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

net/publication/316155534

Machine Learning based Prediction of Thermal Comfort in Buildings of

Equatorial Singapore

Conference Paper · July 2017

DOI: 10.1109/ICSGSC.2017.8038552

CITATIONS READS

90 3,201

4 authors, including:

Tanaya Chaudhuri Y.C. Soh

Agency for Science, Technology and Research (A*STAR) Nanyang Technological University
13 PUBLICATIONS 731 CITATIONS 546 PUBLICATIONS 15,578 CITATIONS

SEE PROFILE SEE PROFILE

Lihua Xie
Nanyang Technological University
1,171 PUBLICATIONS 48,332 CITATIONS

SEE PROFILE

All content following this page was uploaded by Tanaya Chaudhuri on 28 June 2018.

The user has requested enhancement of the downloaded file.

 2017 IEEE International Conference on Smart Grid and Smart Cities

Machine Learning Based Prediction of Thermal Comfort in Buildings of Equatorial

Singapore

Tanaya Chaudhuri1,2, Yeng Chai Soh1,2, Hua Li3, Lihua Xie2

1
Energy Research Institute @ NTU (ERIAN), Interdisciplinary Graduate School
2
School of Electrical and Electronic Engineering
3
School of Mechanical and Aerospace Engineering
Nanyang Technological University, Singapore

Abstract—Majority of energy consumption in Singapore between energy use and thermal comfort- a vital objective of
buildings is due to air-conditioning, because of its hot and smart city buildings.
humid weather. Besides attaining a healthy indoor The most widely adopted method to predict thermal
environment, a prior knowledge about the occupant’s thermal comfort is the Fanger’s PMV (Predicted Mean Vote) Model
comfort can be beneficial in reducing energy consumption, as it [2]. According to this model, a person’s thermal comfort
can save energy which is otherwise spent in extra cooling. This depends upon six factors – air temperature, mean radiant
paper proposes a data-driven approach to predict individual temperature, relative humidity, air speed, clothing rate
thermal comfort level (‘cool-discomfort’, ‘comfort’, ‘warm- (clothing insulation of the person), and metabolic rate
discomfort’) using environmental and human factors as input.
(activity level of the person). The PMV model uses these six
Six types of classifiers have been implemented- Support Vector
Machine (SVM), Artificial Neural Network (ANN), Logistic
factors as inputs to predict actual thermal comfort by an
Regression (LR), Linear Discriminant Analysis (LDA), K- index termed PMV. The PMV index ranges on a scale of -3
Nearest Neighbors (KNN), and Classification Trees (CT), on a (Cold) to +3 (Hot), according to the ASHRAE thermal
publicly available database of 817 occupants for air- comfort scale (Fig. 1), where neutral is the desirable level [3].
conditioned and free-running buildings separately. Results However, previous works have revealed discrepancies
show that our approach achieves prediction accuracies of between the predicted vote PMV and the actual thermal
73.14-81.2%, outperforming the traditional Fanger’s PMV sensation vote reported by the occupants [4, 5]. This could
