See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/292946065

Advanced occupancy sensing for energy

efficiency in office buildings
Article in Proceedings of the Institution of Mechanical Engineers Part I Journal of Systems and Control
Engineering May 2016
DOI: 10.1177/0959651815625001

CITATIONS

READS

24

4 authors, including:
Tobore Ekwevugbe

Neil Brown

De Montfort University

De Montfort University

6 PUBLICATIONS 29 CITATIONS

71 PUBLICATIONS 303 CITATIONS

SEE PROFILE

SEE PROFILE

Vijay Pakka
De Montfort University
19 PUBLICATIONS 164 CITATIONS
SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Agent-based Modelling of Electricity Networks View project

All content following this page was uploaded by Vijay Pakka on 19 December 2016.
The user has requested enhancement of the downloaded file. All in-text references underlined in blue are added to the original document
and are linked to publications on ResearchGate, letting you access and read them immediately.

Original article

Advanced occupancy sensing for

energy efficiency in office buildings

Proc IMechE Part I:

J Systems and Control Engineering
2016, Vol. 230(5) 410423
IMechE 2016
Reprints and permissions:
sagepub.co.uk/journalsPermissions.nav
DOI: 10.1177/0959651815625001
pii.sagepub.com

Tobore Ekwevugbe1, Neil Brown1, Vijayanarasimha Pakka1 and Denis Fan2

Abstract
Control systems for heating, ventilation and air conditioning in non-domestic buildings often operate to fixed schedules,
assuming maximum occupancy during business hours. Since lower occupancies usually mean less demand for heating,
ventilation and air conditioning, energy savings could be made. Air quality sensing, often combined with temperature sensing, has performed sufficiently in the past for this if maintained properly, although sensor and control failures may
increase energy use by as much as 50%. As energy costs increase, coupled with increased complexity in building services
and reduced commissioning time, all placing ever higher demands on sensing, building controls must meet increasingly
stringent environmental requirements, whilst also improving reliability. Sensor fusion offers performance and resilience
to meet these demands, while cost and privacy are key factors which are also met. This article describes a neural network approach to sensor fusion for occupancy estimation. Feature selection was carried out using symmetrical uncertainty analysis, while fusion of sensor features used a back-propagation neural network, with occupant count accuracy
exceeding 74%.

Keywords
Sensors/sensor applications, intelligent sensing, neural networks, software engineering, system integration, fuzzy logic
control, sensor fusion

Date received: 23 March 2015; accepted: 7 December 2015

Introduction
Following the Kyoto Agreement, the UK government
adopted a target to reduce CO2 emissions by 60%, of
the current levels (1990s), by 2050. Achieving this target
is a key policy goal for the government, and energy efficiency is expected to play a significant role in achieving
this.1 Worldwide building energy use is expected to
grow 45% over the next 20 years2 and buildings use up
to 40% of total energy use in many countries including
the United Kingdom, but much energy is wasted servicing unoccupied buildings: In the United Kingdom, up
to 23%30% of service sector electricity use can be
from unoccupied lighting3 while 39% of US domestic
building energy wasted due to unoccupied heating and
cooling.4 With roughly 50% of building energy use due to
heating, ventilation and air conditioning system (HVAC),5
there is clearly massive potential for energy savings;
indeed, improvements in HVAC operation and management can lead to energy savings approaching 56%.6,7
People spend about 80% of their time indoors, and
so any building energy management system (BEMS)
must provide comfort while using energy efficiently.
While badly controlled HVAC can mean problems

ranging from minor discomfort through to illness,

improved building controls can mean improved health,
well-being and productivity. Improving HVAC system
operations can contribute to reductions in building
energy use.6,7 Beyond automatic controls, for successful
building and energy management, a reliable occupancy
count also becomes invaluable.

Building occupancy detection

Conventional occupancy detection systems have several
short-comings; passive infra-red (PIR) sensors used the
most commonly for lighting control,8 less so for
HVAC, may fail to detect stationary occupants,
1

Institute of Energy and Sustainable Development, Faculty of Technology,

De Montfort University, Leicester, UK
2
Technology Centre, E.ON Technologies (Ratcliffe) Limited, Nottingham,
UK
Corresponding author:
Neil Brown, Institute of Energy and Sustainable Development, Faculty of
Technology, De Montfort University, Leicester LE1 9BH, UK.
Email: nbrown@dmu.ac.uk

