Comfort-efficiency-equilibrium
A Proactive, at Room Level Individualized Climate Control System for Smart
Buildings
Antonios Karatzoglou
1,2
, Julian Janßen
1
, Vethiga Srikanthan
1
, Yong Ding
1
and Michael Beigl
1
1
TECO Research Group, Karlsruhe Institute of Technology (KIT), Karlsruhe, Germany
2
Robert Bosch GmbH, Corporate Sector Research and Advance Engineering, Renningen, Germany
Keywords:
Energy Efficiency, Thermal Comfort, Proactivity, Personalization, Smart Building.
Abstract:
Energy efficiency and thermal comfort depict two key topics in indoor climate controlling domain. HVAC
systems are one of the biggest energy consumers in nowadays’ households and yet they have difficulties in
reaching the users’ optimal comfort. We are presenting SVReCLCE, a proactive two-fold climate controlling
approach that takes explicitly both energy consumption, as well as comfort in consideration. A user study in
an office environment shows that our solution can in practice achieve up to 49% energy savings by keeping
the personal comfort level high at the same time. Therefore, SVReCLCE sets a solid basis for future work in
the field of climate control for low-energy buildings.
1 INTRODUCTION
Heating, ventilation and air conditioning systems
(HVAC) represent the largest energy consumers in
buildings. In expectation of hotter summers in the
future, the energy requirement in the climate sector
is expected to increase. Hence, optimizing HVAC
systems is an important step towards decreasing our
ecological footprint. Ideally, an optimal HVAC sys-
tem would be fully aware both of environmental and
contextual conditions as well as of individual charac-
teristics of the occupants and their behavior. Recent
products like Hive
1
and Nest
2
signalize the start of
tangible intelligent climate control in buildings with
focus on energy saving and a better human-machine-
interaction. Still, there is room for improvement.
In this work, we introduce a proactive controlling
concept capable of adapting and reacting on both the
users’ direct feedback of sense of comfort as well as
the actuators energy consumption. In addition, a ma-
chine learning based model is built, trained and used
for modeling and predicting the residents’ attendance.
Finally, a rule base considering common sense users’
behavior is created to support additionally our system.
We tested and evaluated our approach by comparing
1
https://www.hivehome.com/
2
https://nest.com/thermostat/meet-nest-thermostat/
it with two other systems serving as a reference on
a building cooling scenario. For this purpose we de-
signed and conducted a 4-week experiment in an of-
fice environment. The details of our work are pre-
sented as follows. Our SVR enhanced controller is
introduced in section 3. Section 4 describes our ex-
perimental setup. The results of our experiment are
presented in section 5. The last section concludes the
work.
2 RELATED WORK
Collecting a variety of sensor data and utilizing them
for accurate temperature control plays a core role in
many current researches. The art in which intelli-
gent controllers handle these data varies widely. (Ed-
wards et al., 2012) explore seven different machine
learning techniques for predicting the next hour res-
idential energy consumption, such as Feed Forward
Neuronal Network (FFNN) and Support Vector Re-
gression (SVR) a.o. Depending on the data set some
techniques perform better than others. Overall, SVR
represents one of the best methods applied across all
the data sets. (Megri and El Naqa, 2016) use SVM
to predict thermal comfort indices. For this purpose,
the SVM was trained with a group of different factor
258
Karatzoglou, A., Janßen, J., Srikanthan, V., Ding, Y. and Beigl, M.
Comfort-efficiency-equilibrium.
DOI: 10.5220/0006308002580265
In Proceedings of the 6th International Conference on Smart Cities and Green ICT Systems (SMARTGREENS 2017), pages 258-265
ISBN: 978-989-758-241-7
Copyright © 2017 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

data sets influencing the thermal comfort, such as air
temperature, relative humidity, clothing value etc. An
interesting SVM based approach close to our own but
applied in a complete different domain is the one pro-
posed by (Deng et al., 2014), which is used to improve
the position precision of a servo system. They model
and predict the position error with SVM and then feed
it back to the system, thereby improving the precision.
