Occupant-Detection-Based Individual Control of Four-Way Air Conditioner for Sustainable Building Energy Management

Abstract

In this study, individual control of a four-way air conditioner was developed based on the distribution of occupants to prevent unnecessary energy consumption during room-wide control. An occupancy detection algorithm was created in Python using YOLOv5 object recognition technology to identify the occupants’ distribution in space. Recorded video data were used to test the algorithm. A simulation case study for a building energy model was conducted, assuming that this algorithm was applied using surveillance cameras in commercial buildings, such as cafés and restaurants. A grey-box model was established based on measurements in a thermal zone, dividing one space into two zones. The temperature data for the two zones were collected by individually turning on the air conditioner for each zone in turns for a specific period. Manual closure was applied to each supply blade using a tape to provide cooling to the target zone. Finally, through energy simulations, the decreased rates in energy consumption between the proposed individual control and existing room-wide controls were compared. Different scenarios for the occupants’ schedules were considered, and average rates in energy savings of 21–22% were observed, demonstrating the significance of individual control in terms of energy consumption. However, marginal comfort violations were observed, which is inevitable. The developed control method is expected to contribute to sustainable energy management in buildings.

1. Introduction

1.1. Background

Recently, thermal comfort has become one of the foremost requirements for occupancy because of improvements in occupants’ living standards. The relationship between occupants and building systems is emphasized in terms of comfort and energy [1]. Therefore, energy consumption in the building sector continues to increase globally. According to the 2021 building energy consumption data released by the Korea Energy Agency, the electricity consumption rate for buildings increased by approximately 14%, rising from 57.1% to 72.0% between 2010 and 2022 [2]. In addition, according to energy statistics in the United States in 2023, the energy consumption rate of buildings accounts for more than 37% of the primary energy, and those related to heating, ventilating, and air-conditioning (HVAC) account for approximately 40% of the total energy consumption of buildings [3]. This energy consumption directly contributes to carbon dioxide emissions and the corresponding portion of the building sectors is 26%. Therefore, reducing building energy consumption is an urgent task that must be resolved for environmental conservation and sustainability of the earth.

However, typical HVAC systems operate over an entire zone without considering the presence or absence of occupants, leading to unnecessary energy use. Hence, various studies have been conducted to predict occupancy density using passive infrared (PIR) sensors [4,5,6] and CO₂ sensors [7,8]. The development of the Internet of Things in buildings has facilitated the deployment of various sensors and the accumulation of data for analysis [9]. If more granular control within the zone based on occupancy and its distribution is possible, the amount of unnecessary energy used in building energy consumption can be reduced by approximately 20–45% [10]. This potential privacy challenge can be resolved because the two methods do not record videos of the occupants. However, in the case of CO₂ sensors, it is impossible to correctly locate the occupant, as CO₂ is emitted into space while the occupants exhale [11]. Therefore, this method is unsuitable for the occupancy-based individual control of HVAC systems. Previous studies have shown that energy savings can be increased when the space is compartmentalized, and space control is achieved using the location details of the indoor occupants rather than the distribution rate of the indoor occupants [12]. Consequently, it is more effective to implement individual controls based on occupancy data regarding the spatial positioning of occupants, rather than merely accounting for their numbers. Acquisition of indoor occupancy data necessitates the use of cameras; however, concerns regarding occupant privacy persist in this domain [13]. Privacy-safeguarding technologies may be implemented to proactively prevent potential privacy issues. This includes image-blurring technology, which makes it impossible to distinctly identify occupants in recordings, and a data deletion protocol after information utilization. These measures yielded favorable survey outcomes for video documentation [10]. Furthermore, the survey results were more positive when the participants were aware of their contributions to reducing building energy consumption.

1.2. Literature Review

One of the most recent review studies reported the possibility and applicability of computer vision for detecting the occupants of buildings [14]. Through an extensive review, the recent advancements in computer vision technologies and their control applications in air quality, thermal, and lighting systems were summarized. Another review surveyed vision-based-sensing systems for information on the occupants of buildings [15]. The most recent sensing technologies and vision-based control strategies and analyzed energy-saving potential with occupant information were reviewed.

