AI-Driven Multi-Objective Optimization and Decision-Making for Urban Building Energy Retrofit: Advances, Challenges, and Systematic Review

Abstract

Urban building energy retrofit (UBER) is a critical strategy for advancing the low-carbon and climate-resilience transformation of cities. The integration of machine learning (ML), data-driven clustering, and multi-objective optimization (MOO) is a key aspect of artificial intelligence (AI) that is transforming the process of retrofit decision-making. This integration enables the development of scalable, cost-effective, and robust solutions on an urban scale. This systematic review synthesizes recent advances in AI-driven MOO frameworks for UBER, focusing on how state-of-the-art methods can help to identify and prioritize retrofit targets, balance energy, cost, and environmental objectives, and develop transparent, stakeholder-oriented decision-making processes. Key advances highlighted in this review include the following: (1) the application of ML-based surrogate models for efficient evaluation of retrofit design alternatives; (2) data-driven clustering and classification to identify high-impact interventions across complex urban fabrics; (3) MOO algorithms that support trade-off analysis under real-world constraints; and (4) the emerging integration of explainable AI (XAI) for enhanced transparency and stakeholder engagement in retrofit planning. Representative case studies demonstrate the practical impact of these approaches in optimizing envelope upgrades, active system retrofits, and prioritization schemes. Notwithstanding these advancements, considerable challenges persist, encompassing data heterogeneity, the transferability of models across disparate urban contexts, fragmented digital toolchains, and the paucity of real-world validation of AI-based solutions. The subsequent discussion encompasses prospective research directions, with particular emphasis on the potential of deep learning (DL), spatiotemporal forecasting, generative models, and digital twins to further advance scalable and adaptive urban retrofit.

1. Introduction

In dense urban areas dominated by aging building stocks, Urban Building Energy Retrofit (UBER) has become a crucial strategy for decarbonization and climate resilience, especially in city centers and old districts facing outdated facades, poor indoor comfort, rising energy demand, and evolving functional needs [1,2]. The UBER process entails complex trade-offs among energy performance, occupant comfort, environmental impact, and economic feasibility, requiring urban-scale energy modeling with high spatial-temporal granularity [3,4]. Urban Building Energy Modeling (UBEM), essential for planning, must account for heterogeneous stocks, urban interactions (e.g., mutual shading, heat island effects), and diverse occupant behaviors—factors whose omission may lead to energy savings underestimation of up to 7% [1]. Yet, limited high-resolution data on geometry, usage patterns, and operations remains a key barrier [3,4,5]. To overcome these challenges, recent UBEM workflows integrate AI-enhanced surrogate models and ML techniques such as ANN, RF, and XGB, which accelerate simulation, predict energy use, and optimize retrofits with minimal physical inputs [6,7,8,9,10]. ML also supports MOO across energy, cost, and comfort while enabling district-scale forecasting, occupant modeling, and smart grid integration [11,12,13]. In this shift from isolated retrofits to systemic solutions, the convergence of ML, GIS, and UBEM offers scalable, cost-effective pathways toward climate-neutral cities [14,15,16].

Over the past two decades, UBEM and retrofit analysis have been extensively reviewed. As summarized in Table 1, 22 relevant review articles published between 2009 and 2025 span a wide range of themes, including simulation platforms, archetype calibration, ML-based forecasting, and retrofit decision-making. These works can be classified into five primary types: foundational UBEM modeling reviews, information modeling and simulation tool surveys, ML-based energy prediction frameworks, retrofit strategy evaluations, and hybrid or interdisciplinary reviews.

Swan and Ugursal [2] and Reinhart and Cerezo-Davila [14] established the early foundations of UBEM through bottom-up modeling approaches and energy archetype definitions. Ang et al. [17] reviewed modeling techniques and building stock aggregation strategies. At the same time, Salvalai et al. [16] and Malhotra et al. [15] provided tool-based taxonomies involving CitySim, TEASER, and UMI, with emphasis on simulation workflows and GIS/BIM data integration. Johari et al. [18] offered a comparative analysis of existing UBEM platforms, highlighting usability and scope limitations. Despite their value in systematizing the field, these reviews often lacked direct links to intelligent decision-making or retrofit strategy optimization.

