Review of Data-Driven Personal Thermal Comfort Modeling and Its Integration into Building Environment Control

Abstract

With the increasingly prominent demand for building energy efficiency and occupant-centric design, accurate and reliable personal thermal comfort models (PTCMs) are playing an important role in various residential and energy applications (e.g., building energy-saving design, indoor environmental regulation, and health and well-being improvement). In recent years, data-driven and artificial intelligence (AI) technologies have attracted considerable attention in the field of personal thermal comfort modeling. This study systematically reviews recent progress in data-driven personal thermal comfort modeling, emphasizing contact-based and non-contact data collection ways, correlation analysis of feature data, modeling methods based on machine learning and deep learning. Considering the high cost and limited scale of collection experiments, as well as noise, ambiguity, and individual differences in subjective feedback, special attention is put on the data-efficient thermal comfort modeling in data scarcity scenarios using a transfer learning (TL) strategy. Characteristics and suitable occasions of four transfer methods (model-based, instance-based, feature-based, and ensemble methods) are summarized to provide a deep perspective for practical applications. Furthermore, integration of PTCM into building environment control is summarized from aspects of the integration framework, modeling method, control strategy, actuator, and control effect. Performance of the integrated systems is analyzed in terms of improving personal thermal comfort and promoting building energy efficiency. Finally, several challenges faced by PTCMs and future directions are discussed.

1. Introduction

People spend approximately 80% to 90% of their time in various indoor environments of buildings [1], such as residences, office buildings, shopping malls, and hospitals, etc. One of the main functions of buildings is to provide a comfortable indoor environment, as it can affect the physical health, emotional state, and productivity of occupants. Thermal comfort, defined as “condition of mind which expresses satisfaction with the thermal environment” [2], has been a hot topic in indoor environment research. Thermal discomfort in the indoor environment may cause irritability, reduced attention, poor working productivity, and even exacerbate the symptoms of some chronic diseases [3,4].

Table 1. Highly cited papers on thermal comfort, from database of Web of Science Core Collection.

Reference (Year)	Journal	Main Contributions
[5] (2019)	Building and Environment	Assess PMV-PPD accuracy using ASHRAE Global Thermal Comfort Database II. Summarize that PMV can predict thermal sensation correctly only one out of three times, and PPD is unable to predict the dissatisfaction rate. Conclude that PMV-PPD accuracy varies strongly between ventilation, building types and climate.
[6] (2022)	Sustainable Cities and Society	Evaluate outdoor thermal comfort in an urban park in a hot-summer and cold-winter region of China (Chengdu). Apply machine learning to determine the relationship between thermal sensation vote and meteorological factors. Conclude that the effect of landscape spaces on human thermal comfort varies across the season.
[7] (2023)	Applied Energy	Present a data-driven predictive control method for smart HVAC operations. Develop and validate 16 LSTM models with bi-directional processing, convolution, and attention mechanisms. Integrate the optimal model with a reinforcement learning agent to analyse sensor metadata and optimise the HVAC system. Improve 17.4% energy efficiency and 16% thermal comfort in IoT-enabled smart building.
[8] (2021)	Journal of Cleaner Production	Summarize that personal thermal comfort monitoring relies on connected sensors, recording and wearable devices. Obtain that readings from sensing devices are highly sensitive on specific position on human body. Conclude that skin temperature and metabolic rate are direct indicators of personal thermal comfort. Point out that data processing methods request further investigation towards proper categorization.
[9] (2021)	Renewable and Sustainable Energy Reviews	Provide the first systematic review of thermal comfort with individual interactions into control loop. Present a holistic view of the complexities of delivering thermal comfort in buildings in an energy efficient way. Point out that AI implementation in building industry is still an ongoing endeavor. Discuss research challenges faced by AI-based modeling in buildings due to the lack of data.

summarizes highly cited papers (ranked in the top 1% globally in terms of citation frequency among papers in the same year and ESI subject category) on thermal comfort in recent years, including evaluation of traditional predicted mean vote and predicted percentage dissatisfied (PMV-PPD) models, data-driven predictive control for thermal comfort and energy saving, development trends in intelligent building control, etc.

1.1. Thermal Comfort Models

Thermal comfort is easily influenced by multiple factors (e.g., environmental factors, subjective feelings of occupants, individual adaptation mechanisms), which brings significant challenges to its precise modeling and assessment [10]. The PMV and adaptive models are two typical thermal comfort models. The PMV model, proposed by Fanger [11], was based on laboratory studies and underpinned by heat balance theory between the human body and metabolic heat production, assuming thermal comfort occurs when the body reaches equilibrium with the environment and treating occupant responses as largely passive. The PMV model was established with six parameters: air temperature, relative humidity, air speed, mean radiant temperature, metabolic rate, and clothing insulation. The PMV model employed the 7-point scale of the ASHRAE 55-2004 standard. It is typically used to assess the average thermal sensation of the population in a steady-state thermal environment.

Different from the PMV model that considers thermal comfort as a consequence of the heat balance between the human body and surrounding environment, the adaptive model treats occupants as active participants to maintain thermal preferences by behavioral, physiological, and psychological adaptations [12]. Within this framework, adaptive models typically present a linear relationship between acceptable indoor operative temperature and outdoor temperature, with such relationships derived from field studies. This kind of model takes into account the human inherent ability to adapt to variable environmental conditions in naturally ventilated buildings. For examples, Nicol and Humphreys [13] proposed the EN 15251 adaptive model, suggesting that occupants in real environments do not just passively accept the thermal environment, but actively adapt to external conditions through behavioral adjustments (e.g., opening windows, changing clothes, and adjusting fans), physiological adjustments (such as sweat regulation), and psychological adjustments. Yao et al. [14] introduced the adaptive coefficient to quantify the dynamic influence of factors such as culture, climate, and social background on thermal comfort. Zhou et al. [15] significantly improved the prediction accuracy of the PMV model in low-pressure environments by modifying the metabolic rate, evaporative heat dissipation, and convective heat transfer coefficient at low pressure. PMV and adaptive models have been developed and successfully adopted by international standards to provide the acceptable indoor environment for occupants. However, these models still have the following inherent limitations [16,17]: (1) They both exhibit poor prediction performance when applied to individuals since they are designed to predict the average thermal comfort of populations. (2) They rely on complex input parameters that are difficult to obtain in real time, limiting their practical applications. (3) Due to their fixed frameworks, it is difficult to integrate new variables (e.g., physiological characteristics and behavioral data) or adapt to complex dynamic environments, restricting further model optimization.

Personal thermal comfort model (PTCM) can predict the individual’s thermal comfort response, rather than the average response of populations. The key characteristics of PTCMs lie in the following aspects [16]: (1) Use individuals rather than groups as analytical units. (2) Train the model by combining individual direct feedback with other related data, such as environmental, physiological, and behavioral data. (3) Prioritize easily accessible and cost controllable data. (4) Have the capability to adapt by incorporating new data. Moreover, it has been verified that the PTCM significantly outperforms the traditional PMV and adaptive models in terms of higher prediction accuracy and flexibly constructed input features [18].

Table 2. Comparison between three types of thermal comfort models.

Model	Theoretical Basis	Advantages	Limits
PMV	Heat balance theory	Standardized calculation, easy integration, strong generalizability	Poor adaptability, low accuracy
Adaptive model	Behavioral, physiological, and psychological adaptations	Dynamic PMV correction, strong adaptability	Dependence on long-term climate data
PTCM	Multiple data drive	High accuracy, flexible structure, personalized prediction	Weak interpretability

summarizes three types of thermal comfort models in terms of theoretical basis, advantages, and limits. Based on the above analysis, it is seen that the PTCM can fundamentally change the conventional “one-size-fits-all” thermal comfort management by offering individual-specific and context-relevant comfort prediction for the occupant-centric environment control.

Personal thermal comfort modeling can be defined as the process of predicting or evaluating the individual thermal comfort sensation in a specific thermal environment through relevant data and algorithms. In recent years, with the development of sensing technology and artificial intelligence (AI), advanced data-driven methods have gradually been introduced into the field of personal thermal comfort modeling. These methods adopt machine learning (ML) [19,20,21,22,23,24,25,26,27] or deep learning (DL) [28,29,30,31,32,33,34] algorithms to establish the PTCMs. The framework of personal thermal comfort modeling is shown in Figure 1, which will be further discussed in subsequent Section 2 and Section 3.