(Predicted Mean Vote) model, which has accuracies of only probably be because certain other factors such as age, gender
41.68-65.5%. Age, gender, and outdoor effective temperature, and the outdoor climate conditions are not included in the
which are not included in the PMV model, are found to be PMV model.
important factors for thermal comfort. The proposed approach Metabolic ability decreases with age and so it changes
also outperforms modified PMV models- the extended PMV the person’s sensitivity to heat or cold [6]. In case of gender,
model and the adaptive PMV model which attain accuracies of women have lower metabolic rate and hence lower
61.75% and 35.51% respectively. evaporative heat loss as compared to men. It is found that
women usually prefer higher air temperature [7]. Several
Keywords— thermal comfort; smart building; PMV model; works have studied the impact of age and gender, but
adaptive model; machine learning
resulting in different conclusions [2, 6-8]. Outdoor weather
I. INTRODUCTION can have psychological effect on thermal comfort in air-
conditioned (HVAC) buildings, and direct effect on thermal
Singapore is a hot and humid country due to its close comfort in naturally ventilated (NV) buildings. This paper
proximity to the equator at 1.3521°N, and is characterized investigates the importance of these three factors (age,
by uniformly high temperatures and humidity throughout the gender and outdoor weather) on thermal comfort in building
year. About 50% of the electricity consumption in Singapore indoors.
is due to buildings, and about 60% of the electricity In this paper, we have proposed a data-driven approach to
consumption in buildings is due to cooling [1]. Hence predict individual thermal comfort in real time using several
substantial amount of energy and capital resources is spent environmental and human factors including the six Fanger’s
every year in cooling to achieve a thermally comfortable factors and the three new proposed factors- age, gender and
environment in buildings. Maintaining a healthy indoor outdoor weather. Machine learning algorithms [9-13] have
climate is crucial to prevent health hazards to the occupants been used for this approach because they can deal with big
such as thermal stress and hypothermia [2]. A healthy indoor data, and yet have high computational speed as compared to
environment also improves productivity at work. A prior PMV model, which has a long and complex iterative
knowledge of the occupant’s thermal comfort level can help computation process. HVAC and NV buildings are studied
to control how much energy is to be spent in cooling, thus separately. The proposed model has been compared with
limiting the energy consumption as per requirement. Fanger’s PMV model, as well as modified PMV models
Prediction of thermal comfort can therefore, not only (ePMV, aPMV) that are based on the adaptive model [14,
maintain a favorable indoor climate but also aid in reducing 15].
energy consumption. It can bring about an optimal balance

