Smart Office Lighting Control Using Occupancy

Sensors
Xin Wang Tjalling Tjalkens Jean-Paul Linnartz
Department of Electrical Engineering Department of Electrical Engineering Philips Lighting Research and
Eindhoven University of Technology Eindhoven University of Technology Department of Electrical Engineering
Eindhoven, The Netherlands Eindhoven, The Netherlands Eindhoven University of Technology
Eindhoven, The Netherlands

Abstract-Nowadays, despite the use of efficient LED lighting, based on the Doppler effect. They send high frequency sound
lighting consumes a considerable amount of energy. To reduce the waves and microwaves and check the reflected patterns. If the
energy consumption, many office lighting systems are equipped
reflected pattern is changing continuously, then the office is
with occupancy sensors. Since these sensors have a limited relia­
bility in detecting presence, usually very conservative strategies
regarded as occupied[3], [6].
are used such as keeping the lights on for fifteen minutes after the However, the sensors are not always reliable. For instance,
last detected presence. In this paper, we propose a novel lighting non-moving users are often ignored by ultrasonic sensors and
control strategy for an office that considers the sensor output microwave sensors. Thus the lighting system has to face a
as a noisy observation of the real occupancy status. Simulation
dilemma that it either takes the risk to turn off the light
results show that compared with conventional "on/off" strategy,
it saves more energy.
when there are non-moving users in the office, which is very
annoying, or to keep the light on when the office is empty,
I. INTRODUCTION which wastes energy.
In the past few decades, lighting systems developed rapidly. As a compromise, a simple and straightforward idea used by
According to [1], lighting consumes 19% of global energy many primary presence-based lighting systems is to switch on
and is responsible for 6% of greenhouse emissions worldwide. the light immediately after presence is detected and turn off the
Particularly, 75% of the consumed electric light is used for light if no presence is detected during a certain time period [6].
residential and conunercial buildings, which results in almost Usually the period is set to fifteen to twenty minutes, which
1.9 x 109 tons of CO2 emissions [2]. is used to reduce the false negative rate, i. e. , the probability
To reduce the energy consumption of an office lighting that an occupied office is regarded as empty. But this method
system, there are several approaches. The most straightforward has another problem that the sensor only passes the binary
approach is to replace low energy efficiency lamps such as occupancy status to the lighting system. A lot of information
incandescent lamps by high energy efficiency lamps such as is lost in this process. For example, a presence status with
compact fluorescent lamps and LED lamps. For the same 99% confidence and a presence status with 51% confidence
illuminance level, this will decrease the energy cost by 47% will lead to the same control strategy, which is obviously sub­
to 55% [2]. optimal.
Further, using occupancy sensors, an intelligent lighting As an improvement, instead of using only a binary status,
system can monitor the occupancy status of the office [3]. Then sensors can also share the estimated occupancy probability
following a certain strategy, the light can be dimmed or turned of the office with the lighting system. According to the
off for unoccupied areas. Compared with a traditional manu­ estimated occupancy probability, the lighting system decides
ally controlled lighting system, this will avoid the situation whether to turn on the light or not. One possible solution
that the users forget to turn off the light when they leave the is to set a probability threshold, above which the light will
office, which will reduce the unnecessary energy consumption. be turned on and below which the light will be switched
Moreover, with advanced machine learning algorithms, the off [6]. However, the fundamental problem is still there. An
performances of intelligent lighting systems can be further estimated occupancy probability just above the threshold and
improved [4], [5]. an estimated occupancy probability close to one would still be
In this paper, we mainly focus on discussing the second the same for the lighting system. The information brought by
approach. In modern lighting systems, there are usually one or the estimated occupancy probability is not fully utilized.
multiple occupancy sensors in the office which are responsible In this paper, we propose an idea to further exploit the
for detecting the occupancy status of the office. Commonly information contained in the estimated occupancy probability.
seen occupancy sensors in office lighting systems include Instead of either switching the light on or turning the light
passive infrared sensors, ultrasonic sensors, and microwave off, the system dims the light in the office according to the re­
sensors. The passive infrared sensors can detect infrared ra­ ceived occupancy probability. According to this (time varying)
diation radiated by human body within the observation field. occupancy probability, the system optimizes the illuminance
The ultrasonic sensors and the microwave sensors are both level in the office so that the energy consumption is reduced.