Ekwevugbe et al.
switching off services (notably lighting) falsely. One
approach is also to measure sound levels: for instance,
Padmanabh et al.9 used a combination of microphones
and PIR sensors to gather occupancy information for
conference rooms. The room was considered occupied
(meeting ongoing) if a microphone value exceeded a
threshold twice in a 5 min interval. Jianfeng et al.10
built a standalone novelty system, based on microphone and PIR real-time sound events, using a hidden
Markov model to 87% accuracy. Both systems though
were prone to external audio interference. Dodier
et al.11 proposed a Bayesian belief network (BBN) of
three PIR sensors and a telephone sensor to probabilistically infer occupancy, of an individual office, with
accuracy reaching 76%. In another study, Zhao et al.12
also utilised a BBN for information fusion from various types of occupancy-related measurements for realtime occupancy detection at room level and working
zone level. Sensors deployed for the room level occupancy measurement comprises of PIR sensor, door sensor, light switch sensor, chair sensor, keyboard and
mouse. For the working zone level, real-time GPS location and Wi-Fi connection from smart device like smart
phones and occupancy access information from BEMS
were utilised. Occupancy detection accuracy reached
96.7%. Rana et al.13 developed a novel approach to
occupancy sensing utilising accelerometer and microphone data from smart phones, and fusing these with a
sparse random classifier for occupancy detection in an
office setting. Their system accuracy was 90%, which
was 10% better than that achieved using support vector
machine (SVM) and k-nearest neighbour models.
These systems all detected occupancy often more effectively than conventional sensing, but did not count
occupants.
To count occupants, Dong and colleagues1416 proposed using information from CO2, acoustic and PIR
sensors in an open-plan office. Using information theory, the most relevant information was extracted from
sensor data and fused with SVM, artificial neural networks and a hidden Markov model. An average
reported 73% accuracy was achieved. Meyn et al.17
improved occupancy detection accuracy even further,
using a sensor network of CO2 sensors, digital video
cameras and PIR detectors as well as historic building
use data. The system used a receding-horizon convex
optimisation algorithm to infer occupancy numbers,
with detection accuracy reaching 89%, although working only at whole building level.
At the electrical appliance level, it is possible to log
case temperatures18 as a low cost and non-invasive procedure to establish usage pattern (e.g. desktop PCs use)
to 97% precision. Occupancy can in turn be inferred
from these patterns. Martani et al.19 studied the relationship between building occupancy and energy use
using Wi-Fi use as an occupancy proxy. At whole
building level, occupancy accounted for 63%69% of
electricity use (both systems relied of course on occupants using IT). Both systems could not detect

411
occupants not using a computer. Li et al.20 used radio
frequency identification (RFID) tags, measuring realtime occupancy in offices. A K-nearest neighbour algorithm reached accuracies of 88% and 62% for stationary and moving occupants, respectively. Of course,
RFID needs buy-in from building users and is unsuitable for public buildings/areas. Vision-based systems
have also been used,21,22 although occupants privacy is
a concern, as well as a clear line-of-sight. Besides, their
applicability is limited in heavily partitioned spaces
(including intrusiveness and privacy concerns).
Sensor fusion offers a way to make up for limitations described above, aiming for a more reliable occupancy count using various indoor climatic variables. It
can merge various sensors advantages, while minimising their weaknesses.

Test area instrumentation

The test area is a university admissions office an
open-plan office in a 1990s advanced naturally ventilated building which forms part of the De Montfort
University campus in the English Midlands. This test
area has been chosen because of its multi-occupancy
nature, and indoor environmental conditions can be
considered as heavily dynamic. Six staff work here, each
with a computer, plus a small kitchen area. Windows
are openable, and there is one door. Visitors are frequent. Airflow rate in this room is such that there is
build-up of indoor parameters during the day, although
temperature averages 21 C. Figure 1 shows instrumentation deployed. PIR sensors were placed in close proximity to the existing BEMS ones to monitor motion
while HOBO temperature loggers were attached to desktop PC cases. HOBO U series portable data-loggers were
used to monitor temperature, humidity and illumination
levels. volatile organic compound (VOC) and CO2 levels
were monitored using an Aerasgard rlq-series unit, and
four GE Sensing Telaire CO2 sensors, respectively, logging using HOBO units. Ambient sound monitoring used
bespoke circuits, logging events when sound reached a
threshold. Sensor placement is shown in Figure 2.
Data collection ran for two winter months, and logging was done every minute. Data were downloaded to
a PC and then uploaded to a MySQL database using
MATLAB scripts. Occupancy estimates from the data
were subsequently analysed in MATLAB and Waikato
Environment for Knowledge Analysis.23 A FLIR A40
infra-red camera was used as a control (infra-red
images preserve anonymity, and illumination is not a
problem) video capture also followed 1 min intervals.
The occupant count from these images became the
ground truth, used for model training and testing.

Data processing
Data from candidate features from various sensors are
combined to produce occupancy information that may

412

Figure 1. Test area instrumentation.

Figure 2. Sensor placement.

Proc IMechE Part I: J Systems and Control Engineering 230(5)

Ekwevugbe et al.

413

Figure 3. Fusion strategy.

not be achievable using a single sensor type. Indoor

events (PIR and sound level measurements) and indoor
climatic data (CO2, VOCs, case temperature) were processed to determine occupancy profile. Feature selection involved ranking individual features predictive
capacity, and a search to determine optimal features for
fusion. Redundant features were considered as features
with the same predictive strength, as per symmetrical
uncertainty evaluation measured value. These features
have been excluded from the feature selection analysis.24 With candidate features identified, the next step
involved implementing the fusion strategy for occupancy estimation, see Figure 3).25,26

Information theory
Feature ranking for each variable in the sensing domain
finds association strength between each feature and the
number of occupants. Information theory is a widely
used non-linear correlation measure for feature relevance analysis in machine learning applications.27 It
tends to be more computationally efficient and less time
consuming than other feature selection models such as
wrapper filters.24,28 Here, entropy indicates a variables
uncertainty. Entropy of Y with a probability mass

function is shown in equation (1). For a pair of random

variables (X, Y), if the outcomes of X are known, then
the amount of uncertainty for realisation of variable Y
is given by the conditional entropy, HYjX defined by
equation (2)
X
HY = 
log2 pi yi
1
i

HYjX = 

X  X 



p xj
p yi jxj log2 p yi jxj
i

wherepyi  is the prior probabilities for all values of Y

and p yi jxj is the conditional probability of yi given xj .
Suppose Y is treated as the actual occupancy data,
and X as the features extracted from various indoor
environmental sensors. HYjX = 0, if all feature subsets belong to the same class, indicating that there is no
uncertainty between both variables, while, HYjX = 1,
if a feature subset is totally random to a class. Entropy
of Y decreasing indicates the additional information
about Y provided by X and is called information gain
as shown in equation (3). Information gain measures
the dependence or common uncertainty between a feature and actual occupancy data, and it is defined as
IY, X = HY  HYjX

414

Proc IMechE Part I: J Systems and Control Engineering 230(5)

Table 1. Mathematical descriptions of features investigated.

