Big data analytic for energy management – Predictive control of air conditioning system at HKIA

By EG Division

Climate change has become a pressing concern around the world. In order to combat climate change, concerted efforts are crucial to reducing greenhouse gas emission to limit global warming in the 21st century to below 2ºC, preferably below 1.5ºC, compared to pre-industrial levels¹. In Hong Kong, the Government is committed to achieving carbon neutrality before 2050. As buildings account for 90% of the electricity used in Hong Kong and generate over 60% of Hong Kong’s carbon emissions, “energy saving and green building initiatives” is one of the main decarbonisation strategies in Hong Kong^2,3.

Hong Kong has a subtropical climate with hot humid summers and relatively dry mild winters. Similar to other major cities, against the background of climate change and urbanisation, Hong Kong has been experiencing adverse effects of climate change. This includes significant increasing trend in local temperatures over the last century with sharp increases in the number of very hot days (daily maximum temperature at 33.0ºC or above) and hot nights (daily minimum temperature at 28.0ºC or above) in recent decades⁴. In 2021, there were 61 hot nights and 54 very hot days in Hong Kong, both ranking the highest since records began in 1884⁵. Future projection of temperature suggests an even warmer climate for the rest of the 21st century⁶. Since temperature has long been considered as one of the key factors influencing building energy consumption, it poses a great challenge to the building energy management and implementation of related energy saving initiatives^7,8.

Annual mean temperature recorded at the Hong Kong Observatory Headquarters (1885-2021). Data are not available form 1940 to 1946.

To help combat climate change, Airport Authority Hong Kong (AAHK), CLP Power Hong Kong Limited (CLP Power), Hong Kong Observatory (HKO) and Hong Kong Green Building Council Limited (HKGBC) recently joined forces to develop an accurate cooling load predictive model for Hong Kong International Airport’s passenger terminals by utilising weather, electricity consumption and airport specific data as well as load profile for chiller plant proactive control. Each entity has specific involvement in climate change actions as described below.

Hong Kong International Airport is the first airport in the world using this mathematical model and machine learning to predict and regulate the precise amount of cooling required without wasting energy.

Collaboration with CLP Power, Hong Kong Observatory and Hong Kong Green Building Council

AAHK is responsible for the development and maintenance of Hong Kong International Airport (HKIA). AAHK and its key aviation-related business partners are committed to achieving Net Zero Carbon by 2050, with a midpoint target of 55% absolute emissions reduction by 2035 from a 2018 baseline. The target aligns with the Intergovernmental Panel on Climate Change (IPCC) 1.5°C scenario, the HKSAR Government’s 2050 carbon neutrality target and the Airports Council International (ACI) Long Term Carbon Goal of Net Zero Carbon.

CLP Power is a major Hong Kong power utility wholly owned by CLP Holdings Limited, a company listed on the Hong Kong Stock Exchange and one of the largest investor-owned power businesses in Asia. CLP Power operates a vertically integrated electricity supply business in Hong Kong to supply highly reliable electricity to over 80% of Hong Kong’s population. The carbon intensity of CLP Power’s electricity supply has fallen significantly over the last 15 years and is expected to fall a further 50% over the next 15 years. CLP Power is committed to helping its customers use trending technologies to enhance energy efficiency.

HKO is the meteorological authority of Hong Kong, responsible for monitoring and forecasting weather as well as issuing warnings on weather-related hazards. For weather monitoring, HKO has expanded the areal coverage and enhanced the temporal resolution of its meteorological observation significantly over the years⁹. Moreover, using post-processed outputs from state-of-the-art numerical weather prediction models, HKO now provides location-specific weather forecast for the next nine days to the public and stakeholders through the Automatic Regional Weather Forecast (ARWF) service. ARWF provides hourly forecasts of different weather elements at selected locations, including temperature, humidity, wind speed and direction, state-of-sky (cloudiness) and the daily chance of precipitation. These weather data and forecast products are well-received by the public through the ARWF web portal and HKO’s mobile app “MyObservatory”. Furthermore, ARWF products have been widely used by climate sensitive sectors in various applications, including energy saving and management.

HKGBC is a non-profit, member led organisation established in 2009 and became a public body under the Prevention of Bribery Ordinance in 2016. HKGBC promotes the standards and developments of sustainable buildings in Hong Kong. It also aims to raise green building awareness by engaging the public, the industry, and the government, plus develop practical solutions for Hong Kong’s high-rise, high density urban area of green building development.

AAHK, CLP Power and HKGBC are also partners of the HKSAR Government’s Carbon Neutrality Partnership.

