Design of an Edge-Cloud IoT System for Dynamic Thermal Sensation Control and Energy Optimization

Abstract

Improving HVAC energy efficiency while maintaining collective thermal comfort remains challenging in multi-occupant shared indoor environments, where occupants differ in thermal sensation, activity level, clothing condition, and spatial distribution. This study develops and field-validates an integrated edge-cloud IoT framework that connects non-invasive occupant-state sensing, INT8 edge thermal-sensation inference, and group-comfort-oriented HVAC setpoint optimization for classroom-based shared spaces. The proposed system integrates localized temperature–humidity sensing, vision-derived occupancy, posture, and clothing estimation, cloud-based thermal sensation model training, and edge-deployed real-time control on a HUB 8735 ULTRA device. A 4-day model-training data collection campaign with structured questionnaires was first conducted to obtain occupants’ Thermal Sensation Votes (TSVs) as ground-truth labels. The trained model was compressed from Float32 to INT8 through post-training quantization and deployed on the edge device for real-time inference. Predicted individual TSV values were then transformed into a PPD-inspired TSV-derived dissatisfaction index and used to determine the HVAC setpoint through rolling-horizon group comfort optimization. A separate eight-school-day single-blind daily-block A/B field validation was conducted, with four validation days assigned to the proposed smart control strategy and four days assigned to a fixed 25 °C baseline. The validation dataset included 2194 valid TSV questionnaire responses, which were aggregated into 116 valid 30 min classroom sessions for statistical comparison. The proposed control achieved a session-level mean TSV of −0.13, compared with −0.66 under the baseline, with Welch’s t(100) = 11.2, p < 0.001 and Cohen’s d = 2.11. Daily HVAC energy use decreased from 2.61 to 2.32 kWh/day, corresponding to a cumulative reduction of 1.16 kWh, or 11.1%, over the validation period. These results support the short-term feasibility of the proposed classroom-level human-centric HVAC control framework. However, because the validation was limited to a short-term classroom setting without full weather/load normalization, longer multi-season and multi-room studies are required to further evaluate generalizability and long-term energy performance.

1. Introduction

Heating, ventilation, and air conditioning (HVAC) systems account for approximately 40% to 50% of the total energy consumption in modern commercial and residential buildings. Therefore, improving HVAC energy efficiency while maintaining acceptable indoor environmental quality remains an important issue in building operation [1,2]. Traditional HVAC control commonly relies on fixed temperature setpoints. However, fixed setpoints are difficult to adapt to changing occupancy, heterogeneous thermal preferences, and transient indoor thermal variations in shared spaces. This limitation may lead to unnecessary cooling or heating energy use and may also reduce occupant thermal comfort. Recent advances in Internet of Things (IoT), cloud computing, and artificial intelligence (AI) provide opportunities to develop HVAC control systems that respond to both environmental conditions and occupant-related information [3,4,5].

Thermal comfort assessment is commonly based on standards such as ISO 7730 and ASHRAE Standard 55, which consider both environmental and personal parameters [6,7,8,9,10]. The main environmental variables include air temperature, relative humidity, air velocity, and mean radiant temperature, while the main personal variables include metabolic rate and clothing insulation. Fanger’s Predicted Mean Vote (PMV) and Predicted Percentage of Dissatisfied (PPD) indices have been widely used for thermal comfort evaluation [4,11,12,13]. However, the direct use of PMV in real-time control is difficult because personal variables such as metabolic rate and clothing insulation are often unavailable or difficult to measure continuously. Therefore, recent studies have increasingly adopted data-driven personal comfort models that combine sensor data and machine-learning methods to estimate occupants’ thermal sensation [5,14,15,16,17,18].

Although personal comfort models have improved individualized thermal comfort estimation, many practical indoor environments are multi-occupant shared thermal zones. In such spaces, occupants may have different activity levels, clothing conditions, and thermal preferences, making fixed setpoints or simple average-based control strategies insufficient for group comfort management [19]. This study focuses on classroom-based shared indoor environments, where multiple occupants stay in the same thermal zone during scheduled learning activities. The proposed framework may also be relevant to small offices, meeting rooms, and other shared spaces with similar occupancy patterns. However, the field validation in this study was conducted in a classroom setting; therefore, broader application to other building types requires further validation.

Existing IoT-based thermal comfort systems have commonly focused on environmental monitoring, occupancy detection, or cloud-based data visualization [20,21,22]. These systems provide useful sensing information, but they do not always close the loop between occupant-state estimation and physical HVAC actuation. Optimization-based and MPC-based HVAC studies have shown potential for reducing energy consumption, but many of them are implemented mainly in simulation environments, building-level supervisory platforms, or cloud-centered architectures [1,21,23,24,25]. Recent MPC-based HVAC studies have further shown that predictive control can reduce energy use while maintaining thermal comfort constraints, including EnergyPlus-based MPC, multi-objective MPC, and real-life data-driven MPC implementations. However, these approaches often require detailed building models, simulation environments, supervisory control platforms, or non-trivial optimization solvers, which may limit direct deployment in low-cost classroom-level applications [26,27,28,29]. Compared with these approaches, the present study emphasizes a field-deployed edge-cloud implementation that connects multimodal sensing, edge thermal sensation inference, and localized HVAC setpoint control in a classroom environment.

From the hardware and data-pipeline perspective, edge-cloud hybrid architectures have increasingly been adopted in IoT and building automation systems to reduce latency, network load, and dependence on continuous cloud connectivity [30,31,32,33,34]. Recent edge-computing studies have also highlighted the importance of task allocation and computation offloading for latency-sensitive and resource-constrained applications [35]. These findings support the use of distributed edge-cloud architectures for HVAC control systems that require both real-time local responsiveness and cloud-assisted model training. Edge deployment is important for HVAC control because real-time thermal sensation inference and setpoint adjustment require timely local processing and actuation. If all sensing data and control decisions are processed only through the cloud, the system may be affected by communication delay, intermittent connectivity, data transmission overhead, and privacy concerns related to occupant-state information. In the HVAC and thermal comfort domain, edge-based optimization has also been explored for localized thermal comfort control, showing that edge computing can support faster environmental decision-making while reducing dependence on centralized cloud processing [36]. In the proposed framework, the cloud is mainly used for model training, model optimization, and long-term data storage, whereas the HUB 8735 ULTRA edge device performs local thermal sensation inference, group discomfort calculation, and HVAC setpoint search [30,37,38,39,40].

Despite these advances, three practical research gaps remain. First, from the sensing perspective, many thermal comfort control systems still rely on fixed assumptions or active user feedback for occupant-related variables such as metabolic rate and clothing insulation. Such feedback may interrupt occupants’ normal activities and may be difficult to maintain during long-term classroom operation. Therefore, non-invasive sensing methods that combine environmental measurements with vision-derived occupant-state features require further field validation.

Second, from the modeling perspective, data-driven thermal comfort models are often trained and evaluated off-line, whereas their performance after compression and deployment on resource-constrained edge devices is less frequently reported. This limits the practical assessment of whether thermal sensation models can support real-time embedded HVAC control.

Third, from the control and deployment perspective, many optimization-based or MPC-based HVAC studies are implemented in simulation environments, building-level supervisory platforms, or cloud-centered architectures. Although these approaches have demonstrated energy-saving potential, practical classroom-level deployment still requires a localized framework that can connect sensing, edge inference, and physical HVAC actuation under real operating conditions.

To address these gaps, this study develops and field-validates a cloud-trained but edge-deployed HVAC control framework using the HUB 8735 ULTRA platform. The system integrates environmental sensing, non-invasive visual extraction of occupant-state features, INT8 thermal sensation inference, and localized rolling-horizon setpoint optimization. The main contributions of this study are summarized as follows:

• A classroom-oriented edge-cloud IoT framework is developed by integrating indoor temperature and humidity sensing with non-invasive visual extraction of occupancy, posture-related metabolic rate, and clothing-related insulation.
• A cloud-trained thermal sensation model is compressed from Float32 to INT8 using post-training quantization and deployed on the HUB 8735 ULTRA edge device to evaluate the feasibility of real-time embedded thermal sensation inference.
• An event-triggered rolling-horizon HVAC control strategy is implemented using a TSV-derived group discomfort index, a 10% comfort deadband, and a 30 min lockout interval. The strategy is field-validated against a fixed 25 °C baseline in an eight-school-day classroom A/B experiment.

The remainder of this paper is organized as follows. Section 2 describes the proposed hardware deployment, data-driven modeling method, and control strategy. Section 3 presents the system validation and experimental results. Section 4 discusses the findings, limitations, and future work.

2. Materials and Methods

To develop a dynamic thermal sensation control system based on edge and cloud computing, this study utilizes the HUB 8735 ULTRA development board (specifications detailed in

Table 1. Technical Specifications of the HUB 8735 Ultra Development Board.