Air temperature, relative humidity, case temperature, CO2 and VOC measurements
First-order difference (FDIFF_notation)
Second-order difference (SDIFF_notation)
Time (t) moving average (AVR_notation)

raw (i) - raw (i) -1

raw_FDIFF (i) - raw_FDIFF
! (i) -1
i
P
raw i=tminutes
it1
t
i

Approximate area under the curve for between

two data instances (A_notation)

FDIFF iFDIFF i  FDIFF i  1

ti1

 FDIFF(i  1)
(f (FDIFF(i))
+ (FDIFF(i  1))=2
Sound and PIR data
s
N
P
1
AVRi  mi
U

Variance (VAR_notation)Where U = number of

sensors, and m = mean

U
P

Occupied times: Total duration of occupancy as

detected by the sound or PIR sensor in a given
time interval (TOS_notation)
Number of pulses: Total number of state changes
as detected by the sound or PIR sensor in a
given time interval (TONP_notation)
High to low: total number of state changes from
high to low as detected by the sound or PIR
sensor in a given time interval
(THI_LO_notation)
Low to high: total number of changes from low
to high as detected by the sound or PIR sensor
in a given time interval (TLO_HI_notation)

i
U
P
i
U
P
i

U
P
i

AVRt minutesi
AVRt minutesi
AVRt minutesi

AVRt minutesi

The mutual information gain between two random

variables X and Y is symmetric, and this is desirable
when measuring correlations between features.
However, it is biased for features with more values, so
symmetric uncertainty (SU) is employed to determine
features predictive strength, ultimately for an occupancy count (equation (4)). SU compensates for information gains bias towards features with more values
and normalises them for comparability. It also treats
pairs of features symmetrically and averages the values
of two random variables, removing bias.29,30 If the SU
evaluation measure of a feature to the actual occupancy data is low, it implies poor predictive ability for
occupancy levels, and vice versa
SUY, X = 2IY, X=HX + HY

4
Figure 4. CO2, case temperature and VOC sensors feature
ranking.

Feature selection
Features analysed alongside their mathematical description are given in Table 1. A rank of extracted features
for each individual sensing domain was carried out
using symmetrical uncertainty analysis, to gauge correlation with occupant count, as shown in Figures 46.
From Figure 4, VOC had poor predictive ability for
occupancy, with a maximum value of 0.065 for the features investigated. This supports previous research that

occupants are not necessarily the major source of VOCs

in buildings. CO2 levels have a strong correlation with
occupancy numbers, as captured by features such as
FDIFF_CO2 and AF_DIFF_CO2, with SU values of
0.306 and 0.300, respectively. This trend is expected as
the test area has a relatively small space volume, does
not have ventilation stacks, and windows that are
mostly kept shut, all of which may impact on CO2 levels

Ekwevugbe et al.

415
THI_LO_PIR with an SU value of 0.240. This may be
due to the spatial arrangement of the sensors, as these
were positioned close to existing BEMS PIR sensors,
which may not capture the movement of all six occupants at all times. Sound features were effective for
occupancy monitoring, the highest SU values reaching
0.339 achieved by VOS_SND. This suggests that sound
sensing may be useful for capturing the spread of occupancy information in an observed space.
Features from each individual sensing network were
passed on successively to a genetic based correlation
feature selection (CFS) filter. This was used to determine an optimal combination of features to be applied
as inputs for the occupancy estimation model. The reason for this being that redundant features are likely to
have similar rankings or predictive power, and SU as
with most feature weighting/ranking algorithms is
incapable of removing this redundancy.27 CFS uses a
correlation-based heuristics called merit to evaluate
the worth of features

Figure 5. PIR features ranking.

krcf
MeritS p
k + kk  1rff

Figure 6. Sound features ranking.

in the space. All desktop computers in the test area were

instrumented for case temperature measurements, and
indeed captured sufficient occupancy information to
make some of its features good predictors of occupancy, as shown by AVR_CAS and VAR_CAS, with
SU values reaching 0.330 and 0.218, respectively.
VAR features had the highest SU values for PIR
and sound measurements, see Figures 5 and 6. The best
predictive performance of PIR features was by

MeritS is the heuristic merit of

Psubset S conPa feature
taining k features, and rcf = fi 2 S k1 (fi , C) is the
mean feature class correlation and rff is the average feature inter-correlation. CFS explores the feature space
for an optimal combination of a feature subset with the
highest MeritS value using a genetic algorithm based
search, with a stopping criterion set at when this value
does not increase.30
Let n Rule represent a combination with n features.
The MeritS score for each heterogeneous multi-sensory
combination was computed using CFS,23 where n
Rule = 1, 2, 3,., 6. Table 4 shows the features selected
for each successive combination. The most effective
features were evaluated as those selected with the highest number of times, appearing over 80% (or five of the
six iterations) throughout the n Rule combination process. These were considered as the best set of features
for occupancy number detection within the observed
environment and form the inputs for a machine learning model used for fusion. Using this feature selection
methodology,26 candidate features for occupancy estimation are given in Table 2.

