Predicting Human Thermal Comfort During Winter Heating Using Multi-Class Machine Learning Algorithms

Tongwen Wang; Weijie Huang; Haiyan Yan; Jingyuan Gao; Yawei Li; Yongxuan Guo

doi:10.3390/pr14050875

Abstract

To address the critical need for accurate human thermal comfort prediction in winter heating environments, this study established a comprehensive thermal comfort dataset containing 2089 valid samples through experiments. On this basis, thermal comfort prediction models were constructed using three multi-class machine learning algorithms: Support Vector Classification, K-Nearest Neighbors, and Random Forest. The predictive performance of 63 different feature combinations was systematically evaluated. The results indicate that the feature subset comprising indoor air temperature, forehead temperature, cheek temperature, dorsal hand temperature, heart rate, and systolic blood pressure yields the optimal prediction performance. Among the evaluated models, the Random Forest model demonstrated superior overall performance, achieving an accuracy exceeding 90% and an AUC ranging from 96% to 99%, significantly outperforming the SVC and KNN models. Compared with the traditional Predicted Mean Vote (PMV) model, the machine learning models developed in this study showed a substantial improvement in prediction accuracy under identical conditions; notably, the Random Forest model improved accuracy by approximately 40% over the PMV model. Based on these findings, a smart heating system framework integrating environmental sensors, wearable devices, and intelligent control valves is proposed, providing a theoretical basis and technical approach for realizing personalized and energy-efficient heating control.

Keywords:

thermal comfort prediction; multi-class machine learning; random forest; smart heating; feature optimization

1. Introduction

According to data from the International Energy Agency, the three major sectors contributing to global energy consumption and carbon emissions are the building industry, transportation, and other industries. Notably, the building sector accounts for 35% of total energy consumption and 38% of carbon emissions [1]. In China, energy consumption in the building sector constitutes nearly 20% of the total national energy consumption. The growing demand for heating has led to substantial building energy consumption and significant carbon emissions [2], making winter heating a matter of widespread public concern [3].

Currently, most cities in northern China adopt a centralized heating mode [4]. The working principle of this system involves transporting hot water from a centralized heat source to various users via a thermal pipe network, which ensures stable and safe continuous heating 24 h a day. With the rapid development of urbanization in China, the application of centralized heating systems has become increasingly extensive. However, due to a lack of intelligent control, this heating mode often suffers from issues such as supply-demand imbalances and energy wastage [5]. Consequently, it is imperative to construct a more efficient heating regulation system, which fundamentally requires the ability to accurately predict the thermal comfort of residents.

Since the 1970s, scholars have progressively established a series of thermal comfort models to predict human thermal sensation. Among them, the Predicted Mean Vote (PMV) model [6], established by Fanger based on experimental data, has been widely recognized. However, in practical applications, numerous studies have found that PMV predictions often deviate from subjective experiences [7,8,9,10,11,12,13,14]. Furthermore, some input variables required by the PMV model are difficult to collect in actual building environments. Therefore, the development of new thermal comfort prediction models is necessary.

In research related to model prediction, data-driven machine learning methods have demonstrated excellent performance in prediction accuracy, becoming a crucial component of intelligent heating systems. However, despite progress in data-driven thermal comfort prediction, several issues in the existing literature may still warrant further investigation. First, a considerable portion of existing machine learning studies tend to rely on large-scale public datasets [15,16,17]. While such data are valuable for identifying global trends, it may be challenging for them to fully capture the nuanced effects of specific building operation modes (e.g., centralized heating).

Second, in current predictive modeling practices, “thermal sensation” is more frequently adopted as the model output than “thermal comfort”, the latter being more closely related to subjective pleasure [18,19]. This is because thermal sensation, as an objective description of one’s thermal state [20,21], is often less difficult to predict. However, models based solely on sensation may not fully reflect the subjective satisfaction of occupants.

Third, the rationale for selecting input features is sometimes context-specific or constrained by the dataset used. There remains room for discussion regarding how different combinations of multi-source features (environmental, physiological, and behavioral) affect the performance of machine learning algorithms.

Against this background, this study attempts to construct a localized thermal comfort dataset specifically for winter centralized heating conditions in the cold region of China. In the research design, we tend to employ a thermal comfort scale rather than a thermal sensation scale. Furthermore, using a consistent evaluation protocol, we conduct a systematic comparative analysis of 63 multi-factor, multi-type coupled feature combinations across three representative classifiers. The aim is to provide a data reference for building a deployable intelligent closed-loop control framework that balances individual comfort and energy efficiency.

2. Methodology

The research methodology is divided into five main components: data collection, data processing and feature engineering, model training, baseline model comparison, and control framework development. Figure 1 illustrates the overall workflow of this study. First, experiments were conducted according to the research objectives, and thermal comfort data were collected. The data were then integrated and cleaned. The processed data were used to train three different machine learning models, after which the machine learning models were compared with the PMV model. Finally, based on the requirements of intelligent heating, an intelligent heating framework equipped with a machine learning prediction system was proposed. The research procedure of this study is illustrated in Figure 1.

Figure 1. Research procedure.

2.1. Experimental Site and Participant Recruitment

The experiment was conducted from October 2020 to April 2021 in a laboratory at a university in Jiaozuo, Henan Province, China. The laboratory was equipped with typical centralized heating terminals. To minimize the potential confounding effects of gender, age, and social background on the results, junior university students were recruited as participants. The cohort consisted of 32 students (16 males and 16 females), all of whom had lived locally for at least one year and were acclimatized to the local climate. This study has obtained verbal informed consent from the participants to publish this paper along with any accompanying data and images. Prior to the experiment, participants were instructed to avoid alcohol, caffeine, strenuous exercise, and smoking, and to ensure adequate sleep the night before. During the experiment, participants remained in a sedentary seated position, with an estimated metabolic rate of 1.2 met. The basic information of the participants is presented in Table 1.

Table 1. Basic information of occupants.

2.2. Experimental Design

During the experiment, staff monitored the completion of questionnaires and physiological measurements in real time. The experimental procedure is illustrated in Figure 2. The collected data included basic participant information, indoor and outdoor environmental parameters, physiological parameters, and subjective evaluations.

Figure 2. Experimental flow chart.

Regarding environmental parameters, indoor air temperature (Ta) was measured. In accordance with ISO 7726 [22] and ASHRAE 55-2020 standards [23], the measurement instruments were positioned within 0.3 m of the participants. The recorded values represent the temperature at the waist (0.6 m) level.

Regarding physiological parameters, local skin temperatures (forehead, cheek, and dorsal hand; all measured on the left side), heart rate (HR), and systolic blood pressure (SBP) were recorded using a multi-channel heat flow tester and an automated blood pressure monitor, respectively. The instrument specifications are provided in Table 2. The skin measurement points are shown in Figure 3a, and the field test setup is depicted in Figure 3b.

Table 2. Field test parameters and measuring instruments.

Figure 3. Measurement setup. (a) Schematic diagram of skin measurement points. (b) Measurement height of environmental parameters.

During the test, participants reported their real-time clothing status in the questionnaire. To simulate realistic conditions, participants were not required to wear a fixed uniform. The thermal resistance values for single garments were determined based on ISO 7730-2005 [24] and GB 50785-2012 standards [25]. Given the sedentary posture of the participants, a correction of 0.15 clo [26] was applied to the clothing insulation values to account for the thermal insulation of the chair.

The subjective questionnaire assessed thermal comfort using the voting scale presented in Table 3.

Table 3. Thermal comfort voting scale.

A total of 2210 data samples were obtained. After removing 121 samples containing errors or significant omissions, 2089 valid samples remained for analysis.

2.3. Data Preprocessing

Before model training, the raw dataset was preprocessed to ensure data quality and consistency. Outliers in the input variables were removed using the interquartile range (IQR) criterion, and the cleaned dataset was used for all subsequent modeling. Samples with missing values in any required input feature or in the target label (thermal comfort) were excluded. The thermal comfort labels were treated as a multi-class target and encoded accordingly. For models sensitive to feature scales (SVC and kNN), input features were standardized, while Random Forest does not require scaling.

2.4. Algorithm Selection and Performance Evaluation

Machine learning is generally categorized into supervised and unsupervised learning. Supervised learning involves training a model to predict outputs for given inputs and includes classification and regression tasks. Unsupervised learning is primarily used for clustering, dimensionality reduction, and association. Given the objective of this study, supervised learning was selected. Common supervised learning methods are listed in Table 4.

Table 4. Classification of supervised learning methods.

