A multi-year campus-level smart meter database Authors: Mingchen Li, Zhe Wang, Yao Qu, Kin Ming Chui, Marcus Leung-Shea Published: 2024 ABSTRACT With the growing need for precise campus electricity management, understanding load patterns is crucial for improving energy efficiency and optimizing energy use. However, detailed electricity load data for campus buildings and their internal equipment is often lacking, hindering research. This paper introduces an energy consumption monitoring dataset from The Hong Kong University of Science and Technology (HKUST) campus in Hong Kong, comprising data from over 1400 meters across more than 20 buildings and collected over two and a half years. Using the Brick Schema curation strategy, raw data was curated into a research-ready format. This dataset supports various research tasks, including load pattern recognition, fault detection, demand response strategies, and load forecasting. Background & Summary In Hong Kong, residential energy consumption accounts for 24% of total energy use, while commercial sectors (including offices, healthcare, retail, etc.) account for 43%, indicating that building energy consumption has exceeded half of the total energy usage1. Utilizing smart meters to monitor different buildings and spaces within them can effectively identify load patterns and manage loads to enhance energy efficiency2. University campus areas are comprised of various types of buildings, including classrooms, libraries, dormitories, offices, restaurants, laboratories, and retail spaces, each with its unique energy consumption characteristics3. These buildings effectively represent the typical range of architectural types. Open-source databases that encompass various building types are critically important as they provide researchers and engineers with a rich resource of data for analyzing different patterns of energy consumption. Moreover, they enable further detecting unnecessary energy wastes, profiling customer behaviors, benchmarking and optimizing building energy efficiency, and forecasting of building loads. As demonstrated by Panapakidis and others4, Pan and colleagues5 conducted pattern recognition on a campus’s buildings and residential buildings based on electrical load. Furthermore, an increasing number of machine learning algorithms, such as Artificial Neural Network (ANN)6, Recurrent Neural Network (RNN)7, Long Short-Term Memory (LSTM)8, etc., are being used for load forecasting9–11. Datasets covering multiple types of buildings can significantly enhance the robustness of these models. However, currently, datasets on building electricity load often focus on single types of buildings, such as residential12–17 or office buildings18–20. This narrow focus provides an incomplete picture of the energy landscape, as it excludes a wide variety of other building types, such as retail, restaurant, and laboratory, each with distinct energy needs and consumption patterns. Even among the datasets that include multiple building types3, they lack detailed data on the electricity load of internal devices, such as Heating, Ventilation, and Air Conditioning (HVAC) systems, lighting, and office equipment. This omission hinders a comprehensive understanding of energy consumption and the identification of targeted energy conservation strategies. In 2020, an unprecedented global lockdown was implemented worldwide to control the spread of COVID-19 in many countries19. As of March 1, 2023, Hong Kong officially lifted the mask mandate, becoming one of the last regions in the world to do so, marking the end of its pandemic prevention policies21. During the pandemic, changes in travel frequency, epidemic prevention policies, and working behaviors significantly affected the energy consumption patterns of different types of buildings. A comparative study of energy consumption data collected during and after COVID-19 can reveal how occupant behaviors and energy usage patterns change as a result of the pandemic22. This paper presents a dataset on the electricity load of a campus, along with the development of its associated metadata using the semantic Brick23 model. The uniqueness of this dataset lies in:Two years of detailed, high-resolution time interval data from a real campus building complex, including data from one year of the COVID-19 pandemic and one year after lifting restrictions to return to normal life.The diversity of the dataset as it encompasses electricity load data from various types of buildings within the campus area, as well as from devices within these buildings. This diversity offers rich research opportunities by allowing the investigation of load variation patterns in response to different operational times of buildings and the behavior of occupants3.The dataset has been used for the campus’s annual energy usage analysis report, demonstrating its capability to accurately reflect the electricity usage situation of the