Eleventh International IBPSA Conference

Glasgow, Scotland
July 27-30, 2009

SENSOR-BASED OCCUPANCY BEHAVIORAL PATTERN RECOGNITION FOR

ENERGY AND COMFORT MANAGEMENT IN INTELLIGENT BUILDINGS
Bing Dong1, Burton Andrews2
1

Center for Building Performance and Diagnostics, Carnegie Mellon University, Pittsburgh,
PA, U.S.A. 15213
2
BOSCH Research and Technology Center, Pittsburgh, PA, U.S.A. 15212

ABSTRACT
There has been extensive research focusing on
developing smart environments by integrating data
mining techniques into environments that are
equipped with sensors and actuators. The ultimate
goal is to reduce the energy consumption in buildings
while maintaining a maximum comfort level for
occupants. However, there are few studies
successfully demonstrating energy savings from
occupancy behavioural patterns that have been
learned in a smart environment because of a lack of a
formal connection to building energy management
systems. In this study, the objective is to develop and
implement algorithms for sensor-based modelling
and prediction of user behaviour in intelligent
buildings and connect the behavioural patterns to
building energy and comfort management systems
through simulation tools. The results are tested on
data from a room equipped with a distributed set of
sensors, and building simulations through
EnergyPlus suggest potential energy savings of 30%
while maintaining an indoor comfort level when
compared with other basic energy savings HVAC
control strategies.

INTRODUCTION
Occupant presence and behavior in buildings has
been shown to have large impacts on space heating,
cooling and ventilation demand, energy consumption
of lighting and space appliances, and building
controls (Page, 2007). Several stochastic models
have been developed to model occupant presence and
interactions with space appliances and equipment.
Fritsch et al. (1990) proposed a model based on
Markov chains to model the random opening of
windows by occupants. Degelman (1999) developed
a Monte Carlo modeling approach for space
occupancy prediction based on survey statistics.
Reinhart et al. (2004) determined occupant presence
for lighting software by using a simplified stochastic
model of arrival and departure. Wang et al. (2005)
applied Poisson distributions to generate daily
occupancy profile in a single-occupied office.
Mahdavi et al. (2006) explored the possbilities of
identifying general patterns of user control behavior
as a function of indoor and outdoor environmental
parameters such as illuminance and irradiance.

Bourgeois et al. (2006) integrated an occupancy

model based on Reinharts algorithm into ESP-r to
investigate lighting use. However, most of the
previous occupancy presence models were either
tested on a single person office or presented a
specific application such as lighting control. Only
recently, Page et al. (2008) targeted individual
occupancy behavior by developing a generalized
stochastic model for the simulation of occupant
presence with derived probability distributions based
on Markov chains. However, the occupancy behavior
derived from stochastic model was based on the
assumption that occuapncy will interact with
different appliances in the space and the validation
was conducted in single-occupied offices. Most of
the previous works are based on supervised
approaches, which require ground truth occupancy
information. In addition, the latest models are all
based only on motion sensors, which often fail to
detect occupants that are sitting or standing still, and
thus have been shown in some cases to provide
insufficient accuracy for occupancy detection (Lam,
et al., 2008). Because occupants generate heat, water
vapour, CO2 and odors, a richer sensor enviroment
allows for a more robust means of detecting
occupancy presence and behaviour than motion
sensors alone.
Recently, there has been extensive research focusing
on developing smart environments by integrating
data mining techniques into environments that are
equipped with rich sensors and actuators. A smart
environment is defined as an environment able to
acquire and apply knowledge about the resident and
the physical surroundings to improve the residents
experience (Cook et al, 2004). Duong et al. (2006)
used hidden semi-Markov models for modeling and
detecting activities of daily living such as cooking,
eating, etc., and Youngblood et al. (2007) introduced
a new method of automatically constructing
hierarchical hidden Markov models using the output
of a sequential data mining algorithm to control a
smart environment. Other work investigates HVAC
preconditioning and device automation via mined
location and device interaction patterns, and the
energy savings potential is estimated through a
relatively simple consumption model (Roy et al.
2007). The objective of this study is to develop and
simulate unsupervised algorithms for ambient sensor-