They show that combining SVM with a PID controller
can raise significantly the prediction accuracy. Occu-
pancy prediction is one way of reducing energy con-
sumption by switching down the heating or cooling,
if no one is in the room like in (Koehler et al., 2013)
and (Scott et al., 2011). However, prediction is not
restricted in occupancy when it comes to energy effi-
ciency and user comfort. (Ellis et al., 2013) use tem-
perature sensors, gas meter sensors, outdoor tempera-
ture sensors and boiler or furnace time in order to pre-
dict the indoor temperature. By predicting the tem-
perature they are able to control the boiler or furnace
more accurately. (Oldewurtel et al., 2012) use the
weather forecast to control the temperature more effi-
ciently. Feldmaier et al. swear by a great number of
fixed and portable sensors in (Feldmeier and Paradiso,
2010). Beyond that they also collect the user’s feed-
back about their thermal comfort by letting them push
one of three buttons (hot, cold and neutral). A control
module collects all information and decides whether
to open or close the window or change the tempera-
ture. Thermal comfort describes the feeling that peo-
ple have about the ambient temperature. There are
different components that influence thermal comfort
like air temperature, air velocity, air humidity, radia-
tion, clothes, activity, outdoor temperature and other
(Dentel and Dietrich, 2013). Fanger created an index
to represent the comfort level - the Predicted Mean
Value (PMV) (Fanger, 1970). It considers energy ex-
change, clothing, air temperature, humidity, velocity
and radiation temperature. The Predicted Percentage
of Dissatisfied (PPD), also developed by Fanger, de-
scribes the satisfaction of a group of people. PMV
and PPD are widely used indices as in (Kalz et al.,
2013) and (Yun and Won, 2012).
3 SVR ENHANCED CLOSED
LOOP COMFORT-ENERGY
CONTROLLER (SVRECLCE)
Our proposed closed loop comfort-energy controller
consists of three core components: A PID based com-
ponent, a SVR based prediction component and a rule
based component that pieces i.a. the previous two to-
gether. The system takes both the comfort feedback
of the users, as well as the energy consumption into
account to control the temperature. Moreover, it is
capable of adapting to occupancy patterns describing
how, when and how long people do occupy their of-
fice or living space. The individual components and
their part in the overall system will be discussed in
the next paragraphs. An illustration of SVReCLCE is
shown in figure 1.
3.1 PID-based Component
This section presents in detail the core of our comfort-
energy controller. It is designed to control the room
temperature, depending on both the user comfort and
the energy consumption. To achieve this we built a
twofold comfort-energy controller consisting of two
independent PID controllers (PID
E
and PID
C
) respec-
tively. One for each control variable. While the first
aims at reducing the energy consumption, lies the
scope of the second in optimizing the users’ comfort.
PID
E
(t)=K
p,E
∗e
C,E
(t)+K
i,E
∗
R
t
0
e
C,E
(t)∂t+K
d,E
∂e
C,E
(t)
∂t
(1)
PID
C
(t)=K
p,C
∗e
C,E
(t)+K
i,C
∗
R
t
0
e
C,E
(t)∂t+K
d,C
∂e
C,E
(t)
∂t
(2)
PID
Fused
(t)=w
E
∗PID
E
+w
C
∗PID
C
(3)
w
E
=1−w
C
(4)
In 1 and 2 we can see the output function of the two
PID controllers. They describe the cumulative way,
in which the value of the manipulated variable of the
energy and the comfort PID controller respectively
is being determined. Each function consists of three
K
x
-weighted terms: a proportional, an integral and
a derivative term. The overall behavior of the con-
troller is defined by the summation of the terms and
consequently by the combination of the aforemen-
tioned coefficients. So, the steadiness of the system
can be adjusted by tuning the integral term. Modi-
fying the derivative term fine tunes the speed of the
controller. PID
Fused
represents the overall output of
the comfort-energy core component, which is in our
case the room temperature. This value stems from
the weighted fusion of the two single PID controllers
PID
E
and PID
C
. Configurating the weights w
C
and
w
E
allows a twofold adjustment of the controller’s be-
havior based on both comfort and energy efficiency
accordingly. This enables us to shift from a comfort
based system towards a more energy efficiency aim-
ing one and vice versa.