978-1-5090-4429-0117/$3l.00 ©2017 IEEE

 Intuitively speaking, the higher the presence probability
is, the brighter the office will be. Yet it is possible that an
occupied office might be only illuminated to a low illuminance
level due to improper sensor installations or inappropriate
dimming strategies. In this case, the user(s) in the office will
intervene the light system to illuminate himself. The aim of
this paper is to find the optimal dimming strategy given the
detected presence probability.

II. PROBLEM FORMULATION

We consider an office lighting system equipped with a

switch and several occupancy sensors which pass the occu­
pancy probability p to the decision module of the lighting
system. Then according to the sensed occupancy probability p,
the system illuminates the office to a certain illuminance level Fig. 1: The probability of users turning on the light plotted
x. If the user feels x as insufficient, he/she will fully turn on against the minimum illuminance on the working plane. The
the light to illuminate the office to a default light level Xo. The circles are Hunts experimental results [7]. The curves are the
target is to design an optimal lighting control strategy such that log-normal functions with different (T2 and Xo 500 Ix.
=

the expected energy consumption of the lighting system after

being intervened is minimized.
When the office is empty, the lighting system will not be As indicated in [8], this user behavior can be quantitatively
intervened so the illuminance level of the office will still be x. described by a parameterized log-normal function
When the office is occupied, we assume that with probability
q(x) the user in that office will feel insufficientand turn on
q(x) =
{1- exp ( - ;2 (Inx - lnxo) 2) , if 0 � x � Xo;
the light to illuminate the office to the standard illuminance 0, if x> Xo.
level Xo Ix; with probability - q(x) the user will accept the
1 (4)
illuminance level x. In (4), (T2 is a non-negative parameter describes the tendency
For an office with presence probability p which is firstly of targeted users to turn on the light which is determined by
illuminated to x, the expected illuminance level after the the contexts, conditions, users, etc. The default illuminance
possible intervention is level Xo is selected as 500 Ix according to several standards
and other recommendations [9], [10]. Note that choosing a
E(x) = xopq(x) + x(l - pq(x)). (1)
different Xo such as 400 Ix has no fundamental change in the
optimization because the expected illuminance level E(x) can
Then the expected energy consumption of the lighting
be rewritten as
system is
C(x) = o:E(x) + Eo, (2) E(x) = (
Xo pq(x) + :0 (1 - pq(X))) (5)

where 0: and Eo are the energy efficiency and the standby

and q(x) is also only determined by the ratio between x and
power of the lighting system, respectively.
Xo. This means in (3), what we optimize is the ratio between
In (2), we assume that for a specific lighting system, the
x and Xo.
energy efficiency and the standby power are both independent
In Figure 1, we plot q(x) with different (T2 and a fixed Xo =

with the illuminance level x. Then the optimum illuminance

500 Ix. It can be seen that the larger (T2 is, the more likely the
level can be derived by
targeted users accept relatively low light levels. Considering
that the experiment was conducted in the late 1970s, most of
x* = argmin C(x) = argminE(x). (3)
x x Hunt's data are less than 300 lx, which may be rather low for
common office tasks today. Actually as the energy efficiency
III. A QUA NTITATIVE USER BEHAVIOR MODEL
of luminaires increases, the recommended illuminance level
In practice, q(x) is determined by the group behavior of for office work also increases in the past few decades [11].
target users in the office. For example, in 1979, Hunt measured
the probability of people turning on the light for different IV. THE OPTIMAL DIMMING STRATEGY
minimum illuminance on the working plane [7]. As can be seen from (1), E(x) consists of two parts. The
However, Hunt's experiment only observed whether users first term El(x) xopq(x) is the expected illuminance level
=

turn on the light after they just enter the office. Considering when the office is occupied and the user needs more light
the fact that different groups of users have different needs so the light is turned fully on. The second term E2(x) =

and preferences, we want to parameterize Hunt's results. x(l -pq(x)) is the expected illuminance when the pre-assigned
 X800r-�-r��.--����--�-�-�
�
� 700
C
<IJ

�1: 600 �4-�-' __L

� 500
c
1l400�_.....
'­
____
<IJ
'Iii 300
=>
== 200
� �.-�-.����
tJ 100
<IJ
c.

