evaluation and modeling of the occupants' behavior
the monitoring system, not all data was always collected correctly.
For the year 2012 (chosen for the here proposed analysis), at least In this section, the results of the regression analysis applied at
90% of the data was correctly collected; the crashes of the moni- room (window) level are illustrated. The method explained in
toring system were spread out over the entire monitoring period. Section 2 was applied to the 300 windows located in buildings B2
Table 2 shows the specification of the sensors adopted to collect and B3. The only categorical variable used was “Time range”, which
the variables used within the LRA (the choice of those variables is distinguishes between low, medium and high probability of a state
explained in the next section). change of a window, and is grouped as follows:
A description of the measured data used for the logistic
regression analysis is offered in Fig. 6. The subfigures (a)e(c) pro- 1. Night, low probability of action: 7 h, between 11:00 p. m. and
vide information related to the observed weather conditions in 5:59 a. m.
2012. February was particularly cold, with daily average tempera- 2. Morning, high probability of action: 3 h, between 7:00 a. m. and
ture down to 12  C; On average, the outdoor temperature in 2012 9:59 a. m.

Fig. 5. (a) RMU ready for plugging. The border on the bottom of the RMU is fixed to the wall. The circular opening permits the light penetration. (b) Upper part of an open RMU: VOC
sensor (bottom), CO2 sensor (in the center), temperature and relative humidity sensor (top right outside the board), and illuminance sensor (behind the black tape, bottom left).
60 D. Calì et al. / Building and Environment 103 (2016) 54e69

Table 2
Specifications of the installed sensors.

Metric Sensor Range Accuracy

Room CO2 SenseAir CO2EngineTM(K30) 0e500 ppm ±20 ppm

Temperature Sensirion SHT75 40  C ± 124  C ±0.3  C
Relative humidity Sensirion SHT75 0%e100% ±4%
Weather Temperature MWS 5MV/MWS8 40  C ± 60  C ±0.3  C
Relative humidity MWS 5MV/MWS8 10%e100% ±2%
Wind speed MWS 5MV/MWS8 0e42 m/s ± m/s

Fig. 6. Monitored data (2012). Outdoor measurements (minutes and daily average): outdoor air temperature (a), outdoor relative humidity (b) and wind speed (c). The numbers on
the x-axis indicate the months of 2012. Interruptions in the graphs are caused by missing values. Indoor measurements: box plots of the values of air temperature (d), relative
humidity (e) and CO2 concentration, measured in the rooms of each entrance. The boxes contain 50% of the observations; the whiskers are up to 1.5 the size of the boxes, the rest of
the data is included in the flyers.

3. Rest of the day, medium of action: 14 h, between 6:00 a. m. and 4. Daily average outdoor temperature (DAT) [ C];
6:59 a. m. and between 10:59 a. m. and 22:59 p. m. 5. Wind speed (WS) [m/s];
6. Outdoor relative humidity (ORH) [%];
The following continuous variables, measured each minute,
were used: The categorical variable “Time range” was used in order to take
into account the differences in behavior, depending on the time of
1. Room air temperature [ C]; the day. The categories were chosen based on the observations of
2. Room carbon dioxide concentration (CO2) [ppm], transformed the window opening activity in Building 2 entrance 1. Those are
through the reciprocal function to obtain a more suitable dis- illustrated in Fig. 8.
tribution for the use in the regression analysis [ppm1]; In the selection of the variables to be used for the LRA, corre-
3. Room relative humidity (RRH) [%]; lating variables were avoided. The carbon dioxide concentration
 D. Calì et al. / Building and Environment 103 (2016) 54e69 61

Fig. 7. Number of times windows were opened, and duration of the state open, by month, for B3E2.

variables, selected with the procedure explained in Section 2, and

the number of models using each of them. In this way, the occu-
pants' behavior related to the operations of 300 windows (corre-
sponding to 300 analyses resulting in 300 models for the opening
operation and 300 models for the closing operation) can be eval-
uated. The results shown are related to those windows, for which at
least one state change per year was observed (some of the sensors
were defect and did not register the changes of state): thus, the
results were based on 273 windows.

4.1. Example of an opening and a closing model of one window

One model for the opening and one model for the closing action
of one window were selected, with the aim to illustrate the way the
models work; the models presented within this section were
Fig. 8. Observed no. of times windows were opened in B2E1, depending on the time of selected randomly from the 546 models (273 for opening, 273 for
the day, and thresholds for the choice of the action probability levels. closing), and therefore those models are not representative. The
presented models were obtained from the data monitored in a
living room of one apartment located in B2E1. The models were
was chosen instead of the volatile organic compounds (VOC) as the defined by Eqs. (5) and (6), respectively for the opening and the
indicator for air quality. This choice is justified by the fact that it is closing action. The terms of the equations were defined as
not possible to distinguish between good and bad odors through following:
the VOC (some VOCs have pleasant odors e.g. used in cleaning
products and perfumes while others have unpleasant odors), while 1. a is the intercept;
the CO2 concentration is a good indicator for human bioeffluents 2. b are the coefficients, free of interactions;
[40]: hence, a high concentration of CO2 is an indicator for bad air  
quality. In addition, the continuous variable “global radiation” was po
ln ¼ ao;Time þ bo;f ðCO2 Þ xf ðCO2 Þ þ bo;RT xRT (5)
not included in the evaluation since it correlated to the variable 1  po
“room temperature”. The variable “outdoor temperature” was
 