campus.A Brick model has been developed, with the goal of assisting data users in easily understanding the data, and machine learning algorithms to automatically query the data. Brick schema is an open-sourced standardized semantic metadata model that describes physical, logical, and virtual assets in buildings and their interrelationships24. Two years of detailed, high-resolution time interval data from a real campus building complex, including data from one year of the COVID-19 pandemic and one year after lifting restrictions to return to normal life. Two years of detailed, high-resolution time interval data from a real campus building complex, including data from one year of the COVID-19 pandemic and one year after lifting restrictions to return to normal life. The diversity of the dataset as it encompasses electricity load data from various types of buildings within the campus area, as well as from devices within these buildings. This diversity offers rich research opportunities by allowing the investigation of load variation patterns in response to different operational times of buildings and the behavior of occupants3. The diversity of the dataset as it encompasses electricity load data from various types of buildings within the campus area, as well as from devices within these buildings. This diversity offers rich research opportunities by allowing the investigation of load variation patterns in response to different operational times of buildings and the behavior of occupants3. The dataset has been used for the campus’s annual energy usage analysis report, demonstrating its capability to accurately reflect the electricity usage situation of the campus. The dataset has been used for the campus’s annual energy usage analysis report, demonstrating its capability to accurately reflect the electricity usage situation of the campus. A Brick model has been developed, with the goal of assisting data users in easily understanding the data, and machine learning algorithms to automatically query the data. Brick schema is an open-sourced standardized semantic metadata model that describes physical, logical, and virtual assets in buildings and their interrelationships24. A Brick model has been developed, with the goal of assisting data users in easily understanding the data, and machine learning algorithms to automatically query the data. Brick schema is an open-sourced standardized semantic metadata model that describes physical, logical, and virtual assets in buildings and their interrelationships24. The potential applications of this data include:Load Pattern Recognition25,26: Through data analysis, identify energy usage patterns and load variation trends in various building types or zones to optimize power consumption and reduce costs.Fault Detection and Diagnostics27,28: Monitor for atypical energy consumption that could signal equipment failure, enabling timely corrective measures.Demand Response Strategy29: Develop data-driven strategies to curtail energy use during peak demand times, thereby improving grid reliability and energy efficiency.Load Forecasting26,30: Utilize statistical analyses and neural network techniques to forecast future energy demands based on historical data, ensuring a reliable power supply and efficient energy management. Load Pattern Recognition25,26: Through data analysis, identify energy usage patterns and load variation trends in various building types or zones to optimize power consumption and reduce costs. Load Pattern Recognition25,26: Through data analysis, identify energy usage patterns and load variation trends in various building types or zones to optimize power consumption and reduce costs. Fault Detection and Diagnostics27,28: Monitor for atypical energy consumption that could signal equipment failure, enabling timely corrective measures. Fault Detection and Diagnostics27,28: Monitor for atypical energy consumption that could signal equipment failure, enabling timely corrective measures. Demand Response Strategy29: Develop data-driven strategies to curtail energy use during peak demand times, thereby improving grid reliability and energy efficiency. Demand Response Strategy29: Develop data-driven strategies to curtail energy use during peak demand times, thereby improving grid reliability and energy efficiency. Load Forecasting26,30: Utilize statistical analyses and neural network techniques to forecast future energy demands based on historical data, ensuring a reliable power supply and efficient energy management. Load Forecasting26,30: Utilize statistical analyses and neural network techniques to forecast future energy demands based on historical data, ensuring a reliable power supply and efficient energy management. Methods Campus The dataset utilized in this study originates from real-world electricity consumption data from 1432 meters at the Hong Kong University of Science and Technology, situated in the rural coastal area