978-1-5386-0504-2/17/$31.00 ©2017 IEEE 72

 comfort is perceived [15]. The modified PMV models,
namely the ePMV and aPMV models take into account these
extraneous factors while predicting the comfort level [14, 15,
19]. These modified models have been discussed in later
Figure 1. ASHRAE Thermal Comfort scale
sections.
Several work [2, 13, 16] have been done to predict
thermal comfort, but this study differs from them in the
following ways: this is a data-driven approach unlike most of
them; it is specific to equatorial climate and buildings of
Singapore (different climatic regions have different building
envelopes and properties suited to their climates); air-
conditioned and naturally ventilated buildings are studied
separately; three new features, namely age, gender, and
outdoor weather have been introduced, and lastly, outdoor
effective temperature has been used for outdoor weather data,
because unlike air temperature, effective temperature is
representative of the overall weather.
Figure 2. Overview of proposed approach
II. PROPOSED PREDICTION MODEL
We propose a data-driven approach to real time The rest of the paper is as follows: section III describes
prediction of individual thermal comfort level. The thermal the methodology used: the dataset, its processing and the six
comfort reported by occupants on the ASHRAE scale (Fig. 1) machine learning algorithms studied, section IV presents the
is first classified into three levels: Cool discomfort [-3,-1), results and discussion for HVAC and NV buildings and
Comfortable/neutral [-1, 1], Warm discomfort (1, 3]. This the importance of proposed features- age, gender, and
level is taken as ground truth or actual thermal comfort level. outdoor weather; lastly section V concludes the findings.
Our approach is based on supervised machine learning III. METHODOLOGY
algorithm wherein a classifier learns the pattern between the
inputs and the outputs, and gains the ability to predict the A. Data Preprocessing
output for unseen future input. The overview of the approach ASHRAE is one of the main organizations that set
is presented in Fig. 2. The first step is the selection of the thermal comfort standards and guidelines. For this study, we
best classifier. Each classifier takes in several environmental have used the database of thermal comfort experiments
and human factors as input, and the actual comfort level; performed in Singapore by de Dear et al., which is publicly
trains itself, and its accuracy is tested on new/test dataset. available as part of the ASHRAE RP-884 project [18]. A
The classifier with the minimum classification error is total of 818 people were surveyed about their thermal
selected as the best. The selected classifier with its optimal sensation, across several buildings in Singapore. There were
tuning parameters forms the prediction model, which can be two types of buildings used in the survey: air-conditioned
implemented in real-time. The predicted comfort level can be (HVAC) and naturally ventilated (NV) buildings. A total of
sent as a signal to the HVAC controller in air conditioned 235 occupants were surveyed across 12 different air-
buildings, or in case of naturally ventilated buildings, it can conditioned office buildings. These were all high-rise
be displayed, and used by the occupant to manually control buildings with open plan design, with centrally controlled
the cooling device (fans etc.). The inputs are called as HVAC system. For naturally ventilated buildings, a total of
features and a set of inputs is called a feature set. In this case, 583 residents were surveyed across 4 public residential
feature selection is not done as we consider all the mentioned buildings. These buildings were high-rise as well, and the
inputs as important. Unlike several indoor thermal studies, mode of thermal control consisted mainly of manually
we have not discarded mean radiant temperature as studies operating ceiling fans and standing fans. For each occupant,
show that it is in fact one of the most important factors, following measurements were taken during the experiments
especially in a wet equatorial region as Singapore [17] where among others: indoor and outdoor environmental parameters
solar radiation is high. (air temperature, air velocity, mean radiant temperature,
There are two types of thermal comfort models- static relative humidity), personal parameters (gender, age,
and adaptive. Fanger’s PMV model is a static model, metabolic rate, clothing rate), and thermal comfort
applicable to both HVAC and NV case, while the adaptive questionnaire.
model is more suited for NV buildings. In this paper, HVAC For our study, we are interested in the data of Fanger’s
and NV buildings have been studied separately because they six parameters, thermal sensation vote, gender, age and
have different thermal dynamics, not only in terms of control, outdoor weather. Hence, the datasets that had any of these
but also the influencing factors. According to the adaptive missing values were excluded. This amounted to 229
model, in NV buildings, several other factors such as the datasets and 583 datasets for HVAC and NV buildings
occupant’s adaptation abilities, psychological state, thermal respectively.
preferences and control also play a role in the way thermal

