University of Southern California Dissertations and Theses

Exploration for the prediction of thermal comfort & sensation with application of building HVAC automation

(USC Thesis Other)

Exploration for the prediction of thermal comfort & sensation with application of building HVAC automation

PDF

Download

Open document

Flip pages

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content i

Exploration for The Prediction of Thermal Comfort & Sensation with Application of
Building HVAC Automation

By

Hanxun Liu

Presented to the
FACULTY OF THE
SCHOOL OF ARCHITECTURE
UNIVERSITY OF SOUTHERN CALIFORNIA
In partial fulfillment of the
Requirements of degree
MASTER OF BUILDING SCIENCE

MAY 2019

ii

ACKNOWLEDGEMENTS

My deepest gratitude goes to National Science Foundation for its financial support.

I would also like to offer my sincere thanks to the following people: Dr. Shrikanth Narayanan,
who offered professional guidance and patient supervision; Gideon Susman, who provided the
experimental environment for our validation test; and my friend Han Gao, who provided technical
support for our experiment and control design.

iii
COMMITTEE MEMBERS

Joon-Ho Choi
Assistant Professor
University of Southern California
joonhoch@usc.edu

Marc Schiler
Professor
University of Southern California
marcs@usc.edu

Yolanda Gil
Research Professor
University of Southern California
gil@isi.edu

iv
ABSTRACT
Thermal comfort optimization is crucial for Heating, Ventilation and Air Conditioning (HVAC)
automation, which adjusts thermal conditions automatically, based on occupants’ real-time
thermal preferences. However, most standard thermal comfort models, such as Predicted Mean
Vote (PMV), have not considered individual thermal preferences. In many cases, it is reported
that the prediction of PMV is not accurate enough, especially for individual use, where thermal
preference is relatively subjective for everyone. Previous research focuses on how some
physiological indices indicate people’s thermal comfort states, while psychological factors such
as heart rate variability, which also affects their preferences, are practically ignored because of
the difficulty in quantifying occupants’ psychological activities.

The algorithms use human indexes to predict an individual’s thermal comfort, and estimate their
thermal sensation state based on environment parameters (indoor temperature). The estimation
results are regarded as signals that trigger a work-mode change in HVAC devices to optimize the
indoor thermal environment to comply with the preferences of a building’s occupants.

A series of human subject experiments were performed in an environmental chamber at USC to
collect different types of data, such as human body indexes, including skin temperature (°F),
electrode activity (μS), and the rate of low-frequency to high-frequency (LF/HF) for heart rate
variability. The last two represent the occupants’ reflections of outside stimuli and their stress
levels, respectively. Environmental parameters, including indoor temperature (°F), relative
humidity, radiant temperature (°F), and CO2 concentration (ppm), were also defined. The study
used machine learning to construct an algorithm set, and used it to build an HVAC auto-control
process based on real-time sensor data.

The research outcome shows the potential correlations between the human indexes/environment
factors that we chose, and the thermal comfort/thermal sensation levels, as well as the control
design. It presents applications for an HVAC auto-control based on designed thermal condition
estimation and incorporating real-time data from wearable sensors.

KEY WORDS: Thermal preference; Machine learning; Automation; Real time, HVAC control

HYPOTHESIS
Machine learning can help identify skin temperatures as signals to estimate thermal comfort
level.
Real-time human index and environment data can be used to achieve HVAC system automation.

RESEARCH OBJECTIVES
Establish the relationship between skin temperature and thermal comfort.
Investigate the correlations between electrode activity and thermal comfort.
Better understand psychological factors to eliminate subjective factors.
Develop an accurate thermal comfort prediction algorithm based on machine learning for data
analysis.
Achieve device control automation.

v

TABLE OF CONTENTS
ACKNOWLEDGEMENTS ............................................................................................................................... II
COMMITTEE MEMBERS .............................................................................................................................. III
ABSTRACT ...................................................................................................................................................... IV
1 INTRODUCTION ............................................................................................................................................ 1
1.1 Problems in Current Thermal Model Application ............................................................................................... 1
1.2 Skin Temperatures and Thermal Comfort ............................................................................................................ 1
1.3 Electrode Activity ................................................................................................................................................ 2
1.4 Heart Rate and Heart Rate Variability ................................................................................................................. 4
1.5 Computer Control System ................................................................................................................................... 4
2 BACKGROUND AND LITERATURE REVIEW ............................................................................................ 8
2.1 Current Thermal Estimation and Control Strategy .............................................................................................. 8
2.2 Skin Temperature and Thermal Comfort ............................................................................................................. 8
2.3 Thermal Comfort and Psychological Activity ................................................................................................... 10
2.4 Relationship Between Stimulation and Thermal Comfort ................................................................................. 11
2.5 Relationship Between HRV and Thermal Comfort ........................................................................................... 12
2.6 Summary ............................................................................................................................................................ 13
3. METHODOLOGY ........................................................................................................................................ 14
3.1 Methodology diagram ........................................................................................................................................ 14
3.2 Data Collection .................................................................................................................................................. 18
3.2.1 Data Type .................................................................................................................................................. 18
3.2.2 Collection Method ..................................................................................................................................... 18
3.3 Algorithm Design (From Methodology Diagram) ............................................................................................. 19
3.4 Real Time Control (From Methodology Diagram) ............................................................................................ 19
3.5 Software and Hardware ..................................................................................................................................... 19
3.5.1 Heart Rate Variability Generator .............................................................................................................. 20
3.5.2 BIOPAC EDA Devices .................................................................................................................................. 21
3.5.3 LabVIEW ................................................................................................................................................... 22
3.5.4 Matlab and LabVIEW Machine Learning Toolkit ....................................................................................... 23
3.5.5 Power Relay Hardware ............................................................................................................................. 24
3.5.6 General Sensors ........................................................................................................................................ 24
3.6 Data Analysis and Algorithm Definition ........................................................................................................... 26
3.6.1 Artificial Neural Network for Machine Learning ....................................................................................... 27
3.6.2 Logistic Regression for Feature Generation .............................................................................................. 28
3.7 Summary ............................................................................................................................................................ 29
4. DATA COLLECTION AND PROCESSING ................................................................................................. 30
4.1 Experiment for Training Data Collection .......................................................................................................... 30
4.2 Logistic Regression Process for Algorithm Design ........................................................................................... 31
4.2.1 Logistic Regression for Thermal Comfort Feature Selection ...................................................................... 31
4.2.2 Logistic Regression for Thermal Sensation Feature Selection ................................................................... 37
4.2.3 Logistic Regression for ANN Feature Determination ................................................................................. 40
4.3 Conclusion ......................................................................................................................................................... 45
5 THE APPLICATION AND V ALIDATION BY ARTIFICIAL NEURAL NETWORK .................................. 48
5.1 Validation for Real Office Application .............................................................................................................. 48
5.2 Validation for HV AC Control Automation ........................................................................................................ 50
5.3 Validation Results Analysis ............................................................................................................................... 52
5.3.1 Results of Real Office Application ............................................................................................................. 52

vi
5.3.2 Results of Real-Time Control ..................................................................................................................... 60
5.4 Optimization for ANN Application ................................................................................................................... 64
5.4.1 Renewing Training Data for ANN Simulation ............................................................................................ 64
5.4.2 New Feature Selection for ANN ................................................................................................................ 66
5.5 Conclusion ......................................................................................................................................................... 76
6 DISCUSSION, FUTURE WORK AND CONCLUSION ............................................................................... 78
6.1 Discussion .......................................................................................................................................................... 78
6.1.1 Feature Determination for Thermal Estimation ........................................................................................ 78
6.1.2 Validation and Real Application for ANN Model ....................................................................................... 78
6.1.3 Limitation .................................................................................................................................................. 79
6.2 Future Work ....................................................................................................................................................... 80
6.3 Conclusion ......................................................................................................................................................... 80
REFERENCES ................................................................................................................................................. 82
1

1 INTRODUCTION
1.1 Problems in Current Thermal Model Application
With the rising attention to occupant comfort in architecture design, more and more researchers
have focused on thermal comfort in order to complete the thermal comfort definition with higher
accuracy (Taleghani, Tenpierik, Kurvers, & Van Den Dobbelsteen, 2013). Based on ASHRAE-
55, thermal comfort can now represent the level of thermal sensation, which is related to the heat
balance between internal and external body environments. Predicted Mean Vote (PMV), which
describes the rule for human thermal equilibrium, is currently a widely-accepted model for
defining the thermal comfort level of building occupants. This model was developed using the
principle of heat balance and experimental data for indoor occupants.

However, there are limitations to the application of this model. First, PMV relies on a probability
distribution of the temperature preferences of the occupants of an enclosed space, and to some
degree, it ignores the thermal comfort of individuals (Ghahramani, Tang, & Becerik-Gerber,
2015). Second, experiments with it have been conducted in stable indoor environments, and it
may be not suitable for dynamic ones (Maiti, 2014).

Furthermore, PMV requires that a uniform thermal environment be applied, and ignores the
dissimilarities in thermal conditions for different body parts (d’Ambrosio Alfano, Ianniello, &
Palella, 2013). So, when considering local exposure, which means some body parts are exposed
to different thermal environments, PMV is not able to predict occupants’ thermal sensation
appropriately, while research shows specific body parts can affect the thermal comfort of
individuals.

In addition, although there is some research revealing the relationship between thermal comfort
and human body indices, such as skin temperature (Chaudhuri, Zhai, Soh, Li, & Xie, 2018), there
are still large gaps in the research quantifying people’s thermal sensations.

What is more, the research about psychological influences on occupants’ experiences of thermal
conditions is very limited, which increases the difficulty of making make thermal estimations
properly. In order to explore the understanding of thermal conditions, we need to consider
occupants’ psychological activities, and apply related factors such as EDA and heart rate
variability to estimate occupants’ current mental influences, such as stress level, to make the
results of research into thermal sensation more objective. Given the rapid development of
intelligent wearable devices, we now have instant access to more kinds of human data (heart rate,
respiration activity, etc.), and there is huge potential to make more accurate estimations or
prediction of human sensations in real time.

Under this circumstance, more research about individual thermal comfort is needed to describe
occupants’ thermal sensations more precisely, verify the related data measured in the lab, and
apply it in practice.

1.2 Skin Temperatures and Thermal Comfort
Current studies have shown that specific skin temperatures can be regarded as signals of human
thermal sensation (Liu, Wang, Liu, & Di, 2013). As the largest organ of the human body, skin

2
plays a significant role in energy exchange, and its temperature changes within limits to keep the
heat balance between external and internal body environments. As a result, the energy flow
direction is a critical factor affecting skin temperature, which determines our thermal sensation of
hot and cold. Skin temperature can potentially be used as a signal in thermal comfort estimation
(Sakoi et al., 2007).

Dr. Choi’s previous study was based on the measurement of skin temperature for 10 specific
body points in one lab. Consideration was given to gender, age and body mass index (BMI).

As we can see from the methodology below, Dr. Choi created a gradual change in the
temperature in the indoor lab environment, and the experimental subject was in the center of the
laboratory. The skin temperature measurement was recorded each time while the indoor
temperature was stable, and the occupant was asked to finish a thermal sensation questionnaire at
the same time. Dr. Choi wanted to explore whether skin temperature could be used to show
thermal comfort level, and then find out which local part had the skin temperature that best
represented the overall human sensation.

The result shows an obvious relationship between the change of specific skin temperature and
thermal comfort, as well as a tight correlation between local body parts and overall thermal
sensation. Therefore, there should be potential for skin temperature to indicate the thermal
comfort levels of occupants with a high degree of accuracy.

1.3 Electrode Activity
Electrode activity (EDA) refers to the electrical characteristics of skin. This activity includes two
related phenomena. The first is skin potential, which is intrinsic and self-generated. Such
potential is produced when dissimilar skin layers are placed in proximity and coupled with the
presence and movement of interstitial fluids. The biological significance of EDA is that it is
correlated to eccrine sweat gland activity. EDA measurements represent the changes in electrical
conductance of the skin because of such gland activity, which is a physiological signal that
Figure 1.1 Lab Setting of Dr . Choi’s
Research (Choi & Yeom, 2017)

Figure 1.2 Methodology of Dr . Choi’s
Research (Choi & Yeom, 2017)

3
shows increased sympathetic nervous system activity. Sympathetic nervous activity increases
sweat gland secretions, and increased EDA is an indicator of increased emotional response.

The change in skin potential is produced by electrical, auditory, or magnetic stimulation. This
gives us the chance to measure it when stimuli are introduced to the body. As eccrine glands are
in nearly all skin locations, and are found mostly on the palms of the hands, the fingertips, and
the soles of the feet, sensors should be placed in these locations to make test results more
accurate.

There are two kinds of skin conductance measurements: the phasic and tonic. The tonic value
represents the average skin conductance level, which is relatively consistent over time, and is
related to emotional response. The phasic value relates to the change in skin conductance as a
result of a reaction to a stimulus. When the stimulus is introduced, skin conductance rises for a
short time and then returns to tonic levels.

Skin conductance responses are described by four parameters: the response peak amplitude, the
latency of response, the rise time to the response peak, and the recovery time after peak. The
following pictures show an EDA measurement made with BIOPAC EDA testing devices. It is
obvious that EDA is a fluctuating parameter, which suggests a reflection of outside stimulation.
This factor can help us understand the relationship between stress level and thermal sensation. In
addition, from its strong reaction to outside stimulation from, we can determine whether external
influences affect occupants’ instantaneous psychological activity and their thermal sensations.
This, in turn, helps us to understand more about thermal investigation results at each time point,
and thus to make more accurate predictions.

EDA can show emotional responses to outside stimuli, and represent the mental states of the
occupants at that time, which help us make clear about different thermal sensations of occupants
with the consideration of psychological activities, and then make our thermal prediction more
precisely.

Figure 1.3 The BIOP AC EDA Measurement Interface
(Retrieved from: BIOP AC: EDA INTRODUCTORY
GUIDE)

4
1.4 Heart Rate and Heart Rate Variability
As thermal sensation is a relatively subjective parameter, psychological factors need to be
considered. To achieve a more accurate result, we use heart rate variability (HRV) as a measure
in order to estimate the psychological states of study volunteers.

HRV is the variation in the time interval between heartbeats. Physically speaking, HRV is a
measure of neurovegetative activity and the autonomous function of the heart. In a clinical
application, HRV shows the risk of injuries or overtraining, and can be used to guide
performance improvements. However, more and more studies have shown the relationship
between HRV and psychological activities. Specifically, HRV is related to emotional response
and mental parts of humanity, which makes it available to be applied in psychophysiology as a
representation of stress.

Basic study has shown the connection between HRV and easy-measured factors such as HR, R-R
interval and photoplethysmography (PPG), which makes it possible for us to access this data. In
cases where psychological or stress-related characteristics are relevant, HRV is critical to our
ability to make correct estimations about the psychological activities of occupants, and helps to
make our predictions of thermal sensation more useful to them.

1.5 Computer Control System
Currently most of the studies about the human body index focus on how to describe clinical
relationships more precisely. However, the focus should be on how to apply this information to
practice. With the rapid developments in computer science, the automation of control systems
based on machine learning has become a new growth point. Based on indexes above, there is
potential to make more precise predictions about occupants’ preferred thermal conditions, and
then to transform the predictions into signals that can be identified by control system, in order to
change the mode of heating/cooling devices.

To achieve this goal, we need to apply machine learning to define the process for making an
accurate estimation of current thermal comfort based on a human index, and then use signal
transferring technology to connect this estimation to building devices and systems to achieve an
auto-control process.

In our project, it is quite possible that the relationship between our independent features and
outputs is non-linear, which means the multiple linear regression might not match the practical
situation for our project. Accordingly, we need a new method to describe more complex relations.
A neural network is most appropriate.

For example, the basic mathematical expression of linear regression, one of the simplest
regressions in machine learning, can be expressed as follows (we define x0=1, which corresponds
to Θ0):

In our project, we can define skin temperature, EDA and HRV index as x1, x2 and x3 respectively,
while ℎ
"
(𝑥) represents thermal sensation.

5

When considering this analysis, we need to think about the cost function, which shows the average
deviation of this model. In the multiple linear regression, cost function is expressed as:

This function is used to compare the average difference between the squared value of each model
output and the true value of its corresponding points. In order to get the most precise model for
our prediction, we need to find a proper ℎ
"
(𝑥) to minimize the value of J(Θ), which has the
smallest deviation from the reality. We can regard therefore regard it as the multiple linear
regression model we need for this simulation.

In order to obtain the adoptable weights’ values, we need to find a valid mathematical process to
adjust them repeatedly until they are close to the required ones. This process is called iteration.

In linear regression, a gradient descent is usually used as iteration process.

In a gradient descent, the basic idea is that the shape cost function J(Θ) of linear regression is
convex, like a valley. For a function, the direction of gradient is always the fastest direction in
which the function value increases. If we can find a small step α, and make the step along the
opposite direction of its current gradient each time, several stages later, when the gradient is at its
lowest point, the function value will tend to be stable (“机器学习,” n.d.).

It is obviously difficult to get the appropriate smallest point. In this case, when J(Θ) is less than
10
-3
in one iteration, we think that the function is a convergence that can be applicable to our
project.

We prefer to use linear regression for the algorithm referring to environmental parameters and
thermal sensations, because the temperature is continuous and we need to find the consecutive
relationship.

However, as noted, it is possible that the relationship between our independent features and outputs
is non-linear, which means multiple linear regression might not be sufficient for our purposes. We
need a new method to describe more complex relations, such as a neural network.