taken into account, but only as daily average outdoor temperature pc
ln ¼ ac;Time þ bc;DAT xDAT þ bc;f ðCO2 Þ xf ðCO2 Þ
to prevent correlation issues with the time and with the outdoor 1  pc
relative humidity.
þ bc;ARH xARH þ bc;RRH xRRH þ bc;RT xRT (6)
As previously discussed in the introduction, the presence of
occupants is a prerequisite for occupants' actions. A method to The coefficients to solve Eq. (5) and Eq. (6), and their confidence
reveal occupants' presence using the carbon dioxide concentration interval, are listed in Table 3. When the coefficients b were positive,
in a room, and the mechanical ventilation patterns was developed an increase of the related explanatory variable caused an increase
and introduced in Ref. [16]. However, since the door position of the of the probability of the opening/closing action, while, when
rooms as well as the occupants' settings regarding the ventilation negative, an increase of the explanatory variable caused a decrease
system were unknown, the uncertainty on the results of the pro- of the probability of the opening/closing action. Since the CO2
posed algorithm was too big, hence the algorithm was not used. As concentration was transformed through the reciprocal function, an
a consequence, the evaluation and modeling activity was executed increase of the CO2 concentration caused a decrease of the recip-
for the entire measured time, without distinction between occu- rocal of the CO2 concentration: hence a negative sign of the coef-
pied/not occupied timeslots (as those were unknown). ficient of f (CO2) meant that with increasing CO2 concentration in
In the next section, an example of an opening and a closing the room, the probability of opening/closing a window increased.
model, as result of the LRA, are introduced. Hence, in Section 4.2 Which time slot is analyzed, depends on the chosen intercept:
and in Section 4.3 the results of all the windows are presented. the explanatory categorical variable “time” only affected the
The results are organized in graphics, showing the explanatory intercept, since no interaction terms between this variable and
62 D. Calì et al. / Building and Environment 103 (2016) 54e69

Table 3
Intercept a and coefficients b of the explanatory variables for the living room's window model of an apartment in B2E1 (n.p. implies that the variable was “not present” in the
model).

B2E1, models of a living room's window

Opening action Closing action

Coef. Conf. interval Coef. Conf. interval

Expl. variable 2.50% 97,5% 2.50% 97,5%

aNight 10.089 11.262 8.916 2.539 0.559 4.519

aMorning 8.214 9.431 6.998 3.317 1.309 5.325
aRest of the day 7.795 9.002 6.587 3.955 1.956 5.954
bf((CO1
2 )) 551.15 696.32 405.98 785.70 934.88 636.52
bRT 0.134 0.082 0.187 0.268 0.348 0.188
bRRH n.p. n.p. n.p. 0.058 0.071 0.044
bDAT n.p. n.p. n.p. 0.089 0.105 0.074
bARH n.p. n.p. n.p. 0.027 0.022 0.031

other continuous variables were present in these particular models. sign of the coefficients: a negative coefficient shows a negative
For the presented opening action model, the probability of opening correlation to the probability of action). An increase in carbon di-
a window grew with increasing CO2 concentration and with oxide concentration leads to an increase of the probability of
increasing room temperature. In the closing action model, the window opening for more than 45% of the windows; However,
probability of closing a window grew with: contrary to this result, the carbon dioxide concentration negatively
influenced the opening action for approximately 6% of the models;
1. Increasing CO2 concentration, this is not necessarily a contradiction, and could be related to the
2. Decreasing room temperature, occupancy (which is not included in the models) and window
3. Decreasing relative humidity in the room, opening behavior upon arrival. The probability of opening windows
4. Decreasing daily average outdoor temperature, increased with increasing indoor temperatures, and by increasing
5. Decreasing outdoor relative humidity. room relative humidity and daily average outdoor temperature
(each of these explanatory variables was present in more than 35%
The probability of the opening action in the model described in of the models). The only remarkable explanatory variable that
Eq. (5) is graphically illustrated in Fig. 9. The probability of the negatively influenced the opening of windows (more than 30% of
opening action for this particular window was very low at night, the windows) was the outdoor relative humidity.
high in the morning and very high during the rest of the day. The
probability of opening varied between 0 and 0.02. The small values 4.3. Evaluation of the window closing action
were a consequence of the very short measuring interval (60 s). In
essence, the model calculated the probability of opening and Fig. 13 shows the explanatory variables of the logistic regression
closing a window within the next minute. This window was in models of closing action. Figs. 14 and 15 show the drivers that
“open status” 32.1% of the time during 2012 (The analyses of this positively and negatively influenced window opening action
room were based on a time period of 329 days and had a 10% failure respectively. The daily average outdoor temperature was the most
rate of the monitoring system) and the window was closed (and common driver (in almost 70% of the models); in particular, as
opened, when it was in status closed) 1014 times. On average, the Fig. 15 shows, the probability of closing windows increased with
probability of opening that window in 2012 was equal to 0.0032 decreasing daily average outdoor temperature. The variable time
(1014 opening events in 321,683 min, 67.9% of 329 days), while the was present in almost half of the models.
probability of closing that window was equal to 0.0067 (1014 Further, an increase in the carbon dioxide concentration was
closing events in 152,077 min, 32.1% of 329 days). associated with increasing probability of the closing action: the
carbon dioxide concentration correlated with the presence of oc-
4.2. Evaluation of the window opening action cupants, and the presence of occupants is a necessary condition for
the window to be closed. In almost 40% of the models, a decrease of
Fig. 10 shows the explanatory variables of the models of opening the room temperature corresponded to an increase of the proba-
a window. The only categorical variable used, “time”, was included bility of closing the window.
in over 70% of the windows. All the interaction terms with the
variable “time” were used by less than 10% of the models with the 4.4. Differences in behavior related to room typology
consequence that the variable time commonly influenced the
intercept, but not the coefficients of the continuous explanatory In this section, the explanatory variables of the logistic regres-
variables. sion models for the opening and the closing action are illustrated,
The most common continuous explanatory variable was the depending on room typology. The names “living room”, “bedroom”
carbon dioxide concentration, present in over 50% of the models. and “children room” only refer to the nomenclature used in the
The room temperature, the relative humidity of the room, the daily floor space of Fig. 3; the real use of those rooms is unknown. In
average outdoor temperature, and the outdoor relative humidity contrast to the other rooms, the living room has a balcony-door. The
were used by over 35% of the models. The wind speed turned out to most common drivers for the opening and the closing action,
be mostly irrelevant for the opening action. “None” indicates that depending on room typology, are listed in Table 4 and Table 5
no explanatory variables were found for the fitting of the model. respectively.
Figs. 11 and 12 show the drivers which directly (positively) and The daily average outdoor temperature influenced the opening
inversely (negatively) influenced the probability of the action models of kitchen windows less than the window models of the
“window opening” respectively (this can be verified based on the other rooms. In addition, the relative humidity was present in two
 D. Calì et al. / Building and Environment 103 (2016) 54e69 63