of Sai Kung District, Hong Kong (22.3375 N, 114.263E). The campus spans 60 hectares, and the data is overseen by the University Sustainability Office. The detailed collection and analysis of electricity data aim to scrutinize current usage patterns across various building types and uncover potential energy efficiencies within the campus. For better management and analysis, the campus is segmented into three zones as depicted in Fig. 1.Fig. 1Overview of HKUST Campus. Overview of HKUST Campus. The Main Campus comprises the Academic Building and the Cheng Yu Tung Building. The Academic Building is a multifaceted complex that includes offices, classrooms, a library, dining facilities, and more. The Cheng Yu Tung Building is integrated with the Academic Building through a heat exchange link and is also part of the Main Campus. The Lee Shau Kee (LSK) Campus hosts the Lee Shau Kee Business Building, the HKUST Jockey Club Institute for Advanced Study/Lo Ka Chung Building, and the Li Dak Sum Yip Yio-Chin Kenneth Li Conference Lodge. The remainder of the buildings are categorized under the Rest Campus. For more detailed information about the campus, please refer to the webpage at HKUST Maps (https://mtpc.hkust.edu.hk/resources/maps). Building The Table 1 provides a comprehensive overview of the building types within the HKUST campus. It categorizes each building based on its primary function, such as educational, research, recreational, residential, cultural, or utility. This classification is crucial for analyzing building electricity consumption patterns.Table 1Building Types in the Campus.Building NameBuilding TypeAcademic BuildingEducational, ResearchBoat HouseRecreationalCheng Yu Tung BuildingEducational, ResearchCoastal Marine LabResearchIASR PlantEducational, ResearchIndoor Swim PoolRecreationalJockey Club HallResidentialLSK Business BuildingsEducationalNFF2 LabResearchOutdoor PoolRecreationalSea Water Pump HouseUtilitySeafront Indoor Sport CentreRecreationalSeafront Sport Centre Sport FieldRecreationalShaw AuditoriumCulturalStaff Quarters SSQ T1-19ResidentialStaff Quarters SSQ APT1-48, 1-49ResidentialStudent Hall UG1-UG9, PGResidentialStudent Hall GGTResidentialTower A-DResidentialWind Wave Tunnel BuildingResearch Table 1Building Types in the Campus.Building NameBuilding TypeAcademic BuildingEducational, ResearchBoat HouseRecreationalCheng Yu Tung BuildingEducational, ResearchCoastal Marine LabResearchIASR PlantEducational, ResearchIndoor Swim PoolRecreationalJockey Club HallResidentialLSK Business BuildingsEducationalNFF2 LabResearchOutdoor PoolRecreationalSea Water Pump HouseUtilitySeafront Indoor Sport CentreRecreationalSeafront Sport Centre Sport FieldRecreationalShaw AuditoriumCulturalStaff Quarters SSQ T1-19ResidentialStaff Quarters SSQ APT1-48, 1-49ResidentialStudent Hall UG1-UG9, PGResidentialStudent Hall GGTResidentialTower A-DResidentialWind Wave Tunnel BuildingResearch Building Types in the Campus. By understanding the building types, one can predict electricity usage trends and identify areas with higher consumption. For example, research buildings may require constant, high electricity for equipment, whereas residential buildings might have peak usage during mornings and evenings. This categorization enables more accurate and targeted analysis of electricity meter data, facilitating efficient energy management and conservation strategies across the campus. Zone To better manage and optimize energy consumption across the campus, the school has been divided into 104 Energy Zones. This approach offers several benefits. Firstly, by segmenting the campus into smaller areas, energy usage in each zone can be more precisely monitored and analyzed, enabling the identification and resolution of energy wastage in specific areas. Secondly, this granular management allows for the implementation of targeted energy-saving measures based on the unique consumption characteristics of each zone, thereby enhancing the overall energy efficiency of the campus. Lastly, by managing the campus in smaller sections, any issues of abnormal energy consumption can be quickly located and addressed, facilitating timely interventions. Within these Energy Zones, varying numbers of electricity meters have been deployed to monitor energy consumption. Some zones have one or more meters that directly measure the entire zone’s energy usage, while others do not have such direct measurements. To facilitate understanding, data for 34 zones within the Academic Building have been extracted, listing their names, the number of meters each zone has, and whether each zone has meters that directly measure its energy consumption, as shown in Table 2. Due to space constraints, it is impractical to list all 104 zones individually. For detailed query examples to get similar information, please refer to our GitHub repository at GitHub (https://github.com/LiMingchen159/HKUST_Meter_Brick).Table 2Zones in the