With the emergence of urban-scale data, researchers began to explore the integration of ML into UBEM. Fathi and Mahdavinejad [8], Manandhar et al. [19], and Zhang et al. [20] reviewed forecasting algorithms such as ANN, RF, and XGB for predicting building energy use. Mousavi et al. [21], Li et al. [22], and Ferrando et al. [23] investigated ML applications for energy retrofit optimization. At the same time, Kong et al. [24] focused on physics-informed neural networks for simulation efficiency. Kristensen and Hedegaard [25] emphasized the role of flexibility potential in neighborhood-scale UBEM. Wei et al. [26] synthesized DL models for occupant behavior prediction and control strategy generation. These studies advanced methodological diversity but often failed to unify ML capabilities with retrofit prioritization or policy applications.

Retrofit-focused reviews have also gained momentum. Suppa and Ballarini [27], Shu and Zhao [28], and Ahmad et al. [29] evaluated multi-criteria decision-making (MCDM), cost-benefit analysis, and environmental indicators for building stock retrofit planning. Abbasabadi and Ashayeri [30] incorporated socio-behavioral uncertainty into retrofit potential assessments. Xu et al. [31] and Mondal et al. [32] addressed climate adaptation, while Mauree et al. [33] reviewed urban-scale sustainability assessment methods. Nevertheless, these works treated ML as an ancillary tool rather than a strategic core of retrofit workflows.

Table 1. Summary of review articles on urban building energy retrofit and modeling.

	Author	Year	Review Type	ML	UBEM	Retrofit	Key Contributions
[2]	Swan & Ugursal	2009	Foundational UBEM Review	×	✓	×	Early archetype-driven simulation of residential energy use; foundation for bottom-up UBEM
[14]	Reinhart & Cerezo Davila	2016	Pioneering UBEM Field Review	×	✓	×	First field-wide UBEM review; scoped archetypes, scales, limitations and outlook
[25]	Kristensen & Hedegaard	2018	Archetype Calibration Review	×	✓	△	Developed calibration workflow for building archetypes; useful for UBEM parameter tuning
[26]	Wei et al.	2018	UBEM Simulation and Framework Integration	✓	✓	×	Combined ML and UBEM simulation; discussed real-time forecasting and policy alignment
[29]	Ahmad et al.	2018	Energy Equity and Urban Energy	✓	✓	×	Reviewed urban energy equity with ML and socio-demographic modeling; emerging UBEM link
[33]	Mauree et al.	2019	Climate Adaptation Review	×	✓	×	Assessment methods for climate-responsive cities; UHI and outdoor-indoor coupling focus
[30]	Abbasabadi & Ashayeri	2019	Urban Energy Systems Review	×	✓	×	Energy system modeling at urban scale; partial UBEM scope with techno-economic emphasis
[17]	Ang et al.	2020	UBEM Use Case Review	×	✓	×	Reviewed practical applications of UBEM; focused on policy and planning use cases
[8]	Fathi et al.	2020	ML for Urban Energy Forecasting	✓	✓	×	Systematic ML method review for urban energy forecasting; highlighted DL and ensemble models
[18]	Johari et al.	2020	Tool-Centered UBEM Review	×	✓	×	Reviewed UBEM tools (CitySim, TEASER, etc.); simulation comparatives without ML discussion
[23]	Ferrando et al.	2020	LCA-Oriented Urban Retrofit Review	△	✓	✓	Mapped LCA models with retrofit strategies in urban context; focused on emissions tracking
[15]	Malhotra et al.	2022	GIS/BIM Workflow Review	×	✓	×	Classified information modeling approaches and toolchains for UBEM
[22]	Li et al.	2023	Hybrid ML + Physics Modeling Review	✓	✓	×	Combines physical simulation with DL for urban-scale prediction and design optimization
[21]	Mousavi et al.	2023	ML for Retrofit Optimization	✓	×	✓	Summarized ML applications for energy prediction and optimization in retrofitting
[24]	Kong et al.	2023	General UBEM Framework Review	×	✓	×	Introduced a structured UBEM typology; emphasized integration and classification of approaches
[27]	Suppa & Ballarini	2023	Multi-Criteria Retrofit Review	△	✓	✓	Reviewed retrofit scenarios with environmental and economic trade-offs; limited ML framing
[28]	Shu & Zhao	2023	Embodied Carbon Stock Modeling	×	×	✓	Stock-level carbon modeling; relevant to retrofit lifecycle assessments, not UBEM
[19]	Manandhar et al.	2023	Urban Energy Use Method Review	×	✓	×	Tool-based review of urban energy modeling; partial UBEM; focused on simulation methods
[16]	Salvalai et al.	2024	Simulation Tool Comparison	×	✓	×	Compared UBEM platforms and simulation tools; ML integration not emphasized
[20]	Zhang et al.	2024	Data-Driven UBEM Forecasting	✓	✓	×	Analyzed ML methods in UBEM from prediction to optimization; lacks socio-technical integration
[32]	Mondal et al.	2024	Data-Driven UBEM under Extreme Heat	✓	✓	×	Focused on data-driven UBEM in extreme heat; emphasized ML method adaptation and gaps
[31]	Xu et al.	2024	District Energy Modeling Comparison	×	△	×	Compared district energy models; UBEM considered as one of several simulation layers
	This Study…	2025	ML-focused Retrofit Strategy Framework	✓	✓	✓	Builds an ML-centered framework for optimizing UBER strategies, bridging prediction, MOO, and scalable urban planning