Parameter	Specifications
Main SoC	Realtek Ameba RTL8735 (NCC certified)
AI Accelerator	Embedded Neural Processing Unit (NPU) for edge inference
Camera Module	Integrated 2-megapixel image sensor
Video Resolution	Up to 1080P Full HD video streaming
Audio Interface	Built-in microphone and audio mixer (with Buzzer PWM support)
Wireless Connectivity	Dual-band Wi-Fi (2.4 GHz/5 GHz, 802.11 a/b/g/n) and BLE
Storage Expansion	MicroSD card slot (supports up to 512 GB)
Peripheral Interfaces	GPIO, I2C, SPI, UART, PWM
Development Environment	Arduino IDE compatible
Power & Data Interface	Native USB Type-C

) as the core hardware to propose a multimodal smart thermal sensation control system leveraging an Edge-Cloud Hybrid Architecture. At the perception layer, this system not only integrates traditional environmental sensors (for real-time measurement of air temperature, relative humidity, air velocity, and mean radiant temperature) but also introduces Artificial Intelligence (AI) visual recognition technology to dynamically extract features such as occupancy, occupant behavior, and clothing insulation values. All multimodal sensing data first underwent real-time preliminary processing on the edge AI chip. Only environmental data and structured occupant-state variables were then transmitted to the cloud platform for long-term storage and model analysis, whereas raw image frames were processed locally and were not uploaded. To maintain a low-cost and easily deployable hardware configuration, the final deployed prototype continuously uploaded indoor air temperature, relative humidity, and vision-derived occupant-state variables. Before finalizing the sensing configuration, preliminary field checks were conducted using a TM-402 anemometer and a TM-188 black-globe thermometer. These checks indicated near-still-air conditions and no evident radiant heat source or strong radiant asymmetry in the tested classroom. Therefore, air velocity was treated as a fixed near-still-air boundary condition, and mean radiant temperature was not included as a continuous online input in the present low-cost deployment. The TM-402 and TM-188 instruments were retained as auxiliary reference instruments for preliminary checking and future high-precision system extension. The experimental framework of this study comprises four phases, with the architectural flow illustrated in Figure 1, detailed as follows:

(1)

Hardware System Implementation and Validation: The initial phase focuses on testing the accuracy and stability of the proposed system. By comparing the sensing data from the custom-built system with those from calibrated standard measurement instruments, the reliability and validity of the hardware measurements are established.

(2)

Data-Driven Modeling: A thermal sensation prediction model is trained using synchronized environmental, vision-derived occupant-state, and questionnaire-derived TSV data. The trained model is then quantized and deployed for edge inference.

(3)

Control Strategy Design: Based on the predicted individual TSV values, a TSV-derived nonlinear dissatisfaction index is computed and minimized through a deterministic rolling-horizon grid search to determine the HVAC setpoint.

(4)

Field Validation and Performance Evaluation: In the final phase, the constructed smart control system is deployed in a real-world environment and benchmarked against the traditional setpoint based control strategy. This phase comprehensively quantifies and compares the specific performance of the proposed system in enhancing overall thermal comfort and achieving HVAC energy savings.

(1) 

System Hardware Setup, Deployment, and Validation

The experimental site for this study was a multi-purpose classroom with a capacity of 20 occupants. To ensure that the sensory data precisely reflects the complex thermal distribution within the space and complies with international measurement standards, this section delineates the hardware architecture, spatial deployment strategy, and accuracy validation procedures.

A.1. 

Multimodal Perception Hardware Architecture

The system employs the HUB 8735 ULTRA development board, equipped with visual edge computing capabilities, as the core node to construct an edge-cloud perception network. The implemented low-cost environmental perception layer continuously collected indoor air temperature and relative humidity. Air velocity and mean radiant temperature were evaluated during preliminary field checking using auxiliary reference instruments. Because the tested classroom showed near-still-air conditions and no evident radiant asymmetry, these two parameters were not integrated into the continuous cloud-uploaded data stream in the present prototype. Meanwhile, the occupant-state perception layer utilized edge AI visual algorithms to non-invasively extract occupancy count, clothing category, and behavioral posture, which were subsequently mapped to $I_{clo}$ and $MET$ values. The detailed technical specifications of core sensors and auxiliary reference instruments used in the proposed system are listed in

Table 2. Core sensors and auxiliary reference instruments used in the proposed system.

Parameter	DHT22
	Range	Resolution	Accuracy
Temperature	−40~80 °C	0.1 °C	±0.5 °C
Humidity	0–100% RH	0.1% RH	±2% RH
Role in this study	Online sensing input; air temperature and relative humidity
	TM-188 WBGT
	Range	Resolution	Accuracy
Dry-Bulb Temperature ($T_a$)	0~50 °C	0.1 °C	±0.8 °C
Globe Temperature ($T_G$)	0~80 °C	0.1 °C	±0.6 °C
Relative Humidity (%RH)	1~99% RH	0.1% RH	±3.0% RH (20–80% RH)
Dew-Point Temperature (DEW)	−35.3~48.9 °C	0.1 °C	-
Wet-Bulb Temperature (WET)	−21.6–50.0 °C	0.1 °C	-
Role in this study	Auxiliary preliminary checking; radiant condition/MRT spot checking, not continuous online input
	TM-402 Anemometer
	Range	Resolution	Accuracy
Air Velocity	0.4–20 m/s	0.1 m/s	±(2% + 0.2) m/s
Airflow Calculation	0–9999 CMM	1 CMM	-
Temperature	−20–60 °C	0.1 °C	±1 °C
Role in this study	Auxiliary preliminary checking; air velocity spot checking, not continuous online input

, and the overall system architecture and physical system are illustrated in Figure 2 and Figure 3.

A.2 

Vision-based Occupant-State Estimation Pipeline

The vision-recognition module was based on the edge-deployed YOLOv4-tiny detection function available on the HUB 8735 ULTRA platform, combined with rule-based occupant-state mapping for extracting occupancy count, behavior/posture category, and clothing category. The computer-vision model was not trained from scratch in this study. The image frames were processed directly on the edge device, and the system output only structured occupant-state variables, including occupancy count, behavior/posture category, and clothing category. These variables were not interpreted as direct measurements of metabolic rate or clothing insulation. Instead, behavior/posture categories were mapped to metabolic equivalent ($MET$) values, while clothing categories were mapped to clothing insulation ($I_{clo}$) values according to standard thermal comfort references. The resulting $MET$ and $I_{clo}$ values were then integrated with environmental variables as inputs to the thermal sensation prediction model. The mapping information is listed in

Table 3. Mapping rules from vision-recognized occupant states to $MET$ and $I_{clo}$ values.

Vision-Recognized Variable	Recognized Class	Assigned Value	Reference Basis
Behavior/posture	Seated/writing/reading	1.0 MET	ASHRAE Standard 55
Behavior/posture	Standing	1.2 MET	ASHRAE Standard 55
Behavior/posture	Walking slowly	1.7 MET	ASHRAE Standard 55
Clothing category	Short-sleeved shirt	0.15 clo	ASHRAE 55/ISO 9920 [41]
Clothing category	Long-sleeved shirt	0.25 clo	ASHRAE 55/ISO 9920
Clothing category	Light jacket	0.22 clo	ASHRAE 55/ISO 9920

.

B. 

Spatial Deployment Strategy

In the present prototype, one HUB 8735 ULTRA-based integrated edge-sensing node was installed at the central representative seating zone of the 20-person classroom. The node was positioned away from direct HVAC supply airflow, window-side solar radiation, and heat-emitting electronic devices. The placement was selected to represent the main occupant zone rather than a local thermal disturbance region. The sensor height was fixed at 1.1 m above the floor, corresponding to the seated occupant head-level sampling height.

Camera Placement and Privacy Handling:

The camera module was installed to cover the main seating area of the classroom while avoiding unnecessary close-up facial capture. The viewing angle was selected to observe occupants’ upper-body posture and clothing category for thermal comfort estimation. To reduce privacy risks, raw images were processed locally on the edge device and were not uploaded to the cloud server. Only structured and anonymized variables, including occupancy count, posture/behavior category, clothing category, estimated $MET$, and estimated $I_{clo}$, were stored and used for thermal sensation modeling and HVAC control.

To clarify the physical deployment of the experimental sampling field, Figure 4 illustrates the classroom layout, including the seating area, HUB 8735 ULTRA edge-sensing node, camera viewing direction, HVAC outlet/indoor unit, windows, door, and the reference instrument position used during calibration.

C. 

Accuracy Validation and Vision-Based Recognition Stress Testing

To substantiate the measurement reliability of the custom-built IoT nodes, the following two validation procedures were executed prior to the formal experiments:

Physical sensor accuracy calibration: A 24 h synchronous comparative analysis was conducted between the proposed DHT22-based edge-sensing node and the calibrated TM-188 WBGT reference instrument for air temperature and relative humidity. During the calibration test, the proposed sensor node and the reference instrument were placed adjacent to each other at the representative seated-occupant height of 1.1 m to ensure comparable exposure conditions. A total of 59 synchronized samples were used for the validation analysis. Measurement errors were evaluated using Bias, Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and linear regression. In compliance with ISO 7726 [42] requirements for thermal comfort measurements, the temperature error was maintained within ±0.5 °C and relative humidity within ±5%, thereby ensuring data reliability. The TM-402 anemometer and TM-188 black-globe function were additionally used for preliminary spot checking of air velocity and radiant conditions, respectively, rather than as continuous online inputs.

