2540 IEEE INTERNET OF THINGS JOURNAL, VOL. 6, NO.

2, APRIL 2019

Thermal Comfort Modeling for Smart Buildings:

A Fine-Grained Deep Learning Approach
Wei Zhang , Member, IEEE, Weizheng Hu , and Yonggang Wen , Senior Member, IEEE

Abstract—The emerging Internet of Things (IoT) technology The most prevalent thermal comfort model should be the
enables smart building management and operation to improve predict mean vote (PMV) proposed by Fanger et al. [5] and
building energy efficiency and occupant thermal comfort. In this adopted in the ASHRAE Standard 55. PMV relates thermal
paper, we perform data analysis using the IoT generated build-
ing data to derive accurate thermal comfort model for smart comfort with six comfort factors, four indoor environmental
building control. Deep neural network (D NN) is used to model ones including temperature, humidity, mean radiant temper-
the relationship between the controllable building operations and ature (MRT), and air velocity and two vital ones including
thermal comfort. As thermal comfort is determined by multiple metabolic rate and cloth insulation. Based on the factor val-
comfort factors, a fine-grained architecture is proposed, where ues, the thermal comfort score can be calculated ranging from
an exclusive model is trained for each factor and accordingly the
corresponding thermal comfort can be evaluated. The experimen- −3 to 3, where the seven integers in ascending order within
tal results show that the proposed fine-grained D NN outperforms the range indicate cold, cool, slightly cool, comfort, slightly
its coarse-grained counterpart by 3.5× and is 1.7×, 2.5×, 2.4×, warm, warm, and hot, respectively.
and 1.9× more accurate compared to four popular machine Despite PMV’s prevalence of thermal comfort modeling, it
learning algorithms. Besides, D NN’s performance promotes with is not readily applicable to smart buildings. The comfort fac-
deeper network topology and more neurons, and a simple topol-
ogy with the same number of neurons per network hidden tors such as indoor temperature and humidity are not directly
layer is sufficient to achieve high modeling accuracy. Finally, controllable variables of the HVAC system. Rather, they are the
the derived thermal comfort model reveals a linear relation- results of the complex synergy between the outdoor environ-
ship between comfort and air conditioning setpoint. The linear ment, HVAC controllable operations and the other parameters.
property helps quickly and accurately search for the optimal For example, as we will show in the later part of this paper,
controllable setpoint with the desired comfort.
the difference between the HVAC temperature setpoint and
Index Terms—Deep learning, smart building, smart city, indoor temperature varies in time and can be significant in
thermal comfort. practice. Thereby in a typical smart building, the building man-
agement system (BMS) would like to emulate the resulting
comfort level and energy consumption of a variety of control-
I. I NTRODUCTION
lable HVAC settings before the real deployment. Accordingly,
UILDING accounts for 40% of worldwide energy usage
B and 60% electricity usage [1]. A dominate fraction of
the usage is contributed by the building’s heating, ventila-
the BMS can choose the best setting with satisfied comfort
and minimized energy usage and deploys the corresponding
controllable operations in real settings. Thus, new thermal
tion, and air conditioning (HVAC) system. In subtropical areas comfort models which link comfort and the controllable HVAC
like Australia, HVAC even consumes about 70% of the build- operations need to be developed to realize the accurate smart
ing energy [2]. The chief purpose of the HVAC system is building control.
to maintain thermal comfort for the building occupants, who Recent development of the Internet of Things (IoT) tech-
nowadays spend 90% of their time in buildings [3]. With poor nology helps derive the new comfort models, where a wide
comfort in buildings, occupants tend to suffer from sick build- spectrum of the IoT solutions already exist in the mar-
ing syndrome, absenteeism, and cognitive degradation [4]. ket [6]–[8]. Many nowadays buildings are well instrumented
Thus, it is important to maintain a healthy and comfortable with the IoT facilities to monitor the building conditions such
indoor environment for the well-being of the occupants and as HVAC status, energy usage, and ambient environment [9]–
meanwhile minimize the building energy usage. A critical [11]. The building data is collected using the associated IoT
step toward this goal is to create accurate models for thermal data acquisition network and can be accessed by the BMS
comfort. for online data analysis. In this paper, we aim to advance the
Manuscript received July 24, 2018; revised August 29, 2018; accepted thermal comfort modeling for smart buildings using the IoT
September 17, 2018. Date of publication September 20, 2018; date of current generated building data and the deep neural network (D NN).
version May 8, 2019. This work was supported by the Singapore National As a popular deep learning algorithm, D NN has proved its
Research Foundation (NRF) via the Green Buildings Innovation Cluster
(GBIC) administered by the Building and Construction Authority under Grant excellent capability in the applications such as image process-
NRF2015ENC-GBICRD001-012. (Corresponding author: Wei Zhang.) ing, computer graphics, and signal processing [12], [13]. This
The authors are with the School of Computer Science and paper tries to extend D NN’s excellence to comfort modeling
Engineering, Nanyang Technological University, Singapore 639798 (e-mail:
[email protected]; [email protected]; [email protected]). and investigates its feasibility for bridging the gap between
Digital Object Identifier 10.1109/JIOT.2018.2871461 the controllable HVAC operations and thermal comfort.
2327-4662 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2021 at 15:55:36 UTC from IEEE Xplore. Restrictions apply.
ZHANG et al.: THERMAL COMFORT MODELING FOR SMART BUILDINGS 2541

