AI Agent-Based Intelligent Urban Digital Twin (I-UDT): Concept, Methodology, and Case Studies

Highlights

What are the main findings?

• In the developed AI agent-based intelligent digital twin (I-DT), UBEM overcomes the limitations of the traditional UBEM approach and enables efficient analysis of urban building energy.
• GPT-based UBEM effectively performed the core functions of UBEM, serving as a key technology in I-UDT applications and services.

What are the implications of the main findings?

• The I-UDT enables more accurate and comprehensive urban energy management, supporting the development of sustainable cities and carbon-neutral strategies.
• Implementing I-UDTs enables urban policymakers to make data-driven decisions, improve energy efficiency, and enhance the scalability of digital twin applications.

Abstract

The concept of digital twins (DTs) has expanded to encompass buildings and cities, with urban building energy modeling (UBEM) playing a crucial role in predicting urban-scale energy consumption via modeling individual energy use and interactions. As a virtual model within urban digital twins (UDTs), UBEM offers the potential for managing energy in sustainable cities. However, UDTs face challenges with regard to integrating large-scale data and relying on bottom-up UBEM approaches. In this study, we propose an AI agent-based intelligent urban digital twin (I-UDT) to enhance DTs’ technical realization and UBEM’s service functionality. Integrating GPT within the UDT enabled the efficient integration of fragmented city-scale data and the extraction of building features, addressing the limitations of the service realization of traditional UBEM. This framework ensures continuous updates of the virtual urban model and the streamlined provision of updated information to users in future studies. This research establishes the concept of an I-UDT and lays a foundation for future implementations. The case studies include (1) data analysis, (2) prediction, (3) feature engineering, and (4) information services for 3500 buildings in Seoul. Through these case studies, the I-UDT was integrated and analyzed scattered data, predicted energy consumption, derived conditioned areas, and evaluated buildings on benchmark.

1. Introduction

1.1. Background

Cities account for over 75% of global primary energy consumption, with buildings contributing a significant portion to this usage [1]. They are also some of the main sources of greenhouse gas emissions, and as the urban floor area expands, energy consumption will continue to rise [2]. Consequently, many countries have outlined specific goals and strategies for achieving carbon neutrality in the building sector [3]. In South Korea, the Ministry of Land, Infrastructure, and Transport is promoting a 2050 carbon neutrality initiative based on energy performance data [4]. As part of the “2050 Carbon Neutrality Strategy” initiated in 2021, policies have been introduced to manage carbon emissions throughout the life cycles of buildings, including mandatory zero-energy construction for new buildings and retrofitting existing buildings, aimed at creating carbon-neutral cities [4]. Improving the energy performance of buildings and continuously evaluating net-zero energy and retrofitted buildings have thus emerged as essential challenges for sustainable urban development.

Energy consumption in cities arises from interactions between multiple sectors, including buildings, mobility, industry, and infrastructure, necessitating a holistic and comprehensive analysis [5,6]. Within the realm of buildings, energy consumption considerations should extend beyond intra-building usage to include inter-building physical interactions. These inter-building physical interactions encompass factors such as shading effects from neighboring structures, variations in wind flow due to building spacing and height differences, and the influence of thermal radiation exchange between buildings. Such interactions significantly impact energy performance, making their consideration essential for accurate urban-scale energy modeling. Urban building energy modeling (UBEM) has recently become a vital tool for predicting and analyzing building energy consumption on an urban scale [7,8,9]. Through UBEM, it is possible to predict and analyze citywide energy consumption patterns by considering individual building energy characteristics, inter-building interactions, and relationships with surrounding environments. One of UBEM’s main strengths is its ability to comprehensively incorporate diverse building types, ages, energy systems, and climatic factors [10]. This capability plays a critical role in accurately assessing building energy performance and scientifically formulating energy-saving strategies applicable across various urban zones [11]. Moreover, UBEM continues to evolve with advancements in technology, providing new insights and information on urban building energy consumption characteristics through the integration of artificial intelligence, machine learning, and diverse data analysis techniques [12].

In the existing research investigating UBEM through a top-down approach, researchers have employed accumulated energy consumption data and building features using various visualization tools to monitor and analyze building energy. As the utilization of AI (artificial intelligence) agents continues to increase, new studies on the application of AI agents for UBEM are also emerging [13]. Top-down approaches are constrained by significant limitations, including challenges in integrating large-scale fragmented data, a lack of detailed and actionable information, and restricted practical applicability. Furthermore, the black-box nature of the generated results and benchmarks often impedes interpretability and transparency. In contrast, AI agent-based methods offer substantial advancements, such as the automated integration of urban building big data, the incorporation of new data sources and insights provided by AI, and improved services for users. Additionally, the explainability of result derivation has been enhanced through AI-driven learning processes, addressing critical deficiencies inherent in traditional top-down methodologies. The studies in this field can be categorized into three areas: (1) UBEM and data fusion, (2) urban digital twins, and (3) the application of AI agents for UBEM. Guarino et al. [14] developed a data fusion approach using climate models to predict future heating and cooling energy demand in buildings under climate change conditions, achieving more reliable estimations compared to individual model outputs. Luo et al. [15] proposed an adaptive AI-based DNN model optimized through feature extraction for building energy prediction at the urban level. Also, Wang et al. [16] developed a systems theory-based framework to predict the urban-scale electricity consumption of residential buildings. Sun et al. [17] proposed a data fusion framework integrating heterogeneous traffic data to improve vehicle-matching accuracy and enable applications like fuel consumption estimation and real-time traffic analysis in smart cities. Roth et al. [18] conducted urban building energy benchmarking using U.S. commercial building data (CBECS) and open data. Their findings showed that the critical data to collect in urban areas include building area, property type, and conditioned area. This study highlighted the need for changes in current benchmarking practices, specifically by (1) establishing a transparent benchmarking process and (2) encouraging energy savings through the public disclosure of benchmarking results. Ali et al. [19] proposed a data-driven machine learning approach to predict building energy performance and analyze retrofitted buildings. To address the issue of data scarcity, they conducted parametric simulations based on existing residential building data and evaluated the performance of the machine learning model using synthetic building datasets. Johari et al. [20] utilized UBEM to analyze large-scale energy retrofits and their impact on the power grid for 22,000 residential buildings in Varberg, Sweden. They developed an urban building energy model using Energy Performance Certificate (EPC) data and simulated various retrofit scenarios to assess their outcomes. Velamazán et al. [21] developed a nationwide UBEM approach using XGBoost and AI technologies, predicting energy consumption for 1,264,864 residential buildings across Spain based on building features. By comprehensively considering factors such as total floor area and energy usage, they categorized buildings into energy performance grades ranging from A to G.