Table 2. Optimal features subset from various sensing networks.

PIR sensor network

Case temperature network

Sound network

CO2 network

Heterogeneous multi-sensory
network

TOS_PIR
VOS_PIR
THI_LO_PIR
FDIFF_ THI_LO_PIR
VHI_LO_PIR
AF_DIFF_TLO_HI_PIR

AVR_CAS
AF_DIFF_CAS
VAR_CAS

TOS_SND
VOS_SND
VTONP_SND
VHI_LO_SND
TLO_HI_SND

AVR_CO2
FDIFF_CO2
AF_DIFF_CO2
AS_DIFF_CO2

FDIFF_CO2
AF_DIFF_CO2
AVR_CAS
VAR_CAS
VOS_SND
TOS_SND

PIR: passive infra-red.

416

Proc IMechE Part I: J Systems and Control Engineering 230(5)

Figure 7. ANN architecture used for the study.

Occupancy estimation using various

sensing network configurations
A back-propagation artificial neural network (NN) was
applied for occupancy level estimation (Figure 7).
Generally, inputs for the NN were candidates from feature selection, as shown in Table 2. Various NN architectures were tested, and an appropriate model structure
was adopted when the error variation between estimated
and actual occupancy data became sufficiently small.
The log-sigmoid transfer function was used in both hidden layers, while a linear function was used in the output layer. Fifteen neurons were used in each of the
connecting layers, with other parameters such as a
learning rate of 0.05, number of epochs of 500 and
momentum of 0.9. The training phase is repeated for 10
times to increase the probability of reaching a global
solution. The resulting average from the outputs of the
training phases was used for analysis in the next stages.
The main statistics used for evaluating the model
performance is the root mean square error (RMSE)
and accuracy. RMSE measures the difference between
estimated and actual occupancy data, while accuracy
was obtained by applying mean absolute percentage
error (MAPE), which gives a term-by-term comparison
of the relative error in the model estimations with
respect to actual occupancy data. It makes more sense
to apply MAPE only to measurements during the occupied periods than otherwise. Relative absolute error
(RAE) measures the total absolute term-by-term error
and normalises it by dividing with the total absolute
error of the estimated occupancy data. Results are presented for PIR, case temperature, CO2 and sound

Figure 8. Occupancy results using PIR sensor network.

sensor networks. Results for a typical week (excluding

weekend days) are presented and compared against
each network. In this analysis, data from the first
2 weeks were used for model training, while data from
the remaining week were used for model testing.

PIR sensor network

Studies have shown that a PIR sensor network can
work well for occupancy presence detection for building controls (especially lighting control), as opposed to
single point PIR occupancy sensing, as it tends to minimise the incidence of false-offs.31 Figure 8 shows estimated and actual occupancy data for two typical days

Ekwevugbe et al.

417

Table 3. Model performance for a typical week using a PIR

sensor network.
Weekdays

RMSE

R2

RAE

Accuracy

14th
17th
18th
19th
20th

1.313
1.945
1.683
1.432
1.216

0.569
0.402
0.503
0.634
0.732

0.484
0.579
0.491
0.390
0.341

49.71
41.08
43.23
50.17
68.33

RMSE: root mean square error; RAE: relative absolute error.

in a week. Occupancy levels were estimated with an

RMSE of 1.683 on the 18/12/2012, and RMSE of 1.432
on the 19/12/2012, with low accuracies of 43.23% and
50.17% during occupied period, respectively.
An average RMSE of 1.518, and accuracy of
50.50% for occupied periods was achieved using this
sensory network, see Table 3. The average relative error
and R2 was 0.457 and 0.568, respectively. These suggest
poor predictive ability of a PIR sensor network for
occupancy.

Case temperature sensor network

In most modern office buildings, each occupant work
station is often fitted with a computer. For control strategies based on occupancy patterns, usage patterns of
electrical appliances can provide insights into physical
space use, forming control strategies for ventilation.
Features such as AVR_CAS, AF_DIFF_CAS, and
VAR_CAS, shown in Table 2, which produced the
highest predictive capacity, were used as the model
inputs for occupancy level estimation.
From Figure 9, model estimations track fairly well
with actual occupancy data, with an RMSE of 1.141
and an accuracy of 62.58%. As presented in Table 4,
average RMSE of 1.334, R2 value of 0.598, RAE of
0.407, and accuracy of 55.70% were achieved for the
week examined.

Sound sensor network

An in-depth literature review strongly suggests that a
sound sensing network has not been applied for occupancy numbers estimation. The effectiveness of this
network for occupancy number estimation was examined by combining features such as TOS_SND,
VOS_SND, VTONP_SND, VHI_LO_SND and
TLO_HI_SND, as shown in Table 2. Table 5 presents
the performance metrics for the sound sensor network
for a typical week. An average RMSE of 1.225, R2
value of 0.676, RAE of 0.371 and accuracy 60.60%
achieved, suggesting the model estimation was fairly
close to the actual occupancy data. Figure 10 shows
model estimations and actual occupancy data. These
results are promising, given that the sensors are fairly
simple (in operation). Besides, sound measurements

Figure 9. Occupancy results using case temperature sensor

network.