In addition, a fine-grained modeling architecture is proposed presented in Section VI. Finally, Section VII concludes this
to promote the modeling accuracy. Most existing thermal paper and suggests several future research directions.
comfort models adopt coarse-grained architecture, using only
one model to link the input attributes and thermal comfort II. R ELATED W ORKS
directly [14], [15]. In the proposed fine-grained architecture, Thermal comfort modeling has received much attention in
an exclusive model is created and trained for each comfort fac- the recent decades. The most widely adopted model should be
tor like temperature. Then the factor values from the multiple PMV proposed by Fanger et al. [5] in the 1970s. ASHRAE
fine-grained models are combined to calculate the final ther- Standard 55 [16] recommends the buildings to maintain a
mal comfort score. To valid the soundness of our proposed PMV within ±0.5 for a comfortable environment.
fine-grained deep learning approach, extensive data analytics While PMV score can be mathematically calculated with
is performed using D NN and several other popular machine the six comfort factors, recent studies also try to use ML
learning (ML) algorithms, including neural network (N N), to model the relationship between the factors and score.
linear regression (L R), and support vector machine (S VM), Megri et al. [14] achieved the goal using ε-S VM. Tested on
regression with linear kernel (S VR - L), and nonlinear RBF a small-scale dataset with less than 800 training samples and
kernel (S VR - R). Key research findings are as follows. less than 20 testing samples, the proposed ε-S VM comfort
1) Fine-grained modeling outperforms the coarse-grained modeling can approximate the PMV calculations by up to
modeling for D NN by 3.5× and the other tested ML 99%. Similarly, Atthajariyakul and Leephakpreeda [15] inves-
algorithms by over 20%. tigated the feasibility of using N N. A network topology with
2) D NN’s modeling accuracy promotes with more hid- two hidden layers are adopted and the derived model also
den layers and more neurons in the network topology, approximates the PMV model well. Besides, other comfort
although a diminishing return of adding more layers and factors such as gender and age are incorporated to improve
neurons is observed. Also, a simple topology with the the modeling accuracy [17], [18]. Hu et al. [19] monitored
same number of neurons per hidden layer is sufficient to the occupant vital signals like heartbeat rate using wearable
produce remarkable performance compared to the more devices and predict thermal comfort accordingly using N N.
complex topologies. PMV model is also inverted to derive the comfortable tem-
3) Thermal comfort model is not trivial in nature. Linear perature, which is often assumed to be equal to the HVAC
ML algorithms L R and S VR - L fail to deliver good setpoint. Javed et al. [20] adopted the random N N, where the
modeling accuracy compared to the nonlinear algo- indoor temperature is the target to model and the PMV score
rithms D NN, N N, and S VR - R. Besides, D NN achieves as well as the rest five comfort factors are used as the input
the best modeling accuracy with moderate time usage attributes. Besides, some works avoid the complex PMV calcu-
on model training. Compared to N N, L R, S VR - L, and lations by assuming certain simplified thermal comfort models,
S VR - R, D NN is 1.7×, 2.5×, 2.4×, and 1.9× more i.e., a quadratic function of the difference between the indoor
accurate, respectively. Also, thermal comfort is lin- temperature and the most comfortable temperature [21].
early related to the HVAC setpoint, which enables to Despite the progress of thermal comfort modeling, the above
compute the optimal setpoint for the desired comfort works are parameterized by the comfort factors unable to be
instantaneously. controlled directly. To realize smart building control, this paper
While the main goal of this paper is thermal comfort aims to model the relationship between the controllable HVAC
modeling and smart building control, the work also has other operations and thermal comfort. So that the BMS can use the
benefits like demand response and power grid stabilization. derived model to emulate the thermal condition with different
For example, during the peak hours with high energy demand operational settings, and accordingly choose the best one for
and electricity price, the thermal comfort requirement can be real deployment.
relaxed to a certain extent with less energy usage, and vice
verse. Corresponding accurate building control can be realized III. DATA D ESCRIPTION AND S YSTEM
based on the derived thermal comfort model in this paper. This A RCHITECTURES
not only helps stabilize the power grid but also reduces the
electricity bill of the building owners. Such benefits can be In this part, a building dataset adopted in this paper is intro-
substantial as building serves as a dominant energy consumer duced first. Then, the system architectures of thermal comfort
of the grid. modeling are presented.
The rest of this paper is organized as follows. We review
related works in Section II. In Section III, we describe A. Dataset
the adopted dataset and present the system architectures. This paper adopts a publicly available dataset from an IoT
Section IV investigates different modeling granularity and instrumented building [22], and the results shown in this paper
compares the performance of fine-grained modeling and can be reproduced using the dataset. The dataset is collected at
coarse-grained modeling. Section V discusses the modeling an office building in Center City Philadelphia, USA, and spans
performance using different network topologies for D NN. In for one year from July 2013 to July 2014. In total there are
Section VI, the modeling performance using different algo- 678 621 data samples in the dataset, and the attributes asso-
rithms are studied. Based on the derived thermal comfort ciated with the data samples are grouped into the following
models, a case study of the smart building control is also categories.

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2021 at 15:55:36 UTC from IEEE Xplore. Restrictions apply.
2542 IEEE INTERNET OF THINGS JOURNAL, VOL. 6, NO. 2, APRIL 2019