73
B. Feature Set and Output Network (ANN), the datasets were divided into three sets:
The following factors were included in the feature set: training, validation, and testing in the ratio 70%: 15%: 15%.
 Fanger’s parameters: The six Fanger’s parameters Thus the numbers of samples in each case were: training
are well established as definite influencers of thermal (161), validation (34) and testing (34) for HVAC case; and
comfort: air temperature (Ta), mean radiant training (409), validation (87), and testing (87) for NV case.
temperature (MRT), relative humidity (RH), air We have tested six different machine learning classifiers
speed (Va), clothing rate (Clo), and metabolic rate on the datasets of each type of building. This is repeated for
(M). each of the two feature sets per building type, HVAC:
(fHVAC_wo, fHVAC_w), and NV: (fNV_wo, fNV_w). The different
 Outdoor effective temperature (ETout): The effective
classifiers studied are discussed briefly as follows:
temperature indicates the combined effect of air
temperature, relative humidity and air velocity [18]. 1) Support Vector Machine (SVM): SVM is an
Thus it better represents overall outdoor weather, as extension of support vector classifier which is based on the
compared to using only outdoor air temperature. concept of a hyperplane that separates different classes of
 Age (A) training observations with a soft margin, using kernels [9,
 Gender (G) 10, 12]. In SVM, cost (C) is a non-negative tuning
To study the importance of the three new factors ETout, A parameter that is used to control the bias-variance tradeoff.
and G on thermal comfort, we created two different feature In this study, we have experimented with different types of
sets, one including and one excluding these three factors. For kernels, and Radial basis Function (RBF) worked best
NV case, the feature sets are: among them. We found the best tuning parameters of cost
fNV_wo = {Ta, MRT, Va, RH, M, Clo}, and and gamma () using 10-fold cross-validation, the values of
which are presented in Table I. We have used a C-type
fNV_w = {Ta, MRT, Va, RH, M, Clo, A, G, ETout}. classification (C-SVC) with one-versus-one approach.
However, information about occupant’s age were not 2) Artificial Neural Networks (ANN): ANN is based on
available for the HVAC buildings, hence the feature set for the human biological neuron networks, wherein signals are
HVAC case are: passed between a defined numbers of neurons over
fHVAC_wo = {Ta, MRT, Va, RH, M, Clo}, and weighted connection links, with each neuron using an
activation function to determine the output signal [10, 11].
fHVAC_w = {Ta, MRT, Va, RH, M, Clo, G, ETout}. In our study, we have used a two-layer feed forward ANN,
The ratio of gender proportion in HVAC case was: 61% with sigmoid function as activation function for hidden and
(male), 39% (female); and for NV case it was 49.2% (male) output neurons. Scaled conjugate gradient backpropagation
and 50.8% (female). The ages were divided into four groups has been used to train the network. The optimal number of
in the NV case. The proportion of ages were: 15% (17-20 neurons (N) is mentioned in Table I.
years old), 41.3% (21-40 years old), 28.3% (41-60 years old), 3) Logistic Regression (LR): Logistic Regression uses
and 15.4% (>60 years old). maximum likelihood method to fit the data according to a
The comfort experiments required the occupants to fill a
logistic function model. Once the coefficients are estimated,
questionnaire/survey wherein each occupant was supposed to
express their thermal sensation on the ASHRAE thermal it can calculate the probabilities of the observation
scale (Fig. 1). This reported sensation is called as the Actual belonging to each class. LR then predicts the class
Mean Vote (AMV) [2, 3].The AMVs were classified into 3 according to the highest probability [9].
comfort levels: Cool discomfort [-3,-1), Comfortable/neutral 4) Linear Discriminant Analysis (LDA): LDA is similar
[-1, 1], Warm discomfort (1, 3]. This thermal comfort level is to Logistic Regression in the aspect that it also produces
considered the actual or ground truth for the classifiers. linear decision boundaries, and takes a probabilistic
C. Machine Learning Algorithms approach to classification. However, unlike LR, LDA
estimates the coefficients using estimated mean and
This study is a case of supervised learning, as inputs and
variance from a normal distribution, using the least squares
outputs are both provided. Our aim is to implement several
classification algorithms and select the method with the best method [9].
classification accuracy. To avoid scaling issue, all the feature 5) K- Nearest Neighbours (KNN): The KNN classifier is
data were standardized by normalizing them to a mean of one of the simplest yet powerful classifiers [9, 10]. It
zero and standard deviation of one, prior to learning. Each requires a user-input constant K. It first identifies K number
sample consisted of a feature set and its corresponding of points in the training data that are closest to the test
thermal comfort level. The datasets were then divided into observation, estimates the conditional probability of each
training and testing sets in the ratio 70%:30%. Thus there class as a fraction of points among the K points belonging to
were 160 samples for training and 69 samples for testing in that class. KNN then classifies the test observation to the
HVAC case (n=229); and 408 samples for training and 175 class with the highest probability using Bayes rule. K=1
samples for testing in the NV case (n=583); where, n is the
gives the optimal result in this study.
total number of samples. However, for Artificial Neural

