Professional Documents
Culture Documents
by users
David Lindelöf∗ and Nicolas Morel
Solar Energy and Building Physics Laboratory (LESO-PB)
EPFL, CH-1015 Lausanne, Switzerland
July 15, 2007
This paper describes how data collected during a continuously running data
acquisition program on the LESO building in Lausanne, Switzerland, was used
to measure the intermediate light switch probability by users as a function of
current illuminance levels, i.e. the probability for a given timestep that the user
will switch on or off the electric lighting, excluding such actions that happen
upon user entry to or exit from the office. We assume such a probability
to be independent of the user’s history and further derive some theoretical
consequences of this postulate. In particular, we show how a history-less user
leads naturally to patterns of behaviour already observed in real buildings.
1 Introduction
Understanding the way users interact with building services (blinds, electric lighting, cool-
ing, ventilation, window opening, etc), and the impact of their use on the building’s total
energy consumption, helps us attain two goals. First, we may elaborate better models of
the user’s behaviour for simulation software that will help building planners to predict and
optimize the energy use or the comfort provided to the user. Secondly, advanced control
algorithms may use this information to increase their acceptance by users and help achieve
energy savings.
Several building simulation software packages that need a good simulation of user’s
bahaviour are available (1; 2) or will shortly be (3; 4). The software packages just mentioned
are all based on the Lightswitch-2002 algorithm described in (2). An underlying assumption
∗
Corresponding author. Tel.: +41 21 693.55.56; fax: +41 21 693.27.22; email: david.lindelof@epfl.ch
1
behind this algorithm is that users use the manual controls at their disposal in a conscious
and consistent way, which allows us to predictively model their behaviour.
The algorithm seeks also to model the intermediate light switch-on probability, i. e. the
probability that a user switches on the artificial lighting without leaving or arriving in the
office. It uses a probability function that depends on the workplane illuminance, derived
from previous work by the author of the algorithm (5). For five-minute timesteps, it finds
that the intermediate switch-on probability is about 5% between 0 and 200 lux workplane
illuminance, and sharply drops to about 0.002 for higher illuminances. One purpose of this
paper is to verify this model.
More specifically, we will focus on the following themes concerning lighting actions (day-
light or electric lighting):
The lapse of time between the entry of the user into the room and the use of controls,
or between use and subsequent exit from the room;
The correlation between the delay before the user action and the illuminance level.
We will also discuss some theoretical consequences of the modelling of users’ actions and
their relationship to experimentally observed data.
2 User simulation
Some models of user behaviour assume that the time between user actions, given constant
environmental conditions, is a random variable distributed according to an exponential
distribution with sole parameter λ satisfying λ = 1/T where T is the average time before
the action. In other words, its probability density function is given by
fλ (t) = λ exp(−λt).
This distribution is believed to hold, with different λ parameters of course, for most of
the user’s actions, such as use of artificial lighting controls, window opening or closing, and
exit or arrival.
The exponential distribution function is used for modelling the occurence of events rang-
ing from earthquakes to phone calls. Similarly, the time remaining until, for instance, the
user’s next opening of windows can be modelled in much the same way as the time remain-
ing until the user’s next phone call.
This postulate is justified by strong evidence that the number of user actions of a given
kind for constant or near-constant environmental conditions in a given time frame follows
2
a Poisson distribution. From that fact follows that the intervals between events are dis-
tributed according to an exponential distribution. In the case of user entry and exit, Wang
(6), for example, has recently verified experimentally that the duration of user absences
from a room indeed follows an exponential distribution.
The problem is that environmental conditions are seldom constant in an office. Temper-
atures, air quality and illuminance levels change over time. How are we to compute the
probability density function for user events under varying conditions, assuming we know
how to do so for constant conditions?
Consider the example of the use of artificial lighting controls. If we could find a relation-
ship between the time a user tolerates given visual conditions and the variables describing
these visual conditions (e.g., the illuminance levels), then we would be in an advantageous
position to simulate the behaviour of the user in a computer simulation of the building.
This section will describe how.
Let us assume that such a relationship exists between the average time T before user
action and the illuminance E, and that we have found it. We do not specify this rela-
tionship; we are in no position to do so yet. But we assume that the probability density
function of the time the user spends before switching on the lights follows the exponential
distribution, and that its λ parameter is given by some function of E.