The comfort of the user is represented by a com-
fort index described in chapter4. It is mapped to a
numerical scale and is subsequently passed on the
Comfort-efficiency-equilibrium
259

Figure 1: SVReCLCE control loop and its elements.
PID
comfort
controller, or more precisely, it is the sys-
tem deviation from the user comfort set point e
C
(t),
which is routed back to the input of the PID controller.
e
C
(t)=C
set
(t)−C
meas
(t) (5)
Analogous to PID
C
the energy focused controller
PID
E
takes the system deviation from the energy set
point e
E
(t) as input.
3.2 SVR-based Prediction Component
The twofold PID-based core component is extended
by a prediction component. The predictor capacitates
the overall controller to behave proactively according
to the expected attendance of the user and thus en-
ables the system to create a comfortable environment
before the users arrives the temperature controlled
area. The underlying presence model is built on the
basis of a support vector machine/regression (SVM)
(Abe, 2005) with a radial basis function (RBF) kernel
described in 10. The goal in SV-based regression is
to find a function f (x), which shows a maximal devi-
ation ε from the obtained objects y
i
for all the given
training data (x
i
, y
i
). The quality of the estimation of
SVR is measured by the loss function shown in for-
mula 6. The actual estimation is given in 9. G(x
n
,x)
represents the kernel function of the SVM. ε defines
the space in which no penalty is added in the loss
function. The penalty error is defined by the parame-
ter C. Finally N is the number of the support vectors
and b gives the bias. The notation used in 6 to 9 corre-
sponds to the dual formulation of SVR. Our presence
model is built upon two features. The time of day and
an indicator that distinguishes between workdays and
non workdays. This basic day differentiating indica-
tor yields a better true positive prediction rate despite
the small amount of the training data. The prediction
component adapts to the current situation by provid-
ing both a short- and a long-term prediction. For this
purpose it considers two separate prediction horizons
of different size. While a short horizon is sufficient
during a normally busy workday in an office, longer
inactivities, e.g. before the begin of the workday and
in holidays, are handled and compensated by a larger
horizon. The selection of the horizons type is deter-
mined by a set of rules comprised in a rule base.
L(α)=
1
2
∑
N
i=1
∑
N
j=1
(α
i
−α
∗
i
)(α
j
−α
∗
j
)G(x
i
,x
j
)
+ε
∑
N
i=1
(α
i
+α
∗
i
)−
∑
N
i=1
y
i
(α
i
−α
∗
i
)
(6)
∑
N
n=1
(α
n
−α
∗
n
)=0 (7)
∀n:0≤α
n
≤C;∀n:0≤α
∗
n
≤C (8)
f (x)=
∑
N
n=1
(α
n
−α
∗
n
)G(x
n
,x)+b (9)
G(x
1
,x
2
)=exp−
kx
1
−x
2
k
2
2σ
2
) (10)
3.3 Rule-based Component
The role of the rule based component lies in extend-
ing the residential behavior model built by SVR and
therefore supporting the prediction and optimizing
the overall control behavior. The rules cover situa-
tions like long absence, short lunch breaks, and even
shorter absences like quick toilet stops during the day
and help to revise inaccurate predictions by defining
two core behavior modes concerning the automatic
start and stop of the system respectively: automatic
on and automatic off. The controller alternates be-
tween these two states during the day.