This study selected three classification algorithms—Support Vector Classification (SVC), K-Nearest Neighbors (KNN), and Random Forest (RF) (an ensemble method based on decision trees)—to model the preprocessed thermal comfort data. These algorithms were chosen according to the following criteria: (1) suitability for multi-class classification on tabular data with potentially non-linear decision boundaries, which is typical for combined physiological and environmental measurements; (2) complementary learning principles to support a robust comparison—SVC is a margin-based classifier, KNN is an instance-based method, and RF is an ensemble tree approach that can capture non-linear interactions; and (3) practical computational complexity for real-time or near-real-time implementation in intelligent control applications. These three methods are also widely used as strong classical baselines in indoor comfort-related classification tasks [32,33,34].

Model performance was evaluated using Accuracy (reflecting the proportion of correctly classified samples), f1-score (a comprehensive metric balancing precision and recall), and AUC (Area Under the ROC Curve, assessing the overall discriminative ability of the classifier). The detailed calculation methods for each metric are provided in Appendix A.

To obtain a robust and reproducible performance estimate, we adopted a two-stage evaluation protocol. First, the cleaned dataset was split into a stratified training set and an independent hold-out test set (80/20), and the test set was reserved for the final unbiased assessment (reported in Section 2.6 and Section 3.1). Second, we performed stratified k-fold cross-validation (CV) on the training set to compare classifiers (SVC, kNN, and RF) and feature-group configurations, reporting mean ± standard deviation of Accuracy and Macro-f1. Because the least frequent thermal-comfort class contains only two samples in the training set, a fixed 5-fold stratified CV is mathematically infeasible; therefore, k was constrained by the minimum class size in the training set (k ≤ 5), which resulted in stratified 2-fold CV for this dataset. Macro-f1 was emphasized in CV because it is more informative for multi-class imbalanced data.

Considering the various combinations of input features and the distinct characteristics of the three algorithms, this study conducted a comparative analysis. This included a horizontal comparison of different feature subsets and a vertical comparison of the three algorithms, providing a theoretical reference for feature selection in intelligent building control and the application of classification algorithms in thermal comfort research.

2.5. Feature Engineering

Thermal comfort prediction models based on SVC, KNN, and RF were constructed using Python 3.12.1. To fully account for the comprehensive impact of environmental, physiological, and behavioral factors on human thermal comfort, a multi-factor, multi-type coupled classification approach was adopted. The specific combinations of input feature parameters are detailed in Table 5.

Table 5. Input feature parameter combinations.

Table 5 lists the feature-group configurations used in this study. The groups were constructed based on a deployment-oriented assumption that indoor air temperature (Ta) is always available and should be included as a mandatory input for thermal comfort prediction and subsequent control. Under this assumption, the remaining six variables were combined using all possible non-empty subsets, resulting in 2⁶ − 1 = 63 feature groups. Hence, the 63 groups in Table 5 constitute an exhaustive enumeration of all combinations subject to the Ta-mandatory constraint.

2.6. Algorithm Establishment

The thermal comfort prediction models were developed on the Jupyter Notebook 6.5.4 platform using Python. The input features comprised three categories: environmental parameters (indoor air temperature, Ta), physiological parameters (forehead temperature T_skforehead, cheek temperature T_skcheek, back of hand temperature T_skBH, heart rate HR, systolic blood pressure SBP), and clothing thermal resistance (clo). The pseudo-code for the modeling process is shown in Table 6.

Table 6. Pseudo-code of multi-factor and multi-type coupling thermal comfort prediction algorithm.

As indicated in the pseudo-code, the algorithm divides the dataset into a training set (D_train) and a testing set (D_test). The three algorithms were used to learn the initial classification patterns on the training set to obtain the models (M). These models were then tested and adjusted on the testing set to obtain the final models (Madjust), with Accuracy, f1-score, and AUC calculated simultaneously.

3. Model Performance Comparison

3.1. Class Distribution

Table 7 summarizes the distribution of thermal comfort classes in the overall dataset and in the stratified 80/20 train–test split. The labels are clearly imbalanced: class “1” and class “2” account for 38.34% and 43.14% of all samples, respectively, while minority classes are scarce (e.g., class “−3” includes only 2 samples overall). This imbalance motivates the use of stratified splitting and the reporting of Macro-f1 in addition to accuracy. It also constrains stratified cross-validation; therefore, k was set according to the minimum class size in the training set (k ≤ 5), which resulted in stratified 2-fold cross-validation for this dataset.

Table 7. Distribution of thermal comfort classes in the overall dataset and in the stratified 80/20 train–test split.

3.2. Performance Results of Support Vector Classification (SVC)

Thermal comfort models based on Support Vector Classification (SVC) were established for the 63 different feature parameter combinations listed in Table 5. The evaluation metrics (Accuracy, f1-score, and AUC) for these combinations on the test set are presented in Table 8.

Table 8. Evaluation criteria for classification support vector machine thermal comfort model.

Based on the performance results, specific input feature combinations—including Group 12, Group 16, Group 22, Group 26, Group 32, Group 39, Group 41, Group 42, Group 49, Group 51, Group 53, Group 55, Group 57, Group 58, Group 61, and Group 62—were selected. For these combinations, all three evaluation metrics exceeded 60%.

Overall, the Accuracy and f1-score of the SVC models were concentrated in the range of 60–64%, while the AUC ranged from 63% to 70%. This indicates that the predictive performance of the thermal comfort models established using SVC falls within the 60–70% range.

3.3. Performance Results of K-Nearest Neighbors (KNN)

Thermal comfort models based on the K-Nearest Neighbors (KNN) algorithm were established for the 63 feature combinations listed in Table 5. The evaluation metrics for the different combinations on the test set are presented in Table 9.

Table 9. Evaluation standard of k-nearest Neighbor algorithm thermal comfort model.

Based on the performance results, feature combinations such as Group 22, Group 23, Group 35, Group 38, Group 43, Group 44, Group 46, Group 52, Group 55, Group 56, Group 57, Group 59, Group 62, and Group 63 were selected. For these combinations, all three evaluation metrics exceeded 75%.

Overall, the Accuracy and f1-score of the KNN models were concentrated in the range of 75–79%, while the AUC metrics were primarily concentrated in the range of 80–87%. This indicates that the predictive performance of the thermal comfort models based on KNN falls within the 75–87% range.

3.4. Performance Results of Random Forest (RF)

Thermal comfort models based on the Random Forest (RF) algorithm were established for the 63 feature combinations listed in Table 5. The evaluation metrics for the different combinations on the test set are presented in Table 10.

Table 10. Evaluation criteria of thermal comfort model based on random forest algorithm.

Based on the performance results, feature combinations such as Group 23, Group 29, Group 30, Group 31, Group 41, Group 44, Group 45, Group 47, Group 49, Group 50, Group 51, Group 53, Group 55, Group 56, Group 57, Group 58, Group 59, Group 60, Group 61, Group 62, and Group 63 were selected. For these combinations, all three evaluation metrics exceeded 90%.

Overall, the Accuracy and f1-score of the RF models were concentrated in the range of 90–95%, while the AUC metrics were primarily concentrated in the range of 96–99%. This demonstrates that the predictive performance of the thermal comfort models established using RF consistently falls within the 90–99% range.

3.5. Comparison of Features and Model Performance

This study employed three machine learning algorithms—SVC, KNN, and RF—to analyze thermal comfort prediction across different input feature combinations. The selection of the optimal feature combination and prediction model was based on a comprehensive consideration of feature convenience, rationality, and model performance metrics (Accuracy, f1-score, and AUC).

The results indicated that the performance differences between certain feature combinations were minimal (approximately 1–2%). Consequently, to prioritize practical application convenience, non-essential features could be excluded. For instance, Group 57 does not include clothing insulation, whereas Group 63 does. A comparison of the models reveals that the differences in all three performance metrics for SVC and KNN were less than 1%. For the RF model, the difference in AUC was less than 1%, and the differences in Accuracy and f1-score were less than 4%. This suggests that winter clothing insulation has a limited impact on the performance of the thermal comfort prediction model. Considering that winter clothing insulation is relatively stable and its regulation exhibits hysteresis in reality, it was ultimately excluded from the input features.

On the other hand, skin temperatures of the forehead, cheek, and dorsal hand are sensitive to environmental changes, while heart rate and systolic blood pressure directly reflect physiological thermoregulatory states. These features were included in most high-performance combinations (e.g., Group 57), yielding superior predictive performance.

Comparing the three algorithms, the Random Forest model significantly outperformed the other two (performance metrics reaching 96–99%), whereas the SVC and KNN models were concentrated in the 60–70% and 75–87% ranges, respectively. These results demonstrate that Random Forest can more effectively identify the complex relationships between features and thermal comfort, making it suitable for the prediction tasks in this study.

