Received February 3, 2017, accepted February 28, 2017, date of publication March 9, 2017, date of current version October

25, 2017.
Digital Object Identifier 10.1109/ACCESS.2017.2679807

A Data-Driven Daylight Estimation

Approach to Lighting Control
STEFANO BORILE1,2 , ASHISH PANDHARIPANDE1 , (Senior Member, IEEE), DAVID CAICEDO1 ,
LUCA SCHENATO2 , (Fellow, IEEE), AND ANGELO CENEDESE2
1 Philips Lighting Research, 5656 AE Eindhoven, The Netherlands.
2 Department of Information Engineering, University of Padova, 35122 Padua, Italy.
Corresponding author: Ashish Pandharipande ([email protected])

ABSTRACT We consider the problem of controlling a smart lighting system of multiple luminaires with
collocated occupancy and light sensors. The objective is to attain illumination levels higher than specified
values (possibly changing over time) at the workplace by adapting dimming levels using sensor information,
while minimizing energy consumption. We propose to estimate the daylight illuminance levels at the
workplace based on the daylight illuminance measurements at the ceiling. More specifically, this daylight
estimator is based on a model built from data collected by light sensors placed at workplace reference
points and at the luminaires in a training phase. Three estimation methods are considered: regularized least
squares, locally weighted regularized least squares, and cluster-based regularized least squares. This model
is then used in the operational phase by the lighting controller to compute dimming levels by solving a linear
programming problem, in which power consumption is minimized under the constraint that the estimated
illuminance is higher than a specified target value. The performance of the proposed approach with the
three estimation methods is evaluated using an open-office lighting model with different daylight conditions.
We show that the proposed approach offers reduced under-illumination and energy consumption in compar-
ison to existing alternative approaches.

INDEX TERMS Lighting control systems, daylight estimation, occupancy and daylight adaptation,
least squares, linear programming.

I. INTRODUCTION values from the sensors, is able to provide a total illuminance

A major proportion of the electrical energy consumption in level that is higher than specified values at certain control
commercial office buildings is due to artificial lighting [1]. points at workplaces. The total illuminance is an aggregation
Portions of lighting energy are often misspent due to ineffi- of both artificial light and daylight. A key challenge in the
cient management of ambient conditions [2], [3]. To reduce design of the control law is the lack of knowledge of the day-
lighting energy consumption, control of artificial lighting light mapping from ceiling (where measurements are made
has been an active topic of recent research, in particular by the light sensors) to the workplaces.
by adapting to occupancy and daylight changes [6]- [19].
Such lighting control systems require dimming ability in A. RELATED WORK
luminaires, which can be achieved flexibly with light emitting Since the illuminance over workplaces is of interest, a direct
diodes (LEDs) [20]. approach is to measure the illuminance at certain control
In this paper, a smart lighting system consisting of multi- points in this plane. The works in [4] and [5] considered opti-
ple luminaries, with collocated occupancy and light sensors, mization techniques for lighting control assuming knowledge
and a central controller is considered. These sensors provide of light distributions at the workplaces as well as occupant
binary occupancy and net illuminance values within their locations. In [8]–[10] light sensors were placed at workplaces.
fields-of-view respectively. The local sensor information at Lighting control for daylight adaptation was performed using
each luminaire is used to control the luminaires individually sensor measurement feedback transmitted to a controller by
using a control law at the central controller. The main objec- wireless communication [8], [9]. In [6] and [7], light sen-
tive is to design a control law that, taking into account the sors were carried by occupants. Such sensor configurations

2169-3536
2017 IEEE. Translations and content mining are permitted for academic research only.
VOLUME 5, 2017 Personal use is also permitted, but republication/redistribution requires IEEE permission. 21461
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 S. Borile et al.: Data-Driven Daylight Estimation Approach to Lighting Control

however have limitations. A commissioning step is needed to the achieved illuminance value. This optimization framework
properly associate light sensor data to control the luminaires. is described in Section IV.
Moreover, temporary physical obstructions may occur that Finally we evaluate the performance of our proposed
impact the quality of the sensor measurements as well as the approach using data from an open-office lighting model
wireless connectivity between the sensors and the controller. in Section V. As comparison, we consider a lighting sys-
It is thus common practice in lighting control systems tem that is controlled solely on the basis of measurements
to deploy light sensors at the ceiling [12], [16], [17], [18], at the ceiling-based light sensors (no training phase), with
[21], [22]. In particular, collocated sensors at the luminaires illuminance constraints defined at these light sensors. We
simplifies the commissioning step and have been considered show that the proposed approach is able to achieve reduced
for lighting controls [11], [18], [21]. This however means that under-illumination, while also obtaining substantial energy
direct measurements at the control points at workplaces are savings.
not available. As such, a simple night time calibration [21]
using the artificial lighting is used to establish a relation II. SYSTEM MODEL AND PROBLEM FORMULATION
between illuminance at the ceiling and illuminance at the We consider a smart lighting system in an open office with
workplaces. N workplaces. The smart lighting system has M LED lumi-
The contribution of daylight at the workplace and at the naires with collocated occupancy and light sensors that pro-
ceiling can vary with time, depending on the incidence of vide binary occupancy information and illuminance levels,
daylight (influenced by time of day, weather conditions, etc.) respectively. At each workplace, a control point is defined
in the indoor space. To the best of our knowledge, this impor- where a minimum illuminance level is desired. This scenario
tant insight was first reported in [22] and [23], to show that is illustrated in the lighting models shown in Figure 1 and 2.
maintaining a constant output at a ceiling based light sensor
does not result in constant illumination at the workplace.
In these works, an offset was proposed to be included in a
closed-loop proportional control algorithm for a single light
sensor driven lighting system. This offset related to the ceiling
sensor to workplace daylight ratio was determined using a
one-time daylight calibration. An assumption made here was
that the ceiling sensor to workplace daylight ratio during
calibration is a good fit to the overall set of possible ratios.
Extensions of such an approach to a distributed lighting sys-
tem with multiple light sensors were studied in [11]. In [17],
we showed that a constrained optimization approach results
in a lower energy consumption compared to [11], while still
FIGURE 1. Open-office lighting system model with multiple luminaires
achieving the light sensor set-points. All the aforementioned and collocated sensors.
methods cannot effectively deal with the daylight mismatch
problem since light sensor measurements are done at ceiling Let y(k) ∈ RM and w(k) ∈ RN be the vectors that contain
locations. the illuminance levels at time instant k at the light sensors at
the ceiling and at the control points at the workplace plane,
B. CONTRIBUTIONS AND OUTLINE OF THE PAPER respectively. We also define d(k) ∈ RM and p(k) ∈ RN as the
In this work, we propose a data-driven daylight estimation daylight contribution to the light sensors at the ceiling and at
approach to lighting control. We consider a training phase the control points at the workplace plane, respectively.
wherein light sensors are placed at workplaces in addition to We consider that the output power of the luminaries is
those at the ceiling. In this phase, daylight values are collected controlled using pulse width modulation (PWM). Denote by
at both sets of sensors. The data is then used to obtain an esti- u(k) ∈ RM the vector containing the PWM duty cycles of
mate of the mapping between the ceiling measurement points the luminaires, representing their dimming values. Due to
and the control points at the workplaces. Three methods are physical limits on dimming, each element of u(k) takes a
investigated to obtain the estimate: regularized least squares, value between zero and unity, i.e.
locally weighted regularized least squares, and cluster-based
regularized least squares. These methods are described fur- 0 ≤ u(k) ≤ 1, (1)
ther in Section III.
We then formulate an optimization problem for minimizing where the inequality should be interpreted component-wise,
the power consumption of the lighting system under the and 0 = [0 0 · · · 0]T ∈ RM and 1 = [1 1 · · · 1]T ∈ RM
constraints that the achieved illuminance at certain control represent vectors of all zeros and all ones, respectively. Under
points at the workplaces is higher than specified values and PWM dimming, lighting energy is proportional to the control
the dimming levels are within physical limits. The estimated input [21]. Therefore, J (u(k)), the total energy consumed at
mapping is used in the first constraint to obtain an estimate of time instant k is proportional to the sum of the dimming