The architecture in Fig. 1(a) aims to model a direct rela-

tionship between the input attributes and thermal comfort.
(a) Differently, the second architecture in Fig. 1(b) explores the
relationship at finer granularity. An exclusive ML model
is introduced to model the relationship between the input
attributes and each of the thermal comfort factors, including
indoor temperature, indoor humidity, and indoor MRT. Here,
the historical data of the three indoor attributes are used as
the ground-truth of the target values. Then, the modeled fac-
tor values can be inputted into the PMV calculator to derive
the corresponding comfort score. The details of the PMV
(b) calculations can be found in [5].
The derived thermal comfort model helps realize the smart
Fig. 1. System architectures of the (a) coarse-grained and (b) fine-grained building control. Given any datetime and ambient condition,
thermal comfort modeling.
we can use the model to test the resulting thermal comfort
scores of different building configurations, i.e., HVAC settings.
Then, the configuration with satisfied comfort and high energy
1) Datetime Attributes: The raw dataset includes a string-
efficiency can be selected to deploy in real setting.
based datetime attribute. To facilitate the data processing in
The following part will compare and discuss the
ML algorithms, datetime is digitalized into three attributes,
performance difference of the coarse-grained modeling and
including the integer-based year, month, and weekday.
the fine-grained modeling.
2) Outdoor Attributes: Three outdoor attributes available in
the dataset are the temperature in Celsius, relative humidity as
IV. T HERMAL C OMFORT M ODELING G RANULARITY
a percentage and air velocity in m/s.
3) HVAC Attributes: There are two HVAC attributes: In this section, the two system architectures are investigated
a) cooling setpoint and b) heating setpoint, both in Celsius. for their comfort modeling performances using D NN.
Cooling and heating setpoints represent the configured ideal
indoor temperature under HVACs cooling mode and heating A. Experimental Settings
mode, respectively, and normally the former is higher than the The algorithm and experiment in this paper are implemented
later. in Python 3.6. The data samples are divided into the training
4) Indoor Attributes: The dataset also has three attributes dataset with 75% samples and the testing dataset with 25%
to evaluate the indoor environment, including temperature and samples. The input attributes for both architectures are nor-
MRT in Celsius and relative humidity in percentage. malized. D NN is configured with one input layer with eight
In this paper, we define a typical valid range for each neurons to match with the eight input attributes, ten hidden
attribute, i.e., weekday between 1 and 7 with 1 indicating layers each with 300 neurons, and one output layer with one
Monday, indoor temperature between 10 ◦ C and 35 ◦ C, rela- neuron for the modeling target. 10% of the training data is used
tive humidity between 10% and 90% and outdoor temperature as the validation data in each D NN iteration. Experiments are
between −15 ◦ C and 40 ◦ C. The data samples with any conducted on a laptop running 64-bit Windows 10 Pro on an
attribute value beyond the predefined range are treated as Intel Core i5-7300U CPU and using 16 GB of memory.
outliers and filtered out. We use the datetime, outdoor and The performance is evaluated using two metrics: 1) mean
HVAC attributes as the input attributes for training the ther- absolute error (MAE) and 2) improvement. The later is defined
mal comfort models. The indoor attributes are taken as the as follows. Given two solutions A and A , achieving the
ground-truth of the target attributes for model training. Details modeling error of ε and ε , respectively, A’s improvement to
will be presented in the following part. A is ε /ε. An improvement larger than 1 means A performs
better than A with smaller modeling error, and vice verse.
B. System Architectures
In this part, two thermal comfort modeling architectures are B. Algorithms With Different Modeling Granularity
introduced and illustrated in Fig. 1. The coarse-grained one is Although the same D NN algorithm is used, the comfort
shown in Fig. 1(a), where the data system feeds the attributes modeling performance varies with different modeling granu-
for datetime, outdoor, and HVAC to the ML algorithm as the larity. Here, the following eight algorithm settings are adopted
inputs and the PMV-based comfort scores as the target. to investigate the impact of modeling granularity.
Here, the PMV score can be computed based on the three The first setting denoted as DALL is based on the fine-
indoor attributes. Since an office building is considered in the grained architecture as shown in Fig. 1(b), where three comfort
dataset, this paper follows the general assumption that indoor factors are modeled. Then, DALL is modified to model two of
air velocity, occupant metabolic rate, and cloth insulation are the three factors and three settings D TH, D TM, and D HM can
unchanged during a long enough period and can be deemed as be derived, where the subscripts T, H, and M represent indoor
given. Specifically, air velocity and metabolic rate are assumed temperature, humidity, and MRT, respectively. For example,
to be 0.1 m/s and 1.2, respectively, in a typical office setting. D TH implies that temperature and humidity are modeled and

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2021 at 15:55:36 UTC from IEEE Xplore. Restrictions apply.
ZHANG et al.: THERMAL COMFORT MODELING FOR SMART BUILDINGS 2543

D H, where only humidity is modeled. D PMV’s fine-grained

counterpart DALL achieves a median improvement of 3.5×.
Overall, fine-grained modeling at comfort factor level is the
right choice for advancing the comfort modeling accuracy. In
this scenario, the thermal comfort model can be decomposed
into several building blocks, such as temperature model and
humidity model, which are relatively easier to be modeled with
(a) (b) higher accuracy. It is also important to explore more between
the input attributes and comfort factors, as modeling more
Fig. 2. Modeling error of D NN with different modeling granularity in (a) factors reveals more performance advantage than modeling
and DALL’s improvement to the other settings in (b). The improvement of 1 fewer.
is indicated by a red horizontal line in (b).

V. D NN T OPOLOGY
such implication applies to D TM and D HM also. If not mod- In this section, D NN topology is investigated in detail to
eled using D NN, temperature and MRT are assumed to be find out the ones with remarkable modeling performance.
equivalent to the middle value of the HVAC’s cooling setpoint
and heating setpoint, and humidity is assumed to be a moder- A. Network Depth and Width
ate value of 50%. Similarly, three more variants D T, D H, and
D M are introduced, which models only temperature, humidity, Two critical D NN topology parameters are the network
and MRT, respectively. Lastly, the coarse-grained architecture depth and width or the number of hidden layers and the num-
in Fig. 1(a) is considered to model the PMV-based comfort ber of neurons per hidden layer. Generally, small network due
score directly, denoted as D PMV. to its limited capacity cannot contain sufficient information
to approximate the real system model. This is especially true
when the modeled system is complex. However, large network
C. Results With Different Modeling Granularity
sometimes also fails to derive good performance due to the
The results of the above-introduced algorithm settings are reasons like over-fitting, where the trained model is biased
shown in Fig. 2. Fig. 2(a) shows the algorithms’ MAE of ten by the training data and becomes unfit for the testing data. In
independent runs. Fig. 2(b) selects a random run and reports this part, different D NN topologies are tested with the network
DALL’s improvement to the remaining settings for every test depth varies between 1 and 10 and the width varies between
data sample in the random run. 30 and 300. The results are shown in Fig. 3.
In Fig. 2, DALL achieves the best modeling accuracy with 1) Network Depth: The performance impact of the network
an average MAE of merely 0.13 and a median MAE of 0.08. depth is shown in Fig. 3(a), (c), and (e). Seen from the figures,
For the PMV variance within such tiny range, even occupants D NN’s modeling performance improves with more hidden lay-
may not be able to tell the difference subjectively. ers. The median MAE for indoor temperature improves from
Removing any comfort factor from modeling, the accuracy 0.78 ◦ C with one hidden layer to 0.56 ◦ C with ten hidden lay-
suffers. For D TH, D TM, and D HM, the median1 MAEs are ers, with around 40% improvement. Similar values also apply
0.10, 0.14, and 0.23, respectively, larger than DALL’s 0.08. to MRT. For humidity, it improves from 5.4% to 3.4%, around
Among the factors, the temperature has the highest impact on 60% improvement. The reason could be that more layers allow
the overall modeling accuracy. DALL is 2.5× better than D HM the modeled system to be decomposed into more information
without modeling temperature. When MRT and humidity are building blocks easier to be interpreted.
not modeled, DALL is only 7.1% and 39.0% better than D TH However, the improvement slows down with more layers
and D TM, respectively. This is reasonable, as the temperature added. For temperature and MRT, the MAE improvement
has a more direct impact on the thermal environment compared becomes insignificant with more than five hidden layers. While
to humidity and MRT. the median MAE for temperature modeling shows a 32%
The modeling performance further degrades when two fac- decrease from 0.78 ◦ C with one hidden layer to 0.56 ◦ C with
tors are not modeled, where D T, D H, and D M are generally five layers, doubling the number of hidden layers to ten only
noncompetitive to DALL and D TH, D TM and D HM without achieves a 5.4% decrease with 0.03 ◦ C modeling accuracy
only one factor to model. The median MAEs for D T, D H, and variation. For humidity, such phenomenon occurs even ear-
D M are 0.15, 0.39, and 0.27, respectively, and DALL outper- lier since three hidden layers. After over 50% improvement
forms them by 1.5×, 4.4×, and 2.9× regarding the median from one hidden layer to three, the median MAE of humidity
improvement. modeling merely varies from 3.5% to 3.4% with seven more
Lastly, the coarse-grained modeling D PMV fails to model hidden layers added. Such diminishing return also implies that
thermal comfort accurately and the performance gap to the the temperature and MRT models are more complex com-
fine-grained DALL is significant. Seen from Fig. 2, D PMV pared to humidity and the models’ return from more layers
achieves a median MAE of 0.34 which is better than only diminishes much slower.
1 Median values are selected for results analysis and discussion in some Besides, none of the factor models suffers from over-fitting
parts of this paper, as the mean values could be drifted substantially by the and the modeling performance does not degrade with more
results outliers, which however only occur rarely. hidden layers. One reason should be that our dataset is large