We begin at time t = 0. What is now the probability P (T ) that the user has not turned
on the lights yet at a time t = T ? Let us discretize the time between t = 0 and t = T
into n equal timesteps ∆t = T /n. Let us assume that these timesteps are sufficiently small
that the workplane illuminance can be taken as constant during each timestep, noted Ei ,
with i running from 0 to n − 1. The corresponding λ parameters are noted λi .
The probability that the user did not switch on his lights during the first timestep is
high because ∆t is small, but not quite equal to unity. It is given by
Z ∆t
P (∆t) = 1 − fλ0 (t)dt
0
= exp(−∆tλ0 )
By the exponential distribution’s lack of memory, the probability that the user did not
switch on the lights between ∆t and 2∆t knowing that he did not do so between 0 and ∆t
is similarly equal to exp(−∆tλ1 ), and so on.
The probability that the user has not switched on his lights by the time T is thus the
product of all these probabilities. We obtain
n−1
Y
P (T ) = exp(−λi ∆t)
i=0
n−1
X
= exp(−∆t λi )
i=0
= exp(−T λ),
3
Pn−1
where λ = i=0 λi /n is the average of all λi .
But if p(t) denotes the probability distribution function of the time at which the user
switches on the lights for non-constant environmental conditions, then the following must
hold: Z ∞
p(t)dt = P (T ),
T
since the left-hand term is the probability that the user switches on the lights at a time
between T and ∞ and the right-hand term is the probability that the user did not switch
the lights on between 0 and T . They are, of course, the same thing.
Replacing with the value for P (T ) found previously, and deriving with respect to T on
both sides, we obtain
p(t) = λ exp(−λt)
Note that along the way we lost any reference to ∆t, so this formula holds for ∆t van-
ishingly small. In fact, it does not hold if ∆t is too large for the environmental conditions
to be considered as constant.
We thus have an expression for the distribution function of the time of user action for
varying environmental conditions. It is however difficult to use this expression in practice
since it cannot readily be integrated over time, due to the non-constant λ parameter.
A computer running a building simulation should therefore rather compute the evolution
of environmental conditions in steps of ∆t, and compute at each timestep P (∆t) with
parameter λ = λi where i indexes the timestep.
The question remains, how does one measure this λ parameter? One cannot put a
user in his office, bid the sun and the clouds not to move too much for the next hour
and wait until the user gets up and switches on his lights, repeat N times, and derive
the average waiting time. Rather, since a process obeying an exponential distribution
implicitly assumes that the probability that the user should use his controls within the
next few minutes is independent of the user’s history, one should rather measure for a
given illuminance level what is the probability π that the user should use his controls in a
given time window ∆t. Then the average waiting time between two user actions, if ∆t is
small enough, is 1/λ = ∆t/π.
This probability p can be obtained over the course of a measuring campaign by counting
how many times the user finds himself in a given illuminance, and how many times that
illuminance leads to a user action.
3 Methodology
The data representing the base material for the study discussed in this paper is part of a
continuous recording program by LESO-PB (Solar Energy and Building Physics Labora-
tory, EPFL, Lausanne, Switzerland), on the LESO building. The LESO building is a small
office building (20 office rooms of around 20 m2 floor area each, about half of them with a
single user and the other half with two), hosting the activities of LESO-PB (for a detailed
4
description of the building, see (7)). Additionally, this inhabited building has been used
for the experimentation of new passive solar systems and advanced control algorithms for
building services (heating, blinds and electric lighting). It is equipped with a commercial
EIB building bus system, with the following sensors and actuators for each room:
Sensors: inside and outside air temperatures, diffuse and global solar radiation, solar illu-
minance, wind speed and direction, interior illuminance1 , occupancy, window open-
ing;
Finally, the building is also equipped with a legacy central data logger used for recording
miscellaneous data such as room temperatures at different positions (measuring, in effect,
the stratification of the temperature) or the electricity consumption for each room (counting
separately the heating and the other appliances).
The EIB was installed in 2000. Since then, two research projects focused on control
systems have been carried out at LESO-PB that made use of mesurements on the LESO
building, AdControl (8; 9; 10) and Ecco-Build, for which no publication is available yet.
The data acquisition has been running independently of any research project and been
stored to a MySQL database (11), representing slightly more than two years of continuous
monitoring. For data due to user actions alone, we have accumulated at the time of writing
about three million datapoints.