Vision-based occupant-state extraction and validation: The HUB 8735 ULTRA edge board was used to capture classroom images and extract occupant-state information, including occupancy count, behavior/posture category, and clothing category. To validate the practical reliability of this procedure, the edge-generated recognition outputs were compared with real-time logs recorded by on-site observers during classroom deployment. The evaluation focused on the consistency between the edge-recognized outputs and observer-recorded ground truth for occupancy count, behavior/posture category, and clothing category. The validation covered normal classroom operation and practical boundary cases, including partial body occlusion caused by desks, chairs, or other occupants, overlapping occupants in the seating area, and ambiguous clothing layers. In the supplementary face-occlusion stress test, the face-region occlusion ratio was estimated as the proportion of the manually identified face region visually blocked in each selected challenging frame, and the empirical recognition-failure boundary was examined. For low-confidence or unstable recognition outputs, the system retained the previous valid observation or applied conservative default $MET$/$I_{clo}$ values to avoid abrupt control actions. The validation was conducted under normal classroom lighting conditions; controlled low-light, backlight, and non-uniform illumination tests were not included in the present study.

(2) 

Data-Driven Modeling

This section elucidates how the system constructs an individual thermal sensation prediction model through multimodal sensory data fusion and a “cloud training, edge inference” architecture. Furthermore, it proposes a “Minimize $D_{Group}^{TSV}$” control strategy tailored for multi-occupant shared scenarios. The overall algorithm implementation and optimal control workflow are divided into three core phases, as illustrated in Figure 5 and detailed below:

A. 

Multimodal Data Acquisition and Temporal Feature Alignment

In the final low-cost implementation, the model input vector consisted of indoor air temperature, relative humidity, vision-derived $MET$, and vision-derived $I_{clo}$. This system introduces Edge-AI Vision algorithms for non-invasive feature extraction, detailed as follows:

• Vision-based $MET$ Mapping: The edge-deployed detection module and rule-based occupant-state mapping were used to classify behavior/posture categories, which were then mapped to $MET$ values according to ASHRAE Standard 55.
• Object Detection-based $I_{clo}$ Mapping: Employs an object detection algorithm to identify occupants’ clothing styles, dynamically accumulating the thermal resistance coefficient based on visual recognition results (e.g., short-sleeved shirts $+0.15I_{clo}$).

B. 

Cloud-Based Machine Learning Model Training and Hyperparameter Setting

To construct a data-driven thermal sensation prediction model, a dedicated preliminary field-data collection phase was conducted before full system deployment. During this period, occupants’ subjective thermal sensation responses were recorded using structured questionnaires together with synchronized environmental and occupant-state data. The questionnaires were distributed at 30 min intervals to capture occupants’ immediate thermal perceptions while avoiding excessive response burden. The subjective responses were quantified using the ASHRAE seven-point thermal sensation vote (TSV) scale (+3: Hot, +2: Warm, +1: Slightly Warm, 0: Neutral, −1: Slightly Cool, −2: Cool, −3: Cold). The cloud server temporally synchronized the questionnaire-derived TSV labels with the corresponding feature vector, including indoor air temperature, relative humidity, vision-derived metabolic rate, and vision-derived clothing insulation.

The model-training phase was conducted on the Google Colaboratory cloud computing platform using the TensorFlow 2.x deep learning framework and cloud-based hardware accelerators. The data-driven thermal sensation model was developed using an independent four-day preliminary dataset collected before the A/B field validation. This dataset contained 894 synchronized samples and was used for model training and post-training quantization.

The thermal sensation prediction model was implemented in TensorFlow 2.x as a supervised feed-forward neural-network regression model. The input layer contained four features: indoor air temperature, relative humidity, vision-derived metabolic rate, and vision-derived clothing insulation. Before model training, the input features were standardized using the mean and standard deviation calculated from the training dataset. The network consisted of two hidden dense layers with 16 and 8 neurons, respectively, using ReLU activation functions. The output layer contained one neuron with a linear activation function for predicting the continuous TSV value. The model was trained using the mean squared error loss function with the Adam optimizer. The learning rate was set to the default Keras Adam learning rate of 0.001, the batch size was 16, and the model was trained for 100 epochs.

No separate internal validation/test split or occupant-disjoint split was performed because personal identifiers were not recorded. Therefore, the four-day preliminary dataset was used to learn the nonlinear mapping between indoor environmental conditions, vision-derived occupant-state features, and occupants’ thermal sensation responses. This limitation is explicitly acknowledged in the interpretation of the model results. Accordingly, the model-development phase aimed to learn the following data-driven nonlinear mapping under the near-still-air classroom condition:

$$TSV=f_{ML}\left(\left.T_a,RH,MET,I_{clo}\right|near-still-aircondition\right)$$

(1)

To satisfy the hardware and memory constraints of the localized environment, the cloud-trained thermal sensation prediction model was compiled into a lightweight TensorFlow Lite (.tflite) binary via a post-training optimization pipeline. Using the tf.lite.TFLiteConverter operator, a Post-Training Quantization (PTQ) was applied to convert the network weights from 32-bit floating-point (Float32) to 8-bit integers (INT8) representation. This edge-optimization process reduced the computational complexity and model footprint while maintaining acceptable prediction accuracy, allowing the embedded inference engine ($f_{TFLite}$) to operate directly on the HUB 8735 ULTRA development board.

After the Float32 model was trained and converted into the INT8 TensorFlow Lite model, the fixed INT8 model was deployed on the HUB 8735 ULTRA edge device. No additional retraining, model updating, or parameter tuning was performed during the subsequent four-block A/B field validation. Therefore, the field-validation data were independent of the model-development data and were used only to evaluate the real-time control performance of the proposed edge-cloud HVAC control framework.

C. 

Receding Horizon Group Thermal Comfort Optimization Strategy

In multi-occupant shared thermal zones, the primary objective is to minimize collective dissatisfaction while preventing micro-fluctuations in sensor readings. Let ${\tau }_s=5\min$ be the edge data sampling interval. At any given decision time step $t$ (triggered every $\mathsf{\Delta }t=30\min$), the edge controller aggregates the historical $N=6$ data packets from the past window to compute a moving average, smoothing out transient environmental noise:

$${\overline{\mathrm{X}}}_i\left(t\right)={\displaystyle \frac{1}{N}}{\displaystyle \sum _{m=0}^{N-1}}{\mathrm{X}}_i\left(t-m\cdot {\tau }_s\right)$$

(2)

where ${\mathrm{X}}_i=\left[T_a,RH,ME{T}_i,I_{{clo}_i}\right]$ represents the smoothed baseline parameters for occupant $i$.

By introducing a discrete virtual search space $\mathcal{T}=\left\{22.0°\mathrm{C},22.1°\mathrm{C},\dots ,28.0°\mathrm{C}\right\}$, the edge controller iteratively substitutes a candidate temperature $T_{test}\in \mathcal{T}$ as a variable into the embedded INT8 inference engine to predict individual thermal sensations:

$${\widehat{TSV}}_i\left(T_{test}\right)=f_{TFLite}\left(T_{test},{\overline{\mathrm{X}}}_i\left(t\right)\right)$$

(3)

The predicted TSV was not treated as an analytical PMV value. Instead, because PMV and TSV are expressed on a comparable seven-point thermal sensation scale, the machine-learning-predicted TSV was transformed through a PPD-inspired nonlinear penalty function to derive a TSV-based dissatisfaction penalty [9,10,43,44,45]. This penalty was used only as a control-oriented discomfort index, rather than as a formal ISO 7730 PMV–PPD calculation.

$${\widehat{D}}_i^{TSV}\left(T_{test}\right)=100-95\times \exp \left(-0.03353\cdot {\widehat{TSV}}_i^4\left(T_{test}\right)-0.2179\cdot {\widehat{TSV}}_i^2\left(T_{test}\right)\right)$$

(4)

The optimal setpoint $T_{opt}$ that yields the global minimum of the group dissatisfaction objective function $J\left(T_{test}\right)$ is determined as

$$J\left(T_{test}\right)=D_{Group}^{TSV}\left(T_{test}\right)={\displaystyle \frac{1}{k}}{\displaystyle \sum _{i=1}^k}{\widehat{D}}_i^{TSV}\left(T_{test}\right)$$

(5)

$$T_{opt}=\arg \underset{T_{\mathit{test}}\in \mathcal{T}}{\min }J\left(T_{test}\right)$$

(6)

where $k$ represents the total number of occupants detected within the shared indoor environment.

(3) 

Control Strategy Design

To mitigate thermal preference conflicts among multiple occupants while preventing redundant start-stop cycles and excessive hunting of the HVAC compressor, an Event-Triggered Rolling-Horizon Control (ET-RHC) mechanism with a built-in thermal comfort deadband was implemented. The deadband threshold was set to ${\theta }_D=10\%$ by referring to the commonly adopted thermal comfort acceptability criterion in ISO 7730 and ASHRAE Standard 55, where the PMV range of approximately −0.5 to +0.5 corresponds to PPD below 10%. Since the present study used a TSV-derived dissatisfaction penalty rather than a formal analytical PMV–PPD calculation, this 10% threshold was adopted as a standard-informed reference for defining the comfort deadband rather than as an optimized parameter. In addition, a 30 min lockout interval was applied between consecutive setpoint revisions. This interval was empirically configured based on the 30 min questionnaire sampling interval, the thermal inertia of the classroom, and practical compressor-protection considerations. Therefore, the 10% deadband and 30 min lockout interval should be interpreted as practical control settings rather than globally optimized control parameters. The adopted control parameters are summarized in

Table 4. Control parameter settings and rationale.