1.2. Literature Selection

This study presents a comprehensive review on data-driven thermal comfort modeling. The literatures are searched in the database of Web of Science Core Collection, with the time interval limited to the past five years. The searching keywords are selected as follows: “(thermal comfort model or thermal comfort prediction or thermal comfort assessment or thermal comfort identification) and (personal or individual) not PMV”. Based on the above database, time interval, and keywords, the initial search is conducted. Subsequently, the searched literature is sorted by relevance, and the top 50% are selected. Then, further literature screening is carried out based on the following criteria: (1) detailed description of the acquisition of modeling datasets (such as from thermal comfort experiments or public datasets) and (2) inclusion of one/several data-driven modeling methods. Finally, combined with the previous related work of our research team, a total of 113 pieces of the literature are identified.

1.3. Research Gap Analysis

Over the past five years, several review of the literature have been published on thermal comfort models. Zhao et al. [35] reviewed the evolution of thermal comfort models, explored the adaptability of different models in various scenarios (e.g., sleeping scenario and outdoor scenario) and for different populations, and summarized data-driven models based on machine learning. Martins et al. [36] concluded that current research mainly focused on office environments, using ML methods to build the PTCM. They pointed out three limitations of current research, including strong sample homogeneity, inconsistent evaluation standards, and insufficient validation. Fard et al. [37] comprehensively analyzed the application of ML in thermal comfort research. It was shown that ML models could achieve the 58.5% energy consumption reduction in heating, ventilation, and air conditioning (HVAC) optimization while improving indoor thermal comfort. Future directions, such as developing non-contact sensing technologies and multi-parameter coupling analysis methods, were also provided.

Table 3. Summary of the previous literature reviews.

Contents	[35] (2021)	[36] (2022)	[37] (2022)	[38] (2022)	[39] (2023)	[8] (2021)
Data collection way	×	×	×	×	×	×
Feature correlation analysis	✓	✓	✓	✓	✓	×
Modeling method	✓	✓	✓	✓	✓	✓
Model evaluation indicator	×	✓	✓	✓	×	×
Data-efficient modeling using transfer learning	×	×	×	×	×	×
Integration into building environment control	×	✓	✓	×	×	×

is presented to understand the research gap between this study and previous reviews. This table is organized based on the main contents of this study as follows: (1) data collection way (contact-based or non-contact); (2) feature correlation analysis; (3) modeling method (including ML and DL); (4) model evaluation indicator; (5) Data-efficient modeling using transfer learning (TL); (6) integration into building environment control. From Table 3, it is known that most previous reviews did not discuss the integration of PTCM into building environment control systems, which aims to investigate whether the PTCM can be applied to real-world scenarios to achieve the desired control performance in improving thermal comfort and enhancing building energy efficiency. Moreover, the high cost and limited scale of collection experiments, as well as ambiguity and individual differences in subjective feedback, usually lead to difficulty in obtaining sufficient high-quality datasets. As an efficient ML paradigm, TL provides an important technical path for solving small sample modeling problems. However, none of these literature reviews addressed the TL-based data-efficient thermal comfort modeling in sample scarcity scenarios. This study systematically reviews the advanced data-driven technologies currently used for personal thermal comfort modeling. Main parts could be summarized as follows: (1) We present insights into thermal comfort data collection, including data type, collection way, and data correlation analysis. (2) We comprehensively review various thermal comfort modeling methods based on ML and DL, as well as commonly used evaluation indicators for thermal comfort models. (3) On the basis of research accumulation of our group, we analyze the data-efficient thermal comfort modeling using the TL strategy and discuss the characteristics of different transfer methods. (4) We summarize the application potential of integrating the PTCM with thermal control to achieve the expected performance in actual building environment control systems. (5) We discuss some challenges and future directions in the field of PTCM.

2. Thermal Comfort Data Collection and Preprocessing

2.1. Data Type

For personal thermal comfort modeling, data collection has a significant influence on model accuracy and generalization performance. According to the collected content, thermal comfort data are usually classified into the following four types: environmental data, physiological data, behavioral data, and subjective feedback data. The first three types serve as input features for the model, while the latter type is typically used as output labels. Environmental data usually include air temperature ($T_{air}$), relative humidity (RH), air speed, radiant temperature ($T_r$), light intensity, carbon dioxide concentration, etc. These data reflect the objective state of the external environment. Physiological data usually include skin temperature, core body temperature, heart rate (HR), heart rate variability (HRV), electrodermal activity (EDA), sweating rate, etc. These data can intuitively reflect the physiological regulation process and sensation differences of the human body in the external environment.

Behavioral data usually include regulating HVAC setpoints, opening/closing windows, putting on/taking off clothes, wiping sweat, etc. The behavioral data concern strategies that the human body adopts to adapt to the environment through active behaviors. Li et al. [40] introduced six typical thermal adaptation actions (e.g., wiping the forehead, fanning, and crossing arms, etc.) into the personal thermal sensation (TS) prediction. Results showed that the prediction accuracy of seven-point TS could be increased from 82.86% to 86.8%. Liu et al. [41] used cold behaviors (such as wearing a coat) and hot behaviors (such as wiping sweat) as input features to build the thermal preference model. Results demonstrated that the model accuracy reached 0.81, which was superior to the model using environmental variables as input features. When the above behaviors were combined with some environmental data, model accuracy could be improved to 0.85. Due to the fact that behavioral data usually reflect behaviors in specific environments and causal inference of behaviors is difficult, they are not as widely applied in thermal comfort modeling as environmental and physiological data.

Based on the identified literature, during the experimental phase, most studies (approximately 80%) collect both physiological and environmental data as input variables for thermal comfort models, aiming to more comprehensively depict the dynamic interaction mechanism between the human body and environment. For examples, Jiang et al. [42] systematically analyzed the coupling relationship between environmental changes and physiological responses by monitoring $T_{air}$, RH, and multi-area skin temperatures. Aryal and Becerik-Gerber [43] investigated the spatiotemporal characteristics of personal thermal sensation using indoor ambient temperature and local skin temperatures of the forehead, nose, and cheeks. Yu et al. [44] constructed the machine learning-based thermal comfort model by integrating skin temperature, environmental parameters, and individual factors (such as age and gender). Wang et al. [45] utilized skin temperatures of multiple facial regions, together with environmental temperature and humidity as the input variables, and established machine-learning models for elderly thermal sensation prediction. Some studies only collect environmental data for thermal comfort modeling [46,47,48,49,50,51,52,53,54,55,56,57,58]. These studies are simple for experimental implementation. However, due to the lack of physiological data, the established models have certain limitations in depicting personal thermal comfort perception.

Subjective feedback data, as the output of thermal comfort training models, are typically obtained through questionnaire voting. Although questionnaires may vary in form, the ASHRAE thermal sensation scale is commonly used. The 7-point scale (−3 for “cold”, 0 for “neutral”, and +3 for “hot”) is widely used in both laboratory and field investigations, as in [43,59,60,61,62]. The 5-point scale [63,64] or 3-point scale [65,66] has simplified gradations, reducing the cognitive burden on the subjects. In addition to the thermal sensation scale, other scales, such as the thermal comfort scale [67,68], the thermal preference scale [69], and the thermal acceptability scale [70], are adopted to comprehensively assess the subjective feelings of the human body in a thermal environment. Examples of these scales are shown in Figure 2. Note that subjective voting inevitably introduces inherent noise, ambiguity, and individual differences, which greatly affect the data consistency and stability. To address this issue, the following measures can be adopted: (1) Using physiological signals as objective agents of subjective perception can compensate for and calibrate the delay, inaccuracy, and volatility of purely subjective reports, forming a more consistent multimodal dataset. (2) Apply statistical methods (such as smoothing based on time series) or clustering algorithms to identify abnormal votes that are significantly inconsistent with individual long-term patterns or contemporaneous group trends.

2.2. Collection Way

According to the degree of contact, input data collection for thermal comfort modeling can be classified into two ways: contact-based and non-contact. Contact-based collection usually requires direct contact with subjects to obtain high-precision physiological signals. Thermocouples, attached temperature probes, finger pulse oximeters, etc. [42,43,60,62,71,72,73,74,75,76], can be used for contact-based collection. Although this collection way has high measurement accuracy, the direct contact of equipments with body may affect the subject’s behavior or psychological state, thereby affecting the authenticity of the data. Non-contact collection has received considerable attention in recent years, with advantages of good subject experience and ease of large-scale deployment in real world. Thermal imaging devices, such as FLIR Lepton and FLIR ONE Pro, could measure the temperature of face or body surface in non-contact manners [45,65,66,77]. Web cameras were utilized for non-contact collection of facial expressions [78]. Then, combined with computer vision algorithms, individualized thermal comfort state could be accurately predicted. However, accuracy and stability of non-contact collection are easily influenced by environmental conditions, changes in orientation, and variations in lighting [79]. Consequently, there is a trade-off between these two collection ways in terms of measurement accuracy, subject experience, and deployment flexibility.