You Only Look Once (YOLO), an object recognition technology, is a deep-learning-based framework that identifies a specified object and verifies its location using a bounding box. The input image is partitioned into specific grids, processed using a neural network, and class predictions are generated, which affect object detection [16]. This technique is used to detect the number of people based on video camera operations [17]. Moreover, it has been applied in various engineering fields, such as the construction industry for the safety of site workers [18,19] and built environments for occupant-related research.

Faster region-based convolutional neural network (R-CNN) technology was used to detect and recognize the activities of office occupants in real-time, achieving an accuracy of up to 97.32% [20]. The study also investigated the effects of activity detection and the corresponding internal heat gain from occupants. The heat-gain profiles of the static schedule and the proposed prediction-based schedule showed a difference of approximately 55%, which is equivalent to 8.3 kW. However, faster R-CNN technology has disadvantages, including slow detection speed and the need for high memory owing to its complex structure in comparison to YOLO technology. Therefore, the faster R-CNN is unsuitable for implementation in occupancy-detection-based HVAC operations that require instant control actions.

Recent studies have focused on air quality and viral transmission in buildings. Some of these studies focused on efficient ventilation systems to investigate viral transmission in buildings [21,22,23]. A hardware prototype with a Raspberry Pi was developed to provide appropriate ventilation on the HVAC side of airports and train stations crowded with passengers [21]. The email alert system was integrated with coding, and the infection probability and energy consumption were analyzed. A similar approach to the hardware prototype using a Raspberry Pi and camera was adopted to proactively provide fresh air with ventilation systems in a campus office space [22]. Demand-response cooling strategies have been proposed using cameras with deep-learning algorithm applications (for example, YOLO) [23]. Chamber experiments were conducted to evaluate the thermal comfort of the subject to validate the proposed method.

Other researchers focused on evaluating the comfort of occupants using a robot [24] and a camera [13]. A robot system with a kinetic and thermal camera captured occupants’ details, such as gender, age, clothes, and body temperature, to estimate thermal comfort [24]. This estimation is not limited to a single person but can detect people in a room; thus, the thermal comfort distribution can be determined. In another study, a camera was used to estimate the metabolic rate and clothing insulation to determine thermal comfort based on deep vision technology [13]. The researchers focused more on HVAC operation with chamber tests and obtained more satisfactory results in terms of the thermal sensation of occupants between the proposed comfort-based control and fixed-setpoint control.

Various building energy models have been established, including white, grey, and black boxes. White-box models are building energy simulation tools, such as Energy Plus [25] and TRNSYS [26]. They are based on prior knowledge of heat transfer via the envelope between the outside and conditioned zones. Detailed algorithms for heat transfer, such as conduction, convection, and radiation, have been developed. A disadvantage of white-box methods is that only one node for each zone is calculated. Artificial walls or vertical holes should be set, which might not have been sufficiently validated, to create different thermal environments, such as setpoint temperatures. Black-box models are based on the input-output relationship and are currently regarded as machine-learning methods [27]. However, a significant amount of data to ensure robustness of the prediction and simulation results is required. Moreover, the black-box approach neglects the physical meaning of the building parameters. Grey-box models, which are positioned between these two box models mentioned above, have advantages such as the physical interpretation of parameters and reasonable prediction performance [28]. They are based on the heat balance equation of each node, such as the room air and envelope, and the granularity can be set based on the design of the conditioned zone. This is suitable for modeling the thermal zone, where different thermal conditions are provided based on its occupants without an internal wall. Various studies have used this model to evaluate the performance of model-based predictive control [29,30].

1.3. Research Gap and Objectives

Energy savings based on occupancy-based HVAC operations are expected to be reasonable. Various techniques based on computer vision have been developed to identify the distribution of occupants in spaces. However, the saving potential of occupancy-based prioritized zone control has not been investigated in the context of vision-based occupancy density quantification. Moreover, this study was not limited to simulation cases; experiments for developing the occupancy detection algorithm and building model were conducted to improve the credibility of the study.

The objective of this study was to evaluate the energy-saving potential of occupancy-based prioritized zone control based on occupant detection using cameras and artificial intelligence. The detection algorithm was developed in Python with YOLO based on an experiment using a camera in a campus classroom. In addition, a building model with a grey-box model structure with two split zones was developed based on an experiment using a four-way ceiling air conditioner. Subsequently, a simulation study was conducted to quantify the improved energy-saving potential of the occupancy-based HVAC operation.