Parameter	Adopted Value	Basis/Rationale	Interpretation
Comfort deadband threshold ${\mathit{\theta }}_{\mathit{D}}$	10%	Standard-informed reference based on ISO 7730 and ASHRAE Standard 55, where PMV values of approximately −0.5 to +0.5 correspond to PPD below 10%.	Used as a comfort deadband reference rather than an optimized parameter.
Lockout interval	30 min	Empirically configured based on the 30 min questionnaire sampling interval, classroom thermal inertia, and compressor-protection considerations.	Used as an engineering heuristic to reduce frequent setpoint revisions and compressor switching.

and the ET-RHC control flowchart as illustrated in Figure 6.

The temperature range of 22–28 °C was used as a bounded virtual search envelope for classroom cooling operation, preventing the optimizer from selecting unrealistically low or high cooling setpoints. The 0.1 °C interval was adopted only as an internal numerical grid resolution for evaluating the TSV-derived group dissatisfaction objective. It should not be interpreted as the physical actuation resolution of the HVAC unit. In the deployed system, the optimized virtual setpoint was converted to the nearest HVAC-supported command before actuation. Because the tested HVAC unit accepted integer setpoint commands, the final command was rounded to the nearest 1 °C value:

$$T_{cmd}={round}_{1.0°\mathrm{C}}\left(T_{opt}\right)$$

(7)

Accordingly, the deterministic grid search provides a computationally simple method for locating the minimum of the predicted group dissatisfaction curve, whereas the final HVAC actuation remains constrained by the actual equipment resolution.

The localized actuation logic executed inside the HUB 8735 ULTRA edge chip operates under a dual-mode finite state machine:

A. Standby/Maintenance Mode ($D_{Group}^{TSV}\le 10\%$): If the collective group dissatisfaction remains within the deadband, the indoor thermal environment is deemed acceptable. The edge controller suppresses fine-tuning commands and locks the current operational state to extend the mechanical lifespan of the HVAC hardware:

$$T_{set}\left(t\right)=T_{set}\left(t-{\tau }_s\right)$$

(8)

B. Optimization Trigger Mode ($D_{Group}^{TSV}>10\%$): When the group dissatisfaction violates the comfort deadband threshold, it indicates collective thermal discomfort. To grant occupants a sufficient physiological buffer period and prevent control instability, a Time Hysteresis Mechanism enforces a strict lockout interval ($\mathsf{\Delta }t=30\min$) between consecutive setpoint revisions. This specific interval was determined by considering both the thermal inertia of the 20-person classroom and the mechanical constraints of the HVAC compressor. This ensures sufficient time for the indoor microclimate to reach a new steady-state while protecting the hardware from high-frequency start-stop cycles. Upon triggering, the system executes the deterministic grid-search optimizer (Algorithm 1) to locate the candidate setpoint with the minimum objective value $T_{opt}$, directly overwriting the setpoint:

$$T_{set}\left(t\right)=T_{cmd}$$

(9)

By utilizing the nonlinear exponential characteristic of the TSV-derived dissatisfaction penalty function, the proposed framework assigns relatively larger penalties to occupants with stronger deviations from thermal neutrality. Therefore, compared with a simple mean-TSV objective, the controller can reduce, but not eliminate, the risk of overlooking occupants experiencing higher thermal discomfort. Nevertheless, the current formulation minimizes an aggregated group-level dissatisfaction index and does not mathematically guarantee individual fairness. Accordingly, this study interprets the proposed strategy as a discomfort-mitigation approach rather than a strict fairness-guaranteed control formulation.

Fault-Detection and Fail-Safe Handling:

To improve practical robustness, a fail-safe handling logic was incorporated into the edge-side control procedure. Before each control update, the HUB 8735 ULTRA edge node screens environmental and visual inputs for missing packets, physically implausible values, frozen readings, and abrupt temporal jumps. Isolated invalid environmental readings are rejected and replaced by the latest valid moving average within the 30 min sampling buffer. If abnormal sensing persists over a complete control horizon, the controller suppresses new optimization commands and maintains the current HVAC setpoint. If reliable sensing cannot be restored, the system reverts to the conventional fixed-temperature baseline or manual operation.

For vision-derived occupant-state variables, low-confidence or unstable recognition outputs are not directly used for setpoint optimization. The system retains the previous valid observation or applies conservative default $MET$/$I_{clo}$ values to avoid abrupt control actions. If occupancy counting is completely unavailable for a sustained period, occupancy-dependent group optimization is temporarily disabled, and the controller enters the fail-safe baseline mode.

For cloud-network disconnection, the time-critical control loop remains executable because the trained INT8 TensorFlow Lite model, group dissatisfaction calculation, and deterministic setpoint search are deployed locally on the edge device. Therefore, temporary loss of cloud connectivity mainly affects data upload, storage synchronization, and future model updating, rather than immediate HVAC actuation. After reconnection, cloud synchronization and model maintenance can be resumed.

Algorithm 1 Localized Receding Horizon Search for Group Comfort Optimization

Input: Historical 5 min (${\tau }_s$) sampled data matrices for $k$ occupants over the past 30 min, ${\mathrm{X}}_i$; Previous setpoint $T_{set}\left(t-1\right);$ where N = 6. Output: Optimal HVAC temperature setpoint $T_{opt}$. Initialize minimum group dissatisfaction: $J_{min}\leftarrow 100.0$ Initialize optimal setpoint default: $T_{opt}\leftarrow 24.0°\mathrm{C}$ Compute smoothed feature vectors ${\overline{\mathrm{X}}}_i$ for all occupants $i\in \left\{1,\dots ,k\right\}$ via moving average: $${\overline{\mathrm{X}}}_i\leftarrow {\displaystyle \frac{1}{N}}{\displaystyle \sum _{m=0}^{N-1}}{\mathrm{X}}_i\left(t-m\cdot {\tau }_s\right)$$ Define discrete virtual test space: $\mathcal{T}=\left\{22.0,22.1,22.2,\dots ,28.0\right\}$ For each virtual temperature $T_{test}\in \mathcal{T}$ do     Initialize accumulated zone dissatisfaction: $Sum\_PPD\leftarrow 0$     For each occupant $i=1$ to $k$ do      Execute INT8 edge inference: ${\widehat{TSV}}_i\leftarrow f_{TFLite}\left(T_{test},{\overline{\mathrm{X}}}_i\right)$      Map predicted TSV to a PPD-inspired TSV dissatisfaction penalty: $${\widehat{D}}_i^{TSV}\leftarrow 100-95\times \exp \left(-0.03353\cdot {\widehat{TSV}}_i^4-0.2179\cdot {\widehat{TSV}}_i^2\right)$$      $Sum\_D^{TSV}\leftarrow Sum\_D^{TSV}+{\widehat{D}}_i^{TSV}$     End For     Compute current virtual group objective: $J\left(T_{test}\right)\leftarrow {\displaystyle \frac{1}{k}}\cdot Sum\_D^{TSV}$     If $J\left(T_{test}\right)<J_{min}$ then      $J_{min}\leftarrow J\left(T_{test}\right)$      $T_{opt}\leftarrow T_{test}$     End If End For Convert $T_{opt}$ to the nearest HVAC-supported command setpoint $T_{cmd}$ and actuate $T_{cmd}$ via the local HVAC controller. Clear historical 30 min data buffer; restart sampling window.

(4) 

Field Validation and Performance Evaluation

To evaluate the practical performance of the proposed system, a field comparative experiment was conducted in a classroom environment under normal teaching activities. The validation was performed over eight school days using a single-blind daily-block A/B protocol. In each two-day block, one day was operated using the traditional baseline control and the other day was operated using the proposed smart control. This two-day block was repeated four times, resulting in four validation days for each control strategy. The traditional baseline control used a fixed cooling setpoint of 25 °C, which represented the historical default cooling setting adopted by the facility management for the classroom. The control assignment followed a predefined daily-block alternating schedule rather than full randomization. The overall field validation workflow, including the daily-block A/B assignment, TSV questionnaire collection, and HVAC energy monitoring procedure, is illustrated in Figure 7.

During the experiment, the participants were not informed of the specific HVAC control strategy used on each day in order to reduce subjective expectation bias. However, because the experiment was conducted under real classroom operation, the number and composition of occupants varied across sessions. Individual participant identities were not recorded to preserve anonymity. Demographic information such as gender and age was not collected because the field validation was conducted under anonymized classroom operation. Clothing-related and activity-related variations were represented through the vision-derived clothing category and posture/behavior category used in the thermal sensation model. Therefore, the validation dataset was treated as a time-blocked classroom-level field comparison rather than a subject-level repeated-measures dataset. The daily validation schedule, control strategy assignment, valid questionnaire counts, occupancy variation, available outdoor weather information, and HVAC energy consumption are summarized in

Table 5. Summary of the daily-block field validation protocol and operating conditions.