Note: ✓ = covered; × = not covered; △ = partially or indirectly discussed.

Table 1. Summary of review articles on urban building energy retrofit and modeling.

	Author	Year	Review Type	ML	UBEM	Retrofit	Key Contributions
[2]	Swan & Ugursal	2009	Foundational UBEM Review	×	✓	×	Early archetype-driven simulation of residential energy use; foundation for bottom-up UBEM
[14]	Reinhart & Cerezo Davila	2016	Pioneering UBEM Field Review	×	✓	×	First field-wide UBEM review; scoped archetypes, scales, limitations and outlook
[25]	Kristensen & Hedegaard	2018	Archetype Calibration Review	×	✓	△	Developed calibration workflow for building archetypes; useful for UBEM parameter tuning
[26]	Wei et al.	2018	UBEM Simulation and Framework Integration	✓	✓	×	Combined ML and UBEM simulation; discussed real-time forecasting and policy alignment
[29]	Ahmad et al.	2018	Energy Equity and Urban Energy	✓	✓	×	Reviewed urban energy equity with ML and socio-demographic modeling; emerging UBEM link
[33]	Mauree et al.	2019	Climate Adaptation Review	×	✓	×	Assessment methods for climate-responsive cities; UHI and outdoor-indoor coupling focus
[30]	Abbasabadi & Ashayeri	2019	Urban Energy Systems Review	×	✓	×	Energy system modeling at urban scale; partial UBEM scope with techno-economic emphasis
[17]	Ang et al.	2020	UBEM Use Case Review	×	✓	×	Reviewed practical applications of UBEM; focused on policy and planning use cases
[8]	Fathi et al.	2020	ML for Urban Energy Forecasting	✓	✓	×	Systematic ML method review for urban energy forecasting; highlighted DL and ensemble models
[18]	Johari et al.	2020	Tool-Centered UBEM Review	×	✓	×	Reviewed UBEM tools (CitySim, TEASER, etc.); simulation comparatives without ML discussion
[23]	Ferrando et al.	2020	LCA-Oriented Urban Retrofit Review	△	✓	✓	Mapped LCA models with retrofit strategies in urban context; focused on emissions tracking
[15]	Malhotra et al.	2022	GIS/BIM Workflow Review	×	✓	×	Classified information modeling approaches and toolchains for UBEM
[22]	Li et al.	2023	Hybrid ML + Physics Modeling Review	✓	✓	×	Combines physical simulation with DL for urban-scale prediction and design optimization
[21]	Mousavi et al.	2023	ML for Retrofit Optimization	✓	×	✓	Summarized ML applications for energy prediction and optimization in retrofitting
[24]	Kong et al.	2023	General UBEM Framework Review	×	✓	×	Introduced a structured UBEM typology; emphasized integration and classification of approaches
[27]	Suppa & Ballarini	2023	Multi-Criteria Retrofit Review	△	✓	✓	Reviewed retrofit scenarios with environmental and economic trade-offs; limited ML framing
[28]	Shu & Zhao	2023	Embodied Carbon Stock Modeling	×	×	✓	Stock-level carbon modeling; relevant to retrofit lifecycle assessments, not UBEM
[19]	Manandhar et al.	2023	Urban Energy Use Method Review	×	✓	×	Tool-based review of urban energy modeling; partial UBEM; focused on simulation methods
[16]	Salvalai et al.	2024	Simulation Tool Comparison	×	✓	×	Compared UBEM platforms and simulation tools; ML integration not emphasized
[20]	Zhang et al.	2024	Data-Driven UBEM Forecasting	✓	✓	×	Analyzed ML methods in UBEM from prediction to optimization; lacks socio-technical integration
[32]	Mondal et al.	2024	Data-Driven UBEM under Extreme Heat	✓	✓	×	Focused on data-driven UBEM in extreme heat; emphasized ML method adaptation and gaps
[31]	Xu et al.	2024	District Energy Modeling Comparison	×	△	×	Compared district energy models; UBEM considered as one of several simulation layers
	This Study…	2025	ML-focused Retrofit Strategy Framework	✓	✓	✓	Builds an ML-centered framework for optimizing UBER strategies, bridging prediction, MOO, and scalable urban planning