Fig. 9. Probability of opening action for the window in the living room of an apartment located in B2E1, at night (c and f), in the morning (b and e) and during the rest of the day (a
and d).

Fig. 10. Drivers for opening and the number of rooms/models with the driver used as explanatory variable.

thirds of the models of bathroom windows, and in more than half of the models is compared depending on the retrofit solution of the
the models of kitchen windows. Finally, wind speed was found in buildings (as illustrated in Table 6 and Table 7). The CO2 con-
13% of the opening models of the living room windows, against an centration was a driver for 53.5% of the windows, varying be-
average value over the entire field test equal to 4%; for the closing tween 41.3% for B2E2 and 70.8% for B3E3: however, these two
models, wind speed was found in 55.6% of the models, against an entrances have the same ventilation system. Similar consider-
average value over the entire field test of 20.9% This could be due to ations can be applied to the other variables. In conclusion, there
the fact that the living room windows are much bigger than the were no clear trends showing differences in the drivers for win-
windows in the other rooms; therefore wind may have a stronger dow opening behavior between the buildings' different ventila-
impact on the comfort of occupants. tion systems.

4.5. Differences in behavior related to HVAC system 4.6. Differences in behavior related to level of the apartments

Within this section, the share of explanatory variables within In this section, the share of explanatory variables within the
64 D. Calì et al. / Building and Environment 103 (2016) 54e69

Fig. 11. Drivers for opening and number of rooms/models with a positive correlation between the variable and the probability of opening.

Fig. 12. Drivers for opening and number of rooms/models with a negative correlation between the driver and the probability of opening.

Fig. 13. Drivers for closing and number of apartments/models with the driver used as explanatory variable.

Fig. 14. Drivers for closing and number of rooms/models with a positive correlation between the variable and the probability of closing.

Fig. 15. Drivers for closing and number of rooms/models with a negative correlation between the driver and the probability of closing.

models is compared depending on the level of the apartments (as possible to find systematic patterns in the variation of the behavior.
illustrated in Table 8 and Table 9). The variables varied strongly The change in behavior may have been a result of distribution of
depending on the level of the apartments. However, it was not occupants within the different levels.
 D. Calì et al. / Building and Environment 103 (2016) 54e69 65

Table 4
Share of the drivers for the opening action, of all the models and of the models of each room typology (e.g. Time 73.7% for the kitchen indicates that 73.7% of the 60 models
obtained by analyzing the windows in the 60 kitchens, located in the demonstration buildings, used the explanatory variable, time).