21462 VOLUME 5, 2017

S. Borile et al.: Data-Driven Daylight Estimation Approach to Lighting Control

j-th luminaire is turned on to its maximum, i.e. uj (k) = 1,

while all the others are off, i.e. uh (k) = 0, h 6 = j, and under
the absence of daylight, i.e. d(k) = 0. A similar consideration
holds for Hi,j ≥ 0.
The illuminance levels at the control points at the work-
place plane, w(k), can only be measured by placing light
sensors at each control point. However, in practice, this might
not be possible or desirable due to reliability issues, e.g.
the user may unintentionally block the light sensor. Note
that using (3) we can estimate the illuminance level at time
instant k for each control point at the workplace plane, w(k),
if we have knowledge of: (i) illuminance gain matrix, H ,
(ii) dimming vector, u(k − 1), and (iii) daylight contribution
term, p(k). The illuminance gain matrix H can be measured
during a calibration phase and the dimming vector is known
at the lighting control. Thus, the only remaining unknown is
the daylight contribution term p(k).
The objective of this work is to propose a method to obtain
an estimate of the daylight contribution at each control point
at the workplace plane at each time instant k, b p(k), based
on daylight contribution measurements at light sensors at the
ceiling, d(k). Using the estimate b p(k), we propose a lighting
control method that minimizes the energy consumption of
the lighting system while providing the desired illuminance
levels at each control point at the workplace plane, i.e.
FIGURE 2. Top view of the simulated office model in Figure 1 showing
location and numbering of the 80 luminaries and collocated ceiling u∗ (k) = arg min Emax 1T u
sensors (depicted by blue boxes and red dots respectively) and of the u
36 light sensors at the workplaces (red dots). Windows span the right
side of the office. 0 ≤ u ≤ 1

s.t. w(k) = Hu + b p(k) (4)

w(k) ≥ wr (k),

values, i.e.
where u = [u1 · · · uM ]T . Here, wr (k) ∈ RN is the vector of
M
X reference illuminance values at the workplace plane at time
J (u(k)) = Emax um (k) = Emax 1T u(k), (2)
instant k, which is a function of the presence of a person and
m=1
possibly his/her personal desired illuminance level.
where Emax is the maximum energy consumption of each
LED luminaire and u(k) = [u1 (k) · · · uM (k)]T . III. DATA-DRIVEN DAYLIGHT ESTIMATION
The illuminance levels at time instant k at the ceiling and In this Section, we propose a method to estimate at each
workplace plane (y(k) and w(k), respectively) can be written time instant k the daylight contribution term p(k) based on
as a linear combination of the artificial illumination due to the the current value of daylight contribution at the ceiling d(k).
lighting system and daylight contribution [11], Note that d(k) can be estimated from the light sensor mea-
( surements y(k) and the previous control input u(k − 1) as
y(k) = Gu(k − 1) + d(k),
(3) d(k) = y(k)−Gu(k −1). More formally, we want to compute
w(k) = Hu(k − 1) + p(k), an estimator function f : RM → RN such that
where p(k) = f (d(k)),
b (5)
y(k) = [y1 (k) · · · yM (k)] , T
where b p(k) is an estimate of p(k) at time instant k.
d(k) = [d1 (k) · · · dM (k)]T , This function will be computed via an identification-based
w(k) = [w1 (k) · · · wN (k)]T and approach [30] based on experimental data collected during
p(k) = [p1 (k) · · · pN (k)]T . a training phase. During the training phase, we assume that
it is possible to collect not only daylight measurements at the
Matrices G ∈ RM ×M and H ∈ RN ×M are the illumi- ceiling d(k), but also daylight measurements at the workplace
nance gain matrices for the ceiling and workplace illumi- plane p(k). In the next subsections we describe in detail the
nation, respectively. More specifically, Gi,j ≥ 0 represents proposed approach. In section V-B the performance of the
the illuminance measured by the i-th light sensor when the estimation techniques will be evaluated and compared.