Based on the 113 reviewed papers, the statistics chart of the above two collection ways is shown in Figure 3. It is known that the contact-based collection way accounts for approximately 67.1%, while the non-contact way accounts for 32.9% (excluding the literature that directly uses public datasets). Generally speaking, in the field of thermal comfort data collection, researchers are faced with two main choices: high-precision contact-based collection and easy-to-deploy non-contact collection. Researchers can achieve a flexible balance between collection accuracy and convenience, and choose the most suitable way. In addition to self-conducted experimental collection, some public datasets are also used for thermal comfort modeling, such as the ASHRAE RP-884 Dataset [80,81,82,83] and the ASHRAE Global Thermal Comfort Database [30,84,85,86,87,88,89]. These datasets have large sample sizes, diverse sources, and a wide range of environmental conditions, providing reliable support for model accuracy evaluation and generalization performance validation.

2.3. Data Preprocessing

In data-driven modeling tasks, data preprocessing helps to reveal the intrinsic relationships among data, thereby promoting the improvement of model accuracy. Correlation analysis is widely applied in the field of thermal comfort modeling. For instance, the Pearson correlation coefficient [61,63,75] and Spearman rank correlation coefficient [42,62,71,74] are frequently used to quantitatively assess the potential relationships between variables such as physiological indicators, environmental parameters, and subjective thermal sensation. In [90], based on experimental data from 15 male subjects in East China, the authors employed Pearson and Spearman correlation analysis to investigate the relationships between nine environmental and physiological indicators and thermal sensation voting (TSV). Results revealed that $T_{air}$, skin temperature at the wrist, HR, and HRV were significantly correlated with TSV, with skin temperature at the wrist showing the strongest correlation. Furthermore, the feature selection methods are introduced to identify the variables that have a significant influence on thermal comfort. For example, in [44], the Boruta method was employed to analyze the importance of features for thermal comfort prediction using data with professional and practical settings. The K-nearest neighbor (KNN) method was further adopted to filter the initial features according to their contributions to model accuracy enhancement. Results showed that the local skin temperatures were highly related to thermal comfort for professional settings, while local skin temperature at the hand and age were the most related features for practical settings.

2.4. Comparison and Summary

Based on data types, collection ways, and correlation analysis,

Table 4. Studies on data collection for personal thermal comfort modeling.

Reference	Data Type		Collection Way		Collected Data	Collection Device	Correlation Analysis
	Physiological	Environmental	Contact-Based	Non-Contact
[59]	✓	✓	✓		$T_{air}$, RH, $T_g$, $T_{forehead}$, $T_{abdominal}$, $T_{elbow}$, $T_{hand}$, $T_{thigh}$, $T_{leg}$, $T_{foot}$, Age, Sex, Height, Weight, BMI, Skin surface area, Seven-point TSV	HOBO U12-012, HQZY-1, DS1921H
[75]	✓	✓	✓		$T_{air}$, RH, Air speed, $T_r$, $T_{skin}$, Age, Sex, BMI, Seven-point TSV	WBGT-2009, Pt1000, DT-830LN, Kata thermometer	Pearson
[70]		✓		✓	$T_{air}$, RH, Air speed, $T_g$, Seven-point TSV, Four-point TCV, Two-point TUV	NOAA
[63]	✓	✓		✓	$T_{head}$, $T_{top}$, $T_{foot}$, $T_{back}$, $T_{chest}$, Sex, Five-point TSV	Flir Lepton 3.5, ETA1006T, TA622B	Pearson
[60]	✓	✓	✓		$T_{air}$, RH, $T_{floor}$, $T_{roof}$, $T_{wall}$, $T_r$, Illumination, $C{O}_2$ level, EDA, HR, Sex, Age, Birthplace, Seven-point TSV	BioHarness 3.0, BITalino
[64]	✓	✓	✓		$T_{skin}$, EDA, HRV, $T_{air}$, RH, $C_{PM2.5}$, Air speed, Five-point TCV	SS6L, SS57LA, SS2LB
[42]	✓	✓	✓		$T_{air}$, RH, Air speed, $T_{skin}$, Sweat rate, $T_{core}$, Nine-point TSV, Long-wave radiation, Short-wave radiation	RM YOUNG 41382, RM YOUNG 81000, Kipp & Zonen CNR-4, elfin VapoMeter SWL5, BodyCap eCelsius	Spearman
[61]	✓	✓	✓		$T_{air}$, RH, $C{O}_2$ level, Illumination, $T_{bg}$, $T_{left-back-hand}$, $T_{left-forearm}$, $T_{right-upper-arm}$, $T_{left-chest}$, $T_{left-thigh}$, $T_{right-calf}$, $T_{temple}$, Seven-point TSV	HOBO, WBGT, ppbRAE 3000, iButton	Pearson
[68]	✓	✓			$T_{wrist}$, HRV, $T_{air}$, Seven-point TSV, Five-point TCV	DS18B20, DHT22, Arduino Pro Mini 5V, Pulse sensor
[44]	✓	✓	✓	✓	$T_{hand}$, $T_{head}$, Blood pressure, $C{O}_2$ level, RH, $T_{air}$, Air speed, $T_{back}$, $T_{chest}$, $T_{arm}$, $T_{leg}$, $T_{thigh}$, HR, Age, Weight, Height, BMI, Body fat Seven-point TSV	TMC6-HD, MLX90640, HOBO UX0.006-006M, OMRON U30, Polar A300
[65]	✓	✓		✓	$T_{air}$, RH, $T_{nose}$, $T_{right-cheek}$, $T_{left-cheek}$, $T_{mouth}$, $T_{chin}$, Sex, Three-point TSV	HOBO UX100-003, HIKVISION K20
[71]	✓	✓	✓	✓	$T_{skin}$, HR, $T_{air}$, RH, $T_{breath}$, Sex, Age, Height, Weight, Seven-point TSV	MLX90614, DHT-22, NXFT15XH103FA2B130, Ear clip	Spearman
[74]	✓	✓		✓	$T_{air}$, RH, Air speed $T_{head}$, $T_{upper-arm}$, $T_{lower-arm}$, $T_{chest}$, $T_{back}$, $T_{hand}$, $T_{leg}$, $T_{thigh}$, $T_{finger}$, Sex, Age, Height, BMI, Weight, $T_r$, Seven-point TSV	iButton	Spearman
[29]	✓	✓		✓	$T_{air}$, RH, Thermal image, Air speed, Sex, Age, Weight, Height, Five-point TCV	Flir Lepton 3.5, T-type thermocouple
[48]		✓		✓	$T_{air}$, RH, MRT, Age, Weight, Clo, Air speed, $T_g$, $T_{wall}$, Seven-point TSV	Testo 174H, RS-HQ-USB, Testo 425, FOTRIC 226
[91]	✓	✓		✓	$T_{air}$, RH, Thermal image, $T_{hvac}$, Sex, Age, Height, Weight, Five-point TSV	Flir Lepton 3.5
[92]		✓		✓	$T_{air}$, RH, MRT, Air speed, $T_g$, Seven-point TSV	Hobo U30	Pearson, Spearman
[69]	✓	✓		✓	Action, $T_{hvac}$, $T_{air}$, RH, Air speed, $T_g$, Age, Sex, Seven-point TSV, Three-point TPV	HD32.2 WBGT Index, ANA-AN00
[77]	✓	✓		✓	Thermographic videos, Clo, Age, Sex, Race, Three-point TSV	FLIR ONE Pro
[93]		✓		✓	$T_{air}$, Air speed, MRT, Age, Height, Weight, Four-point TCV, Seven-point TSV	HOBO UX100-011A, WWFWZY-1, HQZY-1	Pearson
[67]	✓	✓		✓	$T_{forehead}$, $T_{check}$, $T_{nose}$, $T_{mouth}$, $T_{neck}$, Three-point TCV	FLIR Lepton
[94]		✓		✓	$T_g$, $T_{air}$, $T_{dry-bulb}$, MRT, Air speed, Clo, Seven-point TSV, Three-point TPV	Tasco Anemometer, Utron Heat Index WBGT Meters
[95]		✓		✓	$T_{out}$, $T_{in}$, Thermal discrepancy, $T_{max-out}$, $T_{mean-out}$, Age, Sex, Three-point TPV	NOAA
[62]	✓	✓	✓		$T_{air}$, Air speed, RH, MRT, $T_{hand}$, Pulse, Metabolic rate, Clo, Seven-point TSV	Exacon D-S18JK, Pulsioximeter, Velocicalc 9545, LM35	Spearman
[66]	✓	✓		✓	$T_{air}$, Facial expression, Three-point TSV	UX100-003, Web camera