- 1444 -

Episode Discovery
Minimum Description
Length

Raw
Sensor
Data

Event
Detector

Candidate
Episode
Generator

Energy Weight
Factor

Occupancy
Patterns

Period
Detection

Hidden Semi
Markov Model

Energy & Comfort

Management

Figure 1 A holistic view of this study

OCCUPANCY PATTERN
RECOGNITION
Event Detector
We first discuss the detection of events from a
variety of different sensors. We denote each single
event with a symbol and an episode as a sequence of
symbols. Table 1 shows symbol assignments; an
example of an episode is agghkjhk. A detailed
explanation of the event definitions for each sensor is
discussed below. All parameter values used in the
definitions are determined empirically for the testing
environment used in this work; however, variations
in these values are possible while still producing
meaningful sensor events.
Table 1
Definition of important events from sensors
State
Transitions

Acoustics 1. Low acoust.

2. Loud acoust.
Illumination 1. Off-On

Motion

Sensors

CO2

State
Transitions
1. Increasing

b
2. Decreasing
c Temperature 1. Increasing

2. On-off

1. Off-on
(motion)

2. On-off
(no motion)

2. Decreasing
Relative 1. Increasing
Humidity
2. Decreasing

Code

Sensors

Code

based modelling and prediction of user behaviour

within the context of intelligent buildings and
connect the derived user behavioural patterns to
building energy and comfort management. We base
our approach on the work of Youngblood et al.
(2007) in that a behavioral pattern model is
constructed by mining sensor events for significant
patterns (Episode Discovery), and then a Markov
model is generated from the resulting patterns. Our
contributions lie first in the integration of a rich
environmental sensor network using acoustics,
temperature, relative humidity etc. into the datadriven model of occupancy behavioural patterns.
Sensor event definitions account for significant
behavioral changes and energy events, and the
resulting (semi-)Markov models incorporate duration
to capture behavioral transitions over larger time
scales. Second, we develop a formal method to
connect the discovered patterns with energy and
comfort management in buildings and demonstrate
through simulation the energy-savings potential on
real data from a conference room in an office
building. In particular, we propose a dynamic
occupancy schedule for use in both EnergyPlus
(Crawley, 1999) simulations as well as more
commonly used energy management strategies, thus
providing a first step to truly integrating smart
building concepts into the building management
community.
An overview of the approach used in this study is
shown in Figure 1. The outline of the paper is as
follows. Section II describes sensor event detection.
Section III presents the approach for frequent pattern
detection using Episode Discovery, minimum
description length (MDL), period detection (PD) and
energy weight factors. Section IV introduces a semiMarkov model for occupancy duration models.
Section V discusses the connection of these frequent
patterns with building and energy comfort
management, and Section VI presents conclusions
and possibilities for future work.

g
h
i
j
k
l

a. Acoustics
The acoustics sensor outputs a calibrated percentage
of the acoustics level in the space. Figure 2 shows an
example acoustics profile for a typical day in a
conference room. The acoustic events are categorized
into two types: (1) ventilation noise or background
noise, defined as an acoustics level between 15% and
20% that is accompanied by at least a 5% increase
from the previous level (event a); (2) human
activity (e.g., voice or door opening/closing), defined
as an acoustics level above 20% accompanied by at

- 1445 -

least a 10% increase from the previous level (event

b). A smoothing method based on a root mean
square approach is implemented to reduce noise
(Smaton and McHugh, 2006).
100

Acoustic data
Acoustic events

90
80

Acoustics level (%)