To address potential split-dependence of a single hold-out evaluation, we additionally report stratified cross-validation results on the training set. Table 11 summarizes the CV performance (mean ± std) of the three models under the selected best feature group (Group 57), together with the corresponding hold-out test performance for direct comparison. Although CV typically provides a more conservative estimate than a single split, the relative ranking of the models is consistent across CV and the independent test set (RF > SVC and kNN), supporting the robustness of the model selection and the conclusions drawn in this study.

Table 11. Stratified cross-validation (CV) and hold-out test performance.

3.6. Comparative Analysis of Machine Learning Models and the PMV Model

This section compares the parameter combinations and prediction models derived from machine learning algorithms with those used in the traditional PMV model under identical conditions to determine which is superior. Accuracy was used as the comparative metric. The results are shown in Table 12.

Table 12. Comparative analysis of thermal comfort prediction models.

As indicated in the table, the thermal comfort prediction models established using machine learning algorithms in this study consistently achieved higher prediction accuracy than the traditional PMV equation. Compared to the six factors of the traditional PMV model (indoor air temperature, mean radiant temperature, air velocity, relative humidity, clothing insulation, and metabolic rate), the parameter combinations derived via machine learning algorithms are more aligned with human perception of current thermal environment changes. Specifically, the predictive performance of the SVC model improved by 9.49% over the PMV model, the KNN model by 26.31%, and the RF model by 40.18%.

4. Discussion

To address the prevalent issues in the centralized heating sector, such as severe energy wastage, supply-demand imbalances, and the lack of intelligent control, this study proposes a targeted solution. By integrating the Random Forest-based thermal comfort prediction model established herein with sensors, wearable devices, and imaging technology, we designed an intelligent heating control device and system. The intelligent heating control device is a novel smart control valve. Its regulation feedback mechanism primarily relies on real-time data collection via indoor environmental monitoring systems and wearable wristbands. Based on the thermal comfort levels predicted by the Random Forest model, the valve opening angle is adjusted to achieve the objective of intelligent heating regulation.

4.1. Analysis of the Mechanisms Underlying Model Performance Differences

The experimental results in Section 3 demonstrate that the Random Forest (RF) model significantly outperforms both Support Vector Classification (SVC) and K-Nearest Neighbors (KNN) in predicting thermal comfort. While the previous section presented the performance metrics, this part delves into the underlying reasons for this performance gap, focusing on how the intrinsic characteristics of each algorithm interact with the specific nature of the thermal comfort dataset.

The superior performance of the RF model can be attributed to several key factors inherent to its ensemble learning mechanism. First, thermal comfort is a complex phenomenon resulting from non-linear interactions between environmental (e.g., Ta), physiological (e.g., skin temperatures, HR), and personal (e.g., clo) factors. As an ensemble of decision trees, RF is inherently adept at capturing these complex, non-linear relationships and high-order feature interactions without requiring extensive feature engineering or data transformation [35,36]. Each decision tree in the forest learns different splitting rules based on different feature subsets, and the final prediction is an aggregation (e.g., majority voting) of these trees. This process effectively models the multifaceted nature of the perception of the thermal environment. Second, RF provides a robust measure of feature importance. This inherent capability allows the model to automatically weigh the predictive power of different inputs (e.g., it might find that skin temperature is a stronger indicator than heart rate for certain comfort levels), leading to a more nuanced and accurate decision boundary. This aligns with our feature selection process in Section 3.4, where models including physiological signals like skin temperature consistently performed better. Third, the bagging (Bootstrap Aggregating) procedure used by RF introduces randomness into the training process, which significantly reduces the model’s variance and makes it less prone to overfitting compared to a single decision tree [37]. This robustness is critical for generalizing well to unseen data, as reflected in the high and stable performance metrics (Accuracy > 90%, AUC 96–99%).

In contrast, the weaker performance of SVC and KNN can be explained by their respective algorithmic limitations when applied to this specific problem. The SVC model’s performance, concentrated in the 60–70% range, suggests that its underlying principle of finding an optimal separating hyperplane may be less suited for this dataset. While SVC with a non-linear kernel (like the RBF kernel used in this study) can model complex boundaries, its performance is highly sensitive to the choice of kernel parameters and feature scaling. The high-dimensional feature space created by the combination of environmental and physiological data may make it challenging for the SVC to find a clean, maximal-margin hyperplane that separates the six comfort categories effectively [30]. Furthermore, the one-vs-one or one-vs-rest strategies required for multi-class classification with SVC can lead to ambiguous regions and reduced confidence in predictions, especially when class boundaries are fuzzy, as is often the case with subjective thermal comfort votes.

The KNN model, while outperforming SVC, still lagged significantly behind RF. As an instance-based learner, KNN’s prediction relies on measuring the distance (e.g., Euclidean distance) between a new data point and its nearest neighbors in the training set [27]. Its performance is therefore highly dependent on the quality of the distance metric and the local structure of the data. In a dataset with 2089 samples across six comfort classes, the feature space is dense but not infinite. The “curse of dimensionality” can affect KNN, where distances become less meaningful in high-dimensional spaces. Moreover, KNN assigns equal weight to all features unless feature weighting is explicitly implemented. In contrast to RF, which automatically learns feature importance, KNN may be equally influenced by a highly predictive feature (e.g., forehead temperature) and a less relevant one, potentially diluting its predictive accuracy. This explains why KNN’s accuracy (75–79%) is good but not exceptional, as it lacks the sophisticated feature interaction modeling and importance weighting that RF provides.

In summary, the RF model’s ensemble nature, its ability to model non-linear interactions, and its inherent robustness to overfitting make it exceptionally well-suited for the complex, multi-class classification task of predicting human thermal comfort from a mixed set of environmental and physiological features.

4.2. Feasibility Analysis of Intelligent Heating

Traditional studies have predominantly correlated environmental parameters with thermal comfort for prediction [38,39,40]. However, relying solely on environmental parameters is insufficient, as individuals may experience different thermal sensations under identical environmental conditions [41,42]. Physiological parameters—particularly skin temperature [43], which is sensitive to environmental changes—are closely linked to thermoregulatory mechanisms [44] and can better reflect individual differences. Therefore, incorporating them into the model is rational.

With technological advancements, the development of data collection and data-driven methods has made the precise prediction of thermal comfort using new variables possible [45,46]. Meanwhile, the mature application of sensors and wearable devices (such as wristbands capable of monitoring heart rate and skin temperature) [47,48] provides a viable technical pathway for non-invasive, real-time data acquisition and personalized prediction. These technologies can be integrated into the intelligent control of HVAC systems via the Internet of Things (IoT).

4.3. Design of Intelligent Heating Scheme

Based on the feasibility analysis, the intelligent heating scheme designed in this study primarily comprises intelligent heating control valves, indoor environmental sensors, wearable devices, a central control panel, and mobile terminals. The operational principle is illustrated in Figure 4.

Figure 4. Working principle diagram of intelligent heating.

The proposed framework consists of four components, namely the data acquisition layer, the smart central processing unit, the control/actuation layer, and the self-learning module. Its working principle is described as follows.

(1) Multi-source sensing and data uploading (Sensing & Uploading): The system collects data from an indoor temperature sensor (Temperature Sensor), a wearable wristband (Wearable Wristband), and user behavior collection (Behavior Collection). The indoor temperature reflects variations in the thermal environment; the wearable device captures physiological signals such as skin temperature; and behavior collection is used to characterize activity states and changes in subjective preferences. These data streams are preliminarily aggregated on the mobile device or gateway and then transmitted via Data Upload to provide inputs for subsequent prediction and control.

(2) Data fusion and model prediction (Fusion & Prediction): After uploading, the data enter the Smart Central Processing Unit. This unit performs data cleaning, alignment, and fusion across sources (e.g., outlier handling, missing-value processing, and time synchronization), and then feeds the fused features into the multi-class thermal-comfort prediction model developed in this study (Model Prediction). The model output represents the user’s current (or short-term predicted) thermal-comfort category, which quantifies the control demand, i.e., whether heating should be increased, decreased, or maintained. To improve practical usability, the mobile App simultaneously conducts App Data Collection and allows users to provide preference inputs or manual settings (User Control via Mobile Phone). Such information can be used as a correction term in the control strategy (e.g., warm/cool preference or target comfort level) to accommodate individual differences.

(3) Control decision and actuation (Decision & Actuation): The control layer comprises two complementary pathways.

Macro-regulation pathway (Macro-regulation of Heating Temperature): The heating regulation department (Heating Regulation Department) determines a baseline heating temperature at a slower time scale based on historical operation and statistical analyses (Data Comparison and Analysis). This pathway is designed to address low-frequency, large-amplitude changes in demand, such as outdoor weather variations and building-wide load shifts.