T_g: globe temperature, TCV: thermal comfort voting, TUV: thermal unacceptability voting, TPV: thermal preference voting, MRT: mean radiant temperature, BMI: body mass index, Clo: clothing insulation, T_bg: black globe temperature, T_forehead: forehead temperature, T_abdominal: abdominal temperature, T_elbow: elbow temperature, T_hand: hand temperature, T_leg: leg temperature, T_thigh: thigh temperature, T_foot: foot temperature, T_skin: skin temperature, T_top: temperature at top of the head, T_back: back temperature, T_chest: chest temperature, T_floor: floor temperature, T_roof: roof temperature, T_wall: wall temperature, T_core: core temperature, T_{left-back-hand}: left hand back temperature, T_left-forearm: left forearm temperature, T_{right-upper-arm}: right upper arm temperature, T_left-chest: left chest temperature, T_left-thigh: left thigh temperature, T_right-calf: right calf temperature, T_temple: temple temperature, T_wrist: wrist temperature, T_head: head temperature, T_nose: nose temperature, T_right-cheek: right cheek temperature, T_left-cheek: left cheek temperature, T_mouth: mouth temperature, T_chin: chin temperature, T_breath: exhaled breath temperature, T_upper-arm: upper arm temperature, T_lower-arm: lower arm temperature, T_finger: finger temperature, T_hvac: HVAC outlet air temperature, T_check: cheek temperature, T_neck: neck temperature, T_dry-bulb: dry-bulb temperature, T_out: outdoor temperature, T_in: indoor temperature, T_max-out: max daily outdoor temperature, T_mean-out: mean daily outdoor temperature.

summarizes some of the representative literature on thermal comfort data collection in recent years. It can be seen that (1) Almost all studies collect environmental data $T_{air}$ and RH for thermal comfort modeling. In addition, the collection of physiological data (e.g., skin temperature, HR, and EDA) has increased in recent years, reflecting the emphasis on the relationship between individual differences and subjective feelings. (2) With the development of non-contact equipment (e.g., infrared cameras, video cameras, and smart wristbands), more researchers prefer non-contact data collection for thermal comfort modeling. (3) Pearson and Spearman’s correlation coefficients are commonly used to measure correlation. Moreover, some ML-based feature selection methods can be adopted to reduce redundant inputs and computational costs. Overall, the trend in thermal comfort data collection is characterized by multi-source data fusion, suitable collection ways, and diversified analysis methods, which provide support for the construction of more accurate and personalized thermal comfort models.

3. Model Construction and Evaluation

3.1. Model Input and Output

The inputs of thermal comfort models are typically composed of environmental variables and personal variables. Compared with most adopted $T_{air}$ and RH, other environmental variables such as $T_r$, time of day [96], and the operating status of the personal comfort system (PCS) [96,97] are less frequently considered. Personal variables usually reflect an individual’s physiological/behavioral characteristics (e.g., skin temperature, HR, adaptive actions, thermal image [28,29,91]) and personal information (e.g., age, gender, body mass index, and clothing insulation). The reviewed literature indicates that over half of the personal comfort models incorporate at least one personal variable as their input. The outputs of thermal comfort models are usually quantified as thermal comfort-related indicators, including thermal sensation (TS), thermal comfort (TC), thermal preference (TP), etc. For examples, seven-point TS [28,74,97,98,99,100,101], five-point TS [91,102], three-point TS [96,103], five-point TC [64,68], three-point TC [29,67,96], and three-point TP [30,41,67,104,105,106] were used as output variables of PTCMs.

3.2. Data-Driven Modeling Method

The data-driven modeling methods for thermal comfort can be divided into two categories: ML and DL methods. According to the reviewed 113 literature, approximately 64.6% of them adopt the ML method.

(I)

Modeling method using ML

In the reviewed literature, the top three most commonly used models are support vector machine (SVM), random forest (RF), and KNN. For example, in [74], the authors demonstrated that, based on the hand and finger temperature features, the SVM model achieved the RMSE of 0.6921–0.9091 in the prediction of seven-point TS. In [41], indoor/outdoor environmental parameters and individual information were utilized for thermal preference prediction. Results showed that the prediction accuracy of SVM ranged from 0.57 to 0.87. In [41], it was reported that the RF model achieved an accuracy of 0.64–0.88 in the prediction of three-point TP. In [107], it was found that, using facial temperatures as input features, prediction accuracy of RF ranged from 0.726–1.000 (TC) and 0.817–1.000 (TS). In [41], the authors indicated that the KNN model achieved an accuracy of 0.71–0.85 in the prediction of three-point TP. In [102], it was proved that, when using physiological signals such as electroencephalogram (EEG), photoplethysmography (PPG), and skin temperature, accuracy of KNN ranged from 0.62 to 0.74 for prediction of five-point TS.

(II)

Modeling method using DL

Compared with traditional ML, DL has a deep structure of ANNs with the capability of processing very large-scale data [108,109], thus demonstrating stronger flexibility, better generalization, and stronger robustness in modeling complex nonlinear relationships [110,111,112]. In the field of thermal comfort modeling, the convolutional neural networks (CNN) and long short-term memory network (LSTM) are two commonly used DL methods. CNN is composed of convolutional, pooling, and fully connected layers, in which the convolutional layer is used for feature extraction and the pooling layer for data dimensionality reduction. LSTM is mainly composed of a series of repeating cells. Main advantages of the LSTM lie in dividing the information flow into long-term memory and short-term memory and solving the gradient disappearance of the typical recurrent neural network (RNN) through the dynamic adjustment mechanism. More details of CNN and LSTM can be referred to [113,114,115,116].

According to the reviewed literature, many encouraging results have been obtained in the field of thermal comfort modeling with the aid of CNN and LSTM. Based on thermal imaging data, Zakka et al. [28] collected thermal imaging data and TSV data from 10 subjects under different temperature conditions, and developed the CNN-based prediction models for both 3-point and 7-point thermal sensation scales. Validation results showed that the 3-point model achieved the high average accuracy of 99.51%, while the 7-point model also reached the accuracy of 89.90%, both outperforming many existing models that relied on complex feature engineering and multi-sensor fusion. Cho et al. [31] proposed a personalized prediction method based on multi-head LSTM (Mh-LSTM) networks. Experiments were conducted to collect hand skin temperature, environmental temperature and humidity, and TSV data from six subjects. Results demonstrated that the Mh-LSTM model could achieve the root mean square error (RMSE) of 0.2225, which was significantly better than other comparison models such as the feed-forward neural network (FFNN) and single-head LSTM. Chennapragada et al. [104] proposed a time-series DL model based on regularized LSTM (R-LSTM) for predicting personal thermal preference. L1 regularization was combined with an attention mechanism to suppress overfitting while capturing long-term dependencies. Using physiological data (HR, wrist temperature, ankle temperature) and environmental data ($T_{air}$, RH) from 14 participants collected over 2–4 weeks, 3D training samples were constructed with a 120 min sliding window and evaluated via 5-fold time-series cross-validation. Results verified that the R-LSTM model achieved an average accuracy of 78%, higher than the typical RF model.

