Energy & Buildings 319 (2024) 114507

Contents lists available at ScienceDirect

Energy & Buildings

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

Data-efficient comfort modeling: Active transfer learning for predicting

personal thermal comfort using limited data
Zeynep Duygu Tekler, Yue Lei, Adrian Chong ∗
Department of the Built Environment, National University of Singapore, 4 Architecture Drive, 117566, Singapore

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

Dataset link: https://github.com/ideas-lab- Personal comfort models are used to predict thermal comfort responses at the individual level rather than
nus/ComfortPredict-ActiveTL predicting the average thermal comfort responses for large populations. These models, data-driven in nature, need
to be trained on large amounts of occupant comfort feedback and sensor data to achieve accurate predictions.
Keywords:
However, collecting such data is often expensive and labor-intensive in reality. To address this, we proposed a
Data-efficient machine learning
Active learning data-efficient active transfer learning (ATL) framework to improve the performance of personal comfort models
Transfer learning under limited data. To demonstrate the validity of this framework, we developed a base Convolutional Neural
Personal thermal comfort models Network-Long Short-term Memory (CNNLSTM) model alongside two transfer learning models utilizing feature
CNNLSTM extraction (TL-CNNLSTM-FE) and fine-tuning (TL-CNNLSTM-FE) approaches, enhanced by a novel active learning
strategy. Using these models, three comfort prediction tasks (i.e., thermal preference, thermal acceptability, and
air movement preference) were performed by transferring the knowledge from the ASHRAE Global Thermal
Comfort Database II to a limited dataset collected in the tropics. Empirical results indicate that the active transfer
learning framework proposed was able to consistently outperform the base and transfer learning models using
only less than 10% of the training data for all personal comfort tasks, highlighting the effectiveness of this
strategy. The implications of this work are especially useful for the research community working on the practical
applications of data-efficient machine learning approaches for personal thermal comfort predictions.

1. Introduction as occupant-centric Heating, Ventilation and Air Conditioning (HVAC)

controls [6] based personal thermal comfort and occupancy to improve
There is a growing trend in learning and predicting individual ther- the efficiency and effectiveness of HVAC systems. Past studies suggest
mal comfort needs through data gathered from everyday environments. that smart HVAC systems can improve energy savings and thermal com-
Personalized thermal comfort models play a pivotal role in this shift, fort by more than 15% [7], while employing an appropriate comfort
aiming to predict thermal comfort responses at the individual level model for HVAC controls can lead to further energy savings exceeding
rather than relying on the average response of large populations. Unlike
10% during simulations [8]. Meanwhile, in real-world implementations,
traditional models, including Predictive Mean Vote (PMV) and adap-
personal preference-based HVAC control systems have demonstrated re-
tive models [1], personal comfort models are designed to account for
markable energy savings, ranging from 28% to 35% [9], proving the
unique differences between each occupant, allowing for a more pro-
effectiveness of personal comfort models in improving energy efficiency.
found comprehension of their distinct thermal comfort needs in the
Personal comfort models, being inherently data-driven machine
buildings. Such information not only enhances occupants’ comfort but
also contributes to energy efficiency by ensuring that the building sys- learning models, require two primary categories of data for their de-
tems operate optimally to meet specific comfort requirements [2]. A velopment. (1) The first category includes objective measurements,
notable example is occupant-centric control applications, which inte- including indoor and outdoor environmental conditions, HVAC oper-
grate occupants’ personal comfort and preferences in building controls ations, building occupancy, and occupants’ physiological conditions.
[3]. These include personalized lighting control [4] to improve visual These measurements are obtained through various sensing and com-
satisfaction and lighting energy use, personalized plug load control [5] munication technologies. (2) The second category involves collecting
based on occupancy and personalized plug load use preferences, as well occupant comfort feedback through subjective surveys. This feedback

* Corresponding author.
E-mail addresses: [email protected] (Z.D. Tekler), [email protected] (Y. Lei), [email protected] (A. Chong).

https://doi.org/10.1016/j.enbuild.2024.114507
Received 27 February 2024; Received in revised form 25 May 2024; Accepted 29 June 2024
Available online 3 July 2024
0378-7788/© 2024 Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
Z.D. Tekler, Y. Lei and A. Chong Energy & Buildings 319 (2024) 114507

indicates occupants’ thermal preference, thermal acceptability, or ther- of different models [10], emphasizing a significant improvement in pre-
mal sensation based on the thermal conditions they experience. dictive accuracy (17-40%) when compared with conventional comfort
While recent developments in the Internet of Things (IoT) and sens- models (i.e., PMV and adaptive models). This underscores the need for
ing technologies have enabled the cost-efficient and effective collection adopting more personalized approaches for predicting thermal comfort.
of various sensor measurements, challenges persist during the collection The data used for developing personal comfort models were obtained
of occupant comfort feedback, also known as user-labeled data. These through (1) various sensor measurements and (2) occupant comfort
challenges are primarily due to the intrusive and labor-intensive nature feedback via surveys. Sensor measurements can be categorized under
of the data collection process, often leading to survey fatigue during five broad categories: indoor environmental measurements, outdoor
the data labeling. Researchers also reported difficulties determining an weather measurements, occupant-related measurements, temporal mea-
appropriate frequency to collect comfort feedback from the study par- surements, and system (i.e., HVAC) related measurements. Indoor envi-
ticipants [10]. A recent study further quantified this challenge, finding ronmental measurements include indoor temperature (𝑇𝑖𝑎 ), indoor rel-
that approximately 250–300 data points per study participant are nec- ative humidity (𝑅𝐻𝑖 ), indoor (𝐶𝑂2 ) levels, indoor radiant temperature
essary for achieving accurate predictions of personal thermal comfort, (𝑇𝑟 ), indoor air velocity (𝐴𝑉𝑖 ). Outdoor weather measurements often in-
highlighting the significant volume of data needed [11]. Therefore, it is volve collecting outdoor humidity, outdoor air temperature (𝑇𝑜𝑎 ), solar
essential to strike a balance between gathering sufficient data to de- radiation, wind speed, and wind direction. Both indoor and environ-
velop accurate personal comfort models while minimizing the effort mental measurements have been widely used to predict thermal pref-
required from respondents. This challenge highlights the crucial need erences [16], thermal sensations [17], and thermal acceptability [18].
for developing more data-efficient approaches that can accurately pre- Occupant-related measurements encompass both physiological and be-
dict personal comfort with less reliance on extensive comfort feedback. havioral data. To incorporate occupant physiological measurements into
In this paper, our objective is to reduce the need of collecting large thermal comfort modeling, specialized sensors are utilized to capture pa-
sums of sensor and occupant comfort survey data to develop data- rameters such as skin temperatures, heart rate, and metabolic rate [19].
efficient personal comfort models with satisfactory model performance. Furthermore, occupant activity levels (i.e., exercise habits) and clothing
The study contributions are outlined as follows: levels are collected as part of behavioral measurements [20]. Temporal
measurements, such as the day of the week or specific hours of the day,
• Introduced a novel active transfer learning framework that inte- represent the variations over time in the thermal comfort levels experi-
grates active learning and transfer learning approaches to improve enced by occupants Finally, HVAC-related measurements represent the
the performance of personal comfort predictions using limited data. operational status of indoor environments, encompassing variables such
• Presented a comprehensive comfort evaluation for three comfort as damper position, supply airflow rate, and variable air volume control
tasks (i.e., thermal acceptability, thermal preference, and air move- settings [10].
ment preference) by transferring the knowledge from one of the In addition to sensor measurements, the development of personal
largest global thermal comfort databases (i.e., ASHRAE Global thermal comfort models necessitates gathering feedback on occupants’
Thermal Comfort Database II) to our limited dataset collected in comfort using various methods. Surveys are frequently employed, uti-
the tropics. lizing a range of interfaces such as wearables or thermostats to collect
• Developed two transfer learning approaches (i.e., feature extraction comfort feedback efficiently [21]. Occupants’ comfort feedback can be
and fine-tuning) based on a CNNLSTM architecture and evaluated collected based on their thermal acceptability (Acceptable, Unaccept-
them with different amounts of training data under air-conditioned able), thermal preference (Cooler, No change, Warmer), and thermal
(AC) and natural ventilation (NV) conditions. sensation (7-point scale Predicted Mean Vote (PMV) index). Further-
more, air movement preferences (More, No Change, Less) and air move-
The remainder of this paper is structured as follows: Section 2 re- ment acceptability (Acceptable, Unacceptable) were also considered as
views the existing personal comfort models and data-efficient machine comfort measures for mixed-mode buildings in recent studies [22].
learning approaches for personal comfort models. Section 3 provides
details on the active transfer learning methodology, including the uti- 2.2. Active learning and transfer learning approaches for personal comfort
lized datasets, the architecture of the base CNNLSTM model, transfer models
learning approaches, and the novel active learning strategy developed.
Section 4 presents the results and analysis, evaluating the impact of ac- Data-efficient machine learning approaches have been increasingly
tive transfer learning strategy and different amounts of training data gaining attention for personal comfort modeling, addressing the chal-
on the performance of personal comfort models. Finally, Section 5 con- lenges associated with collecting large amounts of training data to
cludes the paper by summarizing the key findings of this research. achieve accurate personal comfort models [23]. Among the small num-
ber of studies available, these approaches enhance the performance of
2. Related work traditional machine-learning models when dealing with limited training
data. This is especially useful as collecting frequent occupant comfort
2.1. Development of personal comfort models feedback (i.e., user-labeled data) is expensive, labor-intensive, and of-
ten impractical for extended data collection periods.
In recent years, there has been a notable rise in the popularity of Active learning (AL) is a promising approach for predicting personal
data-driven thermal comfort modeling, leading to many efforts leverag- comfort with limited training data. AL is used to identify the most in-
ing machine learning (ML) techniques for predicting personal thermal formative data points for user labeling, thereby potentially reducing the
comfort. Popular machine learning models, such as gradient boosting frequency of user comfort feedback collection. A pioneering study [24]
[12], random forest (RF) classification [13], logistic regression (LR) demonstrated that AL could reduce user labeling efforts by up to 46%
[14], support vector machine (SVM) [15], and have been employed in personalized thermal comfort models during a field implementation.
to predict personal thermal comfort based on different thermal met- The same study also implemented a feature selection algorithm [25] to
rics. Notably, Luo et al. [12] performed a comprehensive comparison of understand and filter the useful input features before applying AL as
nine common machine learning algorithms for predicting thermal sen- part of a data-efficient minimum sensing strategy. The application of
sation using the ASHRAE Comfort Database II. Their findings indicated AL has been extended to HVAC control systems, with results indicating
that machine learning models exhibited superior accuracy compared to a substantial decrease in labeling effort (about 31.0%) while still main-
traditional PMV models, with Random Forest outperforming other al- taining high levels of personal comfort and energy savings compared to
gorithms. This observation particularly aligns with Kim et al.’s review conventional controls [26].