Table 4. Model performance for a typical week using a case

temperature sensor network.
Weekdays

RMSE

R2

RAE

Accuracy

14th
17th
18th
19th
20th

1.601
1.352
1.141
1.346
1.231

0.427
0.581
0.725
0.570
0.688

0.552
0.414
0.331
0.378
0.360

44.67
56.57
62.58
54.71
60.00

RMSE: root mean square error; RAE: relative absolute error.

Table 5. Model performance for a typical week using a sound

sensor network.
Weekdays

RMSE

R2

RAE

Accuracy

14th
17th
18th
19th
20th

1.304
1.205
1.176
1.306
1.132

0.585
0.662
0.697
0.656
0.778

0.452
0.349
0.352
0.376
0.328

54.07
60.61
60.26
58.71
69.34

RMSE: root mean square error; RAE: relative absolute error.

can be subject to interference from sources other than

occupants which can easily degrade the model performance. Low-cost sound sensors clearly show potential
for occupancy measurement.

CO2 sensor network

CO2 sensing networks are usually deployed for ventilation control. Such a network may be an expensive
option for occupancy level monitoring, with BEMS
CO2 sensors costing around 200.00. It was difficult to
establish a reliable threshold for CO2 concentration levels with corresponding occupancy numbers, as CO2 levels in the space are affected by various factors including

418

Proc IMechE Part I: J Systems and Control Engineering 230(5)

Table 6. Model performance for a typical week using a CO2
sensor network.
Weekdays

RMSE

R2

RAE

Accuracy

14th
17th
18th
19th
20th

0.863
1.255
1.120
1.137
0.880

0.837
0.634
0.707
0.783
0.821

0.256
0.384
0.366
0.340
0.273

71.08
58.54
59.78
61.55
70.92

RMSE: root mean square error; RAE: relative absolute error.

Table 7. Model performance for a typical week using a

heterogeneous multi-sensor network.

Figure 10. Occupancy results using sound sensor network.

Weekdays

RMSE

R2

RAE

Accuracy

14th
17th
18th
19th
20th

0.842
1.161
0.815
1.064
0.845

0.849
0.706
0.859
0.827
0.830

0.244
0.349
0.229
0.288
0.246

73.20
62.24
74.67
68.53
71.89

RMSE: root mean square error; RAE: relative absolute error.

Figure 11. Occupancy results using CO2 sensor network.

air infiltration (caused by window opening) and outdoor CO2 levels. However, with an average RMSE of
1.067, R2 value of 0.756, RAE of 0.324, and accuracy
64.37% achieved for the week analysed, model estimations track well with actual occupancy data, suggesting
occupants may be largely responsible for CO2. Due to
slow CO2 decay rates, results show rather noisy estimates during unoccupied times as in Figure 11. Table 6
shows the network performance in a typical week.

Heterogeneous multi-sensory network

Figure 12 shows model estimations and actual occupancy data for two typical days using this network. It
is clear from these plots that model estimations are in
good sync with actual occupancy numbers, with an
RMSE of 0.815 on the 18/12/2012, and RMSE of 1.064
on the 19/11/2012. The model particularly performs
well during unoccupied periods; this is not surprising as
measured indoor climatic variables do not show any

Figure 12. Occupancy results using heterogeneous multisensor network.

significant temporal variation. Although for some days

in the week examined during unoccupied period, model
estimations indicated there were occupants in the space.
This may be due to the slow CO2 decay rates in the test
area, which can sometimes take till about 8:00 a.m. the
next morning for CO2 levels to completely decay.
During occupied periods, model estimation accuracy
reached 74.67%, showing close tracking with ground
truth data. Although there are certain days where accuracy was relatively low (e.g. 62.24% on the 17/12/2012).
This shows that some amount of occupancy variation
can occur between days. Table 7 presents the model
performance for a typical week. Although model outputs are decimal and may not represent practical

Ekwevugbe et al.

Figure 13. RMSE values for various sensing network

configurations in a typical week.

419

Figure 14. Occupied times accuracy for various sensing

network configurations.

observations, that is, number of occupants cannot be

4.12, outputs may require quantisation. However, the
outputs are still useful for occupancy-driven HVAC
systems, notably where the fractional occupancy count
may be useful for proportionalintegralderivative
(PID) control.
Overall, the model sometimes struggles when there
are abrupt changes in occupancy, which again may be
linked to the slow CO2 decay/increase rates. For
occupancy-driven HVAC control operations, the need
for such rapid responses may be moot.

Comparison of various sensing

configurations
In general, the RMSE measure increased with RAE
and varied inversely with R2 and the model accuracy
for the week examined. In Figure 13, the best average
daily testing RMSE of 0.945 was achieved using a heterogeneous combination of sensors. This is considered
good, since the number of occupants varied mostly
between 0 and 6. This suggests that the model estimations and actual occupancy number usually differ not
more than 1. Overall, PIR sensor network showed the
largest RMSE values for days in the week, with values
reaching 1.945; this was followed by the case temperature network with a peak RMSE of 1.601. Both sound
and CO2 sensor network showed good performance
with their RMSE not exceeding 1.255.
Figure 14 shows the accuracy of the test results for
different days in a typical week using various sensor
network configurations between 14/12/2012 and 20/12/
2012. Different accuracies for days of the week were
recorded, with variation between the highest and lowest
values reaching up to 27.25% for the PIR sensor network, 17.91% for the case temperature sensor network,
15.27% for the sound sensor network, 12.54% for the
CO2 sensor network and 12.43% for the heterogeneous
multi-sensory network. These variations may be influenced by the occupancy dynamics for a particular day.