Block/Day	Control Mode	Valid TSV Responses	Occupancy	Outdoor Condition	HVAC Energy Use
Block1-Day1	Baseline	$N$ = 228	15–22 mean 17.5	$T_a$ = 22.2–28.7 °C $RH$ = 65–94%	2.32 kWh
Block1-Day2	Proposed	$N$ = 241	10–22 mean 18.5	$T_a$ = 22.4–30.0 °C $RH$ = 53–92%	2.48 kWh
Block2-Day1	Baseline	$N$ = 284	15–23 mean 18.9	$T_a$ = 22.4–27.6 °C $RH$ = 73–90%	2.69 kWh
Block2-Day2	Proposed	$N$ = 316	13–22 mean 18.6	$T_a$ = 20.6–24.7 °C $RH$ = 76–97%	2.05 kWh
Block3-Day1	Baseline	$N$ = 268	17–24 mean 20.6	$T_a$ = 21.9–32.8 °C $RH$ = 46–95%	3.03 kWh
Block3-Day2	Proposed	$N$ = 314	13–23 mean 18.5	$T_a$ = 23.6–31.5 °C $RH$ = 54–86%	2.43 kWh
Block4-Day1	Baseline	$N$ = 258	14–22 mean 18.4	$T_a$ = 24.1–29.9 °C $RH$ = 57–84%	2.40 kWh
Block4-Day2	Proposed	$N$ = 285	14–24 mean 20.4	$T_a$ = 23.3–31.8 °C $RH$ = 52–87%	2.32 kWh

Note: Baseline represents the traditional fixed 25 °C setpoint control, whereas Proposed represents the proposed smart control strategy. Valid TSV responses indicate the number of usable questionnaire records collected on each validation day. Occupancy is reported as range/mean for each validation day. Outdoor condition refers to outdoor temperature and relative humidity obtained from the nearest Central Weather Administration weather station during the operating period. These weather data were used as supplementary context rather than for full weather normalization.

.

Subjective thermal sensation votes were collected at 30 min intervals using a standard 7-point TSV scale. To reduce pseudo-replication caused by multiple responses collected within the same classroom period, individual questionnaire responses were first aggregated into 30 min session-level mean TSV values. After aggregation, 55 valid sessions were obtained for the proposed smart control and 61 valid sessions were obtained for the fixed 25 °C baseline control. The comparison between the two control strategies was conducted using these session-level mean TSV values as the analysis units. Welch’s independent-samples t-test was adopted because the number and composition of occupants varied across sessions and equal variance could not be assumed. Cohen’s d and the 95% confidence interval of the mean difference were also reported. Because individual participant IDs were not tracked, a participant-level repeated-measures analysis or a mixed-effects model with participant-specific random effects was not applicable in this study.

For energy-saving performance evaluation, electrical energy meters were installed on the dedicated circuits of the HVAC systems to continuously record electricity consumption. Daily HVAC energy use was calculated from the cumulative kWh readings recorded during each operating period. The cumulative electricity usage and daily electricity usage were compared between the two control strategies. Outdoor temperature and relative humidity during the operating period were obtained from the nearest Central Weather Administration weather station to provide supplementary weather context. Because the field experiment was conducted under real classroom conditions and full cooling-load, occupied-hour, or weather normalization was not performed, the observed energy-saving result was interpreted as a classroom-level reduction under the tested schedule, occupancy, and available weather conditions, rather than a fully normalized annual energy-saving estimate.

3. Results

3.1. System Hardware Implementation and Validation

A validation campaign comprising 59 synchronized baseline samples was conducted to evaluate the measurement reliability of the proposed sensing system, with the descriptive statistical summaries presented in

Table 6. Descriptive statistics of environmental measurements obtained from the reference instrument and the proposed system.

Parameter	Measurement System	Sample ($\mathit{N}$)	Mean	SD	Min	Max
Ambient temperature (°C)	Reference instrument	59	25.29	0.68	24.09	27.37
	Proposed system	59	24.98	0.67	23.73	26.98
Relative humidityc (%RH)	Reference instrument	59	74.54	6.66	62.00	91.00
	Proposed system	59	74.54	6.66	61.47	91.02

. The field testing conditions covered an air temperature range of 23.7 °C to 27.1 °C and a relative humidity range of 59.5%RH to 91.5%RH, providing a representative environmental envelope for sensor validation. To benchmark the sensing fidelity, linear regression and residual error analyses were performed between the proposed system and the reference instrument, as shown in Figure 8. Figure 8a presents the temperature validation result, whereas Figure 8b and Figure 8c show the relative humidity results before and after software-based calibration, respectively.

For ambient temperature measurement, the localized edge node demonstrated high agreement with the reference instrument, achieving a coefficient of determination of $R²=0.993$, a mean absolute error ($MAE$) of 0.306 °C, and a root mean square error ($RMSE$) of 0.311 °C. These results satisfied the predefined validation threshold of ±0.5 °C, indicating that the proposed temperature sensing module is suitable for subsequent thermal comfort analysis.

For relative humidity measurement, the raw edge sensor exhibited strong linearity with the reference instrument, with $R²=0.998$. However, the raw readings showed a systematic negative bias of −8.980%RH, with an $MAE$ of 8.980%RH and an $RMSE$ of 8.984%RH, exceeding the predefined acceptable threshold of ±5.0%RH. This indicates that the raw humidity sensor was able to capture the relative variation trend but required calibration to correct its absolute measurement bias. Based on the regression relationship between the raw sensor output and the reference measurement, expressed as $R{H}_{raw}=0.992R{H}_{ref}-8.35$, a software-based linear calibration function was implemented in the HUB 8735 ULTRA edge firmware:

$$R{H}_{calibrated}={\displaystyle \frac{R{H}_{raw}+8.35}{0.992}}$$

(10)

After applying the calibration function, the humidity measurement error was substantially reduced, with the MAE decreasing to below 1.0%RH. Nevertheless, the pre-calibration results indicate a hardware limitation of the low-cost humidity sensor. Although the raw sensor output captured the relative variation trend of indoor humidity, it was not sufficiently accurate for absolute RH measurement before compensation. Therefore, all RH values used in the subsequent thermal sensation inference and HVAC control process were based on the software-calibrated output implemented in the edge firmware. The calibrated humidity measurement satisfied the predefined ±5.0%RH threshold, demonstrating that the proposed sensing platform can provide reliable environmental telemetry after software compensation. Overall, the validation results confirm that the calibrated edge hardware platform provides sufficiently accurate temperature and humidity measurements for subsequent cloud-based thermal sensation inference and localized receding-horizon optimization.

In addition to the 24 h temperature and humidity validation, preliminary field checks were conducted using a TM-402 anemometer and a TM-188 black-globe thermometer to examine whether continuous air-velocity and radiant-temperature sensing should be included in the low-cost prototype. The spot-check results indicated that the classroom was characterized by near-still-air conditions and no evident radiant heat source or strong radiant asymmetry during the tested periods. Therefore, these two measurements were used only to justify the field-specific simplification of the deployed low-cost sensing configuration and were not included as continuous online model inputs. The calibration coefficient was kept fixed during the subsequent short-term field validation. However, repeated reference-instrument comparisons across multiple days were not conducted in the present study. Therefore, the long-term drift stability of the humidity calibration remains a limitation and should be further evaluated in extended deployments.

3.2. Validation of Vision-Based Occupant-State Extraction

To evaluate the practical reliability of the vision-based occupant-state extraction module, the recognition outputs generated by the HUB 8735 ULTRA edge board were compared with on-site observer logs during classroom deployment. The validation focused on the consistency between the edge-recognized outputs and the observer-recorded ground truth for three structured occupant-state variables used in the proposed control framework: occupancy count, behavior/posture category, and clothing category. For occupancy counting, the validation sample unit was the image frame. For behavior/posture and clothing-category recognition, the validation sample unit was the person-level annotation because each image frame could contain multiple occupants. In addition, challenging visual conditions, including partial body occlusion, overlapping occupants in the seating area, and ambiguous clothing layers, were considered to assess the robustness of the edge-deployed recognition module under practical classroom conditions. To further examine recognition failure under challenging visual conditions, 10 face-occlusion/overlap frames were selected as a supplementary stress-test set. In these frames, the face-region occlusion ratio ranged from approximately 38% to 52%. Recognition was generally successful when the face-region occlusion ratio was below approximately 44%, whereas recognition failure occurred when the occlusion ratio reached or exceeded approximately 44%. Therefore, this value should be interpreted as an empirical failure boundary observed in the supplementary stress test, rather than as a generalizable threshold. The validation was conducted under normal classroom lighting conditions. Controlled low-light, backlight, and non-uniform illumination tests were not included in the present study. The validation results are summarized in

Table 7. Validation results of the vision-based occupant-state extraction module.