Real-time regulation pathway (Real-time Regulation): On top of the macro-regulation baseline, the system generates real-time control commands according to the model prediction and user inputs, and applies them to the Smart Flow Control Valve. By adjusting the heating flow rate/valve position, the valve enables Real-time Heating Temperature Regulation to respond to rapid changes in human state and local thermal conditions. Together, the two pathways provide both global stability (ensured by macro-regulation) and individualized, fast responsiveness (achieved by real-time regulation).

(4) Feedback and self-learning updates (Feedback & Self-learning): During operation, changes in the thermal environment after actuation and user feedback (including physiological responses and possible manual adjustment records) are fed back to the Self-learning module within the smart central processing unit. This module periodically updates model parameters or optimizes the mapping from predicted comfort categories to control actions (e.g., valve-adjustment magnitude/thresholds), allowing the system to gradually adapt to individual differences and environmental variations over long-term operation and to improve the stability and generalization of both prediction and control. In this way, a closed-loop iterative process of “data–model–control–feedback–re-learning” is formed.

In summary, the framework in Figure 4 integrates multi-source sensing data with the thermal-comfort prediction model, establishes coordination between the macro-level heating baseline and local real-time regulation, and continuously improves performance through self-learning, thereby forming an intelligent closed-loop control system for thermal comfort.

4.4. Limitations and Future Outlook

Although this study has made progress in establishing machine learning-based thermal comfort prediction models and designing an intelligent winter heating scheme, certain limitations remain. The dataset was primarily derived from a student cohort at a university in Jiaozuo, Henan, which limits the representation of climate zones, building types, and demographic diversity. This may affect the model’s generalization capability in broader scenarios. Future work could involve collecting heterogeneous data across multiple climate zones and building types and utilizing transfer learning techniques to construct more adaptive models.

The PMV model is a physics-based heat-balance approach, primarily used to estimate thermal sensation under steady-state assumptions with representative values of metabolic rate and clothing insulation. In contrast, the machine-learning models developed in this study are trained to predict thermal comfort; although the two constructs are related, they are not equivalent. The PMV-related results are presented only as a reference to a conventional physical index and to highlight the potential of data-driven models for predicting more complex, preference-related outcomes, rather than to claim that PMV “underperforms” the proposed ML models on the same target. Establishing a rigorous and fair comparison framework between different constructs such as thermal sensation and thermal comfort remains an open problem. Future work will explore more advanced approaches to enable more principled cross-metric comparisons and interpretations.

Additionally, the proposed intelligent heating system relies on reliable multi-source data acquisition and transmission; its cost-effectiveness, user acceptance, and long-term stability in actual deployment require further verification.

5. Conclusions

Addressing the issue of imprecise regulation in the thermal environment of centralized heating buildings, this study provided a solution for intelligent thermal comfort prediction and control through experiments and data-driven modeling. The conclusions are as follows:

(1) A comprehensive dataset containing indoor air temperature, multi-site skin temperatures, heart rate, and systolic blood pressure was constructed through experiments. Three machine learning models—Support Vector Classification (SVC), K-Nearest Neighbors (KNN), and Random Forest (RF)—were systematically compared. The Random Forest model exhibited optimal performance, with a prediction accuracy exceeding 90%, significantly outperforming the other algorithms.

(2) The study identified that the Group 57 parameter combination (indoor air temperature, heart rate, systolic blood pressure, forehead temperature, cheek temperature, and dorsal hand temperature) performed optimally across all three machine learning models. This combination demonstrated high stability within the models.

(3) Under identical conditions, the prediction accuracy of the data-driven models established in this study comprehensively surpassed that of the traditional PMV model, with the optimal model showing a performance improvement of over 40%.

(4) Based on the aforementioned models, a system scheme integrating environmental sensors, wearable devices, and intelligent control valves was proposed.

Future research will explore model generalization, algorithmic interpretability, lightweight system deployment, and multi-modal comfort fusion to promote the practical application of these research findings.

Author Contributions

T.W.: Writing—review & editing, Supervision. W.H.: Writing—original draft, Writing—review & editing, Methodology, Visualization, Conceptualization. H.Y.: Writing—review & editing, Resources, Supervision, Conceptualization, Project administration. J.G.: Writing—review & editing, Methodology, Visualization, Conceptualization. Y.L.: Writing—review & editing. Y.G.: Writing—review & editing. All authors have read and agreed to the published version of the manuscript.

Funding

This research was supported by the National Natural Science Foundation of China (Grant No. 52378095), the Key Research Project Support Program for Higher Education Institutions in Henan Province (Grant No. 24A410001 and 26A410002), the International Science and Technology Cooperation Program of Henan Province (Grant No. 252102521004), and the Natural Science Foundation of Henan (Grant No. 252300423404).

Institutional Review Board Statement

This study did not involve any direct medical or psychological interventions on the participants. According to Article 32 of the Measures for Ethical Review of Life Sciences and Medical Research Involving Humans deliberated and adopted by the National Science and Technology Ethics Committee of the People’s Republic of China, ethical review may be exempted for research involving humans that uses human information data or biological samples under the following circumstances, provided it does not cause harm to the human body and does not involve sensitive personal information or commercial interests (original text available at: https://www.nhc.gov.cn/qjjys/c100016/202302/6b6e447b3edc4338856c9a652a85f44b.shtml, URL accessed on 10 February 2026). This study falls under such circumstances; therefore, formal ethical review approval was not required.

Informed Consent Statement

Informed consent has been obtained from all participants to publish the details and/or images contained in this manuscript. All data in the manuscript and in any public datasets have been anonymized to minimize the risk of identification, and the participants are aware that the article will be published in an open-access journal under a Creative Commons CC BY 4.0 license. (1) Consent for Minors: Minors were not included as participants in this study. (2) Data Anonymization: All content, relevant data, and images presented in the manuscript have been anonymized to the greatest extent possible. (3) Information Provided to Participants: Participants were fully informed about the purpose of this study, the potential risks and benefits of publication, and the consequences of disclosing their personal information. (4) Voluntary Participation: Participants were informed that their consent and participation in the publication of this research are entirely voluntary. They were made aware of their right to withdraw consent at any time. (5) Compliance with Local Laws: This study complies with local laws and regulations regarding informed consent and privacy protection. (6) The entire experimental process complied with the principles of the Declaration of Helsinki.

Data Availability Statement

Data associated with our study has been deposited into a publicly available repository. The dataset from this study has been stored in the Open Science Framework (OSF, https://osf.io) repository, with the access number https://doi.org/10.17605/OSF.IO/4B72C, (URL accessed on 9 February 2026).

Acknowledgments

The first draft of this study was entirely prepared manually by the author. Upon completion of the draft, Gemini 3.0 was used for partial language polishing and information proofreading. The authors have reviewed and edited the output and take fully responsible for the content of this publication

Conflicts of Interest

The authors declare no conflicts of interest.

Appendix A

To facilitate reproducibility and comparison, we evaluate classification performance using Accuracy, Precision, Recall, f1-score, and ROC–AUC. Suppose there are

K

classes and the total number of test samples is

N

. Let the confusion matrix be

C \in R^{K \times K}

, where

C_{i j}

denotes the number of samples whose true class is

i

but are predicted as class

j

. For any class

i

, we define the number of true positives (

T P_{i}

), false positives (

F P_{i}

), and false negatives (

F N_{i}

) as follows:

T P_{i} = C_{i i}, F P_{i} = \sum_{j \neq i} C_{j i}, F N_{i} = \sum_{j \neq i} C_{i j} .

(A1)

(1) Accuracy

The overall Accuracy represents the proportion of correctly predicted samples across all classes:

A c c u r a c y = \frac{\sum_{i = 1}^{K} C_{i i}}{N} .

(A2)

(2) Precision and Recall (Class-wise)

For each individual class

i

, Precision_i and Recall_i are computed as:

{P r e c i s i o n}_{i} = \frac{T P_{i}}{T P_{i} + F P_{i}}, {R e c a l l}_{i} = \frac{T P_{i}}{T P_{i} + F N_{i}} .

(A3)

(3) f1 score (Class-wise)

The f1_i score is the harmonic mean of Precision_i and Recall_i:

{f 1}_{i} = \frac{2 \times {P r e c i s i o n}_{i} \times {R e c a l l}_{i}}{{P r e c i s i o n}_{i} + {R e c a l l}_{i}} .

(A4)

(4) Macro and Weighted Aggregation

To summarize performance under class imbalance, we employ two averaging strategies.