Academic Building.CategoryEnergy ZoneNumber of MetersZone Consumption MeterLibraryLibrary_Canteen_Building_Elect36—SportHallLG1_Sport_Hall_Elect2—CarParksCarpark_Elect1—Carpark_MCC_from_SB1_8A_Elect4—Carpark_TX3_MCC_From_SB_1_1A_Elect4—Indoor_Carpark_Elect13—Indoor_Carpark_Switch_Board_2_Elect14—AZone_A1_A2_Lift_3_4_Elect39Meter_D0614Zone_A1_Lift_4_Elect1—Zone_A2_Lift_3_Elect3Meter_D0620, Meter_D0621Zone_A3_A4_Lift_1_2_Elect19Meter_D0613, Meter_D0619Zone_A5_A6_Lift_17_18_Elect36Meter_D0606Zone_A7_Lift_19_Elect12—DZone_D_Lift_25_26_Elect25Meter_D0637, Meter_D0649EZone_E_Lift_27_28_Elect25Meter_D0640, Meter_D0650HZone_H1_Lift_23_Elect22Meter_D0635, Meter_D0647Zone_H2_Lift_24_Elect23Meter_D0636, Meter_D0648JZone_J1_Lift_20_Elect26—Zone_J2_Lift_21_Elect29—Zone_J_Lift_20_21_Elect6Meter_D0651, Meter_D0690, Meter_D0691, Meter_D0692LZone_L1_Lift_22_Elect41Meter_D0633, Meter_D0645Zone_L2_Lift_22_Elect22Meter_D0634, Meter_D0646AnnexAnnex_Lift_29_30_Elect12—Enterprise CenterEnterprise_Centre_Elect39—Animal Care FacilityAnimal_Care_Facility_7_F_Elect65—ITSC Data CentreInformation_Technology_Services_Centre_Data_Centre_RM2113_Elect6—Information_Technology_Services_Centre_Data_Centre_RM2124_Elect14—Service roomBoiler_Water_Plant_LG1_BSU_Elect5—Building_Service_Unit_LG5_Elect6—Plumbing_Drainage_Switch_Board_Elect19—CanteenG_F_Canteen_Lift_13_15_Elect1—LG1_G_F_Canteen_Lift_13_15_Elect1—OthersPhase_1_Academic_Building_Elect24—Phase_2_Academic_Building_Elect11— Table 2Zones in the Academic Building.CategoryEnergy ZoneNumber of MetersZone Consumption MeterLibraryLibrary_Canteen_Building_Elect36—SportHallLG1_Sport_Hall_Elect2—CarParksCarpark_Elect1—Carpark_MCC_from_SB1_8A_Elect4—Carpark_TX3_MCC_From_SB_1_1A_Elect4—Indoor_Carpark_Elect13—Indoor_Carpark_Switch_Board_2_Elect14—AZone_A1_A2_Lift_3_4_Elect39Meter_D0614Zone_A1_Lift_4_Elect1—Zone_A2_Lift_3_Elect3Meter_D0620, Meter_D0621Zone_A3_A4_Lift_1_2_Elect19Meter_D0613, Meter_D0619Zone_A5_A6_Lift_17_18_Elect36Meter_D0606Zone_A7_Lift_19_Elect12—DZone_D_Lift_25_26_Elect25Meter_D0637, Meter_D0649EZone_E_Lift_27_28_Elect25Meter_D0640, Meter_D0650HZone_H1_Lift_23_Elect22Meter_D0635, Meter_D0647Zone_H2_Lift_24_Elect23Meter_D0636, Meter_D0648JZone_J1_Lift_20_Elect26—Zone_J2_Lift_21_Elect29—Zone_J_Lift_20_21_Elect6Meter_D0651, Meter_D0690, Meter_D0691, Meter_D0692LZone_L1_Lift_22_Elect41Meter_D0633, Meter_D0645Zone_L2_Lift_22_Elect22Meter_D0634, Meter_D0646AnnexAnnex_Lift_29_30_Elect12—Enterprise CenterEnterprise_Centre_Elect39—Animal Care FacilityAnimal_Care_Facility_7_F_Elect65—ITSC Data CentreInformation_Technology_Services_Centre_Data_Centre_RM2113_Elect6—Information_Technology_Services_Centre_Data_Centre_RM2124_Elect14—Service roomBoiler_Water_Plant_LG1_BSU_Elect5—Building_Service_Unit_LG5_Elect6—Plumbing_Drainage_Switch_Board_Elect19—CanteenG_F_Canteen_Lift_13_15_Elect1—LG1_G_F_Canteen_Lift_13_15_Elect1—OthersPhase_1_Academic_Building_Elect24—Phase_2_Academic_Building_Elect11— Zones in the Academic Building. Metered space and equipment As previously mentioned, the HKUST campus is divided into 104 distinct electrical zones. Each zone encompasses a variety of equipment and spaces. Given the extensive number of zones, it is impractical to list every piece of equipment contained within each zone in this article. However, we provide a request method to obtain this information, as illustrated in Listing 1. For more detailed code and implementation, please refer to our GitHub repository at (https://github.com/LiMingchen159/HKUST_Meter_Brick). By utilizing the provided request methods, researchers and facility managers can access specific data relevant to their needs, ensuring precise and actionable insights into the campus’s electrical consumption. Listing 1. An example SPARQL query to list equipment in a specific zone. Electric meter data The campus has deployed a variety of electricity meters across its campus to monitor energy consumption in different types of buildings. The specific details of these meters, including their providers, models, communication protocols, accuracy, resolution, and maximum readings, are summarized in Table 3.Table 3Power meter specifications and deployment locations.ProviderMain measurement buildingsPower meter modelCommunications protocolAccuracyResolutionMax readingMunHeanAcademic Buildings and Remote BuildingSKYDER-MH E63-LLoRaWANIEC 62053-21 Class 10.1 kWh9999999.9 kWhHensen & CET Electric TechnologyStaff quarters and Student HallPMC-350-CLoRaWANIEC 62053-21:2003 Class 10.01 kWh—PMC-53AModbus RTUIEC 62053-22 Class 0.5S0.1 kWh— Table 3Power meter specifications and deployment locations.ProviderMain measurement buildingsPower meter modelCommunications protocolAccuracyResolutionMax readingMunHeanAcademic Buildings and Remote BuildingSKYDER-MH E63-LLoRaWANIEC 62053-21 Class 10.1 kWh9999999.9 kWhHensen & CET Electric TechnologyStaff quarters and Student HallPMC-350-CLoRaWANIEC 62053-21:2003 Class 10.01 kWh—PMC-53AModbus