2
Z.D. Tekler, Y. Lei and A. Chong Energy & Buildings 319 (2024) 114507

Fig. 1. The ASHRAE Global Thermal Comfort Database II (left) and Singapore BCA Dataset (right) serve as the source and target domains in this study, respectively.

Transfer Learning (TL) presents an alternative strategy involving the comfort tasks (i.e., thermal preference, thermal acceptability, and air
application of pre-trained models from an information-rich (source) do- movement preference) using a real-world comfort dataset collected in
main to a related task in a different (target) domain where labeled data the tropics.
is limited or unavailable. Gao et al. [27] investigated TL for predicting
thermal comfort across different cities. They utilized multilayer per- 3. Methodology
ception (MLP) models trained on data from various climatic regions
and applied them to a (target) building containing scarce data. Their 3.1. Dataset description
findings indicated that models pre-trained on comparable climate zones
demonstrated improved prediction accuracy. Similarly, Somu et al. [28] Two thermal comfort datasets were utilized for this study. The first
developed a transfer learning-based Convolutional Neural Networks- dataset, the ASHRAE Global Thermal Comfort Database II [32], serves as
Long Short Term Memory (CNNLSTM) model for thermal preference the information-rich source domain, containing an extensive amount of
prediction. The resulting model reported 56% accuracy, using the same thermal comfort data collected from multiple buildings across different
source and target domains with [27]. Furthermore, Park et al. [29] in- climatic zones. The second dataset, the Singapore Building Construction
troduced an ensemble TL approach, leveraging knowledge from diverse Authority (BCA) testbed dataset [22], represents the information-poor
thermal conditions and physiological data to predict thermal comfort target domain, consisting of a limited amount of thermal comfort data
preferences accurately for specific subjects with sparse data. Finally, Li collected from a single building in a tropical climate. Further elaboration
et al. [30] proposed a transfer learning model using TrAdaBoost and in- on the data collection process and the fields present in both datasets is
troduced an automatic weighting strategy to adjust the weights of the provided in the subsequent subsections and depicted in Fig. 1.
training data from the source domain for performance improvement.
To date, only one study [31] has attempted to combine both active 3.1.1. Source domain: ASHRAE global thermal comfort database II
and transfer learning for predicting personal thermal comfort responses. The ASHRAE Global Thermal Comfort Database II currently stands
The study involved using a ridge regression model to learn the thermal as one of the most comprehensive open-source repositories for ther-
comfort distribution of a group of occupants within a building before mal comfort research, encompassing a compilation of 34 field-measured
transferring the model to a different group of occupants from the same thermal comfort studies [32]. This database comprises data collected
building by leveraging limited queries. However, the study is limited as from over 800 commercial and residential buildings, with contributions
it still requires collecting large amounts of comfort feedback data from from researchers representing 23 countries and 39 institutions across
other occupants in the building initially before it can be transferred to 16 different climate zones. Each dataset incorporated into the database
a new occupant within the same building, significantly reducing the ap- undergoes a rigorous quality assurance process to ensure its suitabil-
proach’s scalability when applied at a large scale. ity for robust hypothesis testing. Various steps are taken to maintain
Our study extends upon past studies by proposing a novel active quality control, including visualizing field variables to identify outliers,
transfer learning framework that leverages one of the largest global generating cross-plots between highly correlated variables (e.g., ther-
comfort datasets, the ASHRAE Global Thermal Comfort Database II, to mal comfort and thermal sensation) to detect inaccurately coded data,
train the initial model on the personal comfort preferences of occu- and randomly sampling from each study to ensure consistency between
pants from different buildings. The model is subsequently transferred the original and standardized database.
to a new group of occupants using a limited comfort dataset collected In cases where certain data fields were absent from the dataset,
from the new building. This proposed method significantly improves the those fields were replaced with null values to prevent misrepresentation.
approach’s scalability when applied at a large scale. Additionally, we Through this process, a total of 81,846 data entries, containing informa-
further demonstrated the robustness and effectiveness of the approach tion about the building’s indoor and outdoor environmental conditions,
by being the first study to perform a comprehensive evaluation of dif- as well as study participants’ demographic information, occupancy pat-
ferent active transfer learning strategies based on a diverse range of terns, and subjective comfort votes (i.e., thermal acceptability, thermal

3
Z.D. Tekler, Y. Lei and A. Chong Energy & Buildings 319 (2024) 114507

Table 1
Summary of operating conditions in the Singapore BCA Dataset.