Macro-averaging assigns equal weight to each class:

{P r e c i s i o n}_{m a c r o} = \frac{1}{K} \sum_{i = 1}^{K} {P r e c i s i o n}_{i},

(A5)

{R e c a l l}_{m a c r o} = \frac{1}{K} \sum_{i = 1}^{K} {R e c a l l}_{i},

(A6)

{f 1}_{m a c r o} = \frac{1}{K} \sum_{i = 1}^{K} {f 1}_{i} .

(A7)

Weighted-averaging weights each class by its support

n_{i}

, where

n_{i} = \sum_{j = 1}^{K} C_{i j}

:

{P r e c i s i o n}_{w e i g h t e d} = \sum_{i = 1}^{K} \frac{n_{i}}{N} {P r e c i s i o n}_{i},

(A8)

{R e c a l l}_{w e i g h t e d} = \sum_{i = 1}^{K} \frac{n_{i}}{N} {R e c a l l}_{i},

(A9)

{f 1}_{w e i g h t e d} = \sum_{i = 1}^{K} \frac{n_{i}}{N} {f 1}_{i}

(A10)

Given the class imbalance, we emphasize

{f 1}_{m a c r o}

in our comparisons to ensure that performance on minority classes is adequately reflected.

(5) ROC–AUC (One-vs-Rest)

For multi-class classification, we adopt a one-vs-rest

(O v R)

strategy. The True Positive Rate (

T P R_{i}

) and False Positive Rate (

F P R_{i}

) for class

i

at threshold

t

are defined as:

T P R_{i} (t) = \frac{T P_{i} (t)}{T P_{i} (t) + F N_{i} (t)}, F P R_{i} (t) = \frac{F P_{i} (t)}{F P_{i} (t) + T N_{i} (t)} .

(A11)

The Area Under the Curve for class

i

(

{A U C}_{i}

) is calculated as:

A U C_{i} = \int_{0}^{1} T P R_{i} (F P R_{i}) d (F P R_{i}) .

(A12)

The final reported

{A U C}_{m a c r o}

is the average over the set of classes

K^{'}

present in the test set, where

K^{'} = | K^{'} |

:

A U C_{m a c r o} = \frac{1}{K^{'}} \sum_{i \in K^{'}} A U C_{i}

(A13)

Appendix B

Additional model performance metrics for the selected best feature combination are listed in Table A1.

Table A1. Additional evaluation metrics for the optimal feature set on the test set.

Model	Precision (Macro)	Recall (Macro)	Precision (Weighted)	Recall (Weighted)
SVC	0.2563	0.2374	0.4262	0.5000
KNN	0.3721	0.3112	0.4924	0.4977
RF	0.5216	0.3590	0.5627	0.5724

Macro-averaging (equal weight for each class) and weighted-averaging (weighted by sample size per class).

To more comprehensively characterize the discriminative ability of the models in a multi-class task, we plotted ROC curves for SVC, kNN, and RF using a one-vs-rest (OvR) strategy under the selected best feature group and the same test-set split (see Figure A1), and reported the AUC for each class. It should be noted that, because class −3 has zero samples in the independent test set (see the class distribution table in the main text), the ROC/AUC for this class is not defined on the test set and was therefore automatically skipped in the OvR plotting. Overall, the three models show relatively strong discrimination for class 0 (AUC ≈ 0.92–0.98), whereas their discrimination for minority classes (e.g., −2) is weaker (AUC ≈ 0.49–0.63), which is consistent with the imbalanced class distribution and the limited sample size of minority classes. These ROC results are provided to supplement the class-wise discriminative differences among the models.

Figure A1. ROC Curves of 3 models. The three subfigures show the ROC curves of different models ((a). kNN; (b). RF; (c). SVC), used to compare the classification performance for each class; the closer a curve is to the upper-left corner, the better the model distinguishes that class, while the pink dashed line indicates the chance level for random classification.

Appendix C

The key hyperparameters of the three classifiers were specified as follows. For SVC, an RBF kernel was used (kernel = “rbf”), with probability = probability = False, and the random seed was fixed (random_state = 42); other SVC parameters were kept at their default values. For KNN, the number of neighbors was set to n_neighbors = 5; other parameters (e.g., distance metric and neighbor weighting) followed default settings. For RF, the number of trees was set to n_estimators = 300, with random_state = 42 and n_jobs = −1; remaining parameters (e.g., max_depth, max_features, and min_samples_split) were left as defaults. In addition, input features were standardized for SVC and KNN using z-score normalization, while RF was trained without requiring feature scaling. Unless otherwise stated, all unspecified hyperparameters followed the default configuration of the adopted implementation.

Appendix D

The PMV-calculated thermal sensation indices are reported in Table A2.

Table A2. Thermal sensation voting scale.