The automatic on mode relates to the wakening
procedure of the system. During this mode the sys-
tem monitors simultaneously both the current output
SMARTGREENS 2017 - 6th International Conference on Smart Cities and Green ICT Systems
260

of the present detection mechanism, as well as the
outcome of the prediction component and reacts ac-
cordingly. If at least one of them triggers an event,
the system is turned on and its mode is being set to
automatic off, waiting for the right conditions to turn
itself off again. The presence detection mechanism
applies a detection time window to prevent the system
from reacting too sensitive to the output of the motion
detection sensor. An event in an area is only triggered
if the duration of the motion in the particular area
exceeds the detection window length. Some factors
need to be taken into consideration when determining
the length of the detection window. For example too
short visits in the corresponding area should not start
the system. A long enough observation window can
filter this kind of short actions out. On the other hand
if the window is too long, it can lead to discomfort of
the user, due to a slower response time of the system.
In addition to presence detection, a SVR based pre-
diction model is used to estimate the day and time of
the next appearance of a user. Such predictions sup-
ply the system with an extra series of triggering events
enabling it to react in a forward-looking manner.
In the automatic off mode the system monitors
continuously the presence of the users. After a certain
period of time of motionlessness the system is turned
automatically off and the automatic on mode is set.
The length of the positive detected absence observa-
tion window plays again a significant role. On the one
hand, short absence of the user should not be con-
sidered for sending a triggering event. On the other
hand choosing a very long interval would result to a
late turning-off and would therefore affect the over-
all energy efficiency of the system. A correct balance
among these two extremes should be sought.
The user has at any time the possibility to bypass
the automatic process. Sending a manual start signal
would result in starting the system and setting the au-
tomatic off mode. In the case of manually turning the
system off, the system can only be restarted automat-
ically after an additional positive absence detection
that sets it first to the automatic on mode. This pro-
motes further the behavior stability of the controller.
4 USER STUDY AND
EXPERIMENTAL SETUP
We designed and conducted a 4-week user study in
five different rooms of an office building in order to
evaluate our system in practice. Each office contained
one to three people. All in all we had nine partic-
ipants for the study, whereby one of them was our
control person, serving as a comparison. We used the
Figure 2: Sensor and actuator deployment.
first week for testing our infrastructure while collect-
ing training data at the same time. The remaining 3
weeks were then used to test the performance of our
approach among the one of two further reference sys-
tems: one week for each. Figure 3 gives an overview
of our server-based experimental infrastructure.
A mobile air-conditioning device (AC) was in-
stalled in every room, except in the one of our test
person. A Raspberry Pi served both as remote control
unit for the ACs, as well as gateway for collecting and
forwarding all sensor data to our server. In addition to
the AC devices, each office was also equipped with a
temperature and humidity sensor, one to two motion
detectors, depending on the size of the room, and a
smart meter for capturing the energy consumption of
the ACs. A temperature and humidity sensor was also
installed outside of the building in the same floor to
gauge the outdoor temperature. Our server hosted two
different databases. A time series database (TSDB)
was used to store the sensor values. We have chosen
an open solution called influxdata
3
for this purpose.
At the same time, a MySQL Database was used to
store the filled surveys of the users. To safeguard a
flawless operation of our system, we monitored the
data flow with Kapacitor and we used Chronograph
(both from influxdata), a visualization tool, to con-
duct further periodical sample checks. In order to
achieve room level controlling, a separate instance of
the tested climate controller was initiated for each sin-
gle room.