Mode Description Operating Room Temperature Ceiling Fan

Condition Setpoint (◦ 𝐶 ) Speed (m/s)

AC Activating the ambient VAV system 1 24 Off (0.15)

with traditional ceiling diffusers 2 24 Off (0.15)
3 26 Low (0.45)
4 26 Medium (0.90)

NV HVAC System switched off with the 5 Free-Floating Off (0.15)

plenum windows fully open 6 Low (0.45)
7 Medium (0.90)
8 High (1.15)

comfort, thermal preference, air movement preference, and air move- provide information about the building’s indoor and outdoor environ-
ment acceptability), were retained. Additionally, the data from the origi- mental conditions, such as indoor room temperature (𝑇𝑖𝑎 ), indoor rela-
nal ASHRAE RP-884 database is also included, resulting in 107,463 total tive humidity (𝑅𝐻𝑖 ), air velocity (AV), globe temperature (𝑇𝑔 ), outdoor
entries. temperature (𝑇𝑜𝑎 ), outdoor relative humidity (𝑅𝐻𝑜 ), and the room’s
ventilation mode (i.e., air-conditioned or naturally ventilated).
3.1.2. Target domain: Singapore BCA testbed dataset Following this step, we addressed the missing data issue in some of
The BCA testbed is a 50 𝑚2 experimental facility consisting of six the fields within the ASHRAE Global Thermal Comfort Database II. An
desks, an extra desk for a researcher, and an operable window located imputation step is employed to populate the missing data. Despite var-
on the facade’s west side. The operable window allows the testbed to ious imputation algorithms proposed previously [33], we opted for the
operate as an air-conditioned or naturally ventilated space. The testbed MissForest imputation algorithm due to its widespread acceptance and
is conditioned by a variable air volume (VAV) system and a stand-alone effectiveness in handling missing tabular data [34]. The imputation pro-
air-handling unit with its own building management system for greater cess begins by focusing on the column with the lowest amount of missing
precision in controlling space conditions. Additionally, the indoor envi- values (referred to as the candidate column) and initially imputing the
ronmental conditions are monitored by a network of advanced sensors missing values in other columns using their mean values. Subsequently,
collecting information about indoor relative humidity, indoor air tem- an RF model is trained, where the candidate column’s missing values
perature, air velocity, globe temperature, carbon dioxide levels, total are predicted based on the values in the other columns. This process is
volatile organic compounds (TVOC), and indoor fine particulate matter repeated for all of the columns with missing values, iteratively improv-
(𝑃 𝑀2.5 ). Real-time measurements of outdoor weather conditions are ob- ing the imputed dataset until the difference between successive imputed
tained through a weather station, including outdoor relative humidity, datasets becomes negligible.
outdoor air temperature, outdoor fine particulate matter (𝑃 𝑀1 , 𝑃 𝑀2.5 , After imputing the missing data within the ASHRAE Global Thermal
𝑃 𝑀10 ), and atmospheric pressure. Comfort Database II, an analysis was performed on the subjective votes
To collect occupants’ subjective comfort votes under different venti- reported by the study participants in both the ASHRAE Global Database
lation conditions, a data collection effort was conducted in June 2022 II and BCA Testbed Dataset (refer to Fig. 2). By examining the distri-
bution of reported labels in Fig. 2, it can be observed that the study
[22]. Fifty-eight tropically acclimatized (residing in Singapore for at
participants in the ASHRAE Global Database II tend to more frequently
least the past three years) participants (50% males and 50% females),
report that they were satisfied with the current environmental condi-
aged between 21 and 60, took part in the data collection. Through-
tions by indicating “Acceptable” (72%) when asked about their thermal
out the data collection period, the indoor conditions in the testbed
acceptability, followed by “Unacceptable” (28%). Similarly, for thermal
were altered every thirty minutes, following a randomized sequence,
comfort preferences, 50% of the responses indicated “No Change” fol-
to one of eight distinct settings. Half of these settings were categorized
lowed by “Cooler” at 32.9% and finally “Warmer” at 17.1%. A similar
under Natural Ventilation (NV) mode, while the remaining fell under
trend is observed when asked about their air movement preferences,
Air-Conditioned (AC) mode. At the beginning of each setting, adjust-
with the majority of responses indicating “No Change” (55%), followed
ments were made manually to the room temperature setpoint and ceiling
by “More” at 36% and “Less” at 9%. This trend is also reflected in the
fan speed, following the specifications outlined in Table 1, to establish
subjective votes reported by the study participants in the BCA testbed,
distinct operating conditions. A brief thermal comfort survey was dis-
with the majority of responses reporting “No Change” when asked about
tributed via wearable devices assigned to participants at the onset of
their air movement and thermal preferences, and “Acceptable” when
each setting, at the 5th and 15th minute intervals, to gather information
asked about their thermal acceptability. This imbalance between the
regarding their thermal acceptability, preference, and air movement
comfort votes reported by the study participants may lead to the de-
preference under the current conditions. Additionally, a comprehensive
velopment of a comfort prediction model that is heavily biased toward
Indoor Environmental Quality (IEQ) survey was conducted at the 25th-
the majority class (i.e., No Change or Acceptable), and may fail to cor-
minute mark using the Qualtrics online platform to assess participants’
rectly identify conditions which are uncomfortable for the occupants.
current levels of thermal, air quality, acoustic, and visual comfort. This
Therefore, we addressed the imbalance in the subjective labels by ap-
data collection effort resulted in 1,745 entries collected over the course
plying the Synthetic Minority Over-sampling Technique (SMOTE) on the
of the study period.
ASHRAE Global Thermal Comfort Database II to rebalance the instances
belonging to the minority classes.
3.2. Data processing and analysis
3.2.1. Synthetic minority over-sampling technique (SMOTE)
In this section, we outline the data processing and analysis steps un- The SMOTE algorithm, first proposed by [35], is an oversampling
dertaken in our study. method designed to address the issues associated with class imbalance
We initiate the data processing step by retaining features common by creating synthetic instances from the minority class. The algorithm
to both the ASHRAE Global Thermal Comfort Database II and the BCA begins by sampling the 𝑘 nearest neighbors of a instance in the minority
Testbed Dataset, and those features that offer valuable insights into oc- class 𝑋𝑗 and creating a new instance 𝑃𝑘 through a random interpola-
cupants’ thermal and air movement comfort conditions. These features tion process between 𝑋𝑗 and one of the nearest neighbors, denoted as

4
Z.D. Tekler, Y. Lei and A. Chong Energy & Buildings 319 (2024) 114507

Fig. 2. Label distribution of the comfort feedback for thermal acceptability, thermal preference, and air movement preference reported in the ASHRAE Global Thermal
Comfort Database II and the Singapore BCA Dataset. (For interpretation of the colors in the figure(s), the reader is referred to the web version of this article.)