Figure 15. RAE values for various sensing network

configurations in a typical week.

However, a heterogeneous multi-sensory network produced the least variation in the model daily performance. From Figures 15 and 16, the heterogeneous
multi-sensory network configuration produced the least
relative error and variations between model estimations
and actual occupancy data in weekdays examined. On
the 14/12/2012, CO2 and multi-sensory networks show
similar relative errors, at 0.256 and 0.244, respectively,
and their R2 values at 0.837 and 0.849 also vary accordingly. Other sensory configurations showed poor RAE
and R2 values relative to both. This trend was similar
for 19/12/2012 and 20/12/2012. However, on the 17/12/
2012 and 18/12/2012, the CO2 sensor network showed
larger relative error and lower R2 values compared to
other days in the week. This may be due to noisy model
estimates recorded during unoccupied times caused by
slow CO2 decay rates in the space. Sound network with
an RAE of 0.349 and R2 of 0.662 performed better than
CO2 sensor network on the 17/12/2012, while case temperature network outperformed sound and CO2 network on the 18/12/2012, with a RAE of 0.331 and R2

420

Figure 16. R2 values for various sensing network

configurations in a typical week.

Proc IMechE Part I: J Systems and Control Engineering 230(5)

Figure 17. Comparison of NRMSE of all four models.

value of 0.725. Overall, the PIR sensor network showed

the poorest performance, only performing better than
case temperature network on the 14/12/2012 and 20/12/
2012.
In summary, a heterogeneous multi-sensory network
clearly outperformed others. The CO2 sensor network
also shows good performance, although CO2 decay
rates may impact on the model estimations and control
strategies, especially when unoccupied. The use of
the PIR sensor configuration produced the worst performance. Case temperature monitoring is useful for
occupancy counting, although further testing is recommended using bigger data samples from differing
spaces.
Figure 18. Comparison of RMSE of all four models.

Comparison of various occupancy

estimation models
Several representative machine learning techniques
were applied for occupancy estimation as a robustness
test. Fusion techniques selected include SVM, radial
basis neural network (RBF), and linear regression
(LR). Use of these techniques reflects established work.
RMSE, MAPE, RAE and R2 alongside the normalised
root mean square error (NRMSE) were used to evaluate performance. NRMSE is scale dependent and more
suitable for comparing model estimations obtained
using different algorithms and has been applied to compare occupancy detection models performance.32
With respect to all metrics applied in the analysis, all
machine learning models showed similar trends on any
particular day. The NRMSE for the NN, LR, SVM
and RBF showed an error rate of 13.58%, 15.68%,
14.10% and 14.82% and the corresponding RMSE values of 0.815, 0.941, 0.846 and 0.869, respectively, on
the 18/12/2012 (Figures 17 and 18).
SVM, LR and RBF generated rather noisy estimates
with frequent variation as against the true occupancy

number, especially during the early hours of the morning. This may be so in part, because both SVM and
RBF assume each data instance is independent and
identically distributed. This may not always be the case,
as parameters such as CO2 levels tend to have strong
inherent temporal correlations. It is also clear that
while LR model can produce decent results, it does not
achieve optimal occupancy estimations. However, it
does provide some information about the model. For
example, on the 18/12/2012, the LR model suggests
that case temperature and CO2 features are highly correlated to the number of occupants. Surprisingly, sound
features produced poor correlation. The model learns
the following coefficients for estimating observed number of occupants. Equation (6) shows this relationship
^ = 0:2935  AVR CAS  0:2506  VAR CAS
O
+ 3:2807  FDIFF CO2  2:3487  AF DIFF CO2
+ 0:0069  TOS SND + 0:068  VOS SND
 5:5864

Ekwevugbe et al.

Figure 19. Comparison of occupied times accuracies for all

four models.

^ is estimated occupancy number.

where O
On the 19/12/2012, SVM had the least error rate at
13.22%, followed by RBF and LR at 14.52% and
14.94%, respectively. NN showed the worst performance with an error rate of 15.20%, showing large fluctuations during occupied times. As already mentioned
in this section, the rather noisy estimates may be linked
to the assumption that the NN model also treated each
data instance as independent and identically distributed. The average accuracies achieved for NN, SVM,
LR and RBF in the week tested were 70.15%, 71.34%,
67.25% and 68.22%, respectively, with SVM providing
the best accuracy (Figure 19). All four models struggle
to track abrupt changes in occupancy numbers within a
short duration. This may be linked to the slow response
of the CO2 sensors, whose features contribute significantly to the model predictive capacity. However, for
occupancy-driven HVAC control, results from all four
models may be considered sufficient, since sudden
occupancy level changes may not be of any control significance, unless the space is switching from occupancy
to vacancy or vice versa.
All models do show a certain level of fluctuations
and spikes in their estimation profiles as against actual
occupancy data. This may be attributed to the models
use of features inputs (such as FDIFF_CO2,
AF_DIFF_CO2), which may be sensitive to noise. The
average daily R2 values indicated that LR model
showed the largest variation with a value of 0.775, and
SVM model showed the least with a value of 0.822.
NN and RBF were 0.815 and 0.795, respectively.
Results for the relative error show similar pattern.
Figure 20 shows how the R2 values vary for all four
models in a typical week, while Figure 21 presents that
of the RAE values.
Extensive fluctuations in the estimated occupancy
profile is highly undesirable for building operations, as
this would lead to frequent adjustment of HVAC
equipment, which may impact negatively on the

421

Figure 20. Comparison of R2 of all four models.