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2021 at 15:55:36 UTC from IEEE Xplore. Restrictions apply.
2544 IEEE INTERNET OF THINGS JOURNAL, VOL. 6, NO. 2, APRIL 2019

(a) (b) (a) (b) (c)

Fig. 4. Factor modeling performance using D NN with the same amount of

3000 neurons and different network topologies. (a) Temperature. (b) Humidity.
(c) MRT.

hidden layers and n neurons per layer. D NN’s modeling capa-

bility normally scales in the number of neuron links, which are
(c) (d) the basic units to record model information. Such link amount
scales linearly in the network depth, but exponentially in the
network width. As a result, the performance change is more
significant as the network widens and relatively less significant
as the network deepens.

B. Neuron Allocation
More neurons promote the comfort modeling performance
(e) (f)
as shown above. In this part, let us discuss the optimal
Fig. 3. D NN’s modeling performance with different network topologies. allocation of the neurons to each hidden layer.
(a), (c), and (e) Factor modeling results with 300 neurons per hidden layer 1) Topology Description: Five topologies with different
and the number of layers varies between 1 and 10. (b), (d), and (f) Results neuron allocations are investigated given the same total num-
with ten hidden layers and the number of neurons per hidden layer varies
between 30 and 300. ber of neurons, i.e., 3000. The first one, denoted as BASE,
evenly distributes the neurons to ten hidden layers, so each
layer has 300 neurons. The second one denoted as FAT
enough. With over half million data samples, it is less likely squeezes the topology into less hidden layers but each layer
that certain characteristics are only contained in the training has more neurons; specifically, five hidden layers each with
data instead of existing in both training data and testing data. 600 neurons. The third one denoted as T HIN stretches the
So, the derived models are not biased by the training data and topology of BASE into 20 hidden layers each with 150 neu-
keep improving with more neurons. rons. The rest two topologies do not assume the evenly
2) Network Width: Following the discussion on network distributed neurons. One is D ED, which assumes a deductive
depth, this part studies the performance impact of network topology where the information decomposes into finer units for
width, or the number of neurons per hidden layer. deeper hidden layers. The number of neurons of D ED increases
Fig. 3(b), (d), and (f) shows the results with ten hidden layers as 100, 100, 200, 200, 300, 300, 400, 400, 500, and 500 from
and the width varies between 30 and 300. Same as the finding the first hidden layer adjacent to the input layer to the tenth
for network depth, the modeling performance benefits from hidden layer. The other one I ND assumes an inductive infor-
more neurons as the network widens. Also, the improvement mation diffusion with the number of neurons in an reverse
saturates with more neurons added. order as D ED for the ten hidden layers.
One interesting fact is that the convergence speed in network 2) Performance Comparison: The modeling performance
width seems faster than that of the network depth. Take tem- for the three comfort factors using different topologies intro-
perature for example. Doubling one hidden layer to two layers duced above is shown in Fig. 4. For the temperature modeling
decreases the median MAE from 0.78 ◦ C to 0.74 ◦ C by 5.4%, results shown in Fig. 4(a), BASE achieves the best overall
and further doubling to four layers improves the MAE to performance with a median MAE of 0.56 ◦ C, which is 1.6%,
0.63 ◦ C by 17%. Regarding the network width, the first dou- 1.7%, 3.3%, and 0.9% more accurate than FAT, T HIN, D ED,
bling, from 30 neurons to 60 neurons, reduces the MAE to and I ND, respectively. Similar observation can also be found
0.69 ◦ C by 16% and further doubling to 120 neurons achieves for the humidity and MRT modeling.
a 0.60 ◦ C median MAE with 15% improvement. Similar One implication is that BASE evenly distributes the same
behavior can also be observed for MRT. One reason is that number of neurons to each of the ten hidden layers is sufficient,
the increase in the number of hidden layers and the num- or even the best, for D NN to achieve a superior modeling
ber of neurons per hidden layer have the different impact on performance. While D ED and I ND make certain assumptions
D NN. In D NN, the neurons between two adjacent layers are such as deductive and inductive information diffusion, it turns
fully connected, producing O(mn2 ) neuron links in total for m out that the assumptions are not true enough and the diffusion

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2021 at 15:55:36 UTC from IEEE Xplore. Restrictions apply.
ZHANG et al.: THERMAL COMFORT MODELING FOR SMART BUILDINGS 2545