Figure 1.4 Basic Idea of Gradient Descent

6

A neural network is a classification method that involves adding extra hidden layers between the
input layer and the output layer. The features in hidden layers are directly related to features in
the previous ones, which helps to create more complicate mathematical relationships between
initial inputs and outputs, and makes the simulation model closer to the real situation.

An artificial neural network (ANN) classifies thermal conditions into different states, which helps
to identify how to use HVAC devices to optimize indoor environments.

However, ANN is a relatively complex machine-learning algorithm that has some drawbacks, such
as problems with the gradient descent. For example, it is difficult to defining the value of step α.
If α is too small, the iteration process is very slow and time-consuming. However, if α is too large,
we cannot guarantee that J(Θ) is decreasing for every iteration. In addition, gradient descent needs
feature-scaling, which means getting every feature into a close variable range, in order to increase
the speed of convergence. The biggest problem is that there are so many iterations in the gradient
descent process, and it might contain a large calculation that costs time. These problems obviously
increase the difficulty of ANN’s application in real life.

In order to make the machine-learning process valid and proper, therefore, the application of ANN
is usually achieved with the help of a relevant machine-learning toolkit, in which professional
work has been done to guarantee the model’s accuracy. Under this circumstance, ANN is usually
applied as a black box, and we can only decide the input /output to be analyzed. Before using ANN,
we need to find a simulation model whose process of analysis is similar to ANN’s, and can help
to identify the correlations between all collected human index and environmental parameter data
and occupants’ thermal feedback. Based on comparisons between these correlations’ strengths, we
can find out the right features to use as ANN’s input data.

Like the ANN model, logistic regression also uses a sigmoid function as an activation function in
order to achieve estimation. The function is expressed as follows:

Where the value of hθ(x) means the possibility of specific events. When θ
T
x > 0, the outcome of
hθ(x) is greater than 0.5; when θ
T
x < 0, the outcome of hθ(x) is less than 0.5.
Figure 1.5 Basic Idea of Neural Network

7

The difference is that logistic regression has Boolean outputs, and its logistic cost function is
expressed as:

Where the sub-definition is:

In order to get the proper value of Θ with the smallest cost function value, we also apply a
descending gradient to have the solution.

In the logistic regression model, the outcome should be Boolean. The model divides all situations
into two groups, represented by two outcomes (0 and 1). When θ
T
x < 0, the model defined the
circumstance as a situation represented by outcome 0. When θ
T
x > 0, the model defines the
circumstance as a situation represented by outcome 1.

Figure 1.6 Outcome of Sigmoid
Function

8
2 BACKGROUND AND LITERATURE REVIEW
2.1 Current Thermal Estimation and Control Strategy
Traditionally, the target for thermal estimation is to approach a group’s thermal comfort level.
Currently, Predicted Mean Vote (PMV) is the most-widely applied thermal prediction model all
over the world. It uses a seven-point scale (-3~+3) to represent the group’s thermal sensations,
from cold to hot.

However, PMV lacks the capacity to account for individuals’ thermal preferences. Factors such
as age, gender and race can impact people’s requirements for their surrounding thermal
conditions (Deng, Wang, Li, Miao, & Zhao, 2017). Previous research reveals the PMV model is
not valid for predicting Indians’ subjective thermal responses (Maiti, 2014), which suggests the
importance of considering an individual thermal estimation model.

Besides, PMV has obvious drawbacks for predicting different thermal situations (ter Mors,
Hensen, Loomans, & Boerstra, 2011). Experiments focusing on the accuracy of PMV prediction
in everyday thermal environments suggests PMV results are easily biased because of factors
including different clothing insulation and metabolic rates (Humphreys & Nicol, 2002), and the
experiment result shows a much narrower range in which PMV can be applied as compared to
with its design standards. A new thermal estimation model is needed in order to get a more
acceptable level of prediction accuracy.

In fact, the question of how to satisfy occupants’ subjective requirements has become a
breakthrough point for relevant optimizing work. More and more relative projects have begun to
focus on the idea of “user-centered control,” where cross-disciplinary work has been applied to
create control strategies that promote the expression of occupants’ subjective needs in the control
process with the help of interfaces or other tele-controllers. Relevant attempts have proceeded in
many ways. In 2018, one study used RGB video images to create a visual-based thermal comfort
quantification for individuals in order to optimize thermal control, which helped apply
individuals’ preferences about their thermal comfort levels into the control processes of HVAC
devices (Jung & Jazizadeh, 2018).

Also, user-centered control for individuals shows advantages in accuracy. One project in
University of Southern California applied machine learning to achieve a human-based
environment control, which showed an obvious satisfaction improvement of individual
preference when compared with the PMV model (Zhong & Choi, 2017). Strengthening the
connection between control strategy and users is necessary, and it is valuable to create a human-
based control environment to achieve individual thermal satisfaction (Yeom, Choi, & Zhu,
2017).

2.2 Skin Temperature and Thermal Comfort
Previous research has shown the relationship between skin temperature and thermal comfort,
especially local skin temperatures, including forehead, neck, chest, back, arm, belly, waist and
wrist (front and back), with accuracy rates from 88.21% to 94.39% when gender and BMI are
considered (Choi & Yeom, 2017).

9

As is shown in the result, local skin temperature has a tight connection with the whole thermal
sensation, although the changing rate of skin temperature needs to be taken into account. This
finding shows the potential to utilize the localized skin temperatures of specific body parts to
describe thermal sensations, and local skin temperature have the potential to represent the
thermal conditions of the current indoor environment based on occupants’ sensations.

Another article shows the investigated relationship between thermal comfort and skin
temperature for different local body parts (Metzmacher, Wölki, Schmidt, Frisch, & van Treeck,
2018). It suggests an obvious influence from clothing, which might have a linear relationship to
skin temperature decreasing with different degrees, and it argues that clothing should be a
controlled factor during experiments because there is large fluctuation when changing it.

Moreover, this article shows a big difference in chest-skin temperature based on different
clothing.

Table 2.1 The Relation Strength between Skin Temp.
and Thermal Sensation in Dr . Choi’s Research (Choi
& Yeom, 2017)
Figure 2.1 The Effect of Skin Temp.
(Chest) Caused by Different Clothing
(Metzmacher et al., 2018)

10

2.3 Thermal Comfort and Psychological Activity
As a subjective estimation of the ambient environment, the possibilities that thermal comfort is
related to psychology has been considered in previous research for a long time. As early as 1998,
there is a paper in literature review form which shows the psychological effect on thermal
sensation (I & Dear, 1998). This paper drew the conclusion that behavioral adjustment and
expectation have a much bigger impact on thermal adaption than physiological acclimatization.
Also, the correlation between thermal comfort and psychological factors has been explored
recently (Klemm, Heusinkveld, Lenzholzer, & van Hove, 2015), and experiments were executed
in a specific environment, in which researchers discovered the impact of psychological activity
on occupants’ thermal comfort (Klemm, Heusinkveld, Lenzholzer, Jacobs, & Van Hove, 2015).

There is also a paper presenting the idea of a self-assessment manikin (SAM), which a
represents self-perceived level of pleasure, arousal and dominance (Schweiker, Brasche, Bischof,
Hawighorst, & Wagner, 2013). This article mentions that occupants feel better in their
surrounding thermal condition if they have ways to achieve control, which might improve their
self-perceived levels of dominance.

Similar result were achieved in another recent study that explored psychological effects on
thermal comfort(Luo et al., 2016). It set a comparison between test groups with or without
perceived control, which meant the occupants could ask researchers to adjust indoor temperature
when they feel uncomfortable, but the actual temperature did not change. The comparison result
can be used to show the psychological effect.

Figure 2.2 The Difference in Skin Temp.
(Chest) Caused by Different Clothing
(Metzmacher et al., 2018)

11

There are obvious differences between these two groups. The thermal comfort level of occupants
has been improved with perceived control, which shows a psychological tendency for humans to
estimate their thermal comfort level.

2.4 Relationship Between Stimulation and Thermal Comfort
Although thermal condition is traditionally recognized in physical ways, more and more studies
have shown that psychological activities have a huge impact on people’s thermal sensations
(Choi, Loftness, & Lee, 2012).

A previous research paper presented the idea of “psychological adaption,” which refers to
changes in physiological responses resulting from repeated exposure to a stimulus. These
changes decrease following the exposure (Nikolopoulou & Steemers, 2003).

This paper focuses on the influence of psychological adaption on thermal comfort in exterior
urban spaces, and it indicates that physical parameters account for just 50 percent of actual
thermal sensation votes, while psychological adaption is likely to be more significant. As is
shown in the following picture, this paper has marked six interactive factors of psychological
adaption in thermal satisfaction. One of these factors, environmental stimulation, has been
regarded as having a particularly important impact on people’s thermal comfort (Sato & Usui,
2018), because of the increasing belief that a variable environment is more acceptable for people
than a stable one. The paper says such stimulation might be the reason that people prefer to stay
outside.

Figure 2.3 Effect of perceived control on comfort assessment: (a) thermal comfort; (b) thermal
satisfaction (Luo et al., 2016)

12

Although it pays attention to outside space, this paper was published very early (in 2003) and the
authors indicated the problem was that they lacked the technology to quantify psychological
factors, such as the influence of environmental stimulation. They suggested that finding ways to
quantify the interrelationships of these psychological factors could help improve understandings
of even indoor environments.

Another recent paper has shown the connection between local body warm stimulation and
thermal comfort (Sato & Usui, 2018). This study planned to use a device to generate warm
stimulation to the waist. Although it focused on local stimulation by devices, it found that
increasing area of warm stimulation can raise not only perceived intensity but also the comfort
level, and that perceived intensity has more impact on thermal comfort than the intensity of the
stimulation. The authors concluded that the increasing of stimulated area is more important for
improving thermal comfort than the increasing of stimulation intensity, which might be
considered as a potential for our project to generate stimulation by AC adjustment.

2.5 Relationship Between HRV and Thermal Comfort
HRV has been introduced into the process of quantifying psychological effects on thermal
sensations, where it is seen as a factor representing the state of emotional change.

There is one paper that attempts to determine the relationship between HRV and psychological
activities (Lischke et al., 2018). However, this paper finds that an increase in HRV can represent
lower stress in an individual’s social life, while the relationship seems to be much weaker in the
test of family life. It mentioned that this might be explained by concluding that occupants feel
less stress when in a more familiar environment. In any event, it shows the potential to regard
HRV as a reflection of psychological activity.

Other papers related to our focus more directly. For example, one of Hui Zhu’s studies has
revealed the relationship between HRV and thermal comfort and thermal sensation (Zhu et al.,
2018). This paper started with the concept that HRV contains the signal of the autonomic
Figure 2.4 Lines of influence between the different
parameters of psychological adaptation (Nikolopoulou
& Steemers, 2003)

13
nervous system, and it introduced an HRV-related idea of “LF/HF,” which represents the ratio of
the low-frequency component (0.04-0.15HZ) to the high-frequency component (HF). The
relationship between LF/HF and thermal sensation/comfort has been shown as follows:

The lower value of LF/HF shows a more acceptable thermal sensation/comfort level for tested
occupants. This study revealed that the high LF/HF value might indicate obvious sympathetic
nerve activities, which impact the subjective estimation of people’s thermal comfort/sensation
level. According to the paper, the reason might be the dominated state of sympathetic nerve
activities while LF/HF is high.

2.6 Summary
From the literature review, we can see that both occupants’ physical conditions and their
psychological one’s reflection in both physical and psychological way can present signals for
their thermal sensation and comfort level, and the connection with factors such as local skin
temperature, outside stimulation and LF/HF of HRV has been shown in previous studies. Based
on our measurement of skin temperature, HRV related index and EDA, which can represent the
body’s reflection of outside stimulation quantitatively, there should be a mathematical method
that can present the thermal-feeling information captured by these factors. In addition, previous
research has shown the accuracy for skin temperature for different local parts, which helps us to
design our measurements more precisely.

Figure 2.5 Relationship between the thermal sensation (a), comfort level (b) and the LF/HF
(n = 6, p b 0.05) (Zhu et al., 2018)

14
3. METHODOLOGY
3.1 Methodology diagram
The high-level methodology diagram is shown as follows, our project can be mainly separated
into three parts: data collection, algorithm design and real-time control.

Figure 3.1: High-Level Methodological Stages

We collected data for both the human indexes we choose and environmental parameters such as
temperature and relative humidity. We also had questionnaires to record people’s thermal
comfort and sensation level. All data was collected in the LabVIEW interface, and then we used
the LabVIEW ANN Toolkit to apply machine learning based on our data collection. We used the
principle of supervised learning, in which we designed our algorithm based on our input
laboratory data, called training data. After that we tested our algorithm in real life, in order to see
whether its accuracy was acceptable.

Our algorithm contains three loops: the estimation loop to detect whether people feel
comfortable; the sensation loop, to explore whether the indoor environment is hot, cold or
Figure 3.2: High-Level Steps of Loop

15
neutral; and the control loop, to determine the HVAC operation strategy based on the comfort
and sensation states.

Figure 3.3: Methodology of Data Collection and Algorithm Design

16

Figure 3.3 shows the more detailed workflow of data collection and algorithm design. Here, we
needed to obtain data for skin temperature, LF/HF value of HRV and EDA for studied human
indexes. Environmental temperature, relative humidity, CO2 concentration and radiant temperature
were our chosen environmental parameters.

Based on the data collection, we created two algorithms. One was for HVAC triggering based on
the relationship between thermal comfort and human indexes, while the other was for deciding the
adjustment direction for the HVAC system based on the relationship between environmental
parameters and thermal sensation results.

17

Figure 3.4 shows the more detailed workflow of real-time control, in which we performed practical
tests on our two designed algorithms. The real time data was collected once every 10 minutes, and
the algorithm made indoor environmental change by HVAC devices based on our monitoring data.
The result of thermal comfort investigation was also recorded and used for comparison to
determine whether our auto-control system was correct.
Figure 3.4: Methodology of Real-Time Control

18

3.2 Data Collection
The first step of my project was data collection. Data regarding human factors like skin temperature,
electrode activity and psychological factors like HRV were collected and processed by machine
learning to make a thermal comfort sensation prediction algorithm, and help find potential
applications based on real-time data. In addition, environmental parameter data were considered
to make result more objective.

3.2.1 Data Type
The data needed were divided into three groups: human parameters, environmental parameters,
and questionnaire survey results. Human parameters included local skin temperature, electrode
activity and heart rate (HR), which could help get the LF/HF parameter of heart rate variability
(HRV). Environmental parameters included temperature, CO2 concentration, radiant temperature,
and relative humidity.

For human parameters, local skin temperature and electrode activity were regarded as direct
representations of occupants’ thermal sensations, while psychologically-related parameters
including HR and HRV allowed us to promote psychological analysis and eliminate part of the
subjective sensation for thermal comfort. The human parameters were expected to function as
signals to predict the thermal estimation of occupants.

The values of the specified environmental parameters should be regarded as a comprehensive
index representing different environmental conditions for indoor space. Data collected at this stage
included temperature, relative humidity, CO2 concentration and radiant temperature. Each set of
data recorded and described current environmental conditions for occupants’ specific thermal
sensation, which helped us to find out how to improve people’s thermal comfort level by adjusting
these environmental parameters.

For the questionnaire survey, we recorded the thermal sensation and thermal comfort level results
for occupants in each specified environmental condition. The questionnaire used number -2, -1, 0,
1, 2 to represent different degrees of these factors. For thermal sensation, the indicated numbers
described the degree as very cold, cold, neutral, hot, and very hot, respectively. For thermal
comfort, the sequence indicated the degree was very uncomfortable, uncomfortable, slightly
uncomfortable, neutral, slightly comfortable, comfortable, and very comfortable. Questionnaire
survey results helped us to determine occupants’ subjective estimations of indoor thermal
conditions, which was considered key to determining the adjustment routine of building devices.

3.2.2 Collection Method
The methods of data collection were based on the nature of monitoring devices, and were different
for each group of parameters. For human parameters, wearable and wireless sensors were applied
to collect the data.

The brief process of human data collecting was as follows:

19

Sensors measured constant physical body index data, and then transmitted the record as a digital
signal to a data receiver. The interface showed the recording results, which were saved in tabular
data format, allowing us to make data analysis using a LabVIEW Machine Learning Toolkit.

For environmental parameter sensor, a tower-shaped monitoring device was fixed near the
occupants in the center of our lab. Environmental sensors were installed on the device to make
records.

A survey questionnaire was completed during the experiment. Occupants were asked to describe
their thermal-related feelings for each different set environment.

3.3 Algorithm Design (From Methodology Diagram)
The object of the algorithm design was to create a model to predict the thermal feelings of
occupants based on collected real-time data. We needed to make conclusions and find the
mathematic relationship between the thermal estimation and the data collected in the lab. Data
analysis was performed after collecting process was finished, and we expected to use machine
learning to promote the algorithm-defining process.

3.4 Real Time Control (From Methodology Diagram)
The final section of our project was to achieve the self-control of heating/cooling devices based
on real-time monitoring data. With the defined algorithm, we could obtain the thermal preference
of environment condition with certain fixed types of human and environment parameters.
Accordingly, we planned to use wearable sensors we applied for experiments to get these types of
data and use the algorithm to make thermal condition estimation. A computer system interpreted
the thermal preference results and gave commands to indoor heating/cooling devices to change
their operation mode through a power delay design.