The remainder of this paper is organized as follows. Section 2 describes the methodology of the study, and Section 3 explains the occupancy detection algorithm. Section 4 presents the grey-box model development and a simulation case study, assuming that the detection algorithm is applied to typical schedules of occupants. Section 5 presents the conclusions, and Section 6 explains the limitations and discusses the study.

2. Methodology

The aim of this study was to investigate the energy-saving potential of occupancy-based control within indoor spaces by dividing conditioned zones. This objective was realized by reducing the unnecessary energy provided to unoccupied zones while implementing prioritized-zone control in areas with the highest occupant density. This concept is illustrated in Figure 1.

Figure 2 depicts the methodology used in this study. First, object recognition technology was used to estimate the distribution of occupants in open spaces. An algorithm was developed using YOLOv5 object recognition technology to detect the presence of indoor occupants and locate spaces within a partitioned area with the highest number of people. The presence or absence of occupants was determined, and priority control zones were designated. Second, the energy-saving potential and efficiency were quantified for individualized control within spaces. The experiment was conducted in a single open space (26.4 m²) divided into two subzones, where the four-way ceiling air conditioner was installed. This was achieved by providing cooling to each zone in turns. A grey-box building model was established based on the experimental data, and a simulation case study was conducted with an occupancy schedule to demonstrate occupancy-based prioritized-zone control.

3. Occupancy Detection

3.1. Algorithm Development Subsection

The open-source YOLOv5 framework was used in this study [31]. The existing code was customized in a Python environment to determine the number of occupants while filtering out other objects in the divided thermal zones. First, the video data were loaded, and the thermal zones were divided based on predefined axis information. In this study, it is divided into four thermal zones (A, B, C, and D, as shown in Figure 3) [32]. Each zone was set as a confined area using the polygon function. YOLOv5 was run to detect the occupants and set the bounding rectangles only for the person. A circular object was added at the bottom of the rectangle to locate the exact location of the occupants in each zone. This process was repeated for a specific period (1 min) to constantly monitor occupants’ distribution.

3.2. Actual Demonstration of Developed Algorithm

This algorithm was demonstrated through real-time experiments. The classroom was set up with a camera connected to a laptop. Six occupants entered the classroom and moved, whereas YOLOv5 identified the number of occupants (the red rectangles in Figure 3) in each zone and exported the zone with the maximum number of occupants. The “turn off” was returned when the zone was unoccupied, which indicated that the air conditioner turned off.

The experiment was conducted for 20 min with four occupants. Five different occupancy distributions were set: four occupants entered the classroom and stayed for 4 min, and then moved randomly and stayed for another 4 min. This procedure was repeated three times. The detection performance was as high as 90%, which is reasonable. Some mispredictions were observed because the zone division configuration and desk and chair arrangements were not synchronized. Moreover, when the occupants were not fully visible through the camera (for example, standing at the edge of the zone), the detection rate decreased.

4. Energy Simulation Case Study

4.1. Feasibility Experiment Occupancy Detection

Preliminary experiments were conducted to validate the feasibility of the individual control of a four-way air conditioner in a single zone. The target zone was an individual office space (26.4 m²) on the 11th floor of a campus building conditioned using a four-way air conditioner that was installed on the ceiling (Figure 4). The single open space was divided into two zones (Zones 1 and 2) without partitions. The experiment was conducted using four temperature sensors installed in two zones (at a desk height of 0.6 m) and two supply diffusers, as shown in Figure 4. The four-directional blades of the existing air-conditioning system were sealed with tape (retrotec). This sealing tape is one of the decent products used for typical research and practice for the building infiltration domain such as blower doors test. During the control periods for Zones 1 and 2, only the blade in the corresponding direction was opened for measurements.

The experiments were conducted from 17 May to 2 June 2023. This is shown for three separate periods in Figure 5. The air conditioner was operated in only one zone in turns, and the average temperature difference between the two zones was quantified. The results revealed temperature differences of 0.9–1.7 °C during the three periods. The temperature in the uncontrolled area decreased because of air movement following the operation of the control zone. However, the temperature differences were maintained with a constant deviation, demonstrating the feasibility of individual control of the air conditioner in open spaces.