All Kitchen Bathroom Livingroom Bedroom Childrenroom

Time 72.2% 73.7% 76.7% 83.3% 57.7% 68.0%
Room CO 2 concentration 53.5% 47.4% 56.7% 68.5% 40.4% 54.0%
Room air temperature 41.8% 47.4% 51.7% 50.0% 19.2% 38.0%
Room relative humidity 39.9% 52.6% 66.7% 33.3% 17.3% 24.0%
Daily average outdoor temperature 38.8% 24.6% 31.7% 51.9% 42.3% 46.0%
Outdoor relative humidity 35.9% 45.6% 36.7% 40.7% 26.9% 28.0%
Time & Room air temperature 5.5% 3.5% 8.3% 7.4% 0.0% 8.0%
Time & Room CO 2 concentration 4.8% 1.8% 1.7% 3.7% 7.7% 10.0%
Wind speed 4.0% 3.5% 1.7% 13.0% 0.0% 2.0%
Time & Daily average outdoor temperature 2.9% 0.0% 1.7% 3.7% 5.8% 4.0%
Time & Room relative humidity 2.6% 1.8% 6.7% 1.9% 0.0% 2.0%
None 1.8% 1.8% 0.0% 0.0% 3.8% 4.0%
Time & Outdoor relative humidity 0.7% 1.8% 0.0% 1.9% 0.0% 0.0%

Table 5
Share of the drivers for the closing action of all the models and of the models of each room typology.

All Kitchen Bathroom Livingroom Bedroom Childrenroom

Daily average outdoor temperature 67.8% 64.9% 75.0% 74.1% 57.7% 66.0%
Time 49.1% 40.4% 36.7% 72.2% 46.2% 52.0%
Room CO 2 concentration 42.5% 26.3% 43.3% 72.2% 36.5% 34.0%
Room air temperature 40.7% 42.1% 35.0% 44.4% 40.4% 42.0%
Outdoor relative humidity 33.0% 22.8% 25.0% 55.6% 30.8% 32.0%
Room relative humidity 29.3% 19.3% 23.3% 50.0% 21.2% 34.0%
Wind speed 20.9% 8.8% 3.3% 55.6% 15.4% 24.0%
Time & Daily average outdoor temperature 7.3% 7.0% 5.0% 13.0% 7.7% 4.0%
Time & Room CO 2 concentration 5.9% 0.0% 0.0% 20.4% 1.9% 8.0%
None 3.3% 5.3% 1.7% 1.9% 3.8% 4.0%
Time & Room air temperature 2.9% 1.8% 0.0% 9.3% 0.0% 4.0%
Time & Room relative humidity 1.1% 0.0% 0.0% 1.9% 0.0% 4.0%
Time & Outdoor relative humidity 0.7% 0.0% 0.0% 3.7% 0.0% 0.0%
Time & Wind speed 0.7% 0.0% 0.0% 1.9% 0.0% 2.0%

Table 6
Share of the drivers for the opening action, of all the models and of the models of each retrofit layout.

All B2E1 B2E2 B2E3 B3E1 B3E2 B3E3

Time 72.2% 77.8% 76.1% 84.6% 70.8% 66.0% 60.4%
Room CO 2 concentration 53.5% 57.8% 41.3% 64.1% 45.8% 42.6% 70.8%
Room air temperature 41.8% 46.7% 37.0% 41.0% 35.4% 36.2% 54.2%
Room relative humidity 39.9% 42.2% 47.8% 61.5% 43.8% 29.8% 18.8%
Daily average outdoor temperature 38.8% 37.8% 41.3% 48.7% 35.4% 42.6% 29.2%
Outdoor relative humidity 35.9% 28.9% 47.8% 41.0% 43.8% 23.4% 31.3%
Time & Room air temperature 5.5% 8.9% 8.7% 5.1% 4.2% 4.3% 2.1%
Time & Room CO 2 concentration 4.8% 4.4% 6.5% 7.7% 8.3% 2.1% 0.0%
Wind speed 4.0% 2.2% 4.3% 2.6% 2.1% 4.3% 8.3%
Time & Daily average outdoor temperature 2.9% 0.0% 0.0% 7.7% 4.2% 4.3% 2.1%
Time & Room relative humidity 2.6% 2.2% 2.2% 5.1% 4.2% 2.1% 0.0%
None 1.8% 2.2% 0.0% 0.0% 4.2% 2.1% 2.1%
Time & Outdoor relative humidity 0.7% 0.0% 2.2% 0.0% 2.1% 0.0% 0.0%
66 D. Calì et al. / Building and Environment 103 (2016) 54e69

Table 7
Share of the drivers for the closing action, of all the models and of the models of each retrofit layout.

All B2E1 B2E2 B2E3 B3E1 B3E2 B3E3

Daily average outdoor temperature 67.8% 77.8% 80.4% 74.4% 58.3% 61.7% 56.3%
Time 49.1% 53.3% 43.5% 64.1% 47.9% 57.4% 31.3%
Room CO 2 concentration 42.5% 51.1% 32.6% 51.3% 39.6% 51.1% 31.3%
Room air temperature 40.7% 35.6% 41.3% 35.9% 43.8% 44.7% 41.7%
Outdoor relative humidity 33.0% 37.8% 32.6% 48.7% 31.3% 31.9% 18.8%
Room relative humidity 29.3% 17.8% 13.0% 35.9% 41.7% 36.2% 31.3%
Wind speed 20.9% 17.8% 21.7% 23.1% 25.0% 23.4% 14.6%
Time & Daily average outdoor temperature 7.3% 2.2% 8.7% 10.3% 12.5% 2.1% 8.3%
Time & Room CO 2 concentration 5.9% 8.9% 6.5% 2.6% 6.3% 6.4% 4.2%
None 3.3% 2.2% 4.3% 0.0% 4.2% 6.4% 2.1%
Time & Room air temperature 2.9% 2.2% 0.0% 2.6% 2.1% 10.6% 0.0%
Time & Room relative humidity 1.1% 2.2% 0.0% 0.0% 2.1% 2.1% 0.0%
Time & Outdoor relative humidity 0.7% 0.0% 0.0% 0.0% 2.1% 0.0% 2.1%
Time & Wind speed 0.7% 0.0% 0.0% 0.0% 2.1% 2.1% 0.0%

Table 8
Share of the drivers for the opening action, of all the models and of the models of each level.