3.5 Software and Hardware
For our program, we needed software to deal with collected data and generate specific parameter
values, output the dataset in a specified format, and define the algorithm. In addition, we needed
hardware to complete the index monitoring and power relay process.
Figure 3.5: Human monitoring workflow Figure 3.6: Component of human monitoring
devices

20

3.5.1 Heart Rate Variability Generator
We used Vernier Heart Rate sensor to get continuous heart rate and R-R interval data, and we used
Kubios software to generate heart rate variability (HRV) the value.

Kubios is an HRV analysis program. In this project, we input tabular-format input HR and R-R
interval data into this software. It then calculated the value of HRV and analyzed it to produce
output.

The pictures above show the basic workflow of Kubios. Vertical tabular-formed R-R interval value
text files were used as input, and the software generated output representing an HRV triangular
index.

Figure 3.7: Kubios HRV
Measurement interface
Figure 3.8: Input of Kubios
Figure 3.9: Output of Kubios

21
3.5.2 BIOPAC EDA Devices
In our project, we used the psychologically-related factor EDA (electrode activity) to help analyze
the thermal reflection of occupants. Although, as mentioned before, we could measure HRV using
HR sensors, Kubios needed the per-second heart-rate data of occupants, while the period of our
data collection in experiments was five seconds. Accordingly, in addition to HRV, we also
considered EDA for helping analyze occupants’ psychological activities.

BIOPAC device setting: As mentioned, we needed a series of devices to make a complete EDA
measurement, and BIOPAC EDA setting could help measure the EDA data directly. In addition,
the BIOPAC setting could also achieve the measurement of photoplethysmography (PPG) data,
which could be transformed into HRV data directly, and we kept the PPG measurements during
our experiments.

Figure 3.10 shows the EDA and PPG sensor of BIOPAC setting. Electrodes were sensors applied
to occupants’ skin to detect change of skin sweat gland. PPG wireless pulse ear clips were
attached to people’s ears to get PPG data, which was an alternative for heartbeat interval
measurements, and helped us to obtain the value of occupants’ HRV index.

Other hardware we used for experiments included MP160 and transmitter/receiver pairs. The
electrodes and ear clip transmitted digital signals to a transmitter, which was a relatively small
device, as shown in Figure 3.11. Then the digital information was conducted to interface device
mp160. After that, the signal amplifier sent the digital signal to our computer system for data
recording.

Figure 3.10 EDA Electrodes and PPG
wireless pulse earclip

22

3.5.3 LabVIEW
Before testing, we used software LabVIEW to create a platform that helped process and save data
from sensors. LabVIEW is system engineering software, designed to be applied in projects that
need test, measurement, and control with quick access to hardware and data insights. We used it
as a type of visual programming to help us input the local body part data we recorded.

As is shown in Figure 3.13, LabVIEW software was needed to combine extensive hardware
integration, which dealt with collected data from sensor and input it into computer system. Figure
3.14 shows a concept view of LabVIEW programming interface. In our project, it was used to
define the data dealing process and helped generate different data in an acceptable format for other
analysis software (“What Is LabVIEW?,” n.d.).

Figure 3.15 shows the setting interface for Dr. Choi’s research about the relationship between skin
temperature and thermal comfort.

Figure 3.11 Interface Device
MP160
Figure 3.12 Transmitter
and Receiver
Figure 3.13 Labview Workflow
Figure 3.14 Labview Programming
Conceptial

23

During our experiments, this interface presented specified local body parts for measurement, along
with the result of skin-temperature test results. The left part showed the heart rate test, while the
lower part showed different indoor surface temperatures.

For this study, data was generated in a tabular format, which was acceptable for LabVIEW
Machine Learning Toolkit and Matlab, to help finish the algorithm definition.

3.5.4 Matlab and LabVIEW Machine Learning Toolkit
In order to achieve the machine-learning process, our projects needed data analysis software to
define algorithms based on our collected data.

Matlab is mathematical analysis software applied to data analysis, model creation and algorithm
development. Importantly, Matlab helps promote machine learning that defines algorithms to help
make better decisions, which assists in completing processes such as clustering, regression, and
classification. The simple workflow elements related to machine learning in Matlab include
importing, cleaning, exploration, training, testing and deployment. Matlab has interactive tools to
help with data collection and importing, and it removes rows with missing values automatically to
produce a data set that is analyzable. After that, the work is about choosing the best algorithm for
the project, and then training and testing the collected data to see if can be deployed dependably.

The LabVIEW Machine Learning Toolkit is a LabVIEW plug-in that contains the defined process
for data analyzing for machine learning, and it automatically uses machine-learning algorithms
(such as ANN) to process data collected in this software.

For this project, we applied the LabVIEW Machine Learning Toolkit and Matlab to help achieve
a machine learning process and define algorithms about thermal comfort prediction. The input data
were our collected human index and environmental parameter data mentioned before, and we used
such analysis software to evaluate the mathematical relationship between these parameters and the
questionnaire results to figure out if it was possible to make precise predictions about occupants’
thermal sensations.

Figure 3.15 Labview Result Interface

24
3.5.5 Power Relay Hardware
For this project, we needed relay hardware to transform electric signals, and achieve the
connection between HVAC devices and computer instruments. We chose Ogrmar SSR-25 DA
with 3-32V DC input and 24-280V AC output as our power relay (“Ogrmar SSR-25 DA 25A 3-
32V DC / 24-380V AC Solid State Relay and Heat Sink,” n.d.).

3.5.6 General Sensors
We used the Vernier collection of sensors to set environmental parameters. As for sensors for
human indexes, we used BIOPAC devices for EDA testing, and applied Vernier sets for skin
temperature and HRV indexes.

Environment parameter sensors are shown as follows. All the environment sensors were fixed in
a sensor holder. Environment data was first recorded by different sensors, and then the data was
used as an input to the data acquisition interface (DAQ) for data collecting and processing. Lastly,
data were saved in tabular form so they could be applied in the computer system (“Vernier,” n.d.).
Figure 3.16 Ogrmar SSR-25 DA

25

In addition to the above, we applied one sensor DAQ, a USB data-acquisition interface, during our
experiments. It helped collect data from all sensors and process the data to be applicable.

We used the surface temperature sensors above to test skin temperature for the human indexes. we
also adopted the Vernier heart rate monitor to get obtain rate data, which was processed into an
HRV related index using Kubios.

Figure 3.17 Environmental sensor
devices

Figure 3.19 Relative
Humidity Sensor
Figure 3.20 Vernier
CO2 Gas Sensor
Figure 3.21 Vernier
Surface Temperature
Sensor
Figure 3.18 Stainless Steel
Temperature Probe User Manual
Figure 3.22 Vernier Sensor DAQ

26

3.6 Data Analysis and Algorithm Definition
After data collection, we started to find methods for analyzing the relationship between human
indexes and thermal sensation.

We knew the value of our input (human indexes and environmental parameters) and the result of
our output (thermal sensation/comfort) for our collected data, so we considered the supervised
learning for our data analysis, which is outlined below. We adopted 75 percent of the data as
training data, and the remaining 25 percent as testing data.

In order to pick up specific human indexes and environment parameters to achieve accurate
estimation, we considered both thermal comfort and thermal sensation. There were two
correlations we needed to explore. The first was the relationship between human indexes and
thermal comfort, which decided whether people felt comfortable with their thermal condition. The
results determined of whether they needed to trigger the HVAC system to optimize the indoor
environment. We called the algorithm needed to express the correlations between thermal comfort
and human indexes the “HVAC Triggering Algorithm Part.”

The second correlation was the relationship between environmental parameters and thermal
sensations. Th corresponding algorithm, named the “Set Point Definition Algorithm Part,”
determined the indoor temperature that occupants preferred in a given room. This process indicated
how we needed to change our indoor environment (heating/cooling).

Figure 3.23 Vernier Heart Rate Monitor
Figure 3.24 Workflow of Supervised
Learning

27
3.6.1 Artificial Neural Network for Machine Learning
As suggested in the Background section, we chose ANN for this process.

As was mentioned before, our collected data was in tabular form. In order to process raw data
correctly, the machine-learning software used knowledge of matrices to achieve ANN
classification. The software defined this as x
(i)
, which represented the i-th data point for every data
point.

In this expression, x0 was defined as 1.

Based on this, we supposed that we had collected m data points, and we defined X=&
(𝑥
(')
)
(
…
(𝑥
(*)
)
(
+, and
defined y=&
𝑦
(')
…
𝑦
(*)
+. In this way, the software could adopt and analyze our tabular-form data.

The application of the ANN model was achieved using the LabVIEW Toolkit, which identified
the correlations between collected data (divided into input and output). The black-box form of
the Toolkit only allowed us to decide input, output, and the function of the algorithm, and each
algorithm only allowed 1 type of output, where the toolkit’s inner mathematical structure was
unalterable. For this reason, we had to determine appropriate input features for analysis and then
combine our thermal comfort and thermal sensation feedback as one output.

In the practical application, before using ANN, we finished combining the results of thermal
comfort and thermal sensation and got one unified output. In order to improve its accuracy, we
narrowed the output to a three-point scale, and ANN divided the results of thermal preference
classifications as follows:

.
1
0
0
1 .
0
1
0
1 .
0
0
1
1
Uncomfortable: cold Neutral Uncomfortable: hot

To simplify the analyzing process for training data, we used “-1” to represent the position of
“uncomfortable: cold,” where “0” and “1”were used to represent “Neutral” and “Uncomfortable:
hot,” respectively. The toolkit identified the meaning of each number automatically.

28

3.6.2 Logistic Regression for Feature Generation
As mentioned in the Background section, we started determining input features for ANN analysis
before using the machine-learning toolkit. We considered logistic regression, which had ta similar
estimating process as the artificial neural network model and which expressed the thermal
comfort/sensation level of occupants.

Logistic regression was for situations in which the output was Boolean (e.g. 1/0, true/false), and

In this hypothesis, we divided our model into two parts: thermal comfort and thermal sensation.
The result determined whether we needed to trigger the operation of our HVAC system. If the
simulation result showed an uncomfortable situation, the HVAC device was triggered, and the
operation mode of HVAC was based on the output of thermal sensation estimation.

In order to achieve this process, we made a Python code for logistic regression, in which the
activation function and iteration process of logistic regression were defined. We input all collected
data (including input features and outputs) into this code. It generated the weights of all input
features, and we determined the most important features by comparing their weight values.

Data was randomly divided into two parts: 75 percent was used as training data, and 25 percent
as test data.

Figure 3.25 Logistic Regression Python Code

29

Given that measuring five skin temperatures was too difficult in practice, we decided to choose
only two of these five for making estimations of thermal comfort. Generally, we opted for
10,000,000 iterations, though a minority of these 20 dataset samples were iterated 20,000,000
times because their correlations between inputs and outputs (hot/cold,
comfortable/uncomfortable) were relatively weak when the iteration time was 10,000,000.

First, all five skin temperatures and thermal comfort feedbacks were put into the logistic
regression model. The model generated a weight for each body part’s skin temperature. The two
highest-weight body parts were then chosen for individual model analysis.

3.7 Summary
This project focused on the computing design of intelligent self-control heating/cooling systems
by using machine learning to deal with lab-collected data and create an algorithm to define a
formula for thermal preference predictions according to different values of specified
human/environmental factors.

The number of volunteers was 20 (10 men and 10 women), and we used wireless sensors for
human testing. For environmental parameters, we applied fixed monitoring devices to make
records. All the recording data were saved as digital signals transmitted into our computer
system.

After collecting, all saved data were input to LabVIEW to get spreadsheet data file, and then
Matlab and LabVIEW Machine Learning Toolkit were applied in a machine-learning process in
order to define an algorithm. The algorithm was then applied to create thermal preference
estimations.

Lastly, we combined the building control system and monitoring devices. We used wearable
sensors to detect the real-time data of occupants, and our algorithm got a thermal preference
result based on the real-time data, which sent electrical signals to HVAC devices by power relay
and adjusted the operation mode of heating and cooling devices to optimize indoor environment.

Figure 3.26 Logistic Regression Model
Outline

30
4. Data Collection and Processing
4.1 Experiment for Training Data Collection
The first step, collecting data for machine learning, was executed at the USC Watt Hall
laboratory chamber B11. The purpose of training data collection was to express the value of skin
temperatures and indoor temperature when occupants were under different thermal conditions
that led to different thermal preferences. We planned to simulate a real work situation, and
occupants were expected to sit in front of a desk, where they were allowed to do simple work on
their computers.

Our experiments included 20 participants (10 men and 10 women). In each experiment, only one
participant sat in the center of our laboratory. We used healthcare tape to fix sensors to their
body, including their upper arm, the back of their neck, the front and back of their wrist, and
their waist.

As shown in Figure 4.1, we allowed participants to do some slight computer work during the
experiment to simulate an office work situation. They were not allowed to play games or watch
videos, as these activities might have had noticeable effects on their psychological conditions.
We provided bottled water during each experiment. Participants were required to stay in their
chairs because of the limitations of the sensor connections.

In addition to skin temperature, we also measured electrode activity through fingers, and we had
a questionnaire to get people’s immediate feedback about their thermal comfort and thermal
sensation feelings. The questionnaire also asked participants to estimate their current stress
levels, and recorded whether they wanted to use HVAC to change current indoor temperature in
order to help us judge their thermal comfort level.

Figure 4.1 Participant During Experiment

31
Because the machine learning needed a larger number of valid data points to analyze, we set two
prerequisites for picking up continuous data points: (1) Each had to be taken in the three minutes
before each test point. (2) The temperature change from the temperature at the test point had to
be less than one degree Celsius.

When the continuous data points met both prerequisites, we thought the indoor thermal condition
was relatively stable, and we assumed that during this period the thermal comfort and sensation
level for the occupant was the same. We selected all data points that met these conditions as our
valid data for machine learning.

Table 4.1 required feedback about occupants’ current thermal comfort/sensation level and
psychological feelings. It also asked whether people need to manually change indoor thermal
conditions.

4.2 Logistic Regression Process for Algorithm Design
4.2.1 Logistic Regression for Thermal Comfort Feature Selection
As was shown in the Methodology section, we applied logistic regression for selecting
appropriate features (specific skin temperatures and indoor temperature) in order to predict
people’s thermal preferences. As mentioned, the activation of ANN and logistic regression is the
same, so we planned to select the best input features (skin temperatures) from the logistic
Table 4.1 Experiment Questionnaire

32
regression for thermal preference prediction and then apply these features in the ANN model.
Logistic regression generated the weight of each input factor, which helped us determine which
factor had the most obvious impact on thermal comfort estimation, and the weight was treated as
the criterion for us to select the ANN input feature.

Before the machine learning stage, we had to decide whether a general model (one trained with
all 20 participants’ data) and an individual model (one trained with its corresponding occupant’s
data) were both meaningful for thermal preference estimation.

However, we found an obvious difference among our 20 participants. For instance, during the
experiment involving Occupant 1, the lowest measured wrist-front temperature was 34.39℃,
while for Occupant 2, the highest measured wrist-front temperature was 33.24℃. In this case,
the combination of all occupants’ data would have misled the machine learning process for using
skin temperatures to classify people’s thermal preferences. Accordingly, we decided to build an
individual model for each occupant.

Logistic regression used the same activation function as an artificial neural network model
(sigmoid function). The difference between these two models was that logistic regression
removed the hidden layer between input and output, where it processed input data by a single
mathematical step and adopted Boolean output format (0 and 1) to achieve a dividing effect.

There were two algorithms created by logistic regression: the algorithm for identifying people’s
thermal comfort (with outputs as yes/no), and the algorithm for identifying people’s thermal
sensation (with outputs as hot/cold).

Logistic regression needed the output of training data to be Boolean. For thermal sensation, data
with output of “very cold” and “cold” were arranged in “cold” group, and data with output of
“very hot” and “hot” were arranged in “hot” group. For data with sensation feedback of
“neutral,” if the occupant wanted to manually turn down the indoor temperature, we arranged the
data into “hot” group; if the occupant wanted to manually raise indoor temperature, we arranged
the data into “cold” group.

For thermal comfort, data with output of “very comfortable” and “comfortable” were arranged in
a “comfortable” group, and data with output as “very uncomfortable” and “uncomfortable” were
arranged in an “uncomfortable” group. For data with sensation feedback as “neutral,” if the
occupant wanted to manually change indoor temperature, we arranged the data into the
“uncomfortable” group; if the occupant preferred the current indoor temperature, we arranged
the data into the “comfortable” group.

In order to decide the input feature of our algorithms, skin temperatures and skin-temperature
change rates (temperature difference between adjacent data points) were designated as input
features.
For five skin-temperature change rates, the data sets of Occupant 1 and Occupant 2 were
regarded as input features for logistic regression respectively, where 75 percent of data were
used as training data and the remaining 25 percent were used for model testing. The weights for
each skin-temperature change rate were as follows:

33

However, in the test process, the accuracy was only 59.61 percent and 59.57 percent. It seemed
to be difficult for the algorithm to show correlations between skin-temperature change rates and
thermal comfort.

The datasets of Occupant 1 were used in logistic regression for algorithm modeling for five skin
temperatures, and the same processes for model training and testing were executed. The logistic
regression results were as follows (note that accuracy climbed to 81.13 percent):

Based on the testing results, the correlations between skin temperature and thermal comfort were
more obvious than correlations between skin-temperature change rates and thermal comfort.
Accordingly, skin temperatures were chosen as input features for thermal comfort prediction.