4.2. Grey-Box Building Modeling

Grey-box modeling was conducted after the feasibility test. Approximately 27 days (16 June–10 July 2023) were selected, and the air conditioner was separately operated for each zone in turns, as performed in the feasibility test.

The grey-box model consisted of three states for the envelope, zone 1 air and zone 2 air, as shown in Figure 6. The heat balance equations for each state are expressed by Equations (1)–(3), as follows:

$${Cp}_{env}\cdot {\dot{T}}_{env}={\displaystyle \frac{T_{out}-T_{env}}{R_{oe}}}+{\displaystyle \frac{T_{z1}-T_{env}}{R_{ez1}}}+{\displaystyle \frac{T_{z2}-T_{env}}{R_{ez2}}}+{\alpha }_{sol.env}·Q_{sol}+0.7·Q_{IH}$$

(1)

$${Cp}_{z1}\cdot {\dot{T}}_{z1}={\displaystyle \frac{T_{env}-T_{z1}}{R_{ez1}}}+{\displaystyle \frac{T_{z2}-T_{z1}}{R_{zz}}}+0.9·{\alpha }_{sol.room}·Q_{sol}+0.15·Q_{IH}+Q_{z1}$$

(2)

$${Cp}_{z2}\cdot {\dot{T}}_{z2}={\displaystyle \frac{T_{z1}-T_{z2}}{R_{zz}}}+{\displaystyle \frac{T_{cor}-T_{z2}}{R_{zc}}}+{\displaystyle \frac{T_{env}-T_{z2}}{R_{ez2}}}+0.1·{\alpha }_{sol.room}·Q_{sol}+0.15·Q_{IH}+Q_{z2}$$

(3)

Cp, T, and R represent the thermal capacity, temperature, and resistance, respectively. The subscripts of the capacity and temperature (state) were named based on the location. The subscripts of the resistance were named using the first heading of the neighboring states. It was assumed that 70% of the internal heat gain from the lighting and equipment would go to the envelope while 30% of those would go to two zones evenly (i.e., 15% each). Those were not estimated with optimization not to have many model parameters but roughly assumed based on ASHRAE fundamentals [33].

The heat balance equations were input into the matrices of the state-space formulation, which are typically used for control engineering, as follows.

$$\dot{x}=Ax+Bu$$

Here, x, u, A, and B are defined as follows.

$$\begin{array}{l}x=\left[\begin{array}{c}{\dot{T}}_{env}\\ {\dot{T}}_{z1}\\ {\dot{T}}_{z2}\end{array}\right],u=\left[\begin{array}{c}{T}_{out}\\ Q_{sol}\\ Q_{IH}\\ Q_{z1}\\ Q_{z2}\end{array}\right],A=\left[\begin{array}{ccc}-{\displaystyle \frac{1}{C{p}_e}}\left({\displaystyle \frac{1}{R_{oe}}}+{\displaystyle \frac{1}{R_{ez1}}}+{\displaystyle \frac{1}{R_{ez2}}}\right)& {\displaystyle \frac{1}{C{p}_e}}\left({\displaystyle \frac{1}{R_{ez1}}}\right)& {\displaystyle \frac{1}{C{p}_e}}\left({\displaystyle \frac{1}{R_{ez2}}}\right)\\ {\displaystyle \frac{1}{C{p}_{z1}}}\left({\displaystyle \frac{1}{R_{ez1}}}\right)& -{\displaystyle \frac{1}{C{p}_{z1}}}\left({\displaystyle \frac{1}{R_{ez1}}}+{\displaystyle \frac{1}{R_{rr}}}\right)& {\displaystyle \frac{1}{C{p}_{z1}}}\left({\displaystyle \frac{1}{R_{rr}}}\right)\\ {\displaystyle \frac{1}{C{p}_{z2}}}\left({\displaystyle \frac{1}{R_{ez2}}}\right)& {\displaystyle \frac{1}{C{p}_{z2}}}\left({\displaystyle \frac{1}{R_{rr}}}\right)& -{\displaystyle \frac{1}{C{p}_{z2}}}\left({\displaystyle \frac{1}{R_{ez2}}}+{\displaystyle \frac{1}{R_{rr}}}+{\displaystyle \frac{1}{R_{rc}}}\right)\end{array}\right],\\ B=\left[\begin{array}{ccccc}{\displaystyle \frac{1}{C{p}_e}}\left({\displaystyle \frac{1}{R_{oe}}}\right)& {\displaystyle \frac{1}{C{p}_e}}{\alpha }_{sol,env}& {\displaystyle \frac{1}{C{p}_e}}0.7& & \\ & {\displaystyle \frac{1}{C{p}_{z1}}}0.9\cdot {\alpha }_{sol,r}& {\displaystyle \frac{1}{C{p}_{z1}}}0.15& & {\displaystyle \frac{1}{C{p}_{z1}}}\\ & {\displaystyle \frac{1}{C{p}_{z2}}}0.1\cdot {\alpha }_{sol,r}& {\displaystyle \frac{1}{C{p}_{z1}}}0.15& {\displaystyle \frac{1}{C{p}_{z2}}}\left({\displaystyle \frac{1}{R_{rc}}}\right)& {\displaystyle \frac{1}{C{p}_{z1}}}\end{array}\right]\end{array}$$