70
60
50
40
30
20
10
0
00:00

06:00

12:00

18:00

00:00

Figure 2 One day example of acoustic events

b. Lighting
Lighting events are defined as: (1) light turned on
(event c); (2) light turned off (event d).
c. Motion
Motion sensor events are defined in the obvious way
for a binary motion sensor with an event each for
motion switching (1) on (event e) and (2) off (event
f). However, to avoid capturing high frequency
fluctuations that occur naturally when occupants are
inside the room and to obtain a more informative
signal, a 10 minute time window is used to smooth
the signal. A motion off event must be followed by
no motion activity within this window. Figure 3
shows an example motion profile and the
accompanying events.
2

Sensor output (normalized)

Motion data
Motion events

0
06:00

09:00

12:00

15:00

18:00

Figure 3 Example motion events

d. CO2
According to the results from Lam et al. (2008), an
increase of 50 ppm CO2 level in 10 minutes is found
to have high correlation with human presence. This,
however, clearly depends on the location of the
sensor; in this study, the CO2 sensor is located above
the conference table in the center of the room at
roughly nose level. The events are then defined as:
(1) CO2 increase of 50 ppm in a 10 minute time
window (event g); (2) CO2 decrease of 50 ppm in
10 minutes (event h).

e. Temperature and relative humidity

In a room without any windows, as is the conference
room test-bed, individual human-based temperature
fluctuations are minimal or on vary slow time scales.
Large changes in temperature (1 C) in a short time
frame (10 minutes) are more likely associated with
high energy activities such as large group presence,
the HVAC system being turned, or a projector.
Hence, the events for temperature are defined as (1) 1
C increase in 10 minutes (event i); (2) 1 C
decrease in 10 minutes (event j). Relative humidity
fluctuates very little under the test-bed conditions
unless there are occupants inside the space or the
HVAC system brings in outside air. Hence RH
events are defined as: (1) 10% increase in 10 minutes
(event k); (2) 10% decrease in 10 minutes (event
l).
Episode Discovery
Episode Discovery (Heierman et al., 2004) is the
process of discovering significant patterns in the data
sequence by first generating candidate sequences and
then pruning this set to obtain a final set of important
sequences. Time series sensor event sequences
generated according to our definitions in the previous
section are mined for potentially significant
candidate episodes using a sliding time window.
Briefly, in every episode window, the event codes are
ordered according to the time of occurrence. If the
codes happen at the exact same time, they are
ordered by alphabetical order for consistency. For
each episode window, all possible subsets of the
episode are generated. The generation of these
subsets as additional candidates accounts for
fluctuations in event order or the occurrence of
spurious events. For example, if the episode pattern
in a 3 minutes time window is {c,e,f,g,d}, then the
candidate episode patterns are {null}, {c,e,f}, {g,d}
and so on. However, to make this problem more
tractable and avoid considering the superset of the
episode as candidates, subsets are pruned using the
following rule (Heierman et al., 2004). The subset
candidates of a candidate episode that have the same
episode occurrences as the parent episode do not
need to be generated as candidates. An example
resulting candidate episode is cef, which, for our
event definitions, corresponds to light on followed
by motion on and motion off and is most likely
representative of someone entering a room.
After candidate episodes are generated, significant
episodes to be included in the behavioural model are
determined using the minimum description length
(MDL) criteria and periodicity (PD) as described
below. Additionally, because our focus is on energy
consuming behaviour, we use a weighting factor in
both the MDL and PD steps to increase the
importance of episodes containing high energy
impact events, namely lighting, temperature, and
humidity events.