Among DL-based thermal comfort models, the attention mechanism can be applied to enhance the feature representation ability, which has attracted great interest from researchers. A simple diagram of the spatial attention mechanism (SAM) is given in Figure 4. The channel-refined feature F’ is first processed through both average pooling and max pooling operations along the channel dimension to extract complementary spatial information. The two pooled feature maps are then concatenated and passed through a convolution layer to generate the spatial attention map (M_s). The generated M_s enables the model to automatically focus on thermal comfort-sensitive regions such as nose, lips, and eye area, while suppressing interference from background and irrelevant information. Miao et al. [91] combined residual network with 34 layers (ResNet34) with the SAM for personal thermal comfort prediction. Results showed that the SAM based ResNet34 (SAM-ResNet34) achieved the accuracy of 93.75% in prediction of three types of thermal comfort states. Kang et al. [29] constructed the attention-based residual network with 50 layers (Attention-ResNet50) for thermal comfort prediction in different genders, in which the spatial attention module was embedded in the deep network architecture. Results demonstrated that the constructed model could significantly improve the prediction accuracy, with accuracy close to 100% for female subjects. Note that in the above two studies, the residual learning effectively alleviated the vanishing-gradient and performance degradation problems in deep networks through “shortcut connections”.

3.3. Model Evaluation Indicators

The performance of PTCM can be measured by various indicators. Accuracy is one of the most commonly used indicators [28,38,41,91,97,102,104,107]. It represents the proportion of samples correctly predicted by the model to the total samples, as follows:

$$\mathrm{Accuracy}={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$$

(1)

where $TP$, $TN$, $FP$, and $FN$ denote true positive, true negative, false positive, and false negative, respectively. Note that when data categories are imbalanced, especially in cases with extremely skewed data, accuracy is difficult to objectively evaluate the model’s performance.

Precision [28,29,91] and recall are a pair of mutually restrictive evaluation indicators, as follows:

$$\mathrm{Precision}={\displaystyle \frac{TP}{TP+FP}}$$

(2)

$$\mathrm{Recall}={\displaystyle \frac{TP}{TP+FN}}$$

(3)

During the model construction process, if too much emphasis is placed on precision, some actual positive samples may not have been identified, thereby reducing recall. Conversely, if a higher recall is pursued, it may increase the false positive rate, resulting in a decrease in precision. Therefore, in practical applications, it is usually necessary to seek a balance between these two. To address this issue, F1 Score [28,30,91,104] is proposed, as follows:

$$\mathrm{F}1\mathrm{Score}=2·{\displaystyle \frac{\mathrm{Precision}·\mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}}$$

(4)

The F1 Score can reflect the robustness of the classification model, and is particularly suitable for scenarios with imbalanced category distributions.

In addition to the above commonly used indicators, other indicators, such as the RMSE [31,74] and area under the curve (AUC) [104,117], can be employed to evaluate the model performance. RMSE measures the difference between predicted and actual values, reflecting the overall fitting degree of the model. AUC is often utilized to evaluate the discrimination ability of binary classification models, with a value closer to 1 indicating better discrimination between positive and negative samples. These indicators can provide the multi-angle evaluation criteria for comprehensive performance of thermal comfort models.

Table 5 summarizes some recent studies on data-driven personal thermal comfort models from aspects of modeling method, input and output, evaluation indicator, etc. It is concluded the following from Table 5: (1) In addition to commonly used ML models (e.g., SVM, RF, KNN), other ML models such as linear regression (LR) [72,74,118,119] and ANN [31,107,120] also demonstrate unique advantages in specific prediction scenarios. (2) Advanced DL models (CNN, LSTM, Attention-ResNet, etc.) outperform traditional ML models in complex nonlinear modeling and extraction/processing of high-dimensional features (e.g., thermographic images, multiple physiological and environmental features). (3) Various evaluation indicators are adopted for thermal comfort models. For classification tasks, accuracy, precision, recall, and F1-score are commonly utilized, with the latter being more representative in scenarios of data imbalance. For regression tasks, RMSE is often utilized to measure the model’s fitting ability to real feedback data.

Table 5. Some studies on data-driven personal thermal comfort models.

	Reference	Modeling Method	Model Input	Model Output	Evaluation Indicator
Machine Learning	[74]	SVM	$T_{forehead}$, $T_{hand}$ $T_{lower-arm}$, $T_{finger}$	Seven-point TS	RMSE: 0.6921–0.9091
	[41]	SVM, KNN, DT, RF	$T_{in}$, $R{H}_{in}$ $T_{out}$, $R{H}_{out}$ Gender, Height, Weight Cold behaviors, Hot behaviors	Three-point TP	Accuracy: SVM: 0.57–0.87 KNN: 0.71–0.85 DT: 0.69–0.83 RF: 0.64–0.88
	[102]	KNN, LRG, GNB, RF	EEG, PPG EDA, $T_{skin}$	Five-point TS	Accuracy: KNN: 0.62–0.74 LRG: 0.68–0.79 GNB: 0.67–0.80 RF: 0.67–0.76
	[103]	DT, LRG, RF, GB, ANN	$T_{right-forehead}$, $T_{chin}$, $T_{left-forehead}$ $T_{left-cheek}$, $T_{right-cheek}$	Three-point TS Three-point TP	TP: Accuracy: DT: 0.698–1.000 LRG: 0.393–1.000 RF: 0.726–1.000 GB: 0.715–1.000 ANN: 0.944–1.000 TS: Accuracy: DT: 0.805–1.000 LRG: 0.585–1.000 RF: 0.817–1.000 GB: 0.847–1.000 ANN: 0.947–1.000
	[97]	RF, LRG, SVM	PCS status, $T_{air}$ $\Delta T_{\mathrm{fh}-\mathrm{cheek}}$, $\Delta T_{\mathrm{fh}-\mathrm{nose}}$ $\Delta T_{\mathrm{nose}-\mathrm{cheek}}$	Three-point TS	Accuracy: RF: 0.69 LRG: 0.69 SVM: 0.67 F1-Score: RF: 0.61–0.77 LRG: 0.57–0.83 SVM: 0.60–0.79
	[96]	SVM	$T_{skin}$, $T_{hand}$, $T_{air}$ RH, $T_r$ Time, PCS status	Three-point TS Three-point TC	TC: Accuracy: 0.814–0.849 AUC: 0.80 RMSE: 0.42 TS: Accuracy: 0.936–0.952 AUC: 0.78 RMSE: 0.34
	[121]	RF	$T_{in}$, $T_{out}$, $T_{set}$ H, Month, $\Delta T_{\mathrm{set}-\mathrm{in}}$, $\Delta T_{\mathrm{out}-\mathrm{in}}$, ACSA, ACOM	Three-point TP	Accuracy: 0.753–0.873
Deep Learning	[28]	CNN	Thermographic images	Three-point TS	Precision: 0.9786–0.9991 Recall: 0.9491–0.9995 F1 Score: 0.9640–0.9993 Accuracy: 0.9765–0.9994
	[29]	Attention-ResNet50	Thermographic images, $T_{head}$, $T_{body}$ $T_{foot}$	Three-point TC	Precision: 0.897–1.000 Recall: 0.833–1.000 Specificity: 0.914–1.000
	[91]	SAM-ResNet34	Thermographic images	Three-point TS	Accuracy: 0.9375 Precision: 0.9416 Recall: 0.9341 F1 Score: 0.9378
	[31]	Mh-LSTM	$T_{keyboard}$, $H_{keyboard}$ $H_{left-hand}$, $H_{wall}$, $T_{right-hand}$$,H_{right-hand}$, $T_{left-hand}$, $T_{wall}$	Seven-point TS	RMSE: 0.2225–1.0940
	[30]	CNN-LSTM	RH, $T_{air}$, air speed, $C{O}_2$ level, $T_{bg}$, PM2.5, $T_{out}$	Three-point TP	Accuracy: 0.675–0.698 F1 Score: 0.636–0.658
	[104]	LSTM	RH, $T_{air}$, $T_{wrist}$ $T_{breath}$, HR, $T_{ankle}$	Three-point TP	Accuracy: 0.68–0.86 F1 Score: 0.66–0.81 AUC: 0.65–0.84

DT: decision tree, LRG: logistic regression, GNB: Gaussian naive Bayes, GB: gradient boosting, ANN: artificial neural network, Mh-LSTM: multi-head long short-term memory network, ΔT_fh-cheek: temperature difference between forehead and cheek, ΔT_fh-nose: temperature difference between forehead and nose, ΔT_nose-cheek: temperature difference between nose and cheek, ACSA: HVAC setting airspeed, ACOM: HVAC operation mode, H: hours of the day, T_body: core body temperature, T_keyboard: temperature in front of the keyboard, H_keyboard: humidity in front of the keyboard, H_wall: humidity in front of the wall, T_left−hand: left hand temperature, H_left−hand: left hand humidity, T_right−hand: right hand temperature, H_right−hand: right hand humidity, T_ankle: ankle temperature, T_in: indoor temperature, T_out: outdoor temperature, RH_in: indoor relative humidity, RH_out: outdoor relative humidity, T_{right-forehead}: right forehead temperature, T_{left-forehead}: left forehead temperature, T_set: setpoint temperature of HVAC, ΔT_set-in: difference between HVAC setpoint and indoor temperatures, ΔT_out-in: difference between outdoor and indoor temperatures.