The data we consider covers the period from mid-November 2002 to mid-January 2005.
Any time one of the physical variables changes by more than an adjustable threshold, that
change is logged to the database together with the time of the event. It is trivial to use
this data to reconstruct a time series of any variable for any given constant timestep.
The data was analyzed with the open-source R data analysis environment (12).
4 Data analysis
4.1 Definition of a user action
In the following a user action is defined as a set of interventions on individual controls
available to the user not more than one minute apart from each other. For example, a
user might come in in the morning, switch on the lights, and open the blinds. During the
day, he or she might decide that the sunlight is enough to illuminate the office so he or
she switches off the light but lowers the blinds sufficiently to prevent direct sunlight from
1
We use Siemens brightness sensors GE 252, which are actually roof-mounted luminance sensors shielded
from the window’s luminance. They were calibrated by Guillemin (8) with LMT reference luxmeters
with an estimated accuracy of ±1.9%. The conversion from the workspace’s luminance to its illuminance
is a programmable feature of the sensor.
5
hitting the computer screen. These would be considered two user actions. As long as the
individual interventions are not spaced more than one minute apart, they are considered
the same user action.
The data recorded on the building bus is used each night to rebuild a table with all
user actions since data acquisition began. Each entry of that table records the time of
the beginning of the action, the time of its end, as well as the time when the user came
into the office before the action occured and the time at which the user left the office
afterwards. We also record which control(s), among blinds, artificial lighting, or windows,
were affected.
2
I.e., the values below which we have respectively 25% and 75% of the total number of events
6
1000
800
600
[s]
400
200
0
001 002 003 004 101 102 103 104 105 106 201 202 203 204
Office
Figure 1: Boxplots of the times between user entry and his use of manual controls, per
office.
about 75% of all user actions happen about five minutes before the user’s departure. In
other words, again, users use their controls mostly when coincidentally passing by.
This has important consequences for the choice of placement of the user’s controls in
an office, which should be as close and as convenient as possible to the user, who will
otherwise simply not use them. This observation also highlights again the obvious need for
smart building control systems since users, unless particularly energy-conscious, usually
will not adjust their controls if the only benefit is the saving of energy. A clear discomfort
is required for the user to take action.
7
1000
800
600
[s]
400
200
0
001 002 003 004 101 102 103 104 105 106 201 202 203 204
Office
Figure 2: Boxplots of the time between user’s use of manual controls and his departure,
per office.
Their non-linear behaviour for illuminances above 500 lux has been corrected by the EIB
monitoring software.
That the artificial lighting should increase the amount of available workplane illuminance
is hardly surprising. Neither should one ascribe too much importance to the differences
in the distribution of illuminance after the action. Past certain hours, in particular those
hours where lighting is needed most, the lighting can only provide so many lux and we doubt
that the users in office 101 have deliberately and systematically fiddled with the dimming
controls to get to the shown average of about 300 lux. That value represents more likely a
rough estimate of the maximum workplane illuminance the lighting can provide.
More interesting are the disparities seen in the distribution of illuminance right before
the action. Some users (e.g. 104) never allow the illuminance to go below about 200 lux
before turning the lights on. Others (such as the people in office 001) seem less bothered
and tolerate even very low light levels before turning the lights on.
Only office 004 shows odd results, but the measurement of this office’s illuminance values
is known to be faulty. A new user moved in during 2003, as a result of which the main lumi-
naire (a lamp projecting its light on the ceiling) has been moved right under the luminance
sensor doubling as an illuminance sensor. This mistake has now have been corrected, but
the data taken on this office will be excluded from further analysis in this paper and will
not be included in plots obtained by pooling together all data. The histogram obtained by
lumping together all offices except office 004 are given in Fig. 5.
8
Illuminance office 001 Illuminance office 002 Illuminance office 003 Illuminance office 004
8
40
12
20
6
30
15
Counts
Counts
Counts
Counts
8
4
20
10
6
4
10
2
5
2
0
0
0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500
Illuminance office 101 Illuminance office 102 Illuminance office 103 Illuminance office 104
35
8
20
25
6
4
15
Counts
Counts
Counts
Counts
3
4
10
15
2
2
5
0 5
0
0
0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500
Illuminance office 105 Illuminance office 106 Illuminance office 201 Illuminance office 202
10
6
8 10
9
8
25
4
Counts
Counts
Counts
Counts
6
3
15
4
2
2
1
5
0
0
0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500
8 10
Counts
Counts
6
4
2
0
0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500
[lux] [lux]
Figure 3: Distribution of illuminance levels before user actions, per office. The histograms are first filled in black with
events where the workplane illuminance increased after the user action. On top of this the events where the
illuminance decreased or stayed constant are added in white.