The multidisciplinary collaboration is definitely an unprecedented success in realising the innovative idea to drive energy saving by utilising meteorological data, big data analytic and machine learning.

The Cooling Load Predictive Model

Building cooling load is composed of both external load and internal load. External load is affected by the daily and seasonal changes in weather throughout the year whilst internal load is due to occupants, artificial lighting and plug load within the building that change on a daily or weekly basis. Since the thermal inertia of buildings is large and the activities of occupants would not change very much within a 30-minute interval, the hourly changes in building cooling load would not be significant.

The carbon intensity of CLP Power’s electricity supply is expected to fall more than 50% in the next 15 years.

Traditional control of a chiller plant is reactive because indoor temperature takes time to change after additional chiller is switched in or out. To avoid customer complaints on thermal comfort, especially in summer, some operators may overcool the building, causing energy wastage. The objective of the cooling load predictive model is to predict the cooling load and apply it to proactively control the chiller plant to reduce energy wastage.

Cooling load prediction can be very effective for optimising the chiller sequencing control by reducing unnecessary switching actions (i.e. control action and reaction). Using proactive control strategies to determine the supply of cooling to match the predicted cooling demand can minimise the energy consumption of the air conditioning system by reducing the number of action and reaction responses in temperature regulation by the chiller plant. The resultant energy saving can be substantial for air conditioning systems with large inertia of changes.

Nowadays, advance statistical models using artificial neural network, fuzzy logic and genetic algorithm are commonly adopted for the development of prediction models especially for modelling parameters with complicated relationships¹⁰. However, such modelling techniques usually require many types of data for tuning and updating which is not normal practice in building services facilities. Alternatively, regression analysis is much more practical and robust.

The Cooling Load Predictive Model for chiller plant proactive control at Terminal One of Hong Kong International Airport

The development team of the cooling load predictive model is made up of experts from diverse professional disciplines. These include building services engineering experts who are familiar with the characteristics of the airport’s operations, meteorology experts who are familiar with climate characteristics and weather nowcasting, and retro-commissioning experts who are familiar with data analytics and chiller optimisation. Therefore, AAHK, CLP Power, HKO and HKGBC have collaborated to develop the cooling load predictive model for chiller plant proactive control at Terminal 1 of HKIA. The model is a combination of multiple regression and autoregressive analysis for enhancing the accuracy of cooling load prediction^{11, 12}. There are five major development stages, namely, the identification of necessary data of plant and meteorology, the collection of necessary big data, development and testing of the cooling load predictive model, development of chiller control rules and estimation of energy saving.

Project team with members from AAHK, CLP Power, HKO and HKGBC

Development of the Cooling Load Predictive Model

The most important consideration for adopting proactive control on a chiller plant is the accuracy of the cooling load predictive model and the corresponding control strategies for managing the system response. As such, stability and reliability of the cooling load predictive model are very important. A multi-variable prototype model has been built. Significance of parameters is tested by t-Stat and the accuracy of the prediction is tested by the correlation coefficient (R²). The prototype model has been repeatedly fine-tuned and tested.

During the fine-tuning process, parameters in the model have been added/modified/removed based on the t-Stat of their regression coefficients. The improvement of the overall prediction accuracy has been monitored by the R² of the model. Finally, eight basic parameters in descending order of their relative magnitudes of regression coefficients have been identified, namely, Dry Bulb Temperature, Number of Flights, Relative Humidity, Solar Radiation, Cloud Amount, Seawater Temperature, Wind Speed, and Wind Direction. The R² of the cooling load predictive model given by the simple regression of the above parameters has exceeded 0.8.

Challenges encountered

The accuracy, stability and reliability of the prediction model are most important because they will affect the accuracy of the control strategies and the energy consumption of the chiller system, particularly in summer when the energy consumption by air conditioning is much higher than that in winter. To raise the prediction accuracy, the project team has critically reviewed the architectural design and operations of HKIA, the characteristics of the airport’s electricity consumption profile in different months and years, and the possibility of combined effect between various weather parameters. Different hypothesis have been initiated, tested, modified and retested.

Since the structure of the passenger terminal building is a huge box with large surface-area-to-volume ratio, the external load plays an important role in cooling load variation especially when sunny weather conditions change to cloudy or rainy. Apart from ambient temperature, solar radiation, relative humidity and wind factors all become the major weather parameters included in the regression portion of the predictive model. HKO can provide localised nowcasting weather parameters except solar radiation, which can be replaced by the calculated solar irradiance together with the nowcasting cloud coverage data. Wind speed and wind direction are data that may fluctuate vigorously as compared with the cooling load variation. This data may be removed from the prediction model during its development process. Cross-product terms and higher order terms are also included in the prediction model for enhancing the non-linear effect on the cooling load.