UBEM is also utilized as a virtual model in developing urban digital twin frameworks. Weil et al. [22] conducted a systematic review of urban digital twin (UDT) implementation challenges, identifying eight key areas, including data quality, modeling, and governance, to address bottlenecks in sustainable smart city management. Mazzetto [23] analyzed the role of urban digital twins (UDTs) in fostering sustainable smart cities. There are relatively few studies on digital twin (DT) development at the urban scale for advancing UBEM. It is difficult to understand the urban context and employ urban environments as virtual models, and few technological and theoretical foundations have been established. Lastly, there are studies focusing on the application of GPTs for UBEM. Feng et al. [24] proposed CityGPT, a framework that fine-tunes large language models (LLMs) using urban knowledge via the City Instruction dataset, enabling them to solve urban tasks effectively and achieve competitive performance regarding the CityEval benchmark. Bentley et al. [25] proposed an image-to-data approach using large language models combined with domain-specific knowledge to estimate urban environment and building sustainability features and energy consumption, achieving better accuracy than human experts. The accuracy was evaluated based on key metrics, including the average error in building age prediction, the correctness of building- and heating-type identification, the precision of energy source estimation, and the estimation of window types and energy consumption. Table 1 presents a summary of the research on UBEM, DT, and AI agent-based UBEM applications.

One of the core technical strengths of UBEM lies in its ability to integrate a wide range of variables comprehensively, including building types, ages, energy systems, and climatic factors. This capability is essential for accurately assessing building energy performance and establishing scientifically sound energy-saving strategies. With ongoing advancements in technology and extensive research, UBEM continues to evolve, leveraging AI, machine learning, and diverse data analysis techniques to provide new insights into urban building energy consumption. However, traditional UBEM approaches predominantly rely on simulation-based, bottom-up methods that focus on individual buildings [26]. Even with the application of top-down methods, the analyses often remain fragmented and lack a holistic perspective. These limitations are largely due to the multi-source and vast-scale nature of urban data, combined with the absence of scientifically robust and multifaceted analytical approaches. Additionally, despite significant research efforts, the practical application of UBEM remains limited, with fewer implementations compared to its theoretical and experimental developments.

In the context of digital twins (DTs), studies focusing on their urban-scale application are relatively scarce [22,27]. Challenges in this regard include the difficulty of collecting high-quality urban data and integrating heterogeneous data sources. As a result, modeling efforts have predominantly focused on individual buildings rather than achieving comprehensive urban-scale modeling. To overcome these issues, it is crucial to unify dispersed data across buildings, regions, and collection institutions into a cohesive framework and develop models that account for the entire urban building energy system. The reliance on predominantly bottom-up approaches further constrains UBEM’s potential. These methods often face difficulties in handling large-scale data integration and scalability challenges and capturing the complex interdependencies between buildings and their environments. Practically, UBEM approaches based on bottom-up methods are not well suited for use as virtual models within digital twin frameworks. This is because bottom-up approaches heavily depend on detailed building-level data, which are often challenging to collect, integrate, and scale to an urban level. Additionally, these approaches cannot adequately reflect the complex inter-building interactions and dynamic urban contexts, limiting their broader applicability within comprehensive DT frameworks.

Table 1. Summary of the research on UBEM, DTs, and AI agent-based UBEM applications.