- 1446 -

1. Minimum description length

The intent of MDL is to discover event patterns that
best represent the original input stream. Event
patterns may be thought of as a code table for
encoding the original input sequence. The optimal
code table is the one that minimizes both the size of
the code table plus the length of the encoded original
sequence. A brief algorithm is shown below (for a
detailed algorithm, see Bathoorn, 2006):
Let candidate episodes ={P1, P2,.. Pn}, where Pn is
the nth episode.
1. Ordering according to
a. Length; b.Frequency
2. Compress ()
CodeTable = allSinglePatterns;
minSize = computeSize(CodeTable)
for each Pi
CodeTable.add(Pi)
newSize = computeSize(CodeTable)
if newSize<minSize
minSize = newSize;
else
CodeTable.remove(Pi)
return CodeTable

2. Periodicity detection
Often, behaviors with the most utility for smart
building or building automation systems are those
that exhibit some periodicity. In a time series data
set Dorg, a symbol s or an episode p is said to be
periodic with a period l, if s or p exists every l time
steps. We compute episode periodicity using a
convolution-based approach, where the time series is
shifted l positions and the shifted series Dnew is
compared with Dorg (Mohamed et al., 2005). This
amounts to conducting a frequency spectrum analysis
using a Fourier transform. A detailed algorithm can
be found in Elfeky et al. (2005).
Semi-Markov model generation
One of the most important inputs in designing an
optimal room controller is the duration of occupancy
in the room. To this end, we investigate an
occupancy duration model from the discovered event
patterns. Specifically, we employ a semi-Markov
model that allows for duration in each state before
transitioning to the next state (Murphy, 2002). Duong
et al. (2006) applied HSMMs for pattern recognition
of daily human activities. In this study, as in
Youngblood et al. (2007), we treat each discovered
important pattern as a state in the Markov model. We
learn the semi-Markov model using a forwardbackward algorithm (Yu and Kobayashi 2003). Note
that in our current approach, states are not considered
hidden and thus the typical HSMM framework is not
needed. Hence, the model parameter estimation
algorithm is greatly simplified.

EXPERIMENT IN A CONFERENCE
ROOM
Sensor data collection
Six different types of wireless and wired sensors are
installed in a conference room of a commercial
building in Pittsburgh. Data is collected every one
minute from May 1st to August 31th, 2008. Figure 4
shows a picture of the conference room and its
installed sensors.

Figure 4 Test-bed in a conference room, wireless

mote and occupancy counting input device
Results of event detection
Figure 5 illustrates an example day of sensor events
generated according to the definitions described
earlier. Event numbers on the y axis indicate which
event occurred for the given sensor according to the
codes in Table 1. For example, at 5:40am, the
temperature decrease event (Event_2 for the
temperature decrease event) occurred when the air
conditioning system turned on. As is typical with
most days in the conference room, numerous motion
and acoustic events occur from 10:00am to 11:00am
when the room is active with meetings. At 11:00 pm,
a custodian enters the room, generating lighting and
acoustics (vacuum cleaner) events.
Events Generator

Acoustics
Illumination
Motion
CO2
Temp
RH

Event_2

Event_1

0
00:00

03:00

06:00

09:00

12:00
Time

15:00

18:00

21:00

00:00

Figure 5 A one day example results of event

detection
Results of MDL and PD
Based on a time window of 10 minutes, a summary
of important patterns resulting from the MDL and PD
selection criteria are shown in Table 2. It is noted
here that the MDL component discards some very
long patterns due to highly infrequent occurrence
(once every week or every few days). The final set of

- 1447 -

important patterns are those resulting from both

MDL and PD.
Table 2
Results of Patterns from MDL and PD

#of
Patterns

MDL

PD

Longest
Pattern

ebbfe(24)

dfcedf,
bebdf,
ebbdf,fefe,
aa,ghg,gge
bgfb,feg,
hbe,aec,
fhd

bg(84)