Table 5. Some studies on data-driven personal thermal comfort models.

	Reference	Modeling Method	Model Input	Model Output	Evaluation Indicator
Machine Learning	[74]	SVM	$T_{forehead}$, $T_{hand}$ $T_{lower-arm}$, $T_{finger}$	Seven-point TS	RMSE: 0.6921–0.9091
	[41]	SVM, KNN, DT, RF	$T_{in}$, $R{H}_{in}$ $T_{out}$, $R{H}_{out}$ Gender, Height, Weight Cold behaviors, Hot behaviors	Three-point TP	Accuracy: SVM: 0.57–0.87 KNN: 0.71–0.85 DT: 0.69–0.83 RF: 0.64–0.88
	[102]	KNN, LRG, GNB, RF	EEG, PPG EDA, $T_{skin}$	Five-point TS	Accuracy: KNN: 0.62–0.74 LRG: 0.68–0.79 GNB: 0.67–0.80 RF: 0.67–0.76
	[103]	DT, LRG, RF, GB, ANN	$T_{right-forehead}$, $T_{chin}$, $T_{left-forehead}$ $T_{left-cheek}$, $T_{right-cheek}$	Three-point TS Three-point TP	TP: Accuracy: DT: 0.698–1.000 LRG: 0.393–1.000 RF: 0.726–1.000 GB: 0.715–1.000 ANN: 0.944–1.000 TS: Accuracy: DT: 0.805–1.000 LRG: 0.585–1.000 RF: 0.817–1.000 GB: 0.847–1.000 ANN: 0.947–1.000
	[97]	RF, LRG, SVM	PCS status, $T_{air}$ $\Delta T_{\mathrm{fh}-\mathrm{cheek}}$, $\Delta T_{\mathrm{fh}-\mathrm{nose}}$ $\Delta T_{\mathrm{nose}-\mathrm{cheek}}$	Three-point TS	Accuracy: RF: 0.69 LRG: 0.69 SVM: 0.67 F1-Score: RF: 0.61–0.77 LRG: 0.57–0.83 SVM: 0.60–0.79
	[96]	SVM	$T_{skin}$, $T_{hand}$, $T_{air}$ RH, $T_r$ Time, PCS status	Three-point TS Three-point TC	TC: Accuracy: 0.814–0.849 AUC: 0.80 RMSE: 0.42 TS: Accuracy: 0.936–0.952 AUC: 0.78 RMSE: 0.34
	[121]	RF	$T_{in}$, $T_{out}$, $T_{set}$ H, Month, $\Delta T_{\mathrm{set}-\mathrm{in}}$, $\Delta T_{\mathrm{out}-\mathrm{in}}$, ACSA, ACOM	Three-point TP	Accuracy: 0.753–0.873
Deep Learning	[28]	CNN	Thermographic images	Three-point TS	Precision: 0.9786–0.9991 Recall: 0.9491–0.9995 F1 Score: 0.9640–0.9993 Accuracy: 0.9765–0.9994
	[29]	Attention-ResNet50	Thermographic images, $T_{head}$, $T_{body}$ $T_{foot}$	Three-point TC	Precision: 0.897–1.000 Recall: 0.833–1.000 Specificity: 0.914–1.000
	[91]	SAM-ResNet34	Thermographic images	Three-point TS	Accuracy: 0.9375 Precision: 0.9416 Recall: 0.9341 F1 Score: 0.9378
	[31]	Mh-LSTM	$T_{keyboard}$, $H_{keyboard}$ $H_{left-hand}$, $H_{wall}$, $T_{right-hand}$$,H_{right-hand}$, $T_{left-hand}$, $T_{wall}$	Seven-point TS	RMSE: 0.2225–1.0940
	[30]	CNN-LSTM	RH, $T_{air}$, air speed, $C{O}_2$ level, $T_{bg}$, PM2.5, $T_{out}$	Three-point TP	Accuracy: 0.675–0.698 F1 Score: 0.636–0.658
	[104]	LSTM	RH, $T_{air}$, $T_{wrist}$ $T_{breath}$, HR, $T_{ankle}$	Three-point TP	Accuracy: 0.68–0.86 F1 Score: 0.66–0.81 AUC: 0.65–0.84

DT: decision tree, LRG: logistic regression, GNB: Gaussian naive Bayes, GB: gradient boosting, ANN: artificial neural network, Mh-LSTM: multi-head long short-term memory network, ΔT_fh-cheek: temperature difference between forehead and cheek, ΔT_fh-nose: temperature difference between forehead and nose, ΔT_nose-cheek: temperature difference between nose and cheek, ACSA: HVAC setting airspeed, ACOM: HVAC operation mode, H: hours of the day, T_body: core body temperature, T_keyboard: temperature in front of the keyboard, H_keyboard: humidity in front of the keyboard, H_wall: humidity in front of the wall, T_left−hand: left hand temperature, H_left−hand: left hand humidity, T_right−hand: right hand temperature, H_right−hand: right hand humidity, T_ankle: ankle temperature, T_in: indoor temperature, T_out: outdoor temperature, RH_in: indoor relative humidity, RH_out: outdoor relative humidity, T_{right-forehead}: right forehead temperature, T_{left-forehead}: left forehead temperature, T_set: setpoint temperature of HVAC, ΔT_set-in: difference between HVAC setpoint and indoor temperatures, ΔT_out-in: difference between outdoor and indoor temperatures.

3.4. Data-Efficient Modeling Using Transfer Learning

In recent years, TL has provided effective solutions for data-driven modeling problems in small sample scenarios [122,123,124,125]. TL can be defined as: given a source domain and a target domain, as well as a learning task, relevant knowledge is extracted from the source task to assist the model in completing the target task. For personal thermal comfort modeling, high cost and limited scale of collection experiments, and noise, ambiguity, and individual differences in subjective feedback may lead to considerable difficulties in obtaining high-quality data. As a result, training data-driven models directly on the target building data often faces issues of insufficient samples and overfitting. Considering that the human body’s perception mechanism of thermal environments has certain commonalities across different buildings, introducing TL into thermal comfort modeling tasks can improve the model’s accuracy and generalization under limited sample scenarios [126].

Some researchers have employed TL strategies to address the decline in modeling accuracy caused by the scarcity of high-quality thermal comfort samples. Table 6 summarizes some of the representative literature on TL-based thermal comfort modeling. Considering the high cost, complexity, and long duration of data collection, most studies in Table 6 directly adopt public datasets as the source domain. For instance, the ASHRAE RP-884 and Scales Project datasets were used in [81,127,128]. The ASHRAE Global Thermal Comfort Database II was used in [30]. In terms of transfer methods, the majority of these studies adopt the model-based transfer learning (MBTL). This transfer method pre-trains thermal comfort models with source domain data, and then fine-tunes parameters of the pre-trained models with a small amount of labeled target domain data to enhance the modeling accuracy of the target task. It should be pointed out that the MBTL method consumes significant computational costs in building the pre-trained models and relies on a certain amount of target domain data for parameter adjustment, which may lead to overfitting in cases of extremely scarce data. From the perspective of modeling methods, hybrid modeling methods incorporating CNN, such as CNN-LSTM and CNN-SVM, are commonly used.

Table 6. The representative literature on TL-based personal thermal comfort modeling.

Reference	Source Domain	Modeling Method	Transfer Method
[81]	ASHRAE RP-884, Scales Project	CNN-LSTM	MBTL
[30]	ASHRAE Global Thermal Comfort Database II	CNN-LSTM	MBTL, IBTL
[129]	Experimental Acquisition	CNN-SVM	MBTL
[127]	ASHRAE RP-884, Scales Project	MLP	MBTL
[128]	ASHRAE RP-884, Scales Project	CNN-LSTM	MBTL

Table 6. The representative literature on TL-based personal thermal comfort modeling.