All 1st Floor 2nd Floor 3rd Floor 4th Floor 5th Floor
Time 72.2% 73.6% 68.5% 81.5% 74.1% 63.8%
Room CO 2 concentration 53.5% 47.2% 51.9% 51.9% 50.0% 65.5%
Room air temperature 41.8% 54.7% 20.4% 46.3% 29.6% 56.9%
Room relative humidity 39.9% 37.7% 42.6% 44.4% 37.0% 37.9%
Daily average outdoor temperature 38.8% 34.0% 37.0% 48.1% 38.9% 36.2%
Outdoor relative humidity 35.9% 28.3% 37.0% 51.9% 37.0% 25.9%
Time & Room air temperature 5.5% 9.4% 5.6% 3.7% 1.9% 6.9%
Time & Room CO 2 concentration 4.8% 0.0% 7.4% 0.0% 3.7% 12.1%
Wind speed 4.0% 9.4% 1.9% 1.9% 0.0% 6.9%
Time & Daily average outdoor temperature 2.9% 1.9% 0.0% 3.7% 5.6% 3.4%
Time & Room relative humidity 2.6% 1.9% 1.9% 1.9% 3.7% 3.4%
None 1.8% 0.0% 1.9% 0.0% 3.7% 3.4%
Time & Outdoor relative humidity 0.7% 0.0% 0.0% 1.9% 1.9% 0.0%

Table 9
Share of the drivers for the closing action, of all the models and of the models of each level.

All 1st Floor 2nd Floor 3rd Floor 4th Floor 5th Floor
Daily average outdoor temperature 67.8% 62.3% 66.7% 70.4% 68.5% 70.7%
Time 49.1% 52.8% 46.3% 59.3% 37.0% 50.0%
Room CO 2 concentration 42.5% 43.4% 46.3% 50.0% 29.6% 43.1%
Room air temperature 40.7% 43.4% 31.5% 48.1% 35.2% 44.8%
Outdoor relative humidity 33.0% 28.3% 35.2% 37.0% 25.9% 37.9%
Room relative humidity 29.3% 24.5% 29.6% 42.6% 27.8% 22.4%
Wind speed 20.9% 20.8% 16.7% 25.9% 20.4% 20.7%
Time & Daily average outdoor temperature 7.3% 11.3% 3.7% 11.1% 3.7% 6.9%
Time & Room CO 2 concentration 5.9% 11.3% 5.6% 7.4% 1.9% 3.4%
None 3.3% 0.0% 1.9% 3.7% 5.6% 5.2%
Time & Room air temperature 2.9% 3.8% 0.0% 3.7% 3.7% 3.4%
Time & Room relative humidity 1.1% 1.9% 1.9% 0.0% 1.9% 0.0%
Time & Outdoor relative humidity 0.7% 0.0% 0.0% 3.7% 0.0% 0.0%
Time & Wind speed 0.7% 0.0% 0.0% 0.0% 1.9% 1.7%
 D. Calì et al. / Building and Environment 103 (2016) 54e69 67