In a real-life application, it would be hard to measure the skin temperatures of five body parts.
Therefore, we chose to pick the two parts that had the strongest correlation to peoples’ thermal
feelings.

There were 193 valid processed data points obtained from Occupant 1. Selections from the
collected data and occupants’ feedback are presented below. For feedback records, the number
represented the number sequence shown in the questionnaire above. The overall number of’ valid
data points in our experiments was 3,647.

Table 4.2 shows the five-point data set (℃). Here, “Air1.1” means indoor air temperature
recorded at a height of 1.1 meters.

Waist
front
Upper
Arm
Wrist front Wrist back Neck Air1.1
33.97 35.42 34.54 33.59 35.81 31.04
33.94 35.42 34.55 33.60 35.82 31.02
33.92 35.42 34.56 33.60 35.84 31.00
33.91 35.38 34.55 33.62 35.84 30.99
Figure 4.2 Logistic Regression Result for Skin Temperature Change Rates
Figure 4.3 Logistic Regression Result of 5 Skin Temperatures

34
33.91 35.37 34.54 33.60 35.87 30.97

Table 4.3 shows one feedback record for occupants (the number from -3 to 3 means “very cold”
to “very hot” for “sensation”; “very uncomfortable” to “very comfortable” for “comfort”; and
“very relaxed” to “very stressful” for “stress.” The numbers from 1 to 3 for “change” mean
“wish to trigger heater,” “wish to trigger cooler” and “no need for HVAC,” respectively).

First, feedback for all five skin temperatures and thermal comfort levels were put into the logistic
regression model. The model generated a weight for each body part’s skin temperature. The two
highest-weight body parts were chosen for individual model analysis.

The logistic regression defined the formula: θ
T
x =w0x0+w1x1+w2x2+w3x3+w4x4+w5x5, where x1,
x2, x3, x4, and x5 represented local body points (waist front, upper arm, wrist front, wrist back and
neck), and w1, w2, w3, w4, and w5 represented the weight of each measurement point generated
by the logistic regression. The regression defined x0=1, so w0 was a constant in this mathematical
process.

The weights of each body part for every individual model are shown below. The number in the
left line represents the order number of the occupants, while the remainder of the table shows the
values of weights for the constant, waist front, upper arm, wrist front, wrist back and neck).

Constant
Waist
Front
Upper
Arm
Wrist
Front
Wrist
Back
Neck
No.1 566.97 49.92 360.62 618.91 -618.47 -441.23
No.2 122.64 -494.82 537.85 298.36 -641.32 248.80
No.3 405.14 -66.48 178.16 663.34 -683.54 -126.50
No.4 -2.21 -17.67 -18.24 106.53 -40.81 -32.55
No.5 -4411.36 -111.52 -11.72 -161.16 165.15 262.78
No.6 -645.02 -8.08 175.21 -343.34 145.25 57.48
No.7 -35.85 -372.87 253.64 363.24 -255.91 28.43
No.8 -349.51 885.27 182.54 -107.03 129.44 -1042.96
No.9 1.11 -197.79 684.27 -199.36 -39.85 -208.16
No.10 -4228.11 -484.73 759.21 986.08 -808.85 -323.63
No.11 7442.15 -6.01 -156.68 462.71 -277.48 -255.87
No.12 3520.04 -488.78 80.61 43.73 -374.44 653.88
No.13 708.20 900.89 -339.92 -1869.51 187.52 1078.83
No.14 -921.48 -215.72 -774.39 448.27 -142.93 659.26
Table 4.3 Sample of Thermal Preference Recording
Table 4.2 Temperature Recording
Sample for One Occupant

35
No.15 0.71 35.68 -19.82 -57.54 -65.59 97.86
No.16 19.17 411.48 -215.67 -84.67 167.59 -291.41
No.17 510.58 300.56 147.54 -493.01 11.29 32.82
No.18 1478.28 -319.21 -206.38 880.89 -27.02 -351.99
No.19 -0.26 -3.74 10.49 -21.25 -8.80 18.37
No.20 0.21 -90.32 26.43 -300.75 234.14 129.94

Our local body points measurements are shown in Figure 4.4. These the included wrist front,
wrist back, waist, upper arm, and the back of the neck. In order to simulate a real office situation,
we tried to install sensors on the left side of body, so that an occupant could use their right hand
for writing and computer operation.

Based on the regression result, we got the logistic regression weight values for different body
points. We then compared the weight values of our five measured body points, which helped us
select those that had the most obvious influence on an individual’s thermal comfort level.

Table 4.5 shows that overall, wrist-front, neck, and wrist-back temperatures were the three most
important skin temperatures for thermal comfort estimation.

Table 4.4 Weight of Each Body Part’ s Skin Temperature
for Thermal Comfort Estimation
Figure 4.4 Measurement Body Points

36
In order to verify whether two skin temperatures could be applied for estimating thermal comfort
level, we ran the logistic regression to define the formula θ
T
x =w0+w1x1+w2x2, where x1 meant
skin temperature 1, and x2 meant skin temperature 2 (The order of skin temperature was waist
front, upper arm, wrist front, wrist back and neck, which was the same as shown in Table 4.5).
When θ
T
x>0, the estimated result was “comfortable,” while when θ
T
x<0, the estimated result
was “uncomfortable.”

The result of logistic regression model is as follows:

Constant
Skin
Temp.1
Skin
Temp.2
Accuracy
No.1 -1830.39 486.47 -446.67 77.08%
No.2 2022.42 18.35 -86.64 77.08%
No.3 -413.05 -51.24 65.09 68.00%
No.4 -4756.43 -6.13 137.45 100%
No.5 -2271.26 5.23 61.32 100%
No.6 -1151.5 47.73 -12.53 74%
No.7 1116.17 -222.87 197.75 75%
No.8 169.36 249.45 -249.57 70.45%
No.9 4260.79 124.82 -233.63 45.10%
No.10 -4834.62 259.91 -128.74 78.43%
No.11 639.14 0.8 -22.72 75%
No.12 1505.64 -203.53 162.88 77.78%
Table 4.5 Body Parts Selected for Thermal Comfort
Estimation
Waist Front Upper Arm Wrist Front Wrist Back Neck
No.1 ✔ ✔
No.2 ✔ ✔
No.3 ✔ ✔
No.4 ✔ ✔
No.5 ✔ ✔
No.6 ✔ ✔
No.7 ✔ ✔
No.8 ✔ ✔
No.9 ✔ ✔
No.10 ✔ ✔
No.11 ✔ ✔
No.12 ✔ ✔
No.13 ✔ ✔
No.14 ✔ ✔
No.15 ✔ ✔
No.16 ✔ ✔
No.17 ✔ ✔
No.18 ✔ ✔
No.19 ✔ ✔
No.20 ✔ ✔
Overall 5 4 12 9 10

37
No.13 1628.77 -373.09 330.14 75.92%
No.14 1933.69 87.35 -139.27 63.64%
No.15 0.88 -57.08 52.28 100%
No.16 2308.98 -24.99 -42.88 80.39%
No.17 -449.73 116.39 -103.32 60.42%
No.18 3325.28 109.27 -198.83 85.45%
No.19 -1.37 -28.61 24.01 100%
No.20 596.7 -76.08 58.41 100%
Average 79.19%

The accuracy for each occupant is shown above. While the overall accuracy was 79.19 percent.
rate for different occupants varied widely. For No.4, No.5, No.15, No.19 and No.20, the
accuracy rate was 100 percent, where for No.9, it was only 45.10%.

The difference in thermal comfort estimation accuracy for different genders is shown in Figure
4.5. The overall average thermal comfort accuracy for our 20 occupants was 79.19 percent.
When taking gender into consideration, we noticed that the accuracy of our individual models for
male occupants was 76.56 percent, while it was 81.82 percent for female ones. However, based
on the result of one-way ANOVA, the p-value is 0.445. It was therefore not significant that there
were correlations between thermal comfort estimation and gender.

4.2.2 Logistic Regression for Thermal Sensation Feature Selection
There were two main machine learning objectives for thermal sensation feature selection:
(1) To find out whether skin temperature was related to thermal sensation prediction.
Table 4.6 Weight of Thermal Comfort Model
Figure 4.5 Thermal Comfort Accuracy Distribution by Gender

38
(2) To determine if an environmental parameter (indoor temperature) improved the accuracy of
thermal sensation prediction.

For thermal sensation, first we randomly chose two occupants (No.10 and No.20) for analysis.
Only their selected skin temperatures were chosen as input features. The weighted result of
logistic regression was as follows:

Based on the weight calculation value in Figure 4.6, we created individual thermal sensation
models for No.10 and No.20.The weight value of No.10 and No.20 for thermal sensation
estimation is as follows:

Constant
Wrist
Front
Wrist
Back
No.10 -1537.04 21.95 21.62
No.20 -723.78 10.58 11.14

Table 4.8 shows the process of thermal sensation estimation without indoor temperature, where
only two skin temperatures were applied for thermal sensation prediction. For output, “1”
represented “hot” and “0” represented “cold”.

Wrist
Front
Wrist
Back
Prediction Feedback
34.23 34.21 1 1
34.20 34.24 1 1
34.20 34.25 1 1
33.15 33.70 0 0
33.13 33.74 0 0
33.08 33.75 0 0
33.10 33.75 0 0

The prediction outputs were compared with people’s feedback. The test results showed that the
accuracy of thermal sensation estimation for No. 10 was 100 percent, where for No.20 the
accuracy was 90 percent, and it showed the obvious correlation between skin temperature and
thermal sensation.
Figure 4.6 Logistic Regression Result for Occupants No.10 and No.20
Table 4.7 Thermal Sensation Weight Value of No.10 and No.20
Table 4.8 Sample of Thermal Sensation Estimation for One Occupant (No Indoor
Temperature)

39

However, for environmental parameters, we considered using indoor temperature as an input for
thermal sensation prediction. The estimation process is shown in Table 4.9. Wrist-front
temperature, wrist-back temperature and air temperature were all considered inputs for thermal
estimation.

Wrist
Front
Wrist
Back
Air Prediction Feedback
34.21 34.16 28.22 1 1
34.22 34.15 28.27 1 1
34.23 34.15 28.30 1 1
34.22 34.17 28.31 1 1
34.23 34.21 28.31 1 1
34.20 34.24 28.30 1 1
34.20 34.25 28.29 1 1
33.15 33.70 24.89 0 0
33.13 33.74 25.09 0 0
33.08 33.75 25.26 0 0
33.10 33.75 25.49 0 0
33.03 33.75 25.74 0 0
32.46 32.24 24.53 0 0
32.36 32.31 24.69 0 0

We re-trained Occupant 20’s data set with three input features (two selected skin temperatures
and an indoor temperature), and the test accuracy was improved to 100%.

Based on this result, we regarded indoor temperature as one feature for machine learning that
helped improve the accuracy of thermal sensation estimation.

Some of the occupants only provided hot or cold feedback regarding thermal sensation, and it led
to some difficulty in building their individual sensation models, because the computer system
could not catch both hot and cold conditions of these occupants, and it could not identify the
similar situation in the future. According to the initial data collected, only Occupants 2, 3, 7, 8, 9,
10, 11, 12, 14, 17, 18 and 20 provided both hot and cold feedback during their experiments, so
we created individual thermal sensation models for them.

For thermal sensation level, the matrix form of θ was as (w0, w1, w2, w3) and the matrix form of
x was as (1, x1, x2, x3), where x1 and x2 represented the skin temperatures of two specific body
parts, and x3 represented the indoor environment temperature.

When determining valid input features, we kept using 75 percent of the data as training data and
using the rest as testing data.
Table 4.9 Sample of Thermal Sensation Estimation for One Occupant (with
Indoor Temperature)

40

The weight calculation results of the logistic regression model were as follows:

Constant
Skin
Temp. 1
Skin
Temp. 2
Indoor
Temp.
Accuracy
No.2 -0.41 -18.44 -10.48 34.97 100%
No.3 -2.04 -2.93 -10.48 34.97 100%
No.7 -1.79 -9.71 -5.07 10.78 100%
No.8 -0.74 -3.46 -3.57 10.21 100%
No.9 -0.04 -1.41 -2.22 4.95 100%
No.10 0.02 -16.9 6.46 14.96 100%
No.11 0.07 -4.38 -8.77 14.87 100%
No.12 -508.89 -59.35 42.9 42.98 84.40%
No.14 -1.03 -3.21 -5.88 11.68 100%
No.17 1.41 -2.18 -3.73 7.75 100%
No.18 0.12 -3.57 -4.61 10.75 100%
No.20 -732.61 -37.24 17.04 53.7 100%
Average 99%

In order to estimate thermal sensation level, we used the logistic-regression-defined formula θ
T
x
=w0+w1x1+w2x2+w3x3. When θ
T
x>0, the estimated result was “hot,” and when θ
T
x<0, the
estimated result was “cold.”

As noted, 75 percent of each occupant’s valid data points were used for model training, while the
remaining 25 percent were used as accuracy testing(Venkatesh, 2018). The two charts above
show the accuracy of each individual model based on the results obtained using 25 percent of
their data points.

Besides the individual model, we also considered creating a general model to achieve sensation
estimation. As noted before, however, the skin temperatures of different people were quite
different, and we could only regard indoor temperature as input.

4.2.3 Logistic Regression for ANN Feature Determination
The logistic regression results suggested that two skin temperatures were able to be used to
estimate an occupant’s thermal comfort level. They could also predict the occupant’s thermal
sensation with a high accuracy when indoor temperature was added.

As mentioned before, the ANN application for our project was as a black-box model, and we
could only determine the input features and output of the individual ANN model, so we needed
to combine output feedback for both thermal comfort and thermal sensation. The feature
Table 4.10 Weight of Thermal Sensation Model

41
selection of our combined ANN model for thermal preference was based on both thermal
comfort and thermal sensation models. According to the results of the logistic regression, the
input features of the ANN model were supposed to include two specific skin temperatures and an
indoor temperature.

As for the result of logistic regression shown in Table 4.5, the number of occurrences of each
skin part being the most-weighted part was: wrist front (12), neck (9), wrist back (8), waist front
(5) and upper arm (4). According to the result, the skin locations with the highest correlations to
thermal comfort were the wrist front, neck, and wrist back.

Because of the obscure difference between the weight of “neck temperature” and “wrist-back
temperature”, and to reduce the inconvenience of the sensor-wearing process and decrease the
measured parts, “wrist front” and “wrist back” were selected for validation.

To explore whether the combination of “wrist-front temperature” and “wrist-back temperature”
had any correlation with thermal preference, we applied WEKA (Yadav, Malik, & Chandel,
2014) for ancillary analysis.

WEKA is software that can execute simple machine-learning analysis for data classification, and
the mathematical process of its Multilayer Perceptron function is similar to the process of ANN
and logistic regression (“Weka – GUI way to learn Machine Learning,” n.d.).

As shown in Figure 4.7, WEKA’s Multilayer Perceptron function included hidden layers (labeled
in red), where it could provide extra data-processing work, just like ANN. It was therefore a
good ancillary tool for us to check correlations made with the logic of the artificial neural
network.

Figure 4.6 WEKA for Ancillary

42

In this project, we found the top three skin temperatures (wrist front, back of neck and wrist
back) for thermal comfort level prediction, and for convenience, we selected the first and last for
model generation. WEKA was adopted as an ancillary method for verifying whether there were
correlations between wrist front & back temperatures and thermal comfort levels based on the
similar mathematical analysis of ANN.

We collected wrist-front and -back temperatures and feedback for each occupant as individual
training data sets. All training data sets were entered in WEKA for testing, and the results were
as follows:

The result showed an average 88.62 percent accuracy when using “skin temperature of wrist
front & back” to estimate occupants’ thermal comfort levels, which suggested it was possible to
regard the wrist-front and wrist-back as the two body parts most appropriate for thermal
preference estimation.
Table 4.11 WEKA Simulation Results
Figure 4.7 ANN Structure in WEKA

43

We considered the possibility that some physiological factors had the potential to affect people’s
thermal comfort and thermal sensation level. We collected BMI index and gender information of
all occupants, and we applied ANOVA to see whether these two factors affected our simulation
result.

The collected BMI index and gender information is shown below:

It is worth noting, for the purposes of any potential international applications, that BMI standards
differ. According to The Hospital Authority of Hongkong, the BMI index is divided into 3
categories: Underweight (<18.5), Normal Range (18.5-23) and Overweight (≥23). Based on this
standard, we divided our occupants into three groups, as follows:

Range

Sub Total
Under
Weight
Normal
Weight
Over
Weight
Sub Total

(<18.5 ） (18.5-23） (≥23 ）
Number Male 10 0 8 2 10

Female 10 5 5 0 10

Total 20 5 13 2 20
Table 4.13 Occupant Subgroup Based on BMI and Gender

As we can see from Table 4.13, female occupants tended to have lower BMI value than male
occupants. Half of them were placed in the “Underweight” group, while two male occupants
were placed in the “Overweight” group. In order to explore the impact of BMI on thermal
comfort estimation, we counted thermal comfort estimation accuracy for different BMI groups.
The result is shown in Figure 4.8.
ID Gender
No.1 No.2 No.3 No.4 No.5 No.6 No.7 No.8 No.9 No.10 No.11 No.12 No.13 No.14 No.15 No.16 No.17 No.18 No.19 No.20 Table 4.12 BMI and Gender Information

44

Underweight occupants had higher thermal comfort estimation accuracy, while normal-weight
occupants had the lowest accuracy. Because most of our occupants (70 percent) were in the
“Normal Weight” group, the sample scale should have affected their average accuracy results.