Note: ✓ = covered; × = not covered; △ = partially or indirectly discussed.

Despite notable progress, most existing reviews remain fragmented—focusing on UBEM modeling, ML forecasting, or retrofit policy in isolation—without offering an integrated perspective. In contrast, this review proposes a comprehensive framework that positions ML as a predictive engine and a strategic enabler throughout the retrofit lifecycle, from data acquisition and urban energy profiling to performance forecasting, MOO, and policy-relevant decision-making. Emphasizing the convergence of physics-based simulation and data-driven modeling, this work provides a unified reference for researchers, policymakers, and practitioners seeking scalable, climate-adaptive, and cost-effective strategies to accelerate the transition toward carbon-neutral and resilient cities.

This review addresses a critical gap in the literature by systematically synthesizing advances at the intersection of AI, ML, and UBER. The specific objectives are fourfold: First, to provide a structured taxonomy and classification of AI-driven methods and optimization objectives in UBER, spanning urban morphology, envelope, systems, and district-level strategies. Second, to benchmark recent progress in ML/DL-based MOO frameworks, including surrogate modeling, clustering, scenario generation, and integration with physical simulation platforms. Third, to highlight emerging trends and methodological frontiers—such as DL, spatiotemporal models, generative approaches, and XAI—for adaptive, scalable retrofit planning. Fourth, to identify persistent challenges regarding model transferability, data heterogeneity, and cross-scale validation, and to propose future research directions for robust and context-aware AI-enhanced UBER workflows.

This paper is organized as follows: Section 2 presents the review methodology and outlines the search strategy and data collection processes. Section 3 systematically reviews recent advances in AI-driven MOO frameworks for UBER, including the classification of retrofit strategies, benchmarking of ML/DL algorithms, and representative applications. Section 4 synthesizes emerging trends and future research frontiers, highlighting the roles of DL, spatiotemporal models, generative approaches, and explainable AI. Section 5 discusses outstanding challenges, methodological limitations, and prospects for scalable, robust, and adaptive UBER workflows. Finally, Section 6 concludes with a summary of key findings and recommendations for future research.