𝑍𝑗 𝑘. This process can be mathematically represented using the follow- 3.4. Base model (CNNLSTM) architecture
ing equation [36]:
The base model architecture chosen for this study is based on the
𝑃𝑘 = 𝑋𝑗 + 𝑟𝑎𝑛𝑑𝑜𝑚(0, 1) ∗ (𝑍𝑘𝑗 −𝑋𝑗 ). (1) Convolutional Neural Network-Long Short-term Memory (CNNLSTM)
In essence, the SMOTE algorithm functions by creating new instances model. This decision is due to the model’s ability to effectively cap-
along the line segments connecting 𝑥𝑗 with the selected neighbor 𝑍𝑗 𝑘. ture the spatio-temporal relationship between different time-series vari-
Hence, the fundamental assumption underlying the algorithm is that the ables of the building’s indoor and outdoor environmental conditions to
instances lying between two neighboring instances in the minority class perform accurate comfort modeling [28]. Additionally, the model also
also belong to the same minority class. allows us to freeze different parts of the CNN and LSTM layers to facil-
Given that the ASHRAE Global Thermal Comfort Database II con- itate the implementation and evaluation of different transfer learning
tains a combination of both nominal and continuous variables, this strategies.
study adopted the SMOTE-Nominal Continuous algorithm, which ex- Comprising a convolutional layer, two LSTM layers, two dense lay-
tends upon the SMOTE algorithm, to perform oversampling of the mi- ers, and an output layer, the CNNLSTM model architecture is designed
nority class within the dataset. The oversampled version of the ASHRAE to capture the spatiotemporal characteristics of the building’s environ-
Global Thermal Comfort Database II is used for model training in the mental features 𝑋 to predict the most likely response in terms of the
subsequent sections of this study. occupants’ thermal preference, thermal acceptability, or air movement
preference 𝑦.
A forward pass into the CNNLSTM model involves passing an input
3.3. Problem formulation
feature tuple 𝑋 of dimension (7, 1) into a 1D-convolutional layer with
a filter size of 128 and kernel size of 5, to result in an output tensor of
The task of predicting occupants’ thermal preference, thermal ac- dimension (7, 128). In this case, the convolutional layer functions as a
ceptability, and air movement preference can be framed as a classifica- feature extractor, discerning the spatially invariant structures within in-
tion problem. This entails defining a probabilistic function that assigns put 𝑋 . To prevent overfitting, a dropout layer with a probability of 0.1
a probability score 𝑃 (𝑦|𝑋), indicating the likelihood that a sample be- is introduced after the 1D-convolutional layer. The spatial features ex-
longs to a specific class label 𝑦 given the sample’s input features 𝑋 . tracted from the convolutional layer are subsequently passed through
The input feature 𝑋 , in this case, represents a tuple containing the en- two LSTM layers with 256 neurons to learn the temporal structures of
vironmental features, while the class label 𝑦 depends on whether the the input data. A recurrent dropout probability of 0.1 is also applied to
model is predicting the occupants’ thermal preference 𝑦 ∈ No Change, each LSTM layer to prevent overfitting. Following this step, the output
Warmer, Cooler, thermal acceptability 𝑦 ∈ Acceptable, Unacceptable, or tensor from the second LSTM layer is flattened and directed into two
air movement preference 𝑦 ∈ No Change, More, Less. fully connected dense layers. These layers aimed to obtain a higher-
The probabilistic function 𝑓 (.) is obtained by learning from a train- order representation of the processed features and enhance the separa-
ing dataset 𝐷𝑡𝑟𝑎𝑖𝑛 before validating its performance on a test dataset bility of the instances into the different labels found within the training
𝐷𝑡𝑒𝑠𝑡 . In the case of a regular classification problem, both 𝐷𝑡𝑟𝑎𝑖𝑛 and dataset. Both dense layers employ the rectified linear unit (ReLU) ac-
𝐷𝑡𝑒𝑠𝑡 originate from the same domain 𝐷. However, in the context of tivation function and are initialized using the Glorot uniform weight
transfer learning, the problem extends to involve two separate domains: initialization method, with the first dense layer having 64 neurons and
the source domain 𝐷𝑠𝑜𝑢𝑟𝑐𝑒 and the target domain 𝐷𝑡𝑎𝑟𝑔𝑒𝑡 . the second layer containing 16 neurons.
The source domain 𝐷𝑠𝑜𝑢𝑟𝑐𝑒 is typically defined as the information- The output of the second dense layer is finally fed into the final
rich domain where a large number of labeled instances are readily avail- output layer with the same number of neurons as the number of user
able, while the information-poor target domain 𝐷𝑡𝑎𝑟𝑔𝑒𝑡 tends to contain response labels (i.e., 3 neurons for thermal preference and air move-
only a limited number of labeled instances due to the high cost of data ment preference, and 2 neurons for thermal acceptability). The output
collection and human labeling (i.e., |𝐷𝑠𝑜𝑢𝑟𝑐𝑒 | >> |𝐷𝑡𝑎𝑟𝑔𝑒𝑡 |). Therefore, layer utilizes a softmax activation function to yield a probability distri-
the motivation behind transductive transfer learning involves enhanc- bution of the target labels, where the predicted label 𝑦 is determined by
ing the predictive performance of the probabilistic function 𝑓 (.) in the identifying the label with the highest predicted probability scores.
target domain 𝐷𝑡𝑎𝑟𝑔𝑒𝑡 , by leveraging on the vast amount of labeled in- The training process of the base model architecture employs the
stances in the source domain 𝐷𝑠𝑜𝑢𝑟𝑐𝑒 , and applying that knowledge in Adam optimizer with a learning rate of 0.001 and utilizes the cate-
the target domain 𝐷𝑡𝑎𝑟𝑔𝑒𝑡 . gorical cross-entropy loss function to calculate the model’s prediction

5
Z.D. Tekler, Y. Lei and A. Chong Energy & Buildings 319 (2024) 114507

instead of updating the weights for all model layers during the back-
propagation step, only the weights of the final dense layers are updated,
while the convolutional and LSTM layer weights are frozen. This retrain-
ing approach essentially allows the model to retain its feature-extracting
abilities based on what was learned from the source domain.
Despite their differences, both models leverage the knowledge ob-
tained from the information-rich source domain and adapt it to an
information-poor target domain. This is achieved by improving the mod-
el’s ability to extract informative high-level representations from the
input features through parameter sharing in the pre-training step be-
fore enhancing the model’s predictive performance in the target domain
through retraining. Fig. 4 illustrates the model architectures of the TL-
CNNLSTM-FT and TL-CNNLSTM-FE models, respectively.