5. Discussion typology, to the retrofit layout and to the level of the apartment
were investigated. The room typology (kitchen, bathroom, other)
In the analysis the daily average outdoor temperature was used and the window type (balcony-door, normal window) clearly lead
instead of the actual outdoor temperature measured each minute occupants to a diversified behavior. The time of day remained as the
to infer the probability of opening and closing of windows. The single most frequent driver for window opening across all room
wide outdoor temperature variation within a day may have typologies and room relative humidity was the second most
different impact on the probability of opening and closing the frequent driver in the kitchen and bathroom. In the living room and
windows. However, the choice was made to avoid correlation issues children room, the second most frequent driver was the CO2 con-
of this variable with other variables such as room temperature, centration. This indicates that the occupants' window opening
outdoor humidity and time of day. Additionally, this can facilitate behavior was driven by the activities in the home and aimed at
the use of the models for future applications when, for example, removing pollutants from the indoor air (in this case, moisture from
only daily average values of the outdoor temperature are available. showering and cooking and bioeffluents from the rest of the
The models that were used to investigate the drivers were all rooms). No systematic changes in drivers for the observed behavior
developed using the AIC as a means to select variables. It turned out were recognizable, when comparing the retrofit layout and the
that all variables in the models had a statistically significant impact level of the apartments. Interestingly, no pattern was recognized
(p < 0.05) on the window opening/closing probability. This implies when comparing B2 and B3E3 (equipped with exhaust ventilation)
that the observed correlations are not likely to be random. In the to B3E1 and B3E2 (equipped with balanced ventilation with heat
analyses, we investigated the sign of the correlation (positive or recovery). Even though the drivers for the opening and closing of
negative) but we did not investigate the size of the coefficients, windows were not different between the retrofit layouts, the
which is a measure of the size of the effect. As such, the results in average time with window open was different, both between the
this paper only state how the variables affected the window apartments, and between the entrances. For example, in 2012, B2E3
opening/closing probability (positively or negatively). It does not had open windows for 29% of the time, B3E2 for 39% of the time. In
state how big the effects were. order to investigate whether those differences were statistically
Furthermore, in order to keep the interpretation of the results at significant or not (e.g. due to the random distribution of tenants
a manageable level, interactions between continuous variables within the buildings), the distributions of the average positions of
were not included. Interaction terms between the categorical var- the windows of each apartment were used for a double comparison
iable “time” with the continuous variables were included, but were (each comparison for two entrances) in the Wilcoxon-Mann-
used by less than 10% of the models. Whitney test: the results indicated statistical significance of the
The main scope of the work was the investigation of the drivers differences between B2E1 and B3E2, and between B2E3 and B3E2,
for opening and closing windows, as discussed in Sections 4.2 and but only for the year 2012.
4.3 respectively. Time of day was the most frequent driver for Each window was modeled through a logistic function. Those
opening windows and the second most frequent driver for closing models can be implemented in building energy performance
windows. This could be an indication that the observed window simulation software. Since the occupants, and thus the models,
opening behavior was linked to specific activities which occur at react to variables that are influenced by the presence of occupants
specific times. E.g. a state change is less likely to occur when the (e.g. carbon dioxide concentration in the air), those models can only
occupants are asleep than when they are awake. And most occu- be used within a simulation, when occupants' presence profiles are
pants tend to sleep during the night. As such, the lower probability available. Only if presence patterns are provided, the CO2 emissions
of opening and closing windows during the night was probably caused by occupants could be reproduced in the dynamic simula-
linked to the occupants' sleeping patterns rather than the time it- tion, and therefore the model could render the occupants' reaction
self. Likewise, the peaks in state changes during morning and to this air quality parameter correctly. However, this has the
evening in Fig. 8 could be a consequence of typical activities at disadvantage that presence profiles and window models would not
those times of the day (e.g. showering, cooking, or simply moving be congruent, since those would come from different observations
from one room to another). However, since we did not gather in- or only from general assumptions. Furthermore, in the literature, a
formation about the occupants' activities, we cannot point to a model that reproduces the presence of occupants in the residential
direct explanation. sector at room level resolution, and based on observations, does not
In general, the time of day, the CO2 concentration and the exist at the time of writing. To test the models within a building
average outdoor temperature appeared most frequently in the energy simulation software, presence patterns simulating the
models of window opening and closing. These results support the movements of occupants between the rooms of an apartment could
tendency of the results in Ref. [27] and underline the importance of be modeled using an agent based approach through inhomoge-
including indoor air quality indicators in any attempts of modeling neous Markov Chains based on assumptions about the presence of
occupants' window opening behavior in residential buildings. each occupant and the probability of their transition between
The most frequently occurring driver for closing of windows was rooms at given timestamps. The occupants of 34 of the 60 apart-
the daily outdoor average temperature and in the majority of the ments were interviewed, and disclosed the number of occupants in
models the association was negative, i.e. the probability of closing a their own apartment, as well as the average time per day when the
window increased with decreasing daily outdoor average temper- apartment is not occupied, distinguishing between weekdays and
ature. This is in agreement with Fabi et al. [41], who found that weekends. Such information, together with the collected data in
outdoor temperature was negatively correlated with window the rooms, could be used to generate the necessary presence pat-
closing in two of the three best performing models of window terns, making use of the Bayes theorem and the Markov Chains
opening behavior. From this, a general picture emerges: Occupants technique [42].
tend to open the windows at specific times of day (probably asso- The LRA procedure was effective when focusing on the evalua-
ciated to activities) and when the CO2 concentration and relative tion of the drivers but had the disadvantage of generating two
humidity is elevated. They tend to close windows when it is cold models for each window, one for opening and one for closing. In
outside and at specific times of day (probably associated with their order to model the occupants' behavior, it would probably be more
activities). meaningful to make one model per apartment and use the room as
The differences in drivers for the behavior related to the room an explanatory variable [27]; alternatively, it would be possible to
68 D. Calì et al. / Building and Environment 103 (2016) 54e69