2. Methodology and Results

To ensure comprehensive literature coverage, we conducted a systematic search of both the Web of Science Core Collection and Scopus databases, targeting publications from January 2015 to July 2025. The search strategy was structured around two dimensions: methodological approaches—including artificial intelligence (machine learning, deep learning, reinforcement learning, heuristic and metaheuristic algorithms)—and research targets focused on urban-scale energy modeling and retrofit (e.g., urban building energy, building retrofit, energy renovation, low-carbon optimization). The combined search strings employed Boolean logic [TS = (Method) AND (Target)], with filters applied to include only English-language journal articles and to exclude preprints and review articles. This process initially yielded 979 articles (Table 2). After title and abstract screening, 238 articles were retained. Full-text review further narrowed the sample to 67 studies directly addressing AI- or heuristic-based optimization for urban building energy retrofit (Table A1).

Table 2. List of search keywords and screening conditions.

Category	Details
Method (TS =)	(Machine learning OR Deep learning OR Deep Reinforcement Learning OR Neural Network OR Artificial Intelligence OR Data-driven OR Heuristic algorithm OR Metaheuristic OR Genetic algorithm OR Particle swarm optimization OR Ant colony optimization OR Evolutionary algorithm OR Surrogate modeling OR Optimization algorithm)
Target (TS =)	(Urban building energy OR Urban building energy model OR Urban building energy modeling OR Urban energy simulation OR Urban energy system OR Building retrofit OR Building renovation OR Building refurbishment OR Energy retrofit OR Low-carbon retrofit OR Building energy optimization)
Search logic	TS = (Method) AND (Target)
Database	Web of Science Core Collection; Scopus
Filters	Exclude: Preprint Citation Index, review articles
Search period	January 2015–July 2025
Document type	Article
Language	English
Search date	10 July 2025

Figure 1 and Figure 2 present keyword co-occurrence networks that synthesize prevailing research themes and illustrate the progression toward integrated, AI-driven frameworks in urban building energy modeling, indicating a convergence of the previously isolated domains: UBEM, MOO, urban morphology and climate interactions, technological systems integration, and AI-based prediction methods.

Figure 1. Keyword co-occurrence clusters in ML/AI-assisted UBEM (2015–2025).

Figure 2. Keyword co-occurrence and clustering trends highlight key themes on UBEM, MOO, urban morphology and climate interactions, technological systems integration, and AI-based prediction methods.

3. AI-Driven Multi-Objective Optimization for Urban Building Energy Retrofit

Recent studies span a wide array of building types—residential, commercial, campus, and mixed-use—demonstrating the versatility of machine learning (ML), deep learning (DL), and reinforcement learning (RL) approaches across urban contexts [34,35,36,37]. The central focus has shifted from isolated building upgrades toward integrated, district-scale strategies that balance energy efficiency, emissions reduction, cost optimization, and renewable energy integration. To address this expanded spectrum of retrofit goals, recent literature increasingly integrates AI-assisted MOO frameworks [38,39,40]. These frameworks leverage AI models to support comprehensive objectives—such as thermal comfort, life-cycle cost (LCC), and greenhouse gas (GHG) mitigation—while enhancing retrofit decision-making’s tractability and interpretability. Notably, the fusion of ML with heuristic optimization algorithms has enabled efficient exploration of complex solution spaces [41,42,43]. In parallel, the adoption of explainable AI (XAI) techniques, such as SHapley Additive exPlanations (SHAP), has further improved transparency, fostered trust in AI-driven recommendations, and facilitated stakeholder engagement [44,45].

Building upon these advances, we propose a four-stage ML-integrated workflow for AI-driven UBER optimization, as illustrated in Figure 3. The framework includes the following: (1) Design Generation, where contextual parameters (e.g., climate, geometry, systems) define the retrofit scenarios; (2) AI-Driven Prediction, in which ML, DL, and RL models are employed to estimate performance metrics such as energy use, emissions, cost, and comfort; (3) Multi-Criteria Decision Making (MCDM), which leverages GA, PSO, or Pareto-based optimization to balance objectives; and (4) Strategy Determination, translating optimal results into actionable retrofit plans. This structured pipeline connects data preprocessing, prediction modeling, optimization, and decision support, providing a scalable and adaptive framework applicable to various urban contexts.

Figure 3. The general workflow of AI-driven UBER optimization.