3.6. Active transfer learning framework

To further enhance the transfer learning model’s performance in both

data efficiency and predictive accuracy, we introduced the concept of
active learning during the model’s retraining step to form the active
transfer learning framework.
Specifically, within the model retraining step, an active learning
algorithm is applied to the pool of unlabeled instances in the target do-
main to identify the most informative instances for human annotation
and subsequent model retraining. The Query-By-Committee (QBC) sam-
pling algorithm was adopted for this study due to findings from a prior
Fig. 3. Model architecture of the base CNNLSTM model with model hyperpa-
rameters. study, which found that the QBC sampling algorithm was able to out-
perform other popular AL algorithms, such as Uncertainty Sampling, in
reducing human annotation cost when applied to a similar dataset [24].
error against the training data. The detailed architecture of the proposed
The QBC algorithm begins by defining a panel of classifiers, denoted
CNNLSTM model is illustrated in Fig. 3.
as 𝐶 = 𝜃 (1) , ..., 𝜃 (𝐶) , which are trained on random samples of the labeled
pool 𝑃𝑙𝑎𝑏𝑒𝑙𝑒𝑑 . By applying this panel of classifiers to the unlabeled pool
3.5. Transfer learning model architecture
𝑃𝑢𝑛𝑙𝑎𝑏𝑒𝑙𝑒𝑑 , the unlabeled instance with the highest level of disagreement
among the panel’s predictions is considered to be the most informative
Utilizing the base CNNLSTM model architecture described in the
instance and chosen for human labeling. The newly labeled instance
previous subsection, two transductive transfer learning model architec-
is subsequently removed from 𝑃𝑢𝑛𝑙𝑎𝑏𝑒𝑙𝑒𝑑 and incorporated into 𝑃𝑙𝑎𝑏𝑒𝑙𝑒𝑑
tures are developed for this study: transfer learning using fine-tuning
before retraining the panel of classifiers on the updated 𝑃𝑙𝑎𝑏𝑒𝑙𝑒𝑑 for the
(TL-CNNLSTM-FT) and transfer learning using feature extraction (TL-
next selection round. The vote entropy metric by [37] is employed in
CNNLSTM-FE).
this study to quantify the degree of disagreement among the panel. The
The training process of the TL-CNNLSTM-FT model begins by initial-
unlabeled instance 𝑋 ∗ is chosen for human labeling based on the results
izing a base CNNLSTM model, where the weights of each layer are ran-
of the following equation:
domly initialized using the Glorot uniform weight initialization method.
This model is subsequently trained on the large amount of labeled data
𝐿
∑
from the ASHRAE Global Thermal Comfort Database II (source domain), 𝑛(𝑦𝑖 ) 𝑛(𝑦𝑖 )
𝑋𝑄𝐵𝐶
∗
= argmax − 𝑙𝑜𝑔 (2)
where the weights of the CNNLSTM layers are iteratively updated over 𝑋 𝑖
|𝐶| |𝐶|
100 epochs through backpropagation. This process gradually reduces
where |𝐶| represents the number of classifiers in the panel, 𝐿 represents
the model’s prediction error in inferring the occupants’ state of ther-
the number of class labels, and 𝑛(𝑦𝑖 ) denotes the number of classifiers
mal comfort (i.e., thermal preferences, thermal acceptability, and air
categorizing instance 𝑋 under class 𝑦𝑖 .
movement preferences), based on indoor and outdoor environmental
Within our active transfer learning framework, we defined the
conditions. Once the model has been pre-trained on the ASHRAE Global
Thermal Comfort Database II (source domain), the TL-CNNLSTM-FT ASHRAE Global Thermal Comfort Database II as our initial 𝑃𝑙𝑎𝑏𝑒𝑙𝑒𝑑 ,
model is adapted to perform a similar task on the BCA Testbed Dataset where a panel of classifiers was initially trained. Subsequently, we eval-
(target domain) by retraining the model on small amounts of labeled uated the unlabeled instances in the BCA Testbed Dataset, defined as
data from the BCA dataset. During this retraining process, the weights our unlabeled pool 𝑃𝑢𝑛𝑙𝑎𝑏𝑒𝑙𝑒𝑑 . By identifying the most informative in-
of all layers are updated during the backpropagation process to fine-tune stance in 𝑃𝑢𝑛𝑙𝑎𝑏𝑒𝑙𝑒𝑑 using the above equation, the instance was chosen
and improve the model’s predictive performance in the target domain. for human labeling and included in 𝑃𝑙𝑎𝑏𝑒𝑙𝑒𝑑 for panel retraining and the
Given that the predictive task between the source and target domain next evaluation round.
is identical, the final output layer of the pretrained TL-CNNLSTM-FT The incorporation of active learning within the transfer learning
model was retained during the retraining process, instead of being re- framework enables us to focus the model retraining process on the most
placed with new dense layers. distinctive instances (i.e., most informative instances) in the target do-
The TL-CNNLSTM-FE model follows a similar training process as the main compared to the source domain, thereby enhancing the model’s
TL-CNNLSTM-FT model by first pre-training on the labeled data in the transferability in the new domain. Fig. 5 illustrates the model archi-
ASHRAE Global Thermal Comfort Database II (source domain), before tecture of the proposed active transfer learning framework based on
retraining on the BCA Testbed Dataset (target domain). However, the the TL-CNNLSTM-FT and TL-CNNLSTM-FE model architectures (ATL-
key difference between both models lies in the retraining process, where CNNLSTM-FT and ATL-CNNLSTM-FE).

6
Z.D. Tekler, Y. Lei and A. Chong Energy & Buildings 319 (2024) 114507

Fig. 4. Model architectures of the TL-CNNLSTM-FT and TL-CNNLSTM-FE with hyperparameters.

Fig. 5. Model architecture of the Active Transfer Learning Framework, including ATL-CNNLSTM-FT and ATL-CNNLSTM-FT with models hyperparameters.

7
Z.D. Tekler, Y. Lei and A. Chong Energy & Buildings 319 (2024) 114507

Table 2 model performs across all label classes, particularly for the minority
Details of model parameters used during implementation. class, it provides a more holistic assessment of the model’s overall per-
Model parameters Value formance. By comparing this metric with the model’s reported accuracy
score, we could also determine if the model is biased towards the ma-
Convolution Layer Filter Size: 128
Kernel Size: 5
jority class based on the difference between both metrics.
The micro-average F1 score is calculated based on the harmonic
Dropout Layer Dropout Probability: 0.1
mean of the precision 𝑃 𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛𝑖 and recall 𝑅𝑒𝑐𝑎𝑙𝑙𝑖 scores for each la-
LSTM Layer Number of Neurons: 256
Recurrent Dropout Probability: 0.1
bel class 𝑖 as shown in the following equations.
∑𝑁
𝑖=1 𝑇 𝑟𝑢𝑒𝑃 𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑖𝑖
Dense Layer Dense Layer 1: 64
Neurons Dense Layer 2: 16 Precision𝑖 = ∑ ∑ (4)
𝑁
Neurons Activation Function: ReLU 𝑖=1 𝑇 𝑟𝑢𝑒𝑃 𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑖𝑖 + 𝑗≠𝑖 𝐹 𝑎𝑙𝑠𝑒𝑃 𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑖𝑗
Weight Initializer: Glorot uniform weight initialization ∑𝑁
Output Layer Number of Neurons: 2 (thermal acceptability) or 3 𝑖=1 𝑇 𝑟𝑢𝑒𝑃 𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑖𝑖
Recall𝑖 = ∑ ∑ (5)
𝑁
𝑖=1 𝑇 𝑟𝑢𝑒𝑃 𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑖𝑖 + 𝑗≠𝑖 𝐹 𝑎𝑙𝑠𝑒𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑖𝑗
(thermal and air movement preference) depending on
the classification task
∑
Activation Function: Softmax 2⋅ 𝑁 𝑖=1 Precision𝑖 ⋅ Recall𝑖
Micro-average F1 = ∑ (6)
Optimization Function Adam Optimizer with learning rate of 0.001 𝑁
𝑖=1 (Precision𝑖 + Recall𝑖 )
Loss Function Categorical Cross Entropy
Batch Size 128 4.3. Impact of transfer learning approaches for personal comfort models
Epochs Pre-training Step: 100 Epochs
Retraining Step: 30 Epochs with early stopping enabled To evaluate the influence of transfer learning on the model’s pre-
and an evaluation window of 10 epochs dictive performance, we compared a base CNNLSTM model against
the TL-CNNLSTM-FT and TL-CNNLSTM-FE models in predicting the oc-
cupants’ thermal preference, thermal acceptability, and air movement
4. Results and discussion
preferences, as depicted in Table 3. The primary distinction between the
base CNNLSTM model and the TL-CNNLSTM-FT and TL-CNNLSTM-FE
4.1. Model implementation
models lies in their training data, where the base model is trained solely
on the BCA Testbed Dataset (target domain), while the latter models
The models outlined in this study were implemented using the
undergo pre-training on the ASHRAE database (source domain) before
Python programming language and the TensorFlow Keras library. Addi-
retraining on the BCA dataset.
tionally, the hyperparameters chosen for model implementation are de-
Based on the empirical results presented in Table 3, both transfer
termined by experimenting with different hyperparameters on the mod-
el’s predictive performance and findings from past studies that adopted learning models (TL-CNNLSTM-FT and TL-CNNLSTM-FE) consistently
the same model architecture. A summary of the model parameters and outperformed the base CNNLSTM model in terms of accuracy and
hyperparameters is described in Table 2. micro-average F1 score. Specifically, the transfer learning models out-
During the model training process, the transfer learning models are performed the base CNNLSTM model by 1.3%-3.4% in terms of accuracy
pre-trained on 100% of the labeled data from the ASHRAE Global Ther- and 2%-3.5% for micro-average F1 score when predicting the occupants’
mal Comfort Database II. They are subsequently retrained on a limited thermal preference, thermal acceptability, and air movement prefer-
number of labeled data from the BCA Testbed Dataset before being eval- ence. These results clearly underscore the value of transfer learning,
uated on a holdout set from the BCA dataset. The amount of labeled data where the model was able to leverage insights from the information-
used during the retraining step varies based on the experiments con- rich ASHRAE database during the pre-training step and adapt to the
ducted, which are described in the following subsections. The holdout information-poor BCA dataset during the retraining step, ultimately
set is obtained by randomly setting aside 10 labeled instances from each enhancing its predictive performance in the target domain. The im-
study participant from the BCA testbed to ensure equal representation of plication of this finding is that practitioners can effectively use transfer
each participant when evaluating the models’ predictive performance. learning to improve the performance of existing models trained under
information-poor conditions without collecting additional data in the
4.2. Evaluation metrics target building. Additionally, when comparing the model’s predictive
performance when using different transfer learning approaches, both
Two evaluation metrics were chosen for this study: accuracy and TL-CNNLSTM-FE and TL-CNNLSTM-FT exhibited similar performances,
micro-average F1 score. with the TL-CNNLSTM-FT model performing marginally better in terms
Accuracy is a popular classification evaluation metric used in past of model accuracy, while the TL-CNNLSTM-FE model performed con-
thermal comfort studies and is defined as the ratio of true positive sistently better when considering micro-average F1 score. Lastly, given
instances for each label class against the total number of predicted that the reported accuracy and micro-average f1 scores reported in Ta-
instances across all label classes. The accuracy formula is formally ex- ble 3 are comparable for different comfort tasks and conditions, this
pressed as: result provides strong evidence that the models are also not biased to-
∑𝑁 wards the majority class.
𝑇 𝑟𝑢𝑒𝑃 𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑖𝑖
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = ∑𝑁𝑖=1∑𝑁 (3)
4.4. Impact of active learning on transfer learning models’ performance
𝑖=1 𝑗=1 𝑃 𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑 𝑖𝑗