Figure 21. Comparison of RAE of all four models.

equipment life-cycle. Hence, it could be useful to design

appropriate filters to reject outliers in the estimated
results, before integrating such outputs into any existing BEMS.
In summary, all the four machine learning techniques are seen to be showing the same tendency. Their
model estimations show good tracking with actual
occupancy numbers, and the SVM model performed
better than others. This analysis confirms the robustness of the occupancy detection methodology applied in
this study. Model results suggest the potential for generating occupancy information for enhancing building
control operations, especially demand-driven HVAC
systems.

Sensitivity analysis
Sensitivity analysis was carried out to investigate the
effect of time interval on the performance of the proposed system. Since results for occupancy estimation
using the heterogeneous multi-sensory approach

422

Figure 22. RMSE with varying time interval.

outperformed others, this was selected for further

investigation to study the effect of time interval.
Although data were acquired every minute, 15 min and
half-hourly averages were applied for occupancy estimation. RMSE was applied for this analysis.
The RMSE reduces with increasing time interval.
For the two typical days, RMSE reduced by around
3.8% on the 18/12/2012 whilst using the 15 minutes
data and reduced by around 8.5% with the half-hourly
data. This is mainly due to the decrease in the sum of
squares of the errors due to fewer time instances. On
the 19/12/2012 show a similar trend, with RMSE reduction at around 9.2% and 16.2% for 15 minutes data
and half-hourly data, respectively. Figure 22 illustrates
the effect of varying the time interval on the model performance for the weekdays examined. The average
RMSE reduction for a 15- and 30-minutes window was
around 5.3% and 11.1%, respectively. However, on the
17/12/2012 and 19/12/2012, the RMSE reduced relatively significantly at around 18.6% and 16.2%, respectively. On both days, the performance of the model was
relatively poor.
Many demand-driven HVAC systems are operated
within a moving time interval (window), with indoor
climatic control starting from the present time and finishing after a stipulated time interval. Occupancy presence/count detected anytime within this window can be
considered as occupancy detected. In this study, the
average percentage change in RMSE values should normally not cause significant energy waste, as this would
not affect the detection performance drastically, such
that frequent HVAC control actions would be needed.

Conclusion
This study has demonstrated the need for a heterogeneous multi-sensory fusion approach in the development of occupancy detection systems. This approach
seems to have the best potential in addressing the
short-comings of existing occupancy detection systems.

Proc IMechE Part I: J Systems and Control Engineering 230(5)

For system development, an experimental design utilising assorted low-cost indoor environmental sensors
deployed in open-plan office environment was applied
to demonstrate proof-of-concept. This study has proposed an advanced data processing methodology to
detect occupancy numbers in naturally ventilated openplan office buildings based on information from a network of low-cost sensors. This methodology involved
the use of symmetrical uncertainty analysis and a
genetic based search for feature selection, and a
machine learning model for sensor data fusion.
Extensive experimental testing demonstrated that the
system offers a promising approach for building occupancy sensing capable of supporting improved HVAC
operations, and thereby facilitating energy savings.
Estimated occupancy numbers show reasonable tracking with actual occupancy levels. In general, the model
estimations tracked actual occupancy levels with accuracy reaching 74.67% during occupied instances, which
compares favourably with existing occupancy detection
systems performance.
Declaration of conflicting interests
The author(s) declared no potential conflicts of interest
with respect to the research, authorship and/or publication of this article.
Funding
The author(s) received no financial support for the
research, authorship and/or publication of this article.
References
1. DEFRA. e-Digest statistics about: climate change, 2006,
https://www.gov.uk/government/organisations/department
-for-environment-food-rural-affairs
2. World Business Council for Sustainable Development
(WBCSD). Energy efficiency in buildings brief #2- our
vision: a world buildings consume zero net energy. Geneva:
WBCSD, 2008.
3. Brown N. Image processing for overnight lighting quantification in buildings. In: Proceedings of the improving
energy efficiency in commercial buildings (IEECB),
Frankfurt, 1314 April 2010.
4. Meyers RJ, Williams ED and Matthews HS. Scoping the
potential of monitoring and control technologies to
reduce energy use in homes. Energ Buildings 2010; 42(5):
563569.
5. Perez-Lombard L, Ortiz J and Pout C. A review on
buildings energy consumption information. Energ Buildings 2008; 40(3): 394398.
6. Sun Z, Wang S and Ma Z. In-situ implementation and
validation of a CO2-based adaptive demand-controlled
ventilation strategy in a multi-zone office building. Build
Environ 2011; 46(1): 124133.
7. Tachwali Y, Refai H and Fagan JE. Minimizing HVAC
energy consumption using a wireless sensor network. In:
Proceedings of the 2007 33rd annual conference of the
IEEE industrial electronics society (IECON 2007), Taipei, Taiwan, 58 November 2007. New York: IEEE.