3.1. Optimization Objectives and Retrofit Strategies

To systematically advance retrofit initiatives, it is critical to identify and prioritize optimization objectives and to design corresponding intervention strategies that operate at multiple scales and levels of urban complexity. Therefore, a comprehensive understanding of UBER requires clarifying its multi-layered objectives, mainstream demand- and supply-side strategies, and emerging cross-cutting considerations.

Optimization in UBER is inherently multi-objective, requiring the reconciliation of multiple, and sometimes competing, goals. Four primary objectives frequently guide retrofit decision-making. First, energy efficiency minimizes operational energy consumption through enhancements to the building envelope, mechanical systems, and behavioral interventions [5,40]. As buildings constitute the primary energy consumers in cities, reducing heating, cooling, and lighting demands is central to sustainable urban development. Second, carbon emission mitigation has become increasingly critical, focusing on the reduction of greenhouse gas emissions across the building life cycle—including both embodied and operational carbon—by integrating low-carbon materials, renewable energy systems, and explicit carbon valuation into retrofit strategies [46,47]. Third, economic viability ensures that retrofit solutions remain cost-effective by considering both initial capital expenditures and long-term life-cycle costs [1,35,39]. Economic assessments often utilize metrics such as Net Present Value (NPV), Internal Rate of Return (IRR), and Payback Period (PBP) to evaluate project feasibility. Finally, occupant comfort and urban resilience serve as essential objectives, aiming to improve indoor environmental quality—covering thermal, visual, and acoustic aspects—while also ensuring that retrofit strategies enhance resilience to future climate change and support long-term urban sustainability [48,49].

To achieve the aforementioned retrofit objectives, strategies are typically categorized into demand-side interventions—which focus on reducing energy consumption at the source—and supply-side measures that transform how energy is delivered and managed.

On the demand side, envelope upgrades remain a cornerstone, encompassing improvements to insulation, window performance (e.g., U-values and SHGC), window-to-wall ratio (WWR), infiltration control, and roofing systems. These measures can potentially contribute up to 50% of total energy savings [9,39], with particularly pronounced notable benefits in older, inefficient building stocks [38,50]. Urban morphology optimization, which involves the refinement of spatial configurations—including orientation, compactness, shading, and surface-to-volume ratios—serves to maximize passive solar gains, daylighting, and natural ventilation, with substantial evidence demonstrating its impact on energy performance and solar potential [51]. It should be noted, however, that such interventions are primarily applicable to retrofit projects where significant morphological modifications are feasible, as many existing buildings or urban blocks offer limited scope for large-scale form adjustments. To maximize the combined effectiveness of these demand-side measures, MOO algorithms—often assisted by ML—are widely employed to identify optimal combinations of retrofit actions, balancing trade-offs among energy savings, cost, occupant comfort, and carbon reduction [41,42,43]. This optimization-driven approach enables tailored retrofit strategies that account for building-specific constraints and stakeholder priorities, further enhancing the impact and feasibility of demand-side interventions.

Supply-side strategies are complementary to these interventions, involving upgrading heating, ventilation, and air-conditioning (HVAC) systems to high-efficiency technologies such as heat pumps and advanced chillers, while implementing intelligent control systems. Integrating ML methods in this domain enables improved system prediction, operational optimization, and adaptive control [52,53]. Furthermore, the deployment of renewable energy sources—including photovoltaic (PV), geothermal, and wind systems—at both building and district scales is increasingly optimized through ML algorithms, which assist in determining optimal system sizing, siting, and operational strategies, thus enhancing the cost-effectiveness and performance of renewable integration [34,54,55]. Smart energy management has also emerged as a vital component, leveraging advanced energy management systems (EMS), real-time sensor networks, and demand response strategies to coordinate generation, storage, and flexible loads, which are essential for enabling urban energy flexibility and supporting the resilience of the electricity grid [56].