Results of semi-Markov model

The exponential family of distribution functions were
found to best model the durations associated with the
discovered patterns. This is consistent with other
work in speech recognition (Russell, 1985) and
occupancy of single-person offices (Wang et al.,
2005). Figures 6 and 7 show the resulting semiMarkov model of important patterns. Event code
letters are as defined in Table 1. Figure 6 shows a
standard Markov model with numbers on the arcs
indicating the transition probability between states,
Transitions with relatively low probabilities (less
than 15%) are not shown. Parentheses indicate
number of occurrences of the pattern in the training
period. As an example, state ecf has a 25%
transition probability to state eb and a 24%
probability to state def, with ecf occurring 22
times, eb 37 times and def 15 times during the
month. Figure 7 shows the results of including
duration in the model. Each duration distribution is
denoted as X~(time), where time is the expected
duration for the exponential model. For example,
ecf has an expected duration of 30 minutes before
it transitions to state eb and 10 minutes before
transiting to state def. The red-dotted line indicates
a typical 75 minute meeting scenario where an
occupant enters the room, triggers the motion sensor
e, turns on the light c, and sits down, triggering
the motion off f. The occupant continues to stay in
the room, generating acoustics b and moving
around generating motion e. Upon leaving, the
occupant turns off the light e, moves towards the
door e and finally departs f. Another possible
duration path is on average 138 minutes, representing
a longer meeting.
0.2
0.2
0.25

0.25

hdf (14)

0.2
0.5

ecf (22)

eb (37)

he (32)

bgi (12)
0.4

bfe (20)

def (15)

0.2
0.24

Figure 6 Markov model of discovered patterns on 10

minutes maximal window

0.2

0.2

X~(30)
0.25

ecf (22)

eb (37)

X~(30)

X~(48)

0.25

0.2

0.4

hdf (14)

X~(30)
0.5

he (32)

bgi (12)
X~(10)

Other
Patterns

Most
Compressed
Pattern
cedf(19)

bebdf(22)

X~(45)
X~(45)

def (15)

bfe (20)

X~(10)
X~(10)

0.2

0.24

Figure 7 Semi-Markov model of discovered patterns

on 10 minutes maximal window
Additional models representing longer time scales
may be generated by considering a pattern such as
ecf as a new event G and repeating the pattern
discovery process (Youngblood et al., 2007). Results
are shown in Figures 8 to 9 for the resulting model of
this approach using a maximal window of two hours.
0.25(7)
0.5
G (13)

0.5
AB (9)

0.5
FF (7)

0.3

AAB (7)

I (10)

H (11)
0.15

0.5

D (12)

0.5
A: ebstay B: hestay G: ecfenter E: bfestay K: gedleave
D: hdf leave I:defleave H: agstay F: hj HVAC

Figure 8 Markov model of discovery patterns on

patterns
X~(30)
0.25

X~(60)

X~(30)

0.5

G (13)

X~(15)

0.5

AB (9)

X~(10)

FF (7)

X~(20)

0.3

0.5

AAB (7)

I (10)

H (11)

D (12)

X~(45)
0.5

X~(120)

Figure 9 Semi-Markov model of discovery patterns

on patterns

ENERGY AND COMFORT

MANAGEMENT
A dynamic occupancy schedule with expected
durations was developed from the behavioural
pattern recognition results. This dynamic schedule, as
described below, can be connected with a building
energy and comfort management system (BECMS)
through dynamic real-time temperature and
ventilation set point inputs. The BECMS can then
make decisions according to the dynamic schedule.
In order to test the practicality of this approach, we
coupled the dynamic schedule with EnegyPlus, a
widely used energy simulation tool.
There are several current approaches in the literature
for modelling occupancy within the context of energy
simulation. Claridge et al. (2001) suggested that
occupancy diversity profiles might be derived from
lighting diversity profiles through establishing a
strong correlation between observed occupancy
levels. However, other studies suggested diversity
profiles generate misleading information when
occupancy-sensing lighting controls are used
(Degelman ,1999). Bourgeois et al. (2006) developed
a sub-hourly occupancy-based control (SHOCC)
coupled with the ESP-r simulation program. SHOCC