RTUIEC 62053-22 Class 0.5S0.1 kWh— Power meter specifications and deployment locations. The university utilizes meters from two main providers: MunHean and Hensen & CET Electric Technology. MunHean meters are installed in academic buildings and remote buildings, specifically employing the SKYDER-MH E63-L model, which communicates via the Low-power, Long-range Wide Area Network (LoRaWAN) protocol. This meter adheres to the IEC 62053-21 standard, offering Class 1 accuracy, a resolution of 0.1 kWh, and a maximum reading capacity of 99,999,999.9 kWh. In contrast, Hensen & CET Electric Technology meters are deployed in staff quarters and student halls. The PMC-350-C model, used in these areas, also utilizes the LoRaWAN protocol and conforms to the IEC 62053-21:2003 Class 1 accuracy standard, but with a finer resolution of 0.01 kWh. Additionally, the PMC-53A model communicates through the Modbus RTU protocol and meets the IEC 62053-22 Class 0.5S accuracy standard, providing a resolution of 0.1 kWh. However, the maximum reading capacities for both Hensen & CET Electric Technology meters were not specified. Data transmission and storage To facilitate the development of the HKUST Sustainable Smart Campus, an Internet of Things (IoT) Network Services platform is provided for students and researchers to engage with IoT devices, smart applications, and real-time data analytics. This robust and secure “sensor to insight” platform is built on LoRaWAN technology, utilizing the unlicensed 923 MHz spectrum band in Hong Kong. Electric meters that are integrated with LoRaWAN technology transmit data seamlessly to these gateways, where it is subsequently stored on network servers. This setup allows for easy data visualization and application interfacing, as illustrated by the example in Table 3 for the PMC-350-C meter. For electric meters without inherent LoRaWAN functionality, a Modbus RTU RS485 to LoRaWAN Modem facilitates the necessary protocol translation, standardizing the data handling process across different meter types, as shown in Table 3 for the PMC-53A meter. The entire data transmission workflow is outlined in Fig. 2.Fig. 2HKUST LoRaWAN-Based IoT Network Architecture. HKUST LoRaWAN-Based IoT Network Architecture. Data curation In preparing the dataset, we encountered variations in the sampling frequencies due to the electricity meters being sourced from different suppliers and batches. Thus, we identified four distinct sampling frequencies: 15 minutes, 30 minutes, 1 hour, and 1 day. Missing values were replaced with the placeholder ‘NAN’. It is important to note that these missing values were neither imputed nor subjected to outlier detection. This approach was adopted due to the absence of reliable standards for guiding the imputation of missing values or the detection of outliers. Consequently, we decided to provide unaltered raw data, allowing researchers to apply their preferred data-cleaning techniques. Considering that the best methods can vary across different research or practical contexts, this flexibility is crucial. Data Records The dataset is meticulously organized into two main categories: time-series data and metadata. As illustrated in Fig. 3, the time-series data is stored in Excel (.xlsx) format, while the metadata is in Turtle (.ttl) format. This structured organization facilitates efficient data management and analysis. The raw dataset, comprising various equipment and spaces within the HKUST campus, totals approximately 773 MB. After resampling, the dataset size is 830 MB. The total compressed size of the dataset is approximately 1.43 GB. The choice of Excel for time-series data allows for easy manipulation and analysis due to its widespread use and compatibility with various analytical tools. The Turtle format for metadata, adhering to the Brick schema, ensures semantic richness and interoperability, enabling detailed, standardized descriptions and advanced data queries. This dual-format approach enhances the dataset’s accessibility and usability, supporting comprehensive analysis and effective energy management across the campus. The dataset is hosted on the Dryad website31 (10.5061/dryad.k3j9kd5h6).Fig. 3Organization of Raw and Clean Time-Series Datasets. Organization of Raw and Clean Time-Series Datasets. Time-series data The dataset comprises time-series data, meticulously organized into raw and cleaned formats, as shown in Fig. 3. The raw dataset includes multiple Excel files labeled GUI.NO.xxxxx.xlsx, such as GUI.NO.D0001.xlsx, GUI.NO.D0002.xlsx, and GUI.NO.M0604.xlsx, all housed within a ‘Time-series data’ folder. Accompanying these files is a metadata file, HKUST_Meter_Metadata.ttl, providing essential contextual information. The cleaned dataset involves resampling the raw data into specified intervals and categorizing them into separate directories: T15, T30, T60, and T1440, representing different time intervals of 15 minutes, 30 minutes, 