VOLUME 5, 2017 21463

 S. Borile et al.: Data-Driven Daylight Estimation Approach to Lighting Control

A. DATA COLLECTION where P = [p1 p2 · · · pq ] ∈ RN ×q and D = [d1 d2 · · · dq ] ∈

During the training phase, illuminance data is collected at RM ×q are typically ‘‘fat’’ matrices, i.e. q  M , q  N
both the ceiling light sensors and the control points at the obtained by the data available in the training set T .
workplace plane. This data can be collected during periods Matrix DDT may be ill-conditioned if the data elements di
of unoccupancy, e.g. over weekends. The light sensors at the are strongly correlated, i.e. very similar to each other (over-
workplaces do not need to be installed permanently; they fitting problem), or the size of the training set T is small.
only need to be in place over a time period necessary for A good practice in these scenarios is to add a regularization
data collection. Another option is to obtain the data set via term  > 0. The regularized least square problem is formally
realistic simulated light propagation models. For simplicity, defined as follows:
we also assume that this data is collected when all luminaries q
X
are turned off, i.e. u(k) = 0, ∀k during the training phase. Let CRLS () = argmin ||pi − Cdi ||2 + kCk2F (9)
D be the data set containing all the collected data. We assume C i=1
that r samples were collected during the training phase, i.e. = PDT (DD + I )−1 .
T
(10)
D= {(pi , di )}ri=1 , (6) where we made explicit the dependence of CRLS on  and
|| · ||F is the Frobenius norm. Parameter  is often chosen by
where each pair (pi , di ) represents the measurements col- finding the best performance on the validation set V [25], i.e.
lected at some time instant ti , i.e. pi = p(ti ), di = d(ti ). In our
r
approach we will not take into account the specific time of the X
 ∗ = argmin ||pi − CRLS ()di ||2 . (11)
day and the month when samples were collected, but we will ≥0 i=q+1
look for a function that is able to estimate the current value
of the daylight contribution at the workplace based on the Note that the LS solution in (8) is a special case of (9) where
measurement vector at the ceiling. As such, the specific order the regularization parameter  is set to zero.
of the pairs (pi , di ) is irrelevant and the sampling time ti is
not used as part of the data set. The data set is further divided C. LOCALLY WEIGHED REGULARIZED LEAST
in two disjoint subsets: a training set T used to compute the SQUARES (LWRLS) DAYLIGHT ESTIMATOR
estimator, and a validation set V used to evaluate the estimator The LS problem can be extended by possibly weighing each
performance. More formally: pi k2 differently. To do so, a Locally Weighted
residual kpi − b
q Least Squares (LWLS) approach is performed. The problem
D = T ∪ V, T = {(di , pi )}i=1 , V = {(di , pi )}ri=q+1 . formulation is the following
Typically the size of the training data set is larger than the q
X
size of the validation set. In the next subsections, we propose CLWLS (d, λ) = argmin γi (T , d, λ)||pi − C · di ||2
C
several approaches to perform the estimation in (5). i=1
= P0(T , d, λ)DT (D0(T , d, λ)DT )−1 , (12)
B. REGULARIZED LEAST SQUARES (RLS)
DAYLIGHT ESTIMATOR
where P and D are defined as above, γi (T , d, λ) ≥ 0, ∀i and
0(T , d, λ) = diag(γ1 (T , d, λ), . . . , γq (T , d, λ)) ∈ Rq×q is
We now consider estimation of a static linear mapping
referred as a diagonal weight matrix where the weight are
C ∈ RN ×M between the daylight measured at the ceiling and
computed as follows:
the daylight measured at the workplace plane for any time  
instant k, i.e. kd − di k
γi (T , d, λ) = g , (13)
h(T , d, λ)
p(k) = f (d(k)) = Cd(k).
b (7)
where g : R+ → R+ is a weight function,
Using a least-squares (LS) approach [24], C in (7) is
given as the solution of a quadratic optimization problem that h(T , d, λ) = max kd − di k, (14)
i∈N (T ,d,λ)
minimizes the sum of the square of the residual between the
measured illuminance vector at the workplace plane pi and and N (T , d, λ) is a subset of T containing the q̂ closest
the model estimation b pi = Cdi , i.e. neighbors to d such that qq̂ = λ.
Note that the diagonal matrix 0(T , d, λ) selects a frac-
q
X tion λ of measurements from the set T that are closer to
CLS = argmin ||pi − Cdi ||2 (8)
C input vector d; these measurements are assigned a weight
i=1
defined by function g(·). The intuition behind this choice is
where || · || is the `2 norm. If the inverse of DDT exists, then that the estimated workplace luminance b p will be close to the
the closed-form solution to (8) is given by observed workplace luminance b pi whose corresponding ceil-
ing measurement di is close to the current measured value d.
CLS = PDT (DDT )−1 , In particular, if the n-th diagonal entry in matrix 0(·) is close

21464 VOLUME 5, 2017

S. Borile et al.: Data-Driven Daylight Estimation Approach to Lighting Control

to 1, then the n-th measurement in set T will have a higher Let the centroid of the `-th cluster be given by
importance during the estimation of matrix C LWRLS .
[`] 1 X
The weight function g(·) can be any function that satisfies d = di .
the properties as discussed in [31]. A standard choice of g for q`
(di ,pi )∈T`
locally weighted regression problems is the tri-cube weight
function [31]: For the `-th cluster the corresponding regularized regres-
[`]
( sion matrix CCRLS is calculated as follows

3
1 − |x|3 |x| ≤ 1 [`]
X
g(x) = CCRLS (` ) = argmin ||pi − Cdi ||2 + ` kCk2F ,
0 |x| > 1. C (di ,pi )∈T`
The parameter λ ≥ 0 is chosen by computing the best
where ` is optimized using the data in the validation set, i.e.
performance on the validation set
[`]
X
r `∗ = argmin ||pi − CCRLS ()di ||2
≥0
X
λ∗ = argmin ||pi − CLWLS (di , λ)di ||2 . (15) (di ,pi )∈V`
λ≥0 i=q+1
and
Similarly to the LS problem, it is convenient to regularize [`] [j]
V` = {(di , pi ) : kdi − d k < kdi − d k , ∀j 6 = `}.
the problem to obtain a Locally Weighted Regularized Least
Square problem (LWRLS): When a new measurement vector at the light sensors at
the ceiling d is available, the corresponding estimation at the
CLWRLS (d, , λ) = P0(T , d, λ)D (D0(T , d, λ)D + I )
T T −1
.
workplace b p is calculated as
(16)
[`∗ (d)]
p = CRLS d,
In this case  and λ need to be chosen by computing the
b