- 1448 -

tracks individual instances of occupants and

occupancy-controlled objects such as blinds.
However, its application is limited with lighting
controls.
In this study, the dynamic schedule was used toward
lighting and HVAC controls. The control strategy
utilizes the learned Markov model of behaviour and
takes advantage of the fact that some patterns such as
ecfdef only last briefly, corresponding to commonly
found scenarios where users step into the conference
room to, for example, make a cell phone call. In
situations such as this, the HVAC system does not
need to meet the temperature set point and ventilation
rate. Figure 10 illustrates the coupling of an HVAC
control strategy with occupancy pattern recognition.
The term dynamic schedule refers to the time and
state-dependent use of the Markov model in the
HVAC and lighting control strategy. The system
monitors sensor events to determine the current state
of the environment as given by the Markov model. If
an entry state (e.g., one involving lights turning on) is
identified, the system computes the most probable
duration of occupancy based on the model and
responds accordingly.
The control strategy is
updated as the detected state changes. Because our
emphasis here is on illustrating the utility of datadriven
behavioural
modelling
for
energy
management rather than on controller design, we
implement a simple occupancy-dependent on/off
control; however, more advanced controllers can
achieve better performance by utilizing the duration
information contained in the model. For our
simulation, a software link between the dynamic
schedule and EnergyPlus is used so that the time
dependent schedule can be generated automatically
from pattern recognition.

HVAC
Controller
No Action
Short period
Optimal
Control
Long period

Space

Sensor
Network

Raw
Data

Occupancy
Behaviour
Pattern
Recognition

Figure10 HVAC controls based on pattern

recognition
Comparison among current set point strategies
In order to evaluate the energy saving effects and
thermal comfort conditions based on dynamic
scheduling strategies from the occupancy behavioural
patterns, the energy usage of four different set point
strategies are compared. These four possible HVAC
set point schedules, and their advantages and
disadvantages are:
1. Fixed system schedule set point at 24 C from
7:00am to 6:00pm.

Advantage: simplicity for facility manager

Disadvantage: High Energy Cost. No need to
maintain 24 C when there are no people present
2. Outlook schedule based on company outlook
(Barney and Lynne, 2007)
Advantage: exact meeting schedule and possible
meeting duration
Disadvantage:
Many
meetings
occur
spontaneously with no pre-scheduling in Outlook
3. Occupancy (Motion) sensor based
Advantage: Simplicity
Disadvantage: No motion occurs if occupants are
relatively still in the room. Also, motion is
triggered if an occupant enters the room in the
middle of the meeting, generating spurious
events.
4. Dynamic occupancy schedule
Advantage: Dynamic temperature set point; An
explicit meeting duration model; Automatic
lighting control when zero occupancy; Save
energy and maintain comfort
Disadvantage: Need for additional sensors
All schedules have a night setback temperature of 30
C, and, aside from the fixed-point schedule, all have
a daily setback of 27 C at 7:00 am. A lower
temperature setpoint of 24 C during the day is set as
determined by the approach when the room is
considered occupied.
EnergyPlus simulations with three zones were
conducted: a simple conference zone of size 3m by
6m faces east, a Resistive zone before the
conference zone, and a North zone. We focus on
evaluating controller performance in the conference
zone (the other zones are kept at fixed standard
operation schedules). The building simulation is
conducted from June 1st to August 31st, 2008, with
TMY-3 Pittsburgh weather data, and the predicted
occupancy profile used for the controller is based on
training data from May 1st to May 30th, 2008. The
true occupancy profile used for the simulation is
taken from an occupancy counter box (see Figure
4) deployed in the conference room that allows
occupants to keep track of the number of people in
the room at all times by pushing up or down buttons.
Table 3 shows the results from EnergyPlus in terms
of total building loads for the three months.
Table 3
Building loads and comfort based on different HVAC
set point schedules in the conference room
Total Cooling
Loads (kWh)
Total Lighting
(kWh)
Total (kWh)
Time Comfort
Not Met
(ASHRAE-55)
(hour/day)