74
 6) Classification Trees (CT): CT classifier is a decision model reaches an accuracy of 41.68%. PMV model gives
tree that is grown using recursive binary splitting [9, 10]. better results in air-conditioned buildings, which is
The leaves are the class labels and the branches are the consistent with previous studies that the PMV model is
feature subsets leading to the class label. CTs are built using better suited to HVAC case compared to NV case [14, 15].
either classification error rate, Gini index or cross entropy as It is observed that all the classifiers studied outperforms the
the criterion for making the splits. This classifier is popular Fanger’s method at predicting the actual comfort level, with
for its strong interpretation abilities. SVM performing the best at 71.43%, and both SVM and LR
performing the best at 73.14% for each feature set
TABLE I. OPTIMAL TUNING PARAMETERS OF LEARNING respectively. It is evident that inclusion of the proposed
ALGORITHMS features improves the prediction accuracies. This implies
HVAC NV that gender, age and outdoor weather are important factors
fHVAC_wo fHVAC_w fNV_wo fNV_w of thermal comfort in NV buildings.
C=1 C=1 C=1 C=1
SVM TABLE II. PREDICTION ACCURACY (%)– HVAC BUILDINGS
=0.1 =0.1 =0.5 =0.1
ANN fHVAC_wo fHVAC_w
N = 20 N = 20 N = 20 N = 20
SVM 79.70 81.20
KNN K=1 K=1 K=1 K=1
ANN 79.40 85.30
IV. RESULTS AND DISCUSSION LR 76.81 79.71
LDA 73.91 76.81
A. Performance Metrics
KNN 72.46 75.36
We have used Classification accuracy as the criterion for
comparing performance of the algorithms. The classification CT 72.46 75.36
accuracy is basically the percentage of the test samples Fanger’s PMV 65.50
whose thermal comfort level have been correctly predicted
by the learning algorithm. Confusion matrix is a table that TABLE III. PREDICTION ACCURACY (%)– NV BUILDINGS
provides a visualization of a supervised algorithm’s
performance. Each row consists of the true labels of the fNV_wo fNV_w
comfort level, and columns represent the predicted comfort SVM 71.43 73.14
levels. Classification accuracy is the ratio of the sum of the
ANN 70.10 72.40
diagonals in the confusion matrix to the total number of test
samples. LR 70.28 73.14

B. Air Conditioned Buildings (HVAC) LDA 61.14 64.00

Table II presents the results of the predictive KNN 56.00 61.71

performance of the six machine learning classifiers on the CT 62.86 64.57
test dataset of air conditioned buildings. The first and
Fanger’s PMV 41.68
second column presents the classification accuracies for
feature sets fHVAC_wo and fHVAC_w respectively. The Fanger’s
PMV model reaches an accuracy of 65.5%. It is observed The confusion matrices for SVM and LR are presented
that all the classifiers studied outperforms the Fanger’s in Table IV. To decide which method performs better, we
method at predicting the actual comfort level, with SVM have calculated Positive Predictive Value (PPV) or
performing the best at 79.7% and ANN performing the best Precision for each of the comfort levels. PPV is the ratio of
at 85.3% for each feature set respectively. Also, it is the true positives to the sum of true positives and false
observed that the feature set that includes gender and positives. A PPV of 0.8 implies that out of the times a class
outdoor effective temperature along with the six Fanger’s was predicted, 80% of the time the prediction was correct.
parameters results in improved accuracies for all the Both SVM and LR perform equally well on overall accuracy.
methods. This implies that gender and outdoor weather But PPV results (Table V) shows that SVM performs better
environment are important factors for thermal comfort in at each comfort level, with almost same and balanced
HVAC buildings. precision for each level unlike LR which is biased towards
comfort level.
C. Naturally Ventilated Buildings (NV) Naturally ventilated buildings may/may not have open
Table III presents the results of the predictive windows, which allow wind and solar radiation to enter and
performance of the six machine learning classifiers on the influence the indoor climate. Thus it is intuitive to believe
test dataset of naturally ventilated buildings. The first and that in such buildings, outdoor weather have influence on
second column presents the classification accuracies for indoor climate. However, adaptive model says that the
feature sets fNV_wo and fNV_w respectively. The Fanger’s PMV occupant’s control over the environment such as ability to