Very Hot	Hot	Warm	Netural	Cool	Cold	Very Cold
+3	+2	+1	0	−1	−2	−3

References

Global Alliance for Buildings and Construction. Global Status Report for Buildings and Construction; International Energy Agency, UN Environment: Paris, France, 2020; p. 19. [Google Scholar]
Huo, T.; Ren, H.; Zhang, X.; Cai, W.; Feng, W.; Zhou, N.; Wang, X. China’s energy consumption in the building sector: A Statistical Yearbook-Energy Balance Sheet based splitting method. J. Clean. Prod. 2018, 185, 665–679. [Google Scholar] [CrossRef]
Yang, Z.; Dong, X.; Xiao, H.; Sun, H.; Wang, B.; Shi, W.; Li, X. Investigation of thermal comfort of room air conditioner during heating season. Build. Environ. 2022, 207, 108544. [Google Scholar] [CrossRef]
Su, C.; Madani, H.; Palm, B. Heating solutions for residential buildings in China: Current status and future outlook. Energy Convers. Manag. 2018, 177, 493–510. [Google Scholar] [CrossRef]
Ma, S.; Guo, S.; Zheng, D.; Chang, S.; Zhang, X. Roadmap towards clean and low carbon heating to 2035: A provincial analysis in northern China. Energy 2021, 225, 120164. [Google Scholar] [CrossRef]
Fanger, P.O. Thermal Comfort Analysis and Application in Environmental Engineering; Danish Technical Press: Copenhagen, Demark, 1970. [Google Scholar]
de Dear, R.J.; Brager, G.S. Thermal comfort in naturally ventilated buildings: Revisions to ASHRAE standard 55. Energy Build. 2002, 34, 549–561. [Google Scholar] [CrossRef]
Broday, E.E.; Moreto, J.A.; de Paula Xavier, A.A.; de Oliveira, R. The approximation between thermal sensation votes (OTS) and predicted mean vote (PMV): A comparative analysis. Int. J. Ind. Ergon. 2019, 69, 1–8. [Google Scholar] [CrossRef]
van Hoof, J. Forty years of Fanger’s model of thermal comfort: Comfort for all? Indoor Air 2008, 18, 182–201. [Google Scholar] [CrossRef]
Auffenberg, F.; Stein, S.; Rogers, A. A personalised thermal comfort model using a Bayesian Network. In Proceedings of the Twenty-Fourth International Joint Conference on Artificial Intelligence, Buenos Aires, Argentina, 25–31 July 2015; pp. 2547–2553. [Google Scholar]
Nicol, F. Adaptive thermal comfort standards in the hot–humid tropics. Energy Build. 2004, 36, 628–637. [Google Scholar] [CrossRef]
Bouden, C.; Ghrab, N. An adaptive thermal comfort model for the Tunisian context: A field study results. Energy Build. 2005, 37, 952–963. [Google Scholar] [CrossRef]
Kim, J.T.; Lim, J.H.; Cho, S.H.; Yun, G.Y. Development of the adaptive PMV model for improving prediction performances. Energy Build. 2015, 98, 100–105. [Google Scholar] [CrossRef]
Dahlan, N.D.; Gital, Y.Y. Thermal sensations and comfort investigations in transient conditions in tropical office. Appl. Ergon. 2016, 54, 169–176. [Google Scholar] [CrossRef]
de Dear, R.J. Global database of thermal comfort field experiments. ASHRAE Trans. 1998, 104, 1141. [Google Scholar]
de Dear, R.J.; Brager, G.S.; Cooper, D. Developing an adaptive model of thermal comfort and preference—Final report on RP-884. ASHRAE Trans. 1997, 104, 291. [Google Scholar]
Yang, B.; Li, X.; Liu, Y.; Chen, L.; Guo, R.; Wang, F.; Yan, K. Comparison of models for predicting winter individual thermal comfort based on machine learning algorithms. Build. Environ. 2022, 215, 108970. [Google Scholar] [CrossRef]
Schweiker, M.; Schakib-Ekbatan, K.; Fuchs, X.; Becker, S. A seasonal approach to alliesthesia. Is there a conflict with thermal adaptation? Energy Build. 2020, 212, 109745. [Google Scholar] [CrossRef]
Wang, Z.; Wang, J.; He, Y.; Liu, Y.; Lin, B.; Hong, T. Dimension analysis of subjective thermal comfort metrics based on ASHRAE Global Thermal Comfort Database using machine learning. J. Build. Eng. 2020, 29, 101120. [Google Scholar] [CrossRef]
Zhao, Q.; Lyu, J.; Du, H.; Lian, Z.; Zhao, Z. Gender differences in thermal sensation and skin temperature sensitivity under local cooling. J. Therm. Biol. 2023, 111, 103401. [Google Scholar] [CrossRef]
Luo, M.; Wang, Z.; Zhang, H.; Arens, E.; Filingeri, D.; Jin, L.; Ghahramani, A.; Chen, W.; He, Y.; Si, B. High-density thermal sensitivity maps of the human body. Build. Environ. 2020, 167, 106435. [Google Scholar] [CrossRef]
ISO International Standard 7726-2001; Ergonomics of the Thermal Environment Instruments for Measuring Physical Quantities. International Organization for Standardization: Geneva, Switzerland, 2001; pp. 2–49.
ASHRAE Standard 55-2020; Thermal Environmental Conditions for Human Occupancy. American Society of Heating, Refrigerating and Air Conditioning Engineers: Atlanta, GA, USA, 2020; pp. 5–21.
ISO International Standard 7730-2005; Moderate Thermal Environments—Determination of the PMV and PPD Indices and Specification of the Conditions for Thermal Comfort. International Organization for Standardization: Geneva, Switzerland, 2005; pp. 2–26.
GB/T 50785-2012; Evaluation Standard for Indoor Thermal Environment in Civil Buildings. China Architecture & Building Press: Beijing, China, 2012.
Mccullough, E.A.; Olesen, B.W.; Hong, S. Thermal insulation provided by chairs. ASHRAE Trans. 1994, 100, 795–802. [Google Scholar]
Guo, G.; Wang, H.; Bell, D.; Bi, Y.; Greer, K. KNN model-based approach in classification. In On the Move to Meaningful Internet Systems 2003: CoopIS, DOA, and ODBASE; Meersman, R., Tari, Z., Schmidt, D.C., Eds.; Springer: Berlin/Heidelberg, Germany, 2003; pp. 986–996. [Google Scholar] [CrossRef]
Jiang, L.; Zhang, H.; Cai, Z. A novel bayes model: Hidden naive bayes. IEEE Trans. Knowl. Data Eng. 2008, 21, 1361–1371. [Google Scholar] [CrossRef]
Kingsford, C.; Salzberg, S.L. What are decision trees? Nat. Biotechnol. 2008, 26, 1011–1013. [Google Scholar] [CrossRef] [PubMed]
Noble, W.S. What is a support vector machine? Nat. Biotechnol. 2006, 24, 1565–1567. [Google Scholar] [CrossRef] [PubMed]
Mor, B.; Garhwal, S.; Kumar, A. A Systematic Review of Hidden Markov Models and Their Applications. Arch. Comput. Methods Eng. 2021, 28, 1429–1448. [Google Scholar] [CrossRef]
Liu, T.; Jin, L.; Zhong, C.; Xue, F. Study of thermal sensation prediction model based on support vector classification (SVC) algorithm with data preprocessing. J. Build. Eng. 2022, 48, 103919. [Google Scholar] [CrossRef]
Xiong, L.; Yao, Y. Study on an adaptive thermal comfort model with K-nearest-neighbors (KNN) algorithm. Build. Environ. 2021, 202, 108026. [Google Scholar] [CrossRef]
Ren, J.; Zhang, R.; Cao, X.; Kong, X. Experimental study on the physiological parameters of occupants under different temperatures and prediction of their thermal comfort using machine learning algorithms. J. Build. Eng. 2024, 84, 108676. [Google Scholar] [CrossRef]
Breiman, L. Random Forests. Mach. Learn. 2001, 45, 5–32. [Google Scholar] [CrossRef]
Svetnik, V.; Liaw, A.; Tong, C.; Culberson, J.C.; Sheridan, R.P.; Feuston, B.P. Random Forest: A Classification and Regression Tool for Compound Classification and QSAR Modeling. J. Chem. Inf. Comput. Sci. 2003, 43, 1947–1958. [Google Scholar] [CrossRef]
Hastie, T.; Tibshirani, R.; Friedman, J. The Elements of Statistical Learning; Springer: Berlin/Heidelberg, Germany, 2009. [Google Scholar] [CrossRef]
Ghahramani, A.; Tang, C.; Becerik-Gerber, B. An online learning approach for quantifying personalized thermal comfort via adaptive stochastic modeling. Build. Environ. 2015, 92, 86–96. [Google Scholar] [CrossRef]
Daum, D.; Haldi, F.; Morel, N. A personalized measure of thermal comfort for building controls. Build. Environ. 2011, 46, 3–11. [Google Scholar] [CrossRef]
Rana, R.; Kusy, B.; Jurdak, R.; Wall, J.; Hu, W. Feasibility analysis of using humidex as an indoor thermal comfort predictor. Energy Build. 2013, 64, 17–25. [Google Scholar] [CrossRef]
Fountain, M.; Brager, G.; de Dear, R. Expectations of indoor climate control. Energy Build. 1996, 24, 179–182. [Google Scholar] [CrossRef]
Humphreys, M.A.; Nicol, J.F. The validity of ISO-PMV for predicting comfort votes in every-day thermal environments. Energy Build. 2002, 34, 667–684. [Google Scholar] [CrossRef]
Takada, S.; Matsumoto, S.; Matsushita, T. Prediction of whole-body thermal sensation in the non-steady state based on skin temperature. Build. Environ. 2013, 68, 123–133. [Google Scholar] [CrossRef]
González-Alonso, J. Human thermoregulation and the cardiovascular system. Exp. Physiol. 2012, 97, 340–346. [Google Scholar] [CrossRef]
Havenith, G.; Holmér, I.; Parsons, K. Personal factors in thermal comfort assessment: Clothing properties and metabolic heat production. Energy Build. 2002, 34, 581–591. [Google Scholar] [CrossRef]
Ji, W.; Luo, M.; Cao, B.; Zhu, Y.; Geng, Y.; Lin, B. A new method to study human metabolic rate changes and thermal comfort in physical exercise by CO2 measurement in an airtight chamber. Energy Build. 2018, 177, 402–412. [Google Scholar] [CrossRef]
Chaudhuri, T.; Soh, Y.C.; Li, H.; Xie, L. Machine learning driven personal comfort prediction by wearable sensing of pulse rate and skin temperature. Build. Environ. 2020, 170, 106615. [Google Scholar] [CrossRef]
Li, W.; Zhang, J.; Zhao, T.; Liang, R. Experimental research of online monitoring and evaluation method of human thermal sensation in different active states based on wristband device. Energy Build. 2018, 173, 613–622. [Google Scholar] [CrossRef]

Figure 1. Research procedure.

Figure 2. Experimental flow chart.

Figure 3. Measurement setup. (a) Schematic diagram of skin measurement points. (b) Measurement height of environmental parameters.

Figure 4. Working principle diagram of intelligent heating.

Table 1. Basic information of occupants.

	Age	Height/cm	Weight/kg	BMI
Minimum	19	153.00	45.30	16.40
Maximum	22	185.00	81.30	25.50
Mean	20.50	169.00	60.20	21.10
Standard Deviation	0.76	8.02	10.94	2.62

Table 2. Field test parameters and measuring instruments.

Instrument	Model	Measurement Content	Range	Accuracy
Indoor Wireless Temperature Logger (Beijing Century Jiantong Technology, Beijing, China)	JTR08ZI	Indoor air temperature	−40~85 °C	±0.5 °C
Outdoor Wireless Temperature Logger (Beijing Century Jiantong Technology, Beijing, China)	JTR08ZO	Outdoor air temperature	−40~85 °C	±0.5 °C
Automatic Blood Pressure Monitor (OMRON Healthcare, Kyoto, Japan)	HEM-7124	Heart rate	40~180 bpm	±5%
	HEM-7124	Blood pressure	0–299 mmHg	/
iButton Temperature Logger (Maxim Integrated, San Jose, CA, USA)	DS1922L	Skin temperature	−40~85 °C	±0.5 °C

Table 3. Thermal comfort voting scale.

Very Comfortable	Comfortable	Slightly Comfortable	Slightly Uncomfortable	Uncomfortable	Very Uncomfortable
+3	+2	+1	−1	−2	−3

Table 4. Classification of supervised learning methods.