In this equation, 𝑁 represents the number of label classes, This section investigates the impact of incorporating active learning
𝑇 𝑟𝑢𝑒𝑃 𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑖𝑖 is the number of correctly predicted instances for class into the transfer learning framework through two experiments:
𝑖, and 𝑃 𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑𝑖𝑗 is the number of instances predicted as class 𝑗 .
In addition, given the significant imbalance between the occupants’ Experiment 1 (Random selection, 9.8% of data): The first experi-
reported comfort labels in the BCA Testbed Dataset, as highlighted in ment involves randomly selecting two instances per partic-
Section 3.3, the micro-average F1 score was also adopted in this study ipant from the target domain, which constitutes 9.8% of the
to evaluate the comfort model’s predictive performance across all la- total training data. Specifically, one instance was collected
bel classes. As this evaluation metric is sensitive towards how well the under AC (air-conditioning) and the other under NV (natural

8
Z.D. Tekler, Y. Lei and A. Chong Energy & Buildings 319 (2024) 114507

Table 3
Model performance comparison of different Transfer learning models against the Base model (CNNLSTM)
for thermal preference, thermal acceptability, and air movement preference. The reported accuracy and
micro-average f1 scores are averaged across all respondents in the target domain.

Comfort Task Model Source Target % of Training Accuracy Micro-average

Data from Target F1 Score

Thermal CNNLSTM (Base) N.A. BCA 100% 0.828 0.800

Acceptability TL-CNNLSTM-FE ASHRAE 0.832 0.818
TL-CNNLSTM-FT ASHRAE 0.839 0.813

Thermal CNNLSTM (Base) N.A. BCA 100% 0.675 0.636

Preference TL-CNNLSTM-FE ASHRAE 0.698 0.658
TL-CNNLSTM-FT ASHRAE 0.693 0.654

Air Movement CNNLSTM (Base) N.A. BCA 100% 0.619 0.602

Preference TL-CNNLSTM-FE ASHRAE 0.630 0.614
TL-CNNLSTM-FT ASHRAE 0.633 0.593

Table 4
Model performance comparison with and without active learning for thermal preference, thermal acceptability, and air movement
preference. The reported accuracy and micro-average f1 scores are averaged across all respondents in the target domain.

Comfort Task Approach (Random Model Source Target % of Training Accuracy Micro-average
vs Active Learning) Data from Target F1 Score

Thermal Random CNNLSTM (Base Model) N.A. BCA 9.8% 0.782 0.790
Acceptability TL-CNNLSTM-FE ASHRAE 9.8% 0.809 0.806
TL-CNNLSTM-FT ASHRAE 9.8% 0.809 0.777
Active Learning ACNNLSTM (Base Model) N.A. BCA 9.8% 0.789 0.759
ATL-CNNLSTM-FE ASHRAE 9.8% 0.837 0.814
ATL-CNNLSTM-FT ASHRAE 9.8% 0.823 0.804

Thermal Random CNNLSTM (Base Model) N.A. BCA 9.8% 0.587 0.568
Preference TL-CNNLSTM-FE ASHRAE 9.8% 0.630 0.604
TL-CNNLSTM-FT ASHRAE 9.8% 0.619 0.594
Active Learning ACNNLSTM (Base Model) N.A. BCA 9.8% 0.632 0.597
ATL-CNNLSTM-FE ASHRAE 9.8% 0.654 0.633
ATL-CNNLSTM-FT ASHRAE 9.8% 0.668 0.636

Air Movement Random CNNLSTM (Base Model) N.A. BCA 9.8% 0.530 0.496
Preference TL-CNNLSTM-FE ASHRAE 9.8% 0.582 0.563
TL-CNNLSTM-FT ASHRAE 9.8% 0.579 0.550
Active Learning ACNNLSTM (Base Model) N.A. BCA 9.8% 0.523 0.435
ATL-CNNLSTM-FE ASHRAE 9.8% 0.616 0.574
ATL-CNNLSTM-FT ASHRAE 9.8% 0.614 0.571

ventilation). The two instances were then utilized as training active learning in improving the predictive performance of the trans-
data for the retraining step. fer learning models by optimizing the retraining process and selecting
Experiment 2 (Active learning, 9.8% of data): The second experi- the most informative instances in the information-poor target domain
ment involves using active learning (QBC sampling algorithm) for model retraining. In fact, the results show that by combining active
to select the two most informative instances, also constitut- learning with transfer learning, we were able to obtain a model per-
ing 9.8% of the total training data. Similarly, one instance was formance comparable to another model trained on more than 10 times
collected under AC (air-conditioning) and the other under NV the existing training data, significantly improving the model’s data ef-
(natural ventilation). The two instances were then utilized as ficiency. Lastly, given that the reported accuracy and micro-average f1
training data for the retraining step. scores reported in Table 4 are comparable for different comfort tasks
and conditions, this result provides strong evidence that the models are
By pretraining the CNNLSTM model on the ASHRAE database and not biased towards the majority class.
retraining it on the selected instances from the target domain, the
results for both experiments are presented in Table 4. According to 4.5. Impact of different amounts of training data in target domain on
the empirical results presented in Table 4, the active transfer learning predictive performance
models (ATL-CNNLSTM-FT and ATL-CNNLSTM-FE) consistently outper-
formed their transfer-learning-only counterparts (TL-CNNLSTM-FT and After evaluating the impact of transfer learning and active learning
TL-CNNLSTM-FE) both in terms of model accuracy and micro-average on the model’s predictive performance, this subsection aims to evaluate
F1 score. More specifically, the active transfer learning models were able how the model performs with different amounts of training data in the
to outperform the transfer-learning-only models by 1.7%-7.9% in terms target domain for the model retraining step. By evaluating the models’
of accuracy and 1.0%-7.1% for micro-average F1 score when predicting performance when using varying amounts of training data from the tar-
the occupants’ thermal preference, thermal acceptability, and air move- get domain (i.e., 0%, 9.8% ≈ 2 instances/data-points per participant -
ment preference. Notably, the improvement in predictive performance representing one instance for each AC and NV mode, 36% ≈ 8 instances
after incorporating active learning is consistently observed not only in per participant - representing one instance per condition out of 8 given
both transfer learning models (TL-CNNLSTM-FE and TL-CNNLSTM-FT) in Table 1, and 100%), the results are generated and visualized in Fig. 6.
but also from the base CNNLSTM models without transfer learning, One of the observations from Fig. 6 indicates that the predictive per-
thereby demonstrating its generalizability to different model architec- formances of all models generally follow a positive trend as the models
tures. These findings provide sufficient evidence of the effectiveness of are retrained on an increasing amount of training data from the target

9
Z.D. Tekler, Y. Lei and A. Chong Energy & Buildings 319 (2024) 114507

Another notable observation arises from comparing the predic-

tive performance of the base CNNLSTM model with that of the TL-
CNNLSTM-FT and TL-CNNLSTM-FE models. It can be observed that the
most significant improvement from the application of transfer learning
occurs when very limited training data from the target domain is avail-
able for model retraining. This improvement gap gradually declines as
more training data from the target domain becomes available. This re-
sult aligns with our intuition as the transfer-learned models were able
to effectively bridge their knowledge gap in the information-poor target
domain based on the knowledge gained in the source domain.
Finally, based on the performance boost resulting from the integra-
tion of active learning within the transfer learning framework, we can
observe that ATL-based models with less than 10% of the training data
were able to make predictions almost as accurately as the base CNNL-
STM model trained with 100% of the data. In fact, in the case of thermal
acceptability, ATL models with less than 10% of training data were ob-
served in Fig. 6 to outperform traditional CNNLSTM models with 100%
training data. Lastly, in terms of the ATL-based model’s performance
across different participants, the reported accuracy scores for all 58 re-
spondents have a standard deviation of less than 0.16 across different
comfort tasks, thereby demonstrating the personalized nature of the re-
sulting model. This concludes that our proposed active transfer learning
strategy was effective in improving the data efficiency of personal com-
fort models while maintaining adequate predictive performance.

5. Conclusion

We introduced a novel active transfer learning framework that inte-

grates active learning and transfer learning to enhance the performance
of personal comfort models in scenarios with limited data. The pro-
posed framework was evaluated across three comfort tasks — thermal
preference, thermal acceptability, and air movement preference — by
leveraging knowledge from the extensive ASHRAE Global Thermal Com-
fort Database II and applying it to a smaller comfort dataset collected
from the BCA testbed in Singapore.
Through a comprehensive evaluation of the comfort models de-
veloped using the proposed active transfer learning framework, the
empirical results obtained led to multiple findings:
Firstly, the transfer-learning-enhanced models, namely TL-CNN-
LSTM-FT and TL-CNNLSTM-FE, were able to consistently outperform the
base CNNLSTM model in terms of accuracy and micro-average F1 score.
Specifically, these models exhibited improvements ranging from 1.3%-
3.4% in accuracy and 2%-3.5% in micro-average F1 score when predict-
ing the occupants’ thermal preference, thermal acceptability, and air
Fig. 6. Model performances of the base, transfer learning, and active trans-
movement preference. Furthermore, when evaluating the models’ per-
fer learning models after retraining on different amounts of data in the target
domain for all comfort tasks. The reported accuracy scores on the Y-axis are av-
formance using increasing amounts of training data, the most substantial
eraged results across all respondents in the target domain. improvement through transfer learning occurred when very limited
training data from the target domain was available for model retraining,
with diminishing returns observed as more data became accessible.
domain. However, this trend does not follow a linear pattern but instead Secondly, the active transfer learning models, ATL-CNNLSTM-FT
gradually converges, signifying diminishing returns in model improve- and ATL-CNNLSTM-FE, were able to further outperform their transfer-
ment as more data from the target domain is available for retraining. learning-only counterparts (TL-CNNLSTM-FT and TL-CNNLSTM-FE),
This observation is likely due to the model having already captured the achieving a boost of 1.7%-7.9% in accuracy and 1.0%-7.1% in micro-
existing patterns within the dataset such that there are no additional average F1 score. This result clearly demonstrated the effectiveness of
patterns to learn from the new incoming data. This is especially true combining active learning with transfer learning approaches in improv-
when it comes to modeling personal comfort as there is a physiologi- ing the occupants’ personal comfort predictions.
cal limit to the human body’s ability to adapt to external environmental Lastly, an intriguing finding emerged from comparing ATL-based
conditions. As such, the application of active learning is particularly ef- models (ATL-CNNLSTM-FT and ATL-CNNLSTM-FE) and the base CNNL-
fective in these cases as it helps to identify the most informative data STM model. The proposed active transfer learning models were able
points for learning these patterns. An implication of this result is that to make predictions almost as accurate as the base CNNLSTM model,
researchers conducting data collection studies to gather occupants’ ther- using only less than 10% of the training data compared to the base
mal comfort preferences should carefully balance their data collection CNNLSTM model trained with 100% of the data for all comfort tasks.
budget with the marginal improvements from their models. This ensures This concludes that active transfer learning was an effective strategy for
optimal utilization of their limited budget to achieve greater gains in developing data-efficient personal comfort models using limited data
model performance. while achieving sufficient predictive performance.