- 1449 -

Fixed

Outlook

Motion

Dynamic

5483

4050

3794

3833

1150

880

872

872

6633

4930

4666

4705

0.63

3.26

2.38

Table 3 shows that while the fixed schedule achieves

very good comfort conditions (with very little time
when comfort is not met), it is very energy
inefficient. The Outlook schedule does not perform
well because meetings are often either shorter than
scheduled or even cancelled, leaving the HVAC
system running with no one present. The largest total
savings are from the motionbased approach;
however, this comes with a sacrifice in occupant
comfort because of times when occupants are present
with little or no motion, causing the HVAC system to
revert to the higher, less comfortable daily setback
temperature. The dynamic schedule, which is derived
from the data-driven pattern model, achieves energy
savings comparable to that of the motion-based
approach, but with a less amount of time when
comfort is not achieved. The one hour per day of
uncomfort arises mostly from short visits to the
meeting room (approximately 10 minutes) that are
not worth the cooling effort.
Figure 11 shows a daily indoor temperature profile
from these four different set point schedules. The
outlook schedule for the given day is:
9:15am~10:30am and 1:45pm~3:30pm. All three
non-fixed set point schedules reach the daily setback
temperature at 7:00am as scheduled. Beginning at
7:00am, the temperature profiles behave differently
according to the different set point strategies.
Interestingly, during lunch time, the motioned-based
schedule still tried to meet the set point despite only
short visits to the conference room during that time.

CO2, temperature, and relative humidity are

incorporated into an event-based pattern detection
algorithm used for modelling occupant behaviour
toward HVAC system control. Our aim is for an
unsupervised approach that does not require
extensive training or modelling of the environment at
hand. The pattern discovery and model generation
approaches are based on the work of Youngblood et.
al (2007) with extensions for integrating ambient
sensor events with a focus on energy-related
activities and the use of semi-Markov models that
allow for pattern or state durations. Additionally, we
illustrate a connection of the learned behavioural
model with energy control systems through the
generation of a dynamic occupancy schedule. Such a
dynamic schedule was generated from a conference
room environment equipped with a wireless sensor
network and tested as an input to an HVAC control
system in an EnergyPlus simulation. Compared with
other alternative occupancy-based control strategies,
the results of the dynamic schedule show significant
energy savings with minimal comfort sacrifice.
Possibilities for future work include investigations
into data mining techniques for increased
computational efficiency in generating learned
patterns as well as the use of the Markov model for
prediction-based control of energy management
systems. Additionally, future studies may investigate
more advanced control designs for exploiting learned
behaviour for different HVAC systems (e.g., VAV).

ACKNOWLEDGMENT

CONCLUSION
In this study, we demonstrate through simulation the
energy-saving utility of using a data-driven model of
occupant behaviour for energy management.
Ambient sensing data such as lighting, acoustics,

This work was done during the course of an

internship at Robert Bosch, LLC, Research and
Technology Center, Pittsburgh, PA, U.S.A.

Figure 11 Temperature profile on Summer Design Day (July 21) based on different set points