Topic	Title	Author (Year)	Content
UBEM and data fusion	“Data fusion analysis applied to different climate change models: An application to the energy consumptions of a building office”	Guarino et al. [14]	Prediction of heating and cooling energy levels
	“Feature extraction and genetic algorithm enhanced adaptive deep neural network for energy consumption prediction in buildings”	Luo et al. [15]	Energy consumption prediction for each feature cluster defined via k-means
	“Prediction and comparison of urban electricity consumption based on grey system theory: A case study of 30 southern China cities”	Wang et al. [16]	A systems theory-based framework developed to predict urban-scale electricity consumption for residential buildings
	“Batch-based vehicle tracking in smart cities: A Data fusion and information integration approach”	Sun et al. [17]	Capturing individual vehicle movements across a traffic network, aiding in the optimization of traffic management and planning in smart cities
	“Examining the feasibility of using open data to benchmark building energy usage in cities: A data science and policy perspective”	Roth et al. [18]	Benchmarking using CBECS and open data, suggesting a transparent benchmarking process
	“Urban building energy performance prediction and retrofit analysis using data-driven machine learning approach”	Ali et al. [19]	Proposing a data-driven machine learning approach for predicting building energy performance and analyzing retrofits
	“Analysis of large-scale energy retrofit of residential buildings and their impact on the electricity grid using a validated UBEM”	Johari et al. [20]	Analysis of large-scale energy retrofits and power grid impacts using UBEM for 22,000 residential buildings in Varberg, Sweden
	“Predicting Energy and Emissions in Residential Building Stocks: National UBEM with Energy Performance Certificates and Artificial Intelligence”	Beltrán-Velamazán et al. [21]	Prediction of energy consumption for 1,264,864 residential buildings across Spain using building features and classification of energy performance grades based on predicted values
Urban digital twins	“Urban Digital Twin Challenges: A Systematic Review and Perspectives for Sustainable Smart Cities”	Weil et al. [22]	Challenges in urban digital twin (UDT) implementation
	“A Review of Urban Digital Twins Integration, Challenges, and Future Directions in Smart City Development”	Mazzetto [23]	Urban digital twin (UDT) integration, challenges, and future directions in smart cities
GPT application for UBEM	“CityGPT: Empowering Urban Spatial Cognition of Large Language Models”	Feng et al. [24]	Analysis of urban task performance using GPT
	“Zero-Shot Building Age Classification from Facade Image Using GPT-4”	Zeng et al. [28]	Age classification of buildings
	“Automated Real-World Sustainability Data Generation from Images of Buildings”	Bentley et al. [25]	Data generation (efficiency and energy data)

Table 1. Summary of the research on UBEM, DTs, and AI agent-based UBEM applications.

Topic	Title	Author (Year)	Content
UBEM and data fusion	“Data fusion analysis applied to different climate change models: An application to the energy consumptions of a building office”	Guarino et al. [14]	Prediction of heating and cooling energy levels
	“Feature extraction and genetic algorithm enhanced adaptive deep neural network for energy consumption prediction in buildings”	Luo et al. [15]	Energy consumption prediction for each feature cluster defined via k-means
	“Prediction and comparison of urban electricity consumption based on grey system theory: A case study of 30 southern China cities”	Wang et al. [16]	A systems theory-based framework developed to predict urban-scale electricity consumption for residential buildings
	“Batch-based vehicle tracking in smart cities: A Data fusion and information integration approach”	Sun et al. [17]	Capturing individual vehicle movements across a traffic network, aiding in the optimization of traffic management and planning in smart cities
	“Examining the feasibility of using open data to benchmark building energy usage in cities: A data science and policy perspective”	Roth et al. [18]	Benchmarking using CBECS and open data, suggesting a transparent benchmarking process
	“Urban building energy performance prediction and retrofit analysis using data-driven machine learning approach”	Ali et al. [19]	Proposing a data-driven machine learning approach for predicting building energy performance and analyzing retrofits
	“Analysis of large-scale energy retrofit of residential buildings and their impact on the electricity grid using a validated UBEM”	Johari et al. [20]	Analysis of large-scale energy retrofits and power grid impacts using UBEM for 22,000 residential buildings in Varberg, Sweden
	“Predicting Energy and Emissions in Residential Building Stocks: National UBEM with Energy Performance Certificates and Artificial Intelligence”	Beltrán-Velamazán et al. [21]	Prediction of energy consumption for 1,264,864 residential buildings across Spain using building features and classification of energy performance grades based on predicted values
Urban digital twins	“Urban Digital Twin Challenges: A Systematic Review and Perspectives for Sustainable Smart Cities”	Weil et al. [22]	Challenges in urban digital twin (UDT) implementation
	“A Review of Urban Digital Twins Integration, Challenges, and Future Directions in Smart City Development”	Mazzetto [23]	Urban digital twin (UDT) integration, challenges, and future directions in smart cities
GPT application for UBEM	“CityGPT: Empowering Urban Spatial Cognition of Large Language Models”	Feng et al. [24]	Analysis of urban task performance using GPT
	“Zero-Shot Building Age Classification from Facade Image Using GPT-4”	Zeng et al. [28]	Age classification of buildings
	“Automated Real-World Sustainability Data Generation from Images of Buildings”	Bentley et al. [25]	Data generation (efficiency and energy data)

The aim of this study was to address the limitations of existing UBEM approaches by introducing the concept of an intelligent urban digital twin (I-UDT), achieved through the integration of artificial intelligence (AI) technology—particularly OpenAI’s emerging generative pre-trained transformer (GPT) [29]—as an AI agent. This concept builds upon and extends F. Tao’s five-dimension digital twin model [30], enhancing the service entity layer through a GPT interface. By leveraging the selected GPT’s powerful capabilities in natural language and machine language processing, as well as its large-scale data analysis proficiency, integration with UBEM allows for the more effective processing and analysis of building energy data. Figure 1 illustrates the technologically complementary relationship between DTs and UBEM, assisted by the GPT. The GPT automatically integrates and analyzes data and information collected from the real world, enabling it to serve as the virtual model of the DT. Moreover, UBEM acts as a virtual model within the DT framework, allowing various elements depicting the city to describe mathematical behaviors within the urban environment. This enables the real-time exchange of data and information between the real world and the virtual model through UBEM. On the other hand, the service functionality of the five-dimension DT facilitates the realization of service stages that were previously underdeveloped in UBEM. The GPT enables effective user interfacing, allowing data and information obtained from UBEM and the DT database to be efficiently delivered to users, such as citizens and policymakers.

