Application of Artificial Intelligence in Air Conditioning Systems

Abstract

Urbanization has led to a sharp rise in the demand for power over the past 10 years, alarmingly rising greenhouse gas (GHG) emissions. HVAC (heating, ventilation, and air conditioning) systems account for nearly half of the energy used by buildings, and minimizing the energy use of the HVAC systems is essential. However, the common problems, such as hot spots and cold spots in office spaces, experienced in the building need to be addressed. Therefore, this chapter introduces the application of artificial intelligence proactive control to resolve typical office issues. A demonstration testbed was implemented on the Singapore Institute of Technology (SIT) campus. The experiments were conducted in baseline mode and smart mode. In the case study, two big zones were segregated into 43 micro-zones equipped with smart dampers at each diffuser, allowing a localized set point to improve thermal comfort and eliminate hot and cold spots. It has been observed that the proactive AI control reduces cooling provided to the office by 29 percent and AHU electricity usage by 50 percent, respectively, while keeping the area within thermal comfort range of 23 to 25°C and 50 to 63% relative humidity.

1. Introduction

Energy is the most important component for the operation of various sectors, including transportation, business, residential buildings, and many others. Recent technological developments have led to a sharp rise in global energy consumption, which is alarmingly increasing the rate of greenhouse gas emissions. As shown in Figure 1, the world energy consumption by different sources of fuels was about 173,340 Terra-Watt-Hr (TWh) in 2019, while it was 122,073 TWh in 2000. The world’s energy consumption increased by approximately 42% within 19 years. Electricity is the prime energy source that the built environment utilizes. Global electricity generation in 2021 increases approximately twofold compared to 2000 to accommodate the drastic increase in energy consumption in the built environment, as indicated in Figure 2. Primary fuel sources, like coal and gas, account for almost 60% of total primary energy sources, whereas renewable energy makes up only 13% of total primary energy sources. IEA reported that the increase in coal-fired power plants contributes to a sharp rise in carbon dioxide emissions. The electricity demand continues to grow by 4% in 2022. Despite substantial expansions of renewable energy usage, it is anticipated to offset the rise only partially in electricity consumption [2]. Due to the rise of greenhouse gas emissions, the environment is seriously threatened by the continued growth of energy consumption. Authorities from many countries, however, are focused on achieving net-zero carbon emissions and a major increase in the production of renewable and clean energy for end consumers. Figure 3 shows that the total generated capacity will be 38,900 GW in 2050, while the expected rise in electricity output will be roughly 88,000 TWh. Additionally, it is anticipated that implementing the carbon tax will significantly reduce carbon emissions starting in 2025 [3]. In order to achieve net-zero carbon buildings, energy efficiency upgrades made to existing structures and energy-efficient designs for new buildings, including passive and active technology, will be crucial.

Figure 1.

Energy consumption by different fuel sources since 2000 [1].

Figure 2.

Global electricity generation by sources from 2000 to 2021 [1].

Figure 3.

Projected electricity statistics and carbon emissions till 2050 [3].

The built environment is seriously threatened by overpopulation and rapid urbanization. By 2050, the world’s population is expected to reach 9.6 billion, a 21 percent increase from the current number. Therefore, the energy demand, particularly electricity for the built environment, will rise dramatically unless energy-saving options and measures are implemented. Moreover, 59% of the world population, as shown in Figure 4, resided in highly urbanized regions in 2020 because these regions have employment opportunities, living standards, and ease of commute. After 2007, the proportion of urban residents overtook rural residents, sharply increasing the need for cooling and heating systems in residential and commercial structures. Urbanization significantly increased ambient temperature and decreased cooling system effectiveness due to the heat island effect. According to the IEA, two-thirds of homes may have air conditioning units [5]. By 2100, the average worldwide temperature could rise by 4°C due to the sharply increasing trend in the deployment of air conditioning systems in urban areas. Therefore, there is an urgent need to implement smart and energy-efficient air conditioning systems, including both passive and active cooling systems, for existing and new buildings. Doing so will lead to achieving net-zero carbon buildings.

Figure 4.

World’s Population residing in urban and rural areas [4].

Digitalization is a crucial component of the movement toward intelligent and energy-efficient solutions that are required to reach the targets of net-zero carbon emissions. Digitalization enables numerous energy systems to be more interconnected, intelligent, dependable, sustainable, and efficient. Digitalization could reduce energy consumption in buildings by around 10% by using real-time data to increase operational effectiveness. The installation of smart thermostats can also better predict heating and cooling requirements by employing self-learning algorithms, and real-time weather forecasts to predict occupant behavior.

2. Application of artificial intelligence in air conditioning systems

Machine learning (ML), a subset of artificial intelligence, widely applies to various sectors. The development of instrumentation and sensors has led to a significant increase in the amount of data collected per minute. Plotting and analyzing these data is crucial to turn them into insightful information that can be used for planning, operations, and forecasting. Machine learning techniques provide the link between the input parameters and the predicted output variables. Machine learning can be generally categorized into two groups, namely (i) supervised learning and (ii) unsupervised learning.

By deploying the appropriate methods, ML can be applied to the followings:

1. Detecting the sale trends

2. Time series forecasting

3. Multivariate time series forecasting with recurrent neural networks (NNs)

4. Detecting financial fraud using decision trees

5. Convolutional neural networks implementation for car classification

Globally, many countries are embracing digitization, which will help businesses increase productivity, lower operating costs, and improve safety. Additionally, researchers and industry participants can create machine learning and artificial intelligence algorithms using historical data because it is easier to acquire thanks to digitization. However, even though machine learning is developing quickly, there have been difficulties using it in practical applications because it needs a vast amount of data. However, the road to digitization greatly aided the oil and gas industries’ ability to access data, leading to machine learning easily. The built environment has been the focus of extensive research into intelligent control for air conditioning systems since 2000 to increase the effectiveness of these systems. Artificial intelligence applications in the HVAC sectors are made possible by digitalization, which is essential. Therefore, HVAC firms may create smarter systems to make buildings more environmentally friendly, thanks to technological improvements. Artificial neural networks (ANNs) have also been used in HVAC systems to optimize the operation set points of the air conditioning system.

2.1 Review of the application of machine learning (ML) and artificial intelligence (AI) in air conditioning systems

Many researchers have been working on machine learning and artificial intelligence for both the demand and supply side of HVAC systems. A vast majority of research conducted in the last 10 years can be generally categorized into (i) prediction of occupancy and their behavior, energy consumption, and energy management and (ii) control and optimization of HVAC systems.

Aftab et al. designed and implemented a sophisticated occupancy-predictive control system with the aid of recent development in embedded system technologies [6]. The system is cost-effective, has fewer requirements for powerful processors to execute highly sophisticated tasks, and deploys real-time occupancy recognition using video processing and ML techniques. The model can predict the occupancy pattern and allow to control of HVAC systems using real-time building thermal response simulations, achieving significant energy savings. Reeba et al. developed a model that can determine the occupants’ behavior, which generally results in the wastage of energy in the operation of HVAC systems [7]. An ML-based model focused on the space’s heat flow and could capture the energy waste depending on the status of the space, such as occupied or non-occupied. The model could predict the optimal temperature settings utilizing the status of the space, along with predicted mean vote (PMV) and the deployment of motion sensors. The author observed that about 50% of the total energy was wasted due to the suboptimal temperature settings in the space. Esrafilian-Najafabadi also analyzed the impact of different occupancy prediction models using ML techniques [8]. Four different ML techniques, namely decision trees, k-nearest neighbor (KNN), multilayer perceptron, and gated recurrent units, were deployed to predict the occupancy types and patterns and provide an accurate and reliable evaluation of the performance of the occupancy model for coupling with HVAC control systems. The author studied different models that analyze the occupants’ energy savings and thermal comfort. The study included thermal comfort favored mode and energy savings priority mode. Despite having a trade-off between the occupants’ energy savings and thermal comfort, the author observed that equally weighted energy savings and thermal comfort provide the best performance and that the KNN technique outperformed other machine learning techniques. Although numerous studies related to ML techniques that account for occupant patterns and behavior have been conducted, there is a lack of study on effective air distribution due to the dynamics of occupant patterns and their impact on temperature profiles across a spacious open office.

Many researchers emphasize their research on predicting energy consumption and optimizing energy usage by HVAC systems, the most energy-intensive system, utilizing supervised learning methods. For example, Liu et al. applied Deep Deterministic Policy Gradient (DDPG) for short-term energy consumption of HVAC systems [9]. The authors deployed a powerful autoencoder (AE) to process the raw data linked up with the DDPG method to attain high-level space state data for optimizing the prediction model. In this study, the authors set up a ground source heat pump system (GSHP) to supply a small office’s cooling and heating needs. The operation data were used to train the model, and the authors demonstrated the office’s energy consumption verification. The authors also verified that the proposed model predicted the state space variables more accurately than the common supervised learning models, such as support vector machine (SVM) and neural network (NN). The rapid expansion of deep learning techniques has made them promising alternatives to conventional data-driven methods. Vazquez-Canteli et al. developed an integrated simulation environment that links the building energy simulators and TensorFlow, which allows the implementation of various advanced machine learning algorithms [10]. This development enables many researchers to test and formulate optimized control algorithms to accommodate potential energy savings in buildings. The simulation platform also can be easily scaled up to the district or city level to study model-free algorithms and their impact on energy consumption and control strategy. Despite many interesting applications of ML and AI in HVAC systems being conducted, some research focused on energy consumption while other emphasized thermal comfort for small offices. There is a gap to close the loop between energy consumption while maintaining the thermal comfort, along with optimized cooling load predictions. In addition, most of the algorithms operate offline and cannot account for the heat loads in space’s extremely dynamic nature and external parameters such as weather conditions. In order to incorporate artificial intelligence focused control that enables online load forecasting for extremely dynamic environments, this work is motivated by the desire to investigate the performance of HVAC systems, particularly airside systems.

3. Case study: deploying AI solution to airside energy efficiency improvement

3.1 Energy consumption in air conditioning systems

As illustrated in Figure 3, by 2050, worldwide power consumption is expected to have doubled from what it is today. Although it is questionable whether the sharp rise in power use is related to the sharp rise in cooling and heating requirements, the fact that there are currently 1.9 billion air conditioning units worldwide serves as proof. Additionally, it is anticipated that by 2050, cooling and heating requirements will have increased by 37%. Therefore, the road map for net-zero carbon buildings requires immediate effort to increase the efficiency of the air conditioning systems and the occupants’ behavior, incorporate cutting-edge control systems, and embrace passive technologies. This project aims to increase air conditioning system’s efficiency by integrating them with AI-focused self-learning control systems.

A case study was carried out at one of the spacious offices at the Singapore Institute of Technology (SIT) to apply AI to air conditioning systems. Due to its tropical climate, space cooling is required throughout the year, and the building sector accounts for 37% of total energy consumption. Figure 5 depicts the energy consumption of the air conditioning system, which is as high as 50% of building energy consumption due to its hot and humid climate. A detailed breakdown of the energy consumption of HVAC systems is shown in Figure 5, and airside accounted for 34% of the total energy consumption of HVAC systems. Although the airside energy consumption is equally important compared to the waterside, it is mostly overlooked due to the high dynamics in nature.

Figure 5.

Breakdown of energy consumption of buildings in Singapore [11].

3.2 Motivation of the case study

In chiller plants, airside systems account for the second highest energy consumption. In addition, the airside cannot support more control flexibility due to the high dynamics involved. Another thing to consider is that uneven thermal heat maps caused by oversized and undersized air distribution systems deviate from the thermal comfort of the occupants. Figure 6 shows that hot and cold spots in large open offices are prevalent issues in airside systems.

Figure 6.

Uneven thermal heat map due to improper sizing of AHUs and control strategies.

Although there are other potential contributing elements, ineffective air distribution systems are the main problem. Ineffective air distribution systems cause hot areas because of insufficient cold air provided to the space. In addition, cold spots develop in the remaining areas of the office because there is excessive cold air. The control system, however, is unable to respond appropriately. Conventional control systems operate in reactive methods, which is the cause of the control system’s slow response. As indicated in Figure 7, although the main return air temperature is employed as feedback to serve as the control, it does not accurately represent the local zones, leading to unequal temperature distributions. The zoning of the space is one aspect that shows a significant role. The zone size is too large for the control systems to capture all information; thus, they are unable to adapt. Therefore, in this study, the impact of the big zone being segregated as smaller zones (micro-zones), proactive AI control on the performance of the airside system, and energy savings potential will be investigated.

Figure 7.

Typical reactive feedback control system in air handling units (AHUs).

3.3 Details of the testbed located at the Singapore institute of technology

The pilot tests were conducted on one floor of the Singapore Institute of Technology (SIT) campus located at Dover Drive. The testbed occupying 11,000 square feet is located at level two, comprising open offices, meeting rooms, a pantry, an AHU room, and washrooms. The space is fully air-conditioned except for the washrooms. The details of the testbed are tabulated in Table 1. Figure 8 illustrates the layout of the office located at the Singapore Institute of Technology, and the spaces are segregated into two zones, namely, block B and block C. While block B’s cooling requirements are supplied by air handling unit 2-1 (AHU 2-1), block C is served by AHU 2-2. The temperature set point of the space was 24°C throughout the day.

Total floor area	11,000 square feet (sqft)
Seating capacity	Approx. 100
Operation hours	0830 hrs to 1800 hrs
Area types	Enclosed workstations, cabins, cafeteria, and conference rooms
BMS	Yes, Johnson controls
Chilled water actuator	Yes, installed for each AHU

Table 1.

Details of the pilot in the case study.

Figure 8.

Plan view of the office space at level two of the University Service Centre at the Singapore Institute of Technology.

The key issue with the air conditioning system is the thermal comfort of the occupants stationed in the space. From the occupant’s feedback, it is discovered that there are areas with hotspots and overcooling within the office. On occasion, occupants feel uncomfortably hot or extremely cold in the office. It is observed that some of the diffusers are covered with masking tape to restrict airflow. The AHU VFD and actuator set points are changed manually based on complaints from the occupants. In addition, due to the work nature of the academic staff, they are frequently required to leave and return to their desks for lectures and classes, resulting in a dynamic heat map. Therefore, there is a need to resolve the issue without compromising the energy efficiency of the air conditioning system. The primary objective of this study is to develop an intelligent solution to resolve thermal comfort issues without compromising energy efficiency while eliminating the conventional reactive approach to control systems.

The proactive solution would account for the varying occupant numbers throughout the day while creating an optimal condition for their staff. Despite the abundant availability of smart sensors, which work on room levels, an AI algorithm was developed and tested at SIT staff office, along with the collaboration between SIT and Singapore Digital Pte. Ltd., a sole distributor of 75F smart innovation solutions in Singapore.

Dynamic air balancing and chilled water balancing, along with proactive AI predictive control, are the essential components of this study to achieve energy savings while maintaining the thermal comfort of the occupants. In order to optimize the air distribution efficiency, two big spaces, as indicated in Figure 8, are divided into 43 micro-zones, as indicated in Figure 9. Each meeting room is treated as a micro-zone. Each micro-zone in the open office is equipped with an IoT smart sensor that measures the key parameters, such as temperature, relative humidity, CO₂ concentration, and occupancy status, using a passive infrared sensor (PIR), enabling an accurate representation of the local heat load. Moreover, as illustrated in Figure 10, smart dampers are retrofitted between each supply air diffuser/VAV duct and flexible air duct to modulate the amount of airflow based on the actual heat load. This facilitates micro-zonal control, allowing better comfort and energy savings. In addition, the opening of the smart damper is controlled based on the local heat load. The IoT smart sensors and dampers are wirelessly connected smart nodes which communicate wirelessly with central control units (CCUs). A cloud-based proactive AI control powers the algorithm behind the control units, and the architecture of the proactive AI control system is shown in Figure 10. The CCU sends minute-by-minute data regarding temperatures in various building parts to cloud servers. Every night, these servers run proprietary algorithms to crunch the historical data and develop a thermal model of the building. They then predict the thermal load in each part of the building for the next day based on the forecasted weather.

Figure 9.

Two zones are split into 43 micro-zones.

Figure 10.

Architecture of proactive AI control system.

3.3.1 Dynamic airflow balancing

Figure 11 illustrates the dynamic airflow balancing, which optimizes the cold air supply to the most required space. The smart dampers’ opening at micro-zones with cold spots is adjusted to accommodate the cooling needs in that micro-zone. Due to the changes in the opening of the smart dampers, the static pressure in the duct increases. However, cold air is circulated to space (hotspots), which requires more cooling, restoring the static pressure. Therefore, supply air fan speed is not ramped up to supply more cooling to the hot space; instead, air balancing between cold and hotspots progresses, resulting in energy savings in AHUs. This means that the AI control is able to identify which zones require more cooling by deploying dynamics zone priority (DP). Since the system enables minute-by-minute data collection, real-time DP is performed prior to executing the next control phase. Air balance is performed using a weighted average of the local heat load, as shown in Eq. (1) (Figure 12).

Figure 11.

Illustration of dynamics of airflow balancing.

Figure 12.

Dynamic airflow balancing diagram.

$${\dot{Q}}_{\mathit{\text{weighted}}}=\frac{{\sum }_{i=z_1}^{z_n}\left({\dot{Q}}_{i,j}{\mathit{xDP}}_i-{\dot{Q}}_{i,k}{\mathit{xDP}}_i\right)}{{\sum }_{i=z_1}^{z_n}{\mathit{DP}}_i}$$E1

where ${\dot{Q}}_{\mathit{\text{weighted}}}$ and $\dot{Q}$ denote the weighted average of heat load and local heat load in the space, respectively, DP refers to dynamics zone priority, i denote the number of zones in the space, j represents the zones with overcooling, and k is for the zones requiring more cooling. Then, air balancing for the micro-zones is carried out based on the DP value, which identifies how far the current temperature is away from the set point. The AI algorithm identifies and optimizes the air balancing, resulting in the evenly distributed cold air supply to each of the micro-zones, and AHUs can still be operated at a lower speed as compared to the conventional control system because the speed of the AHUs is adjusted, as shown in Figure 13, based on the weighted average of micro-zones after air balancing is carried out. In addition, fresh air optimization is enabled by incorporating a modulating damper in the fresh air duct. The bandwidth of the opening of the fresh air damper ranges from 20 to 100% based on CO₂ concentration in the space, enabling minimal fresh air usage when the indoor CO₂ level is about 900 ppm.

Figure 13.

Block diagram that shows smart damper and AHU VFD relational control.

3.3.2 Dynamic chilled water balancing

Chilled water balancing is achieved by utilizing the micro-zones’ weighted average return air temperature. When the weighted return air temperature falls above the set point, the Al algorithm detects that more cooling is required in the space. However, the steps for air balancing and AHUs speed adjustment are completed to accommodate the cooling requirement. Therefore, the controlled valve will be modulated to a wider position to provide more chilled water to maintain the temperature in micro-zones within the thermal comfort range defined by ASHRAE standard 55 [12]. Therefore, the differential temperature of the chilled water is maintained at the optimal range, while the chiller water pump’s speed is adjusted to provide the required cooling in the space, resulting in energy savings without compromising the thermal comfort of the occupants in the space. The sequence of activating the opening of the chilled water modulating valve is shown in Figure 14.

Figure 14.

Dynamic chilled water balancing control process.

3.4 Deployment of AI solution in airside system of the chiller plant located at SIT

Implementation of measurement of the performance of AI-oriented proactive control solution comprises the following stages:

1. Retrofitting of smart dampers to the existing system that enables dynamics air balancing

2. Installation of a power meter at each AHU to measure the power consumption of the fans

3. Installation of CCU for each AHU to control smart nodes and smart dampers, modulate VFD and chilled water actuator, and act as a cloud gateway

4. Installation of smart dampers at the existing mixing boxes outlet. The opening and closing of the smart dampers are controlled by the smart nodes installed above the false ceiling. The smart nodes communicate wirelessly to the cloud; users can access all data and control through the App or portal.

5. Installation of Intelligent Temperature Mote (ITM) for each zone across the pilot area to sense and collect data (temperature, humidity, and Lux) in real-time every minute, and there are a total of 43 zones, as shown in Figure 15.

6. Installation of chilled water flow meter, the temperature sensors for chilled water return, and supply.

7. Installation of a chilled water actuator controlled by the CCU to modulate chilled water flow to match the optimal set point and ensure optimal flow rate and differential temperature through the chilled water pipe networks, as shown in Figure 16.

8. Installation of a new fresh air damper with Belimo actuator and the fresh air damper is modulated (20–100% opening) based on the CO₂ level, as shown in Figure 17.

Figure 15.

A schematic diagram of dynamic airflow balancing.

Figure 16.

A schematic diagram of dynamic chilled water balancing.

Figure 17.

A schematic diagram of uutside air optimization comfort range.

3.4.1 Baseline measurement and smart mode measurement deployment of AI solution in the airside system of the chiller plant located at SIT

After completing the installation of the required instrumentation, sensors, and IoT devices and commissioning, which includes fine-tuning the parameters, the testbed was operated in two phases: baseline mode and smart mode. Each mode was operated for 10 days, excluding weekends. The baseline mode represents the operation of existing conditions, isolating the proactive AI control, whereas the smart mode enables the proactive AI control, including dynamic air balancing, dynamic chilled water balancing, and fresh air optimization. The AI control overwrites the set point of BMS for existing operating conditions. Outdoor temperature and relative humidity were also recorded in the cloud during the testing of both phases to ensure that the impact of the weather conditions on the airside system’s performance was considered. During both testing phases, data are recorded every minute using the instrumentation and sensor installed during the retrofitting stage, as tabulated in Table 2. During weekdays, AHUs are scheduled to start at 6:30 am, and the chiller and pumps are staged to turn on from 7 progressively for pre-conditioning. Since the building operates from 8:30 am to 6:00 pm, the data analysis only includes this period of the day. Key parameters, such as the temperature and relative humidity of all 43 zones, were recorded every minute in both baseline and smart modes. While AHU 2-1 supplies the cooling requirements to zone 1–23, AHU2-2 serves zones 24–43. The set point for all spaces was maintained during the tests at 24°C.

Location	Parameters measured
Meeting rooms and office space	Temperature Relative humidity CO2 concentration Dynamics occupancy Supply air temperature from each diffuser Smart damper opening in %
AHUs	Fan power Fan speed Chilled water flow rate Chilled water supply and return temperature Supply air temperature Return air temperature Fresh air damper opening in % (20–100%)

Table 2.

List of measured parameters and locations of measurement.

Figure 18 depicts the temperature profiles of supply air measured during the baseline and smart mode tests. It is indicated that supply air temperature fluctuated between 13.6°C and 21.7°C, whereas it was maintained between 15.3°C and 19.3°C. The median temperature of supply air for baseline test and smart modes were 16.3°C and 17.1°C, respectively. Despite maintaining the close median supply air temperature between baseline mode and smart mode, the differential temperature in the interquartile for baseline mode was 1.5°C, and that of smart mode was about 1°C. It is also concluded from the box plot that most of the supply air temperature during the smart mode test fall outside the interquartile, and outliers are beyond 1.5 times the interquartile (upper whisker) due to inefficient air distribution systems and control strategy.

Figure 18.

Supply air temperature profiles during the baseline and smart test.

On the other hand, no outliers were discovered beyond the upper and lower whiskers during the smart mode testing. It is also worth noting that the temperature difference between the minimum and maximum supply air temperature was less than 4°C, assuring that smart mode control performs significantly better than baseline mode in terms of air distribution effectiveness. The space temperature with respect to time during baseline mode and smart mode is presented in Figure 19. During the test period of both modes, the set point temperature was maintained at 24°C, and the results were analyzed by comparing the baseline and smart mode tests. From the temperature and relative humidity profiles during the baseline test, it was observed that the space temperature during the smart mode test fluctuated from 21 to 25°C, while relative humidity in the space varied between 67% and 48%. Furthermore, the difference between space temperature and the set point was found to be considerably huge in some cases; it was as high as 3°C, resulting in cold spots and hotspots in space. However, during the smart mode test, the space temperature varied between 23 and 24.5°C, while relative humidity ranged from 52 to 65%, which falls well within the thermal.

Figure 19.

Average temperature and relative humidity profiles of the space during the baseline and smart mode tests.

In order to analyze further details of the temperature distribution in the space, a temperature bin is created with 2°C range with a total of 551,872 data points, and the results are illustrated in Figure 20. Seventeen percent of the data points that falls under undercooled regions (19–22°C) during the baseline test were shifted to 22–26°C when the smart mode was activated. Moreover, the smart mode delivered 99.97 percent of the events within the bin range of 20–23.9°C, highlighting that proactive AI control works perfectly fine to optimize the airside performance compared to the baseline mode. Therefore, proactive AI control not only achieves a better thermal comfort condition in the space but also improves the efficiency of the airside system, because AI control optimizes the cooling load prediction by adapting the characteristics and activity ongoing in the space, along with the dynamic airflow balancing strategy. Energy consumption should not be overlooked despite improving the thermal performance of the airside. Therefore, energy data, such as electricity consumption and cooling supplied to the building, were monitored and recorded throughout both baseline and smart modes. All energy data were recorded using the Kamstrup BTU (cooling energy) and the Schneider Energy Meter (electrical energy). Data during the weekends of the testing period were excluded from the analysis in both modes. Two AHUs (AHU 2-1 and AHU 2-2) were assigned to supply cooling to the space, and the rated power of AHU 2-1 and AHU 2-2 at the full load are 5.7 kW and 3.7 kW, respectively. During different test modes, weather conditions were normalized to ensure that the deviation in the weather conditions was not affected. The pairs of the daily average ambient temperatures during both modes for comparative analysis are presented in Figure 21.

Figure 20.

Temperature distributions with 2°C range.

Figure 21.

The pairs of average daily outdoor temperatures for the comparative analysis during the baseline and smart modes.

Daily electricity consumption of both AHU 2-1 and AHU 2-2 is illustrated in Figure 22. During the baseline test, it is observed that the daily electricity consumption of AHU 2-2 ranges between 44 kWh and 70.50 kWh, while the electricity consumption of AHU 2-1 varies between 22.6 kWh and 8 kWh. While conducting the test in the smart mode, as indicated in Figure 22, electricity consumption of AHU 2-2 fluctuates between 37.2 kWh and 18.6 kW and that of AHU 2-1 peaks at 14.2 kWh, and its minimum value is 7 kWh. The average electricity consumption of AHU2-1 and AHU 2-1 during the baseline mode was 54.67 kWh and 16.07 kWh, respectively. However, the average electricity consumption of both AHUs during the smart mode was 24.5 kWh [AHU2-1] and 10.43 kWh [AHU 2-2]. Therefore, the total electricity consumption of AHU 2-2 is cut from 492 kWh in the baseline test to 220.5 kWh in the smart test, whereas the electricity consumption of AHU 2-1 is lowered by 50.7kWh from 144.6 kWh to 93.9 kWh, as demonstrated in Figure 23. The results also highlight that electrical energy savings in AHU 2-2 are about 55%, while AHU 2-1 saves approximately 35% of electricity usage when the smart mode is activated. While presenting electricity consumption analysis, cooling energy consumption is also investigated in this case study. Due to some constraints in the installation of BTU meters for each AHU to measure the cooling energy, only one BTU meter was installed at the common chilled water header to log the chilled water flow rates. Therefore, the cooling energy consumption (kWh) is calculated as follows:

Figure 22.

Daily electricity consumption of AHU2-1 and AHU2-2 during the baseline and smart modes.

Figure 23.

Electrical energy savings at different AHUs between baseline and smart modes.

$${\dot{Q}}_{\mathit{\text{cooling}}}={\dot{m}}_{\mathit{chw}}{\mathit{Cp}}_{\mathit{chw}}\left(T_R-T_S\right){\mathit{xN}}_{\mathit{op}}$$E2

In Eq. (2), the first parameter ${\dot{Q}}_{\mathit{\text{cooling}}}$ represents cooling energy consumption in kWh; the second parameter ${\dot{m}}_{\mathit{chw}}$ denotes mass flow rates of chilled water in kg/s, the third parameter ${\mathit{Cp}}_{\mathit{chw}}$ is the specific heat capacity of chilled water in kJ/kg·K, T represents temperature in °C, and $N_{\mathit{op}}$ is the operation time in hours. The subscript R and S represent return and supply, respectively. Figure 24 illustrates the accumulative cooling energy consumption for the baseline and smart mode tests. The smart mode is observed to consume 29% less cooling than the baseline test while maintaining thermal comfort in the space, because the cooling requirements in the office are significantly reduced by optimizing the supply airflow rates to facilitate the cooling load in each micro-zone. The results show that airside energy consumption can be reduced by as high as 50% of electricity consumption in AHUs, while the reduction in cooling supply to the office was also approximately 29%. The results also assure that reduction in the cooling supply and electrical energy consumption do not compromise the thermal comfort of the office.

Figure 24.

Cooling energy consumption during baseline test and smart test.

This case study demonstrated the application of AI-oriented control in airside air conditioning systems to resolve typical issues, such as thermal comfort and high energy consumption due to overcooling and undercooling, in open offices. It also highlights that the improvement on the airside also contributes to the reduction of electricity consumption of the fans, resulting in minimizing the waste energy as compared to the baseline control system while cooling required in the offices is also optimized.

4. Conclusion

This chapter investigates the application of AI solutions in optimizing airside performance while maintaining thermal comfort in the office. Pilot tests were conducted to examine the impact of proactive AI control on resolving common thermal comfort issues in the office, such as overcooling and undercooling. The demonstration testbed was implemented in one of the Singapore Institute of Technology floors, located at Dover, Singapore. The tests were conducted in baseline mode (conventional BMS control) and smart mode (proactive AI control). The results highlight that the proactive AI control solution provides not only the improvement of energy consumption but also an enhancement in thermal comfort by eliminating cold spots and hotspots in the office. Furthermore, it also highlights that the improvement on the airside also contributes to the reduction of electricity consumption of the fans, resulting in minimizing the waste energy as compared to the baseline control system, while cooling required in the offices is also optimized.

Acknowledgments

The author would like to express his gratitude to SP group Pte. Ltd., SP Digital Pte. Ltd. and 75F solution for providing the financial support to accomplish such a live testbed on SIT campus.