best performance on the validation set: where

r [`]
X `∗ (d) = argmin{kd − d k}L`=1
( ∗ , λ∗ ) = argmin ||pi − CLWRLS (di , , λ)di ||2 . (17) `
λ≥0,≥0 i=q+1
is the cluster with the lowest Euclidean distance between its
D. CLUSTER-BASED REGULARIZED LEAST centroid and d.
SQUARES (CRLS) DAYLIGHT ESTIMATOR This approach is computationally more efficient than the
The LWRLS has the potential to provide better performance LWRLS since the computation of the clusters and the matri-
[`]
than standard RLS. The main disadvantage is that at any ces CCRLS along with the optimal regularization parameters
time instant k, when a new measurement d is available, a is performed off-line. Only the computation of the closest
[`∗ (d)]
new matrix C should be calculated using (16). This can be cluster `∗ (d) and the RLS estimation step bp = CCRLS d are
computationally expensive if the training data set T is large. computed on-line.
A possible solution is to compute off-line a set of different Note that the number of clusters needs to be cho-
estimated transfer matrices C which are constant within a sen depending on the size of the training data set. We
certain domain of the ceiling measurement space d. To do consider only the Cluster-based regularized Least Squares
this, during the preliminary training phase, a clustering of because of the overfitting problem: the size of the data set
q [`]
the measurements at the ceiling {di }i=1 is performed. We used to compute CCRLS is smaller than all the previous
define L clusters and for the `-th cluster a RLS estimation approaches.
[`]
matrix CCRLS based only on the data belonging to the `-th
cluster is obtained. A convenient metric, used for the cluster- IV. LIGHTING CONTROL
ing, is the Euclidean distance. In this section we consider two different type of con-
When a new measurement data d becomes available, we trollers: (A) ceiling-based control and (B) workplace-based
find the closest cluster centroid to this measurement and then control.
the corresponding RLS estimation b p is obtained using the
[`]
matrix CCRLS . More specifically, we divide the data set T into A. CEILING-BASED CONTROL
L subsets T` with possibly non-uniform sizes r` , i.e. In ceiling-based control, at each time instant k, the lighting
system is adapted based on:
L
X • daylight contributions at each light sensor at the ceiling,
T = ∪L`=1 T` , |T` | = ri , r` = q.
d(k);
`=1
• illuminance gain matrices for the ceiling, G; and
This can be obtained by applying the k-mean++ algo- • desired illuminance levels at each light sensor at the
q
rithm [32] using the measurements at the ceiling {di }i=1 . ceiling, yr (k).

VOLUME 5, 2017 21465

 S. Borile et al.: Data-Driven Daylight Estimation Approach to Lighting Control

In [17], the ceiling-based control was formulated as an behavior. As such, with knowledge of the daylight contribu-
optimization problem given by tion estimate bp(k) and illuminations gains, H , at the work-
place, the optimization problem (19) can be solved to obtain
u∗ (k) = arg min 1T u the optimal dimming level u∗ (k) and applied to the LEDs to
u achieve the desired value Wr . To ensure smooth changes in
y(k) ≥ yr (k)

dimming levels such that the user is not disturbed by dimming
s.t. y(k) = Gu + d(k) (18) control, methods considered in [34] may be used.

0 ≤ u ≤ 1. We now consider a way for the controller to deal with

errors in daylight estimation. In particular, we consider
The ceiling-based control given by (18) is hereafter over-estimation errors since our main concern is the under-
referred to as REF_CONTROL. In [17], this approach was illuminated workplaces. The optimization problem (19) may
shown to result in a lower energy consumption compared be modified as follows:
to [11], while still achieving the light sensor set-points. Hence
we use this as a reference method for performance compari- u∗ (k) = arg min 1T u
son. u
The ceiling sensor references, yr (k), are determined in w(k) ≥ min{wmax , wr (k) + ξ }

a preliminary night time calibration phase as explained s.t. w(k) = Hu + b
p(k) (20)

0 ≤ u ≤ 1,

in [11] and [21]. In the absence of daylight, the luminaries are
turned to maximum intensity and both the average workplace where wmax is a vector with the maximum achievable illumi-
plane illuminance value and the ceiling sensor measurements nance level at each control point due to artificial light and ξ is
are saved. The ceiling sensor references are then obtained by a offset vector to be determined. The purpose of adding this
suitable scaling to result in the specified reference average offset vector is to provide some robustness against errors in
illuminance at the workplace plane. It is assumed that the daylight estimation. In section V-B, we will explain how ξ is
reference values are feasible, i.e. G1 ≥ yr (k) which implies computed. The improved control given by (20) is hereafter
that there is a set of dimming values u(k) (in the most extreme referred to as WP_CONTROL with Offset, regardless of the
scenario u(k) = 1), that ensures that the illuminance on the estimation method used.
ceiling is no smaller than the reference illuminance. We do not consider adding an offset to the target level for
the REF_CONTROL because in practice we do not know
B. WORKPLACE-BASED CONTROL the offset vector to be added to the target levels at the light
In workplace-based control, at each time instant k, the sensors at the ceiling such that the target illumination at the
lighting system is adapted based on: workspace is achieved.
• daylight contributions at each control point at the work-
place plane, p(k) (or estimates b
p(k)); V. SIMULATIONS RESULTS
• illuminance gain matrices for the workplace, H ; and A. LIGHTING DATA SET
• desired illuminance levels at each control point at the The open-plan office lighting model considered in [11] was
workplace plane, wr (k). used, with the lighting plan depicted in Figure 2. The office
In this paper, we propose the following optimization has length 24 m, width 19 m and height of the ceiling is 2.6 m.
problem for workplace-based control: There are M = 80 luminaires with collocated sensors in a
grid of 10 by 8, and N = 36 workplaces with a collocated
u∗ (k) = arg min 1T u light sensor. The additional light sensors at control points at
u the workplaces were used for data acquisition as explained in
w(k) ≥ wr (k)
 Section III-A. The windows are located on the right side of
s.t. w(k) = Hu + b p(k) (19) the office next to luminaires 71-80; hence the biggest contri-
bution of the daylight is observed in this area. All the artificial

0 ≤ u ≤ 1,

light and daylight distributions were obtained from the office
where b p(k) is the estimate of daylight contribution at the model implemented in lighting software DIALux [38]. The
workplace plane obtained using the estimation approaches lighting control system was implemented in Matlab.
described in Section III. It is assumed that the references are Data from days in different months (January, March, June,
feasible, i.e. H 1 ≥ wr (k). The workplace-based control given August, September and December) and with different sky
by (19) is hereafter referred to as WP_CONTROL, regardless conditions (clear sky, overcast sky and mixed sky) was col-
of the estimation method used. lected at a 15 minute interval. The daylight distributions in
The optimization problems given by (18) and (19) are lin- the office were simulated from 7:00 AM to 7:45 PM for a
ear programming problems. These can be solved, for exam- total of 18 days spread across all four seasons and with the
ple, with the simplex method, interior-point algorithms or three different sky conditions. The data from these 18 days
variants [33]. In this paper, we focus our attention on the was divided into 12 days for the training set T and 6 days for
steady-state behavior and do not consider the entire control the validation set V.