10
Z.D. Tekler, Y. Lei and A. Chong Energy & Buildings 319 (2024) 114507

While our active transfer learning framework proves to be a valu- [5] Z.D. Tekler, R. Low, C. Yuen, L. Blessing, Plug-mate: an iot-based occupancy-
able strategy for enhancing personal comfort models under limited data driven plug load management system in smart buildings, Build. Environ. 223 (2022)
109472.
availability, it is also important to highlight some of the limitations of
[6] Y. Lei, S. Zhan, E. Ono, Y. Peng, Z. Zhang, T. Hasama, A. Chong, A practical deep
this study that will be addressed in future works. Firstly, given that the reinforcement learning framework for multivariate occupant-centric control in build-
proposed active transfer learning framework is evaluated on a single ings, Appl. Energy 324 (2022) 119742, https://doi.org/10.1016/j.apenergy.2022.
target building in the tropics, there is a need to apply the framework 119742.
across different target buildings from a range of climate zones to ensure [7] D. Zhuang, V.J. Gan, Z.D. Tekler, A. Chong, S. Tian, X. Shi, Data-driven predictive
control for smart hvac system in iot-integrated buildings with time-series forecasting
the generalizability of the study’s findings. Additionally, given that the
and reinforcement learning, Appl. Energy 338 (2023) 120936.
adoption of active learning would inevitably increase the computation [8] E. Ono, K. Mihara, K.P. Lam, A. Chong, The effects of a mismatch between thermal
cost during the model training process, there is a need for practition- comfort modeling and hvac controls from an occupancy perspective, Build. Environ.
ers to weigh the costs associated with data collection and the additional 220 (2022) 109255.
computation costs when deciding if the active transfer learning frame- [9] H. Zhang, A. Tzempelikos, X. Liu, S. Lee, F. Cappelletti, A. Gasparella, The impact of
personal preference-based thermal control on energy use and thermal comfort: field
work is appropriate for their use case. Lastly, other potential avenues for
implementation, Energy Build. 284 (2023) 112848.
future research can also include devising more intelligent approaches [10] J. Kim, Y. Zhou, S. Schiavon, P. Raftery, G. Brager, Personal comfort models: pre-
for selecting relevant training data from the ASHRAE Global Database dicting individuals’ thermal preference using occupant heating and cooling behavior
(i.e., source domain) during the initial training phase of transfer learn- and machine learning, Build. Environ. 129 (2018) 96–106.
ing to further enhance the predictive performance of the resulting model [11] F. Tartarini, S. Schiavon, M. Quintana, C. Miller, Personal comfort models based on
a 6-month experiment using environmental parameters and data from wearables,
when applied to the target domain.
Indoor Air 32 (11) (2022) e13160, https://doi.org/10.1111/ina.13160.
To sum up, our active transfer learning framework proves to be a [12] M. Luo, J. Xie, Y. Yan, Z. Ke, P. Yu, Z. Wang, J. Zhang, Comparing machine learning
valuable strategy for enhancing personal comfort models, offering su- algorithms in predicting thermal sensation using ashrae comfort database ii, Energy
perior predictive performance even in scenarios with limited data avail- Build. 210 (2020) 109776.
ability. This study opens avenues for future research in data-efficient [13] Q.Y. Li, J. Han, L. Lu, A random forest classification algorithm based personal
thermal sensation model for personalized conditioning system in office buildings,
comfort modeling for practical applications.
Comput. J. 64 (3) (2021) 500–508.
[14] Y. Wu, J. Fan, B. Cao, Comparison among different modeling approaches for per-
CRediT authorship contribution statement sonalized thermal comfort prediction when using personal comfort systems, Energy
Build. 285 (2023) 112873.
[15] Q. Chai, H. Wang, Y. Zhai, L. Yang, Using machine learning algorithms to predict oc-
Zeynep Duygu Tekler: Writing – review & editing, Writing – origi-
cupants’ thermal comfort in naturally ventilated residential buildings, Energy Build.
nal draft, Visualization, Methodology, Investigation, Conceptualization. 217 (2020) 109937.
Yue Lei: Writing – review & editing, Data curation. Adrian Chong: [16] D. Li, C.C. Menassa, V.R. Kamat, Personalized human comfort in indoor building en-
Writing – review & editing, Supervision, Resources, Project administra- vironments under diverse conditioning modes, Build. Environ. 126 (2017) 304–317.
tion, Funding acquisition, Conceptualization. [17] J.-H. Choi, D. Yeom, Study of data-driven thermal sensation prediction model as a
function of local body skin temperatures in a built environment, Build. Environ. 121
(2017) 130–147.
Declaration of competing interest [18] T.C. Cheung, S. Schiavon, E.T. Gall, M. Jin, W.W. Nazaroff, Longitudinal assessment
of thermal and perceived air quality acceptability in relation to temperature, humid-
The authors declare that they have no known competing financial ity, and co2 exposure in Singapore, Build. Environ. 115 (2017) 80–90.
interests or personal relationships that could have appeared to influence [19] C. Shan, J. Hu, J. Wu, A. Zhang, G. Ding, L.X. Xu, Towards non-intrusive and high
accuracy prediction of personal thermal comfort using a few sensitive physiological
the work reported in this paper.
parameters, Energy Build. 207 (2020) 109594.
[20] Y. Feng, S. Liu, J. Wang, J. Yang, Y.-L. Jao, N. Wang, Data-driven personal thermal
Data availability comfort prediction: a literature review, Renew. Sustain. Energy Rev. 161 (2022)
112357.
[21] P. Jayathissa, M. Quintana, M. Abdelrahman, C. Miller, Humans-as-a-sensor for
The research compendium for this article is available at https://
buildings—intensive longitudinal indoor comfort models, Buildings 10 (10) (2020)
github.com/ideas-lab-nus/ComfortPredict-ActiveTL, hosted on GitHub. 174.
[22] Y. Lei, Z.D. Tekler, S. Zhan, C. Miller, A. Chong, Experimental evaluation of ther-
Acknowledgements mal adaptation and transient thermal comfort in a tropical mixed-mode ventilation
context, Build. Environ. 248 (2024) 111043.
[23] J. Ngarambe, G.Y. Yun, M. Santamouris, The use of artificial intelligence (ai) meth-
The National Research Foundation, Singapore, and the Ministry of
ods in the prediction of thermal comfort in buildings: energy implications of ai-based
National Development, Singapore, provided support for this research thermal comfort controls, Energy Build. 211 (2020) 109807.
project through the Cities of Tomorrow R&D Programme (CoT Award [24] Z.D. Tekler, Y. Lei, Y. Peng, C. Miller, A. Chong, A hybrid active learning framework
COT-V4-2020-5). The views, opinions, findings, conclusions, or recom- for personal thermal comfort models, Build. Environ. 234 (2023) 110148.
mendations expressed in this material are solely those of the author(s) [25] Z.D. Tekler, A. Chong, Occupancy prediction using deep learning approaches across
multiple space types: a minimum sensing strategy, Build. Environ. 226 (2022)
and do not reflect the opinions of these funding agencies. 109689.
[26] Z.D. Tekler, Y. Lei, X. Dai, A. Chong, Enhancing Personalised Thermal Comfort Mod-
References els with Active Learning for Improved hvac Controls, Journal of Physics: Conference
Series, vol. 2600, IOP Publishing, 2023, p. 132004.
[1] ASHRAE, Standard 55, Thermal Environmental Conditions for Human Occupancy, [27] N. Gao, W. Shao, M.S. Rahaman, J. Zhai, K. David, F.D. Salim, Transfer learning for
American Society of Heating, Ventilating, and Air Conditioning Engineers, Atlanta, thermal comfort prediction in multiple cities, Build. Environ. 195 (2021) 107725.
Georgia, 2023. [28] N. Somu, A. Sriram, A. Kowli, K. Ramamritham, A hybrid deep transfer learning strat-
[2] J. Kim, S. Schiavon, G. Brager, Personal comfort models–a new paradigm in ther- egy for thermal comfort prediction in buildings, Build. Environ. 204 (2021) 108133.
mal comfort for occupant-centric environmental control, Build. Environ. 132 (2018) [29] H. Park, D.Y. Park, Prediction of individual thermal comfort based on ensemble
114–124. transfer learning method using wearable and environmental sensors, Build. Envi-
[3] W. O’Brien, A. Wagner, M. Schweiker, A. Mahdavi, J. Day, M.B. Kjærgaard, S. Car- ron. 207 (2022) 108492.
lucci, B. Dong, F. Tahmasebi, D. Yan, et al., Introducing iea ebc annex 79: key [30] K. Li, Y. Liu, L. Chen, W. Xue, Data efficient indoor thermal comfort prediction using
challenges and opportunities in the field of occupant-centric building design and instance based transfer learning method, Energy Build. (2024) 113920.
operation, Build. Environ. 178 (2020) 106738, https://doi.org/10.1016/j.buildenv. [31] A. Natarajan, E. Laftchiev, A transfer active learning framework to predict thermal
2020.106738. comfort, Int. J. Progn. Health Manag. 10 (3) (2019).
[4] J. Xiong, A. Tzempelikos, I. Bilionis, P. Karava, A personalized daylighting control [32] V.F. Ličina, T. Cheung, H. Zhang, R. De Dear, T. Parkinson, E. Arens, C. Chun, S.
approach to dynamically optimize visual satisfaction and lighting energy use, Energy Schiavon, M. Luo, G. Brager, et al., Development of the ashrae global thermal comfort
Build. 193 (2019) 111–126. database ii, Build. Environ. 142 (2018) 502–512.

11
Z.D. Tekler, Y. Lei and A. Chong Energy & Buildings 319 (2024) 114507

[33] R. Low, Z.D. Tekler, L. Cheah, Predicting commercial vehicle parking duration using [36] R. Low, L. Cheah, L. You, Commercial vehicle activity prediction with imbalanced
generative adversarial multiple imputation networks, Transp. Res. Rec. 2674 (9) class distribution using a hybrid sampling and gradient boosting approach, IEEE
(2020) 820–831. Trans. Intell. Transp. Syst. 22 (3) (2020) 1401–1410.
[34] Z.D. Tekler, E. Ono, Y. Peng, S. Zhan, B. Lasternas, A. Chong, Robod, Room-Level [37] I. Dagan, S.P. Engelson, Committee-based sampling for training probabilistic classi-
Occupancy and Building Operation Dataset, Building Simulation, vol. 15, Springer, fiers, in: Machine Learning Proceedings 1995, Elsevier, 1995, pp. 150–157.
2022, pp. 2127–2137.
[35] N.V. Chawla, K.W. Bowyer, L.O. Hall, W.P. Kegelmeyer, Smote: synthetic minority
over-sampling technique, J. Artif. Intell. Res. 16 (2002) 321–357, https://doi.org/
10.1613/jair.953.

12