make one general model for opening and one for closing, using the sachliche Energieberatung?, available at: http://www.buildup.eu/
publications/1810 (accessed 16.05.15).
mixed effect modeling technique, and setting the apartment and
[4] L. Tronchin, K. Fabbri, Energy performance building evaluation in Mediterra-
the room as further explanatory variables [43]. nean countries: comparison between software simulations and operating
To generate congruent models of presence and behavior, with a rating simulation, Energy Build. 40 (2008) 1176e1187.
holistic approach, it would be beneficial to follow the procedure [5] H. Hens, W. Parijs, M. Deurinck, Energy consumption for heating and rebound
effects, Energy Build. 42 (2010) 105e110.
described within this paper for buildings with known occupancy [6] A.C. Menezes, A. Cripps, D. Bouchlaghem, R. Buswell, Predicted vs. actual
patterns. Those patterns could help on the one hand, to better and energy performance of non-domestic buildings: using post-occupancy eval-
more precisely fit the models, and on the other hand, to use the uation data to reduce the performance gap, Appl. Energy 97 (2012) 355e364.
[7] G. Dall'O', L. Sarto, A. Galante, G. Pasetti, Comparison between predicted and
models under the real occupants' presence conditions. actual energy performance for winter heating in high-performance residential
Logistic regression was confirmed to be a strong and robust buildings in the Lombardy region (Italy), Energy Build. 47 (2012) 247e253.
analysis methodology for investigating the drivers for occupants to [8] P. de Wilde, The gap between predicted and measured energy performance of
buildings: a framework for investigation, Autom. Constr. 41 (2014) 40e49.
interact with the built environment. The strengths of the models [9] A. Al-Mumin, O. Khattab, G. Sridhar, Occupants' behavior and activity patterns
are related to the high time resolution of the data, the use of an influencing the energy consumption in the Kuwaiti residences, Energy Build.
indoor air quality parameter as an explanatory variable, and the 35 (2003) 549e559.
[10] A.L. Pisello, C. Piselli, F. Cotana, Influence of human behavior on cool roof
distinction of the room use typology. The weaknesses are mostly effect for summer cooling, Build. Environ. 88 (June 2015) 116e128.
due to the lack of data related to the presence of occupants. [11] O. Guerra Santin, C.A. Tweed, In-use monitoring of buildings: an overview of
Through the logistic regression analysis, each window was data collection methods, Energy Build. 93 (2015) 189e207.
[12] T. de Meester, A.-F. Marique, A. de Herde, S. Reiter, Impacts of occupant be-
modeled singularly. Thus, 546 models were generated. In order to
haviours on residential heating consumption for detached houses in a
reduce the number of models while including the occupants' di- temperate climate in the northern part of Europe, Energy Build. 57 (2013)
versity, Haldi [28] suggested the use of a generalized linear mixed 313e323.
effect logistic regression model. This approach could be investi- [13] D. Müller, T. Osterhage, D. Calì, Quartierskonzept Energieeffizientes Rintheim
e wissenschaftliche Begleitung. Lehrstuhl für Geb€ aude- und Raumklima-
gated in future modeling activities, based on the observed data technik, RWTH Aachen University, 2012. Final Report.
from the demonstration buildings or of new field tests, where the [14] D. Calì, S. Rita, D. Müller, T. Osterhage, Holistic Renovation and Monitoring of
presence of occupants could be monitored. Residential Buildings, ECEEE, 2013.
[15] J. Virote, R. Neves-Silva, Stochastic models for building energy prediction
based on occupant behavior assessment, Energy Build. 53 (2012) 183e193.
6. Conclusions [16] D. Calì, P. Matthes, K. Huchtemann, R. Streblow, D. Müller, CO2 based occu-
pancy detection algorithm: experimental analysis and validation for office and
residential buildings, Build. Environ. (2015) 39e49.
The results of the logistic regression identified the drivers of the [17] I. Richardson, M. Thomson, D. Infield, A high-resolution domestic building
opening and closing of windows. The most common drivers for occupancy model for energy demand simulations, Energy Build. 40 (2008)
opening windows were time of day (for more than 70% of the 1560e1566.
[18] U. Wilke, F. Haldi, J.-L. Scartezzini, D. Robinson, A bottom-up stochastic model
modeled windows), and the indoor carbon dioxide concentration to predict building occupants' time-dependent activities, Build. Environ. 60
(for over 50% of the modeled windows). The most common drivers (2013) 254e264.
for closing windows were the daily average outdoor temperature [19] Energy in Buildings and Communities Programme, IEA-EBC Annex 66-
Definition and Simulation of Occupant Behavior in Buildings, available at:
(for almost 70% of the modeled windows) and the time of day (for http://www.annex66.org/.(accessed 19.01.16).
more than 50% of the modeled windows). [20] J. Page, D. Robinson, N. Morel, J.-L. Scartezzini, A generalised stochastic model
The room typology (kitchen, bathroom, other) and the window for the simulation of occupant presence, Energy Build. 40 (2008) 83e98.
[21] J. Page, Simulating Occupant Presence and Behaviour in Buildings (Ph.D.
type (balcony-door/normal window) affected the frequency with