Reference	Source Domain	Modeling Method	Transfer Method
[81]	ASHRAE RP-884, Scales Project	CNN-LSTM	MBTL
[30]	ASHRAE Global Thermal Comfort Database II	CNN-LSTM	MBTL, IBTL
[129]	Experimental Acquisition	CNN-SVM	MBTL
[127]	ASHRAE RP-884, Scales Project	MLP	MBTL
[128]	ASHRAE RP-884, Scales Project	CNN-LSTM	MBTL

Our research team has also conducted meaningful investigations in TL-based thermal comfort modeling [130,131]. In [130], the instance-based transfer learning (IBTL) method was adopted. The physiological and environmental data were collected as the target domain data from 20 subjects in the winter office environment. The nearest neighbor search (NNS) algorithm was adopted to select samples similar to the target domain data from public datasets. An improved transfer AdaBoost (iTrAdaBoost) algorithm was proposed to dynamically adjust the weights of source domain and target domain samples. Results showed that the developed model based on NNS and iTrAdaBoost algorithms significantly outperformed other models in terms of prediction accuracy when the target domain data were scarce. In [131], various transfer methods for thermal comfort prediction were investigated based on field experiments, including MBTL, IBTL, and feature-based transfer learning (FBTL). On this basis, the above three transfer methods were combined based on a hard-voting approach to establish an ensemble transfer method. Results verified that the FBTL outperformed MBTL and IBTL in terms of prediction accuracy, as it projected the source and target domains into a shared low-dimensional Hilbert space to reduce their distribution differences. Further, the ensemble method was always superior to the other three transfer methods in model accuracy and generalization.

In general, when there is very limited on-site data, the IBTL method shows the most significant improvement in accuracy due to its ability to select similar samples from the source domain to supplement the target domain. As the size of the target domain data increases, prediction efficiency of the MBTL-based model becomes higher. The reason is that by pre-training the model on a certain amount of source domain data, it is possible to extract deep features that characterize the general patterns of human thermal response and then transfer them to the target domain for fine-tuning training, which can effectively overcome the problem of insufficient labels in the target domain. When the data distribution difference between the source and target domains is significant, the FBTL method outperforms commonly used MBTL and IBTL methods. Furthermore, by integrating the above three transfer methods in parallel, more stable and accurate prediction results can be obtained. The cost is an increase in the complexity of the established model.

4. Integration of PTCM into Building Environment Control

From the key features of PTCMs (see Section 1.1), it is known that integration of PTCMs into building environment control offers an opportunity to achieve occupant-centric building design. Furthermore, by providing information on comfort requirements of individuals, PTCMs can help to realize both thermal comfort improvement and building energy savings [16]. A typical integration framework is shown in Figure 5, in which multi-source sensors, communication network, controller (also called a central server), and actuators are included. Compared with data-driven personal thermal comfort modeling, the research on integrating PTCM into environmental control systems is rather limited. This is mainly because of the following: (1) Such integration relies on numerous hardware technologies, such as sensors used for input signal acquisition to predict personal thermal comfort, communication networks used for data transfer from sensors/devices to the controller, the controller used for data analysis, optimization, and actuation commands, etc. These hardware requirements significantly increase the cost and complexity of integration. (2) A significant “lab-to-field gap” exists between laboratory settings and real-world deployments, while the laboratory-trained PTCM can achieve an $R^2$ (determination coefficient) as high as 0.95 under strictly controlled conditions, its performance often degrades in practice due to stochastic factors such as unpredictable occupant activities and distributional shifts in outdoor weather [132]. (3) Python has become the main popular tool for implementing PTCMs due to its rich computing libraries and ease of use. Nevertheless, there are communication difficulties between Python-based PTCMs and legacy building automation protocols, e.g., protocol compatibility (Python libraries have limited support for these legacy protocols), real-time performance of communication, and complexity in data processing.

As a result, most current thermal control systems still adopt traditional temperature ranges or physics-informed comfort models (such as PMV and adaptive models). For examples, Yu et al. [133] proposed an HVAC control method based on multi-agent deep reinforcement learning (MADRL), in which ensuring indoor temperature within the prescribed range (19–24 °C) was one of the main objectives of thermal comfort control. Wu et al. [134] established the indoor environment model through computational fluid dynamics (CFD) simulation and designed the HVAC control strategy to maintain the PMV in the working area within the range of ($[-0.2,+0.2]$). Li et al. [135] presented a MADRL-based multi-zone HVAC control method that optimized energy consumption while guaranteeing the occupant’s thermal comfort. Simulation results based on TRNSYS showed this method could save about 8.5% energy compared with traditional methods under the constraint PMV $\in [-0.5,0.5]$. When the constraint was relaxed to PMV $\in [-1.0,1.0]$, energy savings could reach 15.4%. Xue et al. [136] developed a hybrid strategy for multi-zone HVAC control based on a multi-agent deep deterministic policy gradient algorithm, in which the PMV was utilized for zone comfort assessment. Co-simulation using TRNSYS18.0 and MATLAB2019b demonstrated that the developed strategy could guarantee thermal comfort of zone occupants and reduce the HVAC energy consumption. Er-retby et al. [137] proposed a hybrid control framework for naturally ventilated buildings that combined adaptive thermal comfort standards with probabilistic occupant behavior models. Field test results verified that, by integrating adaptive comfort models with optimized occupancy schedules, energy consumption could be reduced by 0.25 kWh/m² (location: Benguerir) and 0.37 kWh/m² (location: Lyon), while indoor comfort hours be increased by 67% and 39%, respectively.

In recent years, due to the development of sensing, networking, and communication technologies, as well as the superiority of data-driven PTCMs, some researchers have attempted to embed PTCMs into building environment control to achieve more accurate and efficient environmental regulation. Table 7 summarizes some recent studies on PTCM-based control systems. It is seen that RF-based thermal comfort models are mostly utilized, while various control methods are adopted, including model predictive control (MPC), fuzzy logic control, cooperative control, discrete feedback control, etc. From the perspective of actuators, HVAC and different types of PCS were mainly employed. To list a few of the studies, in [97], an automatic PCS was developed to regulate the local environment of occupants according to the TS prediction of the RF model. The facial skin temperature, local temperature difference, air temperature and humidity, and the operating status of the PCS were taken as the input features of the RF model. Results indicated that the developed PCS achieved about a 4% improvement in TS prediction accuracy and realized good automatic operation to guarantee the thermal comfort of occupants. In [138], a PTCM was constructed based on Bayesian multinomial logistic regression (MLR). Combined with a low-cost wireless sensor network and MPC strategy, the collaborative optimization of the individual comfort and energy efficiency of the HVAC system was achieved in office buildings. In [139], a cooperative control system with PCS as the core was proposed, enabling communication between thermal state recognition, HVAC, and PCS. Specifically, the author established the RF-based PTCM using facial skin temperature and environmental parameters as input features. Recognition results of the RF model were used to start and stop the desk fans to regulate the local environment around the human body, and the HVAC was controlled according to the operating status of the PCS. Experimental results demonstrated that the proposed system could automatically regulate environments according to the TS prediction, thereby improving the occupants’ thermal comfort. In [140], a personalized thermal comfort monitoring and control system was developed based on the non-contact TS prediction. Thermal sensitive regions were recognized based on facial thermography and the YOLO algorithm. Average temperatures of these regions were extracted for the establishment of the RF-based personal TS model. Considering possible transient detection interruption, the PMV model was also combined with the established RF model in a parallel ensemble way. On the basis of personal TS prediction, a fuzzy controller was designed to regulate the HVAC through infrared signal matching. Field experiments showed the ensemble TS prediction model achieved the high accuracy of 96.07% (5-point scale). Compared with the fixed setpoint controller, the designed controller could bring about a 50% reduction of thermal stabilization time and HVAC energy consumption.

Overall, the current literature on PTCMs’ integration has obtained rather encouraging results. However, some issues still need to be noted: (1) Most of them (4/6 in Table 7) put focus on occupants’ thermal comfort improvement without considering energy savings issues. This may be attributed to difficulties in accurate quantification or measurement of energy savings in some practical situations. (2) Most of them adopted traditional control methods (MPC, fuzzy logic control, and discrete feedback control). Among them, MPC optimizes the operation of HVAC by predicting the future indoor environment and demonstrates higher efficiency in balancing energy consumption and thermal comfort. However, following shortcomings of MPC restrict its practical applications: strong dependence on building thermal dynamics and physical installation variation between buildings. Compared with MPC, advanced digital twin-based framework [141] (which enables the creation of a virtual replica of the building’s physical environment) and model-free methods (such as neural network control and reinforcement learning control) have the potential to achieve better control performance.