Illuminance office 001 Illuminance office 002 Illuminance office 003 Illuminance office 004
6
40
15
5
6
30
4
Counts
Counts
Counts
Counts
10
3
20
2
5
10
1
0
0
0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500
Illuminance office 101 Illuminance office 102 Illuminance office 103 Illuminance office 104
0 1 2 3 4 5 6 7
35
15
25
Counts
Counts
Counts
Counts
10
15
2
5
0 5
0
0
0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500
Illuminance office 105 Illuminance office 106 Illuminance office 201 Illuminance office 202
10
30
6
8
8 10
5
6
20
4
Counts
Counts
Counts
Counts
3
4
5 10
2
2
1
2
0
0
0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500
10
8
8
Counts
Counts
6
6
4
4
2
2
0
0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500
[lux] [lux]
Figure 4: Distribution of illuminance levels after user actions, per office. See Fig. 3 for the explanation of the black and
white bars.
150
100
Counts
50
0
[lux]
50
0
[lux]
Figure 5: Distribution of illuminance, all data. See Fig. 3 for the explanation of the black
and white bars.
11
Fig. 6 shows a scatterplot for each office of the illuminance level after vs. before user
action. Data points beneath the diagonal represent events when the user found the light
too strong and decreased or turned it off, while points over the diagonal represent events
where the illuminance level was deemed insufficient. Fig. 7 groups together all the events
for all offices, with a small jitter applied to each point, in order to prevent the discrete
illuminance values provided by the measurement from hiding the real data point density.
Office 001 Office 002 Office 003 Office 004 Office 101
1000
1000
1000
1000
1000
800
800
800
800
800
600
600
600
600
600
[lux]
[lux]
[lux]
[lux]
[lux]
400
400
400
400
400
200
200
200
200
200
0
0
0 200 400 600 800 1000 0 200 400 600 800 1000 0 200 400 600 800 1000 0 200 400 600 800 1000 0 200 400 600 800 1000
Office 102 Office 103 Office 104 Office 105 Office 106
1000
1000
1000
1000
1000
800
800
800
800
800
600
600
600
600
600
[lux]
[lux]
[lux]
[lux]
[lux]
400
400
400
400
400
200
200
200
200
200
0
0
0 200 400 600 800 1000 0 200 400 600 800 1000 0 200 400 600 800 1000 0 200 400 600 800 1000 0 200 400 600 800 1000
1000
1000
1000
800
800
800
800
600
600
600
600
[lux]
[lux]
[lux]
[lux]
400
400
400
400
200
200
200
200
0
0 200 400 600 800 1000 0 200 400 600 800 1000 0 200 400 600 800 1000 0 200 400 600 800 1000
Figure 6: Illuminances recorded right after (Y-axis) vs right before (X-axis), per office.
There is a marked tendency for all offices to prefer switching on the lights to switching
them off, something coherent with our personal experience. People are, in general, more
concerned with their visual comfort than with unnecessary energy expenditures3 . Of all
offices, only the occupant of office 104 seems to switch off somewhat regularly the lights
when not necessary.
Note also that for most offices, but most notably in offices 001, 002, 101, 104, 106, 203
and 204, there is a clear clustering of the points along two diagonals running parallel to the
main diagonal, and equally distanced from that diagonal. This reflects the fact that users
usually don’t bother fiddling with the light dimming commands and content themselves
with switching the lights on or off, thus resulting in a constant increase or decrease of
available illuminance of roughly 300 lux, the maximum the current light installation can
3
or even forget they have left the lights on.
12
1000
800
Illuminance after action [lux]
600
400
200
0
13
provide. This shows that even when dimming commands are available, few users will make
use of them in an energetically optimal way when they are placed close to the office’s
entrance rather than close to the user4 .
The occupants of offices 001, 104 and 201 obviously use their electric lighting controls
much more than the other users. Since the placement of the electric lighting controls is
similar in all offices (close to the office’s entrance), one can only conclude that these users
are much more concerned with a rational use of the artificial lighting than the others.