Task	Ground Truth	Validation Samples	Metric	Result
Occupancy counting	On-site observer logs	107 frames	Accuracy	94.8%
Behavior/posture recognition	On-site observer logs	107 person-level labels	Accuracy	94.8%
Clothing-category recognition	On-site observer logs	114 person-level labels	Accuracy	93.8%
Face Occlusion/overlap conditions	On-site observer logs	10 challenging frames; face-region occlusion ratio = 38–52%	Empirical failure boundary	≥44% face occlusion

Note: The on-site observer logs were used as manually recorded ground-truth references. Accuracy was calculated as the proportion of recognition outputs consistent with observer-recorded labels. For the face-occlusion/overlap test, the occlusion ratio was estimated from the visually blocked proportion of the manually identified face region. The 44% value is an empirical failure boundary observed in this supplementary stress test, not a generalizable occlusion threshold or a comprehensive robustness benchmark.

.

3.3. Quantitative Evaluation of Quantized Edge-AI Sensation Model

To verify the inference fidelity of the INT8-quantized TensorFlow Lite thermal sensation prediction model $\left(f_{TFLite}\right)$ deployed on the HUB 8735 ULTRA chip, a comprehensive validation analysis was conducted using an experimental dataset comprising 2194 samples. The model performance was evaluated using exact classification accuracy, tolerance-based accuracy, macro-averaged F1 score, residual error distribution, class-specific error structure, and class-wise prediction performance. The PC-based Float32–INT8 model comparison and the deployment evaluation of the INT8 model are summarized in

Table 8. PC-based comparison between the Float32 and INT8 TFLite models and deployment evaluation of the INT8 model.

Metric	Float32 Model	INT8 TFLite Model	Evaluation Setting
Model size	245 MB	23.1 MB	PC-based model-file comparison; 90.60% reduction
Inference latency	2 s	1 s	PC-based comparison; 50.00% reduction
Peak RAM usage	2.1 GB	1.8 GB	PC-based comparison; 14.29% reduction
Prediction-label agreement with Float32	—	100%	PC-based comparison using the same validation dataset
INT8 edge prediction evaluation	—	Reported in Figure 8 and Table 9	INT8 model deployed on HUB 8735 ULTRA

. The overall diagnostic results are summarized in Figure 9, while the class-wise distribution and prediction performance are listed in

Table 9. Class-wise distribution and prediction performance of the edge-deployed TSV model.

Measured TSV Class	Support (n)	Support (%)	Predicted Mean TSV	MAE	Exact Accuracy/Recall	F1-Score
−3 (Cold)	0	0.0%	—	—	—	—
−2 (Cool)	51	2.3%	−1.03	0.97	0.0%	0.00
−1 (Slightly Cool)	986	44.9%	−0.68	0.34	73.6%	0.78
0 (Neutral)	972	44.3%	−0.03	0.32	81.5%	0.76
+1 (Slightly Warm)	179	8.2%	+0.72	0.41	65.9%	0.62
+2 (Warm)	6	0.3%	+1.37	0.63	16.7%	0.17
+3 (Hot)	0	0.0%	—	—	—	—
Overall	2194	100%	—	0.35	74.6%	Macro-F1 = 0.47

Note: Recall represents the exact classification accuracy within each measured TSV class. The F1-score was calculated from the confusion matrix by considering both precision and recall. Macro-F1 was calculated by averaging the class-wise F1-scores across the observed TSV classes. Predicted TSV values were converted into discrete TSV classes using the same class-assignment rule before constructing the confusion matrix. Classes with zero support were not evaluated.

.

Table 8 summarizes the PC-based comparison between the original Float32 model and the INT8 TFLite model, together with the deployment evaluation of the INT8 model. The PC-based comparison includes model size, inference latency, peak RAM usage, and prediction-label agreement using the same validation dataset. The results show that INT8 quantization reduced the model size, inference latency, and peak RAM usage while producing the same prediction labels as the Float32 model for the tested samples. After this PC-based comparison, the INT8 TFLite model was deployed on the HUB 8735 ULTRA platform for edge inference. The prediction performance of the edge-deployed INT8 model is reported in Figure 9 and Table 9. Therefore, Table 8 should be interpreted as a PC-based Float32–INT8 model comparison combined with an INT8 edge-deployment evaluation, rather than as a full cross-platform benchmark of both models on both PC and edge hardware.

As shown in the confusion matrix in Figure 9a, the edge-deployed model achieved an exact-match classification accuracy of 74.6%. When allowing a tolerance interval of ±1 thermal sensation scale, the prediction accuracy increased to 99.9%, indicating that almost all predictions were located within the adjacent thermal sensation category. To further evaluate the prediction bias and error dispersion, the residual distribution, defined as the difference between the predicted and measured thermal sensation votes, was analyzed in Figure 9b. The residuals showed a mean value of +0.126 and a standard deviation of 0.424, suggesting a slight positive bias with limited dispersion around the measured labels.

The class-specific error analysis in Figure 9c shows that the mean absolute error ($MAE$) remained relatively low in the dominant comfort-related categories, particularly for the Neutral and Slightly Cool levels, with $MAE$ values of 0.32 and 0.34, respectively. Larger errors were observed at the extreme thermal sensation levels, especially in categories with fewer samples. This pattern is further reflected in Figure 9d, where the mean predicted values tend to shift toward the central thermal sensation range, indicating a mild regression-to-the-mean effect. This behavior is expected in imbalanced ordinal classification tasks and suggests that additional samples at extreme thermal sensation levels would be beneficial for improving boundary prediction performance. Nevertheless, the high ±1-scale accuracy indicates that the quantized edge-AI model provides sufficient prediction stability for subsequent group thermal comfort control and localized receding-horizon optimization.

To further evaluate the model under the imbalanced TSV distribution, the class-wise F1-score and macro-averaged F1-score were calculated from the confusion matrix. Macro-F1 was used because it assigns equal weight to each observed TSV class and is therefore more sensitive to weak performance in minority thermal sensation classes than overall accuracy. As shown in Table 9, the TSV dataset was strongly imbalanced toward the near-neutral categories. The Slightly Cool and Neutral classes accounted for 44.9% and 44.3% of the total samples, respectively, representing 89.2% of the dataset. In contrast, the extreme discomfort categories were under-represented, with only 2.3% of samples in the Cool class and 0.3% in the Warm class, while no samples were available for the Cold and Hot classes. This imbalance explains why the model achieved relatively good performance for the dominant near-neutral classes but showed weaker exact classification performance for the extreme categories. For example, the measured Cool class had a predicted mean TSV of −1.03 and an exact recall of 0.0%, indicating that the model tended to shift extreme cold responses toward the Slightly Cool range. Similarly, the measured Warm class had a predicted mean TSV of +1.37 and an exact recall of 16.7%, suggesting a regression-to-the-mean tendency. Therefore, although the overall exact accuracy was 74.6% and the overall MAE was 0.35, the prediction of extreme thermal discomfort should be interpreted with caution. Additional targeted data collection under hotter and colder indoor conditions is required to improve boundary-class prediction performance. Although a full control-decision sensitivity experiment was not conducted in the present study, the potential influence of TSV prediction errors on the final setpoint decision was further discussed. Because the ET-RHC controller uses group-level TSV-derived dissatisfaction and a comfort deadband rather than directly responding to a single individual TSV prediction, adjacent-category prediction errors are expected to have limited influence on the final HVAC command. This interpretation is also supported by the high ±1 TSV accuracy of the INT8 model. However, the weaker prediction performance for under-represented extreme TSV classes may reduce the controller’s sensitivity under rare strong hot or cold discomfort conditions. This issue should be further examined in future studies using control-decision sensitivity analysis.

3.4. Field Experimental Evaluation of Collective Thermal Comfort

The field experimental results obtained from the single-blind daily-block A/B comparison are shown in Figure 10. Figure 10a presents the distribution of individual TSV responses; however, because individual participant IDs were not tracked, occupant-level longitudinal comfort trajectories could not be analyzed in the present study. Figure 10b presents the statistical comparison based on 30 min session-level mean TSV values.

As shown in Figure 10a, the proposed smart control shifted the TSV distribution closer to the neutral category. Under the smart control strategy, the largest proportion of votes was concentrated at the Neutral level, whereas the fixed 25 °C baseline showed a higher proportion of votes in the Slightly Cool category. This indicates that the fixed 25 °C baseline tended to produce an overcooled thermal perception in the tested classroom, while the proposed control strategy reduced this overcooling tendency.

For statistical comparison, individual TSV responses collected within each 30 min interval were first aggregated into session-level mean TSV values to reduce pseudo-replication. After aggregation, 55 valid sessions were obtained for the proposed smart control and 61 valid sessions were obtained for the fixed 25 °C baseline. As shown in Figure 10b, the proposed smart control achieved a mean session-level TSV of −0.13, whereas the fixed 25 °C baseline resulted in a mean session-level TSV of −0.66. Welch’s independent-samples t-test showed a significant difference between the two control strategies, Welch’s t(100) = 11.2, p < 0.001. The mean difference, calculated as Smart Control minus Fixed 25 °C Baseline, was +0.53, with a 95% confidence interval of [0.44, 0.63]. The effect size was large, Cohen’s d = 2.11, indicating a substantial shift in the session-level mean TSV toward the neutral thermal sensation level under the proposed smart control.