Table 7. Studies on PTCM-based building environment control systems.

Reference (Year)	Control Method	Actuator	Thermal Comfort Model	Control Effect Based on Field Experiments
[139] (2022)	Cooperative control	HVAC, PCS	RF	Regulate environments at different spatial scales automatically. Improve thermal comfort.
[97] (2023)	On-off control	PCS	RF	Achieve the TS prediction accuracy of 69%. Ensure the comfort of the personnel.
[138] (2023)	MPC	HVAC	MLR	Increase occupant thermal comfort. Achieve 28–35% energy savings vs. static setpoint control.
[142] (2020)	Multi-level automation	PCS	RF, KNN	Realize different levels of automation control. Improve occupant satisfaction from $-0.643$ to $1.143$.
[143] (2023)	Discrete feedback control	HVAC	RF	Achieve a high TS prediction accuracy of 0.84. Accomplish automatic HVAC control to ensure indoor thermal comfort.
[140] (2026)	Fuzzy logic control	HVAC	Ensemble model using RF and PMV	Achieve the TS prediction accuracy of 96.07% (5-point scale). Reduce both the thermal stabilization time and energy consumption.

Table 7. Studies on PTCM-based building environment control systems.

Reference (Year)	Control Method	Actuator	Thermal Comfort Model	Control Effect Based on Field Experiments
[139] (2022)	Cooperative control	HVAC, PCS	RF	Regulate environments at different spatial scales automatically. Improve thermal comfort.
[97] (2023)	On-off control	PCS	RF	Achieve the TS prediction accuracy of 69%. Ensure the comfort of the personnel.
[138] (2023)	MPC	HVAC	MLR	Increase occupant thermal comfort. Achieve 28–35% energy savings vs. static setpoint control.
[142] (2020)	Multi-level automation	PCS	RF, KNN	Realize different levels of automation control. Improve occupant satisfaction from $-0.643$ to $1.143$.
[143] (2023)	Discrete feedback control	HVAC	RF	Achieve a high TS prediction accuracy of 0.84. Accomplish automatic HVAC control to ensure indoor thermal comfort.
[140] (2026)	Fuzzy logic control	HVAC	Ensemble model using RF and PMV	Achieve the TS prediction accuracy of 96.07% (5-point scale). Reduce both the thermal stabilization time and energy consumption.

5. Discussion

This study systematically reviews the research progress in the field of data-driven personal thermal comfort modeling. Compared with traditional PMV and adaptive models, data-driven models (particularly using DL and TL techniques) demonstrate significant advantages in capturing the individual thermal preference and improving prediction accuracy. This finding not only confirms the great potential of AI technology in handling high-dimensional, non-linear physiological and environmental data, but also lays the foundation for building truly personalized, adaptive indoor environmental control systems. However, pushing PTCM from research to practical applications still faces several challenges worthy of further research. Figure 6 presents a concise roadmap to highlight these challenges and future directions.

5.1. Data Scarcity/Quality Issues

In the field of data-driven thermal comfort modeling, the high cost, limited scale, and low quality of data acquisition (especially personal physiological data) are the core bottlenecks, which directly restrict the accuracy and cross scenario generalization ability of the model. Collecting personal physiological markers usually requires physiological sensors to be used in special experimental scenarios. The entire procedure is expensive and time-consuming. In addition, due to individual differences in thermal sensation, the distribution characteristics of the dataset may vary with different populations and climate regions, resulting in poor generalization of the prediction model.

To overcome this limitation, transfer learning, as an efficient machine learning paradigm, provides an important technical path. Specifically, when local data of the target population is scarce, the performance of the target model can be significantly improved by leveraging large-scale thermal comfort data that already exists in the source domain (such as buildings in other climate zones, public datasets, or simulation platforms). Although transfer learning shows great potential in addressing data scarcity, it has the following several limitations [144,145,146]: (1) Bias propagation across occupants and climates, particularly when the target dataset is rather small. (2) Degradation of comfort prediction under domain shift. (3) Ethical and privacy concerns linked to personal data.

Regarding the aforementioned limitations, further explorations are necessary to improve the performance of transfer learning: (1) Bias-aware transfer learning using multi-branch or multi-task architectures and meta-learning approaches, which can be employed to disentangle occupant- and climate-specific factors and enable rapid personalization with limited target data. (2) Feature selection methods to select model features that are insensitive to domain changes, or domain adaptive methods including feature alignment, model parameter adaptation, and model adversarial training. (3) Data anonymization and de-identification, federated learning with secure aggregation, privacy-preserving feature extraction, and end-to-end encryption with strict access control.

5.2. Low-Resolution Thermal Comfort Assessment Issues

Real-time, personalized thermal comfort regulation relies on continuous and precise individual thermal state perception. Facial temperature or other body skin temperatures can reflect extreme states of “hot” or “cold”. However, it is difficult to accurately distinguish the subtle and most comfortable thermal sensation levels, such as “neutral cool” and “neutral warm”, resulting in insufficient granularity for thermal comfort regulation.

An effective solution is to acknowledge the inherent limitations of non-contact perception and treat humans as “smart sensors” and ultimate judges in a closed loop, compensating for the shortcomings of machines with minimal and most natural interactions. For instance, monitor personal behaviors such as adjusting fans or standing up to open windows, and use them as more reliable signals of thermal discomfort than facial temperature. When the model uncertainty is high or the regulatory effect does not meet expectations, obtain high-quality labels through minimalist interactions (such as smart speakers asking “Do you feel hot now?” or asking users to provide quick feedback with “thumbs up/down” on the phone screen) to fine tune the model online.

5.3. Integration Issues with Building Environment Control

The consensus among PTCM developers is that, eventually, the PTCM should be applied to building environment control. Currently, most PTCM-related literature focuses on the advanced modeling techniques rather than applying the PTCM to actual control scenarios [143]. Some studies have achieved the integration of PTCM with the HVAC system or PCS and demonstrated the potential of integrated systems in improving thermal comfort and energy efficiency. Nevertheless, these studies still remain at the stage of simulation or small-scale experimental validation. From the practical application perspective, quite a few objective difficulties exist in PTCMs’ integration, e.g., cost and complexity of integration, performance gaps between laboratory-based and real-world deployments, and limitations on protocol compatibility (see Section 4). In addition, there are some engineering challenges as follows: (1) field implementations of advanced control algorithms (such as reinforcement learning); (2) long-term robustness of AI-based PTCMs; (3) scalability of PTCMs at different spatial scales (zone, building, and district); and (4) efficiency and reliability of the integrated system.

In response to the above challenges, future research can focus on the following aspects: (1) Extensive verification of advanced control algorithms using long-term validation in real, large-scale building environments. (2) Online learning frameworks and adaptive learning mechanisms for PTCMs. (3) Spatial scalability enhancement of PTCMs, including cross regional data sharing, and deployment from a small number of sensors to large-scale sensor networks. (4) Advanced fault diagnosis and predictive maintenance of the integrated system [147], data-driven HVAC efficiency enhancement, and adaptive thermostat scheduling.

6. Conclusions

This study reviews some of the current literature published in the area of personal thermal comfort modeling using advanced data-driven techniques. Five parts are highlighted as follows: (1) Thermal comfort data collection is discussed, including four data types (environmental data, physiological data, behavioral data, and subjective feedback data), two collection ways (contact-based and non-contact), and data correlation analysis methods. (2) A comprehensive review is conducted on personal thermal comfort modeling using ML/DL and evaluation indicators for PTCMs. It is known that SVM, RF, and KNN are the three most commonly used ML-based models. Compared with these ML-based models, DL-based models (e.g., CNN, LSTM, residual network with SAM) exhibit higher accuracy and stronger robustness in predicting individual thermal comfort. (3) The TL-based thermal comfort modeling in sample scarcity scenarios is investigated. Four transfer methods are analyzed and compared. It is concluded that the FBTL and ensemble methods outperform the MBTL and IBTL methods when the data distribution difference between the target and source domains is significant. (4) From the practical point of view, the PTCM-based building environment control systems are reviewed and discussed. The application potential of integrating the PTCM with thermal control is analyzed in terms of thermal comfort improvement and building energy saving. (5) Some challenges and future directions are proposed from perspectives of data scarcity, low-resolution assessment of thermal comfort, and integration of PTCMs into control systems.