This leads credence to the notion of active vs. passive users found in the Lightswitch-2002
model, which distinguishes users based on their willingness to use the controls at their
disposal.
From now on we shall assume that all user actions on the artificial lights are switch
on/off events and neglect the extremely rare dimming events.
4
A user even told one of the authors that she actually did not even know she could dim her light.
14
Office 001 Office 002 Office 003 Office 004
0.06
0.004
Switch−on probability
Switch−on probability
Switch−on probability
Switch−on probability
0.020
0.04
0.010
0.002
0.010
0.02
0.000
0.000
0.000
0.00
0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500
0.15
0.020
Switch−on probability
Switch−on probability
Switch−on probability
Switch−on probability
0.020
0.10
0.02
0.010
0.010
0.05
0.000
0.000
0.00
0.00
0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500
0.008
0.04
0.04
Switch−on probability
Switch−on probability
Switch−on probability
Switch−on probability
0.00 0.02 0.04 0.06
0.004
0.02
0.02
0.000
0.00
0.00
0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500
Switch−on probability
0.020
0.015
0.010
0.000
0.000
0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500
[lux] [lux]
Switch−off probability
Switch−off probability
Switch−off probability
0.004
0.008
0.002
0.002
0.004
0.000
0.000
0.000
0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500
Switch−off probability
Switch−off probability
Switch−off probability
0.012
0.0010
0.0010
0.0010
0.006
0.0000
0.0000
0.0000
0.000
0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500
0.004
Switch−off probability
Switch−off probability
Switch−off probability
Switch−off probability
0.004
0.010
0.0015
0.002
0.002
0.0000
0.000
0.000
0.000
0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500
0.006
Switch−off probability
Switch−off probability
0.004
0.003
0.000
0.000
0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500
[lux] [lux]
0.015
0.010
0.005
0.000
[lux]
0.0010
0.0005
0.0000
[lux]
17
4.4.1 Switch-on probability
The behaviour of users with respect to swich-on probability shows remarkable consistency.
Most users seem to have an illuminance threshold under which the switch-on probability
sharply rises to a level of between 1% and 10%. As long as the illuminance is above that
threshold, the switch-on probability is negligible. That threshold varies from user to user
but lies between 100 and 200 lux.
Fig. 10, obtained by lumping together the data from all office rooms except rooms 003
and 004 (which both seem to display an abnormal behaviour) shows that our average user
has a switch-on probability of about 3.3% between 0 and 100 lux, which drops to about
1.4% between 100 and 200 lux, then to about 0.6% between 200 and 300 lux, and which
then becomes more or less negligible.
Should the switch-on process be considered as a Poisson process, the above figures would
then correspond to an average switch-on time of about 150 minutes, 360 minutes, and 830
minutes respectively. In terms of “half-lives”, it means that if left in constant conditions,
half the users will have switched their lights on after 104 minutes, 250 minutes and 575
minutes respectively (assuming anyone is still left in the office).
0.04
0.03
Switch−on probability
0.02
0.01
0.00
[lux]
Figure 11: Intermediate switch-on probability for low illuminances, all data.
Fig. 11 shows the combined switch-on probability for low illuminance values, detail-
ing what happens below 100 lux. We see that below 50 lux the intermediate switch-on
probability continues to rise up to about 4%. It is difficult, however, to see whether the
probability should rise up to 1 for vanishing illuminance.
The figures obtained are very comparable to the ones proposed in the Lightswitch-2002
model (2), where the intermediate switch-off probability almost constant and equal to 2%
between 0 and 200 lux, and drops to 0.002 for higher illuminances. This model, and the
Suntool program based on it (3), further set this probability equal to 1 for zero illuminance.
18
This makes arguably sense but cannot be confirmed nor ruled out from our data alone, for
a practical reason: it is simply extremely unlikely that a user would allow the workplane
illuminance to drop to zero without switching on the lights before, and hence, we do not
have in our data these events.