75
close/open the windows, operate and regulate fans, also TABLE VII. ACCURACY (%) OF PMV MODEL, MODIFIED MODELS, AND
PROPOSED MODEL
have a psychological influence on the thermal sensation in
naturally ventilated buildings. Besides thermal control, Fanger’s PMV 41.68
thermal preferences, expectations, and adaptation abilities
ePMV Model 61.75
(physiological, behavioral, psychological), can also
influence thermal sensation. These factors are not taken into aPMV Model 35.51
consideration by the Fanger’s PMV model, which may Proposed Approach 73.14
explain its low prediction accuracy. In order to include these
extraneous factors, modified PMV models were developed V. CONCLUSION
specifically for naturally ventilated buildings.
This paper presents a data-driven approach for real-time
TABLE IV. CONFUSION MATRIX FOR SVM AND LR FOR NV CASE prediction of thermal comfort at individual level. The model
takes in several environmental and human factors as input,
SVM LR
and predicts the occupant’s thermal comfort level (‘cool-
Predicted Comfort Level
discomfort’, ‘comfort’, ‘warm-discomfort’), which can then
-1 0 +1 -1 0 +1 be fed as signal to HVAC controller or used for manual
-1 8 15 0 8 2 0 control. Six different state of the art machine learning
True classifiers have been studied. Results show that our
Comfort 0 2 97 7 17 87 18
Level approach achieves prediction accuracies of 73.14-81.2%,
+1 0 23 23 0 10 33 outperforming the traditional Fanger’s PMV model, which
has accuracies of 41.68-65.5%. Age, gender, and outdoor
TABLE V. POSITIVE PREDICTIVE VALUE (PPV) / PRECISION effective temperature, which are not included in the PMV
model, are found to be important factors for thermal comfort.
Comfort Level SVM LR The proposed approach also outperforms modified PMV
Cool discomfort (-1) 0.8 0.3
models- the extended PMV model and the adaptive PMV
model which attain accuracies of 61.75% and 35.51%
Comfortable (0) 0.7 0.9 respectively. The proposed approach can help to achieve an
Warm discomfort (1) 0.8 0.6 optimal balance between thermal comfort and energy
consumption, which is a vital objective of smart building
Fanger developed an extended PMV model which takes and smart cities. In future, we aim to study the scalability of
into account these extraneous factors, and predicts the actual
this approach to other climatic regions, probably a cold
thermal sensation as in (1), where ‘e’ is called as the
expectancy factor [14, 19]. climate, where a good deal of energy consumption is due to
heating in buildings.
ePMVePMV (1) ACKNOWLEDGMENT
 This work is supported by Singapore’s National Research
Another modified PMV model is the adaptive PMV
Foundation under NRF2011NRFCRP001-090, and partially
model developed by Yao et al. [15, 19]. It uses an adaptive
supported by the Energy Research Institute at NTU
coefficient ‘’ to predict the actual thermal sensation as in (ERI@N).
(2).
REFERENCES
aPMV PMV PMV) (2) [1] The National Research Foundation (NRF) The National Climate
Change Secretariat (NCCS), The Building, and Construction
For this study, we have calculated ‘e’ and ‘’ according Authority (BCA). Building Energy Efficiency R&D Roadmap.
Technical report, 2015.
to the method mentioned in [19], and their values are given
in Table VI. A comparison of our method with the existing [2] Fanger P. O. Thermal comfort, analysis and application in
environmental engineering. Danish Technical Press, Copenhagen,
methods reveals that our method outperforms them at an 1970.
accuracy of 73.14% (Table VII). It is also observed that the [3] ASHRAE. Thermal environmental conditions for human occupancy.
extended PMV model at 61.75% performs almost two times ASHRAE Standard 55, 2013.
better than the adaptive PMV model at 35.51%. [4] Van Hoof J., “Forty years of Fanger’s model of thermal comfort:
comfort for all?” Indoor Air. 18, 3, 182-201, June 2008.
TABLE VI. CALCULATED COEFFICIENTS OF MODIFIED PMV MODELS [5] R. Becker, M. Paciuk, “Thermal comfort in residential buildings –
Failure to predict by Standard model,” Build. Environ., vol. 44, Issue
ePMV e = 0.6 5, May 2009, pp. 948-960, doi: 10.1016/j.buildenv.2008.06.011.
aPMV = -0.3217 [6] Indraganti M., Rao K., “Effect of age, gender, economic group and
tenure on thermal comfort: A field study in residential buildings in
hot and dry climate with seasonal variations”, Energ. Buildings, vol.