This integration facilitates the handling of multi-source data, including urban building energy data, administrative information, unstructured data, and external environmental data, thus overcoming the computational challenges traditionally associated with UBEM and enabling more efficient analysis and visualization. Furthermore, by applying GPT-based UBEM, models can be optimized to provide more precise predictions of building energy consumption characteristics and deliver solutions for urban-wide energy management and optimization. This UBEM approach supports continuous data and information interaction and synchronization among cities, establishing an intelligent urban digital twin (I-UDT) capable of real-time urban energy system analysis and information provision.

1.2. Objective, Novelty, and Contributions

In this study, we establish the concept of a UDT by positioning UBEM as a virtual model within existing DT frameworks. We propose an I-UDT framework, where an AI agent functions as a central processing unit, integrating traditional UBEM technologies with OpenAI’s GPT to implement an intelligent UBEM approach. Intelligent UBEM refers to the GPT-based urban energy model within the I-UDT framework, characterized by the following features:

• Automated data integration: It seamlessly combines fragmented urban building data into a cohesive model, enhancing data accessibility and usability.
• Self-determining analytical tools: It uses a GPT to autonomously select the most suitable analytical tools (or applications) for specific tasks, streamlining the process of data analysis and result derivation.
• User-friendly platform delivery: It increases feasibility and lowers barriers to real-world implementation by offering tailored insights and recommendations to users through an accessible platform format, enhancing decision making for energy efficiency and sustainability.

The proposed I-UDT differs from traditional UDTs in that it enables a dynamic data exchange between virtual urban models and the real world, overcoming limitations in fragmented data handling and actionable insight generation. By leveraging the GPT’s advanced natural language-processing capabilities, the I-UDT synthesizes complex data into user-friendly insights and enhances user interaction through tailored services. It supports predictive analytics, scenario planning, and intelligent feature engineering, making it a transformative tool for sustainable urban development and effective urban energy management. Through four case studies (Cases 1–4) focused on the core technological elements of UBEM—(1) basic data analytics, including integration and preprocessing; (2) energy prediction; (3) feature engineering; and (4) information delivery services for the user—the effectiveness of intelligent UBEM in addressing the limitations of traditional UBEM is demonstrated. Furthermore, this study introduces a framework for intelligent digital twins (I-UDTs), facilitating data and information exchange between GPT-based virtual urban models for UBEM and the real world. Most importantly, this study academically establishes the concept of I-UDT applications, a framework that was previously undefined, while implementing the core technologies of UBEM within it. The evolved I-UDT framework proposed in this study moves beyond merely representing existing urban conditions, as is typical in traditional UBEM approaches, becoming a transformative tool for future urban planning. By enabling the real-time integration and synchronization of diverse urban data, the I-UDT aligns with the ultimate goal of supporting sustainable urban development. It serves as a critical instrument for urban planners, policymakers, and building operators, offering scientifically robust solutions to challenges in urban energy management and sustainability policy development.

2. AI Agent-Based Intelligent Digital Twin (I-DT) for Urban Informatics

2.1. Methodology Behind I-UDTs

The concept of the digital twin, first proposed by Dr. Grieves, has highlighted the potential to replicate physical environments in virtual spaces for monitoring and optimization [31]. Initially, it was a traditional three-dimensional DT model consisting of (1) a physical entity, (2) a virtual entity, and (3) the exchange of data and information. However, as Dr. Grieves recently emphasized, the ultimate goal of digital twins is not merely to replicate physical spaces but to integrate artificial intelligence (AI) technology to create systems capable of automated, active optimization through sophisticated prediction and analysis [32]. This transition involves moving from static representations of physical systems to dynamic, intelligent models that can continuously learn and adapt through real-time data integration, advanced analytics, and predictive capabilities enabled by AI. Building on Dr. Grieves’ intelligent digital twin model and Dr. Tao’s five-dimension digital twin model, in this study, we aim to academically establish and develop an intelligent digital twin model. Specifically, the concepts are applied to develop an I-UDT for the urban building sector. Fei Tao’s five-dimension digital twin model systematically defines the structure and functions of digital twins, consisting of (1) physical space (the actual system), (2) virtual space (digital representation and modeling), (3) data space (data collection and processing), (4) service space (user-centric service provision), and (5) connection space (integration and real-time interaction between dimensions). This model is designed to enable digital twins to reflect the physical world; perform analysis, prediction, and optimization based on data; and provide real-time connectivity and user-focused services. The I-UDT proposed in this study leverages GPT-based natural language processing to enhance data interpretation, automate decision-making processes, and provide actionable insights, transforming the digital twin into an active system capable of addressing complex challenges in urban energy management and sustainability.

Figure 2 illustrates the structure of the GPT-based I-UDT, where UBEM and various technological components (UBEM tools and applications) required for data analysis are integrated with real buildings and a GPT. The I-UDT consists of four main elements:

• User(s): Users are the primary stakeholders at the national or policy level, such as government policymakers or urban energy regulators, who use the system to implement large-scale energy management strategies (for example, a government agency utilizing the I-UDT to optimize energy consumption across urban districts).
• AI agent (GPT): This is the core artificial intelligence engine driving the I-UDT, responsible for data integration, preprocessing, feature engineering, and supporting machine learning models.
• Administrator(s): These are experts or researchers who maintain and update the applications with domain-specific knowledge to align the system with evolving policies and regulations.
• Five-dimension DT, including UBEM as a virtual model and the real world.