19
Office 001 Office 002 Office 003 Office 004
5000
8000
3000
3000
3000
[s]
[s]
[s]
[s]
4000
1000
1000
1000
0
0 100 200 300 400 500 0 100 200 300 400 500 0 100 200 300 400 500 0 100 200 300 400 500
6000
1000 1500
3000
3000
[s]
[s]
[s]
[s]
2000
1000
1000
500
0 100 200 300 400 500 0 100 200 300 400 500 0 100 200 300 400 500 0 100 200 300 400 500
5000
5000
2000
8000
3000
3000
[s]
[s]
[s]
[s]
500 1000
4000
1000
1000
0
0 100 200 300 400 500 0 100 200 300 400 500 0 100 200 300 400 500 0 100 200 300 400 500
3000
1500
[s]
[s]
1000
500
0 100 200 300 400 500 0 100 200 300 400 500
[lux] [lux]
Figure 12: Time between user entry and use of controls vs illuminance levels, per office.
10000
8000
Delay before action [s]
6000
4000
2000
0
Figure 13: Time between user entry and use of controls vs illuminance levels, all data.
20
5 Conclusion
Three-quarters of the LESO users’ use of manual controls occur less than five minutes after
their arrival in their office or before their departure, most likely a direct consequence of
the traditional placement of their controls close to the door rather than close to their desk.
We have measured intermediate switch-on probabilities for the users of the LESO build-
ing as a function of horizontal workplane illuminance for five-minute time intervals. The
increase in probability for lower illuminances is consistent between users and can be taken
as being equal to about 3.3% between 0 and 100 lux, to about 1.4% between 100 and 200
lux, to about 0.6% between 200 and 300 lux, and negligible for higher illuminances.
When it comes to actual use of electric lighting controls, we have observed that the users
(about 30 people) behave quite differently from each other. It is remarkable that some users
give very frequent commands, while others seem more passive and do not bother using the
system. This lends credence to the classification in the Lightswitch-2002 algorithm of users
into different types according to their dynamic or static use of manual controls.
We have also realized that in the LESO building, where lighting controls are placed close
to the offices’ entrances, the users very seldomly use the dimmable feature of their electric
lighting. They almost always switch it completely on or completely off.
Appendix
Link with Nicol’s probit function
For this final section we will try to establish a link between user actions described by an
exponential distribution and the results given by Nicol in his 2001 paper (13).
It will be recalled that Nicol found that the fraction of offices within a given building
exhibiting a certain user behaviour (such as having the windows open, or having the fans
running) was a function of an external stimulus; in his case, he considers only the external
temperature, a hypothesis born out by the study by Fritsch (14). That function was given
by p(x) = exp(a + bx)/ 1 + exp(a + bx) where x was the external temperature and a
and b were parameters to be fitted from experimental data. We will see if we can derive
this result from our assumption about an exponential distribution of delays between user
actions.
Let us consider the case of window opening by the user. We choose this example because,
although the theory is similar, an open window will usually be closed after a short while
and reopened again during the day, whereas a lamp that has been turned on might remain
on until the end of the day. Therefore, we expect the average time between window
openings and closings to be shorter than the average time between the turning on or off of
artificial lighting, and it should be easier to verify this theory with windows during field
measurements. For this reason, we are going to derive an expression for the fraction of
windows open after a long enough time, but which will hold only for “reversible” user
actions, i.e. actions that the user effectively undoes during a day. The user will usually
21
close a window he or she has opened, but will seldomly switch off a light he or she has
turned on before leaving the office.
Let us assume that the mean time before a user opens a window is To , and once the
window is open, the mean time before the user closes it again is Tc . The distribution
function of the time remaining until the next user action is therefore an exponential function
with parameter λo = 1/To or λc = 1/Tc respectively.
For simplicity’s sake, let us assume all windows in the building start closed. We choose a
timestep ∆t sufficiently small so that the probability of the user both opening and closing
the window during that timestep is vanishingly small. If the window begins a timestep
closed, the probability that it should be open at the next timestep is exp(−λc ∆t). Similarly,
if the window begins the timestep open, the probability that it should be closed at the
next timestep is exp(−λo ∆t).