During the study, the participants were asked to
fill out a set of ready-to-use e-surveys on an hourly
basis via a web application. The users were given ad-
ditionally the ability to set a temperature of their de-
sire in their room at any time via the same web inter-
face by adjusting a virtual thermostat. The questions
asked in the surveys aimed mainly at getting feed-
back from the users about their thermal sense of well
being. These were used to derive the corresponding
3
https://www.influxdata.com/
Comfort-efficiency-equilibrium
261

thermal comfort index, which in turn was fed back
to our comfort-energy controller. Our comfort index
is based on the 7-level ASHRAE scale
4
, which de-
scribes 7 states from cold to hot surrounding a neutral
(optimal) value which lies in the middle. Our com-
fort index leverages two different kinds of user feed-
back in parallel. Direct comfort is measured through
the survey and the corresponding queries in which the
user is asked to rank the current felt comfort. Indi-
rect comfort is derived from her interaction with the
virtual thermostat knob. The set value on the thermo-
stat is mapped to the same 7-level scale of ASHRAE
and is finally forwarded to the server after every user
interaction. Some rooms in our testing environment
occupy more than one users. In the case of receiv-
ing different comfort indices from multiple users, we
obtain a mean value by averaging them in order to
achieve the best possible comfort. The floor plan of
one of the offices is shown in figure 2. The remaining
offices show a similar topology.
5 EVALUATION AND
DISCUSSION
In this section, we discuss the performance of our ap-
proach and report the final results of our experimen-
tal study. We implemented and tested three differ-
ent climate controlling approaches during our study:
a basic open-loop (OL) controller, the closed-loop
comfort-energy controller (CLCE) and the SVR en-
hanced closed loop comfort-energy controller (SVRe-
CLCE). Each of them was field-tested for one week.
The open-loop (OL) controller allows the user
to control directly the air conditioner in her room.
Hence, the A/C unit reacts solely to input signals
passed directly through the web thermostat. The basic
principles of the closed-loop comfort-energy (CLCE)
controller are described in section 3.1. It is a two-fold
PID controller, which takes both comfort and energy
consumption into account for regulating the tempera-
ture in a room. We choose a weight of 4/5 in favor
of the comfort. In contrast to the proactive SVRe-
CLCE, starting and stopping of the system is per-
formed here manually by the user. We modeled com-
fort and energy consumption separately with MAT-
LAB Simulink
5
and chose a state space model with
a state vector length of 3. The System Identification
Toolbox was used to identify our models. The PID pa-
rameters were estimated by testing and leveraging the
4
American Society of Heating, Refrigerating and Air-
Conditioning Engineers (https://www.ashrae.org/)
5
https://de.mathworks.com/products/simulink.html
PID Tuner with the goal of obtaining rapid response
and stability at the same time. The following param-
eters have been determined and subsequently applied
on the PID comfort controller: K
p
= 2, K
i
= 0.2, K
d
=
−0.2. Analog, the parameters used by the PID energy
controller are: K
p
= −0.02,K
i
= −0.002,K
d
= 0.0.
These values were utilized for the configuration of
both the CLCE and the SVReCLCE controllers. Neu-
tral comfort and 450W power have been chosen as
reference values for the two controllers respectively.
The energy consumed over a period of 60 minutes
served as input for the energy controller.
After statistically analyzing the users’ behavior
we were able to define the appropriate moving time
window lengths utilized by our prediction algorithm.
So, we selected a 30 minutes time interval as predic-
tion horizon to cover the night, before the workday
begins, while the window during the day is set to 10
minutes due to the increasing change of attendance
state in an office environment. The submission of the
end of day online questionnaire served as an end-of-
workday flag. A detection window of 5 minutes is
used for the actual presence detection and a period of
15 motionless minutes is used to define absence in or-
der to filter short breaks out.
5.1 Energy
Table 1 lists the average energy consumption values
of each room, as well as of all rooms together dur-
ing a period of one week respectively. The average
of the energy consumption over one week and over
both all rooms, as well as rooms B and C (in brack-
ets) for each controller is displayed on the first row,
while the remaining rows below contain the average
one-week consumption in each room. Room D shows
no energy consumption because no A/C unit was de-
ployed there due to being our reference room, where
our control person was. The control person serves as
a reference for the study by providing information re-
garding her sense of comfort from an office without
any climate control. This person is thus only able to
fill out our e-questionnaires but he/she cannot control
the temperature.