Ekwevugbe et al.
8. Delaney DT, OHare GM and Ruzzelli AG. Evaluation
of energy-efficiency in lighting systems using sensor networks. In: Proceedings of the first ACM workshop on
embedded sensing systems for energy efficiency in buildings, Berkeley, CA, 3 November 2009. New York: ACM.
9. Padmanabh A, Adi Malikarjuna V, Sen S, et al. A wireless sensor network based conference room management
system. In: Proceedings of the first ACM workshop on
embedded sensing systems for energy-efficiency in buildings (BuildSys 09), Berkeley, CA, 3 November 2009,
pp.3742. New York: ACM.
10. Jianfeng C, Zhang J, Kam AH, et al. An automatic
acoustic bathroom monitoring system. In: Proceedings of
the IEEE international symposium on circuits and systems,
2005 (ISCAS 2005), Kobe, Japan, 2326 May 2005.
New York: IEEE.
11. Dodier RH, Henze GP, Tiller DK, et al. Building occupancy detection through sensor belief networks. Energ
Buildings 2006; 38(9): 10331043.
12. Zhao Y, Zeiler W, Boxem G, et al. Virtual occupancy
sensors for real-time occupancy information in buildings.
Build Environ 2015; 93(Part 2): 920.
13. Rana R, Kusy B, Wall J, et al. Novel activity classification and occupancy estimation methods for intelligent
HVAC (heating, ventilation and air conditioning) systems. Energy 2015; 93(Part 1): 245255.
14. Dong B, Andrews B, Lam KP, et al. An information
technology enabled sustainability test-bed (ITEST) for
occupancy detection through an environmental sensing
network. Energ Buildings 2010; 42(7): 10381046.
15. Lam KP, Hoynck M, Dong B, et al. Occupancy detection
through an extensive environmental sensor network in an
open-plan office building. In: Proceedings of the eleventh
international IBPSA conference, Glasgow, 2730 July
2009. IBPSA.
16. Lam KP, Hoynck M, Zhang R, et al. Information-theoretic environmental features selection for occupancy detection
in open offices. In: Proceedings of the eleventh international
IBPSA conference, Glasgow, 2730 July 2009. IEEE.
17. Meyn S, Surana A, Lin Y, et al. A sensor-utility-network
method for estimation of occupancy in buildings. In: Proceedings of the 48th IEEE conference on decision and control, 2009 held jointly with the 2009 28th Chinese control
conference (CDC/CCC 2009), Shanghai, China, 1518
December 2009. New York: IEEE.
18. Brown N, Bull R, Faruk F, et al. Novel instrumentation
for monitoring after-hours electricity consumption of
electrical equipment, and some potential savings from a
switch-off campaign. Energ Buildings 2012; 47: 7483.
19. Martani C, Lee D, Robinson P, et al. ENERNET: studying the dynamic relationship between building occupancy

423

20.

21.

22.

23.

24.

25.

26.
27.

28.

29.

30.

31.

32.

and energy consumption. Energ Buildings 2012; 47: 584

591.
Li N, Calis G and Becerik-Gerber B. Measuring and
monitoring occupancy with an RFID based system for
demand-driven HVAC operations. Automat Constr 2012;
24: 8999.
Benezeth Y, Laurent H, Emile B, et al. Towards a sensor
for detecting human presence and characterizing activity.
Energ Buildings 2011; 43(23): 305314.
Tomastik R, Yiqing L and Banaszuk A. Video-based estimation of building occupancy during emergency egress.
In: Proceedings of the 2008 American control conference,
Seattle, WA, 1113 June 2008. New York: IEEE.
Hall M, Frank E, Holmes G, et al. The WEKA mining
software: an update, 2009, http://www.cms.waikato.
ac.nz/;ml/publications/2009/weka_update.pdf
John GH, Kohavi R and Pfleger K. Irrelevant feature
and the subset selection problem. In: Proceedings of
the eleventh international conference machine learning,
Rutgers University, New Brunswick, NJ, 1013 July
1994.
Ekwevugbe T, Brown N, Pakka V, et al. Real-time building occupancy sensing using neural-network based sensor
network. In: Cohen WW and Hirsch H (eds) Machine
learning: Proceedings of the eleventh international conference, Menlo Park, CA, 2426 July 2013, pp.121129. San
Francisco, CA: Morgan Kaufmann Publishers.
Ekwevugbe T. Advanced occupancy measurement using
sensor fusion. Leicester: De Montfort University, 2013.
Yu L, Liu H and Guyon I. Efficient feature selection via
analysis of relevance and redundancy. J Mach Learn Res
2004; 5: 12051224.
Blum A and Langley P. Selection of relevant features and
examples in machine learning. Artif Intell 1997; 97:
245271.
Senthamarai Kannan S and Ramaraj N. A novel hybrid
feature selection via Symmetrical Uncertainty ranking
based local memetic search algorithm. Knowl-Based Syst
2010; 23: 580585.
Hall M. Correlation based feature selection for machine
learning. Hamilton, New Zealand: Department of Computer Science, University of Waikato, 1999.
Tiller DK, Guo X, Henze GP, et al. Validating the application of occupancy sensor networks for lighting control.
Lighting Res Technol 2010; 42(4): 399414.
Erickson VL, Carreira-Perpinan MA and Cerpa AE.
OBSERVE: occupancy-based system for efficient reduction of HVAC energy. In: Proceedings of the 2011 10th
international conference on information processing in sensor networks (IPSN), Chicago, IL, 1214 April 2011.
New York: IEEE.