The I-UDT proposed in this study defines the service entity as a combination of three elements: (1) applications (UBEM tools), (2) the GPT, and (3) users. Users, such as government policymakers or urban energy regulators, represent the primary stakeholders at the national or policy level. They interact with the GPT system through prompt templates to implement large-scale energy management strategies and urban sustainability policies. For example, a government agency could utilize an I-UDT to optimize energy consumption across urban districts. The GPT serves as the core artificial intelligence engine of the I-UDT, acting as a central processing unit responsible for data integration, preprocessing, feature engineering, and supporting machine learning models, which are essential components of UBEM tools. Administrators, such as experts or researchers, maintain and update the applications with domain-specific knowledge to ensure the system aligns with evolving policies and regulations, allowing the delivery of accurate and relevant outputs. This integration of applications, a GPT, and users within the I-UDT framework leverages the five-dimension DT model, where UBEM serves as a virtual model connected to the real world, facilitating intelligent and dynamic decision-making processes at the national and policy levels.

Figure 3 shows the applications of the I-UDT. In these applications, top-down approaches are employed, encompassing data preprocessing, integration, feature engineering, energy consumption prediction modeling, and optimization. Simultaneously, the bottom-up approach ensures simulation accuracy by calibrating and validating models through detailed building energy performance simulations and comparisons with real data. Examples of bottom-up approach calibration methods include environmental simulations, such as shading indices and solar radiation modeling, which assess the impact of neighboring structures on solar exposure and heating or cooling loads, ensuring that models reflect realistic energy usage patterns. By integrating these two approaches within an application, this framework facilitates the collection of hard-to-obtain urban data and applies them to top-down analytical methods. Additionally, basic urban building administrative information, energy consumption patterns, and environmental data can be analyzed, visualized, and stored in the UBEM and DT database, providing users with an information delivery service. Through the natural-language-processing capabilities, advanced analytics, and user-friendly interface of the GPT, the framework ensures the continuous updating of the urban energy model and the database as well as the exchange of data and information between real buildings and virtual models. This enables the implementation of an intelligent urban digital twin (I-UDT). These applications (UBEM tools) focus on the modeling, simulation, and analysis of urban building energy consumption, distinguishing them from other framework components. These tools are designed to integrate and process large-scale urban building data to perform tasks such as energy prediction, feature engineering, and energy efficiency evaluation. In contrast, other framework components, such as the DT database, virtual urban model, and real-world elements, play supporting roles by facilitating data collection, storage, and communication to support the three elements of the service entity. UBEM tools concentrate on the core tasks of UBEM, providing actionable insights and tailored energy management strategies, directly contributing to the goals of the I-UDT framework. Additionally, as the service is divided into three key components—users, the GPT, and applications—the framework emphasizes service-centric functionality, aligning with the main objective of realizing the I-UDT proposed in this study.

Administrators develop the GPT system using pre-generated ontologies and domain-specific information, including documents, expert programming scripts, and newly derived building information model data from the GPT, enabling domain-specific analysis. This approach also allows for the optimization of existing applications. The database integrates these domain-specific resources with multi-source data, acting as a comprehensive repository that stores diverse datasets, such as GIS, climate, vegetation, building features, and real-time energy consumption datasets. The GPT is responsible for interpreting user prompts and autonomously interfacing with the administrator, the DT database, applications, and the urban energy model within the five-dimension DT model including UBEM and the real world. Furthermore, the GPT enables automation and facilitates the seamless execution of I-UDT services based on the provided prompts. Additionally, the digital twin model can serve as a building energy analysis platform accessible to citizens, policymakers, and other stakeholders.

2.2. Materials

In this study, we examine three cases to demonstrate the application of the GPT-based I-UDT, using real-world data from 3500 buildings in Seoul, South Korea. The datasets consist of administrative building data, which provide detailed information about building types, sizes, and construction years; energy consumption data, which track the energy usage patterns of various buildings over time, deriving from the Korea Real Estate Board (REB); and outdoor temperature data from the Korea Meteorological Administration (KMA), which are crucial for evaluating the impact of climatic conditions on building energy performance. To ensure consistency and accuracy, a column definition sheet was utilized to standardize data structures across different sources. Additionally, the energy dataset does not extend to the level of individual household meters; instead, it consists of monthly electricity energy data. Therefore, the data used in this study pose minimal privacy concerns. These open datasets were provided by DataNet, a national research project launched in 2023 and funded by the Korea Agency for Infrastructure Technology Advancement (KAIA) [33]. DataNet was instrumental in supporting the collection and integration of multi-source data, encompassing both structured and unstructured datasets related to urban buildings in South Korea. Furthermore, GPT-4o [29] was employed to explore and implement GPT-based UBEM applications in this study. This tool was utilized to process administrative and technical documents as well as integrate and analyze multi-source data. This use of GPT-4o was critical for ensuring accurate and reliable results, and all GPT outputs were incorporated into the study results as generated, with only minor edits, such as omissions of intermediate outputs. Also, due to the inherent randomness of GPT responses, each task was repeated five times per case, and the results deemed most useful from the user’s perspective were selected and summarized.

3. Case Study

Section 3.1, Section 3.2, Section 3.3 and Section 3.4 present the processes and results of the key functionalities of UBEM within the GPT-based I-DT for urban informatics, focusing on data integration and preprocessing, energy analysis and prediction, and feature engineering. Notably, the GPT was pre-informed about the concept and structure of the GPT-based I-DT for urban building informatics as well as its role within the model, enabling it to better understand the context of the prompts. Figure 4 shows the overall process of the case study.

3.1. Basic Urban Building Data Analytics (Case 1)

Analyzing data from diverse perspectives and generating intuitive visualizations based on multi-source urban datasets are fundamental techniques and essential components of UBEM [34]. Case 1 is a case study demonstrating whether a GPT can perform basic urban building data analysis. Figure 5 shows the results of a basic urban building data analysis conducted by GPT-4o. As shown in the figure, data analysis and visualization were derived based on user prompts. The user prompts were designed to analyze energy usage patterns and distributions, energy usage distribution according to the gross floor area, and energy distribution based on outdoor temperatures. Additionally, the prompts included details about the provided datasets, which consisted of a column definition guide, administrative data, and energy and outdoor temperature data for 2018 and 2019. These years were chosen to avoid the outdated nature of pre-2018 data and the potential impacts of COVID-19 on building energy patterns in post-2019 data, ensuring the model’s accuracy.

As illustrated in GPT-4o’s outputs in Figure 5, the system interpreted the given data structure and described the types of analyses to perform. By representing the energy distribution for 2018 and 2019 using histograms and line graphs, we observed that the average energy consumption and patterns were similar. Furthermore, the building use types were aggregated and visualized, showing that multi-family residential buildings were the most common (with 1995 buildings), followed by type 1 neighborhood facilities (with 483 buildings), single-family houses (with 397 buildings), type 2 neighborhood facilities (with 391 buildings), office buildings (with 79 buildings), and others (amounting to 155 buildings). Additionally, the average outdoor temperature and average energy consumption for 2018 and 2019 were presented as line graphs. Finally, the correlation between the gross floor area and energy usage was analyzed and visualized, revealing that a greater gross floor area generally corresponded to higher average energy consumption.

These analytical results can be stored in the DT database and provided based on user requests. They can be visualized through virtual city models and UBEM while also being integrated with applications to support tasks such as energy prediction, benchmarking, and clustering through unsupervised learning analysis. While the overall analysis results provided a macro-level understanding of the given data, more detailed results are still lacking, such as the correlation between outdoor temperature and energy consumption or between the gross floor area and EUI (energy use intensity). Moreover, the explanations accompanying the graphs are insufficient. These limitations could be addressed by providing the GPT with additional domain knowledge or more specific prompts. A case where domain knowledge was provided to calculate EUI and analyze EUI by building scale is introduced in Case 4. This case, discussed in Section 3.4, demonstrates the effective incorporation of domain-specific metrics through structured prompts and supporting resources.

3.2. Building Energy Prediction at the Urban Scale (Case 2)

Predicting energy consumption for buildings on a city-wide scale plays a vital role in optimizing energy efficiency, minimizing environmental impacts, and formulating urban strategies to achieve these objectives [35]. In Case 2, as shown in Figure 6, the objective was to evaluate whether the GPT could utilize the data provided to predict city-level building energy consumption and analyze the results derived from the model. Similar to Case 1, the dataset included a column definition guide, administrative data, and energy consumption and outdoor temperature data for 2018 and 2019. Notably, as indicated in the user prompt within Figure 6, Case 2 introduced a scenario for energy prediction modeling. The prompt instructed the GPT to predict energy consumption for 2019 using electricity consumption, outdoor temperature, and administrative data from 2018. Additionally, it required the evaluation and visualization of the model’s accuracy, specifying metrics such as R², CVRMSE, and MAE.

As shown in the GPT outputs in Figure 6, the GPT utilized a Random Forest regression model to predict electricity consumption for 2019, achieving high accuracy, with an R² value of 0.9896, a CVRMSE of 0.3616, and an MAE of 0.00077. The choice of Random Forest as the predictive model was initially made by the GPT as we aimed to entrust the entire process of prediction modeling, including data integration, model selection, and hyperparameter optimization, to the GPT. Also, following the initial user prompt, a subsequent prompt requested the GPT to provide details about the structure of the model structure, the use of data for modeling, the calculation of the building-specific CVRMSE, and the visualization of this value’s distribution. The GPT provided a comprehensive explanation in response to the user’s request by addressing the key aspects of the process. First, it identified the input features required for the analysis. These included the monthly electricity consumption data for 2018, the gross floor area of the buildings, and the primary use types. These elements were chosen as the main variables to predict energy consumption. Next, in the data-preprocessing step, the GPT filtered out buildings with gross floor areas less than or equal to zero since these were not valid for analysis. After this, it normalized the remaining data by scaling the values between zero and one. Additionally, it combined multiple data sources by aligning them with building IDs, ensuring it had a cohesive dataset for modeling. For the modeling method, the GPT employed a Random Forest regression model. To validate the performance of the model, it split the dataset into two parts: 80% of the data were used for training the model, while 20% were used for testing its predictions. In the modeling scenario, the GPT used energy data from 2018 to predict electricity consumption for 2019. It calculated CVRMSE values for individual buildings to evaluate the model’s accuracy and created a histogram to display the distribution of these values. Lastly, the GPT provided a Python code script that summarized the steps of the analysis. This script was designed to be easily implemented in a local environment, allowing users to apply the same methodology for similar tasks. This feature made the process adaptable and applicable to practical use cases.

Although the GPT utilized a relatively simple model, the results demonstrated its ability to predict city-level building energy consumption with high accuracy when past data are available. It highlighted that the performance improves when past and present energy patterns are similar. However, to address cases where patterns differ or energy consumption data are unavailable, models that rely solely on the unique features of buildings should also be developed. To achieve this goal, applications for extracting and optimizing building features must be incorporated into the I-UDT framework.

The subsequent case, Case 3, focuses on an application designed to extract and optimize building features, specifically the heating and cooling areas, and presents the results of this analysis.

3.3. Urban Building Feature Engineering (Case 3)

Feature engineering plays a crucial role in UBEM and becomes even more critical when leveraging limited data provided by users to deliver high-quality insights [36]. In the I-UDT, it is also essential for enhancing service delivery as it enables the extraction and optimization of meaningful information from sparse or heterogeneous datasets, ensuring that users receive accurate and actionable outputs tailored to their specific needs. The goal of Case 3 was to derive a new building feature, the heating and cooling area, using existing administrative building data and optimize it to enhance its explanatory power for energy consumption, making it more effective for energy prediction modeling.

As shown in Figure 7, first, the heating and cooling area ratio was determined by evaluating the proportion of conditioned spaces within each building. This process, as outlined in the prompt, involved summing the areas (in m²) designated as conditioned spaces using floor-by-floor usage and area data for individual buildings. The GPT was tasked with identifying whether a given space was conditioned based on the information provided. The GPT utilized the administrative data provided, which included floor-by-floor building usage and area information, along with the user prompt detailing the method for calculating heating and cooling areas, to derive data on the heating and cooling area ratios for individual buildings. Afterwards, a subsequent user prompt requested a comparison between the gross floor area and the heating and cooling area to determine which had a stronger correlation with energy consumption.

The GPT visualized the comparison using scatter plots and calculated the R-value for each variable. While the difference was slight, the heating and cooling area demonstrated a stronger correlation with energy consumption. Based on further prompts, the GPT then made adjustments to the heating and cooling area to enhance its relationship with electricity consumption. This was achieved by applying a correction factor, ensuring that the heating and cooling area better aligned with energy usage. Ultimately, the GPT generated the adjusted heating and cooling area data alongside the applied correction factors. This application can be integrated into urban building energy prediction modeling and has the potential to improve model accuracy. It represents a step towards optimizing features for more precise and reliable energy predictions in the future. However, certain limitations exist. The generated correction factor is derived from the relationship between the heating and cooling area and energy consumption, which inherently includes uncertainty and does not fully account for various environmental and regional conditions. Therefore, it is necessary to verify whether the corrected heating and cooling area closely aligns with the actual heating and cooling area of buildings. However, since there are no ground truth values for heating and cooling areas in South Korea’s urban data, the correction factor can be indirectly validated by using it as an input variable in energy prediction models or directly verified through field surveys and measurements.

3.4. User-Oriented Information Delivery Services (Case 4)

Figure 8 illustrates the application of user-centered information delivery services in Case 4. The user, through a simple prompt containing the phrase “how many buildings in the dataset I provided exceed the standards”, sought to identify the number of buildings in the dataset that exceeded the energy usage standards. The standards presented in this study are national legal standards applicable across South Korea, with energy weighting factors provided for 17 regions, including Daejeon (weight is 1.0), which is in the center of the country. We conducted a case study on buildings located in Seoul, where the energy weighting factor is 1.35. Accordingly, the GPT evaluated the EUI value by multiplying it by 1.35.

Based on the input dataset and a legal document on energy benchmarks provided in advance, the GPT analyzed the data structure and the content of the document. The output revealed that a total of 2915 buildings in the dataset exceeded the energy usage standards specified in the document. Additionally, the user followed up with a prompt requesting a visual representation of the distribution of buildings exceeding the standards arranged by floor area range. In response, the GPT created a bar graph illustrating the number of buildings within each floor area range. These results demonstrate a high level of accuracy, with a 99.32% match compared to manually aggregated results.

This case highlights the efficiency and reliability of the I-UDT’s core service functionality, which includes data analysis and visualization based on user requests. Specifically, the applications incorporated into the I-UDT service entities automatically extract quantitative data on buildings exceeding energy usage standards and present them through intuitive visualizations, delivering clear and meaningful insights to users. These capabilities not only enhance urban energy management but also provide data-driven insights that can be practically applied in policymaking processes, further emphasizing the value of the I-UDT as a service. The analysis and insights generated by the GPT (AI) based on existing documents and real-world data are evidence-based outputs that can comprehensively explain the current state of urban buildings and serve as valuable evidence for updates tailored to changing conditions. Therefore, these outputs are highly suitable for practical policy applications. Additionally, if users (particularly policymakers) utilize I-UDT and urban building ontologies as tools for urban management and direct decision making, the framework can be effectively applied in real-world contexts. For instance, it could be used for updating legal standards or as a preliminary tool for retrofit planning. Since the I-UDT strongly prioritizes service realization, the development of ontologies of existing urban building data databases and new features generated by the GPT and applications are essential. This ontology will robustly support GPTs in delivering their services. Concrete methods for achieving this include developing urban data ontologies and implementing user-specific systems tailored to potential users’ needs.

4. Conclusions

This study establishes the academic concept of a GPT-based I-UDT and demonstrates the technical application of this framework through case studies. By deconstructing the service entities of the traditional five-dimensional DT model [37] into three components—(1) applications, (2) a GPT, and (3) users—the service framework within the UDT is further clarified, with a particular emphasis on the GPT’s role in providing technical support. Through this, the UBEM functions as a virtual model within the DT framework, and the DT enhances functionalities for which UBEM was technically limited, thereby emphasizing the capabilities of UDTs. In Case 2, we achieved a good energy prediction performance, with a CVRMSE value of 0.3616% and an R² value of 0.9896. The GPT independently performed all modeling processes, including data preprocessing, feature selection, and hyperparameter selection. Additionally, in Case 4, legal documents, energy data, and total floor area data were utilized to detect buildings exceeding legal energy usage standards, achieving a detection accuracy of 99.32%.