3.5. Edge Actuation Trajectories and Resource-Saving Performance

The localized closed-loop actuation behavior and resource efficiency of the embedded edge-computing platform are evaluated in this subsection. The time-varying HVAC setpoint ($T_{set}$), group mean Thermal Sensation Vote (Mean TSV), and collective dissatisfaction index ($D_{Group}^{TSV}$) are shown in Figure 11, while the corresponding electrical power profiles and cumulative energy metrics are presented in Figure 12.

As shown in Figure 11a,b, the HUB 8735 ULTRA edge controller executed the rolling grid-search routine described in Algorithm 1 at 30 min intervals and dynamically adjusted the HVAC setpoint within the range of 25.0 °C to 27.0 °C. The setpoint trajectory indicates that the controller increased or decreased the cooling setpoint in response to changes in the group thermal sensation state, while the Mean TSV remained generally close to the neutral range.

The $D_{Group}^{TSV}$ results in Figure 11c,d further show that the proposed smart control strategy reduced collective thermal dissatisfaction compared with the fixed 25 °C baseline. During the representative test day, the smart control strategy maintained a lower $D_{Group}^{TSV}$ profile than the baseline for most time intervals, while avoiding the pronounced dissatisfaction peak observed under fixed-temperature operation. Across the multi-day experimental horizon, the smart control strategy also showed a consistently lower $D_{Group}^{TSV}$ trend than the fixed 25 °C baseline. Although the $D_{Group}^{TSV}$ was not always below the 10% comfort target, the proposed controller reduced the frequency and magnitude of high-dissatisfaction periods, indicating improved group-level comfort stability under dynamic occupancy and thermal load conditions.

Compared with a simple fixed-setpoint rule-based controller, the proposed grid-search controller does not rely on a single predefined temperature or a one-step heuristic adjustment. The baseline fixed 25 °C strategy represents the facility-default rule-based operation, which is easy to implement but cannot respond to dynamic occupancy, clothing, posture, or group thermal sensation variation. In contrast, the proposed controller evaluates multiple candidate setpoints using the edge-deployed TSV prediction model and selects the setpoint that minimizes the TSV-derived group dissatisfaction index. Therefore, the grid-search strategy provides a more adaptive but still computationally lightweight control logic than the fixed rule-based baseline. However, this study did not conduct an additional benchmark against other rule-based variants, such as ±1 °C threshold control or proportional rule-based control. This comparison should be included in future work.

In terms of power consumption dynamics, the full-time instantaneous power profiles in Figure 12a show the temporal variation of electrical demand under the proposed smart control strategy and the fixed 25 °C baseline. The baseline condition exhibited several high-power transient peaks, whereas the smart control strategy generally maintained a lower power profile during multiple operational periods. To further examine the underlying operating characteristics, Figure 12b presents the steady-state power profiles after excluding start-up inrush periods. The results indicate that the proposed edge-based control strategy reduced the steady-state electrical demand for most of the evaluation period, although short-term fluctuations remained due to compressor cycling and dynamic cooling load variations.

The cumulative energy consumption shown in Figure 12c indicates that the proposed smart control consumed less electricity than the fixed 25 °C baseline over the evaluation horizon. The cumulative reduction was 1.16 kWh, corresponding to an observed 11.1% reduction in total HVAC electricity use during the validation period. To further provide statistical uncertainty for the energy comparison, daily HVAC electricity consumption was summarized across the four A/B validation blocks, as shown in Figure 12d. The proposed smart control consumed 2.32 ± 0.31 kWh/day, whereas the fixed 25 °C baseline consumed 2.61 ± 0.51 kWh/day, expressed as mean ± 95% confidence interval. This corresponds to an observed reduction of 0.29 kWh/day, or 1.16 kWh over the four paired validation blocks.

Because each control strategy was tested for only four days and full weather/load normalization was not performed, the energy-saving result should be interpreted as a descriptive classroom-level observation under the tested schedule, occupancy, and available outdoor weather conditions, rather than as a fully normalized or generally transferable annual energy-saving estimate. Overall, the results suggest that the proposed edge-based control framework reduced unnecessary cooling energy use while maintaining improved group-level thermal comfort compared with the conventional fixed-temperature baseline.

4. Discussion

The present study addresses the practical trade-off between HVAC energy saving and collective thermal comfort in multi-occupant indoor environments. This issue is inherently complex because thermal comfort is affected not only by environmental variables, such as air temperature, relative humidity, air velocity, and mean radiant temperature, but also by time-varying human factors, including occupancy, clothing insulation, and behavioral activity [14,18,24,46]. In conventional HVAC systems, several of these human-related parameters are either assumed to be constant or manually estimated, which limits the ability of the controller to respond to real indoor conditions. To overcome this limitation, the proposed system integrates environmental sensing, electrical power monitoring, occupancy detection, and AI-based visual recognition of clothing and behavioral features into a unified edge-cloud control framework. Through this design, dynamic occupant-related variables can be incorporated into the thermal sensation prediction and HVAC control process, rather than being treated as fixed assumptions.

Previous review studies on dynamic thermal comfort have emphasized that both environmental parameters and human factors should be considered in real-time comfort evaluation [8,32,47]. Similarly, IoT-based studies on indoor air quality and smart energy management have highlighted the importance of integrating sensing, communication, and control technologies for HVAC optimization [21]. However, many existing studies remain focused on conceptual frameworks, simulation-based analyses, or partial system implementation. In contrast, the present study provides a field-deployed implementation that links edge data acquisition, cloud-based model training, edge-deployed inference, and localized HVAC actuation. Therefore, the contribution of this work lies not only in proposing a comfort prediction model, but also in demonstrating a complete operational pathway from multimodal perception to physical air-conditioning control.

The proposed framework is also related to previous work on computer-aided environmental thermal comfort sensing [4,11,48], which developed sensing and measurement approaches for evaluating indoor thermal comfort. However, the present study places greater emphasis on closed-loop implementation and energy-saving control. Instead of limiting the system to environmental monitoring or comfort assessment, the proposed architecture uses the sensed and inferred variables to drive a dynamic HVAC control strategy. This distinction is important because a thermal comfort sensing system does not necessarily lead to energy savings unless its outputs are further incorporated into a real-time actuation mechanism.

Compared with optimization-based HVAC control studies, such as ANN–PSO-based cooling management systems that use artificial neural networks to predict temperature or operating modes and particle swarm optimization to select suitable cooling strategies [21], the proposed system adopts a different implementation-oriented architecture. The cloud platform is used mainly for thermal sensation model training and model optimization, whereas the real-time inference and setpoint decision are executed locally on the edge device. This cloud-training and edge-control design reduces reliance on continuous cloud connectivity and improves the feasibility of deployment in classrooms and small shared spaces. Furthermore, the use of an INT8-quantized TensorFlow Lite model allows thermal sensation inference to be performed on resource-constrained embedded hardware, which supports low-latency and localized control.

The sensing validation results indicate that the localized edge-sensing node can provide sufficiently reliable environmental telemetry after appropriate calibration. As illustrated in Figure 8a, the temperature measurements showed good agreement with the reference instrument and satisfied the predefined accuracy threshold for thermal comfort assessment. Although the raw relative humidity sensor exhibited a systematic negative bias, the software-based linear correction substantially reduced the measurement error and brought the calibrated readings within the acceptable range, as shown in Figure 8c. This result is important for practical IoT-based building control because low-cost sensing modules often suffer from sensor-to-sensor variation and systematic bias. The findings suggest that, when a calibration stage is included before deployment, compact edge-sensing nodes can still provide data quality adequate for downstream thermal sensation modeling and control optimization.

The performance of the INT8-quantized thermal sensation model further supports the feasibility of deploying data-driven comfort prediction on resource-constrained edge hardware. The exact classification accuracy of 74.6% indicates that the model was able to identify the precise TSV category for most samples, as shown in Figure 9. More importantly, the ±1-scale accuracy reached 99.9%, showing that almost all prediction errors remained within an adjacent thermal sensation level. Since TSV is an ordinal subjective scale, adjacent categories such as Neutral and Slightly Cool may not always be clearly distinguishable by occupants. Therefore, tolerance-based accuracy is particularly relevant for closed-loop HVAC control. The residual analysis also showed only a slight positive bias, suggesting that the quantized model retained sufficient inference fidelity after conversion from Float32 to INT8. Nevertheless, larger errors at extreme TSV levels reveal a limitation of the current dataset. The model tended to shift predictions toward the central thermal sensation range, which is a common behavior in imbalanced ordinal classification tasks. Future data collection should therefore include more samples under warmer and cooler boundary conditions to improve prediction robustness at the extremes.

The field comparison between the proposed smart control strategy and the fixed 25 °C baseline shows that the proposed system improved group-level thermal sensation while reducing energy consumption in the tested classroom. After aggregation at the 30 min session level, the mean TSV under smart control was −0.13, which was closer to thermal neutrality than the fixed baseline condition with a mean TSV of −0.66, as shown in Figure 10. The mean difference was +0.53 TSV scale units with a 95% confidence interval of [0.44, 0.63], and the effect size was large (Cohen’s d = 2.11). This shift indicates that the conventional 25 °C setting produced a cooler-than-neutral perception in the tested classroom, whereas the proposed controller reduced overcooling by adjusting the setpoint within a moderate range. The observed electricity reduction of 1.16 kWh, corresponding to an 11.1% saving over the four paired validation blocks, suggests that comfort improvement and energy reduction were observed simultaneously under the tested classroom conditions. However, this energy result should be interpreted as a short-term classroom-level observation rather than a weather-normalized annual saving estimate. The observed energy reduction was mainly associated with reduced overcooling through dynamic setpoint adjustment, because the proposed controller often operated at a moderately higher cooling setpoint than the fixed 25 °C baseline while maintaining the group TSV closer to neutrality. The lower steady-state power profile and fewer high-power transient peaks also suggest a possible reduction in unnecessary cooling demand. However, the present short-term field test did not separately quantify the relative contributions of higher setpoints, reduced compressor cycling, and occupancy-aware control. This decomposition should be examined in future seasonal experiments or simulation-based analyses.

The $D_{Group}^{TSV}$ results provide additional insight into the behavior of the proposed controller. Although the $D_{Group}^{TSV}$ was not maintained below the 10% comfort target at all times, the proposed strategy reduced the frequency and magnitude of high-dissatisfaction periods compared with the fixed-temperature baseline, as shown in Figure 11. This result is reasonable in real classrooms, where occupancy, clothing insulation, activity level, and local microclimate conditions vary over time. The use of a comfort deadband and a 30 min lockout interval also means that the controller intentionally avoids excessive setpoint changes in response to short-term fluctuations. Such a design is beneficial for practical HVAC operation because frequent compressor switching may increase mechanical stress and reduce system stability. Therefore, the proposed event-triggered control logic represents a compromise between thermal responsiveness, equipment protection, and energy efficiency.

From an implementation perspective, the proposed edge-cloud architecture offers several practical advantages over cloud-centered control frameworks. The cloud platform provides sufficient computational resources for model training, while the edge device performs real-time data processing, model inference, group discomfort calculation, and HVAC setpoint search. This division of labor reduces dependence on continuous network connectivity, lowers the risk of cloud-side communication delays, and improves the feasibility of deployment in educational or office environments. In addition, the non-invasive visual extraction of occupancy, posture, and clothing-related $I_{clo}$ features reduces the need for frequent manual feedback from occupants. This is particularly relevant for real-world operation, where questionnaire-based or app-based feedback may cause user fatigue and reduce long-term system reliability.

Limitations and Future Work

Although the proposed system demonstrated short-term feasibility in a real classroom environment, several limitations should be noted.

First, the humidity calibration was derived from a single classroom validation condition. The calibration function was kept fixed during the subsequent short-term field validation, but repeated reference-instrument comparisons over multiple days were not conducted in the present study. Therefore, the long-term drift stability and transferability of the humidity calibration remain to be further verified. Because low-cost humidity sensors may exhibit sensor-to-sensor bias, aging drift, and environment-dependent errors, a brief reference-instrument check or recalibration is recommended when the system is deployed in another room, another season, or a substantially different humidity range.

Second, the vision-based occupant-state estimation module was validated only under normal classroom lighting conditions. Dedicated lighting-sensitivity tests under low-light, backlight, and non-uniform illumination conditions were not conducted. In addition, partial body occlusion, overlapping occupants, partial face occlusion, and ambiguous clothing layers may reduce recognition confidence. A supplementary face-occlusion/overlap stress test was conducted using 10 challenging frames with face-region occlusion ratios ranging from approximately 38% to 52%. The results suggested that recognition failure occurred when the face-region occlusion ratio reached or exceeded approximately 44%. However, this value should be interpreted as a preliminary empirical observation rather than a generalizable threshold. The reduced recognition success under these challenging conditions indicates that partial face occlusion remains a major limitation of the current vision-based occupant-state extraction module. In the present implementation, low-confidence visual predictions were handled by retaining the previous valid observation or by applying conservative default $MET$/$I_{clo}$ values to avoid abrupt control actions. Therefore, future work should include controlled lighting tests, larger-scale occlusion scenarios, and more diverse classroom layouts to further quantify the robustness of the vision module.

Third, the present prototype was designed as a low-cost classroom-level deployment. Continuous air-velocity and mean radiant temperature sensing were not included in the online feature vector. Air velocity was treated as a near-still-air boundary condition, and MRT was not continuously measured during model training or control inference. Therefore, the proposed model should be interpreted as a data-driven TSV prediction and control model under near-still-air and low-radiant-asymmetry classroom conditions, rather than as a complete six-parameter PMV-based measurement system. Future high-precision deployments should integrate continuous anemometer and black-globe measurements, especially in spaces with strong local air movement, radiant cooling/heating systems, large glazed facades, or substantial solar radiation.

Fourth, the thermal sensation dataset was imbalanced toward near-neutral TSV categories. As a result, prediction performance for extreme thermal discomfort cases was weaker and should be interpreted with caution. Additional targeted data collection under hotter and colder indoor conditions is needed to improve boundary-class prediction performance. Although the INT8 model showed acceptable short-term edge inference performance, long-term edge-node power consumption, thermal stability, and inference robustness under extended operation were not systematically evaluated. In addition, the deployed INT8 TSV prediction model was kept fixed during the field validation. Potential model drift caused by seasonal changes, changes in occupant composition, clothing habits, classroom schedules, or HVAC operating conditions was not examined in the present study. Future work should therefore investigate periodic model recalibration or privacy-preserving online learning using newly collected environmental, occupant-state, and TSV feedback data, so that the edge model can maintain prediction reliability during long-term deployment. From an ethical and privacy perspective, although raw camera images were processed locally and were not uploaded to the cloud server in the present implementation, broader deployment should further consider informed consent, data-retention policies, access control, and privacy-preserving feature extraction mechanisms.

Fifth, the proposed fail-safe rules improve the practical robustness of the edge-cloud control framework, but systematic fault-injection experiments were not conducted in the present study. Specifically, controlled tests involving sensor outage, complete visual-recognition loss, and cloud-network disconnection were not included. Future robustness tests should quantify recovery time, control stability, and comfort degradation under these abnormal operating conditions.

Sixth, the field validation was conducted over eight school days in a classroom setting, with four days assigned to each control strategy. The number and composition of occupants varied across sessions, and individual participant IDs were not recorded. Therefore, subject-level repeated-measures analysis and participant-level mixed-effects modeling were not applicable. In addition, the observed energy-saving result was not fully normalized by outdoor weather, cooling load, or occupant density. The energy reduction should therefore be interpreted as a short-term classroom-level observation rather than a weather-normalized or generally transferable annual saving estimate.

Finally, the 10% comfort deadband threshold and 30 min lockout interval were adopted as practical control settings rather than globally optimized parameters. Although these values were selected based on standard-informed comfort acceptability criteria, the questionnaire sampling interval, classroom thermal inertia, and compressor-protection considerations, their influence on control performance was not systematically quantified. Future work should conduct sensitivity analyses of different comfort deadband thresholds and lockout intervals to evaluate their effects on comfort responsiveness, energy consumption, and compressor switching frequency.

5. Conclusions

This study developed and preliminarily field-validated an edge-cloud IoT framework for dynamic group thermal sensation control and HVAC energy optimization in classroom-based shared indoor environments. The proposed system integrates calibrated temperature and humidity sensing, non-invasive visual feature extraction, cloud-based thermal sensation model training, INT8 edge deployment, and event-triggered rolling-horizon setpoint optimization. In the implemented low-cost prototype, the online model input was based on indoor air temperature, relative humidity, and vision-derived occupant-state features, including posture-related $MET$ and clothing-related $I_{clo}$. Air velocity was treated as a near-still-air boundary condition, and mean radiant temperature was not included as a continuous online input under the tested low-radiant-asymmetry classroom conditions.

The sensing validation confirmed that the calibrated edge node provided reliable temperature and humidity measurements for downstream comfort inference. The quantized thermal sensation model achieved stable edge inference performance, supporting real-time group comfort estimation on resource-constrained hardware. In the eight-school-day single-blind daily-block A/B field validation, the proposed control strategy shifted the 30 min session-level mean TSV closer to neutrality and reduced overcooling compared with the fixed 25 °C baseline. Meanwhile, cumulative electricity consumption was reduced by 1.16 kWh, corresponding to an observed 11.1% reduction during the experimental period.

These findings demonstrate the short-term feasibility of combining edge intelligence, multimodal sensing, and group-level comfort optimization for human-centric HVAC control in classroom settings. However, the present field validation should be regarded as preliminary because it was conducted in a single classroom over an eight-school-day A/B campaign, with four validation days for each control strategy. Therefore, the observed 1.16 kWh reduction, corresponding to an 11.1% energy saving during the test period, should be interpreted as a promising short-term result rather than a generalized annual energy-saving estimate. Future studies should extend validation across multiple rooms, building types, seasons, and longer operating periods with weather- and load-normalized energy analysis. Further improvement of TSV prediction for extreme hot and cold discomfort categories is also needed to enhance control robustness under more diverse occupant and environmental conditions. Overall, the proposed system can be considered a promising prototype for smart building control, but broader field validation is required before large-scale deployment.