As can be seen from table 1, our proactive ap-
proach SVReCLCE yields the best overall results
when it comes to energy efficiency among the rooms
B and C. CLCE and OL take the second and third
place respectively. Looking the total average, our
SVReCLCE performs better than CLCE, but slightly
worse than the OL approach. A closer look at the ta-
ble, shows that room B clearly stands out in a positive
manner. Our controller here helps saving the most en-
ergy. Nearly 14.37% could be saved by SVReCLCE
SMARTGREENS 2017 - 6th International Conference on Smart Cities and Green ICT Systems
262

Figure 3: Overview of the HW and SW experimental infrastructure.
Table 1: Average power consumption over all rooms B and C only and over each room separately.
OL CLCE SVReCLCE
average [W]
(Avg of B and C only)
104.7426
(113.9577)
109.6741
(95.0499)
107.2958
(90.6457)
room A [W] 86.3123 138.9226 140.5962
room B [W] 185.8407 110.4873 94.6095
room C [W] 42.0747 79.6124 86.6819
room D [W] 0 0 0
0
20
40
60
80
100
120
140
160
180
200
Avg Avg of rooms B and C room B
[W]
OL
CLCE
SVReCLCE
Figure 4: Average Power consumption [W] of our tested
controllers among all three rooms (Avg), among rooms B
and C and at room B.
compared to the CLCE and almost half the energy
(49.09%) consumed by the open loop controller OL.
Our controller does not provide us with similar good
results at room A though. On the contrary, SVRe-
CLCE led to a 62.89% increased energy consump-
tion in comparison with the OL. This relies on the
irregular absence of the associated study participant
that used the particular room during the whole third
week which our training data don’t cover. In addi-
tion to the lack of an appropriate data set, two extra
major influence factors have also to be noted here.
Analysis of the questionnaires showed that the par-
ticipant of the same room (A) felt disturbed by the
operating noise of the A/C unit, resulting in extreme
noise-sensitive behavior in which climate control was
used only when absolutely necessary. This reflects,
for the most part, the low power consumption during
the first week, where the users are able to turn the
system on and off manually themselves. A similar
situation is found in room C, illustrated in 2, which
was used explicitly to test the limits of all 3 control-
ling approaches. Both participants used the room on
a quasi-irregular basis and did not occupy the room
every day during the four weeks of our experiment.
The time of day varied as well. Still, certain atten-
dance days and corresponding times of day remained
almost consistent during the study, building weak pat-
terns. Our approach might consume higher amount
of energy compared to OL, but is still more efficient
than in room A due to recognizing and utilizing the
aforementioned patterns. At the same time, our ba-
sic controller CLCE performs in case of room B bet-
ter than OL, but worst compared to SVReCLCE. In
rooms A and C though, it shows a slightly better be-
havior than SVReCLCE. This can be attributed to the
same reasons as with OL mentioned above. It must
be noted that in case of a direct controller like OL, the
energy consumption depends highly on the users’ be-
havior. The more energy-conscious a person handles,
the higher the efficiency of OL would be. All in all, up
to 49,09% energy could be saved with SVReCLCE.
This is also visualized in figure 4. This extrapolates
to almost 400kWh for a period of 6 months, say from
April to September.
Comfort-efficiency-equilibrium
263

Table 2: Average comfort and standard deviation over all rooms, over rooms B and C only and over each room separately.