21466 VOLUME 5, 2017

S. Borile et al.: Data-Driven Daylight Estimation Approach to Lighting Control

B. MODEL VALIDATION AND PERFORMANCE METRICS

We consider the following metrics for evaluating the perfor-
mance of each estimator described in Section III with respect
to the validation data set V:
• root mean squared error (RMSE) for each control
point n,
s
1 X 2
RMSEn = [ei ]n , (21)
(r − q)
i∈V
where ei = pi − b pi is the vector containing the error in
daylight estimation at the control points at the workplace
plane;
• root average mean squared error (RAMSE) over all
control points,
v
u
u1 X N FIGURE 3. CDF of the daylight estimation error on workplaces for the
RAMSE = t (RMSEn )2 ; (22) considered estimators.
N
n=1
• cumulative distribution function (CDF) of the error in
daylight estimation for each control point n,
1 X
CDFn (e) = 1(e − [ei ]n ), (23)
r −q
i∈V
where 1(x) is the unit step function which is equal to
zero for x < 0 and equal to one for x > 0; and
• CDF of the error in daylight estimation over all control
points,
N
1 X
CDF(e) = CDFn (e). (24)
N
n=1
Using (23), we choose the bias term in (20) as
[ξ ]n = max{0, −αn }, n = 1, . . . , N (25)
FIGURE 4. RMSE of the daylight estimation error for each workplace as
where CDFn (αn ) = 1 − 4 and 0 ≤ 4 ≤ 1 is a parameter that given by (21). LS, LWRLS and CRLS are considered.
indicates the proportion of samples to be compensated during
offset control in (20). We choose 4 = 0.99.
Figure 3 shows the CDF for the proposed daylight esti-
mation methods (RLS, LWRLS and CRLS) over all control
points as given by (24). We can see that the estimation errors
for the three proposed approaches are concentrated within an
error bound of ±20 lux. The error is rather small (less than
5%) as compared to the desired illuminance at the workplace
plane which is around 500 lux. Figure 4 gives a more detailed
insight on how these errors (using (21)) are distributed across
the N = 36 different workplaces (control points). It clearly
shows that the larger errors are on those workplaces near the
windows (29 - 36). This is expected since those are the work-
places over which daylight exhibits the largest variation (up
to several thousands lux as seen in Figure 5). The workplaces
that are further away from the windows (1 - 28) receive small
amounts of daylight even on a bright day as seen in Figure 5. FIGURE 5. Boxplot showing the daylight value at workplace level for our
data set.
For our data set, all approaches (RLS, LWRLS and CRLS)
achieve similar performance for estimating daylight contribu-
tion at control points over the workplace plane. The RAMSE C. SATURATION OF LIGHT SENSORS
as given by (22) is about 3.732 for RLS, 3.004 for LWRLS For most lighting control applications, if the illuminance level
and 3.229 for CRLS. due to daylight is larger than several hundred lux, then it

VOLUME 5, 2017 21467

 S. Borile et al.: Data-Driven Daylight Estimation Approach to Lighting Control

would result in the luminaire dimming down to minimum divides the training set into smaller sets and thus it is not
intensity. Hence, in practice, light sensors are designed to recommended when the training set is small.
measure a limited range of illuminance levels [39], [40].
An illuminance level larger than the maximum range of a
light sensor would result in saturation, i.e. the measurement
is capped to the maximum valid value.
In this section, we compare the performance of the esti-
mators under saturation of the light sensors. We consider a
saturation level for the light sensors of 1000 lux.
The percentage of the samples that are affected by satu-
ration is zero for the data originating from the light sensors
at the ceiling. The percentage of the samples affected by
saturation from the workplace plane is around 2%; if we
consider just the workplaces next to the windows the percent-
age is about 8%. In Figure 5, we plot the daylight values at
each control point at the workplace plane without saturation.
We can see that those control points in the proximity of the
windows (29 to 36) have the largest daylight contribution and
thus they are more likely to be affected by saturation. Hence, FIGURE 7. CDF of the daylight estimation error on all workplaces using
we focus our comparison on only those control points. only two training days: overcast day in September and December. No
saturation.

In Figure 7, we show the CDF of the daylight estimation

error when the training set is small. We consider a training
set with only two days with small variations in daylight condi-
tions: Overcast Sky days in September and December. We can
see that all approaches overestimate the daylight contribution
at control points at the workplace plane. In particular, the RLS
approach with  = 0 overestimates more often (around 15%
of the samples are overestimated by more than 50 lux) than
the other two approaches (less than 10% of the samples are
overestimated by more than 50 lux).
Additionally, we show in Figure 8 the CDF of the daylight
estimation error when the training set includes two days
with large variations in daylight conditions: a single day in
September with clear sky and another single day in December
FIGURE 6. CDF of the daylight estimation error on the workplaces 29 to with overcast sky. We can see that the RLS and LWRLS
36 for the considered estimators. All sensors are saturated at 1000 lux.

In Figure 6, we show the CDF of the daylight estimation

error for all approaches under saturation of light sensors. Note
that the saturation of light sensors introduces a non-linearity
in the relationship between daylight levels at the ceiling and
at the workplace plane. We can see that the performance
of the RLS approach is degraded due to this non-linearity.
In comparison, the LWRLS and CRLS approaches provide a
good estimate of daylight under saturation of light sensors.