Method	Applicable Problems	Model Characteristics	Model Category
K-Nearest Neighbors [27]	Multi-class classification	Feature space	Discriminative model
Naive Bayes [28]	Multi-class classification	Joint probability distribution of features and categories	Generative model
Decision Tree [29]	Multi-class classification	Classification tree	Discriminative model
Support Vector Machine [30]	Binary and multi-class classification	Separating hyperplane	Discriminative model
Hidden Markov Model [31]	Standard	Joint probability distribution model of observation sequence and state sequence	Generative model

Table 5. Input feature parameter combinations.

Group	Parameter Combination	Group	Parameter Combination	Group	Parameter Combination
Group 1	Ta; T_skforehead	Group 22	Ta; T_skforehead; T_skcheek; T_skBH	Group 43	Ta; T_skforehead; T_skcheek; T_skBH; SBP
Group 2	Ta; T_skcheek	Group 23	Ta; T_skforehead; T_skcheek; HR	Group 44	Ta; T_skforehead; T_skcheek; T_skBH; clo
Group 3	Ta; T_skBH	Group 24	Ta; T_skforehead; T_skcheek; SBP	Group 45	Ta; T_skforehead; T_skcheek; HR; SBP
Group 4	Ta; HR	Group 25	Ta; T_skforehead; T_skcheek; clo	Group 46	Ta; T_skforehead; T_skcheek; HR; clo
Group 5	Ta; SBP	Group 26	Ta; T_skforehead; T_skBH; HR	Group 47	Ta; T_skforehead; T_skcheek; SBP; clo
Group 6	Ta; clo	Group 27	Ta; T_skforehead; T_skBH; SBP	Group 48	Ta; T_skforehead; T_skBH; HR; SBP
Group 7	Ta; T_skforehead; T_skcheek	Group 28	Ta; T_skforehead; T_skBH; clo	Group 49	Ta; T_skforehead; T_skBH; HR; clo
Group 8	Ta; T_skforehead; T_skBH	Group 29	Ta; T_skforehead; HR; SBP	Group 50	Ta; T_skforehead; T_skBH; SBP; clo
Group 9	Ta; T_skforehead; HR	Group 30	Ta; T_skforehead; HR; clo	Group 51	Ta; T_skforehead; HR; SBP; clo
Group 10	Ta; T_skforehead; SBP	Group 31	Ta; T_skforehead; SBP; clo	Group 52	Ta; T_skcheek; T_skBH; HR; SBP
Group 11	Ta; T_skforehead; clo	Group 32	Ta; T_skcheek; T_skBH; HR	Group 53	Ta; T_skcheek; T_skBH; HR; clo
Group 12	Ta; T_skcheek; T_skBH	Group 33	Ta; T_skcheek; T_skBH; SBP	Group 54	Ta; T_skcheek; T_skBH; SBP; clo
Group 13	Ta; T_skcheek; HR	Group 34	Ta; T_skcheek; T_skBH; clo	Group 55	Ta; T_skcheek; HR; SBP; clo
Group 14	Ta; T_skcheek; SBP	Group 35	Ta; T_skcheek; HR; SBP	Group 56	Ta; T_skBH; HR; SBP; clo
Group 15	Ta; T_skcheek; clo	Group 36	Ta; T_skcheek; HR; clo	Group 57	Ta; T_skforehead; T_skcheek; T_skBH; HR; SBP
Group 16	Ta; T_skBH; HR	Group 37	Ta; T_skcheek; SBP; clo	Group 58	Ta; T_skforehead; T_skcheek; T_skBH; HR; clo
Group 17	Ta; T_skBH; SBP	Group 38	Ta; T_skBH; HR; SBP	Group 59	Ta; T_skforehead; T_skcheek; T_skBH; SBP; clo
Group 18	Ta; T_skBH; clo	Group 39	Ta; T_skBH; HR; clo	Group 60	Ta; T_skforehead; T_skcheek; HR; SBP; clo
Group 19	Ta; HR; SBP	Group 40	Ta; T_skBH; SBP; clo	Group 61	Ta; T_skforehead; T_skBH; HR; SBP; clo
Group 20	Ta; HR; clo	Group 41	Ta; HR; SBP; clo	Group 62	Ta; T_skcheek; T_skBH; HR; SBP; clo
Group 21	Ta; SBP; clo	Group 42	Ta; T_skforehead; T_skcheek; T_skBH; HR	Group 63	Ta; T_skforehead; T_skcheek; T_skBH; HR; SBP; clo

Table 6. Pseudo-code of multi-factor and multi-type coupling thermal comfort prediction algorithm.

Thermal Comfort Prediction

//1. Data Preprocessing
D_clean ← handle_outliers(D)
//2. One-time Split of Training and Test Sets (Ensuring All Models Use the Same Data)
D_train, D_test ← train_test_split(D_clean, test_size = 0.2, random_state = 42)
//3. Initialize Model List and Results Storage
models ← [SVC(), KNN(), RandomForest()]
results ← empty dictionary
//(NEW) 3.1 Initialize Cross-Validation Setting (on training set only)
k ← 5
cv_splits ← stratified_k_fold_split(D_train, k, shuffle = True, random_state = 42)
//4. Train, Predict, and Evaluate for Each Model
for each model in models do:
//(NEW) Cross-Validation Phase on Training Set
cv_acc_list ← empty list
cv_f1_list ← empty list
cv_auc_list ← empty list
for each (D_tr, D_val) in cv_splits do:
M_cv ← model.fit(D_tr)
y_val_pred ← M_cv.predict(D_val)
y_val_true ← D_val.labels
cv_acc_list.append(calculate_accuracy(y_val_true, y_val_pred))
cv_f1_list.append(calculate_f1_score(y_val_true, y_val_pred))
cv_auc_list.append(calculate_auc(y_val_true, y_val_pred))
end for
cv_accuracy_mean ← mean(cv_acc_list)
cv_accuracy_std ← std(cv_acc_list)
cv_f1_mean ← mean(cv_f1_list)
cv_f1_std ← std(cv_f1_list)
cv_auc_mean ← mean(cv_auc_list)
cv_auc_std ← std(cv_auc_list)
//Training Phase (train on full training set, as original)
M_trained ← model.fit(D_train)
//Prediction Phase (test set, as original)
y_pred ← M_trained.predict(D_test)
y_true ← D_test.labels //Obtain True Labels of the Test Set
//Evaluation Phase (test set, as original)
accuracy ← calculate_accuracy(y_true, y_pred)
f1_score ← calculate_f1_score(y_true, y_pred)
auc ← calculate_auc(y_true, y_pred)
//Store Results (NEW: add CV mean ± std while keeping original outputs)
results[model.name] ← {
‘CV_Accuracy_mean’: cv_accuracy_mean, ‘CV_Accuracy_std’: cv_accuracy_std,
‘CV_f1_mean’: cv_f1_mean, ‘CV_f1_std’: cv_f1_std,
‘CV_Auc_mean’: cv_auc_mean, ‘CV_Auc_std’: cv_auc_std,
‘Accuracy’: accuracy, ‘f1_score’: f1_score, ‘Auc’: auc
}
end for
//5. Output Results for All Models
return results

Table 7. Distribution of thermal comfort classes in the overall dataset and in the stratified 80/20 train–test split.

Class	Overall n (%)	Train n (%)	Test n (%)
−3	2 (0.09%)	2 (0.11%)	0 (0.00%)
−2	15 (0.68%)	12 (0.68%)	3 (0.68%)
−1	234 (10.59%)	187 (10.58%)	47 (10.63%)
0	100 (4.53%)	80 (4.53%)	20 (4.52%)
1	847 (38.34%)	678 (38.37%)	169 (38.24%)
2	953 (43.14%)	762 (43.12%)	191 (43.21%)
3	58 (2.63%)	46 (2.60%)	12 (2.71%)

Table 8. Evaluation criteria for classification support vector machine thermal comfort model.