process can be more complex. As a result, a general topology

as in BASE, although simple and unadorned, can well cover
and model the complexity in the modeled systems of different
comfort factors.
Besides, a fatter or thinner general topology like BASE does
not help improve the modeling performance. Such observation
suggests that the number of hidden layers and the number of
neurons per layer should be configured to moderate values. In (a) (b) (c)
practice, trial tests can be conducted to find out a good value
pair, i.e., ten hidden layers each with 300 neurons. Fig. 5. Comfort factor modeling performance using different ML algorithms.
(a) Temperature. (b) Humidity. (c) MRT.
Specifically for humidity modeling in Fig. 4(b), T HIN per-
forms much worse than its counterparts BASE and FAT with TABLE I
less layers. This observation complies with the finding in T IME C ONSUMPTION IN H OURS OF M ODEL T RAINING FOR THE C OMFORT
Fig. 3(c), where the humidity modeling error converges fast FACTORS U SING D IFFERENT A LGORITHMS . M EAN VALUES AND
S TANDARD D EVIATIONS (STD) A RE R EPORTED
and quickly stops improving since three hidden layers. Thus,
the D NN topology can be application-dependent and the best
topology may differ for different applications.
Overall, the simple topology like BASE by allocating the
same number of neurons to each hidden layer is adequate for
D NN to achieve the performance excellence. In addition, D NN
performs quite stable with different tested topologies, with less
than 5% difference for the majority of the tests. Stability is
often a critical consideration in real applications and imple-
mentations, and D NN for the sake of this can be a suitable
algorithm also.
attributes and target comfort factors with a deeper and wider
VI. M ODELING P ERFORMANCE network. The comparison algorithms, although popular and
have been widely deployed in practice, seem to fail to dis-
Based on the above-derived conclusions, this section adopts
cover the subtle details well as D NN does and makes much
the fine-grained modeling architecture and D NN with a general
higher modeling error compared to D NN.
topology with the same number of neurons per hidden layer.
N N with a shallow network performs the best among the
The corresponding performance is investigated in detail.
four comparison algorithms. This shows that N N-related algo-
rithms, such as D NN and N N are quite suitable for comfort
A. Comfort Factor Modeling Performance factor modeling. The shallow one offers a relatively sound
Four well-known ML algorithms are introduced for basis for D NN to further improve the modeling performance.
performance comparison, including two linear ones L R and Nevertheless, the results also reveal the insufficiency of using
S VR - L and two nonlinear ones N N and S VR - R. Let us describe a shallow network, i.e., one hidden layer only, for fac-
the algorithm settings first. tor modeling and more complex networks are necessary for
1) Algorithm Configuration: D NN is configured with ten achieving higher modeling accuracy.
hidden layers each with the same 300 neurons according to the Besides, the relationship between the input attributes and
conclusions in Section V. The four ML algorithms are imple- target comfort factors should not be trivially linear. According
mented using scikit-learn with the default settings, i.e., to the results, L R and S VR - L assuming a linear relationship
N N is configured with one hidden layer and a hundred neurons perform poorly compared to the rest three algorithms, which
in the layer. Although the default settings may not necessarily assume a nonlinear relationship. Even compared to the worst
achieve the best performance, they are empirically effective performed nonlinear algorithm S VR - R, the two linear ones
for the majority of the general applications. This paper thus have almost 20% higher error for the temperature and MRT
adopts the default settings for generality. modeling, and almost 15% higher for humidity.
2) Modeling Accuracy Comparison: Fig. 5 shows the factor 3) Time Consumption Comparison: Besides modeling
modeling performance for different comparison algorithms. accuracy, modeling time consumption is also a critical con-
D NN achieves the best performance for modeling all the sideration for applying a solution in practice. Thus, this part
three comfort factors, compared to the four comparison algo- presents the time consumption results for the tested algorithms
rithms. D NN’s median MAE is merely 0.56 ◦ C for indoor for modeling training. The time of testing is not discussed here
temperature and such performance is 55%, 96%, 94%, and since testing normally finishes instantaneously. Table I shows
66% better than N N, L R, S VR - L, and S VR - R, respectively. the results.
Similar improvements also apply to the MRT modeling. For Seen from the table, with over half million data samples for
humidity, D NN is 55%, 87%, 86%, and 63% better than the training, D NN still maintains a manageable time consumption
four comparison algorithms, respectively. The results show of around 1 h for one comfort factor. In practice, the mod-
that D NN can well exploit the relationship between the input els are often updated infrequently, i.e., several hours or even

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2021 at 15:55:36 UTC from IEEE Xplore. Restrictions apply.
2546 IEEE INTERNET OF THINGS JOURNAL, VOL. 6, NO. 2, APRIL 2019

(a) (b) (c)

Fig. 6. F-score of each input attribute to the modeled comfort factor in descending order. High score means strong correlation. (a) Temperature. (b) Humidity.
(c) MRT.

days. While the time results here are obtained with a com- and MRT. This motivates us to consider the weather condition
mon laptop, the training process can also be accelerated with in D NN for better modeling the comfort factors.
the high-end computing machines and high-performance com- Another outdoor attribute, outdoor air velocity, also has a
puting technologies. Regarding the time complexity, D NN is significant impact. While the air dynamics outside intuitively
linear in the number of data samples given a fixed number of is not well related to the indoor thermal status, it is corre-
optimization iterations. Thus, D NN can sustain for the even lated with the outdoor temperature, which as shown above
larger dataset in the future. impacts the factor modeling a lot. Indeed, air movement is
Humidity modeling by D NN takes less than 1 h and is faster essentially due to the geographical temperature difference [25].
compared to the temperature and MRT modeling, each taking Despite the remarkable correlation between outdoor temper-
30% and 35% more time, respectively. One reason could be ature and air velocity, the third outdoor attribute, outdoor
that the humidity model is relatively not complex and can be humidity, is hardly correlated with the indoor temperature
well modeled even with only three hidden layers as indicated and MRT.
in Fig. 3(c). b) HVAC attributes: Two HVAC attributes are among
N N and L R run much faster than D NN. Especially for the top ranked attributes. This is reasonable because that the
the later one, which takes only a couple of seconds to indoor air is directly conditioned by the HVAC system using
finish. Nevertheless, it sounds to achieve better modeling the setpoints. Heating setpoint ranks higher than the cool-
performance for evaluating occupant comfort level more accu- ing counterpart, and the reason could be as follows. The data
rately, as long as the time consumption is manageable. used in this paper is collected in Philadelphia, a subtropical
The two S VR variants are significantly slower compared to city, where the cold periods is longer than the hot periods.
the other comparison algorithms. For S VR - L with the linear Consequently, heating plays a more critical role than cooling
kernel, over 3 h are used for humidity modeling and around for air conditioning. Such ranking can vary with different geo-
6 h for temperature or MRT. With a nonlinear kernel, S VR - R graphical areas. Note that either heating setpoint for warming
is even slower, consuming around 10 h for modeling MRT air or cooling setpoint for cooling down air is effective at a
and over 15 h for temperature or humidity. S VR’s slowness time depending on the indoor thermal condition. Considering
is due to its quadratic time complexity in the number of data the full-time scale, the two HVAC setpoints together could be
samples so that S VR is unsuitable for large datasets like the higher correlated with the indoor thermal environment than
one used in this paper. With more data available in the future the individual correlation.
applications, S VR can be even slower and as a result becomes c) Datetime attributes: Three datetime attributes are con-
impractical to be adopted. sidered. The top correlated one is month, while the rest two,
weekday and year, are less correlated. Such correlation could
be due to the fact that outdoor temperature is strongly cor-
related with month, i.e., a high temperature in July and low
B. Attribute Importance for Comfort Factors
in December. Outdoor thermal environment and indoor HVAC
Fine-grained modeling at comfort factor level is adopted in regulation together determine the indoor thermal condition.
this paper. Let us first investigate the importance of each input 2) Indoor Humidity: While the indoor temperature and
attribute to the target factor. Fig. 6 shows the F-score of each MRT share the same attribute ranking, humidity shows a
input attribute to the modeled factor using univariate analysis. totally different ranking except for the outdoor temperature.
1) Indoor Temperature and MRT: Indoor temperature and Outdoor attributes rank higher for modeling humidity than
MRT share the same attribute ranking and let us discuss the temperature and MRT. Outdoor humidity, as the least corre-
results of the two factors first. lated attribute for temperature and MRT, ranks much higher
a) Outdoor attributes: It is interesting to find out that for indoor humidity modeling. Many existing HVAC systems
the outdoor temperature has the highest impact. Many exist- do not operate for adjusting humidity. Indoor humidity change
ing works assume the equivalence among the HVAC setpoint, often is a consequence of air ventilation, where a portion of
indoor temperature, and MRT [23], [24]. However, the results the conditioned air is inhaled from the outside environment to
in Fig. 6 show that the other attributes like outdoor temperature maintain the fresh indoor air. As a result, the outdoor humidity
can have an even stronger impact on the indoor temperature has a substantial impact on the indoor humidity. Due to the

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2021 at 15:55:36 UTC from IEEE Xplore. Restrictions apply.
ZHANG et al.: THERMAL COMFORT MODELING FOR SMART BUILDINGS 2547

(a) (b) (a) (b)

Fig. 7. Thermal comfort modeling performance of D NN with fine-grained Fig. 8. (a) Distribution of the temperature difference between indoor temper-
architecture and the comparison algorithms. N N, L R, S VR - L, and S VR - R ature and HVAC setpoint. Negative (positive) value means indoor temperature
adopt the fine-grained architecture and C N N, C L R, C S VR - L, and C S VR - R is lower (higher) than the HVAC setpoint. The difference of 0, or the equiv-
use the coarse-grained architecture. The improvement of 1 is indicated by a alence between the indoor temperature and HVAC setpoint, is indicated by a
red horizontal line in (b). red vertical line in (a). (b) Smart HVAC control where comfort changes with
HVAC setpoint.

similar reason, HVAC attributes, to the contrary, rank lower

performance on factor modeling as shown in Section VI-A is
for humidity modeling. This again reveals the fact of HVAC’s
well inherited by the subsequent thermal comfort calculation.
low regulation on indoor humidity in the tested case.
The mean MAE for D NN is merely 0.1 for the PMV score
Datetime attributes play an important part for humidity
ranging between −3 and 3 and the other algorithms’ mean
modeling also, where the month and year attributes show
errors are over twice higher. Regarding the median improve-
stronger impact than the weekday attribute. Typically, humid-
ment, D NN is 1.7×, 2.5×, 2.4×, and 1.9× better than N N,
ity does not vary too frequently like temperature does in day
L R, S VR - L, and S VR - R, respectively. Also, D NN not only per-
and night. The variation of humidity often occurs monthly
forms better in general, it also achieves better performance for
or seasonally, i.e., dryer air in winter and moist atmo-
most of the testing data samples compared to the other algo-
sphere in summer. This helps to explain the observation
rithms. As shown in Fig. 7(b), D NN’s improvement seldom
that the long-scale month and year attributes achieve high
goes below 1 and D NN is rarely inferior to the comparison
rankings.
algorithms.
Besides, nonlinear algorithms, D NN, N N, and S VR - L, per-
C. Thermal Comfort Modeling Performance form better than the linear ones, L R and S VR - R. This complies
Based on the derived models for comfort factors, the with the finding in Section VI-A also. Thus, as an application
PMV-based thermal comfort score can be computed and the with a complex system, thermal comfort modeling should bet-
performance of different comparison algorithms is discussed ter be conducted using nonlinear ML algorithms and the trivial
here. Experimental results regarding the MAE of ten inde- linear ones are not suitable enough.
pendent runs and D NN’s improvement on every testing data
samples of a random run are shown in Fig. 7. D. Smart HVAC Control Demonstration
D NN as well as the four comparison algorithms, including In this part, we investigate the divergence between the
N N, L R, S VR - L, and S VR - R, are investigated here, assuming indoor temperature and HVAC setpoint. Then, we demon-
a fine-grained architecture. Also, a variant assuming coarse- strate the smart building control using the D NN-based thermal
grained architecture is implemented for each of the four comfort model and how D NN deals with the divergence in
comparison algorithms, and let the variant names be C N N, reality.
C L R, C S VR - L , and C S VR - R. Seen from Fig. 7, we have the First, the indoor temperature has a significant difference to
following observations. the HVAC setpoint. Fig. 8(a) shows the difference distribution
First, Fig. 7 reaffirms that fine-grained architecture at com- when the outdoor temperature is above HVAC’s cooling set-
fort factor level outperforms its coarse-grained counterpart. point. As we can see, indoor temperature can even be 5 ◦ C
Regarding the mean MAE, the fine-grained N N, L R, S VR - L, lower or higher than the setpoint. Also, the air is often over-
and S VR - R are 21%, 20%, 21%, and 23% better than the cooled by the HVAC system, where indoor temperature is
coarse-grained C N N, C L R, C S VR - L, and C S VR - R, respec- more likely to be much lower than the cooling setpoint. While
tively. Such fact has been observed in Section IV using indoor temperature is often assumed to be equivalent to the
D NN and Fig. 7 shows that it also applies to other ML setpoint [23], [24], Fig. 8(a) shows that the assumption can be
algorithms. This again indicates that fine-grained architec- far from true in reality. To alleviate or even avoid such diver-
ture is good at exploring the detail information within the gence, this paper adopts D NN to model the comfort factors
modeled system, and such information helps model the rela- like indoor temperature, instead of making some unrealistic
tionship between the input attributes and thermal comfort more assumptions.
accurately. Fig. 8(b) shows the comfort score change in HVAC cooling
With the same fine-grained architecture, D NN performs setpoint for a summer workday morning, with 30 ◦ C out-
the best among the tested algorithms for thermal com- door temperature, 60% humidity, and 0.5 m/s air velocity.
fort modeling. The observation shows that D NN’s superior Two schemes are tested. The first one is S ETPOINT, which