76
 42, Issue 3, Mar. 2010, pp. 273-281, doi: [14] Fanger P. O., Toftum J., “Extension of the PMV model to non-air-
10.1016/j.enbuild.2009.09.003. conditioned buildings in warm climates”, Energ. Buildings, vol. 34,
[7] Karjalainen, S., “Thermal comfort and gender: a literature review”, Issue 6, July 2002, pp. 533-536, doi: 10.1016/S0378-7788(02)00003-
Indoor Air, 22:, 2012, 96–109, doi:10.1111/j.1600- 8.
0668.2011.00747.x. [15] R. Yao, B. Li, J. Liu, “A theoretical adaptive model of thermal
[8] Rupp R., Vásquez N., Lamberts R., “A review of human thermal comfort – Adaptive Predicted Mean Vote (aPMV)”, Build. Environ.,
comfort in the built environment”, Energ. Buildings, vol. 105, Oct. vol. 44, Oct. 2009, pp. 2089-2096, doi:
2015, pp. 178-205, doi: 10.1016/j.enbuild.2015.07.047. 10.1016/j.buildenv.2009.02.014.
[9] Hastie, T., Tibshirani, R., Friedman, J. The elements of statistical [16] X. Chen, Q.Wang, J. Srebric, “Model predictive control for indoor
learning: Data mining, inference, and prediction, NY: Springer, 2009. thermal comfort and energy optimization using occupant feedback”,
Energ. Buildings, vol. 102, Sept. 2015, pp. 357-369, doi:
[10] Kotsiantis, S. B., Zaharakis, I. D., Pintelas, P. E., “Machine learning: 10.1016/j.enbuild.2015.06.002.
A review of classification and combining techniques”, Artif. Intell.
Rev., 26(3), 2006, 159-190, doi: 10.1007/s10462-007-9052-3. [17] T. Chaudhuri, Y. C. Soh, S. Bose, L. Xie and H. Li, "On assuming
Mean Radiant Temperature equal to air temperature during PMV-
[11] Kevin Gurney. An Introduction to Neural Networks. Taylor & Francis,
based thermal comfort study in air-conditioned buildings,", 42nd
Inc., Bristol, PA, USA, 1997.
IEEE Annual Conf. on Industrial Electronics-IECON, Florence, Italy,
[12] Nello Cristianini and John Shawe-Taylor. An Introduction to Support Oct. 2016, doi: 10.1109/IECON.2016.7793073.
Vector Machines: And Other Kernel-Based Learning Methods. [18] R. De Dear and G. S. Brager. Developing an adaptive model of
Cambridge University Press, New York, 1999.
thermal comfort and preference. Center for the Built Environment,
[13] Changzhi Dai, Hui Zhang, Edward Arens, Zhiwei Lian, “Machine 1998.
learning approaches to predict thermal demands using skin [19] J. Gao, Y. Wang, P. Wargocki, “Comparative analysis of modified
temperatures: Steady-state conditions”, Build. Environ., vol. 114, Mar. PMV models and SET models to predict human thermal sensation in
2017, pp. 1-10, ISSN 0360-1323, doi: naturally ventilated buildings”, Build. Environ., vol. 92, Oct. 2015, pp.
10.1016/j.buildenv.2016.12.005. 200-208, doi: 10.1016/j.buildenv.2015.04.030.

77

View publication stats