3.2. ML and AI Methods for UBER

3.2.1. Supervised Learning: Mainstream Algorithms and Applications

Supervised learning remains a central analytical engine for UBER, providing scalable, interpretable solutions for energy use forecasting, retrofit value evaluation, and optimal scenario selection. Modern studies predominantly adopt ensemble tree models such as RF and Gradient Boosting Machines (GBM, e.g., XGB, LightGBM), which consistently outperform classical regression and single-tree methods in predictive accuracy, robustness, and feature importance extraction [57,58,59,60,61]. These approaches have become the dominant approach for large-scale energy mapping, regional carbon assessment, and envelope upgrade optimization, offering critical insights for policy and capital investment planning [62].

In more complex, high-dimensional tasks—such as short-term energy load forecasting or multi-objective retrofit optimization—SVR and ANN excel due to their ability to model complex nonlinear relationships and integrate seamlessly with advanced optimization frameworks. Notably, hybrid approaches such as PSO-SVR achieve superior accuracy for carbon, heating, and cooling predictions under uncertain and multi-criteria constraints [36,47]. ANNs, particularly when coupled with genetic algorithms (GA), now routinely function as surrogate models, replacing computationally intensive physical simulations. Recent applications report reductions in computation time from minutes to milliseconds for multi-criteria annual cost or emissions analysis, without loss of accuracy [38,40]. Such surrogate models are increasingly embedded within AI-driven MOO pipelines, enabling rapid scenario generation and large-scale decision support (Figure 4).

Figure 4. A prototype for AI-assisted UBER optimization.

The integration of XAI tools (e.g., SHAP) and Bayesian optimization further enhances interpretability and operational efficiency [31,60], while stacked ensemble methods improve robustness to data gaps and outliers. These advances collectively enable robust sensitivity analysis, uncertainty quantification, and actionable ranking of retrofit strategies, advancing the transition from pilot-scale analytics to automated, scalable UBER decision-making [5].

Despite these successes, supervised models still face challenges related to transferability across regions, climates, and heterogeneous building stocks, often due to overfitting and limited data diversity. The scarcity of open, standardized datasets constrains reproducibility and generalization [62]. Addressing these issues will require cross-domain validation, active learning to accommodate evolving data streams, and tighter integration with physical and unsupervised frameworks, paving the way for more robust, scalable, and trustworthy ML-driven retrofit strategies.

3.2.2. Unsupervised and Hybrid Learning: Archetype Classification and Data Structuring

Unsupervised learning plays an essential role in structuring large-scale UBER analyses, particularly where modeling each building individually is impractical due to data limitations and urban complexity. Clustering algorithms—such as k-means, hierarchical clustering, DBSCAN, and Gaussian Mixture Models (GMM)—have been widely applied to archetype classification, grouping buildings by geometric, operational, or climatic features to streamline scenario simulation and prioritize retrofit actions at city scale [63,64,65]. Dimensionality reduction techniques, including Principal Component Analysis (PCA), t-distributed Stochastic Neighbor Embedding (t-SNE), non-negative matrix factorization (NMF), and autoencoders, further enhance these workflows by integrating heterogeneous data sources and compressing high-dimensional attributes into actionable feature spaces, supporting representative archetype generation and more tractable optimization [66,67]. Recent advances also leverage deep clustering and graph-based learning methods, such as autoencoder-based feature extraction and graph neural networks (GNNs), to capture complex spatial and operational patterns in urban environments.

Recent research increasingly leverages open-access urban datasets and multi-source fusion, enabling transferability of archetype libraries across cities with diverse morphologies and regulatory environments. Hybrid pipelines that combine unsupervised learning with XAI techniques—such as SHAP-based cluster interpretation or attention-guided feature selection—have been introduced to improve transparency, stakeholder trust, and policy relevance in retrofit planning [65,66]. Moreover, large-scale implementations now integrate automated clustering, physical simulation, and policy-driven prioritization into unified platforms, advancing scalable data-driven UBER workflows. Despite these advances, significant challenges persist in harmonizing disparate data streams, ensuring cluster stability in dynamic urban contexts, and generalizing models across varied regions. Future research should focus on robust, interpretable unsupervised frameworks that balance transferability and specificity, supporting real-world deployment and multi-stakeholder decision-making in urban energy retrofitting.