In addition, based on all basic information, we used ANOVA for analyzing the impact of gender
on thermal comfort accuracy. The interval plot of gender with thermal comfort simulation results
is shown as Figure 4.9.

From Figure, 4.8 we can see that the thermal comfort estimation accuracy of female occupants
was higher than that of male occupants. However, for the p-value in ANOVA analysis, it seemed
that there was no obvious correlation. The p-value for “thermal comfort accuracy vs gender” was
0.446, while the correlation is significant only when p≤ 0.05, so it was not significant to say that
there was obvious correlation between thermal comfort accuracy and gender.

We also used ANOVA for analyzing the correlation of BMI and gender with thermal sensation
simulation results. The corresponding interval plot is shown as Figure 4.10.
Figure 4.9 Interval Plot for Gender vs Thermal Comfort
Figure 4.8 Thermal Comfort Accuracy for Different BMI

45

The accuracy of thermal sensation estimation was mostly 100 percent, and it was hard to get the
p-value of “thermal sensation accuracy vs BMI.”

For gender, it seemed that there was still no obvious correlation. The p-value for ANOVA for
“thermal sensation accuracy vs gender” was 0.424, still too large to ensure significance, which
meant there was no obvious correlation.

The result suggested that with logistic regression, there might be no distinct correlation between
BMI and gender on the one hand and thermal comfort and sensation simulation results on the
other. It therefore seemed to be unnecessary to consider people’s BMI and gender when building
individual logistic regression models.

4.3 Conclusion
During the whole algorithm design process, we tried to collect training data in an office-like
environment, in which 10 male and 10 female occupants participated. As mentioned before, we
planned to build both general and individual model for analysis. However, when finished data
collection, we found that there were significant differences between different people’s skin
temperatures, which meant skin temperature was not valid for building general model, and it
would be meaningless for predicting all people’s thermal preference based on one general
algorithm. For this reason, we built individual models for all participants.

ID BMI Gender Thermal Comfort Accuracy Thermal Sensation Accuracy
No.1 21.4 F 77.08% N/A
No.2 20.3 F 77.08% 100.00%
No.3 19.5 F 68.00% 100.00%
No.4 18.2 F 100.00% N/A
No.5 18.3 F 100.00% N/A
No.6 20.5 M 74.00% N/A
No.7 19.3 M 75.00% 100.00%
Figure 4.10 Interval Plot for BMI and Gender vs Thermal
Sensation

46
No.8 23.2 M 70.45% 100.00%
No.9 18.1 F 45.10% 100.00%
No.10 21.3 M 78.43% 100.00%
No.11 18.1 F 75.00% 100.00%
No.12 20.9 M 77.78% 84.40%
No.13 20.3 F 75.92% N/A
No.14 21.3 M 63.64% 100.00%
No.15 19.5 F 100.00% N/A
No.16 24.9 M 80.39% N/A
No.17 19.3 M 60.42% 100.00%
No.18 21.5 M 85.45% 100.00%
No.19 21.1 M 100.00% N/A
No.20 17.1 F 100.00% 100.00%
Overall 79.19% 98.70%

Table 4.14 shows the final information we got. We applied ANOVA for analysis. However, it
seemed that BMI and gender had no obvious correlation with either thermal comfort or thermal
sensation simulation results. For thermal comfort, the overall accuracy for women was higher
than for men. For thermal sensation, the accuracy of individual logistic regression model was up
to 100 percent except in the case of one female occupant (No.12, 84.4 percent).

For thermal comfort, we found correlations between skin temperatures and people’s thermal
comfort level. According to the weight values assigned in the logistic regression, the
temperatures of the wrist-front, back of neck, and wrist-back were the three most closely
correlated with thermal comfort estimation. To estimate thermal sensation, we used skin
temperature to build individual thermal sensation models, and compared the accuracy with &
without indoor temperature. We found the combination of two skin temperatures could estimate
people’s thermal sensation with a relatively high accuracy, while the accuracy was mostly
improved to 100% when indoor temperature was taken into consideration.

Both a general model and an individual model were considered. The individual model was better
than the general one for both feasibility and accuracy.

For ANN feature selection, according to the logistic regression result, we selected the three most
heavily-weighted skin temperatures (wrist-front, back of neck wrist-back) as factors that could
be considered for thermal comfort estimation. To improve convenience, we decided to select the
wrist-front temperature and wrist-back temperature for ANN input features, and we verified the
potential of using these to predict thermal comfort conditions with the help of WEKA.

There were possibilities for using wrist-front and -back temperatures to achieve thermal comfort
estimation. Adding indoor temperature helped us estimate thermal sensation precisely. We
Table 4.14 Conclusion for Thermal Comfort/Sensation
Accuracy

47
needed to select both wrist-front and -back temperatures and indoor temperature as input features
for black box ANN model in order to achieve thermal preference identification.

48
5 The Application and Validation by Artificial Neural Network
As mentioned in Chapter 3, the artificial neural network (ANN) is the most popular machine-
learning model. The distinguishing feature of ANN is that it has multiple hidden layers for
calculation, meaning it can create non-linear mathematical functions via extra steps for data
processing in these hidden layers.

Because of this, ANN is a powerful tool for simulating complex circumstances. We already used
logistic regression to select input features, and we planned to apply the ANN model from the
LabVIEW Toolkit to build individual thermal preference models, and then validate our
individual models in different practical situations.

5.1 Validation for Real Office Application
Based on the machine learning process finished before, wrist fronts and backs showed good
potential for estimating people’s thermal comfort levels. In addition, the machine learning result
showed that environmental temperature was a crucial factor for thermal sensation estimation. In
order to validate our assumptions, we planned to determine:

(1) Whether the ANN model could use specific skin temperatures and environment temperature
to estimate people’s thermal preference in real office environments accurately.

(2) Whether the ANN model could achieve HV AC automation and create a sustained acceptable
environment for individual occupants.

According to previous analysis of the results and the objectives we planned to get, our validation
process was as follows: First, we used an ANN toolkit to generate several occupants’ individual
models based on their previous experimental data, and then we asked the selected occupants to a
real office, where we used their corresponding individual models to test the accuracy of the
thermal preference estimation. Lastly, we asked the same occupants to our laboratory, where we
connected their individual thermal estimation model and HVAC devices with power relay units
to test the control effect.

Firstly, in order to validate our model in a real environment, we tested it in a real office located
in downtown Los Angeles.

49

Figure 5.1 shows the experiment process in a real office. We adopted a LabVIEW Machine
Learning ANN Toolkit to create six individual models (three males and three females), where
their wrist-front temperatures, wrist-back temperature and the indoor temperature were chosen as
input features. We asked the six occupants to the office for validation, and during the experiment
we recorded their wrist-front skin temperature, wrist-back skin temperature and the indoor
environment temperature. We also asked them for their thermal preference feedback every five
minutes. .

One recording sample of occupant’s thermal preference feedback is shown in Table 5.1. Here,
the order from “-3” to “+3” represented the range from “very cold” to “very hot,” respectively, in
thermal sensation, and also represented “very uncomfortable” to “very comfortable,”
respectively, in thermal comfort (the same as the definition in the questionnaire from the
previous laboratory experiment).

The workflow of the real office experiment was as follows:

Table 5.1 shows one individual’s survey results as recorded in the real office. Feedback was
recorded at five-minute intervals. We asked occupants to provide their estimations of current
Figure 5.2 Real Office Experiment Workflow
Figure 5.1 Test Scenario in Real Office

50
thermal comfort and thermal sensation levels, and their preference about whether they wanted
additional heating or cooling to optimize current thermal conditions.

In addition to occupants’ feedback, we also collected their wrist-front and -back temperatures
and the indoor temperature in the office. After the measurement work in the real office, we chose
data-point sets for all feedback times, and then we input these data-point sets into their
corresponding individual ANN model to test their accuracy. Then we compared the compatibility
between ANN results and occupants’ feedback in order to verify the accuracy of our ANN
model.

5.2 Validation for HV AC Control Automation
In order to achieve HVAC automation, we designed our control experiments in our original
laboratory.

We kept three temperature sensors for the control experiment. Two of them were applied for
measuring temperature of wrist-front and wrist-back, and the other one was used for collecting
indoor environment temperature.

The interface of our automation system was configured as follows. The system read real-time
data from sensors, and ANN model automatically generated an occupant’s predicted thermal
preference. “-1” meant “trigger heater,” “0” meant “acceptable environment” and “1” meant
“trigger cooler.”

Table 5.1 Sample of Thermal Preference Feedback in Real
Office

51

The label in the orange box showed the prediction result, and all control orders were based on the
result (“-1” for “trigger heater”; “0” for “acceptable environment”; “1” for “trigger cooler”). The
control system had three parts: a manual control, an automatic control, and a manual validation.
The blue box contained the manual control, where we could manually input the values “-1,” “0”
and “1” for triggering heater, stopping all devices, and triggering the air conditioner,
respectively. The green box contained the automatic control, and the labels in the green box
showed real-time data collected by our three Vernier temperature sensors. The ANN model made
thermal preference predictions based on these three temperature sensors. The black box
contained the validation component, where we could input our collected data from real office
experiment. ANN generated the output according to the input data.

In Figure 5.3, the classification result is shown at the bottom of this interface, where the different
data points were classified into three groups, and all points were labeled with different colors
(orange, red and green). The data points indicated that people were under hot, neutral, and cold
conditions, respectively. The “Classification accuracy: 100%” message in this interface
illustrated that 100 percent of the test data matched the ANN training model, and our automation
experiment made occupants’ real-time thermal estimations based on their corresponding
individual training models.
Figure 5.3 HVAC Real Time Control Interface

52

In order to achieve HVAC automation, we used a DAQ for signal transformation. The blue box
contained the signal input part, a computer connection interface, through which it received
digital signals (“-1”, “0” and “1”) from the computer. The red box contained the signal output
part, which connected power relay hardware, and sent signals to trigger air conditioners (when
receiving the “1” signal) or heaters (when the receiving “-1” signal).

The work flow of HVAC automation is shown in Figure 5.5. We asked six occupants (three male
occupants and three female occupants) who had participated the previous office test in
downtown Los Angeles to take part in this experiment, and we collected their feedback about
their current thermal preferences every five minutes during the experiment.

5.3 Validation Results Analysis
5.3.1 Results of Real Office Application
In the resulting statistical process, we used M (for male occupant) and F (for female occupant) to
identify the occupant’s gender, and we marked our occupants as M1, M2, M3, F1, F2, F3.

We input collected data into the ANN model for validation after the experiment. However, the
ANN model failed to estimate the thermal preference for F3 correctly. It always estimated her
thermal preference as “hot” and “need to trigger heater” at the same time. We thought the small
scale of F3’s training data might be the main reason for the faulty estimation. This required
optimization work for her individual ANN model in our later work.
Figure 5.4 DAQ for Control Automation
Figure 5.5 Workflow of HVAC Automation

53

We considered that there were some sensitive physical variances that we had not thought about
during experiments, such as menstrual status. That would obvious impact female occupants’
physiological and psychological conditions, which might change their thermal preferences even
under the same surrounding thermal conditions.

We decided to use mean absolute error (MAE) to identify the deviance degree of our model,
where:

MAE represents the average deviation of the test model. The larger the MAE result is, the more
deviant our model should be.

We used -1, 0 and 1 to represent people’s thermal preference levels as cold, neutral, and hot,
respectively. We compared occupants’ feedback with their models’ estimation results, and then
we executed the MAE test.

The results of the comparison between feedbacks and ANN results for other occupants are shown
below.

Skin
Temp.1 Air
Skin
Temp.2 Feedback Prediction
31.78 23.45 32.55 0 0
32.16 23.54 31.76 0 -1
32.28 23.65 32.26 0 0
32.78 23.67 32.81 0 0
32.77 23.37 32.58 0 0
33.01 23.60 33.57 -1 0
33.21 23.72 33.18 -1 0
33.09 23.59 33.52 0 0
33.03 23.85 33.31 0 0
32.68 23.94 33.38 0 0
32.83 23.91 33.71 0 0
32.62 23.96 33.54 0 0
32.28 24.20 33.39 0 0
32.31 23.78 33.42 0 0
32.04 23.89 33.06 0 0
32.19 24.00 33.20 0 0
32.01 23.81 33.50 0 0
31.77 23.68 33.17 0 0
31.39 23.61 32.59 0 0
31.16 23.47 33.04 0 -1
30.80 23.46 32.46 0 -1

54
30.60 23.50 32.33 0 -1
30.40 23.83 32.50 0 -1
30.16 23.72 32.37 -1 -1
Accuracy: 70.8%

Table 5.2 shows the temperature and survey records and the prediction result of M1’s individual
ANN model. For occupant M1:
𝑀𝐴𝐸 =
'6'6'6'6'6'6'
78
=0.29

Skin
Temp.1 Air
Skin
Temp.2 Feedback Prediction
34.00 24.01 34.32 -1 -1
34.38 23.95 34.51 -1 -1
34.40 23.78 34.42 -1 -1
34.76 23.71 34.54 -1 -1
34.88 23.84 34.95 -1 -1
34.96 23.77 35.30 -1 -1
34.95 23.63 35.00 -1 -1
35.17 23.67 35.20 -1 -1
34.97 23.67 35.44 -1 -1
34.96 23.76 35.42 -1 -1
35.08 23.60 35.12 -1 -1
35.11 23.61 34.70 -1 -1
34.94 23.65 34.50 -1 -1
34.80 23.54 34.59 -1 -1
34.78 23.54 34.71 -1 -1
34.77 23.57 34.37 -1 -1
34.67 23.72 34.45 -1 -1
34.78 23.99 34.90 -1 -1
34.77 23.97 34.56 -1 -1
34.78 23.79 34.53 -1 -1
34.70 23.66 34.35 -1 -1
34.52 23.61 34.16 -1 -1
34.57 23.75 33.83 -1 -1
34.57 23.80 34.00 -1 -1
Accuracy: 100%

Table 5.3 shows the temperature and survey records and the prediction result of M2’s individual
ANN model. For occupant M2:
𝑀𝐴𝐸 =
9
78
=0
Table 5.2 Real Office Simulation Result of M1
Table 5.3 Real Office Simulation Result of M2

55

Skin
Temp.1 Air
Skin
Temp.2 Feedback Prediction
34.60 23.99 32.85
0 0
34.43 23.94 32.55
0 0
34.39 23.92 34.03
0 0
34.28 24.01 33.00
0 0
34.78 23.93 34.65
0 0
35.03 23.94 34.59
0 0
34.94 24.12 34.66
0 0
34.83 24.05 34.46
0 0
34.54 24.03 34.10
0 0
34.49 23.96 34.25
0 0
34.24 24.00 34.18
0 0
33.99 24.10 34.02
0 0
33.27 23.98 33.82
0 0
33.19 23.96 33.48
0 0
32.63 24.00 33.50
0 0
28.74 24.01 33.39
0 0
34.26 24.89 33.89
0 0
34.14 24.68 33.85
0 0
33.99 24.91 33.97
0 0
33.78 24.84 32.80
0 0
33.43 24.99 33.75
0 0
34.08 24.95 33.60
0 0
34.47 24.21 33.75
0 0
33.84 23.55 33.76
-1 0
Accuracy: 95.8%

Table 5.4 shows the temperature and survey records and the prediction result of M3’s individual
ANN model. For occupant M3:
𝑀𝐴𝐸 =
'
78
=0.04

Skin
Temp.1 Air
Skin
Temp.2 Feedback Prediction
31.02 22.38 31.29 -1 -1
30.86 22.70 31.77 -1 0
30.74 22.70 32.22 -1 0
31.49 22.89 32.60 -1 0
31.41 22.78 32.43 -1 0
31.44 23.04 32.62 -1 0
Table 5.4 Real Office Simulation Result of M3

56
31.43 22.97 32.62 -1 0
31.43 23.08 32.70 -1 0
31.37 23.07 32.63 -1 0
31.47 23.23 32.79 -1 0
31.47 23.40 32.59 -1 -1
31.47 23.32 32.59 -1 -1
31.48 23.31 32.82 -1 0
31.48 23.72 32.44 -1 -1
31.66 23.39 32.83 -1 0
31.67 23.35 32.79 -1 -1
31.86 23.56 33.00 -1 -1
31.83 23.52 33.00 -1 0
31.86 23.56 33.00 -1 -1
32.02 23.68 33.16 -1 -1
31.86 23.86 32.96 -1 -1
31.86 23.71 32.44 -1 -1
31.86 23.73 32.16 -1 -1
31.91 23.68 32.05 -1 0
Accuracy: 45.8%

Table 5.5 shows the temperature and survey records and the prediction result of F1’s individual
ANN model. For occupant F1:
𝑀𝐴𝐸 =
'6'6'6'6'6'6'6'6'6'6'6'6'
78
=0.54

Skin
Temp.1 Air
Skin
Temp.2 Feedback Prediction
32.13 23.86 31.23 -1 0
32.43 23.76 31.33 -1 0
32.21 23.85 31.03 -1 0
32.43 23.93 31.17 -1 0
32.15 23.91 31.10 -1 0
32.46 23.78 31.42 -1 0
32.56 23.84 31.73 -1 0
32.38 23.74 31.51 -1 0
31.68 23.69 30.63 0 0
32.07 23.85 31.16 -1 0
32.46 23.82 31.29 -1 0
32.46 23.82 31.48 -1 0
32.45 23.72 31.25 -1 0
32.32 23.55 30.97 0 0
32.63 23.62 30.56 0 0
Table 5.5 Real Office Simulation Result of F1

57
32.32 23.55 31.00 0 0
32.13 23.61 30.70 0 0
31.76 23.62 30.16 0 -1
31.39 23.73 29.72 0 -1
31.25 23.70 29.94 0 -1
30.85 23.67 29.86 0 -1
31.59 23.64 30.05 -1 -1
31.20 23.67 29.69 0 0
30.92 23.73 29.73 -1 -1
Accuracy: 33.3%

Table 5.6 shows the temperature and survey records and the prediction result of F2’s individual
ANN model. For occupant F2:
𝑀𝐴𝐸 =
'6'6'6'6'6'6'6'6'6'6'6'6'6'6'6'
78
=0.67

Skin
Temp.1
Air
Skin
Temp.2
Feedback Prediction
28.40 23.93 31.43 0 1
28.74 24.06 31.85 0 1
30.05 24.06 32.58 0 1
31.19 24.00 32.55 0 1
30.71 23.88 32.87 0 1
31.15 23.90 33.55 0 1
31.78 23.96 33.94 0 1
31.30 24.12 33.03 0 1
30.92 23.97 33.00 0 1
30.78 23.97 33.39 0 1
30.59 23.94 33.56 0 1
30.74 23.90 33.06 0 1
30.69 24.02 32.99 0 1
30.55 23.97 32.97 0 1
30.78 23.92 32.85 0 1
30.51 24.01 32.85 0 1
30.51 24.09 32.99 0 1
30.73 24.06 33.27 0 1
31.65 23.90 33.88 0 1
31.77 23.93 33.69 0 1
31.28 24.02 33.09 0 1
31.02 24.21 33.19 0 1
Table 5.6 Real Office Simulation Result of F2

58
Accuracy: 0%
Table 5.7 Real Office Simulation Result of F3
For F3, the prediction result was always “1.”

Table 5.7 shows the temperature and survey records and the prediction result of F2’s individual
ANN model. For occupant F3:
𝑀𝐴𝐸 =
'6'6'6'6'6'6'6'6'6'6'6'6'6'6'6'6'6'6'6'6'6'6'6'
78
=1

The simulation results of M1, M2, M3, F1, F2, F3, respectively are shown above. The accuracy
rates achieved are shown in Table 5.7. We can see that the results showed significant accuracy
differences between male and female occupants.