- 1450 -

REFERENCES
Bathoorn, R., Koopman, A. and Siebes. 2006. A.
Frequent Patterns that Compress, Technical
Report, Utrecht University.
Bourgeois D., Reinhart C. and Macdonald I. 2006.
Adding Advanced Behavioural Models in Whole
Building Energy Simulation: A Study On the
Total Energy Impact of Manual and Automated
Lighting Control. Energy and Buildings Vol. 38,
pp.814-823
Barney, L.C. and Lynne, C.C. 2007.Web-based
Enterprise Energy and Building Automation
Systems. Fairmont Press, Inc.
Cook, D. and Das, S. 2004. Smart Environments:
Technology, Protocols and Applications. Wiley.
Crawley, D.B., Lawrie, L.K. Pedersen, C.O., Liesen ,
R.J., Fisher, D.E. Strand, R.K. Taylor, R.D.
Winkelmann, R.C. Buhl, W.F. . Huang, Y.J.
Erdem, A.E. 1999. ENERGYPLUS, A NewGeneration
Building
Energy
Simulation
Program, Proceedings of Building Simulation
99, an IBPSA Conference, Kyoto, Japan, pp. 8188.
Degelman L.O. 1999. A Model for Simulation of
Daylighting and Occupancy Sensors as an
Energy Control Strategy for Office Buildings, in:
Proceedings of Building Simulation 99, an
IBPSA Conference, Kyoto, Japan, pp. 571-578.
Duong, T.V. Phung, D.Q. Bui, H.H. and Venkatesh,
S. Human. 2006. Behavior Recognition with
Generic Exponential Family Duration Modeling
in the Hidden Semi-Markov Model. In
Proceedings of the 18th international
Conference on Pattern Recognition. Vol. 3.
Elfeky, M. G. and Lemagarmid, A.K. 2005.
Periodicity Detection in Time Series Databases.
IEEE Trans. on Knowl. and Data Eng. Vol.17,
875-887
Fritsch R., Kohler A., Nygard-Ferguson M. and
Scartezzini J.-L. 1990. A Stochastic Model of
User Behaviour Regarding Ventilation. Building
and Environment 25 (2): 173-181.
Hagras, H., Callaghan, V., Colley, M., and Clarke, G.
A. Pounds-Cornish, and H. Duman. 2004.
Creating an Ambient-Intelligence Environment
Using Embedded Agents. IEEE Intelligent
Systems 19, 6, 2006.
Heierman, H., Youngblood, G.M. and Cook, D.
2004. Mining Temporal Sequences to Discovery
Interesting Patterns, KDD Workshop on Mining
Temporal and Sequential Data.
Lam, K.P. Hynck, M. Dong, B., Andrews, B. Chiou,
Y.S. Zhang, R. and Benitez, D. 2009. Occupancy
Detection through An Extensive Environmental

Sensor Network in an Open-plan Office

Building. Proceedings of Building Simulation
2009, an IBPSA Conference, Glasgow, U.K.
Mahdavi, A. Kabir, E. Mohammadi, A. Lambeva, L.
and Proglhof, C. 2006. User Interactions with
Environmental Control Systems in Buildings.
Proceedings of
PLEA
2006. Geneva,
Switzerland.
Murphy, K. 2002. Hidden Semi-Markov Models.
Technical Report. University of British
Columbia.
Smaton, A. and McHugh, M. 2006. Event Detection
in an Audio-based Sensor Network. Multimedia
Systems. Vol. 12 pp179-194.
Page J., Robinson D., Morel N. and Scartezzini J.-L.
2007. A Generalised Stochastic Model for the
Simulation of Occupant Presence. Energy and
Buildings. Vol. 40 (2). pp. 83-98.
Reinhart C. 2004. Lightswitch-2002: A Model for
Manual and Automated Control of Electric
Lighting and Blinds. Solar Energy. Vol. 77, pp.
15-28.
Russell, M. J. and Moore, R. K. 1985. Explicit
modelling of state occupancy in hidden Markov
models for automatic speech recognition. In
Proceedings of IEEE Conference on Acoustics
Speech and Signal Processing, pp. 58.
Roy, A., Das, S.K. and Basu, K. 2007. A Predictive
Framework for Location-Aware Resource
Management
in
Smart
Homes.
IEEE
Transactions on mobile computing, Vol. 6 (11).
Wang D., Federspiel C. and Rubinstein F. 2005.
Modeling occupancy in single person offices.
Energy and Buildings 37: 121-126.
Youngblood, G.M. and Cook, D. 2007. Data Mining
for Hierarchical Model Creation. IEEE
Transactions on system, man and cybernetics.
Vol. 37, No.4.
Yu, S.Z. and Kobayashi. 2003 .H. An Efficient
Forward-Backward Algorithm for an Explicitduration HMM, IEEE Signal Processing.

- 1451 -