OL CLCE SVReCLCE
∅ σ ∅ σ ∅ σ
average comfort
(Avg of B and C only)
0.2877
(0.6087)
1.0088
(0.9562)
0.3657
(0.9057)
1.3276
(1.1576)
-
(0.5837)
-
(0.8692)
comfort room A (1p) -0.3542 0.8536 -0.7143 0.9124 - -
comfort room B (2p) 0.4482 0.8544 0.8947 1.0205 0.8947 0.9676
comfort room C (2p) 0.7692 1.1200 0.9166 1.2555 0.2727 0.4453
comfort room D (1p) 1.4857 1.2732 0.0000 0.7071 2.7000 0.4582
5.2 Comfort
On the top of the table 2 we can see the average com-
fort and the corresponding standard deviation values
over the respective 1-week period of use over both all
rooms, as well as rooms B and C (in brackets) for
each controller. The three rows below show the aver-
age comfort and standard deviation measured in office
A, B and C respectively. At last, room D is the room
of our control person. Minus comfort values [-3, 0)
point out the fact that it is too cold for the user and
positive ones (0, +3] state the opposite. The objective
is to get a value close to zero (0). Table 2 indicates
a similar overall performance distribution among the
tested controllers, such the one seen in the energy re-
lated paragraph. Once again SVReCLCE provides
the best results among rooms B and C with an av-
erage comfort value of 0.58. It is followed by OL and
CLCE with a comfort value of 0.61 and 0.91 respec-
tively. As mentioned before, room A was unoccupied
during the third week and was therefore not consid-
ered in the averaging process. By looking closer at
room B and C, one can see that our two test persons
in room C found the temperature by SVReCLCE al-
most 3 (2.820) times more comfortable than by OL
and 3.36 times than CLCE. And this, irrespective of
SVReCLCE using an imperfect occupancy prediction
model (see chapter 5.1). The particular outcome re-
flects in the end our weighting factors chosen for
our 2-fold PID controller, namely 80% comfort and
20% energy efficiency. However in room B, the best
comfort was achieved by the OL controller, where
the users controlled manually the A/C devices them-
selves. SVReCLCE and CLCE take both the second
place with exact the same value. Nevertheless, the
difference between the corresponding comfort values
is fairly small. Comfort value of one (+1) stands for
”slightly warm” on the ASHRAE scale. Hence SVRe-
CLCE results in an environment that is even more
comfortable than slightly warm, which in turn is a
very good outcome for our controller. Looking in the
table at room D we can see that our reference person
felt significantly uncomfortable (in terms of thermal
wellbeing) during week one and three. Furthermore,
thermal comfort is a subjective feeling. Thus, table 2
expresses to a certain degree also the individual satis-
faction of the participants of our study. Our test per-
son in room A for instance felt the temperature in both
first two weeks in average as slightly cold, despite the
warm weather (especially in the first week) and the
fact that he was able to turn off the A/C unit at any
time during this time. The later raises also other is-
sues. Namely, we must not forget that the table 2 gives
only insight about the thermal and not general com-
fort. Users handling by themselves tend to be more
easygoing with their own decisions and less critical
about their own actions. On the other hand, they ex-
pect a perfect outcome from an ”intelligent” system
like ours. This explains additionally the good thermal
comfort values of OL, at the expense of the general
comfort though. SVReCLCE works without the need
of user interaction. This makes it overall more com-
fortable.
6 CONCLUSIONS AND FUTURE
WORK
In this paper, we present SVReCLCE, a two-fold pre-
dictive cooling system that takes explicitly both ther-
mal comfort and energy consumption into considera-
tion in order to provide an optimal balanced outcome
for the inhabitants. At the same time, adaptation and
personalization stand in the foreground. SVReCLCE
is able to shift its focus between energy and comfort
at users’ option. Thereby, our approach meets both
needs of environmental awareness, as well as per-
sonal well being. We tested and evaluated our sys-
tem in comparison to the non-predictive CLCE and
OL, a simple open-loop controller used as baseline,
in practice. We could show that SVReCLCE can sig-
nificantly contribute to saving energy whilst keeping
a high comfort level at the same time. We could also
find a few weak spots in our approach though. Han-
dling irregular attendance is a major issue. While
SVReCLCE performs well by quasi-irregularity (see
chapter 5.1 room C), this is not the case by total un-
SMARTGREENS 2017 - 6th International Conference on Smart Cities and Green ICT Systems
264

expected user behavior such as the visit of a 5-day
conference. This is due to missing knowledge and in-
accuracy of our occupancy model caused in turn by
an incomplete training data set.