ID Accuracy
M1 70.80%
M2 100%
M3 95.80%
F1 45.80%
F2 33.30%
F3 0%

Owing to the obvious problem for F3’s model, we used ANOVA to analyze the correlation
between gender and validation accuracy for M1, M2, M3, F1 and F2. The interval plot is shown
below:

From Figure 5.6, we can see there were obvious differences in validation accuracy between male
and female occupants. The average accuracy rate for male occupants was 88.87 percent, while
for female occupants it was 39.55 percent. The p-value of the one-way ANOVA test was 0.03,
Table 5.8 Real Office Validation Accuracy
Figure 5.6 Interval Plots for "Gender" and "Validation Accuracy"

59
which meant it was significant that there were correlations between gender and real-office test
accuracy.

For M1, M2, M3, F1 and F2, the MAE results were 0.29, 0, 0.04, 0.54, 0.67 respectively, which
revealed that our individual ANN model performed more reliably with male occupants than
female ones.

ANOVA was also applied in analyzing correlations between gender and MAE. The interval plots
are shown below:

From the interval plots, we can see that the MAE of male occupants was obviously less than the
one of female occupants. According to ANOVA result, the p-value of “gender & MAE” was
0.03, which meant it was significant that in our project, MAE was related to gender. The
deviation of the simulation results for male occupants was obviously less than the one for female
ones.

The individual ANN model seemed to predict better for male occupants. We thought the reason
might be that female occupants were more sensitive to changes in their thermal preferences. In
the indoor environment of real office, the temperature was relatively stable. Under these
circumstances, male occupants provided more consistent thermal preference feedback during the
experiment time, while female occupants changed their opinion about their surrounding thermal
condition more frequently. It seemed that female occupants’ thermal preferences were more
subjective in this project.

This finding revealed that it was hard for our ANN model to catch micro-changes in people’s
thermal preferences. We thought that the lack of sufficient training data was what led to our
model’s difficulty in identifying the small difference in people’s thermal preferences with a
small change in the surrounding indoor temperature.

The real office validation showed the difference in people’s thermal preference change rate
based on gender. It also revealed the potential defect in our small set of training data. For these
reasons, we considered finding methods to enlarge the scale of the data. In addition, we had to
consider how physical factors, such as female occupants’ different menstrual statuses, might also
impact their thermal preferences.

Figure 5.7 Interval Plots for "Gender" and "MAE"

60
5.3.2 Results of Real-Time Control
For the real-time control, we recorded the ratio that people felt was acceptable for their
surrounding thermal environment. Their feedback records are shown below:

Table 5.8 shows the survey results for male occupants during the HVAC automation experiment.
For M3, the controlled environment was acceptable for the whole time, while M1 and M2 only 1
provided “cold” feedback once in their whole nine points of survey data.

Table 5.9 shows the survey results for female occupants during the HVAC automation
experiment. However, when compared with male groups, the controlled accuracy for female
occupants was substantially decreased. F2 only provided one “cold” feedback in her nine data
points. However, there were four times F1 provided “cold” feedback for her current
environment.

ID Feedback
0
0
0
0
0
0
-1
0
0
Accuracy 88.89%
0
0
0
0
0
-1
0
0
0
Accuracy 88.89%
0
0
0
0
0
0
0
0
0
Accuracy 100%
M1
M2
M3
Table 5.9 HVAC Automation Records for Male Occupants

61

From the records, it is clear that the individual ANN models performed well for M1, M2, M3,
and F2. However, the accuracy of F1’s model was only 55.56 percent.

Figure 5.8 shows the variations in the nine points of feedback provided by of M1, M2, M3, F1,
F2. It demonstrates that the individual ANN model perfectly optimized the indoor environment
for M3. The controlled indoor environment also seemed to be acceptable for M1, M2 and F1.
However, the ANN did not create a stable, satisfactory indoor environment for F2, and her
thermal satisfaction seemed to fluctuate during her experiment time.

Participants’ overall feedback records are shown below:
-1
0
1
1 2 3 4 5 6 7 8 9
M1
-1
0
1
1 2 3 4 5 6 7 8 9
M2
-1
0
1
1 2 3 4 5 6 7 8 9
M3
-1
0
1
1 2 3 4 5 6 7 8 9
F1
-1
0
1
1 2 3 4 5 6 7 8 9
F2
Figure 5.8 Feedback Variation for Automation Experiment
Table 5.10 HVAC Automation Records for Female Occupants
ID Feedback
-1
0
-1
0
-1
0
-1
0
0
Accuracy 55.56%
0
-1
0
0
0
0
0
0
0
Accuracy 88.89%
F1
F2

62

ID Acceptable Time
M1 88.89%
M2 88.89%
M3 100%
F1 55.56%
F2 88.89%
F3 0%

For F3, the ANN estimation result was always “1”. However, based on their feedback, our
models successfully controlled the indoor space as an “acceptable” environment at the most of
time for all male occupants and even F2. In the case of F1, it seemed to be difficult for her
individual model to provide a stable comfortable thermal environment most of the time.

The temperature variation tendencies for all six participants are shown below:

Table 5.11 Ratio of Acceptable Time during Experiment
Figure 5.9 Temperature Variation Tendencies for All Occupants
15
17
19
21
23
25
27
29
31
33
35
1
28
55
82
109
136
163
190
217
244
271
298
325
352
379
406
433
460
487
514
541
568
595
622
649
676
703
730
757
TemperatureVariationforM1
15
17
19
21
23
25
27
29
31
33
35
1
19
37
55
73
91
109
127
145
163
181
199
217
235
253
271
289
307
325
343
361
379
397
415
433
451
469
487
505
523
541
TemperatureVariationforM2
15
17
19
21
23
25
27
29
31
33
35
1
23
45
67
89
111
133
155
177
199
221
243
265
287
309
331
353
375
397
419
441
463
485
507
529
551
573
595
617
temperature Variation for M3
15
17
19
21
23
25
27
29
31
33
35
1
21
41
61
81
101
121
141
161
181
201
221
241
261
281
301
321
341
361
381
401
421
441
461
481
501
521
541
Temperature Variation for F1
15
17
19
21
23
25
27
29
31
33
35
1
19
37
55
73
91
109
127
145
163
181
199
217
235
253
271
289
307
325
343
361
379
397
415
433
451
469
487
505
TemperatureVariationforF2
15
17
19
21
23
25
27
29
31
33
35
1
10
19
28
37
46
55
64
73
82
91
100
109
118
127
136
145
154
163
172
181
190
199
208
217
226
235
Temperature Variation for F3 (before model
updating)

63
We could see the indoor temperature fluctuated in a small range. However, the individual ANN
model for F3 always estimated her thermal preference as “hot, triggering air conditioner” even
when she felt cold with very low air temperature, and there remained some problems with it that
we needed to solve in our later work.

The temperature variation tendencies suggest the control strategy of ANN model created
relatively stable thermal environments for all occupants except F3. Due to the different initial
indoor temperatures, and because there was time for the HVAC to adjust indoor temperature to a
relatively stable state, we recorded the temperature range of each experiment for the bottom 80
percent of the time. The controlled temperature ranges of M1, M2, M3, F1, F2 are shown in the
following table.

Mark ID
Max Temp.
（℃）
Min Temp.（℃）
Ave Temp.
（℃）
Range
（℃）
M1 28.379 24.137 26.461 4.242
M2 30.288 26.462 28.801 3.826
M3 26.146 23.006 24.85 3.14
F1 29.958 24.699 27.283 5.259
F2 28.194 22.712 25.484 5.482

We used ANOVA for analysis. The interval plot is shown below:

It seemed there was no obvious difference for controlled average temperature and genders.
However, the temperature variation range of men was obviously less than the variation range of
women.

The p-value of “gender vs average temperature” was 0.851. However, the p value of “gender vs
temperature variation range” was 0.031. This suggested that it was significant that the
temperature variation range was related to gender.

The result showed that the controlled temperature ranges of male occupants were relatively
smaller than those of female ones, which also suggested that individual ANN models performed
better for male occupants.
Table 5.12 Controlled Temperature Range
Figure 5.10 ANOVA for "Temperature Range & Average Temperature" and "Gender"

64

In addition, there were some differences between occupants’ thermal preferences and their
individual model’s predictions. Sometimes people already provided “acceptable” for their
surrounding thermal condition, while their model generated “-1” and “1”, which meant the
model still ran the heater and cooler during these times, obscuring the different “comfort zones”
for occupants’ thermal preferences.

During the data preprocessing period, people sometimes provided “neutral” feedback for
surrounding thermal environment, while they also preferred to trigger the heater or cooler at that
time. Under this circumstance, there might be only slight uncomfortable feelings for occupants
and the surrounding thermal condition might be acceptable for them. In order to improve the
precision of control automation, we regarded people’s thermal preference as “uncomfortable” at
these times, and this process narrowed the comfort zone of occupants.

Another reason could be the lack of adequate training data.

At the beginning of our project, we considered using both general model and an individual
model, and we finished experiments for 20 people. Due to the limitation of our experiment time,
however, we could only collect each person’s training data for two hours, and that led to the
decline of machine learning’s accuracy.

Due to the poor estimation result for F3’s model, we also thought that the lack of enough training
data should be another significant factor that limited the accuracy of her ANN model, and we
considered it a major challenge for us to solve for our later work.

In addition, due to the huge difference of our models’ performances in relation to female
occupants, we considered that there might be physical variables that affected female occupants’
thermal preferences.

5.4 Optimization for ANN Application
5.4.1 Renewing Training Data for ANN Simulation
As shown before, the estimation result for occupant F3 was relatively poor, and we considered
the problem of machine learning overfitting due to the lack of training data. In order to optimize
the simulation process, we added a renew function to our real-time control system.

65

The interface of our real time control system is shown in Figure 5.11. The green box contained
the renew function we made for optimization. During the real-time control process, occupants
could provide their feedback by clicking the one of three buttons (COLD, HOT, NORMAL)
below, and the system would record the feedback, current skin temperatures, and indoor
temperatures in the training data file.

This process aimed to enlarge the amount of valid training data for machine learning. The next
time to begin real-time control was initiated, our system read the renewed training data, and
generated a renewed ANN model for real-time estimation.

In order to estimate whether this work helped improve our model, we asked F3 to provide some
new feedback during her experiment, and then we restarted the real-time control system and
input the renewed training data. This time the real-time control system could identify the “cold”
and “neutral” sensation for F3, and it controlled the indoor thermal environment, achieving a
stable state.

Figure 5.11 Renew Setting for HVAC Real-Time Control System
Figure 5.12 Temperature Variation for F3’s Updated Model
15
17
19
21
23
25
27
29
31
33
35
1
10
19
28
37
46
55
64
73
82
91
100
109
118
127
136
145
154
163
172
181
190
199
208
217
226
235
244
Temperature Variation for F3 (after model
updating)

66
Before Model
Updating
-1
-1
-1
-1
Accuracy 0%
After Model
Updating
0
0
0
0
0
Accuracy 100%
Table 5.13 Real-Time Control Experiment Result for F3

According to this result, we considered that the small scale of training data was an obvious
problem limiting the precision of ANN, and that additional tests and experiments with a larger
sample (i.e., human subjects) size should be considered for enhancing the validity of the
machine-learning process. Even given variations in physique, occupants could provide current
thermal preferences on time, and those helped the computer system correct the ANN models for
better estimation.

5.4.2 New Feature Selection for ANN
Besides enlarging the training data scale, another way to improve ANN accuracy was to increase
input features for the machine-learning process. Accordingly, we added two skin temperatures
and an indoor temperature as machine-learning input features. In order to improve prediction
accuracy, more input features were needed. Here, we took psychological factors as consideration.

During the training data collection experiment, our questionnaire contained a question to record
occupants’ subjective stress levels, which required the occupants to indicate how they felt about
their current psychological condition (very stressful, stressful, neutral, relaxed, or very relaxed).

In order to find out whether there were correlations between psychological factors and thermal
preferences, we compared occupants’ psychological and thermal feedback.

Figure 5.13 shows the statistical result of the comparison between thermal sensation and stress
level feedback.

67

In order to make them accord with our ANN output form, we divided occupants’ thermal
sensation levels into a three-part scale and then made statistics. The comparison was made as
follows:

Stressful Neutral Relaxed
Cold 31.6% 23.6% 35.1%
Hot 50.7% 24.0% 20.8%
Neutral 17.7% 52.4% 44.1%

The stress-level results showed a significant difference between people’s thermal sensation
according to their different stress levels. “Stressful” emotion was more likely to be generated
under a “hot” condition. For cold sensation, there was no obvious difference between occupants
with “stressful” or “relaxed” stress level. However, people were less likely to feel hot or cold
when their stress level was “neutral” or “relaxed.”

Table 5.14 Correlation Between Stress Level and Thermal Sensation
Figure 5.13 Correlation Between Stress Level and Thermal Sensation
6%
29%
44%
17%
4%
Thermal Sensation Distribution(Relaxed)
Very Cold Cold Neutral Hot Very Hot
10%
21%
18%
45%
6%
Thermal Sensation Distribution(Stressful)
Very Cold Cold Neutral Hot Very Hot
3%
20%
53%
21%
3%
Thermal Sensation Distribution(Neutral)
Very Cold Cold Neutral Hot Very Hot

68

In order to make them accord with our ANN output form, we also divided occupants’ thermal
comfort levels into a three-point scale and then made statistics. The comparison is provided
below:

Stressful Neutral Relaxed
Comfortable 15.6% 0.0% 46.7%
Neutral 58.2% 60.8% 21.1%
Uncomfortable 40.2% 39.2% 32.2%

In thermal comfort predictions, the correlation between stress level and thermal comfort level
seemed to be more positive.

From “stressful” to “relaxed,” the rate of the “uncomfortable” condition \ gradually decreased,
from 40.2 percent to 32.2 percent. In addition, there were only 15.6 percent of people who felt
comfortable under a “stressful” situation, while the rate for the “relaxed” situation climbed to
46.7 percent. The survey result showed that an increase in stress level corresponded to a decrease
in people’s satisfaction regarding their surrounding thermal condition. Therefore, there was
potential for stress levels to be included as one of the input features for machine learning in order
to improve the accuracy of thermal preference estimations.

In order to find out more potential psychological features for machine learning, we took
electrode activity (EDA) and heart rate variability (HRV) into consideration.
Table 5.15 Correlation Between Stress Level and Thermal Comfort
Figure 5.14 Correlation Between Stress Level and
Thermal Comfort
32%
21%
34%
13%
Thermal Comfort Distribution(Relaxed)
Very Uncomfortable Uncomfortable Neutral Comfortable Very Comfortable
4%
35%
61%
Thermal Comfort Distribution(Neutral)
Very Uncomfortable Uncomfortable Neutral Comfortable Very Comfortable
40%
58%
2%
Thermal Comfort Distribution(stressful)
Very Uncomfortable Uncomfortable Neutral Comfortable Very Comfortable

69

Firstly, in accordance with the BIOPAC EDA measurement explanation, the change in the EDA
wave was taken to represent the outside stimulation of people. In order to quantify the meaning
of an EDA measurement, we looked at every five-minute-long period before each measured
point. We counted the number of wave peaks, and then we calculated the average peak number
of occupants at different stress levels (as recorded in experiment questionnaires).