60 minutes, and 1440 minutes (1 day) respectively. Each directory contains resampled Excel files maintaining the original naming convention. For instance, GUI.NO.D0001.xlsx appears within the T15, T30, T60, and T1440 directories, reflecting the resampled data at 15-minute, 30-minute, 60-minute, and daily intervals respectively. The metadata file, HKUST_Meter_Metadata.ttl, is also included in the cleaned dataset to ensure comprehensive documentation of the processed data. This structured approach facilitates ease of access and analysis, ensuring that both raw and resampled data are systematically categorized and well-documented. This organization is critical for reproducibility and further analysis in time-series research, enabling detailed exploration and utilization of the dataset. Brick metadata Along with the building’s time-series dataset, we used the semantic model as metadata to comprehensively describe the building and its smart meter systems. This model is based on the Brick schema, an open-source framework for standardizing the semantic description of building assets. The schema provides an extensible dictionary of terms and concepts, a set of relationships for linking these concepts, and a flexible data model leveraging semantic web technologies. The detailed Brick model is stored in .ttl file format, facilitating integration with other building management systems and smart analytics applications. This structured metadata is essential for optimizing building operations and developing intelligent control systems. As illustrated in Fig. 4, the model is composed of three key components: the Brick class (in green), Brick entities (in blue), and Entity Properties (in red). Entities are instances of classes, typically named to suit specific projects. At the same time, types ensure both humans and machines can correctly identify each instance, enabling rapid query responses — such as identifying various Sites or Buildings. Generally, properties are consistent across instances; for example, the Normal_Meter_M0270’s properties shown in the figure demonstrate that its unit is kWh and its resolution is 30 minutes, alongside the name of its associated time-series data file. We employ a dynamic location methodology in our modeling process to accommodate the varying measurement locations of each electricity meter. Initially, we convert these locations into Brick format and then add specific location information for each meter, as shown in yellow.Fig. 4Brick model for one Electrical Meter. Brick model for one Electrical Meter. Technical Validation First, we identified the status of all the meters to ensure that those with issues were excluded from further analysis and statistics. Figure 5 illustrates the various conditions of the meters and the corresponding counts. According to the Brick metadata, a total of 1412 meters were recorded. However, only 1394 meters had APIs available for data access. Some of these APIs returned errors when accessed, indicating that the API information was not properly maintained or updated, and there were 65 meters in this category. Additionally, there are meters for which the API can be accessed normally, but the data values retrieved are all zeros. 46 meters fall into this category. Consequently, the number of meters that can be reliably used for analysis is 1283.Fig. 5Meter Data Acquisition Status. Meter Data Acquisition Status. Missing Rate Figure 6 shows the percentage of missing data points for different sampling intervals and quarters from Q1 2022 to Q2 2024. The y-axis lists the sampling intervals: 15 minutes, 30 minutes, 1 hour, and 1 day, while the x-axis covers the different quarters. When calculating missing rates, we included not only meters sampled at each interval but also higher-frequency meters resampled to lower frequencies. For instance, the 15-minute interval includes all 15-minute meters, the 30-minute interval includes all 30-minute meters plus 15-minute meters resampled to 30 minutes, and so on.Fig. 6Missing Rates in Electrical Meter Data Across Different Sampling Times and Quarters. Missing Rates in Electrical Meter Data Across Different Sampling Times and Quarters. The 15-minute interval shows the highest missing rates, ranging from 23.467% to 52.417%. The peak is in 2022 Q4, indicating challenges in data collection at this granularity during this period. As the sampling interval increases, the missing rates decline. The 30-minute interval ranges from 18.548% to 36.680%, the 1-hour interval from 13.988% to 29.656%, and the 1-day interval from 3.377% to 13.584%. This trend highlights the difficulties in maintaining complete data at finer temporal resolutions. Several reasons could explain these missing rates. Higher frequency data collection (15 and 30 Minutes) might face technical issues such as sensor malfunctions or data transmission errors. These issues are less pronounced in lower frequency collections (1 Hour and 1 Day), where data transmission and storage are more manageable. Despite these missing rates, the data retains significant value. The lower missing rates at 1-hour and 1-day intervals still provide comprehensive insights for long-term trends and patterns. Additionally, data imputation techniques can mitigate the impact of missing values, ensuring the dataset remains robust for analysis. Furthermore, the decreasing trend in missing rates over time suggests improvements in data collection processes, reflecting the dataset’s growing reliability. To further elucidate the situation of missing data, we performed an hourly missing rate analysis for the meters in each building. Figure 7 presents the percentage of missing data points for various buildings and quarters, spanning from the first quarter of 2022 to the second quarter of 2024. The y-axis lists all the buildings on campus.Fig. 7Missing Rates in Electrical Meter Data Across Different Buildings and Quarters. Missing Rates in Electrical Meter Data Across Different Buildings and Quarters. Educational and Research buildings, such as the Academic Building, Cheng Yu Tung Building, IASR Plant, LSK Business Buildings, and NFF2 Lab, generally show good data quality in most quarters, with missing rates usually ranging from 0% to 20%. In the case of Residential buildings, the Staff Quarters SSQ APT1-49 and Student Hall GGT exhibit remarkable data quality, with missing rates in most quarters not exceeding 2%. However, the other Residential buildings display a different trend, with a significant increase in missing rates starting from the third quarter of 2022, which was partially addressed by the third quarter of 2023. This issue arose due to data collection problems during the Covid-19 pandemic. Although these problems were partially resolved post-pandemic, further troubleshooting is still necessary. For Recreational buildings, data quality appears to correlate with the importance of the building. For instance, Shaw Auditorium, a landmark building frequently hosting large performances, has very high data quality. In contrast, less significant buildings such as the Indoor Swim Pool and Coastal Marine Lab suffer from poor data quality due to insufficient maintenance. Energy consumption discrepancy analysis Some zones have main meters measuring total area electricity consumption, allowing us to validate our data. For these zones, we compare main meter values with sub-meter values. We first resample all meters to daily data and calculate daily consumption for each meter. We then compare the total consumption of the main meter with the summed consumption of the sub-meters. The relative error is computed by dividing the Mean Absolute Error (MAE) between the two by the larger of the two means. The definition of Relative Error is shown in formula 1:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rm{R}}{\rm{e}}{\rm{l}}{\rm{a}}{\rm{t}}{\rm{i}}{\rm{v}}{\rm{e}}\,{\rm{E}}{\rm{r}}{\rm{r}}{\rm{o}}{\rm{r}}=\frac{\mathop{\sum }\limits_{i=1}^{n}|\mathop{\sum }\limits_{k=1}^{p}{E}_{{\rm{m}}{\rm{a}}{\rm{i}}{\rm{n}}}(i,k)-\mathop{\sum }\limits_{j=1}^{m}{E}_{sub}(i,j)|}{max\,\left(\frac{1}{n}\mathop{\sum }\limits_{i=1}^{n}\mathop{\sum }\limits_{k=1}^{p}{E}_{{\rm{m}}{\rm{a}}{\rm{i}}{\rm{n}}}(i,k),\frac{1}{n}\mathop{\sum }\limits_{i=1}^{n}\mathop{\sum }\limits_{j=1}^{m}{E}_{sub}(i,j)\right)}$$\end{document}RelativeError=∑i=1n|∑k=1pEmain(i,k)−∑j=1mEsub(i,j)|max1n∑i=1n∑k=1pEmain(i,k),1n∑i=1n∑j=1mEsub(i,j)Where, Emain (i,k) is the electricity consumption recorded by the k-th main meter on day i, Esub (i,j) is the electricity consumption recorded by the j-th sub-meter on day i, n is the number of days, p is the number of main meters, and m is the number of sub-meters. In calculating the relative error, we processed the data further to ensure validation. Devices sometimes restart, causing large discrepancies in readings before and after a restart. We used the box plot method to remove such outliers by calculating the quartiles and excluding points more than 1.5 times the interquartile range (IQR) above the third quartile or below the first quartile. Additionally, as mentioned at the beginning of the Technical Validation section, some meters failed to obtain data correctly. Thus, if certain meters within a zone lack corresponding time series data, that zone will not be validated. The relative error results are shown in Table 4.Table 4Relative Error Results for Different Zones.NumberZone NameRelative Error1Tower_C_D_Elect16.46%2Zone_A3_A4_Lift_1_2_Elect9.40%3Zone_H1_Lift_23_Elect24.57%4Zone_H2_Lift_24_Elect30.75%5Zone_J_Lift_20_21_Elect12.06%6Zone_L2_Lift_22_Elect27.75% Table 4Relative Error Results for Different Zones.NumberZone NameRelative Error1Tower_C_D_Elect16.46%2Zone_A3_A4_Lift_1_2_Elect9.40%3Zone_H1_Lift_23_Elect24.57%4Zone_H2_Lift_24_Elect30.75%5Zone_J_Lift_20_21_Elect12.06%6Zone_L2_Lift_22_Elect27.75% Relative Error Results for