�!t..:::±--'�20�0--'�30�0�:"""40�0-----;'50�0-;'�===�70!!'!!0��800 � °0�!E±:"'-o:��..t.:.';30;'0-----;:40"'0---;:50"'--':�60�0::!2�70!!0-�800
Preset Illuminance (Ix) Preset Illuminance (Ix)
(a) (J2 = 0.25. (b) (J2 = 0.5.

� 700
C
<IJ
1: 600
2
c
� 500
c
1l400r__...... �

Q;
'Iii 300
=>
== 200 _ p�0.2 _ p�0.2
"0
<IJ
... p�0.5 ... p�0.5
tJ 100 .... p�0.7
<IJ
C.
X Hp�1
W
600 700 800
(e) (J2 = 1. (d) (J2 = 5.

Fig. 2: Two terms of the expected illuminance after possible intervention (Ix) as a function of x with different CJ2 and p. The
solid curves are El(X) while the dashed curves are E2(x).

light level x is not changed. It is either because the office is more tolerant the targeted users are, the lower the threshold
not occupied or the user is satisfied with the pre-assigned light will be. This is because for darkness-intolerant (small CJ2)
level x. users, it is meaningless to preset a low illuminance level
As illustrated in Figure 2, El(x) is q(x) multiplied by a because with a large probability they will turn on the light
constant XoP, which is monotonously decreasing. Larger x is fully on anyhow.
favored from this perspective. On the other hand, however,
as indicated by E2(x), larger x potentially consumes more
unnecessary energy, especially when the presence probability 500ir===�=C�--'----'---:---�
is low or the tolerance is high. In fact, we are taking advantage
_ ,,2�0.25
•• ,,2�0.5
of the user tolerance to save energy because many research x e,. (72=1
� 400
results indicate that for a typical office, 500 lx is usually more
<IJ
u •..• ,,2�5
C
'"c
than adequate, and many users accept lower light levels [12],
·E 300
[13], [7], [14]. �
. ... .
......
From Figure 2, we can see that the optimal strategy is a 'QJ
<II

trade-off between El(x) and E2(x). High illuminance levels � 200

a.

are favored by El(x) and E2(x) prefers low illuminance E •.........

=>
E
levels. In Figure 3, we plot the optimum illuminance level :c;
c.
100
as a function of the presence probability p for different CJ2. o

It is clearly shown that for every CJ2, there is an optimal

threshold. If the detected presence probability p is less than
this threshold, the light will be turned off. Otherwise the
optimal illuminance level x* increases as p increases. The Fig. 3: The optimal dimming strategy for different CJ2.
Fig. 4: Performance comparison between the optimal strategy and the hard decision making strategy for different (Y2 and Po.
The solid line with circle marker indicates the performance of the optimal strategy. The dashed lines indicate the performance
of the hard decision making strategy with different Po.

For example, in Hunts experiment, the participants are quite light is fully turned on and otherwise the light will be turned
tolerant to darkness; q(x) with (Y2 = 5 fits their behavior off. However, if the light is turned off but the office is actually
quite well. For this group of users, according to Figure 3, occupied, we assume the light will soon be manually turned
the threshold is 0. 24. In other words, when the presence on by the user.
probability of the office is less than 0. 24, the light will As a result, if the hard decision making strategy is applied to
be turned off. Otherwise the pre-assigned illuminance level a lighting system, the expected illuminance level is a function
increases as the detected presence probability increases but it of the detected presence probability P
will not exceed 160 Ix. This is reasonable because according to
Figure 1, when offered 160 lx, the probability that this group if p 2: Po;
(6)
of users will turn on the light is less than 0. 2. if p < Po.