D. SMALL TRAINING SET

In Section III, we discussed that adding a regularization
parameter could be effective if the training data set is small.
In fact, adding the regularization term when we use the entire
training set has little effect on the estimation performance.
Using a small training set, we compare the performance of
the LWRLS and RLS approach. For comparison, we also
FIGURE 8. CDF of the daylight estimation error on all workplaces using
consider the performance of a LS approach, i.e. RLS with only two training days: clear day in September and overcast day in
 = 0. Note that the cluster-based approach would further December. No saturation.

21468 VOLUME 5, 2017

S. Borile et al.: Data-Driven Daylight Estimation Approach to Lighting Control

approaches provide good estimates of the daylight contribu- The WP_CONTROL with Offset provides most of the time
tion while the RLS approach with  = 0 underestimates sufficient illumination at the control points at the workplace
the daylight contribution (around 20% of the samples are plane. Note that the ORACLE_CONTROL always provides
underestimated by more than 50 lux). sufficient illumination.

E. CONTROLLER PERFORMANCE
In this section we evaluate the performance of the differ-
ent controllers described in Section IV using the validation
set V. For the WP_CONTROL and WP_CONTROL with
Offset we consider the LWRLS approach for daylight esti-
mation due to its good performance under large and small
training set, and saturation of light sensors. For comparison,
we consider the WP_CONTROL when the daylight contribu-
tion terms p(k) are perfectly known, hereafter referred to as
ORACLE_CONTROL, and the REF_CONTROL introduced
in (18).
The simulations were done by having all workplaces
in occupied state, which requires an illuminance level at
workplace plane of W = 500 lux in all the workplaces,
i.e. wr (k) = W 1, ∀k.
FIGURE 10. Normalized energy consumption averaged over several days
of the years and weather conditions for different control strategies.

Figure 10 shows the normalized energy consumption aver-

aged over several days of the years and daylight conditions
for each control strategies. The energy consumption is nor-
malized with respect to a lighting system that, under the
absence of daylight, provides an average illumination at the
workplace plane of about W = 500 lux, i.e. all luminaires
are set to a constant
P value of 0.85 throughout the day: u(k) =
0.85 × 1 and N1 N n=1 [w(k)]n = W when d(k) = 0. We
can see that our proposed control strategies have reduced and
similar normalized energy consumption levels with respect to
REF_CONTROL and ORACLE_CONTROL, respectively.

TABLE 1. Normalized energy consumption and percentage of workplaces

that are under-illuminated by more than 1 lux, under different control
FIGURE 9. CDF of the controller error on the workplaces for different
strategies.
approaches with respect to the nominal value of 500 lux as given
by Eqn. (26).

In Figure 9, we show the CDF of the error in achieved

illumination at the control points at the workplace plane with
respect to the nominal value of W = 500 lux for each control
strategy. More specifically, we show for each control strategy:
N K
1 XX In Table 1, we summarize both the average normalized
CDFw (e) = 1(e − [ew (k)]n ),
NK lighting energy consumption and the level of under-
n=1 k=1
illumination. Here, we can see that our proposed con-
ew (k) = w(k) − wr (k). (26)
trol strategies are close in performance to the
In Figure 9, we can see that the proposed controllers ORACLE_CONTROL.
using the daylight estimation approach provides sufficient Finally, we simulate the effect of different occupancy dis-
illumination at the control points at the workplace plane tributions. We consider a scenario wherein probability of
(the under-illumination is less than 10 lux). In compari- a workplace being occupied is equal to 0.5. In Fig. 11,
son, the REF_CONTROL has a larger variation in the error we can see that the proposed controllers outperform the
in achieved illumination at the control points at the work- REF_CONTROL with respect to providing sufficient illumi-
place plane (half of the time the control points are under- nation at the workplaces under different occupancy distribu-
illuminated and the other half they are over-illuminated). tions. Note that the proportion of under-illuminated zones