Group	Accuracy (%)	f1-Score (%)	AUC (%)	Group	Accuracy (%)	f1-Score (%)	AUC (%)
Test 1	58.296	57.419	62.189	Test 33	55.900	53.631	57.597
Test 2	57.128	55.645	63.507	Test 34	57.623	57.226	61.837
Test 3	58.072	57.560	61.164	Test 35	58.421	58.397	62.253
Test 4	57.761	57.705	62.736	Test 36	58.656	58.642	62.716
Test 5	53.363	50.387	49.550	Test 37	50.673	49.437	54.624
Test 6	58.296	58.080	59.232	Test 38	58.647	58.635	63.520
Test 7	59.949	57.828	67.058	Test 39	60.901	60.897	63.988
Test 8	60.762	59.911	65.747	Test 40	54.402	54.047	52.967
Test 9	57.311	57.255	62.906	Test 41	60.439	60.361	62.560
Test 10	52.691	50.468	54.243	Test 42	61.124	61.124	64.042
Test 11	58.969	58.959	59.986	Test 43	54.770	52.616	57.130
Test 12	61.659	61.067	66.567	Test 44	58.744	58.443	62.097
Test 13	57.761	57.705	62.594	Test 45	59.093	59.041	62.432
Test 14	51.794	49.503	54.617	Test 46	58.432	58.406	62.857
Test 15	60.538	60.424	58.995	Test 47	50.673	48.980	54.356
Test 16	60.901	60.897	63.934	Test 48	58.871	58.847	63.501
Test 17	56.135	54.842	55.628	Test 49	60.900	60.899	64.079
Test 18	58.186	57.971	61.190	Test 50	55.072	54.837	55.684
Test 19	58.199	58.192	62.409	Test 51	60.214	60.127	62.605
Test 20	58.432	58.406	62.827	Test 52	58.870	58.836	63.255
Test 21	50.673	49.663	54.465	Test 53	60.227	60.222	63.844
Test 22	63.453	63.006	69.342	Test 54	53.956	53.435	52.941
Test 23	57.985	57.936	62.809	Test 55	60.214	60.127	62.554
Test 24	51.269	50.206	54.319	Test 56	59.092	59.026	63.374
Test 25	58.520	58.515	60.249	Test 57	60.093	60.041	63.297
Test 26	61.126	61.121	64.124	Test 58	60.901	60.897	64.007
Test 27	56.126	53.761	57.894	Test 59	55.076	54.708	51.244
Test 28	58.969	58.684	61.837	Test 60	59.989	59.893	62.572
Test 29	57.972	57.947	62.426	Test 61	60.437	60.324	63.455
Test 30	58.208	58.177	62.887	Test 62	60.212	60.089	63.314
Test 31	50.673	49.434	51.505	Test 63	59.538	59.423	63.379
Test 32	60.224	60.224	63.819

Table 9. Evaluation standard of k-nearest Neighbor algorithm thermal comfort model.

Group	Accuracy (%)	f1-Score (%)	AUC (%)	Group	Accuracy (%)	f1-Score (%)	AUC (%)
Test 1	69.283	69.227	74.042	Test 33	73.055	72.803	80.672
Test 2	71.300	71.280	75.581	Test 34	73.767	73.446	82.048
Test 3	65.247	65.154	70.646	Test 35	75.482	75.217	82.120
Test 4	70.328	70.283	78.241	Test 36	71.300	70.905	79.212
Test 5	69.507	69.432	76.821	Test 37	71.300	70.992	78.042
Test 6	69.669	69.641	76.935	Test 38	75.523	75.362	80.848
Test 7	73.468	73.354	82.199	Test 39	70.135	69.817	76.715
Test 8	72.119	71.981	77.575	Test 40	69.229	69.138	74.534
Test 9	69.731	69.228	77.677	Test 41	72.565	72.372	78.182
Test 10	70.852	70.591	77.978	Test 42	74.888	74.326	84.024
Test 11	74.215	74.037	79.951	Test 43	75.561	75.155	85.608
Test 12	73.094	72.830	80.419	Test 44	78.181	78.006	85.713
Test 13	71.749	71.471	78.265	Test 45	75.249	74.742	83.834
Test 14	69.955	69.758	77.017	Test 46	75.785	75.397	84.220
Test 15	74.888	74.887	80.615	Test 47	74.630	74.333	82.702
Test 16	69.731	69.391	76.082	Test 48	75.336	74.941	83.120
Test 17	69.000	68.920	74.887	Test 49	74.664	74.272	79.996
Test 18	69.058	68.968	74.737	Test 50	73.991	73.479	80.427
Test 19	71.442	71.251	77.334	Test 51	74.440	74.087	81.038
Test 20	71.300	71.286	77.801	Test 52	76.906	76.495	84.946
Test 21	71.237	71.236	75.680	Test 53	71.749	71.228	80.291
Test 22	76.383	76.172	84.464	Test 54	73.060	72.675	81.377
Test 23	75.785	75.397	83.940	Test 55	75.480	75.167	82.678
Test 24	74.857	74.523	82.970	Test 56	75.975	75.778	81.400
Test 25	75.112	74.879	84.320	Test 57	76.906	76.436	85.786
Test 26	73.318	72.875	78.584	Test 58	74.888	74.254	83.868
Test 27	73.318	72.843	79.410	Test 59	76.682	76.360	85.819
Test 28	74.366	74.240	81.583	Test 60	75.336	74.941	84.324
Test 29	73.318	72.994	80.292	Test 61	75.336	74.969	83.123
Test 30	69.192	68.910	78.192	Test 62	76.906	76.466	85.413
Test 31	70.404	69.895	77.446	Test 63	76.682	76.191	86.032
Test 32	72.646	72.298	80.590

Table 10. Evaluation criteria of thermal comfort model based on random forest algorithm.

Group	Accuracy (%)	f1-Score (%)	AUC (%)	Group	Accuracy (%)	f1-Score (%)	AUC (%)
Test 1	85.619	85.618	92.916	Test 33	86.752	86.715	94.414
Test 2	83.184	83.184	90.599	Test 34	87.892	87.880	94.758
Test 3	79.998	79.998	88.453	Test 35	88.992	88.987	96.061
Test 4	83.366	83.361	91.650	Test 36	89.462	89.462	95.982
Test 5	86.771	86.771	93.081	Test 37	89.462	89.461	96.675
Test 6	86.547	86.534	93.874	Test 38	87.415	87.415	95.029
Test 7	86.996	86.993	93.860	Test 39	89.686	89.686	96.421
Test 8	83.857	83.836	92.045	Test 40	88.092	88.089	95.172
Test 9	89.238	89.237	95.324	Test 41	90.135	90.134	96.249
Test 10	87.444	87.444	95.178	Test 42	88.117	88.107	96.292
Test 11	89.238	89.238	96.023	Test 43	89.014	89.011	95.451
Test 12	83.184	83.174	92.128	Test 44	90.333	90.332	96.177
Test 13	86.771	86.771	93.252	Test 45	91.256	91.255	97.591
Test 14	88.341	88.339	95.636	Test 46	89.910	89.910	97.522
Test 15	87.444	87.435	94.811	Test 47	91.928	91.928	97.831
Test 16	85.426	85.424	91.774	Test 48	89.910	89.906	96.367
Test 17	87.411	87.408	92.909	Test 49	91.031	91.031	97.378
Test 18	84.081	84.067	93.098	Test 50	90.135	90.127	96.785
Test 19	87.668	87.661	94.990	Test 51	92.153	92.152	97.896
Test 20	87.668	87.667	95.068	Test 52	89.462	89.461	96.268
Test 21	88.341	88.329	95.124	Test 53	90.359	90.359	96.887
Test 22	84.529	84.520	93.705	Test 54	89.686	89.681	96.857
Test 23	90.807	90.806	96.233	Test 55	91.465	91.459	97.344
Test 24	88.565	88.564	96.369	Test 56	90.135	90.131	96.694
Test 25	89.238	89.238	95.846	Test 57	90.781	90.780	97.201
Test 26	88.117	88.115	94.615	Test 58	92.359	92.359	97.836
Test 27	87.198	87.182	94.200	Test 59	91.256	91.250	97.211
Test 28	87.640	87.640	94.742	Test 60	93.049	93.049	98.704
Test 29	90.807	90.805	96.473	Test 61	91.928	91.928	97.766
Test 30	91.031	91.030	96.835	Test 62	92.153	92.152	98.026
Test 31	90.807	90.805	96.605	Test 63	94.170	94.170	98.465
Test 32	87.668	87.663	94.070

Table 11. Stratified cross-validation (CV) and hold-out test performance.

Model	CV k	CV Accuracy (Mean ± Std)	CV Macro-f1 (Mean ± Std)	Test Accuracy	Test Macro-f1
RF	2	0.5388 ± 0.0004	0.3185 ± 0.0100	0.5724	0.3896
SVC	2	0.4929 ± 0.0092	0.1790 ± 0.0242	0.5000	0.2312
kNN	2	0.4499 ± 0.0116	0.2249 ± 0.0080	0.4977	0.3159

Table 12. Comparative analysis of thermal comfort prediction models.

Model	Parameter Combination	Accuracy
PMV	M, clo, Ta, t_r, V_a, RH	50.60%
SVC	Ta, T_skforehead, T_skcheek, T_skBH, HR, SBP	60.09%
K-NN	Ta, T_skforehead, T_skcheek, T_skBH, HR, SBP	76.91%
RF	Ta, T_skforehead, T_skcheek, T_skBH, HR, SBP	90.78%

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Article Metrics

Citations

Article Access Statistics

Journal Statistics

Multiple requests from the same IP address are counted as one view.