This was then discretized for the time step, which was 1 min in this study.

Nonlinear optimization was performed in MATLAB using the Fmincon function to identify the model parameters (Cp_e, Cp_z1, Cp_z2, R_oe, R_ez1, R_ez2, R_rr, R_rc, a_sol,env, a_sol,r). The objective function was the root-mean-square error (RMSE) between the measured and simulated room-air temperatures of both zones during the total modeling period, as follows.

$$J={\displaystyle \sum _{i=1}^2{\displaystyle \sum _{k=1}^n\sqrt{\frac{{\left(T_{z(i)}^k-{\widehat{T}}_{z(i)}^k\right)}^2}{n}}}}$$

The initial values were set based on the results of a previous study [34], which had only one state for zone air and one state on the envelope, as adopted in this study. The initial resistance between the two thermal zones was set based on the horizontal heat transfer between the two air nodes. The convective heat-transfer coefficient for two air nodes was set as 12.12 W/m² °C.

The modeling results are shown in Figure 7. The zone temperature comparison between the measurement and simulation with control signal (cooling rate) and disturbance inputs (solar radiation and outdoor air temperature) are illustrated. The RMSE of the two zone temperatures were 0.89 and 0.90 °C, respectively. No guidelines are available for evaluating the prediction performance of the grey-box model; however, this error range is acceptable for evaluating the simulation case, as some previous studies obtained values of 0.5–1.0 °C and utilized for actual implementation in real buildings [29,30].

4.3. Simulation Case Study

4.3.1. Occupancy Density Scenarios

A simulation case study was conducted in a real-world context to assess the practical efficiency of the individual controls. The target space was a small café, and a surveillance camera was installed in the café without raising privacy concerns. Surveillance footage can be used to acquire data on occupant distribution without compromising privacy during actual implementation. The zone of 26.45 m² was assumed to be divided into two separate zones, where the maximum occupancy of four individuals could reside, that is, tables for two people are located in each zone. Occupancy distributions for the four scenarios are listed in

Table 1. Scenarios of occupancy distribution and zone control method.

Scenario			8 am–10 am	10 am–12 pm	12 pm–2 pm	2 pm–4 pm	4 pm–6 pm	6 pm–8 pm
1	Occupancy	Zone 1	2	0	2	0	2	0
		Zone 2	2	0	2	0	2	0
	Zone control		Global zone control
2	Occupancy	Zone 1	2	0	2	0	2	0
		Zone 2	2	0	2	0	2	0
	Zone control		Occupancy-based on/off control
3	Occupancy	Zone 1	2	0	2	0	2	0
		Zone 2	0	2	0	2	0	2
	Zone control		Global zone control
4	Occupancy	Zone 1	2	0	2	0	2	0
		Zone 2	0	2	0	2	0	2
	Zone control		Occupancy-based zone control

. Scenarios 1 and 3 were the baseline cases, whereas scenarios 2 and 4 were the proposed cases. In Scenario 1, half of the occupied hours were occupied, whereas the entire zone was conditioned. The same occupants and occupancy distributions in Scenario 1 were applied to Scenario 2; however, the zone in Scenario 2 was controlled based on occupancy (the air conditioner was put off when unoccupied). In Scenario 3, two occupants resided in one zone by taking turns while the entire zone was conditioned. The same occupants and occupancy distributions in Scenario 3 were applied to Scenario 4; however, the zone in Scenario 4 was controlled according to occupancy (the air conditioner was put on only for the occupied zone). A comparison between Scenarios 1 and 2 showed the impact of occupancy-based on/off control; for example, the air conditioner was completely off when the entire zone was unoccupied. A comparison between Scenarios 3 and 4 indicated the influence of occupancy-based zone control; for example, the blade of the air conditioner was open only in the occupied zone.

The scenarios are summarized as follows.

• Scenario 1: global control, half-occupancy (fully occupied and unoccupied)
• Scenario 2: occupancy-based control, half-occupancy (fully occupied and unoccupied)
• Scenario 3: global control, half-occupancy (half-occupied)
• Scenario 4: occupancy-based control, half-occupancy (half-occupied)

4.3.2. Simulation Results

The scenarios were applied to building energy simulations using a grey-box model in a Matlab environment. The two thermal zones were conditioned using feedback control strategies. This study applied only the P gain in the control loop with a time step of 1 min. As the discretized building model had a time step of 1 min, real-time feedback control, as implemented in actual buildings, was unavailable. The setpoint temperature of the conditioned zone was set to 23.5 °C. The initial temperatures of all the states (two air temperatures and envelope temperature) were set to 23 °C. The outdoor air temperature and solar radiation from the actual measured data in 2021 summer (August 14th) were used, as shown in Figure 8.

Figure 9 shows the temperature trajectories and cooling rates for all scenarios. In Scenario 1, the air temperatures in the two zones were maintained under the setpoint, and cooling was provided for the entire period (8 am–8 pm), regardless of occupancy. In Scenario 2, the air conditioner was put on only during the occupied period. The energy consumption in Scenarios 2 was 20.1 kWh whereas that of Scenario 1 was 25.7 kWh. The savings potential of this occupancy-based on/off control was approximately 22%. In Scenario 3, the air temperatures in the two zones were maintained under the setpoint, and cooling was provided for the entire period (8 am–8 pm), regardless of occupancy. In Scenario 4, the air conditioner was put on only during the occupied period. The energy consumption of Scenario 4 was 20.3 kWh, whereas that of Scenario 3 was 25.7 kWh. The savings potential of this occupancy-based zone control was approximately 21%, similar to the previous comparison between Scenarios 1 and 2. When half of the entire area was occupied (Scenarios 1 and 3), the savings potential compared with the global zone control was approximately 21–22%. This could have been 50% if the two zones were fully separated by an internal wall. This limited percentage in savings is attributed to convective air movement between the two zones, which is inevitable in open spaces. However, this range (21–22%) is not marginal and implies the feasibility of energy savings with occupancy-based control.

In Scenarios 2 and 4, comfort violations were observed when conditioning was started in one zone, as indicated in Figure 9. Figure 10 shows the frequency distributions of this violation for Scenarios 2 and 4. The vertical axis represents the number of violations calculated for 1 min. Approximately, a quarter of the total violations was smaller than 0.1 °C for both scenarios (23% and 26% for Scenarios 2 and 4, respectively). The corresponding portions of violations lower than 0.5 °C for both scenarios were 59% and 56%. The violations higher than 1 °C were 19% for both scenarios. The maximum violation was 2.1 °C for Scenario 2, but only for one count, which indicated that this occurred only for 1 min. Moreover, for all periods, the temperature was in the comfort bound of ISO 7730 (less than 25.5 °C in class A, which is the strictest case) [35,36]. When evaluated with ASHRAE standard 55, the temperature was also in comfort bound [37] assuming the air temperature is identical to operative temperature.

These violations are inevitable if the aim is to save energy by applying occupancy-based control. A typical HVAC system requires some time to reach the desired comfort range. Some levels of predictive control, such as model-based predictive control, can be applied to significantly minimize these violations to precondition the zones based on an appropriate prediction of the thermal dynamics of the building and the occupancy schedule. A recent article showed this strategy via a simulation case study based on the grey-box model, which was constructed with experiments in actual buildings [38]. In this study, a perfect balance between comfort and energy consumption was shown.

5. Conclusions

This study demonstrated the feasibility of occupancy-based building control in terms of cooling energy savings and comfort. An occupancy detection algorithm was developed using YOLO based on experiments with actual occupants and recorded video data. Feasibility experiments were conducted to assess the savings potential of the proposed occupancy-based controls by evaluating the temperature deviations of the two zones when only one zone was conditioned by an air conditioner. Next, a grey-box model was constructed based on experimental data that existed in two states with two separate thermal zones in open spaces. This building thermal model was applied to a simulation case study, assuming that occupancy detection was perfectly achieved using the camera. In the synthetic occupancy scenarios, it was assumed that the café was a target zone where the surveillance camera was used; hence, no privacy issues were raised. The main findings of this study are summarized as follows.

• An occupancy detection algorithm for the designated zone was developed and verified experimentally using YOLOv5 object recognition technology in the open spaces of an actual building.
• Meaningful temperature deviations (0.9–1.7 °C) between two thermal zones in open spaces were observed when only one zone was conditioned, demonstrating the feasibility of occupancy-based zone control.
• The grey-box building model was constructed and validated using experimental data from the actual building. The satisfactory prediction performance was shown; for example, an RMSE of 0.9 °C.
• A simulated case study was conducted using an estimated grey-box model with a synthetic occupancy scenario. Approximately 21–22% of the energy-saving potential was confirmed when half of the occupancy was assumed, compared to the existing room-wide global control. Inevitable comfort violation was observed, with more than half of the violations being lower than 0.5 °C.

6. Limitations and Discussion

6.1. Feasibility of Zone Control

In this study, an experiment was conducted using four temperature sensors: one for room air and another for supply air in each zone. The modeling approach used in this study was the grey-box method that assumes the state represents the temperature of the specific part of the building—for example, two states for zone air and another two states for the envelope were set. This means the target zone is explained with only four temperatures, and this is a simplified approach compared to computational fluid dynamics (CFD) simulations. Moreover, when only one zone was conditioned, the granularity of the temperatures in the two zones might not be sufficiently detailed at this sole measurement point. Therefore, future studies should be conducted using more sensor measurements to verify the feasibility of occupancy-based zone control. Furthermore, CFD simulation can be performed to investigate the detailed temperature distribution in the zones. This can be analyzed along with grey-box modeling results to ensure the credibility of the grey-box model, which is simpler than building energy simulations and CFD.

6.2. Privacy Concerns

Potential privacy concerns may arise regarding occupancy-based zone control. Therefore, this study may be limited to zones in which surveillance cameras have already been installed. However, surveillance cameras are deployed for safety and crime prevention in most commercial buildings, such as retail stores and restaurants. Regarding the zone where privacy should be preserved, other engineering methods for detecting occupancy, such as PIR sensors, should be applied.

6.3. Cost Issues

The potential increase in the cost of occupancy-based zone control is negligible. The minimum requirement is a temperature sensor or thermostat for the corresponding number of individual zones. Additional costs might be incurred to configure the control loop between the sensors and AC. This may be embedded in the air conditioner in the form of a microcontroller. Potential future work might include the recent IoT with low-cost devices such as Arduino as shown in a recent study [39]

6.4. Generalization and Real-World Implication

The supply diffuser may be controlled individually to provide cooling for the occupied zone of the small zone, as in the testbed of this study. To realize this, an additional communication module needs to be implemented on the HVAC side not to mention the devices for occupancy detection. For larger open spaces where multiple indoor units are installed, a control signal should be allocated for each indoor unit. Specifically, some indoor units are controlled according to the occupancy while the rest of those remain off. In addition to air conditioners with indoor and outdoor units, the proposed method can be applied to separate HVAC systems, such as standing fans for spaces or occupants. In this case, a more detailed analysis of the comfort model such as predicted mean vote needs to be analyzed.