VOLUME 5, 2017 21469

 S. Borile et al.: Data-Driven Daylight Estimation Approach to Lighting Control

[6] M.-S. Pan, L.-W. Yeh, Y.-A. Chen, Y.-H. Lin, and Y.-C. Tseng, ‘‘A WSN-
based intelligent light control system considering user activities and
profiles,’’ IEEE Sensors J., vol. 8, no. 10, pp. 1710–1721, Oct. 2008.
[7] L.-W. Yeh, C.-Y. Lu, C.-W. Kou, Y.-C. Tseng, and C.-W. Yi,
‘‘Autonomous light control by wireless sensor and actuator networks,’’
IEEE Sensors J., vol. 10, no. 6, pp. 1029–1041, Jun. 2010.
[8] Y.-J. Wen and A. M. Agogino, ‘‘Personalized dynamic design of net-
worked lighting for energy-efficiency in open-plan offices,’’ Energy
Build., vol. 43, no. 8, pp. 1919–1924, 2011.
[9] Y.-J. Wen and A. M. Agogino, ‘‘Control of wireless-networked lighting
in open-plan offices,’’ Lighting Res. Technol., vol. 43, no. 2, pp. 235–248,
Nov. 2010.
[10] M. T. Koroglu and K. M. Passino, ‘‘Illumination balancing algorithm
for smart lights,’’ IEEE Trans. Control Syst. Technol., vol. 22, no. 2,
pp. 557–567, Mar. 2014.
[11] N. van de Meugheuvel, A. Pandharipande, D. Caicedo, and
P. P. J. van den Hof, ‘‘Distributed lighting control with daylight
and occupancy adaptation,’’ Energy Build., vol. 75, pp. 321–329,
Jun. 2014.
FIGURE 11. CDF of the controller error on the workplaces for different [12] S. Afshari and S. Mishra, ‘‘A plug-and-play realization of decen-
approaches under different occupancy distributions as given by Eqn. (26).
tralized feedback control for smart lighting systems,’’ IEEE Trans.
Control Syst. Technol., vol. 24, no. 4, pp. 1317–1327, Jul. 2016,
doi: 10.1109/TCST.2015.2487880.2015.
under REF_CONTROL is smaller when compared to all
[13] M. Fischer, K. Wu, and P. Agathoklis, ‘‘Intelligent illumination model-
zones occupied (Figure 9). This is due to a lower illumination based lighting control,’’ in Proc. Int. Conf. Distrib. Comput. Syst. Work-
requirement (300 lx) in unoccupied zones. Similarly, we can shops, 2012, pp. 245–249.
see that a larger proportion of zones are over-illuminated [14] M. Miki, A. Amamiya, and T. Hiroyasu, ‘‘Distributed optimal control of
lighting based on stochastic hill climbing method with variable neigh-
under all methods. borhood,’’ in Proc. IEEE Int. Conf. Syst., Man, Cybern., Oct. 2007,
pp. 1676–1680.
VI. CONCLUSIONS [15] S. Afshari and S. Mishra, ‘‘Decentralized feedback control of smart
lighting systems,’’ in Proc. ASME Dyn. Syst. Control Conf., 2013,
We considered three methods for daylight mapping estima- pp. V001T13A006–V001T13A006.
tion in a training phase for use in lighting control. The day- [16] S. Afshari and S. Mishra, ‘‘An optimization framework for control of non-
light mapping is used to obtain an estimate of the achieved square smart lighting systems with saturation constraints,’’ in Proc. Amer.
Control Conf., Jul. 2015, pp. 1665–1670.
illuminance at the workplaces. This knowledge is used in the [17] M. Rossi, A. Pandharipande, D. Caicedo, L. Schenato, and A. Cenedese,
control law to obtain the dimming levels of the luminaires ‘‘Personal lighting control with occupancy and daylight adaptation,’’
and adapt the artificial light output to changing daylight con- Energy Build., vol. 105, pp. 263–272, Oct. 2015.
[18] D. Caicedo and A. Pandharipande, ‘‘Distributed illumination control with
ditions. We show that in comparison to the REF_CONTROL local sensing and actuation in networked lighting systems,’’ IEEE Sensors
approach, the proposed solution achieves illuminance values J., vol. 13, no. 3, pp. 1092–1104, Mar. 2013.
closer to the desired values and also saves energy. In practice, [19] A. Peruffo, A. Pandharipande, D. Caicedo, and L. Schenato, ‘‘Lighting
control with distributed wireless sensing and actuation for daylight and
the least squares method for daylight estimation has worse occupancy adaptation,’’ Energy Build., vol. 97, pp. 13–20, Jun. 2015.
performance among the considered methods. In particular, [20] C. Branas, F. J. Azcondo, and J. M. Alonso, ‘‘Solid-state lighting: A sys-
the performance is poor when (i) the training set is small, tem review,’’ IEEE Ind. Electron. Mag., vol 7, no. 4, pp. 6–14, Dec. 2013.
and/or (ii) when non-linearities, for example in the form [21] A. Pandharipande and D. Caicedo, ‘‘Smart indoor lighting systems
with luminaire-based sensing: A review of lighting control approaches,’’
of saturated light sensor measurements, exist. The former Energy Build., vol. 104, pp. 369–377, Oct. 2015.
issue can be solved by considering a regularization term. [22] F. Rubinstein, ‘‘Photo-electric control of equi-illumination lighting sys-
The latter issue can only be handled by the locally weighted tems,’’ Energy Build., vol. 6, pp. 141–150, Feb. 1984.
[23] F. Rubinstein, G. Ward, and R. Verderber, ‘‘Improving the performance
regularized least squares and cluster-based regularized least of photo-electrically controlled lighting systems,’’ J. Illuminat. Eng. Soc.,
squares methods. Note that both these methods assume lin- vol. 18, no. 1, pp. 70–94, 1989.
earity within a (neighboring) subset of the measurements. [24] S. Borile, A. Pandharipande, D. Caicedo, A. Cenedese, and L. Schenato,
‘‘An identification approach to lighting control,’’ in Proc. Eur. Control
Conf., Jul. 2016, pp. 637–642.
REFERENCES [25] T. Hastie, R. Tibshirani, and J. Friedman, The Elements of Statistical
[1] ‘‘Commercial buildings energy consumption survey,’’ Energy Inf. Admin. Learning: Data Mining, Inference, and Prediction, 2nd ed. New York,
Washington, DC, USA, Tech. Rep., 2003. NY, USA: Springer-Verlag, 2009.
[2] O. T. Masoso and L. J. Grobler, ‘‘The dark side of occupants’behaviour [26] J. A. Veitch, G. R. Newsham, P. R. Boyce, and C. C. Jones,
on building energy use,’’ Energy Buil., vol. 42, no. 2, pp. 173–177, 2010. ‘‘Lighting appraisal, well-being and performance in open-plan offices:
[3] A. Williams, B. Atkinson, K. Garbesi, and F. Rubinstein, ‘‘A meta- A linked mechanisms approach,’’ Lighting Res. Technol., no. 40, no. 2,
analysis of energy savings from lighting controls in commercial build- pp. 133–151, 2008.
ings,’’ LEUKOS, J. Illuminat. Eng. Soc. North Amer., vol. 8, no. 3, [27] G. R. Newsham, M. B. C. Aries, D. Mancini, and G. Faye, ‘‘Individual
pp. 161–180, 2013. control of electric lighting in a daylit space,’’ Lighting Res. Technol.,
[4] D. Caicedo, A. Pandharipande, and G. Leus, ‘‘Occupancy-based illumi- vol. 40, no. 1, pp. 25–41, 2008.
nation control of LED lighting systems,’’ Lighting Res. Technol., vol. 43, [28] ‘‘Light and lighting. Lighting of work places. Indoor work places,’’
no. 2, pp. 358–376, 2010. Eur. Committee Standardization, Tech. Rep. E. 12464-1, 2011.
[5] A. Pandharipande and D. Caicedo, ‘‘Daylight integrated illumination [29] Z. Wang, and K. T. Yen , ‘‘Illumination control of LED systems based on
control of LED systems based on enhanced presence sensing,’’ Energy neural network model and energy optimization algorithm,’’ Energy Build.,
Build., vol. 43, no. 4, pp. 944–950, 2011. vol. 62, pp. 514–521, Jul. 2013.