Figure 5.15 shows the EDA records of one participant with different stress-level feedback. The
line shows the variance of the conductivity (microsiemens) of skin, which is related to the gland
variance. It is a signal representing people’s reactions to outside stimulations. Sweat glands
change when people receive stimuli, which increases the conductivity of skin in a short time.
Each such change appears as one wave peak in EDA records.

In order to explore the correlation between EDA and stress level, we considered that the number
of wave peaks representing the outside stimulation frequency, which had potential to affect
people’s stress levels. We selected five-minute-long period before each measurement point as a
measurement period. We counted the number of wave peaks in each measurement period before
analyzing all measurement points for six participants (three males and three females).

Figure 5.15 EDA Wave Change Records

70
Table 5.16 shows the EDA statistic results. In each measurement period, we collected the
number of EDA wave peaks and compared them with people’s stress-level survey results. The
label “N/A” in Table 5.16 means that for a given occupant, there was no such stress level
feedback for that column during his/her experiment. The columns indicate five different stress
levels, and a number in Table 5.16 means the average EDA peak number for one measurement
period under a specific stress level. The result showed that the frequency of EDA wave peaks
differed as occupants’ stress levels changed.

Very
Relaxed
Relaxed Neutral Stressful
Very
Stressful
Mean
P1 N/A N/A 15.25 7 N/A 13.6
P2 15.33 12 N/A N/A N/A 14
P3 20 12 12.5 4 N/A 12.17
P4 12.25 5 N/A N/A N/A 10.8
P5 N/A N/A 20.25 16 N/A 19.4
P6 N/A 6 3 N/A N/A 6
Average 15.86 8.75 12.75 9 N/A 12.66

The mean peak numbers of different people ranged from 6 to19.4, indicating they had different
EDA fluctuation levels.

In individual analyses, it could be inferred that the frequency of wave peaks decreased as stress
levels increased. In our opinion, the reason might be that more wave peaks meant there were
currently more outside stimuli for occupants (air flow, temperature change and so on). These
stimuli might have distracted the occupants, and probably decreased their stress levels.

Table 5.16 EDA Records Statistics
Figure 5.16 Average Peak Number Variation in EDA & Stress Level

71

Figure 5.16 shows the whole average peak number for different stress levels. The results indicate
a tendency toward decrease as stress levels rose, except for the “relaxed” group. We considered
that differences in individual physiological conditions could cause the deviation in EDA data
changes.

Another factor we considered was HRV. As mentioned in Chapter 3, HRV focused on the
variance of R-R interval (the time interval between adjacent heart beats), which had the potential
to illustrate the current stress levels of occupants. We selected the data for every five-minute-
long period before each measured point for HRV analysis.

Table 5.17 outlines the process of getting R-R interval data. The heart-rate data in the left
column represented people’s heartbeats in number of beats per minute, while the R-R interval
(ms) equaled the quotient between the total time (1min, 6000 ms) and total number of heart
beats. We saved our heart-rate data and then transformed it into R-R intervals in txt form.

Heart Rate
(bpm)
R-R
Interval
（ms）
91 659.34
82 731.71
81 740.74
80 750.00
74 810.81
76 789.47
86 697.67
83 722.89
79 759.49
85 705.88
76 789.47
83 722.89
84 714.29
83 722.89
87 689.66

HRV focuses on two nervous systems: the parasympathetic nervous system (PNS) and the
sympathetic nervous system (SNS). The PNS controls homoeostasis and the body at rest, and it
is responsible for the body’s “rest and digest” function. The SNS, on the other hand, controls the
body’s responses to a perceived threat, and it is responsible for the “fight or flight” response. In
clinical research, sympathetic activity tends to increase HR and decrease HRV, while
Table 5.17 Sample of R-R Interval Generating Process for One Occupant

72
parasympathetic tends to decrease HR and increase HRV, so PNS and SNS could be regarded as
factors through which to analyze occupants’ HRV variance.

Moreover, as mentioned before, HRV is recorded as a waveform, and HF and LF represent the
power of the high-frequency peak (0.15-0.40 Hz) and the low-frequency peak (0.04-0.15 Hz)
respectively. Previous research revealed that there is potential for the LF/HF ratio to suggest
people’s stress levels, making it a critical factor in HRV analysis.

In addition, due to the complexity of HRV, it is hard to fully describe the variation of HRV with
only a linear model. In order to avoid the information loss, Kubios introduces Poincare Plot, a
graphical presentation that describes consecutive R-R intervals to capture the standard non-linear
variance of HRV.

Figure 5.17 shows the Poincare Plot report generated by Kubios. In this report, the Poincare Plot
is presented in ellipse form to describe the non-linear variance of consecutive R-R intervals. The
length and width of the ellipse is determined by SD1 (ms). SD1 can be regarded as the ratio of
short-term variability, while the rest variability can be regarded as long-term variability. SD1 can
therefore be applied to represent the non-linear variability of HRV (Tarvainen, Niskanen,
Lipponen, Ranta-aho, & Karjalainen, 2013).

Kubios generated the HRV analysis reports based on the R-R interval data we input. Figure 5.18
shows the HRV report produced in Kubios, which included important HRV-related factors (PNS
index, SNS index, LF/HF and SD1).
Figure 5.17 Poincaré plot for HRV
Analysis

73

The Kubios HRV report showed four potential factors that might be correlated with people’s
stress levels (SD1, PNS index, SNS index, and LF/HF). The result of the Kubios HRV analysis is
shown below.

Sequence SD1 PNS index SNS index LF/HF
Survey
Results
Stress
Level
1 50.10% 3.18 0.01 0.771 3 Neutral
2 44% 1.36 0.27 1.737 3 Neutral
3 52.6 0.56 0.25 0.67 4 Nervous
4 53% 2.55 -0.56 1.063 3 Neutral
5 46.30% 1.03 -0.59 0.659 1 Relax
6 41.50% 0.64 -0.28 2.115 1 Relax
7 42.30% 0.4 0.09 2.174 4 Nervous
8 45.70% 1.13 -0.45 1.184 3 Neutral
9 42.90% 1.05 -0.41 0.626 3 Neutral
10 52.70% 3.25 -1.1 0.307 4 Nervous
11 53.40% 2.68 0.86 0.213 1 Relax
12 50.90% 2.44 1.02 0.188 1 Relax
13 48.80% 2.36 0.29 0.396 1 Relax
14 53.70% 1.88 0.48 0.326 2 Relax
15 44.70% 2.16 0.91 0.251 2 Relax
16 51.40% 1.79 0.19 0.284 1 Relax
Figure 5.18 Kubios HRV Report

74
17 54.40% 0.37 1.31 0.757 2 Relax
18 47.60% 0.34 1.49 0.656 2 Relax
19 47.50% 0.16 1.6 0.174 1 Relax
20 59.60% 1.41 0.89 3.463 2 Relax
21 56% 1.23 0.66 0.229 2 Relax
22 51.90% 0.94 0.61 0.264 3 Neutral
23 50.80% -0.1 1.06 1.023 3 Neutral
24 47.90% 0.8 0.38 0.965 4 Nervous
25 46.40% 0.16 0.98 0.635 2 Relax
26 51% 0.59 0.94 0.647 2 Relax
27 47.50% 0.45 0.79 0.315 1 Relax
28 49.30% 1.73 0.25 0.541 3 Neutral
29 52.90% 0.67 0.38 3.282 4 Nervous
30 45.70% 0.82 -0.08 1.858 3 Neutral
31 43.60% 0.32 0.17 1.744 3 Neutral
32 49.60% 1.42 -0.23 1.354 2 Relax
33 42.40% 0.38 0.26 0.559 3 Neutral
34 48.90% 1.5 -0.03 0.656 3 Neutral
35 46.30% 2.06 0.3 3.007 3 Neutral
36 39.30% 2.06 -1.05 0.477 2 Relax
37 44% 2.39 -1.22 0.856 3 Neutral
38 50.30% 2.8 -1.27 0.433 2 Relax
39 45.80% 2.42 -1.16 1.157 4 Nervous
40 46.60% 2.11 -0.88 0.882 4 Nervous
41 52.80% 2.71 -1.21 0.199 2 Relax
42 50.60% 2.39 -0.5 0.521 1 Relax
43 41.10% 1.72 -0.63 1.785 1 Relax
44 46.40% 2.66 -0.55 0.558 1 Relax
45 50.70% 2.73 -0.8 1.035 2 Relax
46 52% 2.01 -0.72 0.33 1 Relax
47 54.10% 4.16 -1.1 0.763 1 Relax
48 52.70% 1.06 0.31 0.837 2 Relax
49 45.40% 0.86 0.18 1.659 2 Relax
50 46.40% 1.34 0.11 1.139 2 Relax
51 47.30% 1.37 -0.09 0.696 2 Relax
52 44.20% 1.1 -0.09 1.028 2 Relax
53 44.40% 0.47 0.47 1.421 4 Nervous
54 46.80% 1.29 0.19 0.855 4 Nervous
55 50.60% 1.17 0.29 0.459 3 Neutral

75
56 44.80% -0.14 1.26 0.559 3 Neutral
57 39.30% -0.55 0.97 0.824 4 Nervous
58 51% 0.46 0.71 0.399 4 Nervous
59 43.60% -0.17 0.85 0.73 3 Neutral
60 51% 1.03 0.5 0.153 4 Nervous
61 54.10% 1.31 0.34 0.156 2 Relax
62 48% 1.15 0.21 0.831 2 Relax
63 44% 0.56 0.54 2.853 2 Relax
64 48.20% 0.04 0.59 0.623 2 Relax
65 37.50% 0.29 0.2 0.82 2 Relax
66 46.80% 0.35 0.52 0.258 2 Relax
67 46.10% 0.84 -0.07 0.917 2 Relax
68 49.90% 0.29 0.54 0.777 2 Relax
69 42.70% -0.24 0.91 1.512 2 Relax
70 47.10% -0.08 1.18 0.288 1 Relax
71 53.60% 0.66 0.8 0.342 3 Neutral
72 30.60% -0.72 1.22 2.855 3 Neutral
73 55.10% 1.08 0.4 0.295 4 Nervous
74 53.50% 1.3 0.02 0.486 3 Neutral
75 52% 2.13 -0.1 0.462 1 Relax
76 43.10% 0.31 0.4 1.38 4 Nervous
77 53.10% 5.11 -0.7 0.535 4 Nervous
78 53.30% 1.2 -0.1 0.307 2 Relax
79 53.30% 30.97 -1.59 2.39 2 Relax
80 51.30% 1.25 -0.42 0.307 2 Relax
81 52.90% 1.69 -0.7 0.752 2 Relax
82 53.40% -0.07 1.39 0.45 2 Relax
83 49.40% -0.24 1.67 0.042 2 Relax
84 51.20% -0.56 2.14 1.363 2 Relax
85 46.10% -0.41 1.37 0.274 2 Relax
86 50.80% 0.1 1.14 0.169 2 Relax
87 50.20% 0.16 1.29 0.474 2 Relax
88 53.50% 1.5 -0.12 0.639 2 Relax
89 48.30% 1.01 0.03 1.499 2 Relax
90 48.20% 0.95 0.16 0.362 3 Neutral
91 54.20% 1.04 -0.04 0.79 2 Relax
92 52.80% 1.25 -0.03 0.361 3 Neutral
93 52.20% 1.9 -0.45 0.24 2 Relax
94 47.80% 0.52 1 1.155 1 Relax

76
95 47.40% 1.05 0.46 0.224 1 Relax
96 46.40% 0.68 0.46 0.276 3 Neutral
Table 5.18 Kubios HRV Analysis Result for All Occupants

We used ANOVA for analysis. However, the p-value of the factors (SD1, PNS index, SNS
index, and LF/HF) were 0.346, 0.767, 0.810, 0.512, respectively. This meant that the result was
not significant, and it would be hard to use these four factors to represent people’s stress levels.

We thought the reason might be that there were some problems with our heart-rate sensor. Heart
rates fluctuated very widely even though occupants were not engaging in any strenuous exercise.
It seemed that the heart rate data we collected was not valid for HRV generation. Also, we had
only six measurement point for each occupant during their whole experiment, and it was hard to
consider individual analysis for such small scale of data. Future work for is needed in order to
explore further correlations between HRV and stress levels.

Generally speaking, the increasing of EDA peak frequency was related to a lowering in the stress
levels of occupants, while the ANOVA result for the HRV-related index (including the PNS
index, SNS Index, LF/HF and SD1) showed that there was no significance to any correlations
between them and stress levels. We considered whether the huge fluctuation of heart-rate sensor
data led to a deviation in R-R intervals, which limited Kubios HRV analysis.

5.5 Conclusion
In order to valid the application of machine learning results, we executed our validation work in
two directions.

In a real office located in downtown Los Angeles, we recorded current skin temperatures, the
indoor temperature, and the feedback of participants. We compared th feedback to their ANN
model’s estimation results. The result showed that it was hard for our ANN model to catch small
thermal preference changes, possibly because of the lack of adequate training data. Notably, the
accuracy of male occupants’ ANN models was higher than that of female occupants’ models.

In the real-time control experiment, we successfully controlled the indoor thermal environment
as acceptable for all occupants except F3. The temperature change range of male occupants’
experiments was also less than that of female occupants’ experiments.

The two directions of validation seemed to reveal that our individual ANN model performed
better in estimating male occupants’ thermal preferences than it did in estimating those of female
occupants. We theorized that this was due to the more sensitive thermal perceptions of women,
and that physical variances such as menstrual status might be critical factors that affected ANN
models’ performances in relation to female occupants. In addition, we thought that the lack of
enough training data made it difficult for ANN to catch small preference changes.

In optimization work, we considered collecting new training data for occupants by adding a
“renew” function to our control interface. We tried this function for F3, and it obviously
improved the accuracy of her ANN model.

77
In addition, we considered applying psychological features in the future, and we compared
people’s feedback about their stress levels and thermal preferences. We found that there were
correlations between people’s stress levels and their thermal comfort and sensation levels, which
suggested that there were possibilities for adding psychological factors into machine learning.
We considered two psychological factors—EDA and HRV—and the result showed that the
increasing frequency of EDA wave peaks might indicate that occupants were under a more
relaxed psychological condition, while it was hard to use the HRV-related factors we created for
stress level prediction. We checked the original collected data and found that the heart-rate data
we collected was too fluctuant. The heart-rate sensors we used might be not suitable for HRV
analysis.

78
6 Discussion, Future Work and Conclusion
6.1 Discussion
6.1.1 Feature Determination for Thermal Estimation
Our project aimed to use machine learning for analyzing if the combination of specific skin
temperatures and indoor temperature could be regarded as signals for estimating occupants’
thermal preference. During the machine-learning process, we needed to find out which skin
temperature had the biggest effect on thermal preference estimation (input feature
determination).

As the major tool for achieving machine learning, an artificial neural network (ANN) was
regarded as the most precise option for computer analysis. The difficulty for applying ANN was
its complex inner mathematical structure. We found that a LabVIEW software plug-in called the
LabVIEW Machine Learning Toolkit provided ANN analysis services in a black-box process,
which required us to determine input features.

For feature selection, we planned to find a model that had an analysis process like ANN’s. In
order to achieve feature determination, we adopted logistic regression, whose activation function
is the same as ANN’s, for pre-analysis. Our experiment included five measured body parts
(wrist-front, wrist-back, back of neck, waist, and upper arm).

For the thermal comfort model, we planned to select two skin temperatures to estimate people’s
thermal comfort levels. Due to the obvious differences between people’s skin temperature
values, we built individual models for all occupants, and logistic regression generated the weight
values for each occupant in their thermal comfort models. We then made a determination by
comparing each skin temperature’s weight value.

After conducting thermal comfort analysis, we built an individual thermal sensation model for
every occupant. We used the selected two skin temperatures in the thermal comfort model, and
we considered whether indoor temperature could improve the accuracy of thermal sensation
estimations.

According to the comparison result obtained via logistic regression, wrist-front temperature,
back-of-neck temperature and wrist-back temperature had the biggest potential for estimating
occupants’ thermal comfort level. In addition, for thermal sensation, the accuracy of most
individual thermal sensation models was up to 100% when indoor temperature was considered.
6.1.2 Validation and Real Application for ANN Model
We wanted to verify whether the correlations we found between skin temperature, indoor
temperature and people’s thermal preferences were meaningful in practical applications.

We had two issues to determine:
(1) Whether the accuracy of our model’s thermal preference estimation was acceptable.
(2) Whether our model could be applied for real-time HVAC control.

79
In order to validate our model’s accuracy, we designed an experiment in a real office in
downtown Los Angeles. The result showed that male occupants’ thermal preferences were more
consistent than female occupants’ thermal preferences, and our individual ANN models
performed better in the male group.

Our HVAC real-time control experiment also suggested that our ANN model could provide more
acceptable controlled thermal environments for male occupants.

During female occupants’ experiments, the temperature variation ranges became larger. When
more training data for machine learning was added, the accuracy of ANN predictions seemed to
improve. The renewed ANN model also provided a more acceptable controlled thermal
environment, which suggested the importance of enlarging the training data scale.