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2021 at 15:55:36 UTC from IEEE Xplore. Restrictions apply.
2548 IEEE INTERNET OF THINGS JOURNAL, VOL. 6, NO. 2, APRIL 2019

follows the commonly adopted assumption that the setpoint, is 1.7×, 2.5×, 2.4×, and 1.9× better than N N, L R, S VR - L,
indoor temperature and MRT are equivalent. The second one and S VR - R, respectively, for thermal comfort modeling. In
is D NN, which models the complex system among the con- addition, thermal comfort score scales linearly in the HVAC
trollable HVAC setpoints, outdoor and the indoor condition. setpoint, and this allows to search for the optimal setpoint
Seen from the figure, S ETPOINT has a biased estimation of quickly and accurately with the desired comfort.
the indoor environment and tends to assume a warmer situa- In the future, we plan to study the impact of the other fac-
tion compared to D NN. The key reason is that S ETPOINT does tors like occupancy to thermal comfort. We also would like
not capture the divergence between the actual indoor temper- to incorporate the latest technologies like wearable devices
ature and the HVAC setpoint, despite the fact that the air is for thermal comfort modeling. Finally, we want to extend this
often over-cooled in the tested case. Different from S ETPOINT, paper to building energy modeling.
D NN learns the complex thermal dynamics and derives more
accurate thermal comfort model. Then, D NN tends to choose a R EFERENCES
relatively higher setpoint than S ETPOINT to avoid over-cooling
[1] W. H. Sadid, S. A. Abobakr, and G. Zhu, “Discrete-event systems-based
and maintain a comfortable indoor environment. power admission control of thermal appliances in smart buildings,” IEEE
Typically, there exists a tradeoff between energy and Trans. Smart Grid, vol. 8, no. 6, pp. 2665–2674, Nov. 2017.
comfort, where better comfort means more energy usage. [2] A. S. Ahmad et al., “A review on applications of ANN and SVM
for building electrical energy consumption forecasting,” Renew. Sustain.
Assuming a cooling mode, the energy consumption is nor- Energy Rev., vol. 33, pp. 102–109, May 2014.
mally a decreasing function in HVAC setpoint, but discomfort [3] Z. Wu, Q.-S. Jia, and X. Guan, “Optimal control of multiroom HVAC
is increasing. Here, the maximum acceptable comfort level can system: An event-based approach,” IEEE Trans. Control Syst. Technol.,
vol. 24, no. 2, pp. 662–669, Mar. 2016.
be specified, i.e., PMV score of 1.0 indicating a slightly warm [4] R. Maddalena et al., “Effects of ventilation rate per person and per
indoor environment, to minimize the conditioning energy con- floor area on perceived air quality, sick building syndrome symptoms,
sumption as much as possible. Accordingly, the corresponding and decision-making,” Indoor Air, vol. 25, no. 4, pp. 362–370, 2015.
[5] P. O. Fanger et al., Thermal Comfort: Analysis and Applications in
setpoint with the minimized energy usage and satisfied ther- Environmental Engineering. New York, NY, USA: McGraw-Hill, 1970.
mal comfort can be derived using the comfort model. Besides, [6] J. Lin et al., “A survey on Internet of Things: Architecture, enabling tech-
suppose an energy model accurate enough is given, it is also nologies, security and privacy, and applications,” IEEE Internet Things
J., vol. 4, no. 5, pp. 1125–1142, Oct. 2017.
possible to figure out the setpoint given certain energy quota [7] L. Liu, Z. Chang, X. Guo, S. Mao, and T. Ristaniemi, “Multiobjective
for air conditioning. optimization for computation offloading in fog computing,” IEEE
In addition, an approximately linear relationship between Internet Things J., vol. 5, no. 1, pp. 283–294, Feb. 2018.
[8] Y. Wang, S. Mao, and R. M. Nelms, “Distributed online algorithm for
the PMV-based comfort score and setpoint can be observed optimal real-time energy distribution in the smart grid,” IEEE Internet
in Fig. 8(b). Accordingly, given a desired comfort score, the Things J., vol. 1, no. 1, pp. 70–80, Feb. 2014.
optimal setpoint can be computed from testing any two differ- [9] N. Seydoux, K. Drira, N. Hernandez, and T. Monteil, “Autonomy
ent setpoints. For example, the setpoint x can be set to 29.2 ◦ C through knowledge: How IoT-o supports the management of a connected
apartment,” in Proc. Semantic Web Technol. Internet Things (SWIT)
for the PMV of up to 0.5, where the PMV is 0.43 for setpoint CEUR-WS, 2016, pp. 67–78.
29.0 ◦ C and 0.69 for 30.0 ◦ C and (0.5 − 0.43)/(x − 29.0) = [10] K. Zhang et al., “Security and privacy in smart city applications:
Challenges and solutions,” IEEE Commun. Mag., vol. 55, no. 1,
(0.69 − 0.5)/(30.0 − x). Our un-optimized Python code can pp. 122–129, Jan. 2017.
finish such computation instantaneously. [11] M. Chen, S. Mao, and Y. Liu, “Big data: A survey,” Mobile Netw. Appl.,
vol. 19, no. 2, pp. 171–209, 2014.
[12] G. Hinton et al., “Deep neural networks for acoustic modeling in speech
VII. C ONCLUSION recognition: The shared views of four research groups,” IEEE Signal
Process. Mag., vol. 29, no. 6, pp. 82–97, Nov. 2012.
In this paper, we revisit thermal comfort modeling using [13] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521,
deep learning for the IoT instrumented buildings. Fine-grained no. 7553, pp. 436–444, 2015.
[14] A. C. Megri, I. E. Naqa, and F. Haghighat, “A learning machine approach
modeling is proposed to train an exclusive model for each for predicting thermal comfort indices,” Int. J. Ventilation, vol. 3, no. 4,
comfort factor and accordingly the thermal comfort score can pp. 363–376, 2005.
be derived. We show that fine-grained modeling outperforms [15] S. Atthajariyakul and T. Leephakpreeda, “Neural computing thermal
comfort index for HVAC systems,” Energy Convers. Manag., vol. 46,
the coarse-grained counterpart significantly, which only pro- nos. 15–16, pp. 2553–2565, 2005.
duces one model to link the input attributes and the thermal [16] Thermal Environmental Conditions for Human Occupancy, ASHRAE
comfort directly. Such fact not only applies to D NN but also Standard 55-2010, 2010.
[17] T. Chaudhuri, D. Zhai, Y. C. Soh, H. Li, and L. Xie, “Random for-
the other ML algorithms. The impact of D NN topology is est based thermal comfort prediction from gender-specific physiological
investigated. The results show that the modeling performance parameters using wearable sensing technology,” Energy Build., vol. 166,
benefits from more hidden layers and more neurons. Also, a pp. 391–406, May 2018.
[18] E. L. Kruger and P. Drach, “Identifying potential effects from anthropo-
simple topology with the same number of neurons per hid- metric variables on outdoor thermal comfort,” Build. Environ., vol. 117,
den layer is sufficient to achieve higher modeling accuracy, pp. 230–237, May 2017.
compared to the more complex topologies. Thermal comfort [19] W. Hu, Y. Wen, K. Guan, G. Jin, and K. J. Tseng, “iTCM: Towards
learning-based thermal comfort modeling via pervasive sensing for smart
model is shown to be complex and nonlinear. Linear ML buildings,” IEEE Internet Things J., to be published. [Online]. Available:
algorithms such as L R and S VR - L fail to produce stronger https://ieeexplore.ieee.org/document/8423630
performance compared to the nonlinear ones such as D NN, [20] A. Javed et al., “Design and implementation of a cloud enabled ran-
dom neural network-based decentralized smart controller with intelligent
N N, and S VR - R. Among the tested algorithms, D NN achieves sensor nodes for HVAC,” IEEE Internet Things J., vol. 4, no. 2,
the highest accuracy with reasonable time usage. Overall, D NN pp. 393–403, Apr. 2017.