21470 VOLUME 5, 2017

S. Borile et al.: Data-Driven Daylight Estimation Approach to Lighting Control

[30] L. Ljung, System Identification. Boston, MA, USA: Birkhäuser, 1998, DAVID CAICEDO received the Engineer degree in
pp. 13–14. electronics and telecommunications from ESPOL,
[31] W. S. Cleveland, ‘‘Robust locally weighted regression and smoothing Guayaquil, Ecuador, in 2006, the M.S. (cum
scatterplots,’’ J. Amer. Statist. Assoc., vol. 74, no. 368, pp. 829–836, 1979. laude) degree in electrical engineering, with spe-
[32] D. Arthur and S. Vassilvitskii, ‘‘K-means++: The advantages of care- cialization in the field of Telecommunications,
ful seeding,’’ in Proc. ACM-SIAM Symp. Discrete Algorithms, 2007, from the Delft University of Technology, Delft,
pp. 1027–1035. The Netherlands, in 2010, and the Ph.D. degree
[33] S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge, U.K.:
in electrical engineering from the Eindhoven Uni-
Cambridge Univ. Press, 2004.
versity of Technology, Eindhoven, The Nether-
[34] D. Caicedo and A. Pandharipande, ‘‘Daylight and occupancy adaptive
lighting control system: An iterative optimization approach,’’ Lighting lands, in 2014. As part of his engineering program,
Res. Technology, vol. 48, no. 6, pp. 661–675, 2015. he completed an internship at Nokia in 2006, Guayaquil, Ecuador. From
[35] D. H. Li, C. S. Lau, and J. C. Lam, ‘‘Predicting daylight illuminance by 2007 to 2008, he was with Nokia Siemens Networks as a Radio Frequency
computer simulation techniques,’’ Lighting Res. Technol., vol. 36, no. 2, Engineer, Bogota, Colombia. Since 2009, he has been with Philips Lighting
pp. 113–128, 2004. Research, Eindhoven, The Netherlands, working on his master’s thesis, from
[36] P. R. Tregenza and I. M. Waters, ‘‘Daylight coefficients,’’ Lighting Res. 2009 to 2010, on a van der Pol PhD position till 2014, and since 2014, as a
Technol., vol. 52, no. 2, pp. 65–71, 1983. Research Scientist. His research interests are in the areas of sensor signal
[37] J. Benesty, J. Chen, Y. Huang, and I. Cohen, ‘‘Pearson correlation coef- processing algorithms and applications, mobile wireless communications,
ficient,’’ in Noise Reduction in Speech Processing. Berlin, Germany: and wireless networking.
Springer-Verlag, 2009, Ch. 5.
[38] DIAL. Gmbh DIALux 4.11. [Online]. Available: http://www.dial.de/
DIAL/en/dialux/download.html
[39] Leviton. PCIND-000-PCIND-0SV. [Online]. Available:
http://www.leviton.com/OA_HTML/ProductDetail.jsp?partnumber
=PCOUT-SV&section=38452&minisite=10251
[40] Lutron. LRFX-DCRB. [Online]. Available: http://www.lutron.com/
TechnicalDocumentLibrary/3683587_Daylight_Sensor_Design_and_ LUCA SCHENATO (F’17) received the Dr. Eng.
App_Guide_sg.pdf degree in electrical engineering from the Univer-
sity of Padova in 1999 and the Ph.D. degree in elec-
trical engineering and computer sciences from UC
Berkeley, in 2003. He held a post-doctoral position
and a visiting professor position with U.C. Berke-
STEFANO BORILE was born in Este, in 1990. ley, in 2004 and from 2013 to 2014, respectively.
He received the bachelor’s degree in information He is currently an Associate Professor with the
engineering and the master’s degree in automation Information Engineering Department, University
engineering from the University of Padova, Italy, of Padova. His interests include networked control
in 2012 and 2015, respectively. In 2015, he was an systems, multi-agent systems, wireless sensor networks, smart grids and
Intern with Philips Lighting Research, Eindhoven, cooperative robotics. He has been awarded the 2004 Researchers Mobility
working on his master’s thesis. Fellowship by the Italian Ministry of Education, University and Research,
the 2006 Eli Jury Award in U.C. Berkeley, and the EUCA European Control
Award in 2014. He served as Associate Editor of the IEEE TRANSACTIONS ON
AUTOMATIC CONTROL from 2010 to 2014.

ASHISH PANDHARIPANDE (SM’08) received

the B.E. degree in electronics and communications
engineering from Osmania University, Hyderabad,
India, in 1998, and the M.S. degrees in electrical
and computer engineering and mathematics and ANGELO CENEDESE (M ’11) received the M.S.
the Ph.D. degree in electrical and computer engi- and the Ph.D. degrees from the University of
neering from the University of Iowa, Iowa City, in Padova, Italy, in 1999 and 2004, respectively.
2000, 2001, and 2002, respectively. In 2003, he He is currently an Associate Professor with the
was a Post-Doctoral Researcher with the Univer- Department of Information Engineering, Univer-
sity of Florida. From 2004 to 2006, he was with sity of Padova. He has held several visiting posi-
the Samsung Advanced Institute of Technology as a Senior Researcher. tions with international Universities and Research
He has held visiting positions with AT&T Laboratories, the Department of Centers, including UKAEA-JET laboratories,
Electrical Communication Engineering, Indian Institute of Science, India, Culham Research Center, U.K. from 2002 to 2004,
and Nanyang Technological University, Singapore. Since 2006, he has been the JAERI Institute, Japan in 2000, and General
a Senior Scientist with Philips Lighting Research, and now Philips Lighting, Atomics, CA, in 2004, the UCLA Vision Laboratory, CA, USA, in 2010,
Eindhoven, Netherlands. He has authored or co-authored over 160 scientific the F4E European Agency, Spain, in 2014. He has been involved and he is
publications and about 80 patents/application filings in these areas. His currently involved in several projects on distributed systems for monitoring
research interests are in data analytics, sensing, networking and controls, and and localization (sensor/actor, multiagent, and camera networks), control
system applications in domains, such as connected smart lighting, energy of complex systems (adaptive optics systems, fusion devices, aerospace
monitoring and control, and cognitive spectrum sharing. systems), funded by European and Italian government institutions and
Dr. Ashish is currently an Associate Editor of the IEEE TRANSACTIONS ON industries, with different roles of participant and/or principal investigator.
SIGNAL PROCESSING, the IEEE SIGNAL PROCESSING LETTERS, the IEEE SENSORS His research interests include system modeling, control theory and its
JOURNAL, an Editor of the EURASIP Journal on Wireless Communica- applications, sensor and actuator networks, and multiagent systems. On these
tions and Networking, and a member of the International Advisory Board, subjects, he has authored over 100 papers and holds three patents.
Lighting Research & Technology.

VOLUME 5, 2017 21471