At time t = 0, the probability that the window is closed is Pc (0) = 1. At t = ∆t we
have:
Pc (∆t) = exp(−λo ∆t)
At t = 2∆t the probability that the window is closed is given by the probability that
the window was closed at t = ∆t and that it remained so at t = 2∆t, plus the probability
that it was open at t = ∆t but that it closed again at t = 2∆t:
Pc (2∆t) = Pc (∆t) exp(−λo ∆t) + 1 − Pc (∆t) 1 − exp(−λc ∆t)
Expanding the right-hand side all the way down to Pc (0), we finally obtain:
n
Pc (n∆t) = exp(−λo ∆t) + exp(−λc ∆t) − 1
n−1
X i
+ 1 − exp(−λc ∆t) exp(−λo ∆t) + exp(−λc ∆t) − 1
i=0
For large n, and writing T = n∆t → ∞, the first term in that sum vanishes and the
second one is the sum of a geometric serie. We obtain thus
1 − exp(−λc ∆t) 5
Pc (T = ∞) =
1 − exp(−λo ∆t) + 1 − exp(−λc ∆t)
5
We obtain exactly the same result if we consider the state of the window over time as a Markov process
with two states. The transition probability from closed to open is Tco = 1 − exp(−λo ∆t), and the one
from open to closed is Toc = 1 − exp(−λc ∆t). The Markov process asymptotically tends to a state in
which the probability of having the window closed is TcoT+Toc
oc
, which is none other than the equation
above.
22
Taking the limit ∆t → 0, we obtain
λc
lim Pc (T = ∞) =
∆t→0 λo + λc
In his paper, Nicol observed that the fraction of open windows obeyed a relationship with
the outside temperature. Reordering a little bit, he found essentially that the fraction of
closed windows could be written in the form
1
Pc (T = ∞) = ,
1 + exp(a + bt)
where t is the outside temperature (not the time). Identifying his finding with ours, we
see immediately that
λo
= exp(a + bt)
λc
We have just found a theoretical relationship between Nicol’s a and b probit parameters
and the average times between user’s opening and closing of windows.
Notice also that the preceding equation can be rewritten by taking λc = 1/Tc , where Tc
is the average time before window closure, and similarly for average time before window
opening, and taking the logarithm on both sides:
log Tc − log To = a + bt
But since a was an arbitrary constant, and assuming independence between To and t if
the user’s need to open the window is taken as independent of the outside temperature,
we can redefine it as a = log To + a, and we see thus that
log Tc = a + bt
In other words, we find a theoretical affine relationship between the logarithm of the
mean time before window closure, and the outside temperature.
Acknowledgements
We wish to thank Jessen Page (LESO-PB/EPFL) for kindly reviewing this paper and
providing helpful and insightful comments on the more mathematical aspects.
We thank Lee-Ann Nicol (LESO-PB/EPFL) for going through this paper and leaving
no page without comments for improvements of style or english grammar.
We thank Denis Bourgeois (Ecole d’Architecture, Université Laval, Canada) for having
been the catalyst behind some of the more theoretical developments in this work.
We thank Christoph Reinhart (National Research Concil, Canada) for going through
this paper and offering his helpful comments for improvement.
23
References
[1] Christoph F. Reinhart. Daylight Availability and Manual Lighting Control in Office
Buildings—Simulation Studies and Analysis of Measurements. PhD thesis, University
of Karlsruhe, 2001.
[2] Christoph F. Reinhart. Lightswitch-2002: a Model for Manual and Automated Control
of Electric Lighting and Blinds. Solar Energy, 77:15–28, 2004.
[4] D. Robinson and A. Stone. Solar Radiation Modeling in the Urban Context. Solar
Energy, 77:295–309, 2004.
[5] Christoph F. Reinhart and Karsten Voss. Monitoring Manual Control of Electric
Lighting and Blinds. Lighting Research and Technology, 35(3):243–260, 2003.
[7] René Altherr and Jean-Bernard Gay. A Low Environmental Impact Anidolic Facade.
Building and Environment, 37(12):1409–1419, 2002.
[8] Antoine Guillemin. Using Genetic Algorithms to Take into Account User Wishes in
an Advanced Building Control System. PhD thesis, LESO-PB/EPFL, 2003.
[9] Antoine Guillemin and Nicolas Morel. Experimental Assessment of Three Automatic
Building Controllers over a 9-Month Period. In proceedings of the CISBAT 2003
conference, pages 185–190, Lausanne, Switzerland, 2003.
[10] Antoine Guillemin and Simone Molteni. An Energy-Efficient Controller for Shading
Devices Self-Adapting to User Wishes. Buildings and Environment, 37(11):1091–1097,
2002.
[12] R Development Core Team. R: A language and environment for statistical computing.
R Foundation for Statistical Computing, Vienna, Austria, 2004. 3-900051-07-0.
24