6.1.3 Limitation
There were factors that affected the accuracy of our models and restricted the validation process
during the whole project period.

One of the most important factors that limited the accuracy of machine learning process was the
lack of adequate training data. At the beginning of our project, because of our wish to create both
general models and individual models, we asked 20 occupants to participate in our experiments.
Based on the settled schedule of experiments, we could only obtain two hours per occupant in
which to collect training data. However, the measured data showed that the same local skin
temperatures between different people were quite different, and it was difficult to build a general
model to estimate all people’s thermal preferences. For this reason, we could only build
individual models for all participants based on their two hours of measured training data.

The effect of small training scale was revealed in the validation process, where the individual
model could hardly estimate the variation of Occupant F3’s thermal preference. When we
renewed F3’s model by adding more training data, the accuracy improved substantially. This
suggested a training data deficiency.

In addition, there were some hardware and device problems that limited the application of real
time control. During real-time control process, the local temperature of our power relay hardware
was improved drastically after 20 minutes, and this led to a malfunction in the hardware. The
cause of this problem might have been the poor contact between the power relay hardware and
the wires, while the power of HVAC was too large to avoid strong electric current. As well, due
to the limitation of sensor accuracy and experiment process, it was hard to explore further
correlations in extended psychological factor research

Another issue was that our experiment process itself might have affected people’s thermal
preferences. We attached too many sensors to occupants’ bodies, and in order to get rid of sensor
dropping, we pasted sensor wires onto occupants’ skin. This might have changed the clothing
factor for occupants, and led to deviations in their thermal preference levels.

80
Finally, due to the black-box form of the ANN model, we could only combine thermal comfort
and thermal sensation feedback, which might also have increased the deviations in thermal
preference estimation.

6.2 Future Work
According to the result of our project, we found that there is potential for wrist-front and -back
temperatures and indoor temperature to help estimate people’s preferences. There are also some
future efforts to be made that may increase the practical significance of our research.

To begin with, in order to improve the accuracy of our model, we need to enlarge the training
scale for machine learning process. This project already revealed that an individual model is
more practical for estimating thermal preference, so it would be better to collect more training
data for one or two occupants in the future. In addition, more physiological factors (such as
menstrual status for female occupant) should be considered to identify their impact on thermal
preference.

As well, we have already found that there is potential to use only wrist-front and -back
temperatures and indoor temperature to estimate occupants’ thermal preferences. Future studies
can focus on how to improve the estimation accuracy within these three parameters, and decrease
unnecessary sensor installations on occupants’ skin, as the latter may cause deviations in their
thermal preference.

Finally, another important piece of work that needs to be completed in the future is to build
artificial neural network models with clear inner structures. To capture the separate
considerations involved in people’s thermal comfort and thermal sensation estimations, we need
to build ANN models that can predict thermal comfort levels and thermal sensation levels
respectively. The results of our analysis indicate that when adding new features to improve
machine learning accuracy, we need to solve the problem of heart rate accuracy, and we need to
increase the feedback frequency of occupants to collect more data for correlation finding.

6.3 Conclusion
This project aimed to use machine learning to explore the correlations between skin
temperatures, indoor temperature and people’s thermal preferences. While conducting it, we
discovered that there was potential for wrist-front and -back temperatures to be used in
estimating human’s thermal comfort levels. These two skin temperatures could also help
estimate people’s thermal sensation with a very high degree of accuracy when we took indoor
temperature into consideration. In addition, we found that there were obvious differences in skin
temperature values between different people, and individual machine learning models were more
effective for thermal prediction than general models.

Based on the discovered correlations, we executed our validation work in two directions. We
wanted to find out that whether our machine learning model could predict humans’ thermal
preference with a high accuracy, and whether our model’s predictions could be regarded as
signals to achieve HVAC automation. The result showed that machine-learning ANN models
performed better for men, while they could not discover the micro-changes in women’s thermal
preference. We set a renewed system where occupants could provide more feedback in order to

81
increase their training data scales. When more training data was added for one female occupant
through this method, her model was obviously optimized. We considered that the limitation of
training data scale might affect the accuracy of machine learning model. In addition, given the
huge differences in model performance for female occupants, we also thought that different
menstrual statuses could impact those occupants’ preferences with regard to surrounding thermal
conditions.

Generally speaking, wrist-front and -back skin temperatures and indoor temperature can be
applied into machine learning for thermal preference estimation, and individual ANN thermal
preference models can be adopted for real-time HVAC control. ANOVA testing results
suggested that our ANN model predicted male occupants’ thermal preferences precisely, but was
less accurate when it came to female occupants. Increasing the training data scale can optimize
the estimation performance of individual ANN models, and help identify people’s feeling for
different thermal conditions with higher accuracy.

82
REFERENCES
Chaudhuri, T., Zhai, D., Soh, Y. C., Li, H., & Xie, L. (2018). Thermal comfort prediction using
normalized skin temperature in a uniform built environment. Energy and Buildings, 159,
426–440. https://doi.org/10.1016/j.enbuild.2017.10.098
Choi, J. H., Loftness, V., & Lee, D. W. (2012). Investigation of the possibility of the use of heart
rate as a human factor for thermal sensation models. Building and Environment, 50, 165–
175. https://doi.org/10.1016/j.buildenv.2011.10.009
Choi, J. H., & Yeom, D. (2017). Study of data-driven thermal sensation prediction model as a
function of local body skin temperatures in a built environment. Building and Environment,
121, 130–147. https://doi.org/10.1016/j.buildenv.2017.05.004
d’Ambrosio Alfano, F. R., Ianniello, E., & Palella, B. I. (2013). PMV-PPD and acceptability in
naturally ventilated schools. Building and Environment, 67, 129–137.
https://doi.org/10.1016/j.buildenv.2013.05.013
Deng, Q., Wang, R., Li, Y., Miao, Y., & Zhao, J. (2017). Human thermal sensation and comfort
in a non-uniform environment with personalized heating. Science of the Total Environment,
578, 242–248. https://doi.org/10.1016/j.scitotenv.2016.05.172
Ghahramani, A., Tang, C., & Becerik-Gerber, B. (2015). An online learning approach for
quantifying personalized thermal comfort via adaptive stochastic modeling. Building and
Environment, 92, 86–96. https://doi.org/10.1016/j.buildenv.2015.04.017
Humphreys, M. A., & Nicol, J. F. (2002). The validity of ISO-PMV for predicting comfort votes
in every-day thermal environments, 34, 667–684.
I, G. S. B., & Dear, R. J. De. (1998). Thermal adaptation in the built environment : a literature
review. Energy and Buildings, 27, 83–96. https://doi.org/10.1016/S0378-7788(97)00053-4
Jung, W., & Jazizadeh, F. (2018). Vision-based thermal comfort quanti fi cation for HVAC
control. Building and Environment, 142(May), 513–523.
https://doi.org/10.1016/j.buildenv.2018.05.018
Klemm, W., Heusinkveld, B. G., Lenzholzer, S., Jacobs, M. H., & Van Hove, B. (2015).
Psychological and physical impact of urban green spaces on outdoor thermal comfort during
summertime in The Netherlands. Building and Environment, 83, 120–128.
https://doi.org/10.1016/j.buildenv.2014.05.013
Klemm, W., Heusinkveld, B. G., Lenzholzer, S., & van Hove, B. (2015). Street greenery and its
physical and psychological impact on thermal comfort. Landscape and Urban Planning,
138, 87–98. https://doi.org/10.1016/j.landurbplan.2015.02.009
Lischke, A., Jacksteit, R., Mau-Moeller, A., Pahnke, R., Hamm, A. O., & Weippert, M. (2018).
Heart rate variability is associated with psychosocial stress in distinct social domains.
Journal of Psychosomatic Research, 106(August 2017), 56–61.
https://doi.org/10.1016/j.jpsychores.2018.01.005
Liu, Y., Wang, L., Liu, J., & Di, Y. (2013). A study of human skin and surface temperatures in
stable and unstable thermal environments. Journal of Thermal Biology, 38(7), 440–448.
https://doi.org/10.1016/j.jtherbio.2013.06.006
Luo, M., Cao, B., Ji, W., Ouyang, Q., Lin, B., & Zhu, Y. (2016). The underlying linkage
between personal control and thermal comfort: Psychological or physical effects? Energy
and Buildings, 111, 56–63. https://doi.org/10.1016/j.enbuild.2015.11.004
Maiti, R. (2014). International Journal of Industrial Ergonomics PMV model is insuf fi cient to
capture subjective thermal response from Indians. International Journal of Industrial
Ergonomics, 44(3), 349–361. https://doi.org/10.1016/j.ergon.2014.01.005

83
Metzmacher, H., Wölki, D., Schmidt, C., Frisch, J., & van Treeck, C. (2018). Real-time human
skin temperature analysis using thermal image recognition for thermal comfort assessment.
Energy and Buildings, 158, 1063–1078. https://doi.org/10.1016/j.enbuild.2017.09.032
Nikolopoulou, M., & Steemers, K. (2003). Thermal comfort and psychological adaptation as a
guide for designing urban spaces. Energy and Building, 35(1), 95–101. https://doi.org/Pii
S0378-7788(02)00084-1\nDoi 10.1016/S0378-7788(02)00084-1
Ogrmar SSR-25 DA 25A 3-32V DC / 24-380V AC Solid State Relay and Heat Sink. (n.d.).
Retrieved from https://www.amazon.com/Ogrmar-SSR-25-3-32V-24-380V-
Solid/dp/B074FT4VXB/ref=sr_1_1_sspa?ie=UTF8&qid=1548887976&sr=8-1-
spons&keywords=Ogrmar+SSR-25+DA&psc=1)
Sakoi, T., Tsuzuki, K., Kato, S., Ooka, R., Song, D., & Zhu, S. (2007). Thermal comfort, skin
temperature distribution, and sensible heat loss distribution in the sitting posture in various
asymmetric radiant fields. Building and Environment, 42(12), 3984–3999.
https://doi.org/10.1016/j.buildenv.2006.10.050
Sato, K., & Usui, M. (2018). Haptics: Science, Technology, and Applications (Vol. 10893).
Springer International Publishing. https://doi.org/10.1007/978-3-319-93445-7
Schweiker, M., Brasche, S., Bischof, W., Hawighorst, M., & Wagner, A. (2013). Explaining the
individual processes leading to adaptive comfort: Exploring physiological, behavioural and
psychological reactions to thermal stimuli. Journal of Building Physics, 36(4), 438–463.
https://doi.org/10.1177/1744259112473945
Taleghani, M., Tenpierik, M., Kurvers, S., & Van Den Dobbelsteen, A. (2013). A review into
thermal comfort in buildings. Renewable and Sustainable Energy Reviews, 26, 201–215.
https://doi.org/10.1016/j.rser.2013.05.050
Tarvainen, M. P., Niskanen, J., Lipponen, J. A., Ranta-aho, P. O., & Karjalainen, P. A. (2013).
Kubios HRV – Heart rate variability. Computer Methods and Programs in Biomedicine,
113(1), 210–220. https://doi.org/10.1016/j.cmpb.2013.07.024
ter Mors, S., Hensen, J. L. M., Loomans, M. G. L. C., & Boerstra, A. C. (2011). Adaptive
thermal comfort in primary school classrooms: Creating and validating PMV-based comfort
charts. Building and Environment, 46(12), 2454–2461.
https://doi.org/10.1016/j.buildenv.2011.05.025
Venkatesh, M. (2018). What is Training Data, Really? Retrieved from
https://blog.playment.io/what-is-training-data/
Vernier. (n.d.). Retrieved from https://www.vernier.com/products/
Weka – GUI way to learn Machine Learning. (n.d.). Retrieved from
https://www.analyticsvidhya.com/learning-paths-data-science-business-analytics-business-
intelligence-big-data/weka-gui-learn-machine-learning/
What Is LabVIEW? (n.d.). Retrieved from http://www.ni.com/en-us/shop/labview.html
Yadav, A. K., Malik, H., & Chandel, S. S. (2014). Selection of most relevant input parameters
using WEKA for artificial neural network based solar radiation prediction models.
Renewable and Sustainable Energy Reviews, 31, 509–519.
https://doi.org/10.1016/j.rser.2013.12.008
Yeom, D., Choi, J. H., & Zhu, Y. (2017). Investigation of physiological differences between
immersive virtual environment and indoor environment in a building. Indoor and Built
Environment, 0(0), 1–17. https://doi.org/10.1177/1420326X17731945
Zhong, C., & Choi, J. (2017). ScienceDirect ScienceDirect Development of a Data-Driven
Approach for Human-Based Environmental Control. Procedia Engineering, 205, 1665–
1671. https://doi.org/10.1016/j.proeng.2017.10.341

84
Zhu, H., Wang, H., Liu, Z., Li, D., Kou, G., & Li, C. (2018). Experimental study on the human
thermal comfort based on the heart rate variability (HRV) analysis under different
environments. Science of the Total Environment, 616–617, 1124–1133.
https://doi.org/10.1016/j.scitotenv.2017.10.208
机器学习. (n.d.). Retrieved from https://www.coursera.org/learn/machine-learning?

Abstract (if available)

Abstract Thermal comfort optimization is crucial for Heating, Ventilation and Air Conditioning (HVAC) automation, which adjusts thermal conditions automatically, based on occupants’ real-time thermal preferences. However, most standard thermal comfort models, such as Predicted Mean Vote (PMV), have not considered individual thermal preferences. In many cases, it is reported that the prediction of PMV is not accurate enough, especially for individual use, where thermal preference is relatively subjective for everyone. Previous research focuses on how some physiological indices indicate people’s thermal comfort states, while psychological factors such as heart rate variability, which also affects their preferences, are practically ignored because of the difficulty in quantifying occupants’ psychological activities. ❧ The algorithms use human indexes to predict an individual’s thermal comfort, and estimate their thermal sensation state based on environment parameters (indoor temperature). The estimation results are regarded as signals that trigger a work-mode change in HVAC devices to optimize the indoor thermal environment to comply with the preferences of a building’s occupants. ❧ A series of human subject experiments were performed in an environmental chamber at USC to collect different types of data, such as human body indexes, including skin temperature (℉), electrode activity (μS), and the rate of low-frequency to high-frequency (LF/HF) for heart rate variability. The last two represent the occupants’ reflections of outside stimuli and their stress levels, respectively. Environmental parameters, including indoor temperature (℉), relative humidity, radiant temperature (℉), and CO₂ concentration (ppm), were also defined. The study used machine learning to construct an algorithm set, and used it to build an HVAC auto-control process based on real-time sensor data. ❧ The research outcome shows the potential correlations between the human indexes/environment factors that we chose, and the thermal comfort/thermal sensation levels, as well as the control design. It presents applications for an HVAC auto-control based on designed thermal condition estimation and incorporating real-time data from wearable sensors.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Enhancing thermal comfort: data-driven approach to control air temperature based on facial skin temperature

PDF

Developing environmental controls using a data-driven approach for enhancing environmental comfort and energy performance

PDF

Human-building integration based on biometric signal analysis: investigation of the relationships between human comfort and IEQ in a multi-occupancy condition

PDF

Developing a data-driven model of overall thermal sensation based on the use of human physiological information in a built environment

PDF

Human–building integration: machine learning–based and occupant eye pupil size–driven lighting control as an applicable visual comfort tool in the office environment

PDF

Enhancing thermal comfort: air temperature control based on human facial skin temperature

PDF

Multi-occupancy environmental control for smart connected communities

PDF

Human-building integration: Investigation of human eye pupil sizes as a measure of visual sensation in the workstation environment

PDF

Exploring participatory sensing and the Internet of things to evaluate temperature setpoint policy and potential of overheating/overcooling of spaces on the USC campus

PDF

Natural ventilation in tall buildings: development of design guidelines based on climate and building height

PDF

Learning personal thermal comfort and integrating personal comfort requirements into HVAC system control loop

PDF

Development of data-driven user-centered building façade design guideline models: machine learning-based approaches to predict user preferences

PDF

Indoor environmental quality and comfort: IEQ adaptation and human physiological responses in commercial buildings

PDF

Visualizing thermal data in a building information model

PDF

Impacts of building performance on occupants' work productivity: a post occupancy evaluation study

PDF

Nursing comfort study: are hospitals performing for nurses? IEQ analysis of working conditions inside hospitals

PDF

Smart buildings: employing modern technology to create an integrated, data-driven, intelligent, self-optimizing, human-centered, building automation system

PDF

Adaptive façade controls: A methodology based on occupant visual comfort preferences and cluster analysis

PDF

Impact of occupants in building performance: extracting information from building data

PDF

Indoor air quality for human health in residential buildings

Asset Metadata

Creator Liu, Hanxun (author)

Core Title Exploration for the prediction of thermal comfort & sensation with application of building HVAC automation

School School of Architecture

Degree Master of Building Science

Degree Program Building Science

Publication Date 04/24/2021

Defense Date 03/20/2019

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag automation,HVAC control,machine learning,OAI-PMH Harvest,real time,thermal preference

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Choi, Joon-Ho (committee chair), Gil, Yolanda (committee member), Schiler, Marc Eugene (committee member)

Creator Email 531630391@qq.com,hanxunli@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-146206

Unique identifier UC11660611

Identifier etd-LiuHanxun-7264.pdf (filename),usctheses-c89-146206 (legacy record id)

Legacy Identifier etd-LiuHanxun-7264.pdf

Dmrecord 146206

Document Type Thesis

Format application/pdf (imt)

Rights Liu, Hanxun

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA