Renewable and Sustainable Energy Reviews 167 (2022) 112704

Contents lists available at ScienceDirect

Renewable and Sustainable Energy Reviews

journal homepage: www.elsevier.com/locate/rser

A review on occupancy prediction through machine learning for enhancing

energy efficiency, air quality and thermal comfort in the built environment
Wuxia Zhang a, b, *, Yupeng Wu a, John Kaiser Calautit a
a
Department of Architecture and Built Environment, University of Nottingham, Nottingham, NG7 2RD, UK
b
School of Architecture and Civil Engineer, Taizhou University, Jiaojiang, 31800, Zhejiang, China

A R T I C L E I N F O A B S T R A C T

Keywords: The occupants’ presence, activities, and behaviour can significantly impact the building’s performance and
Occupancy prediction energy efficiency. Currently, heating, ventilation, and air-conditioning (HVAC) systems are often run based on
Machine learning assumed occupancy levels and fixed schedules, or manually set by occupants based on their comfort needs.
Thermal comfort
However, the unpredictability and variability of occupancy patterns can lead to over/under the conditioning of
Energy efficiency
Building model simulation
space when using such approaches, affecting indoor air quality and comfort. As a result, machine learning-based
Occupancy detection models and methodologies are progressively being used to forecast occupancy behaviour and routines in
buildings, which may subsequently be used to aid in the design and operation of building systems. The present
work reviews recent studies employing machine learning methods to predict occupancy behaviour and patterns,
with a special focus on its related applications and benefits to building systems, improving energy efficiency,
indoor air quality and thermal comfort. The review provides insight into the workflow of a machine learning-
based occupancy prediction model, including data collection, prediction, and validation. An organised evalua­
tion of the applicability or suitability of the different data collection methods, machine learning algorithms, and
validation methods was carried out.

thermal environment [8]. Even high-performance and energy-efficient

1. Introduction buildings may not be comfortable or healthier than other buildings as
they intended to be [9].
Buildings are responsible for up to 40% of the global total energy [1] Building energy simulation tools and models are used to simulate the
and 30% of greenhouse gas [2]. As a result, reducing the amount of energy consumption at the design stage, ensuring that the building and
energy used by the building industry will considerably benefit the its services match the required standards. However, variations in
overall energy use and carbon concerns [3]. Buildings have a high en­ building construction, operation of building and energy services, usage
ergy consumption since they serve a variety of purposes and consume of ICT and appliances, and occupancy behaviour all contributed to the
energy [4]. Particularly, buildings now combine traditional energy distance between real and expected energy loads [10]. In the past, oc­
services systems like heating, ventilation, and air conditioning (HVAC), cupants’ behaviours were observed [11] or through interviews and
lighting, power distribution, and water systems with on-site power-­ surveys [12] to generate a fixed occupancy schedule [13] which can be
generating systems like solar photovoltaic (PV), wind turbines, and used in building models or simulations for existing buildings. However,
electric vehicle charging systems [5]. Many of these services are the actual occupancy behaviour is difficult to predict since it is
essential for maintaining thermal comfort and air quality [6], and the time-varying and identity in different cases. Therefore, proposing a
main challenge is to find a balance between providing a comfortable and thoroughly and accurately occupancy prediction model is necessary for
healthy indoor environment while minimising the energy demand. building energy conservation and to guide the occupant behaviour
Despite the massive quantity of energy used by buildings, thermal modelling in building energy simulation [14].
comfort is not always achieved. A study showed that in a conditioned In the last decade, new powerful tools, including machine learning
office building, 75% of occupants report that they are dissatisfied with methods and data mining techniques, have been suggested to diagnose
their thermal comfort [7]. Another field study in the US indicated that unnoticed relationships and summarise the data in innovative ways
only 60% of occupants in 60 office buildings were satisfied with their according to large information datasets, as discussed in many studies

* Corresponding author. Department of Architecture and Built Environment, University of Nottingham, Nottingham, NG7 2RD, UK.
E-mail address: [email protected] (W. Zhang).

https://doi.org/10.1016/j.rser.2022.112704
Received 26 January 2022; Received in revised form 28 April 2022; Accepted 10 June 2022
Available online 29 June 2022
1364-0321/© 2022 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
W. Zhang et al. Renewable and Sustainable Energy Reviews 167 (2022) 112704

List of abbreviations ANN Artificial Neural Network

LSTM Long Short-Term Memory
HVAC Heating, Ventilation and Air-Conditioning DNNs Deep Neural Networks
PV Solar Photovoltaic AdB AdaBoost
ICT Information and Communication Technology RF Random Forest
PMV Predicted Mean Vote GB Gradient Boosting
SET Standard Effective Temperature LR Logistic Regression
ASHRAE American Society of Heating Refrigerating and MLP Multilayer Perceptron
Airconditioning Engineer SVR Support Vector Regression
IAQ Indoor Air Quality LMSR Linear Model Stepwise Regression
MPC Model Predictive Control NB Naïve Bayes
ML Machine Learning SVC Support Vector Classification
AI Artificial Intelligence RBFN Radial Basis Function Network
PIR Pyroelectric Infrared HMM Hidden Markov Model
IoT Internet of Things FFNN Feed Forward Neural Network
RFID Radio Frequency Identification Devices TCV Thermal Comfort Votes
HMI Human Machine Interface TSV Thermal Sensation Votes
DT Decision Tree ANFIS Adaptive Neuro-Fuzzy Interference System
SVM Support Vector Machine RMSE Root Mean Square Error
NNARX Nonlinear Autoregressive Network with Exogenous MSE Mean Square Error
KNN K-Nearest Neighbour RMSPE Root Mean Squared Percentage Error
CNN Convolutional Neural Network

[14]. To better understand energy usage in buildings, research tends to accurately reflect real life, regarding unpredictable circumstances in the
study the diversification of occupancy schedules based on big data physical world [26]. Therefore, optimisation of these models was pro­
streams [15]. A lot of research has been conducted to bridge the gap posed to better assess the total comfort sense of occupants [27].
between occupancy prediction and building control while maintaining In addition, the concept of comfort should not only focus on thermal
thermal comfort, which naturally has a significant impact on building environments, as stated above. ASHRAE 62.1–2019 defines acceptable
energy use. One research with an AI-based method achieved energy indoor air quality (IAQ) as there are no known toxins at dangerous
conservation of up to 30% by using occupancy and eight different amounts in the air, and a considerable majority (80% or more) of people
physical sensors [16]. Another paper proposed an integrated framework inside do not express discomfort [28]. Most studies refer to the index of
for an HVAC system that suggested a significant reduction in comfort comfort in IAQ as environment air temperature, humidity [29], airflow
dissatisfaction, going from 25% with the baseline strategy to 0% rate [30], CO2 concentration and pollutants. For example, the permis­
dissatisfaction while decreasing the energy cost by more than 10% [17]. sible concentration of CO2 in closed spaces, according to the World
Therefore, it is necessary to complete a literature review on the associ­ Health Organization (WHO) is 1000 ppm [31] and CIBSE recommends a
ation between energy usage, comfort-improving, and machine learning CO2 concentration of no more than 900 ppm to control human odours
methods to establish a possible state-of-the-art approach to study the and maintain comfort. However, only a few experimental investigations
intercommunication between these topics. have looked at the influence of occupancy behaviour on HVAC system
performance on actual comfort and IAQ [32], which made the index of
1.1. Occupant centric comfort approaches in buildings perceived IAQ still questionable. According to a recent study, an
occupancy-based system can save up to 24% energy-consuming while
Buildings are where people spend more than 85% of their lifetime maintaining thermal comfort and assessing IAQ [33]. From another
[18], provide a comfortable and healthy environment and protect oc­ study, combining CO2 sensors with occupancy-based ventilation control
cupants against outside conditions [19]. Different researchers have might save about 55.8% of outside air ventilation power [34], but the
proposed several definitions of the comfort concept. Usually, it is airflow rate is based on occupancy number and not verified. Therefore,
commonly agreed that comfort is concerned with the occupant’s phys­ the actual perceived thermal comfort and IAQ should be examined by
iology and the physics of the surroundings, in terms of these factors: more experimental studies.
thermal comfort, health, and availability of control [20]. Furthermore, most of these models assume equal comfort for a group
Traditionally, physics-based heat balancing and transmission models of people instead of individual comfort, given the nature of the aggre­
such as standard effective temperature (SET) [21], the predicted mean gation modelling method [35]. Since occupants have significant indi­
vote (PMV) [22], and adaptive comfort models have been used to assess vidual differences, using group-averaged forecasts to control the
and analyse thermal comfort in buildings [23]. These models have building environment may not meet the individual’s thermal comfort
already been integrated into several standards, such as ASHRAE 55 [24] demands [36]. It is shown that in the same situation, people with
and EN 16798–1 [25] for categorising indoor thermal conditions into different body compositions respond differently [37] and occupants’
various comfort categories or classes. The thermodynamic equation gender and age differences would affect their personal thermal comfort.
between occupants and their thermal environment is the basis for these Therefore, more research should focus on the actual perceived
models. They believe that the human being must be in thermal harmony comfort models, which aim to predict the personal comfort of occupants
with its surroundings to feel at ease. Personal factors such as clothing to get a more accurate model while minimising energy consumption.
and activity and physical metrics such as ambient temperature, airspeed, Recent improvements in smart devices (e.g., wearing sensors) have
humidity, and ambient temperature are widely employed. The made it easier to gather data to construct and validate individual ther­
physics-based heat balancing and transfer model has been questioned mal comfort models without being too aggressive [38]. Furthermore,
despite its widespread usage in building standards around the world, developments in ML technology have made it easier to analyse extensive
mostly because it was developed in a stable laboratory that does not data and collect valuable insights that can be summarised into a module

2
W. Zhang et al. Renewable and Sustainable Energy Reviews 167 (2022) 112704

and integrated into a self-learning system. Table 1

Information on existing reviews in the last five years.
1.2. The use of machine learning method in occupancy prediction Ref. Year Journal Research Focus and Gaps

[50] 2019 Indoor Air Focused on the sensors collecting air

Many occupancy models have been created over the last twenty quality index and have not considered
years to simulate occupant unpredictability and variety and generate the occupancy impact.
stochastic occupancy models for making accurate simulations [39]. The [51] 2019 Energy & Buildings Focused on occupancy sensing and
three types of prediction models are the physical model or white-box lack of consideration of future
prediction and validation methods
model, the ML model, also called the data-driven model, and the [52] 2020 Energy & Buildings Mainly focused on occupancy
mixture model [40]. White-box models produce detailed simulations of detection and estimation, not enough
a building’s energy performance, with details such as the building ma­ integration of occupancy information
terial, HVAC control, and management systems [41]. In addition, with models.
[49] 2020 Energy & Buildings Examined papers using machine
creating a white-box model takes time and some building details are
learning in different stages of building
difficult to obtain. Data-driven models are fast to construct and provide life cycle.
acceptable results with good data quality, but they require a large [53] 2021 Building and Focused on the various types of MPC
amount of data, and their parameters and inputs have no obvious Environment and their software implementation
physical meaning [42]. Mixture models combine physical and [47] 2021 Sustainable Energy Focused on prediction on occupant
Technologies and number/level and fail to locate the
data-driven models, inheriting the advantages and disadvantages of Assessments impact of more detailed occupancy
both techniques. Traditional energy models with sets of specified static behaviour.
coefficients multiplied by a maximum room occupancy were white-box [54] 2021 Building Simulation Focused on sensors and algorithms
models with extensive building information and certain occupancy used in occupancy prediction and do
not pay attention to the interaction of
characteristics [43].
occupants with the building systems.
With the rapid advancement of computer technology, data-driven [55] 2021 Renewable and Focused on the energy model but did
approach (black-box) models have shown great potential in building Sustainable Energy not pay enough attention to the
energy models to simulate and predict related appliances, including Reviews occupancy factors and their comfort.
occupancy behaviour, thermal comfort, IAQ and energy consumption. A [56] 2021 Building and Divided the occupancy prediction
Environment models into state/level prediction and
study compared the occupancy prediction model with and without a occupancy activities prediction, but
machine learning algorithm and showed that the accuracy was signifi­ the discussion about activities
cantly improved and 30% energy saving can be achieved with the pro­ prediction is limited.
posed algorithm [44]. Another study using a learning-based model
predictive control (MPC) technique achieved significant energy savings,
can control and optimise the operation of the HVAC regarding energy,
with 40.56% less cooling and 16.73% less heating power while keeping
comfort and health.
occupants comfortable [45].
To address the gaps in the relevant review studies (some detailed
Although these ML algorithms have been widely used and checked in
above), this work will conduct a comprehensive review of occupancy
earlier studies, the algorithm choice differs in each case; the model setup
prediction and evaluate the interrelated applications and benefits to
depends on many factors including the available information, the
building operation, including improving occupancy comfort and indoor
timescale preferred, the time span (from any seconds to years), and the
air quality and reducing energy loads. The review will also provide an
size (a small space to a whole country). As a result, with the growing
insight into the workflow of a machine learning-based occupancy pre­
number of publications produced, it’s more important to examine model
diction model, including data collection, prediction, and validation in
capabilities, problems, and a critical assessment of research gaps. This
Section 2. Section 3 will review the different data collection methods
work will discuss the developments in occupancy behaviour prediction
and technologies. The best-performing algorithms in occupancy pre­
and ML technology and how it enhances thermal comfort and IAQ, and
diction modelling will be highlighted in section 4, and the different
reduces energy consumption.
validation approaches will be investigated in section 5. Finally, the
challenges linked with occupancy prediction models will be discussed,
1.3. Previous reviews, novelty, aim and objectives
and recommendations regarding further research will be made.
With the growing number of articles published on occupancy pre­
2. Method and commonly used occupancy prediction workflow
diction, an in-depth and critical evaluation of the different methods in
based on ML
this area is required. In 2012, a brief review was conducted of the
methods for predicting building energy consumption, including ANNs
Although there is a large amount of literature on building occupancy
and SVM [46]. In 2021 a review compared the AI-based and conven­
prediction using machine learning and a great number of review articles,
tional models employed in building energy consumption prediction with
what is lacking is a straightforward categorisation and organisation of
occupancy factors and proved that AI-based models had better accuracy
the methodologies and technologies, allowing for the definition of a
[47]. Another work reviewed studies on electrical load prediction and
useful (or ideal) “occupancy prediction structure”. Therefore, we
provided an overview of prediction timescale and potential model so­
consider articles published from 2011 to 2021 in the main databases
lutions [48]. The use of machine learning in the various phases of the
such as Scopus and Thomas Reuters’ Web of Science. The keywords
building lifecycle was examined, and research gaps in the design, con­
included “building, occupancy prediction, machine learning” & “ther­
struction, operation and maintenance, and control, were investigated in
mal comfort, occupancy prediction, building”. The keywords “thermal
another paper [49]. Most of these review papers focused on the occu­
comfort, machine learning, artificial intelligence, comfort factor, indoor
pancy detection approach and performance, while in terms of its
air temperature, and control method” were also used to identify more
application in buildings, most of the studies evaluated its impact on
related publications. We focus on papers that employed machine
energy efficiency but not thermal comfort and air quality (as shown in
learning to predict occupancy in buildings and related applications.
Table 1). This work argues that the occupancy behaviour data obtained
Review papers and irrelevant papers were excluded, for example, some
can be employed to minimise energy and at the same time provide a
research only focused on occupancy detection and was not suitable for
comfortable and healthy environment. For example, the occupancy
the review purpose.
prediction method can be integrated into a framework or model which

3
W. Zhang et al. Renewable and Sustainable Energy Reviews 167 (2022) 112704

2.1. The application of reviewed research 3. Data collection for occupancy data

160 papers were selected, and a timely review was proposed, which 3.1. Data collection, methods and privacy preservation
can help guide the future research of occupancy prediction with ma­
chine learning regarding building design, operation, and research ac­ To improve the accuracy of occupancy prediction, plenty of data
tivities and to provide a better understanding of occupancy behaviour collection methods have recently been introduced. According to several
and better the building performance. studies, occupancy sensing can save up to 30% [57] on energy costs
In general, the number of machine learning methods and their ap­ while improving indoor air quality [58]. However, although the use of
plications in built environment research is rising, particularly in the last such technology is promising and provides a glimpse of future smart
five years (Fig. 1). These applications include the prediction of occu­ buildings, privacy issues have to be addressed for wider adoption. More
pancy state, occupants’ interactions with thermal comfort, energy con­ resolution and accurate building prediction models can be achieved by
sumption, indoor temperature, and lighting use. Occupancy state combining adequate monitoring technology of the building environ­
prediction was the most popular application of machine learning models ment with proper HVAC or other systems monitoring.
until 2020, while the number of studies on energy consumption pre­ Because the detection of occupancy status is constantly linked to
diction increased. This could be due to the development of prediction privacy concerns [59], selecting the appropriate sensor is not always
models, which can be specifically used for more detailed problems like simple. Based on the reviewed literature, studies are usually narrowed to
the comfort state and the occupancy actives instead of just predicting if academic buildings (labs or offices in universities/research institutes),
the room is occupied or not. Also, it shows increasing awareness of en­ which could impact the quantity and quality of data obtained, particu­
ergy efficiency and occupancy comfort in the built environment. larly when the prediction method is applied to the industry. As shown in
Fig. 4, 46% of the case studies were conducted in academic buildings.
2.2. The regions of reviewed studies Other case study building types include office (25%), residential (16%),
commercial (8%) and others such as airport terminals [60], museums
The case studies in reviewed papers were mostly conducted in three [61], mosques [44] and metro stations [62].
big geographic regions: Europe, North America, and Asia. Most of the Fig. 5 shows the building types in case studies in different regions.
early studies were in Europe and North America, while studies in Asia Academic buildings play a dominant role in the reviewed studies in all
have increased since 2016, as shown in Fig. 2. In recent years, when the regions because it is easier to conduct, especially when considering
topic became more popular, these three main regions dominated this privacy issues. Office buildings are quite popular in all regions since the
field by turns. Other regions showed less interest in this area until 2017, occupants are usually fixed, and not hard to get permission. In 2020, a
indicating that more studies would be conducted in other regions in the paper conducted a case study in an office building in Stockholm, col­
future. lecting five years of data with multiple sensors installed in the building
The prediction timeframe and model system are different in identi­ [63]. However, the privacy concern may arise when such technology is
fied studies, making it hard to conclude a perfect model for building applied commercially or for widespread adoption in some regions as
occupancy prediction. However, in current studies, a typical occupancy commercial buildings are the least favoured case study type in Europe
prediction model usually consists of several procedures: data collection, and North America. In Asia, the residential building is the least used,
occupancy prediction, and validation (as shown in Fig. 3). Each pro­ indicating the intense privacy concern for households in this area.
cedure contains various options concerning the inputs, data structure Privacy leakage is always a concern when choosing sensors for data
and algorithm, which require dedicated examination based on the target collection. The key privacy risks for occupancy detection include col­
problem and building system. Conversely, the building performance and lecting the identification and location of individuals. Masking, encryp­
occupancy comfort will be impacted by the model proposed. Therefore, tion, noise addition, anonymisation of data, and scrambling of location
this paper will have the following sections: existing data gathering and data to avoid individual identification are all common procedures for
sensor technology, ML techniques for developing occupancy prediction dealing with private data. User/data anonymisation is a simple solution,
models, and model verification methodologies. The best-performing and but it offers no protection against attackers who have direct access to the
popular predictors and ML methods will be labelled, which will help sensing database and fails to provide the room-specific information and
future studies construct suitable models. required room identity [64]. An alternative way is to detect certain
occupancy patterns in a particular zone rather than target individuals
[65]. Also, occupancy location can be inferred from the occupancy data
with some auxiliary information [66]. For instance, a purposely defo­
cused camera that creates a ‘fuzzy’ or ‘warped’ image or out-of-focus
images is also a solution to room occupancy sensing [67].

Fig. 1. An overview of the application of machine learning in the built environment based on the reviewed studies from 2011 to 2021.

4
W. Zhang et al. Renewable and Sustainable Energy Reviews 167 (2022) 112704

Fig. 2. The location of case studies in the reviewed papers conducted from 2011 to 2021.

Fig. 3. The typical procedure of occupancy prediction with machine learning, validation and applications in the built environment.

In general, the two types of data gathering methods are direct applications of occupancy prediction models. Temperature sensors are
counting approaches, which directly track the occupancy number, and the most used sensor in all kinds of studies since they are easy to set up
environmental sensors, which indirectly reveal the occupancy state. and usually pre-installed in HVAC systems or other building systems.
Fig. 6 shows the connection and details of different sensors in various Some sensors are only used in specific applications; for example,

5
W. Zhang et al. Renewable and Sustainable Energy Reviews 167 (2022) 112704

or specialised experimental rooms due to private intrusiveness [71].

In recent articles, wearable sensors, mobile devices, and security
systems have all been used to detect occupancy [72]. The Internet of
Things (IoT) has opened new possibilities for occupancy detection.
Wi-Fi, Bluetooth, RFID, and other technologies are examples of these
strategies. Because Wi-Fi networks are common in modern buildings, it
requires no additional hardware or software instalment and performs
well when it comes to monitoring occupancy. A Wi-Fi-based even­
t-triggered update system for a university lecture theatre was developed
in 2019 to improve detection accuracy from 77.3% to 96.8% [73].
Despite the potential for occupancy monitoring, detection mistakes do
exist, requiring extensive data cleaning methods to filter errors to ac­
quire trustworthy occupancy data. Details of the comparisons between
these data collection methods can be found in Table 2.

Fig. 4. The proportion of building types in the reviewed case studies.

3.3. Environmental sensors for data collection

cameras are only found for occupancy state prediction and energy As shown in Table 2, most direct occupancy counting methods either
consumption prediction. Also, some sensors are more suitable for a cause private intrusiveness or are time-consuming. Compared to direct
particular application, like most studies use energy meters as sensors for occupancy counting methods, environmental sensors often target a
energy consumption predictions. The following sections will explore the smaller group of occupants, which is partly due to the cost of sensors and
benefits and drawbacks of these sensors in terms of precision, price, the detailed data these sensors can collect. Most papers use more than
ethical concerns, unresolved difficulties and future recommendations. one sensor to combine the data and avoid missing data. Also, when
people are aware that they are being watched, they may alter their
3.2. Direct occupancy counting sensing technology behaviour [83]. The idle way of data collection would be employing
existing infrastructures or simple instalments without capturing detailed
The most straightforward way to access occupancy data or profiles personal information that concerns private intrusiveness. In most
that record how occupants use the facilities or their lifestyle is to directly research, the case study is the researcher’s own office or dwelling to
follow the occupants’ status. Many researchers employed question­ avoid private intrusiveness [84]. However, the number of occupants is
naires, especially data from the large-scale survey, and it is convenient always limited, and the behaviour routine is usually fixed, which could
for groups who share the same lifestyle, such as students on campus or make the model defective when applied to larger implementations.
residents of the same culture. A dormitory building with 200 rooms was Therefore, some studies are conducted in public areas like shopping
selected as the target building, and questionnaires were sent to occu­ malls [85] and cinemas [86], while the sensors could miss some data
pants to get their working schedules [15]. Another national survey was with the large group of occupants.
taken in Korea of 5240 single-person households for their daily routines Table 3 summarises some of the recent studies using environmental
[68]. Accurate occupancy data can be obtained through these sensors. Many researchers use physical sensors like motion sensors to
large-scale surveys, and questions about their behaviour and other capture accurate occupancy states without being aware. 20.3% of
evaluation can be easily added to get the full picture as the research did energy-saving was achieved in a 550 m2 office space with motion sen­
in 2007 [68]. The mass data can show the lifestyle of a group of people. sors [87], and another experiment in a smart-home testbed with a mo­
However, these surveys are usually time-consuming and require many tion sensor has achieved around 60% accuracy for occupancy prediction
participants from the same area and extra form-filling while participants [88]. On the other hand, motion sensors are not able to detect nearly
are not always willing to cooperate. stationary individuals, which is common in offices and during the
The most accurate approach for determining the occupants’ state and inactive time at home. Therefore, the occupancy state can only be
the number of inhabitants is camera-based occupancy detection, which identified by the arrival and departure times. Also, non-intrusive sensors
is often used to offer the ground truth of occupants. An experiment in such as pyroelectric infrared (PIR), ultrasonic, and acoustic sensors can
research students’ office rooms with overhead cameras achieved over only be used to assess whether or not a space is occupied, not the oc­
80% accuracy [69], and another monitoring system with cameras was cupants’ number [89]. Therefore, they are suitable for single-occupant
employed to examine the new proposed occupancy prediction algorithm rooms. For example, research conducted in a single-occupant office
[70]. However, most cameras were installed in the researchers’ offices had a 1-h forecast accuracy of 79%–98% [90]. However, due to the air

Fig. 5. Case study building types in different regions of reviewed studies.

6
W. Zhang et al. Renewable and Sustainable Energy Reviews 167 (2022) 112704

Fig. 6. Data collection methods and their related application in the reviewed studies.

mixing process, there were always significant delays for these sensors, and meters that monitor real energy consumption can be used to detect
especially when they were located far away from occupants. the gap and validate the influence of occupancy behaviours [101].
Therefore, environmental sensors, including CO2-based detection, CO2 sensors are a viable technique since they are inexpensive, tiny,
indoor temperature, relative humidity, and energy meters, are proposed. non-intrusive, and non-terminal, making them a popular data collection
The indoor temperature sensor is the most used data collection method method [102]. Since CO2 sensors commonly exist in regular HVAC
in the reviewed papers (as seen in Fig. 6) because they are small and systems, no new infrastructure expenditure is needed. The method cal­
usually already available in standard HVAC systems. Since the indoor culates the number of occupants with an equation using CO2 concen­
temperature is not directly linked to occupancy data, the temperature tration [103], which has the main disadvantage of delayed response and
and humidity sensors are commonly combined with CO2 sensors [87] or possible difficulties in identifying physical parameters. As a result, when
weather data [100]. Also, sensors that record the indoor temperature CO2 sensors are properly installed, and details about observed rooms
and relative humidity are generally used to operate window opening and (room volume and airflow rate) are known, the CO2-based method
thermostat adjustments. These sensors, however, should be kept away performs well, whereas the results were unreliable when the studied
from sources of heat, humidity, and contamination (equipment, spaces were open and irregular, such as an open-plan or naturally
humans, and solar power) to avoid a mixture of their readings [52]. ventilated office [104]. To overcome these weaknesses, more accurate
Smart meters, which can reflect the actual electricity consumption, methods were developed including data mining algorithms.
are also employed in many works. The energy load data is easy to collect Thermal imaging and thermal comfort voting are new contactless
and compare to the simulation or prediction result. Most works exhibit a sensors that have demonstrated the capacity to enhance thermal comfort
significant performance gap between models and observed energy use, while reducing energy consumption. In an office room, using thermal

7
W. Zhang et al. Renewable and Sustainable Energy Reviews 167 (2022) 112704

Table 2 Table 3
Comparison and key findings between different direct occupancy counting Recent studies on occupancy detection using environmental sensors.
methods in recent studies. Ref. Accuracy Testing Study Scale Data Collection
Data collection Year Testing Study scale Ref. Key findings Environment Method (DCM)
mothed environment
[91] Up to 97.4% Office Around 20 3 DCM - Passive
Survey 2021 Dormitory 200 [15] ✓ Get access to m2 infrared sensor
and office on students the full picture (PIR) sensor, an on-
a campus and 90 of the site survey, a
staff occupancy camera
lifestyle. [92] The average An apartment – 6 DCM - CO2
2019 Residential 5240 [73] ✕ Time- accuracy of concentration,
houses occupants consuming and 95.8% motion sensors,
requires a relative humidity,
large number temperature,
of participants heating, and
2019 Apartments 154 [74] lighting
occupants consumption
2007 Residential 60 [12] [84] Average detection Office space 39 m2 AI-powered camera
houses occupants accuracy of
Camera 2020 Office 12.4 m2 [75] ✓ The most 92.2%
accurate [93] The best-adjusted Eight 3-bedroom 5 DCM - Motion
method, R2 is 0.94 apartments apartments sensors, indoor
provides the CO2, indoor
ground truth. humidity,
2020 Student 1400 m2 [42] ✕ The private temperature, and
centre intrusiveness the number of
2018 Student 25 [76] occupants
office residents [94] Up to 84% A house-like 3m×3m 9 DCM -
2017 Student 2 students [69] cubicle Microclimatic
office station air
2017 lecture 876 m3 [77] temperature,
theatre relative humidity,
Internet of 2020 Office 350 [78] ✓ Low cost and net-radiation, air
Things employees requires no speed, the CO2
additional concentration, and
device. illuminance level
[95] Vary from 0.82 to A mixed-use, 7445 m2 3 DCM - Outdoor
0.98 for heat university temperature, and
consumption and building historical energy
2019 Residential 149 rooms [79] ✕ The
0.87–0.97 for consumption data
complex detection error
and need data electricity
cleaning consumption
2019 Office 80 [80] [63] The error of only Office 8 floors, 9 DCM - Water
employees 5% building area: consumption,
2019 Office 200 m2 [81] 19,642 m2 electricity load,
2018 Student 25 [82] room temperatures,
ventilation devices
office residents
and controllers, air
pumping, indoor
air quality
comfort voting to obtain users’ real-time reactions to the environment
[96] Vary from 85.6 to Office Single user 7 DCM - Motion
and then modifying the management goal settings enhances thermal 93.7% and temperature
comfort while saving up to 40% energy [105]. Consequently, subjective sensors, door
responses instead of physical parameters might be an alternative sensors, pressure
sensors on office
approach to occupancy detection that should be paid more attention to.
chairs.
[97] The best accuracy A graduate About 200 3 DCM - CO2
for real-time student office m2 with 25 concentration,
3.4. Data mining technologies
prediction is 86% residents relative humidity,
and temperature
As shown in Table 3, in most studies, data collected from buildings
installed more than one kind of sensor. For reviewed papers in this [98] The highest R2 is Office room The floor 4 DCM - CO2
article, the most widely used method is the combination of indoor 0.9594 area of 152 concentration,
m2 temperature,
temperature sensors and CO2 sensors [97,106,107]. However, raw data relative humidity,
might have a variety of issues, such as missing information or sudden energy
swings if one or more sensors are disrupted. Also, sensor readings could consumption
conflict with each other, and sometimes, the reading in sensors will not [87] The total control Office space 550 m2 5 DCM - Motion
accuracy is 88.1% sensors,
change much, so it provides no valuable information.
temperature
To solve these problems above, data mining technologies have been sensors, relative
introduced by many researchers. For example, missing data were humidity sensors,
replaced with interpolated data, and nonsensical data was either CO2 sensors, and
removed or reset to the sensor’s initial values using the “data cleaning” HMI
[99] Prediction errors A study zone 125 m2, 36 6 DCM - PIR
method [108]. Extraction of the mean, standard deviation, mean abso­ below 7% occupants sensors, cameras,
lute deviation, first, second, and third-order differences and even simple temperature
moving averages are used as post-processing procedures for collected sensors, CO2
original data. For data mining, most researchers use supervised sensors

8
W. Zhang et al. Renewable and Sustainable Energy Reviews 167 (2022) 112704

algorithms like the SVM (Support Vector Machine) and the Decision Section 3 [117]. The great volume of data requires advanced techniques
Tree to categorise samples based on a target variable [109]. Unsuper­ to analyse them which conventional models cannot handle properly. In
vised learning techniques, such as hierarchical clustering and k-means, addition, most behaviours are influenced by several contextual ele­
have recently been adopted in studies to organise data into clusters ments, the best way to mimic them is to either integrate all the pa­
based on the characteristics of all variables without any target variable rameters in one equation or address the factors that influence behaviour
[110]. With the trend of multiple sensors, it is hard to confirm an oc­ separately, allowing them to be split into various formulae. Therefore,
cupancy dataset structure in advance. Therefore, using cluster algo­ powerful methods like deep learning which is suitable for big-data and
rithms is becoming a standard step before sending data to machine computationally intense processes have been introduced in recent years.
learning training. As can be seen in Fig. 7, the neural network-based algorithm (which
occupied more than 40% of reviewed papers after 2018) is the most
4. Machine learning algorithms and their applications popular method in building machine learning prediction. Particularly,
deep learning with a large number of hidden layers that compose the
Supervised learning, unsupervised learning, and reinforcement neural network showed good capacity in image pattern recognition,
learning are the three most typical machine learning approaches used in speech recognition and synthesis, etc. which also indicated possible
occupancy prediction [111]. Supervised learning models include deci­ future development in occupancy prediction models.
sion trees [112] (such as the gradient boosting tree), classifiers (such as The popularity of the neural-network-based algorithm indicated that
the Bayes classifier, kNN, and support vector machine), and neural deep learning is making major advances as typical machine-learning
network-based models [113] (such as the feedforward backpropagation techniques were narrowed in the ability to deal with data in the natu­
network and cascade correlation). Furthermore, these models can be ral form [118]. Deep learning uses graph technologies and neuron
classified as linear or nonlinear based on the data structure. Linear transformations to obtain multilayer learning models and automatically
methods are used when the responding and prediction data are linearly learns the data. The most widely used deep learning models are Con­
linked or converted into a linear relation. With the dramatically volutional Neural Network (CNN) [119] and Recurrent Neural Networks
increasing of variates, data transformation techniques like normal­ (RNN) [120], which are also popular in building occupancy prediction.
isation process, log conversion, and ranking transformation might be Also, the development of deep learning algorithms provides advance­
utilised [114]. In the majority of circumstances, linear models are easy ment in building automation systems as it can convert the data at one
to create and use, and they are frequently used as the first model. Other level (starting with the natural data) into a depiction at a slightly more
nonlinear models can be employed more effectively if the data are un­ abstract level. In 2021, a smart Oracle-based building management
likely to be linearly connected. system was proposed that auto-learns occupancy patterns and leverages
Unsupervised learning methods reduce, summarise, and synthesise spatial organisation to deliver actionable insights on energy savings
data using unlabelled training data [111]. Unsupervised learning algo­ [121].
rithms include cluster analysis learning, like principal component
analysis and parametric analysis, and various ANNs (e.g., autoencoder 4.2. Occupancy prediction
neural network and self-organising map) [115]. Because occupants
behave in a stochastic manner impacted by a variety of parameters, the Occupancy prediction, in general, draws the most attention in the
majority of which are immeasurable and unpredictable, it’s critical to reviewed papers until 2020, which is since the variation of occupants’
figure out which inputs are the greatest influencers and only add those interactions is regarded as the foundation of the uncertainty in building
that significantly increase behaviour. As a result, while unsupervised models. One of the key parameters an occupancy prediction model
learning cannot generate a prediction for a new dataset, it can contribute should consider is the occupancy level. In 2011, a study separated oc­
to the comprehension of the data’s character, allowing for the selection cupancy prediction levels into three major factors: temporal, spatial and
of supervised models for prediction [116]. Since there is no output in occupancy state resolution [122]. The precision with which the timing
unsupervised methods, data linearity is not an issue. Similarly, with of events is modelled is referred to as temporal resolution. The precision
reinforcement learning, a direct match of input and output is not of physical scale is defined by spatial resolution (e.g., a building or a
existing, and it can only estimate how well the output is. zone the model target). The model’s occupancy resolution refers to how
it specifies individuals.
4.1. The trends of machine learning and deep learning For temporal resolution, one of the classifications divided occupancy
prediction models into three categories: real-time recognition, future
In general, there is a rise in machine learning applications because of time-step predicting, and occupancy profile modelling [52]. These ap­
the availability of building automation systems, smart systems and IoT proaches either estimate the number of occupants, determine whether
platforms, which increases the quantity of data available as discussed in they exist in a particular area, or generalise a few occupancy profiles

Fig. 7. Summary of the reviewed studies from 2011 to 2021 using machine learning algorithms.

9
W. Zhang et al. Renewable and Sustainable Energy Reviews 167 (2022) 112704

based on previous occupancy patterns. In most occupancy prediction gains from the facilities operated, contributing to the indoor heat gains.
models, the monitor period has ranged from a day to multiple years, and
the time they mean to predict varies from serval seconds to more than a
year. A study showed 61.5% and 43.6% accuracy for building pre­ 4.3. Indoor air quality (IAQ) prediction
dictions of more than a year and 1 h, respectively [48]. The short-term
prediction has a direct application for quick occupancy demand IAQ has long been an important topic for the health and wellbeing of
response and suits the needs of the industry. However, the seasonal ef­ the occupants in buildings. The previous sections have highlighted the
fect of occupant behaviours requires a full year of monitoring is more importance of a holistic approach to deal with these challenges
reliable, especially in specific cases, such as simulating academic adequately. Traditionally, mechanistic IAQ models have been utilised,
buildings’ holiday schedules regarding energy consumption [95]. and the link between inputs and outputs has been based on mechanisms
In the reviewed studies, the regression-based method based on [133]. However, mechanistic IAQ models do not include the interactions
Random Forest is the most used method for occupancy prediction. Ac­ between the occupants and the indoor environment and the difference
cording to a study, the regression model is primarily used for long-term between individuals, which can impact energy consumption and
forecasting, while ANN is mainly used for short-term forecasting [48]. building performance. The operation of HVAC systems affects both
Different methods should be employed for different types of occupancy comfort and the IAQ. Hence in some studies, IAQ prediction is combined
state prediction. For example, ANN with long short-term memory with thermal comfort prediction and considered as part of the overall
(LSTM) architecture is the most commonly used and suitable method for occupant’s comfort parameter [134]. Therefore, these models, espe­
time series prediction [123]. A study found that Random Forest is the cially ML models, which consider the occupancy interaction and
most suited classifier [124], with at least 90.53% accuracy, after building performance, are increasingly being employed in recent
training data with five different machine learning classifiers (Random research.
Forest, Decision Tree Classifier, Extra-Trees, Gaussian Naive Bayes, and One of the most crucial issues in IAQ prediction is finding the right
Multi-layer Perceptron). Another study achieved 97.27%–98.90% ac­ input to achieve a reliable prediction. Since the model is data-driven, it
curacy in an indoor office by employing several Deep Neural Networks is important to identify the key variables inputs. Many environmental
(DNNs) [125]. The method’s accuracy also depends on the type of data indexes are used to determine the relationship between occupants’
collected. For example, the SVM and k-NN models have lower counting feelings about IAQ, such as door/window opening behaviour, temper­
errors when using Wi-Fi data, whereas the ANN model is more accurate ature, relative humidity, CO2 concentration, solar radiation, rainfall,
when using fused data [82]. wind speed, noise, illumination, and so on [135]. Therefore, in IAQ
Based on the reviewed papers, many studies focus on detecting the prediction models, normally, one or more driving factors are used for
occupancy state, including the occupancy presence, number and loca­ prediction [39]. The inputs may have an indirect and unexpected impact
tion in space, zone or building. However, there are limited studies on the on the behaviour [136], therefore an over-fitted model that has many
detection of the occupancy activities, for example, movement in space inputs is often conducted. Many research used data mining approaches
[75], opening/closing of windows [126], adjustment of HVAC, and use such as stepwise regression, principal component analysis [137], and
of equipment and appliance. Furthermore, significant attention of the partial least squares [138] to uncover the driving components before
existing literature is focused on the performance of developed algo­ developing the models.
rithms, such as their speed and accuracy. Details of different kinds of The algorithm selection is often related to the data structure and
occupancy prediction are listed in Table 4. collection method. However, IAQ is related to a lot of environmental
Limited works focused on evaluating the impact of the detection indexes, as stated before, which can be recorded by various kinds of
technique on the performance of the building and HVAC systems. For sensors and parameters, it is hard to recommend a specific kind of al­
example, a study proposed a vision-based approach for detecting and gorithm without analysing the detail of the model. One review before
recognising the occupants’ activities within building space [75]. Unlike summarised the popular algorithm, for example, ANN, linear regression
previous works which focused on occupancy levels, the study used the models, and Decision Tree developed for predicting different factors of
data to predict the indoor heat gains from the occupants with varying IAQ, but cannot recommend the optimal method and suggested a test
activity levels. Such information would be useful for HVAC controls to and compares different models before choosing the most suitable model
adapt and make a timely response to dynamic changes in occupancy [50].
activities. A recent work used the same detection approach to detect how Most IAQ models are employed to improve the occupants’ overall
the occupants interact with the equipment or appliance such as com­ comfort or lower the concentration of indoor air pollutants. For
puters [126]. Similarly, the proposed approach can predict the internal example, a study tested a control model of a filter for indoor CO2
decreasing in a sports centre while using fuzzy inference to reduce the

Table 4
The information about studies using various algorithms in occupancy state/number/activities prediction.
Prediction Ref. Year Sensor Algorithm Test Environment Accuracy
Classification

Occupancy State [127] 2019 Relative humidity, temperature, and CO2 Linear Regression, Neural A laboratory Higher than 90%
Prediction Networks, and Random Tree
2
[128] 2018 CO data and indoor human occupancy seasonal-trend decomposition An academic office and a An average of
(STD) cinema theatre 94.68%
Occupancy Number [129] 2021 28 Wi-Fi Apps Multilayer Perceptron ANN 5 floors of classrooms RMSPE of 0.29
Prediction [130] 2019 Real-Time Locating System inhomogeneous Markov chain A research laboratory 86% on average
[81] 2019 Wi-Fi probes and indoor air temperature, Gradient tree boosting, Random A large office room, 200 Reached 72.7%
relative humidity, and airflow rate forests, AdaBoost m2
[131] 2019 Camera and motion sensor RNN with LSTM units An exhibition Best RMSE of
10.31
2
Occupancy Activity [84] 2021 Camera CNN Office space, 39 m Average accuracy
Prediction 92.2%
[61] 2020 Social networks Random Forest and XGBoost A public museum RMSE within 30%
[132] 2019 Temperature sensor and PIR sensor Markov model (MM), HMM, and Single-family homes Under 0.80
RNN average accuracy

10
W. Zhang et al. Renewable and Sustainable Energy Reviews 167 (2022) 112704

indoor CO2 concentration [139]. Another research found Multilayer disturbance reaction [147]. Meanwhile, since two behaviours can ach­
Perceptron (MLP) follows the pattern of CO2 changes more quickly and ieve the same goal and thermal comfort often links to serval behaviours,
with higher accuracy compared to other algorithms (Support Vector ensemble models are likely to be introduced in comfort prediction
Machine (SVM), AdaBoost (AdB), Random Forest (RF), Gradient models. A paper using the machine learning approach Bagging, using a
Boosting (GB), Logistic Regression (LR)). It reduced 51.4% of energy multilayer perception network (MLPN) as a learning algorithm, out­
consumption in the total energy usage [140]. Other environmental in­ performed traditional ANN and SVM methodologies [148].
dexes like PM2.5 concentration can also be predicted by neural networks The prediction model of thermal comfort is directly linked to the
(i.e., RNN, LSTM, and gated RNNs) [141]. occupants’ satisfaction with the indoor environment. With new ML
models and data collection methods, the performance gaps will be
4.4. Thermal comfort prediction reduced. Improved models could be linked to a real-time environmental
control system to improve building management without scarifies
The number of thermal comfort prediction studies and approaches occupant comfort. For example, as shown in Fig. 8, the environmental
using the ML methods is limited compared to occupancy and energy information obtained can be used to provide data for the prediction of
consumption prediction, as in Fig. 1. In the existing literature, thermal thermal comfort in real-time, which can be used to adjust the operation
comfort is typically assessed by the PMV model based on extensive of the HVCA system. The occupancy data, such as the occupant’s num­
laboratory tests, which ignored individual comfort [22] and, in some ber and metabolic/activity level, can estimate the indoor CO2 level and
cases, do not provide satisfaction for all occupants [142]. Therefore, minimum ventilation level. Similarly, the thermal comfort prediction
most existing literature uses the ML approach to forecast thermal com­ model can also use the occupancy number and activity level. Such in­
fort and consider all occupants as a whole, disregarding data acquired formation can be used to optimise the HVAC operation while also
from separate occupants [143]. In this scenario, individual occupant minimising the energy demand.
diversity was lost, and occupants were modelled as an “average group”,
which is a statistical construction rather than an actual person [134]. It’s 4.5. Energy consumption prediction
worth noting that occupant comfort differs according to one’s age,
gender, background, and other personal characteristics. Therefore, in­ Prediction of energy usage in buildings is becoming increasingly
dividual comfort is becoming more popular, and personal comfort important, however, it is influenced by interrelated physicals, opera­
models based on data from individual occupant comfort surveys are tional, and behavioural factors such as building material, building
being developed [144]. schedule, and occupant behaviour [149]. In most cases, the
A recent study used two different machine learning algorithms to physics-based building energy simulation tools (white-box models) such
analyse a combination of inputs, including an individual comfort sys­ as DOE-2 and Energy-Plus are often used [150]. However, these tools are
tem, body temperatures, timing, and environmental parameters. Per­ limited for energy analysis since they do not contain uncertain factors
sonal comfort models achieved the best accuracy across all examined like occupancy behaviour, impacting annual energy consumption up to
methodologies and participants, according to their findings [145]. With 75% for residential buildings and 150% for commercial buildings [151].
the advancement of the Internet of Things, it is becoming more conve­ As a result, many researchers use the data-driven method (black-box
nient to collect physiological data using a range of sensors (wearable or models) to forecast energy use and analyse the effects of energy-saving
non-wearable devices). They can forecast thermal experience or satis­ initiatives like energy-retrofit strategies and renewable energy tech­
faction based on users’ physiological data, such as employing wearable nology [152]. Meanwhile, other researchers use the output of occupancy
devices to monitor skin temperature, heart rate, blood pressure, and prediction to generate an occupancy profile as input for physics-based
other physiological parameters at various human body positions (such simulation tools to calculate the energy use result (grey-box model).
as wrist, face, back and legs) [143]. Therefore, these sensors show po­ Existing machine learning-based models, on the other hand, do not
tential for the future development of thermal comfort prediction. adequately account for occupant behaviour. They either ignore occu­
In 2012, a research employed a PMV control model with an RNN pancy behaviour entirely or deal with it in a limited way, such as merely
network and branch-bound boost to the HVAC system [146]. Another examining building operation schedules [153] or simplifying the occu­
study looked at the effectiveness of an ANN-based adaptive PMV control pancy model as occupancy rate [154]. In addition, with the new
algorithm in a residential house and discovered that it was more effec­ development of data collection methods, models that target specific
tive than non-adaptive algorithms for improving control and occupants will be proposed. A model simulating energy consumption on

Fig. 8. The existing workflow of IAQ and thermal comfort prediction and the potential improvement.

11
W. Zhang et al. Renewable and Sustainable Energy Reviews 167 (2022) 112704

the personal level and considering the gender difference was proposed in involve multiple learning algorithms, with its own set of strengths and
Ref. [73] and concluded that females tend to use more energy than disadvantages, making it even more difficult to choose the best method.
males. Making several assumptions and testing various approaches is a frequent
Although HVAC is usually required to provide comfortable, pro­ solution. A more comprehensive estimation can be obtained by training
ductive, and healthy surroundings, it also uses a large amount of energy various models and combining the prediction outcomes. Consequently,
[155]. However, occupants have many adaptive opportunities and other it is vital to summarise the data for various applications to assist re­
energy-relevant behaviours to minimise consumption. Furthermore, two searchers in developing better prediction models. A list of popular ma­
behaviours can achieve the same goal, for instance, adding more chine learning algorithms for different applications in the existing
clothing and turning on or adjusting a heater can both lead to warming a literature is made in Table 5.
person, but at different levels of efficiency, price, and energy inten­ Therefore, new prediction methods that distinguish different types of
siveness. Most machine learning models of energy consumption only activities and the personnel management system are required for future
evaluate and discuss a single behaviour without considering their energy consumption models and fill the research gap. Like the methods
correlated relationship. It could be due to the ML algorithm re­ discussed in the earlier sections, future energy models could benefit
quirements for the data structure and simplifying the model. Therefore, from more advanced occupancy data collection methods or integrated
choosing suitable inputs and model structures is critical for the predic­ sensor systems, which can better capture the dynamic variations and
tion method and affects accuracy and performance. make the necessary adjustments to the HVAC system.
The most often utilised methods for building energy estimates using
historical data are regression and ANN models [156]. The performance 5. Validation of the prediction models: case study and time
of different data-driven models may differ from residential, commercial, series
and office buildings when picking the best strategy for a certain case.
Most researchers would use a trial-and-error method to find the best Most studies include a validation stage or process after obtaining the
model performer for a certain structure instead of assuming a universal results, which evaluates the proposed model’s accuracy and applica­
model and applying it to all building types. In general, ANN prefers bility. The leave-one-out cross-validation approach is the most common
environmental, time index inputs [48]. Ensemble models, which validation method. The entire data set is usually separated into three
combine numerous models due to the nature of energy use in buildings, sections: training stage, verification, and testing. The majority of the
are more likely to produce accurate predictions than single models data is normally used for training (more than 70%), while the rest is used
[157]. A few studies reached better results with the ensemble techniques for testing and model validation [128]. The result from machine
than the single method. For example, the performance of three ANN learning methods will be compared with the validation data collected to
models – Feed Forward Neural Network (FFNN), radial basis function evaluate the method’s accuracy.
network (RBFN), and adaptive neuro-fuzzy interference system (ANFIS) In the reviewed studies, as shown in Fig. 10, most research (90.7%)
– was compared to the ensemble of these three models, and the ensemble conducted field experiments in existing buildings or testbeds to test and
model produced the best accurate prediction results [158]. validate the proposed method, while others used simulation-based in­
One major challenge to the machine learning model is the large vestigations. Using historical occupancy data or other data collected as
number of algorithms available, making it difficult to determine which the input, the prediction accuracy can be up to 95% [104]. For experi­
one should be used for a given task. The type of data provided de­ mental studies, the implementation scale in reviewed studies varies from
termines the learning methods. Statistical models are classified as linear a small testbed [164] to the whole building [15]. Many energy-related
or nonlinear based on whether they are used to solve linear or nonlinear experiments are conducted in a whole building, while most occupancy
problems. After appropriate data transformations, nonlinear issues can prediction models use selected rooms inside a building for the case study
be turned into linear ones. Aside from the differences, one model may (Fig. 9). Some research separates the testbed into zones to compare

Table 5
Summary of the commonly used machine learning algorithms for different applications.
Application Algorithm Suitable Cases Accuracy Ref.

Occupancy State Decision tree and HMM Decision tree is suitable for current state detection and 86.2%-93.2% [159]
Prediction HMM for future state
CNN Good with images 89.39% [160]
DNN Suitable for resource constrained devices used in IoT- Ranging from 97.27% to 98.90%. [82]
based applications
Indoor Air Quality LSTM Outperform other algorithms with real-time collected 96% [161]
Prediction data
2
Markov model and ANN Markov model for comfort assessment and ANN for R = 0.92. [162]
CO2 predictions
SVM, AdB, RF, GB, LR, and MLP MLP outperformed in the study for CO2 forecasting The best RMSE for MLP is 33.78 [140]
Energy Consumption k-means cluster Better fitting for time series withless mobility of 15% error [15]
Prediction occupants or the rooms with larger capacity
ANN four Back-propagation neural Levenberg–Marquardt Back-propagation has better Error rate is 1.07–2.23% [113]
network performance in forecasting electricity consumption
SVR, LMSR, KNN and NB Regression models fit for modelling daily electricity Varies from 0.82 to 0.98 for heat [95]
and heat demand consumption and 0.87–0.97 for electricity
consumption
LSTM and NNARX and MLP LSTM models reduce prediction error by 50%. The error is under 0.35 [163]
Thermal Comfort SVC and ANN Suitable for single room residences with the phone Above 95% [144]
Prediction application
ANNs and SVM, PMV, aPMV, and ANNs model is effective in natural ventilated ANNs model had the highest R (0.6984) and [143]
ePMV residential buildings R2 (0.4872) values
Linear Discriminant Analysis (LDA), Could be combined with the real-time control system Up to 84% [94]
KNN, DT, NB, SVM, and RF classifiers
LSTM Can accurately forecast overheating conditions Over 95% [107]
throughout the year

12
W. Zhang et al. Renewable and Sustainable Energy Reviews 167 (2022) 112704

different methods [165], while others define a small area as a testbed to 6. Discussion and recommendation for future work
check the prediction method [94]. The selection of implementation scale
is often related to experiment design, and the challenges researchers Results of the literature evaluation showed that the application of
faced ranged from communication issues with facility managers to machine learning in building occupancy prediction has significantly
equipment [96] and sensors malfunction, which should be considered grown in recent years. The number of studies focusing on occupancy
before conducting similar experiments [164]. state predictions outnumbered other applications in the early years. The
Some of the studies use public occupancy datasets to test the pre­ focus of occupancy prediction research is shifting from simply deter­
diction models they proposed. For example, one research employed mining whether there are people inside a room toward more compli­
ASHRAE Global Thermal Comfort Database with data from 52 field cated objects such as the occupant’s motion, resulting in more accurate
studies conducted in 160 buildings around the world [166]. This data­ building simulation models and better building service operation.
base is also used in another project to study the subjective metrics used The Internet of Things, which allows affordable deployment of sen­
for the assessment of the occupants’ thermal experience [167]. Another sors and controllers, has also promoted the adoption of occupancy
example is the American time use survey (ATUS) conducted by the U.S. prediction models on a larger scale. In the existing literature, there is
Bureau of Labor Statistics as an annual survey to record the respondent’s currently no one-size-fits-all model in sufficient detail to allow model
activities and locations on a regular day [168]. Another dataset con­ repeating. The way different models describe occupants is inconsistent,
ducted in 2015 in Berkeley, California includes whole-building and making it hard to represent a uniform format for simulation programs.
end-use energy consumption, HVAC system operating conditions, in­ The data collection method and prediction algorithm for different ap­
door and outdoor environmental parameters, as well as occupant counts plications should be selected carefully as the data structure, and algo­
[169]. With the awareness of the importance of occupancy behaviour, rithm adaptation are interconnected with each other.
there will be more datasets available in the future and validated by the The review also highlights the importance of combining different
scientific community. types of data collection methods and sensors to capture the dynamic
Also, the time series they meant to predict can be divided into long- variations within buildings and make the necessary adjustments. For
term and short-term prediction, which varies from a few seconds [165] example, vision-based and environmental sensors can be combined, and
to two years [95]. Across all reviewed studies, about a third summarise the benefits of both strategies will be inherited. The work also high­
the prediction result into 24-h or daily typical profiles, which can be lighted the potential of an all-in-one solution that can detect the occu­
transferred as occupancy profiles for existing building energy modelling pancy information and behaviour/activities and use the data to not only
software [170]. As shown in Fig. 10, short-term, long-term, and 24-hour reduce the energy use but also enhance the IAQ and thermal comfort,
predictions each contribute about one-third of reviewed papers for all which to date, previous works have not addressed.
regions. Short-term predictions are more common in North America and Machine learning implementations in different stages of the occu­
Europe, while long-term predictions are more common in Asia. This pancy prediction workflow were evaluated. One of the most popular
could be due to the sensor chosen and the prediction method design algorithms in building occupancy prediction is the neural-network-
difference and most of the short-term predictions are usually tested based algorithm, particularly ANN - LSTM, which was utilised by
before the longer version. The time series in different regions is shown in more than 10 papers after 2018 [173]. LSTM is a special RNN which has
Fig. 10, as the red columns indicate the time length in implementations. a good effect in dealing with long time sequence problems, with the
The accuracy is an important index for evaluating the model’s per­ combination of ANN, it can quantify the impact of features from the
formance and the baseline could be either raw data collected from sensors and reflect them into the network together with the current time
sensors, or a baseline set before the prediction. However, because of the input to participate in training. In 2020, an experimental result showed
multiple variables that influence their performance, a straight compar­ that the LSTM models exceed multilayer perceptron models by reducing
ison of the study cases may not be the ideal method. Indeed, models are the prediction error by 50% [163]. However, the best method for a
developed for various places and periods, using data of varying quality, specific scenario differs depending on the circumstances. Before imple­
and supplemented by scripts of varying quality. Even the value used to menting the method, an examination of the data structure should be
determine accuracy in different studies differs including mean absolute performed, such as determining whether the data is linear, continuous,
percentage error, mean percentage error, RMSE, and coefficient of or otherwise and whether data mining is required.
variation of RMSE, making comparison impossible. Table 6 shows the According to the study, investigations on thermal comfort and IAQ
algorithms and accuracy index used in some of the reviewed papers, prediction using ML are rather limited compared to other domains such
which indicate the different kinds of the mean for accuracy determina­ as occupancy prediction and energy consumption prediction. According
tion used in various models. to the study, there is a growing trend of research into occupant comfort
and occupancy-centric comfort systems. The concept of thermal comfort
is changing from physical index like PMV to occupant’s overall comfort,

Fig. 9. The implementation scale of different prediction models in reviewed studies.

13
W. Zhang et al. Renewable and Sustainable Energy Reviews 167 (2022) 112704

Fig. 10. The prediction timeframe and experimental methods conducted in different regions based on the reviewed studies.

which needs more attention in future works [174]. Occupants’ behav­ prediction and its applications and the commonly used framework for
iour, including operating the HVAC system, is driven by their satisfac­ the occupancy prediction method. As interest in this area rises, it’s
tion with the overall comfort and leads to changes in energy critical to establish a path for future models that takes a more consistent
consumption. Therefore, advanced models in the future which maintain approach. It was observed that studies were mostly concentrated in the
comfort and minimise energy consumption will have a promising future. US, Europe and Asia, and more research is required in other regions.
Individual occupant diversity should also be considered, and future Furthermore, most studies were conducted in academic/office buildings
models could include exact comfort measures and responses gathered raising the question of the applicability of the proposed methods in other
through thermal-based data collection methods such as thermal cameras indoor environments. Research on thermal comfort and IAQ prediction
and thermal comfort rating apps. using ML is rather limited compared to other domains such as occupancy
prediction and energy consumption prediction. Some promising sensors
7. Conclusion and data collection methods, including vision cameras and thermal
voting applications, are concluded in this review. The neural network-
Overall, this review provides an in-depth investigation of occupancy based algorithm is the most popular method in building machine

14
W. Zhang et al. Renewable and Sustainable Energy Reviews 167 (2022) 112704

Table 6 [12] Rebaño-Edwards S. Modelling perceptions of building quality—a neural network

Summary of the algorithm, prediction time and accuracy in some of the approach. Building and Environment 2007;42(7):2762–77. https://doi.org/
10.1016/j.buildenv.2006.07.018. Building and Environment, vol. 42, pp.
reviewed studies. 2762–2777, 07/01.
Ref. Year Prediction Algorithm Accuracy [13] P. Tuohy, M. Humphreys, F. Nicol, H. Rijal, and J. A. Clarke, "Occupant behaviour
time in naturally ventilated and hybrid buildings," 01/01 2009.
[14] Nastasi B, Markovska N, Puksec T, Duić N, Foley A. Renewable and sustainable
[130] 2021 30 min and Inhomogeneous 86% and 68% for energy challenges to face for the achievement of Sustainable Development Goals.
5 min Markov chain lighting and HVAC Renew Sustain Energy Rev 2022/04/01/2022;157:112071. https://doi.org/
systems 10.1016/j.rser.2022.112071.
[171] 2021 6 months ANN and fuzzy logic Reduce the average [15] Ding Y, Chen W, Wei S, Yang F. An occupancy prediction model for campus
techniques RMSE by 35% buildings based on the diversity of occupancy patterns. Sustain Cities Soc 2021/
[172] 2021 9 months LSTM RMSE reduced from 01/01/2021;64:102533. https://doi.org/10.1016/j.scs.2020.102533.
37% to 24% [16] Turley C, Jacoby M, Henze G, Pavlak G. Development and evaluation of
occupancy-aware model predictive control for residential building energy
[129] 2021 24 h Multilayer Perceptron 86.69% accuracy for
efficiency and occupant comfort. IOP Conf Ser Earth Environ Sci 11/21 2020;588:
ANN classification and
022043. https://doi.org/10.1088/1755-1315/588/2/022043.
RMSPE of 0.29 for [17] Winkler D, Yadav A, Chitu C, Cerpa A. OFFICE: optimization framework for
occupancy counting improved comfort & efficiency. 2020. p. 265–76.
[120] 2021 24 h LSTM cells in RNN RMSE of 4.48% [18] Klepeis NE, et al. The National Human Activity Pattern Survey (NHAPS): a
algorithms resource for assessing exposure to environmental pollutants. J Expo Sci Environ
[162] 2021 2 months Markov model for R2 = 0.92 Epidemiol 2001/07/01 2001;11(3):231–52. https://doi.org/10.1038/sj.
comfort assessment and jea.7500165.
ANN for CO2 [19] Dalamagkidis K, Kolokotsa D, Kalaitzakis K, Stavrakakis G. Reinforcement
predictions learning for energy conservation and comfort in buildings. Build Environ 07/10
[45] 2019 6 months ANN MSE error is 0:003189 2006;42:2686–98. https://doi.org/10.1016/j.buildenv.2006.07.010.
[20] Ortiz MA, Kurvers SR, Bluyssen PM. A review of comfort, health, and energy use:
understanding daily energy use and wellbeing for the development of a new
approach to study comfort. Energy Build 2017/10/01/2017;152:323–35. https://
learning prediction among the reviewed papers. The validation method doi.org/10.1016/j.enbuild.2017.07.060.
and timeframe in reviewed papers are discussed and the future recom­ [21] Ye G, Yang C, Chen Y, Li Y. A new approach for measuring predicted mean vote
mendation for the occupancy prediction method is also made. (PMV) and standard effective temperature (SET*). Build Environ 01/31 2003;38
(1):33–44. https://doi.org/10.1016/s0360-1323(02)00027-6.
[22] P. Fanger, "Thermal comfort analysis and applications in environment engeering,"
Credit author statement 01/01 1970.
[23] deDear RJ, Brager GS. Developing an adaptive model of thermal comfort and
preference. Build Eng: Research 01/01 1998;4106.
Wuxia Zhang: Conceptualisation, Methodology, Writing- Original [24] Refrigerating, A.-C. Engineers, and A. N. S. Institute. In: Thermal environmental
Draft, Yupeng Wu: Supervision, John Calautit: Writing- Reviewing and conditions for human occupancy (no. 2004. American society of heating,
Editing. refrigerating and air-conditioning engineers; 2004.
[25] Olesen BW. The philosophy behind EN15251: indoor environmental criteria for
design and calculation of energy performance of buildings. Energy Build 2007;39
(7):740–9.
Declaration of competing interest [26] Humphreys M, Nicol F. The validity of ISO-PMV for predicting comfort votes in
every-day thermal environments. Energy Build 07/01 2002;34:667–84. https://
doi.org/10.1016/S0378-7788(02)00018-X.
The authors declare that they have no known competing financial [27] Cigler J, Prívara S, Váňa Z, Žáčeková E, Ferkl L. Optimization of predicted mean
interests or personal relationships that could have appeared to influence vote index within model predictive control framework: computationally tractable
the work reported in this paper. solution. Energy Build 2012;52:39–49. https://doi.org/10.1016/j.
enbuild.2012.05.022. 2012/09/01/.
[28] Standard 62.1-2019 – ventilation for acceptable indoor air quality (ANSI
References approved). ASHRAE; 2019.
[29] Goyal S, Barooah P, Middelkoop T. Experimental study of occupancy-based
control of HVAC zones. Appl Energy 2015;140:75–84. https://doi.org/10.1016/j.
[1] Cao X, Dai X, Liu J. Building energy-consumption status worldwide and the state-
apenergy.2014.11.064. 2015/02/15/.
of-the-art technologies for zero-energy buildings during the past decade. Energy
[30] M. Özdamar Seitablaiev and F. Umaroğulları, "Thermal comfort and indoor air
Build 2016;128:198–213. https://doi.org/10.1016/j.enbuild.2016.06.089.
quality," Int J Sci Res Innov Technol, vol. 5, pp. 90-109, 03/03 2018.
2016/09/15/.
[31] Organization WH. Air quality guidelines for Europe. Regional Office for Europe:
[2] Sbci U. Buildings and climate change: summary for decision-makers. In: Buildings
World Health Organization; 2000.
and climate change summary for decision-makers; 01/01 2009. p. 1–62.
[32] Park JY, et al. A critical review of field implementations of occupant-centric
[3] Jacobson M. Review of solutions to global warming, air pollution, and energy
building controls. Build Environ 2019;165:106351. https://doi.org/10.1016/j.
security. Energy Environ Sci 01/21 2009;2. https://doi.org/10.1039/B809990C.
buildenv.2019.106351. 2019/11/01/.
[4] Bosák L, Palko M. Energy saving building in mountain area. Adv Mater Res 10/01
[33] Kong M, Dong B, Zhang R, O’Neill Z. HVAC energy savings, thermal comfort and
2014;1057:27–34. https://doi.org/10.4028/www.scientific.net/AMR.1057.27.
air quality for occupant-centric control through a side-by-side experimental
[5] Šimić G, Devedžić V. Building an intelligent system using modern Internet
study. Appl Energy 2022;306:117987. https://doi.org/10.1016/j.
technologies. Expert Syst Appl 2003;25(2):231–46. https://doi.org/10.1016/
apenergy.2021.117987. 2022/01/15/.
S0957-4174(03)00049-6. 2003/08/01/.
[34] Sun Z, Wang S, Ma Z. In-situ implementation and validation of a CO2-based
[6] Castaldo VL, Pigliautile I, Rosso F, Cotana F, De Giorgio F, Pisello AL. How
adaptive demand-controlled ventilation strategy in a multi-zone office building.
subjective and non-physical parameters affect occupants’ environmental comfort
Build Environ 2011/01/01/2011;46(1):124–33. https://doi.org/10.1016/j.
perception. Energy Build 2018/11/01/2018;178:107–29. https://doi.org/
buildenv.2010.07.008.
10.1016/j.enbuild.2018.08.020.
[35] Kim J, Schiavon S, Brager G. Personal comfort models – a new paradigm in
[7] Erickson V, Cerpa A. Thermovote: participatory sensing for efficient building
thermal comfort for occupant-centric environmental control. Build Environ 2018/
HVAC conditioning. 2012. p. 9–16.
03/15;132:114–24. https://doi.org/10.1016/j.buildenv.2018.01.023.
[8] Karmann C, Schiavon S, Graham LT, Raftery P, Bauman F. Comparing
[36] Wang Z, et al. Individual difference in thermal comfort: a literature review. Build
temperature and acoustic satisfaction in 60 radiant and all-air buildings. Build
Environ 2018;138:181–93. https://doi.org/10.1016/j.buildenv.2018.04.040.
Environ 2017/12/01/2017;126:431–41. https://doi.org/10.1016/j.
2018/06/15/.
buildenv.2017.10.024.
[37] K. Katic, R. Li, B. Kingma, and W. Zeiler, "Modelling hand skin temperature in
[9] Roulet C-A, Flourentzou F, Foradini F, Bluyssen P, Cox C, Aizlewood C.
relation to body composition," J Therm Biol, vol. 69, 07/01 2017, doi: 10.1016/j.
Multicriteria analysis of health, comfort and energy efficiency in buildings. Build
jtherbio.2017.07.003.
Res Inf 09/01 2006;34. https://doi.org/10.1080/09613210600822402.
[38] Z. Wang, R. de Dear, M. Luo, Y. Geng, and Y. Zhu, "Individual difference in
[10] H. Yoshino, T. Hong, and N. Nord, "IEA EBC annex 53: total energy use in
thermal comfort: a literature review," 04/28 2018.
buildings – analysis and evaluation methods," Energy Build, vol. 152, 07/13
[39] E. Kamel, S. Sheikh, and X. Huang, "Data-driven predictive models for residential
2017, doi: 10.1016/j.enbuild.2017.07.038.
building energy use based on the segregation of heating and cooling days,"
[11] Barthelmes V, Becchio C, Corgnati S. Occupant behavior lifestyles in a residential
Energy, vol. 206, p. 118045, 06/01 2020, doi: 10.1016/j.energy.2020.118045.
nearly zero energy building: effect on energy use and thermal comfort. Science
and Technology for the Built Environment; 08/05 2016. p. 1–16. https://doi.org/
10.1080/23744731.2016.1197758.

15
W. Zhang et al. Renewable and Sustainable Energy Reviews 167 (2022) 112704

[40] Foucquier A, Robert S, Suard F, Stephan L, Jay A. State of the art in building Scottsdale; 2014. https://doi.org/10.1145/2666652.2666655 [Online].
modelling and energy performances prediction: a review. Renew Sustain Energy Available:.
Rev 07/01 2013;23:272–88. https://doi.org/10.1016/j.rser.2013.03.004. [67] Y. Wang, Y. Wang, L. Liu, and X. Chen, "Defocused camera calibration with a
[41] D. Coakley, P. Raftery, and M. Keane, "A review of methods to match building conventional periodic target based on Fourier transform," Opt Lett, vol. 44, no.
energy simulation models to measured data," Renew Sustain Energy Rev, vol. 37, 13, pp. 3254-3257, 2019/07/01 2019, doi: 10.1364/OL.44.003254.
pp. 123–141, 09/01 2014, doi: 10.1016/j.rser.2014.05.007. [68] Lee S, Jung S, Lee JJE. In: Prediction model based on an artificial neural network
[42] Meng Y-b, Li T-y, Liu G-h, Xu S-j, Ji T. Real-time dynamic estimation of for user-based building energy consumption in South Korea. vol. 12; 2019. no. 4.
occupancy load and an air-conditioning predictive control method based on [69] W. Wang, J. Chen, and X. Song, "Modeling and predicting occupancy profile in
image information fusion. Build Environ 02/01 2020;173:106741. https://doi. office space with a wi-fi probe-based dynamic Markov time-window inference
org/10.1016/j.buildenv.2020.106741. approach," Build Environ, vol. 124, 08/01 2017, doi: 10.1016/j.
[43] Abushakra B, Haberl JS, Claridge DEJAT. In: Overview of existing literature on buildenv.2017.08.003.
diversity factors and schedules for energy and cooling load calculations. vol. 110; [70] Z. Li and B. Dong, "Short term predictions of occupancy in commercial
2004. p. 164–76. buildings— performance analysis for stochastic models and machine learning
[44] Aftab M, Chen C, Chau C-K, Rahwan T. Automatic HVAC control with real-time approaches," Energy Build, vol. 158, 09/01 2017, doi: 10.1016/j.
occupancy recognition and simulation-guided model predictive control in low- enbuild.2017.09.052.
cost embedded system. Energy Build 08/01 2017;154. https://doi.org/10.1016/j. [71] Chen Z, Jiang C, Xie L. Building occupancy estimation and detection: a review.
enbuild.2017.07.077. Energy Build 2018;169:260–70. https://doi.org/10.1016/j.enbuild.2018.03.084.
[45] Eini R, Abdelwahed S. Learning-based model predictive control for smart building 2018/06/15/.
thermal management. 2019. p. 38–42. [72] Li D, Menassa C, Kamat V, Byon E. Heat - human embodied autonomous
[46] Zhao H-x, Magoulès F. A review on the prediction of building energy thermostat. Build Environ 2020;178:106879. https://doi.org/10.1016/j.
consumption. Renew Sustain Energy Rev 2012;16(6):3586–92. https://doi.org/ buildenv.2020.106879. 05/01.
10.1016/j.rser.2012.02.049. 2012/08/01/. [73] Lee S, Jung S, Lee J. Prediction model based on an artificial neural network for
[47] Ramokone A, Popoola O, Awelewa A, Ayodele T. A review on behavioural user-based building energy consumption in South Korea. Energies 02/15 2019;
propensity for building load and energy profile development – model inadequacy 12:608. https://doi.org/10.3390/en12040608.
and improved approach. Sustain Energy Technol Assessments 06/01 2021;45: [74] D. Maljkovic, "Modelling influential factors of consumption in buildings
101235. https://doi.org/10.1016/j.seta.2021.101235. connected to district heating systems," Energies, vol. 12, p. 586, 02/13 2019, doi:
[48] C. Kuster, Y. Rezgui, and M. Mourshed, "Electrical load forecasting models: a 10.3390/en12040586.
critical systematic review," Sustain Cities Soc, vol. 35, 08/01 2017, doi: 10.1016/ [75] Tien P, Wei S, Calautit JK, Darkwa J, Wood C. A vision-based deep learning
j.scs.2017.08.009. approach for the detection and prediction of occupancy heat emissions for
[49] Hong T, Wang Z, Luo X, Zhang W. State-of-the-Art on research and applications of demand-driven control solutions. Energy Build 08/01 2020;226:110386. https://
machine learning in the building life cycle. Energy Build 2020;212. doi.org/10.1016/j.enbuild.2020.110386.
[50] W. Wei, O. Ramalho, L. Malingre, S. Sivanantham, J. Little, and C. Mandin, [76] W. Wang, J. Chen, T. Hong, and N. Zhu, "Occupancy prediction through Markov
"Machine learning and statistical models for predicting indoor air quality," Indoor based feedback recurrent neural network (M-FRNN) algorithm with WiFi probe
Air, vol. 29, 06/20 2019, doi: 10.1111/ina.12580. technology," Build Environ, vol. 138, 04/27 2018, doi: 10.1016/j.
[51] H. Saha, A. Florita, G. Henze, and S. Sarkar, "Occupancy sensing in buildings: a buildenv.2018.04.034.
review of data analytics approaches," Energy Build, vol. 188–189, 02/01 2019, [77] Z. Sultan, A. Pantazaras, K. A. Chaturvedi, J. Yang, K. Tham, and S. E. Lee,
doi: 10.1016/j.enbuild.2019.02.030. "Predicting occupancy counts using physical and statistical Co 2 -based modeling
[52] Xilei D, Liu J, Zhang X. A review of studies applying machine learning models to methodologies," Build Environ, vol. 123, 07/01 2017, doi: 10.1016/j.
predict occupancy and window-opening behaviours in smart buildings. Energy buildenv.2017.07.027.
Build 05/01 2020;223:110159. https://doi.org/10.1016/j.enbuild.2020.110159. [78] H. Hou, J. Pawlak, A. Sivakumar, B. Howard, and J. Polak, "An approach for
[53] Yao Y, Shekhar D. State of the art review on model predictive control (MPC) in building occupancy modelling considering the urban context," Build Environ, vol.
Heating Ventilation and Air-conditioning (HVAC) field. Build Environ 05/01 183, p. 107126, 07/01 2020, doi: 10.1016/j.buildenv.2020.107126.
2021;200:107952. https://doi.org/10.1016/j.buildenv.2021.107952. [79] Pesic S, Tosic M, Iković O, Radovanovic M, Ivanovic M, Boscovic D. BLEMAT:
[54] Ding Y, Han S, Tian Z, Yao J, Chen W, Zhang Q. Review on occupancy detection data analytics and machine learning for smart building occupancy detection and
and prediction in building simulation. Build Simulat 08/04 2021;15. https://doi. prediction. Int J Artif Intell Tool 09/01 2019;28:1960005. https://doi.org/
org/10.1007/s12273-021-0813-8. 10.1142/S0218213019600054.
[55] H. Fu, J. Baltazar, and D. Claridge, "Review of developments in whole-building [80] Ashouri A, Newsham G, Shi Z, Gunay B. Day-ahead prediction of building
statistical energy consumption models for commercial buildings," Renew Sustain occupancy using WiFi signals. 2019. p. 1237–42.
Energy Rev, vol. 147, p. 111248, 09/01 2021, doi: 10.1016/j.rser.2021.111248. [81] W. Wang, T. Hong, N. Li, R. Wang, and J. Chen, "Linking energy-cyber-physical
[56] M. Esrafilian-Najafabadi and F. Haghighat, "Occupancy-based HVAC control systems with occupancy prediction and interpretation through WiFi probe-based
systems in buildings: a state-of-the-art review," Build Environ, vol. 197, p. ensemble classification," Appl Energy, vol. 236, 11/23 2018, doi: 10.1016/j.
107810, 03/01 2021, doi: 10.1016/j.buildenv.2021.107810. apenergy.2018.11.079.
[57] Lo J, Novoselac A. Localized air-conditioning with occupancy control in an open [82] W. Wang, J. Chen, and T. Hong, "Occupancy prediction through machine learning
office. Energy Build 07/01 2010;42:1120–8. https://doi.org/10.1016/j. and data fusion of environmental sensing and Wi-Fi sensing in buildings," Autom
enbuild.2010.02.003. ConStruct, vol. 94, 07/10 2018, doi: 10.1016/j.autcon.2018.07.007.
[58] Yang Z, Becerik-Gerber B. The coupled effects of personalized occupancy profile [83] Diaper G. The Hawthorne Effect: a fresh examination. Educ Stud 01/01 1990;16:
based HVAC schedules and room reassignment on building energy use. Energy 261–7. https://doi.org/10.1080/0305569900160305.
Build 08/01 2014;78:113–22. https://doi.org/10.1016/j.enbuild.2014.04.002. [84] P. Tien, S. Wei, J. K. Calautit, J. Darkwa, and C. Wood, "Occupancy heat gain
[59] Nguyen TA, Aiello M. Energy intelligent buildings based on user activity: a detection and prediction using deep learning approach for reducing building
survey. Energy Build 01/31 2013;56:244–57. https://doi.org/10.1016/j. energy demand," J Sustain Dev Energy Water Environ Syst, vol. N/A, 11/16 2020,
enbuild.2012.09.005. doi: 10.13044/j.sdewes.d8.0378.
[60] M. Reena, D. Mathew, and L. Jacob, "A flexible control strategy for energy and [85] A. Zeng, S. Liu, and Y. Yu, "Comparative study of data driven methods in building
comfort aware HVAC in large buildings," Build Environ, vol. 145, 09/01 2018, electricity use prediction," Energy Build, vol. 194, 04/01 2019, doi: 10.1016/j.
doi: 10.1016/j.buildenv.2018.09.016. enbuild.2019.04.029.
[61] Lu X, Feng F, Pang Z, Yang T, Zheng O. Extracting typical occupancy schedules [86] Arief-Ang I, Salim F, Hamilton M. In: SD-HOC: seasonal decomposition algorithm
from social media (TOSSM) and its integration with building energy modeling. for mining lagged time series; 2018. p. 125–43.
Build Simulat 2020;14(1). [87] Y. Peng, A. Rysanek, Z. Nagy, and A. Schlueter, "Occupancy learning-based
[62] Massimo V, Giretti A, Tolve L, Casals M. Model predictive energy control of demand-driven cooling control for office spaces," Build Environ, vol. 122, 06/01
ventilation for underground stations. Energy Build 01/20 2016;116. https://doi. 2017, doi: 10.1016/j.buildenv.2017.06.010.
org/10.1016/j.enbuild.2016.01.020. [88] Sama S, Rahnamay-Naeini M. A study on compression-based sequential
[63] Ferrantelli A, Kuivjõgi H, Kurnitski J, Thalfeldt M. Office building tenants’ prediction methods for occupancy prediction in smart homes. 2016. p. 1–8.
electricity use model for building performance simulations. Energies 10/22 2020; [89] Sun K, Yan D, Hong T, Guo S. Stochastic modeling of overtime occupancy and its
13:5541. https://doi.org/10.3390/en13215541. application in building energy simulation and calibration. Build Environ 2014/
[64] 6-10 Alomair B, Clark A, Cuellar J, Poovendran R. Statistical framework for 09/01/2014;79:1–12. https://doi.org/10.1016/j.buildenv.2014.04.030.
source anonymity in sensor networks. In: 2010 IEEE global telecommunications [90] Manna C, Fay D, Brown K, Wilson N. Learning occupancy in single person offices
conference GLOBECOM; 2010. p. 1–6. https://doi.org/10.1109/ with mixtures of multi-lag Markov chains. 2013. p. 151–8.
GLOCOM.2010.5684248. Dec. 2010 2010. [91] Y. Jin, D. Yan, X. Zhang, J. An, and M. Han, "A data-driven model predictive
[65] New York, NY, USA Lee P, Shin E-J, Guralnik V, Mehrotra S, control for lighting system based on historical occupancy in an office building:
Venkatasubramanian N, Smith KT. Exploring privacy breaches and mitigation Methodology Development," vol. 2021, pp. 93-114.
strategies of occupancy sensors in smart buildings. In: Presented at the [92] Rueda L, et al. Online unsupervised occupancy anticipation system Applied to
proceedings of the 1st ACM international workshop on technology enablers and residential heat load management. IEEE Access 01/01 2021;9:109806–21.
innovative applications for smart cities and communities; 2019. https://doi.org/ https://doi.org/10.1109/ACCESS.2021.3098631.
10.1145/3364544.3364827 [Online]. Available:. [93] K. Panchabikesan, F. Haghighat, and M. Mankibi, "Data driven occupancy
[66] Wang X, Tague P. Non-invasive user tracking via passive sensing: privacy risks of information for energy simulation and energy use assessment in residential
time-series occupancy measurement. In: Presented at the proceedings of the 2014 building," Energy, vol. 218, 12/13 2020, doi: 10.1016/j.energy.2020.119539.
workshop on artificial intelligent and security workshop. Arizona, USA:

16
W. Zhang et al. Renewable and Sustainable Energy Reviews 167 (2022) 112704

[94] Pigliautile I, Casaccia S, Morresi N, Arnesano M, Pisello AL, Revel G. Assessing [122] Melfi R, Rosenblum B, Nordman B, Christensen K. In: Measuring building
occupants’ personal attributes in relation to human perception of environmental occupancy using existing network infrastructure. 2011 international green
comfort: measurement procedure and data analysis. Build Environ 04/01 2020; computing conference and workshops; 01/01 2011. Orlando, FL, USA.
177:106901. https://doi.org/10.1016/j.buildenv.2020.106901. [123] D. Mandic and J. Chambers, "Recurrent neural networks for prediction: learning
[95] Li Z, Friedrich D, Harrison G. Demand forecasting for a mixed-use building using Algorithms,Architectures and stability," 01/01 2001, doi: 10.1002/047084535X.
agent-schedule information with a data-driven model. Energies 02/11 2020;13: [124] Haidar N, Tamani N, Nienaber F, Wesseling M, Bouju A. Data collection period
780. https://doi.org/10.3390/en13040780. and sensor selection method for smart building occupancy prediction. 2019.
[96] Y. De Bock, A. Auquilla, A. Nowe, and J. Duflou, "Nonparametric user activity [125] Metwaly A, Peña Queralta J, Sarker V, Nguyen gia T, Nasir O, Westerlund T. Edge
modelling and prediction," User Model User-Adapted Interact, vol. 30, 11/01 computing with embedded AI: thermal image analysis for occupancy estimation
2020, doi: 10.1007/s11257-020-09259-3. in intelligent buildings. 2019.
[97] W. Wang, T. Hong, N. Xu, X. Xu, J. Chen, and X. Shan, "Cross-source sensing data [126] Tien PW, Wei S, Calautit JK, Darkwa J, Wood C. Real-time monitoring of
fusion for building occupancy prediction with adaptive lasso feature filtering," occupancy activities and window opening within buildings using an integrated
Build Environ, vol. 162, 07/14 2019, doi: 10.1016/j.buildenv.2019.106280. deep learning-based approach for reducing energy demand. Appl Energy 2022;
[98] Wu J, Wei Y, Zhang X. Prediction of occupancy level and energy consumption in 308:118336. https://doi.org/10.1016/j.apenergy.2021.118336. 2022/02/15/.
office building using blind system identification and neural networks. 2021. [127] J. Vaňuš, O. Majidzadeh Gorjani, and P. Bilik, "Novel proposal for prediction of
p. 47–79. CO2 course and occupancy recognition in intelligent buildings within IoT,"
[99] Sangogboye F, Arendt K, Singh A, Veje C, Kjærgaard M, Jørgensen B. Performance Energies, vol. 12, p. 4541, 11/28 2019, doi: 10.3390/en12234541.
comparison of occupancy count estimation and prediction with common versus [128] I. Arief-Ang, F. Salim, and M. Hamilton, "RUP: Large room utilisation prediction
dedicated sensors for building model predictive control. Build Simulat 08/01 with carbon dioxide sensor," Pervasive Mob Comput, vol. 46, 03/01 2018, doi:
2017;10:1–15. https://doi.org/10.1007/s12273-017-0397-5. 10.1016/j.pmcj.2018.03.001.
[100] Marchelina J, Chou S-Y, Yu V, Dewabharata A, Sugiarto V, Karijadi I. Two-stages [129] G. Apostolo, F. Bernardini, L. Magalhaes, and D. Muchaluat-Saade, "A unified
occupancy number detection based on indoor environment attributes by utilizing methodology to predict wi-fi network usage in smart buildings," IEEE Access, vol.
machine learning algorithm. 2019. p. 38–43. PP, pp. 1-1, 01/01 2021, doi: 10.1109/ACCESS.2020.3048891.
[101] Ramokone A, Popoola O, Awelewa A. An intelligent approach for assessing [130] S. Salimi, Z. Liu, and A. Hammad, "Occupancy prediction model for open-plan
occupancy and occupant-related activities impact on residential electric load offices using real-time location system and inhomogeneous Markov chain," Build
profiles. 2020. p. 1–6. Environ, vol. 152, 04/01 2019, doi: 10.1016/j.buildenv.2019.01.052.
[102] Y. Wei et al., "Prediction of occupancy level and energy consumption in office [131] S. Kim, S. Kang, K. Ryu, and G. Song, "Real-time occupancy prediction in a large
building using blind system identification and neural networks," Appl Energy, vol. exhibition hall using deep learning approach," Energy Build, vol. 199, 07/01
240, pp. 276-294, 04/24 2019, doi: 10.1016/j.apenergy.2019.02.056. 2019, doi: 10.1016/j.enbuild.2019.06.043.
[103] S. A. Mumma, "Transient occupancy ventilation by monitoring CO2," ASHRAE [132] Huchuk B, Sanner S. Comparison of machine learning models for occupancy
IAQ applications, pp. 21-23, 01/01 2004. prediction in residential buildings using connected thermostat data. Build Environ
[104] Dey A, et al. Namatad: inferring occupancy from building sensors using machine 06/01 2019;160:106177. https://doi.org/10.1016/j.buildenv.2019.106177.
learning. 2016. p. 478–83. [133] Yang L, Ye M, he B-J. CFD simulation research on residential indoor air quality.
[105] Murakami Y, Terano M, Mizutani K, Harada M, Kuno S. Field experiments on Sci Total Environ 2014;472:1137–44. https://doi.org/10.1016/j.
energy consumption and thermal comfort in the office environment controlled by scitotenv.2013.11.118. 2014/02/15/.
occupants’ requirements from PC terminal. Build Environ 2007/12/01/2007;42 [134] Goyal S, Ingley HA, Barooah P. Zone-level control algorithms based on occupancy
(12):4022–7. https://doi.org/10.1016/j.buildenv.2006.05.012. information for energy efficient buildings. 2012. p. 3063–8.
[106] Khalil M, McGough S, Pourmirza Z, Pazhoohesh M, Walker S. Transfer learning [135] Kallio J, Tervonen J, Räsänen P, Mäkynen R, Koivusaari J, Peltola J. Forecasting
approach for occupancy prediction in smart buildings. 2021. p. 1–6. office indoor CO2 concentration using machine learning with a one-year dataset.
[107] Y. Yuan, J. Shim, S. Lee, D. Song, and J. Kim, "Prediction for overheating risk Build Environ 2021;187:107409. https://doi.org/10.1016/j.
based on deep learning in a zero energy building," Sustainability, vol. 12, p. 8974, buildenv.2020.107409. 2021/01/01/.
10/29 2020, doi: 10.3390/su12218974. [136] M. Killian, M. Zauner, and M. Kozek, "Comprehensive smart home energy
[108] Yu Z, Fung B, Haghighat F, Yoshino H, Morofsky E. A systematic procedure to management system using mixed-integer quadratic-programming," Appl Energy,
study the influence of occupant behavior on building energy consumption. Energy vol. 222, pp. 662-672, 07/15 2018, doi: 10.1016/j.apenergy.2018.03.179.
Build 06/01 2011;43:1409–17. https://doi.org/10.1016/j.enbuild.2011.02.002. [137] P. Das et al., "Using probabilistic sampling-based sensitivity analyses for indoor
[109] Das A, Sangogboye F, Kolvig-Raun E, Kjærgaard M. HeteroSense: an occupancy air quality modelling," Build Environ, vol. 78, 08/01 2014, doi: 10.1016/j.
sensing framework for multi-class classification for activity recognition and buildenv.2014.04.017.
trajectory detection. 2019. p. 12–7. [138] Kim M, Kim Y, Sung S, Yoo C. Data-driven prediction model of indoor air quality
[110] Killian M, Kozek M. Short-term occupancy prediction and occupancy based by the preprocessed recurrent neural networks. 2009. p. 1688–92.
constraints for MPC of smart homes. IFAC-PapersOnLine 01/01 2019;52:377–82. [139] B. Omarov, A. Altayeva, A. Demeuov, A. Tastanov, Z. Kassymbekov, and A.
https://doi.org/10.1016/j.ifacol.2019.08.239. Koishybayev, "Fuzzy controller for indoor air quality control: a sport complex case
[111] M. Mohri, A. Rostamizadeh, and A. Talwalkar, "Foundations of machine learning," study," in Advanced Informatics for computing research, Singapore, A. K. Luhach, D.
01/01 2012. S. Jat, K. H. Bin Ghazali, X.-Z. Gao, and P. Lingras, Eds., 2021//2021: Springer
[112] M. Koklu and K. Tutuncu, "Tree based classification methods for occupancy Singapore, pp. 53-61.
detection," IOP Conf Ser Mater Sci Eng, vol. 675, p. 012032, 11/15 2019, doi: [140] A. Razban and S. Taheri, "Learning-based CO2 concentration prediction:
10.1088/1757-899X/675/1/012032. application to indoor air quality control using demand-controlled ventilation,"
[113] Kim MK, Kim Y-S, Srebric J. Impact of correlation of plug load data, occupancy Build Environ, vol. 205, 07/22 2021, doi: 10.1016/j.buildenv.2021.108164.
rates and local weather conditions on electricity consumption in a building using [141] Loy-Benitez J, Vilela P, Li Q, Yoo C. Sequential prediction of quantitative health
four back-propagation neural network models. Sustain Cities Soc 06/01 2020;62: risk assessment for the fine particulate matter in an underground facility using
102321. https://doi.org/10.1016/j.scs.2020.102321. deep recurrent neural networks. Ecotoxicol Environ Saf 2019;169:316–24.
[114] Apostolo G, Bernardini F, Magalhaes L, Muchaluat-Saade D. An experimental https://doi.org/10.1016/j.ecoenv.2018.11.024. 2019/03/01/.
analysis for detecting wi-fi network associations using multi-label learning. 2020. [142] C. T. Cheung, T. Parkinson, P. Li, and G. Brager, "Analysis of the accuracy on PMV
p. 423–8. – PPD model using the ASHRAE global thermal comfort database II," Build
[115] Liang X, Hong T, Shen GQJB, Environment. Occupancy data analytics and Environ, vol. 153, 02/01 2019, doi: 10.1016/j.buildenv.2019.01.055.
prediction: a case study, vol. 102. Jun.; 2016. p. 179–92. [143] Q. Chai, H. Wang, Y. Zhai, and L. Yang, "Machine learning algorithms to predict
[116] N. Amel, B. Marhic, L. Delahoche, and J.-B. Masson, "ALOS: automatic learning of occupants’ thermal comfort in naturally ventilated residential buildings," Energy
an occupancy schedule based on a new prediction model for a smart heating Build, vol. 217, p. 109937, 03/01 2020, doi: 10.1016/j.enbuild.2020.109937.
management system," Build Environ, vol. 142, 06/01 2018, doi: 10.1016/j. [144] Issaraviriyakul A, Pora W, Panitantum N. Cloud-based machine learning
buildenv.2018.06.040. framework for residential HVAC control system. 2021. p. 12–6.
[117] Zantalis F, Koulouras G, Karabetsos S, Kandris D. A review of machine learning [145] Katić K, Li R, Zeiler W. Machine learning algorithms applied to a prediction of
and IoT in smart transportation. Future Internet 2019;11(4):94 [Online]. personal overall thermal comfort using skin temperatures and occupants’ heating
Available: https://www.mdpi.com/1999-5903/11/4/94. behavior. Appl Ergon 2020;85:103078. https://doi.org/10.1016/j.
[118] LeCun Y, Bengio Y, Hinton G. Deep learning. Nature 2015/05/01 2015;521 apergo.2020.103078. 2020/05/01/.
(7553):436–44. https://doi.org/10.1038/nature14539. [146] P. M. Ferreira, S. M. Silva, A. E. Ruano, A. T. Négrier, and E. Z. E. Conceição,
[119] Arvidsson S, Gullstrand M, Sirmacek B, Riveiro M. Sensor fusion and "Neural network PMV estimation for model-based predictive control of HVAC
convolutional neural networks for indoor occupancy prediction using multiple systems," in The 2012 international joint conference on neural networks (IJCNN),
low-cost low-resolution heat sensor data. Sensors 2021;21(4). https://doi.org/ 10-15 June 2012 2012, pp. 1-8, doi: 10.1109/IJCNN.2012.6252365.
10.3390/s21041036. [147] Moon JW. Performance of ANN-based predictive and adaptive thermal-control
[120] Kim M, Jun JA, Kim N, Song YJ, Pyo CS. Sequence-to-Sequence model for methods for disturbances in and around residential buildings. Build Environ
building energy consumption prediction. In: 2018 international conference on 2012;48:15–26. https://doi.org/10.1016/j.buildenv.2011.06.005. 2012/02/01/.
information and communication technology convergence (ICTC); 2018. [148] Wu Z, et al. Using an ensemble machine learning methodology-Bagging to predict
[121] Mitra A, Ngoko Y, Trystram D. Smart Oracle based building management system. occupants’ thermal comfort in buildings. Energy Build 2018;173:117–27. https://
In: IEEE International Conference on Smart Computing (SMARTCOMP); 2021. doi.org/10.1016/j.enbuild.2018.05.031. 2018/08/15/.
p. 61–8. https://doi.org/10.1109/SMARTCOMP52413.2021.00029. 23-27 Aug. [149] Chari A, Christodoulou S. Building energy performance prediction using neural
2021 2021. networks. Energy Effic 2017;10(5):1315–27. https://doi.org/10.1007/s12053-
017-9524-5. 2017/10/01.