A huge spatial void exists inside the building which creates large internal thermal air capacity. Therefore, the heat gain or cooling effect takes time to influence the air mass, and as a result the thermal inertia causes significant delay to change of space temperature. This delay can create a time lag of up to five hours.

The internal load is composed of occupancy load, lighting load and plug load. The building operates 24-hours a day nonstop during which both lighting load and plug load remain steady. The change in lighting and plug load can also be better specified by splitting the model into summer and winter as well as day mode, night mode and midnight mode. The occupancy load is dependent upon the flight schedule, in which occupancy load in the next hour increased with flights arrival, and occupancy load in the current hour decreased with flights departure. All arrival and departure flights are pre-scheduled by the computer system, which is quite accurate and steady unless an accident (e.g. storm) occurs. The change of occupancy load also affects the amount of fresh air load required by the building.

The heat rejection of the chiller plant is done by means of seawater cooling. The temperature of seawater is also another weather parameter which can be obtained from the building management system of the chiller plant. No nowcasting seawater temperature is provided by the Hong Kong Observatory, but is determined by the nowcasting ambient temperature and the past record of seawater temperature profile.

As a result, the following improvements have been added to the model:

• Increase time delay from one hour to six hours to account for the latency in heat gain of the space, which is caused by solar radiation absorbed by the materials of the floor, furniture and building façade. The radiation thus absorbed will be emitted in the form of heat energy and heats up the space later.

• Split the model into summer and winter modes because the electricity consumption characteristics of the chiller system are different in summer and winter. The winter mode covers January to March while the summer mode covers the rest of the year.

• Split the model into day-and-night mode (06:00-24:00) and midnight mode (24:00-06:00) because a steady distinctive number of flights under the respective hours of the modes has been observed.

Cooling Load (t) for day-and-night mode from 06:00 to 24:00 =

k₀ + k₁*WS(t) + k₂*TDB*CC(t) + k₃*CC(t)^2 + k₄*CC(t) + k₅*TDB(t) + k₆*TDB^2(t) + k₇*%RH(t) + k₈*WD(t) + k₉*TSW(t) + k₁₀*NFT(t) + k₁₁*IR(t) + k₁₂*TDB(t)*WS(t) +

k₁₃*WS(t-3) + k₁₄*TDB*CC(t-3) + k₁₅*CC(t-3)^2 + k₁₆*CC(t-3) + k₁₇*TDB(t-3) + k₁₈*TDB^2(t-3) + k₁₉*%RH(t-3) + k₂₀*WD(t-3) + k₂₁*TSW(t-3) + k₂₂*NFT(t-3) + k₂₃*IR(t-3) + k₂₄*TDB*WS(t-3) +

k₂₅*WS(t-6) + k₂₆*TDB*CC(t-6) + k₂₇*CC(t-6)^2 + k₂₈*CC(t-6) + k₂₉*TDB(t-6) + k₃₀*TDB^2(t-6) + k₃₁*%RH(t-6) + k₃₂*WD(t-6) + k₃₃*TSW(t-6) + k₃₄*NFT(t-6) + k₃₅*IR(t-6) + k₃₆*TDB*WS(t-6)

Cooling Load (t) for midnight mode (24:00 to 6:00) =

k₀ + k₁*CC(t) + k₂*TDB(t) + k₃*TDB^2(t) + k₄*%RH(t) + k₅*WD(t) + k₆*TSW(t) + k₇₀*NFT(t) + k₈*IR(t) + k₉*TDB(t)*WS(t) +

k₁₀*CC(t-3) + k₁₁*TDB(t-3) + k₁₂*TDB^2(t-3) + k₁₃*%RH(t-3) + k₁₄*WD(t-3) + k₁₅*TSW(t-3) + k₁₆*NFT(t-3) + k₁₇*IR(t-3) + k₁₈*TDB*WS(t-3) +

k₁₉*CC(t-6) + k₂₀*TDB(t-6) + k₂₁*TDB^2(t-6) + k₂₂*%RH(t-6) + k₂₃*WD(t-6) + k₂₄*TSW(t-6) + k₂₅*NFT(t-6) + k₂₆*IR(t-6) + k₂₇*TDB*WS(t-6)

where,
t is the time stamp
TDB is the ambient dry bulb temperature
RH is the relative humidity
WS is the wind speed
WD is the wind direction
RF is the rainfall
TSW is the seawater temperature
SR is the direct solar radiation
NFT is the number of flights
CC is the cloud coverage
IR is the solar irradiance

After the above improvements have been adopted into the model, the average R² of the day-and-night mode and mid-night mode for summer have increased to 0.9525 and 0.8583 respectively. Therefore, the cooling load in summer is much more accurately predicted.

Time delay terms are used in the model to equate the latency in heat gain of the space. Separate sets of summer and winter cooling load predictive equations are used to raise the predicative accuracy.

When applying the prediction model to determine the hourly cooling load for the building, if all the nowcasting data are within the boundary conditions of the dataset used for developing the model, the accuracy of the model can be maintained. Since the year-round odd hours nowcasting data adopted for developing the model and the year-by-year change of such weather parameters are quite steady and stable, these would not fall outside the boundary conditions. Seasonal variation of seawater temperature is also stable and partially aligned with the variation of ambient temperature. Furthermore, the year-round flight schedule indicates that the number of arrival and departure flights per day is quite stable under normal operations. Lighting and plug load are also steady if all the served areas and occupied areas are under normal operation. Consequently, all the nowcasting parameters are within their boundary conditions and the determined regression coefficients can be adopted for normal operation. If there are operational changes such as significant change in the arrival/departure flight schedule, partial shutdown of served areas and occupied areas, or relocation of the intake of seawater, the regression coefficient is required to be generated again through a moving time process. This would necessitate replacing the oldest set of nowcasting parameters with the newly obtained nowcasting parameters and conducting the regression analysis again to generate a new set of regression coefficients. This regeneration process should be kept ongoing on an hourly basis with the boundary conditions being updating accordingly until the changes become stable and remain steady.

Integration of the cooling load predictive model into the chiller optimisation system

The original chiller optimisation control system is rule-based¹³. Based on energy efficiency and operation parameters, a prioritisation index is developed for the activation of control rules which will be used to determine the number of chillers to be put into operation. After the integration with the cooling load predictive model, the number of chillers to be put into operation is also determined by the predicted cooling load. Since the cooling load is more accurately predicted, the chiller optimisation system can determine the required number of chillers to be put into operation. Excessive cooling is greatly reduced or eliminated. Energy is thereby conserved without impairing passenger comfort.

The above figure shows the typical daily cooling load profile and the cooling capacity delivered to the terminal before and after implementation of the predictive control system. As shown by the grey line in the chart, the cooling demand gradually increases in the daytime owing to the increased passenger flow and solar heat gain until the peak which occurs around 4 pm. The cooling demand then decreases after the peak following the reduction of passenger traffic and solar load at night.

Before implementation of the predictive model, the chillers were started in advance to provide sufficient cool air in the terminal building to suit the cooling load profile as reflected by the orange line in the chart. With the introduction of the predictive control system, based on the predicted cooling load, just-in-time chiller operation is enabled. The system can pre-adjust for starting the chillers and output the most comfortable cooling volume and temperature. Looking at the blue line in the chart, it is much closer to the actual cooling load, achieving energy saving.

In the middle of the day, when the cooling demand increases with the change of weather, the predictive air-conditioning control system allows timely response to the changes and addresses the upcoming increment in cooling demand. Thermal comfort of the terminal building can therefore be enhanced.

In the evening when the cooling load decreases gradually, the predictive air-conditioning control system can allow pre-adjustment of the chiller system to suit the decreasing cooling load profile. Compared with traditional chiller control, the new predictive control system can avoid producing excessive cooled air and thus reduce energy wastage.

The model is synchronised with HKO’s observed and forecast weather data and automatically updates each hour. The predicted cooling load automatically updates accordingly. The system not only saves energy and costs but also provides travellers with a more comfortable environment, as it helps maintain indoor environment more accurately and avoids excessively high or low temperatures caused by sudden weather changes.

To optimise the predictive air-conditioning control system, two existing constant speed chillers rated at 3,500 TR were replaced with high-efficiency chillers rated at 5,000 TR complete with variable speed drives. The application of variable-speed chillers allows delivery of cooling capacity to meet the required cooling demand more accurately. Considering the efficiency of the chillers, the most efficient chiller combination was adopted to improve overall system efficiency and achieve energy saving.

Newly replaced Variable-speed Chiller at Terminal 1 Chiller Plant

Together with the advanced chiller system, the predictive air-conditioning control system has achieved an annual energy saving of 5.1 million kWh, which is approximately 12% of the total energy consumption of the chiller system. Taking the average carbon emission factor of CLP Power Hong Kong in 2020 at 0.37 kg CO₂e per kWh, approximately 1,900 tonnes of carbon emissions can be avoided annually.

Year-round consumption of the chiller system in 2019 (Yr2019) compared with that using predictive air-conditioning control (ML-AR 3500RT), and that using predictive air-conditioning control for advanced chiller system (ML-AR 5000RT) with savings of 1.7 M kWh/annum and 5.1 M kWh/annum respectively.

The above is an artist's impression of the 3RS of the Hong Kong International Airport

Closing remarks

This project showcases the benefits of collaboration between government, public utility and professional body in utilising big data across their respective disciplines for effective building energy saving. Moreover, this initiative serves as an exemplary case for promoting energy saving which echoes relevant policies of the HKSAR Government in climate change mitigation³. The cooling load predictive model can be widely adopted by other airports and commercial buildings with central air-conditioning system. As a way forward, the project team will further enhance the accuracy of the system through machine learning and the introduction of new data such as passenger flow image analysis. Plans are also underway to apply the system in other passenger facilities, including the new Terminal 2 and Terminal 2 Concourse of the Three-runway System (3RS), to further optimise energy efficiency.

In recognition of the dedication and excellent performance of the project in energy efficiency and saving, this project on cooling load predictive model has been awarded the Energy Project of the Year for the Asia-Pacific Rim region by the Association of Energy Engineers (AEE) in the United States. The project also won 1st runner up in the IFMA Asia Pacific Innovation Award in 2021.

Winning the Asia Pacific Rim Region Energy Project of the Year Award for 2021

About the authors: Ir Joseph Lam is from the Airport Authority Hong Kong, Ir Simon Tsui is from CLP Power Hong Kong Limited, Mr W K Wong and Dr T C Lee are from Hong Kong Observatory and Ir Dr Paul Sat is from Hong Kong Green Building Council with coordination of the EG Division.

References

1. UNFCCC (2016). The Paris Agreement. [online]. United Nations Framework Convention on Climate Change. Available at: <https://unfccc.int/process-and-meetings/the-paris-agreement/the-paris-agreement>.
2. HKSAR Government (2017). CLIMATE ACTION PLAN 2030+. [online]. Available at: <https://www.climateready.gov.hk/files/report/en/HK_Climate_Action_Plan_2030+_booklet_En.pdf>.
3. HKSAR Government (2021). Hong Kong’s Climate Action Plan 2050. [online]. Available at: <https://www.climateready.gov.hk/files/pdf/CAP2050_booklet_en.pdf>.
4. Hong Kong Observatory (2022). Climate Change in Hong Kong. [online]. Available at: <https://www.hko.gov.hk/en/climate_change/climate_change_hk.htm>.
5. Hong Kong Observatory (2022). Year’s Weather 2021. [online]. Available at: https://www.hko.gov.hk/en/climate_change/climate_change_hk.htm.
6. Hong Kong Observatory (2022). Climate Projections for Hong Kong. [online]. Available at: https://www.hko.gov.hk/en/climate_change/future_climate.htm.
7. Lee T C, Kok M H & Chan K Y (2010). Climatic Influences on the Energy Consumption in Domestic and Commercial Sectors in Hong Kong. In: 16th Annual International Sustainable Development Research Conference. Hong Kong, China: HKO, Reprint No. 903, 2010.
8. Morakinyo T E, Ren C, Shi Y, Lau K L, Tong H W, Choy C W, Ng E (2019). Estimates of the impact of extreme heat events on cooling energy demand in Hong Kong. Renewable Energy, 142, pp. 73-84.
9. Lee T C, Wong W K and Tam K H (2018). Urban-focused weather and climate services in Hong Kong. Geoscience Letters, 5:18.
10. Fan C, Liao Y and Ding Y (2019). Development of a cooling load prediction model for air-conditioning system control of office building. Internal Journal of Low-Carbon Technologies, 14, pp.70-75.
11. Jia J, Xing J, Ling J, Peng R (2016). A method to predict cooling load of large commercial buildings based on weather forecast and internal occupancy. Frontiers in Energy, 10(4), pp.459-465.
12. Alonso S, Moran A, Prada M A, Reguera P, Fuertes J J and Dominguez M (2019). A Data-Driven Approach for Enhancing the Efficiency in Chiller Plants: A Hospital Case Study. Energies, 12, pp. 827.
13. Stewart I, Aye L, Peterson T(2017). Global optimisation of chiller sequencing and load balancing using shuffled complex evolution. In: AIRAH and IBPSA’s Australasian Building Simulation 2017 Conference. Melbourne.