V. COMPARISON WITH THE HARD DECISION MAKING Now we could see that both the optimal strategy and the
STRATEGY hard decision making strategy have its own threshold. When
In this section, we compare the performance of the optimal the detected presence probability is smaller than the threshold,
strategy described above and the commonly seen hard decision the light is turned off otherwise the office will be illuminated
making strategy mentioned in the introduction section. In the to a certain illuminance level.
hard decision making strategy, there is a pre-defined threshold However, there are two major differences between the two
Po. When the detected presence probability P exceeds Po, the strategies. The first one is that when the detected presence
probability exceeds the threshold, the hard decision making
strategy will turn the light fully on but the optimal strategy will
gradually increase the light level as the presence probability
increases. The second one is that the threshold of the hard
decision making strategy is a pre-determined constant while
the threshold of the optimal strategy is derived by the group
behavior of the users.
In Figure 4, we compare the performances of the optimal
... -
strategy and the hard decision making strategy. First it can "C
... - -
r-
be seen that the dashed lines are always above the solid �
line, which means compared with the hard decision making 'E
w
50

i ..
strategy, the optimal strategy always saves energy. Moreover, . .. .

the more tolerant the users, the more energy can be saved. �
a 0
0.0 0.4 0.6 0.8 1.0
Presence Probability
VI. OP TIM A L DIMMING STR ATEGY W ITH DAY LIGHT
H A RVE STING Fig. 5: The optimum amount of preset artificial illuminance
for different a2 as a function of the presence probability p
Daylight harvesting is a popular idea in recent electric when the amount of daylight XD is 300 Ix.
lighting systems because by properly utilizing the free and
green sunlight, the energy consumption of a lighting system
can be reduced tremendously [15], [16], [17]. Usually a VII. CONCLUSION
daylight harvesting module includes one or multiple photo
In this paper, we propose a novel lighting control strategy
sensors and is capable to estimate the amount of daylight in
for an office with occupancy sensors. According to the pres­
the office [15], [18].
ence probability estimated by the occupancy sensors, the light­
The design of the photo sensors and the daylight estimation ing system will illuminate the office to a certain illuminance
algorithm are beyond the scope of this paper. Here we only level which is usually less than the recommended illuminance
want to discuss how to integrate daylight harvesting into level. Users in the office may turn the light fully on if the light
the optimal dimming strategy. We assume that the daylight is insufficient for them. To reduce the energy consumed by the
estimation works perfectly and the amount of daylight in the lighting system, we calculate the optimum preset illuminance
office is xD Ix. Then if the amount of artificial light offered given characteristics of the users and the presence probability.
by the lighting system is x, the probability that the users need From the results we could see that when the presence
more light and turn the light fully on is q(x + XD) . probability is low, it is wise to completely turn off the light.
Correspondingly, the expected illuminance E(x) emitted by When the presence probability exceeds a certain threshold
the lighting system becomes which is determined by the characteristics of the users, the
optimum illuminance level gradually increases as the presence
E(x) = xopq(x + XD) + x(l - pq(x + XD)) ' (7)
probability increases.
Then we compare our strategy with the commonly seen
Then following a similar optimization procedure, we can
hard decision making strategy. It can be seen that under this
optimize the preset amount of artificial light given the presence
model, our strategy always reduces the energy consumption.
probability p and the amount of daylight XD, which is shown
When daylight is taken into account, the optimum amount of
in Figure 5.
complementary artificial light has a similar behavior.
By comparing it with Figure 3, we could see that due to
Admittedly, in this paper, we only focus on minimizing