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2021 at 15:55:36 UTC from IEEE Xplore. Restrictions apply.
ZHANG et al.: THERMAL COMFORT MODELING FOR SMART BUILDINGS 2549

[21] L. Yu, D. Xie, T. Jiang, Y. Zou, and K. Wang, “Distributed real- Weizheng Hu received the bachelor’s degree in
time HVAC control for cost-efficient commercial buildings under smart computer engineering from Nanyang Technological
grid environment,” IEEE Internet Things J., vol. 5, no. 1, pp. 44–55, University, Singapore, where he is currently pursu-
Feb. 2018. ing the Ph.D. degree at the School of Computer
[22] J. Langevin, P. L. Gurian, and J. Wen, “Tracking the human- Science and Engineering.
building interaction: A longitudinal field study of occupant behavior His current research interests include wearable
in air-conditioned offices,” J. Environ. Psychol., vol. 42, pp. 94–115, devices, smart buildings, and machine learning.
Jun. 2015.
[23] T.-P. Lin, A. Matzarakis, and R.-L. Hwang, “Shading effect on long-term
outdoor thermal comfort,” Build. Environ., vol. 45, no. 1, pp. 213–221,
2010.
[24] R. P. Kramer, H. L. Schellen, and J. W. van Schijndel, “Towards
temperature limits for museums: A building simulation study
for four museum zones with different quality of envelopes,” in
Proc. Healthy Build. Europe (Eindhoven), 2015. [Online]. Available:
https://www.researchgate.net/profile/Rick_Kramer/publication/28071322 Yonggang Wen (S’99–M’08–SM’14) received the
7_Towards_temperature_limits_for_museums_a_building_simulation_ Ph.D. degree in electrical engineering and computer
study_for_four_museum_zones_with_different_quality_of_envelope/lin science (with a minor in western literature) from the
ks/55c1fb7708aed9dff2a62187.pdf Massachusetts Institute of Technology, Cambridge,
[25] R. D. Wooten, “Statistical analysis of the relationship between wind MA, USA, in 2008.
speed, pressure and temperature,” J. Appl. Sci., vol. 11, no. 15, He was with Cisco, San Jose, CA, USA, where
pp. 2712–2722, 2011. he leads product development in content delivery
network, which had a revenue impact of $3 bil-
lion globally. He is an Associate Professor with
the School of Computer Science and Engineering,
Nanyang Technological University, Singapore. He
has authored or co-authored over 140 papers in top journals and prestigious
Wei Zhang (S’10–M’16) received the B.Sc. degree conferences. His research in Multiscreen Cloud Social TV has been featured
in electronic information science and technology by global media (over 1600 news articles from over 29 countries). His cur-
from the School of Science, Xidian University, rent research interests include cloud computing, green data center, big data
Xi’an, China, in 2008, the M.Eng. degree in cir- analytics, multimedia network, and mobile computing.
cuits and system from the School of Electronic Dr. Wen was a recipient of the ASEAN ICT Award 2013 (Gold Medal).
Engineering, Xidian University in 2011, and the His research on Cloud3DView, as the only academia entry, won the Data
Ph.D. degree in computer science from the School Centre Dynamics Award (APAC) in 2015. He was a co-recipient of the
of Computer Science and Engineering, Nanyang 2015 IEEE Multimedia Best Paper Award and the Best Paper Award of
Technological University, Singapore, in 2015. EAI/ICST Chinacom 2015, IEEE WCSP 2014, IEEE Globecom 2013,
He is a Scientist with Experimental Power and IEEE EUC 2012. He serves on the Editorial Boards for the IEEE
Grid Centre, Agency for Science, Technology, and T RANSACTIONS ON C IRCUITS AND S YSTEMS FOR V IDEO T ECHNOLOGY,
Research, Singapore. From 2015 to 2017, he was a Research Fellow with IEEE Wireless Communication Magazine, IEEE C OMMUNICATIONS S URVEY
the School of Computer Science and Engineering, Nanyang Technological & T UTORIALS, the IEEE T RANSACTIONS ON M ULTIMEDIA, the IEEE
University. His current research interests include energy optimization for the T RANSACTIONS ON S IGNAL AND I NFORMATION P ROCESSING OVER
information and communication technology area, as well as the smart build- N ETWORKS, IEEE ACCESS, and Ad Hoc Networks (Elsevier), and was
ings and smart city, multimedia, computer architecture, cloud security, and elected as the Chair for IEEE ComSoc Multimedia Communication Technical
computational intelligence areas. Committee from 2014 to 2016.

Authorized licensed use limited to: Polytechnic University of Bucharest. Downloaded on October 19,2021 at 15:55:36 UTC from IEEE Xplore. Restrictions apply.