The proposed I-UDT framework represents a paradigm shift from traditional digital twins (DTs) by integrating advanced AI capabilities, particularly through the incorporation of a GPT, into UBEM. This paradigm shift is further highlighted by the novel methodologies introduced in this study. Unlike traditional approaches, wherein humans manually conduct analyses using existing tools or code scripts, the GPT autonomously selected and performed the necessary analyses, streamlining the entire process. Moreover, the GPT independently conducted all stages of predictive modeling, including preprocessing, hyperparameter selection, and result analysis. These advancements mark a significant departure from conventional UBEM workflows and offer substantial benefits for future applications. A particularly innovative feature of the I-UDT framework is its ability to process legal documents on energy consumption standards. By providing the GPT with such documents, it was possible to identify buildings exceeding energy thresholds based on their total floor area—entirely without additional coding. This capability not only simplifies regulatory evaluation but also lays the groundwork for leveraging additional passive data on buildings to propose tailored solutions for identified structures.

By combining the service-oriented strengths of the five-dimensional DT model, the domain expertise of UBEM tools, and the technological synergy offered by a GPT, the I-UDT framework has the potential to establish itself as a transformative tool. Its ability to integrate real-time monitoring, intelligent analysis, and practical solution proposals suggests it could surpass the limitations of traditional urban building energy models. Consequently, the I-UDT framework shows promise in enhancing the practicality and feasibility of monitoring urban building energy use and delivering actionable insights for sustainable urban development.

4.1. Limitations

This study has several limitations. First, the case studies relied on a limited dataset, with the analyses having been conducted on approximately 3500 buildings due to the GPT’s computational time-out issues. Using the GPT API can help address time-out issues by offering higher token limits, longer processing times, and support for task segmentation and external tool integration. Additionally, optimizing prompts and integrating external tools can enhance efficiency and scalability for large datasets. To build a more detailed analysis and larger-scale urban building energy model, it is essential to incorporate diverse datasets that explain urban energy dynamics. These datasets should include GIS data, satellite imagery, vegetation data, and administrative and environmental datasets as well as region-specific data and building codes. Such integration would enhance the framework’s adaptability and transferability to different urban contexts.

Another limitation lies in the service delivery framework of the I-UDT. The inputs used in this study were limited to simple prompts and data files in formats such as “.csv” and “.pdf”. To address this, future advancements should aim to develop a comprehensive urban ontology tailored to urban planning and energy modeling. This urban building ontology should incorporate data and contextual knowledge pertaining to three key areas: retrofitting diagnostics for existing buildings, energy prediction for newly constructed zero-energy buildings, and real-time updates on rapidly changing environmental conditions and climate data. By enabling the GPT to leverage such dynamic and evolving ontologies, it can provide more accurate and contextually aligned outputs, meeting user requirements and supporting real-world applications.

Moreover, integrating real-time data from urban sensors remains a critical challenge in establishing a truly dynamic and bidirectional interaction between the GPT and the DT database within the I-UDT framework. While this study primarily relied on pre-existing static datasets, future iterations should explore how urban IoT networks and sensor data can be effectively incorporated. Additionally, the real-time data collected through these sensors can be integrated based on the previously mentioned urban building ontology, enabling continuous updates and enhancing the GPT’s ability to provide rapid responses within the integrated service framework. This integration would facilitate real-time decision making, adaptive energy optimization, and improved responsiveness to fluctuating environmental conditions, ultimately making the I-UDT framework more robust and practical for real-world applications.

Furthermore, while the I-UDT framework was developed using data from Seoul, its core methodology is designed to be transferable to other regions, such as the USA or the UK. However, challenges may arise due to regional variations in data availability, building codes, and climate conditions. These challenges can be addressed by customizing urban ontology to include region-specific data and legal standards. Future research should focus on exploring such region-specific adaptations to ensure the broad applicability and scalability of the proposed framework.

4.2. Future Directions

Future research should also expand the DT database beyond simple data and building features and include causal relationships between datasets, urban building ontologies, and comprehensive metadata [38]. This enhancement would improve the framework’s ability to optimize applications and provide users with high-quality, real-time insights and predictive analytics. An advanced I-UDT would not only represent existing urban conditions, as seen with traditional UBEMs, but also actively analyze and forecast urban energy patterns, improving decision making for future urban planning. Moreover, integrating real-time data from urban sensors will be a crucial step in advancing the I-UDT framework toward a more dynamic and interactive system. At the national level, data measured through meters may be subject to significant time gaps from collection to integration and utilization due to various stakeholder interests. Therefore, to enable real-time data collection, alternative measurement approaches, such as IoT sensors and smart meters, should be explored. Another approach could involve leveraging real-time energy consumption prediction models, wherein the predicted data are continuously collected and used to interact with the GPT for real-time service applications.

Ultimately, for the future advancement of I-UDT, researchers should aim to develop a fully interactive, AI-powered decision support system that bridges the gap between static urban data and dynamic urban processes. By incorporating real-time data streams, predictive analytics, and enhanced ontological frameworks, the I-UDT can become a critical tool for sustainable urban planning, climate-responsive energy management, and data-driven policy development.