the nonlinearity of q(x), when daylight harvesting is taken
energy consumption, which may result in some side effects.
into consideration, the optimum amount of the preset artificial
For example, when the output of the occupancy sensor is not
illuminance is not a simple shift of that shown in Figure 3.
stationary, there might be annoying flickers; if there is a delay
However, it still shows a similar behavior. Moreover, we could
in the sensor output when the occupancy status of the office
see that for those users who are quite tolerant to dark (a2 =5
changes, there will also be a delay in the lighting output. We
, participants of Hunts experiment), it is not necessary to turn
will leave theses problems for further research.
on the light at all when the amount of daylight is 300 lx, which
is enough for them. REF EREN C E S
When the amount of daylight is not 300 lx, the optimum
[1] A. Bahga and V. Madisetti, Internet o f Things: A Hands-on Approach.
curves need to be recalculated following the same procedure. VPT, 2014.
Actually for a practical lighting system, we can calculate the [2] OECD, World energy outlook 2015. OECD/lEA, 2015.
optimum curves for several representative amounts of daylight [3] T. Labeodan, C. De Bakker, A. Rosemann, and W. Zeiler, "On the ap­
plication of wireless sensors and actuators network in existing buildings
in advance and adjust the amount of artificial light according for occupancy detection and occupancy-driven lighting control," Energy
to the real time measurement. and Buildings, vol. 127, pp. 75-83, 2016.
 [4] A. K. Gopalakrishna, T. bz�elebi, A. Liotta, and J. J. Lukkien, "Ex­
ploiting machine learning for intelligent room lighting applications," in
Intelligent Systems (IS), 2012 6th IEEE International Coriference. IEEE,
2012, pp. 406-411.
[5] A. K. Gopalakrishna, T. Ozcelebi, J. J. Lukkien, and A. Liotta, "Rele­
vance in cyber-physical systems with humans in the loop," Concurrency
and Computation: Practice and Experience, vol. 29, no. 3, 2017.
[6] J. D. Jennings, F. M. Rubinstein, D. DiBartolomeo, and S. L. Blanc,
"Comparison of control options in private offices in an advanced lighting
controls testbed," Journal of the Illuminating Engineering Society,
vol. 29, no. 2, pp. 39-60, 2000.
[7] D. Hunt, "The use of artificial lighting in relation to daylight levels and
occupancy," Building and environment, vol. 14, no. 1, pp. 21-33, 1979.
[8] X. Wang and J.-P. Linnartz, "Intelligent illuminance control in a
dimmable led lighting system," Lighting Research and Technology, 2016.
[9] C. CIBSE, "Code for interior lighting," London: T he Chartered Institute
of Building Services Engineers, 1994.
[IO] M. S. Rea, T he IESNA lighting handbook: reference & application.
Illuminating Engineering Society of North America, 2000.
[11] E. Mills and N. Borg, "Trends in recommended illuminance levels:
an international comparison," Journal of the Illuminating Engineering
Society, vol. 28, no. 1, pp. 155-163, 1999.
[I2] P. Boyce, Human factors in lighting. CRC Press, 2010.
[13] G. R. Newsham and J. A. Veitch, "Lighting quality recommendations
for vdt offices: a new method of derivation," Lighting Research and
Technology, vol. 33, no. 2, pp. 97-113, 2001.
[14] S. Chraibi, T. Lashina, P. Shrubsole, M. Aries, E. van Loenen, and
A. Rosemann, "Satisfying light conditions: a field study on perception
of consensus light in dutch open office environments," Building and
Environment, vol. 105, pp. 116-127, 2016.
[I5] R. Leslie, R. Raghavan, O. Howlett, and C. Eaton, "The potential
of simplified concepts for daylight harvesting," Lighting Research &
Technology, vol. 37, no. 1, pp. 21-38, 2005.
[I6] 1. Lu and K. Whitehouse, "Suncast: fine-grained prediction of natural
sunlight levels for improved daylight harvesting," in Proceedings of
the 11th international coriference on lriformation Processing in Sensor
Networks. ACM, 2012, pp. 245-256.
[17] A. Sarkar, M. Fairchild, and C. Salvaggio, "Integrated daylight har­
vesting and occupancy detection using digital imaging," in Electronic
Imaging 2008. International Society for Optics and P hotonics, 2008,
pp. 68 160F-68 160F.
[18] S. Matta and S. M. Mahmud, "An intelligent light control system
for power saving," in IECON 201O-36th Annual Conference on IEEE
Industrial Electronics Society. IEEE, 2010, pp. 3316-3321.