In our future work we plan to increase the quality
of the prediction through a larger training data set. In
addition, we plan to extend our approach and enrich
it through semantic information comprising general,
domain and user-specific knowledge such as the per-
sonal calendar of the inhabitants and their personal
preferences. This would complement the missing in-
formation of our machine learner’s model and help
manage irregular behavior. Furthermore, a dynamic,
situation-dependent balance between the two inde-
pendent PID controllers could also improve our sys-
tem both in terms of energy efficiency and user’s’ sat-
isfaction. The needs and the requirements of the user
vary over the day depending on the situation. Having
an extra learning feature for keeping track of this in-
formation and feeding it back to the prediction model
of our controller would improve further the final solu-
tion.
ACKNOWLEDGEMENTS
This work was funded by the German Federal Min-
istry of Education and Research within the Soft-
ware Campus KEESmartHome project (grant number
01IS12051).
REFERENCES
Abe, S. (2005). Support vector machines for pattern classi-
fication, volume 53. Springer.
Deng, C., Xie, S., Wu, J., and Shao, X. (2014). Position
error compensation of semi-closed loop servo system
using support vector regression and fuzzy pid control.
The International Journal of Advanced Manufactur-
ing Technology, 71(5-8):887–898.
Dentel, A. and Dietrich, U. (2013). Thermische
behaglichkeit–komfort in geb
¨
auden.
Edwards, R. E., New, J., and Parker, L. E. (2012). Predict-
ing future hourly residential electrical consumption:
A machine learning case study. Energy and Buildings,
49:591–603.
Ellis, C., Hazas, M., and Scott, J. (2013). Matchstick:
A room-to-room thermal model for predicting indoor
temperature from wireless sensor data. In Proceedings
of the 12th international conference on Information
processing in sensor networks, pages 31–42. ACM.
Fanger, P. (1970). Thermal comfort: Analysis and applica-
tions in environmental engineering. Danish Technical
Press.
Feldmeier, M. and Paradiso, J. A. (2010). Personalized hvac
control system. In Internet of Things (IOT), 2010,
pages 1–8. IEEE.
Kalz, D., H
¨
olzenbein, F., Pfafferott, J., and Vogt, G. (2013).
Nutzerzufriedenheit mit dem thermischen komfort in
b
¨
urogeb
¨
auden mit umweltenergiekonzepten. Bau-
physik, 35(6):377–391.
Koehler, C., Ziebart, B. D., Mankoff, J., and Dey, A. K.
(2013). Therml: occupancy prediction for thermo-
stat control. In Proceedings of the 2013 ACM inter-
national joint conference on Pervasive and ubiquitous
computing, pages 103–112. ACM.
Megri, A. C. and El Naqa, I. (2016). Prediction of the
thermal comfort indices using improved support vec-
tor machine classifiers and nonlinear kernel functions.
Indoor and Built Environment, 25(1):6–16.
Oldewurtel, F., Parisio, A., Jones, C. N., Gyalistras, D., Gw-
erder, M., Stauch, V., Lehmann, B., and Morari, M.
(2012). Use of model predictive control and weather
forecasts for energy efficient building climate control.
Energy and Buildings, 45:15–27.
Scott, J., Bernheim Brush, A., Krumm, J., Meyers, B.,
Hazas, M., Hodges, S., and Villar, N. (2011). Preheat:
controlling home heating using occupancy prediction.
In Proceedings of the 13th international conference on
Ubiquitous computing, pages 281–290. ACM.
Yun, J. and Won, K.-H. (2012). Building environment anal-
ysis based on temperature and humidity for smart en-
ergy systems. Sensors, 12(10):13458–13470.
Comfort-efficiency-equilibrium
265