Different Zones. Table 4 shows that relative error values across different zones vary significantly, ranging from 9.40% to 30.75%. Several factors could contribute to these discrepancies. Some devices are likely not connected to sub-meters while still being recorded by the main meter, leading to potential discrepancies since sub-meters do not capture all energy usage within a zone. Additionally, data loss or interruptions during long-term monitoring could result in mismatches between sub-meter and main meter readings. Differences in meter accuracy also play a role; sub-meters with lower precision may not accurately record small fluctuations in electricity consumption, causing cumulative discrepancies. These factors together likely contribute to the relative errors. Analysis of metered objects In the 1412 meters discussed earlier, 1268 measure devices, 121 measure rooms, and the rest measure zones. To identify which devices are being measured, we list the top 20 device types in Fig. 8.Fig. 8Top 20 Measured Device Types. Top 20 Measured Device Types. Figure 8 reveals that 11.9% of the meters measure ‘Lighting & Other’. Interestingly, these meters do not exclusively monitor lighting, implying cost-saving measures where a single meter may monitor multiple devices. The next most prevalent category is ‘MCB Board’, indicating that some meters record Miniature Circuit Breaker Board (MCB Board) measurements without specifying other devices. The top 20 device types represent 73.3% of all meters. This suggests that although 104 device types are measured in total, the remaining devices are less common in building equipment, with only a minority of meters measuring them. To further validate our data, we conducted focused data analysis on the most frequently measured device, ‘Lighting’, detailed in Fig. 9.Fig. 9Hourly kW Distribution Over 2 Years: Weekdays vs. Weekends in Academic Building. Hourly kW Distribution Over 2 Years: Weekdays vs. Weekends in Academic Building. Figure 9 depicts the hourly kW distribution over two years, comparing weekdays and weekends in the Academic Building. The Ground Floor, being an open outdoor space, shows lower daytime lighting consumption (7:00 AM to 5:00 PM) compared to nighttime. This pattern is consistent regardless of weekday or weekend. Other floors have similar patterns. Weekdays show increased lighting usage around 8:00 AM, peaking at noon, and a slight dip around midday, typical of lunch breaks. From 2:00 PM to 11:00 PM, lighting usage decreases gradually, indicating reduced evening occupancy. Weekends exhibit minimal lighting changes, with a slight increase in the afternoon, reflecting fewer occupants mainly active during that period. This electricity usage pattern aligns with typical trends in educational and research buildings, validating our data’s reliability. Figure 10 illustrates the electricity consumption distribution in the bedrooms and bathrooms on the first floor of GGT dormitory, categorized by hourly data across the four seasons: spring, summer, autumn, and winter. In Hong Kong, summer temperatures typically range from 80 to 88 degrees Fahrenheit, necessitating prolonged cooling to maintain indoor comfort. Consequently, electricity consumption during summer is significantly higher than in other seasons, while winter sees a marked decrease due to reduced cooling needs. Throughout the day, peak electricity consumption occurs around midnight (12:00 AM), closely linked to the extensive use of air conditioning and lighting as students study or rest. From 12:00 AM to 6:00 AM, consumption gradually declines, indicating that more residents are going to sleep. Usage rises at around 7:00 AM, peaking slightly between 10:00 AM and 12:00 PM as students begin their daily activities, using devices like lights and computers. There is a slight dip in electricity usage during the afternoon; however, consumption increases again in the evening — particularly after 7:00 PM — highlighting that night is a primary period for student activities with frequent use of lights and electronic devices. This consumption pattern aligns with typical trends in student dormitories, validating the reliability of our data. Due to the breadth of buildings and devices included, detailed analysis of each is impractical within the scope of this study.Fig. 10Hourly Power Consumption Patterns by Season for GGT 1 F. Hourly Power Consumption Patterns by Season for GGT 1 F. Usage Notes The time-series data is in xlsx format and can be loaded into any compatible software. The dataset is 1.43GB. Researchers can use Python, Java, MATLAB, or R to process it. Detailed information on data preprocessing, analysis, missing data rates, and query examples is available on the dataset’s GitHub page (https://github.com/LiMingchen159/HKUST_Meter_Brick). This dataset is also useful for research in load pattern recognition, fault detection, demand response strategies, and load forecasting.