thesis), Laboratoire d'Energie Solaire et de Physique du B^ atiment, Ecole pol-
which the drivers appeared in the models of both window openings ytechnique fe
de
rale de Lausanne, 2007.
and closings. The time of day and average outdoor temperature [22] H. Polinder, M. Schweiker, Ad van der Aa, K. Schakib-Ekbatan, V. Fabi,
R. Andersen, N. Morishita, C. Wang, S. Corgnati, P. Hiselberg, D. Yan, B. Olesen,
remained the most frequent drivers for opening and closing,
T. Bednar, A. Wagner, Occupant Behavior and Modeling: Separate Document
respectively, across all room typologies and window types. For the Volume II e Total Energy Use in Buildings e Analysis and Evaluation Methods
opening case, the room relative humidity was the second most e Final Report Annex 53, International Energy Agency e Institute for Building
frequent driver (after the time of day) in the kitchen and bathroom, Environment and Energy Conservation, 2013.
[23] D. Yan, W. O'Brien, T. Hong, X. Feng, H. Burak Gunay, F. Tahmasebi,
while in the living room and children's room, the second most A. Mahdavi, Occupant behavior modeling for building performance simula-
frequent driver was the CO2 concentration. This indicates that the tion: current state and future challenges, Energy Build. 107 (2015) 264e278.
occupants' window opening behavior was driven by the activities [24] European Environment Agency, Household energy consumption per dwelling
by end-use: Chart d Household energy consumption per dwelling by end-
in the home (like showering and cooking activities, which produce use, available at: http://www.eea.europa.eu/data-and-maps/daviz/energy-
moisture) and by pollutants in the indoor air (windows were consumption-by-end-uses-1#tab-chart_1 (accessed 11.02.16).
opened to remove moisture from kitchen and bathroom and to [25] V. Fabi, V.A. Rune, S. Corgnati, B.W. Olesen, Occupants' window opening
behaviour: a literature review of factors influencing occupant behaviour and
remove bioeffluents from the rest of the rooms). models, Build. Environ. 58 (2012) 188e198.
Finally, no systematic changes in drivers for the observed [26] R. Fritsch, A. Kohler, M. Nygård-Ferguson, J.-L. Scartezzini, A stochastic model
behavior were recognizable, when comparing the ventilation of user behaviour regarding ventilation, Build. Environ. 25 (1990) 173e181.
[27] R. Andersen, V. Fabi, J. Toftum, S.P. Corgnati, B.W. Olesen, Window opening
principles and the level of the apartments.
behaviour modelled from measurements in Danish dwellings, Build. Environ.
69 (2013) 101e113.
Acknowledgment [28] F. Haldi, D. Robinson, Interactions with window openings by office occupants,
Build. Environ. 44 (2009) 2378e2395.
[29] R.V. Andersen, Occupant Behaviour with Regard to Control of the Indoor
We gratefully acknowledge the financial support of BMWi Environment (Ph.D. thesis), 2009.
(03ET1105A) and of E.ON New Build and Technology. [30] R.V. Andersen, B.W. Olesen, J. Toftum, Modelling window opening behaviour
in Danish dwellings, in: International Conference on Indoor Air Quality and
Climate, Austin, Texas, 2011.
References [31] S. D'Oca, T. Hong, A data-mining approach to discover patterns of window
opening and closing behavior in offices, Build. Environ. 82 (2014) 726e739.
[1] R. Haas, H. Auer, P. Biermayr, The impact of consumer behavior on residential [32] T.J. Hastie, R.J. Tibshirani, Generalized Additive Models, Chapman and Hall/
energy demand for space heating, Energy Policy (1998) 195e205. CRC, London, 1990.
[2] R. Haas, P. Biermayr, The rebound effect for space heating empirical evidence [33] M. Schweiker, M. Shukuya, Comparison of theoretical and statistical models of
from Austria, Energy Policy (2000) 403e410. air-conditioning-unit usage behaviour in a residential setting under Japanese
[3] H. Erhorn, Bedarf e Verbrauch: Ein Reizthema ohne Ende oder die Chance für climatic conditions, Build. Environ. 44 (2009) 2137e2149.
 D. Calì et al. / Building and Environment 103 (2016) 54e69 69

[34] R. Rovers, J. Kimman, C. Ravesloot, Towards 0-impact Buildings and Built [39] K. Wolfrum, A. Bloch, Messsystem für vielkanaliges Monitoring von Ener-
Environments, Techne Press, Amsterdam, 2010. €men und Raumklimadaten in energetisch sanierten Wohngeb€
giestro auden in
[35] D. Calì, T. Osterhage, D. Müller, Field study of different retrofit solutions for Karlsruhe Rintheim, Forschung aktuell 2012, S. 18e20, Hochschule Karlsruhe
residential housing, in: Contribution to a Conference Proceedings, Euregional e Technik und Wirtschaft, 2012.
Conference Sustainable Building, SB 10, 2010. [40] CEN, Ventilation for Buildings e Design Criteria for the Indoor Environment,
[36] D. Calì, T. Osterhage, D. Müller, Rebound effect related to retrofit solutions for 1998.
residential housing monitoring data from a field test, in: SB11 Helsinki World [41] V. Fabi, R.K. Andersen, S. Corgnati, Verification of stochastic behavioural
Sustainable Building Conference, 2011. models of occupants' interactions with windows in residential buildings,
[37] D. Calì, T. Osterhage, D. Müller, Retrofit solutions for residential buildings, Int. Build. Environ. 94 (2015) 371e383.
J. Sustain. Build. Technol. Urban Dev. 2 (2011) 131e136. [42] H. Han, C. Shin, Bayesian Analysis of Passenger Occupancy Based on the CO2
[38] D. Calì, T. Osterhage, R. Streblow, D. Müller, Advanced renovation of three Concentration Measured in a Subway Station, Clima.
residential buildings e evaluation of a field test, in: Contribution to a Con- [43] F. Haldi, A probabilistic model to predict building occupants' diversity towards
ference Proceedings, Energy Efficiency in Domestic Appliances and Lighting their interaction with the building envelope, in: Conference of International
Proceedings of the 7th International Conference, 2013. Building Performance Simulation Association, 2013.