17
W. Zhang et al. Renewable and Sustainable Energy Reviews 167 (2022) 112704

[150] Fumo N, Mago P, Luck R. Methodology to estimate building energy consumption [162] L. Tagliabue, F. Re Cecconi, S. Rinaldi, and A. Ciribini, "Data driven Indoor air
using EnergyPlus Benchmark Models. Energy Build 2010;42(12):2331–7. https:// quality prediction in educational facilities based on IoT network," Energy Build,
doi.org/10.1016/j.enbuild.2010.07.027. 2010/12/01/. vol. 236, p. 110782, 02/01 2021, doi: 10.1016/j.enbuild.2021.110782.
[151] C. Clevenger, J. Haymaker, and M. Jalili, "Demonstrating the impact of the [163] F. Mtibaa, K.-K. Nguyen, M. Azam, A. Papachristou, J.-S. Venne, and M. Cheriet,
occupant on building performance," J Comput Civ Eng, vol. 28, pp. 99-102, 01/01 "LSTM-based indoor air temperature prediction framework for HVAC systems in
2014, doi: 10.1061/(ASCE)CP.1943-5487.0000323. smart buildings," Neural Comput Appl, vol. 32, 12/01 2020, doi: 10.1007/
[152] Wei Y, et al. A review of data-driven approaches for prediction and classification s00521-020-04926-3.
of building energy consumption. Renew Sustain Energy Rev 2018;82:1027–47. [164] S. Gilani and B. Gunay, "Simulating occupants’ impact on building energy
https://doi.org/10.1016/j.rser.2017.09.108. 2018/02/01/. performance at different spatial scales," Build Environ, vol. 132, pp. 327-337, 03/
[153] Massana J, Pous C, Burgas L, Melendez J, Colomer J. Short-term load forecasting 01 2018, doi: 10.1016/j.buildenv.2018.01.040.
for non-residential buildings contrasting artificial occupancy attributes. Energy [165] Rahaman M, et al. OccuSpace: towards a robust occupancy prediction system for
Build 2016;130:519–31. https://doi.org/10.1016/j.enbuild.2016.08.081. 2016/ activity based workplace. 2019. p. 415–8.
10/15/. [166] Földváry Ličina V, et al. Development of the ASHRAE global thermal comfort
[154] L. Truong et al., "Accurate prediction of hourly energy consumption in a database II. Build Environ 2018;142:502–12. https://doi.org/10.1016/j.
residential building based on the occupancy rate using machine learning buildenv.2018.06.022. 2018/09/01/.
approaches," Appl Sci, vol. 11, p. 2229, 03/03 2021, doi: 10.3390/app11052229. [167] Wang Z, Wang J, He Y, Liu Y, Lin B, Hong T. Dimension analysis of subjective
[155] Fan C, Xiao F, Zhao Y. A short-term building cooling load prediction method using thermal comfort metrics based on ASHRAE Global Thermal Comfort Database
deep learning algorithms. Appl Energy 2017;195:222–33. https://doi.org/ using machine learning. J Build Eng 2020;29:101120. https://doi.org/10.1016/j.
10.1016/j.apenergy.2017.03.064. 2017/06/01/. jobe.2019.101120. 2020/05/01/.
[156] T. Ahmad, H. Chen, Y. Guo, and J. Wang, "A comprehensive overview on the data [168] Statistics US BoL. American Time Use Survey. [Online]. Available: http://www.bls.
driven and large scale based approaches for forecasting of building energy gov/tus/#data.
demand: a review," Energy Build, vol. 165, pp. 301–320, 04/10 2018, doi: [169] N. Luo et al., "A three-year dataset supporting research on building energy
10.1016/j.enbuild.2018.01.017. management and occupancy analytics," Sci Data, vol. 9, no. 1, p. 156, 2022/04/
[157] Wang Z, Wang Y, Srinivasan RS. A novel ensemble learning approach to support 05 2022, doi: 10.1038/s41597-022-01257-x.
building energy use prediction. Energy Build 2018;159:109–22. https://doi.org/ [170] S. D’Oca and T. Hong, "Occupancy schedules learning process through a data
10.1016/j.enbuild.2017.10.085. 2018/01/15/. mining framework," Energy Build, vol. 88, 02/01 2015, doi: 10.1016/j.
[158] Jovanović RŽ, Sretenović AA, Živković BD. Ensemble of various neural networks enbuild.2014.11.065.
for prediction of heating energy consumption. Energy Build 2015;94:189–99. [171] S. Chen, Y. Ren, D. Friedrich, Z. Yu, and J. Yu, "Prediction of office building
https://doi.org/10.1016/j.enbuild.2015.02.052. 2015/05/01/. electricity demand using artificial neural network by splitting the time horizon for
[159] Ryu SH, Moon HJ. Development of an occupancy prediction model using indoor different occupancy rates," Energy AI, vol. 5, p. 100093, 06/01 2021, doi:
environmental data based on machine learning techniques. Build Environ 2016; 10.1016/j.egyai.2021.100093.
107:1–9. [172] A. Chong, G. Augenbroe, and D. Yan, "Occupancy data at different spatial
[160] Tien Paige Wenbin, et al. A deep learning framework for energy management and resolutions: building energy performance and model calibration," Appl Energy,
optimisation of HVAC systems. In: International Conference on Sustainable vol. 286, p. 116492, 01/21 2021, doi: 10.1016/j.apenergy.2021.116492.
Energy and Green Technology, 463. IOP Conf. Ser.: Earth Environ. Sci.; 2020, [173] Jiang L, Wang X, Wang L, Shao M, Zhuang L. A hybrid ANN-LSTM based model
012026. In press. for indoor temperature prediction. In: IEEE 16th conference on industrial
[161] E. Hitimana, G. Bajpai, R. Musabe, L. Sibomana, and K. Jayavel, "Implementation electronics and applications (ICIEA), 1-4; 2021. p. 1724–8. https://doi.org/
of IoT framework with data analysis using deep learning methods for occupancy 10.1109/ICIEA51954.2021.9516151. Aug. 2021 2021.
prediction in a building," Future Internet, vol. 13, p. 67, 03/09 2021, doi: [174] J. Xie, H. Li, C. Li, J. Zhang, and M. Luo, "Review on occupant-centric thermal
10.3390/fi13030067. comfort sensing, predicting, and controlling," Energy Build, vol. 226, p. 110392,
08/01 2020, doi: 10.1016/j.enbuild.2020.110392.

18