3.2.3. Deep Learning and Spatiotemporal Models

DL has emerged as a transformative force in UBER, enabling the modeling of high-dimensional, nonlinear, and spatiotemporally dependent patterns that are difficult for traditional ML methods to capture. Modern UBER workflows increasingly rely on DL to process heterogeneous data sources, from time series and sensor streams to remote sensing imagery and digital twins. Among foundational architectures, multilayer perceptrons (MLPs) have a multilayer option for structured energy demand forecasting tasks, with studies demonstrating notable improvements over shallow models in both accuracy and robustness [42,68]. However, MLPs are limited in representing temporal dependencies—a gap addressed by recurrent neural networks (RNNs), particularly long short-term memory (LSTM) networks and gated recurrent units (GRUs). These temporal models enable multi-step prediction and adaptive control of HVAC or building loads under dynamic conditions [69,70,71], providing critical support for real-time energy management and demand response.

Convolutional neural networks (CNNs) have proven highly effective for spatial and image-driven analysis in archetype classification, solar potential mapping, and extracting building features from satellite data [54,72]. Recent hybrid approaches integrate CNNs with RNNs or metaheuristics to enable spatial-temporal forecasting and design optimization for urban districts. Further, generative adversarial networks (GANs) and variational autoencoders (VAEs) have been used to synthesize high-fidelity occupant profiles and energy patterns in data-scarce or privacy-sensitive settings [73,74,75]. These generative models support uncertainty quantification, scenario analysis, and the augmentation of simulation datasets, providing a foundation for digital twin development and synthetic benchmarking.

A particularly innovative direction is the application of graph neural networks (GNNs) and spatiotemporal graph convolutional networks (ST-GCNs), which model both the topological (e.g., building adjacency, radiative coupling) and dynamic aspects of urban energy systems [76,77]. By learning from graph-structured data, these models excel in district-level prediction, load clustering, and propagation effects analysis, opening up new possibilities for community-scale retrofit and resilience planning. Despite their power, DL models often require extensive training data and computational resources, and challenges remain in generalization, interpretability, and integration with physics-based constraints. Nonetheless, the rapid evolution of spatiotemporal and hybrid DL is reshaping the frontier of AI-driven UBER, enabling truly adaptive, scalable, and data-rich retrofit strategies at both building and urban scales.

3.2.4. Reinforcement Learning and Adaptive Control

The fields of Reinforcement Learning (RL) and Deep Reinforcement Learning (DRL) have emerged as transformative tools for the management of energy in UBER. By formulating retrofit or control as a sequential decision-making problem under uncertainty, RL/DRL enables dynamic optimization of energy use, occupant comfort, emissions, and cost—capabilities often unattainable with conventional methods. Mainstream RL algorithms, including Q-learning, Deep Q-Networks (DQN), Proximal Policy Optimization (PPO), and Actor-Critic methods, are increasingly applied to a range of tasks, such as HVAC control [12], demand response, grid-interactive load shifting [38], and distributed energy management.

In the MOO context, RL/DRL agents are trained with multi-objective reward functions, Pareto front-based approaches, or reward scalarization, allowing them to balance energy, cost, comfort, and emissions dynamically. For instance, in district-level retrofit optimization, RL/DRL agents can be used to dynamically assign and coordinate building-level interventions, learning policies that approach Pareto-optimal trade-offs between carbon, energy savings and comfort. Recent advances in multi-agent DRL have demonstrated the potential for cooperative, multi-objective energy management in building energy systems (BES) [78,79]. Xu et al. developed and benchmarked several DRL algorithms—including TD3—for optimizing battery and heat pump actions in home energy management systems (HEMS), considering both cost savings and PV self-consumption through custom multi-objective reward functions [80]. Their results demonstrated a 13.79% reduction in operating costs and a 5.07% increase in PV self-consumption relative to baseline models, highlighting the potential of DRL-driven control for cost-effective and resilient renewable energy integration. Such frameworks can optimize energy consumption, occupant comfort, renewable energy utilization, and cost simultaneously through agent cooperation and advanced reward design. While most existing studies are focused on the building or subsystem scale, such multi-agent DRL frameworks have shown success in building energy system management and may inspire scalable, district-level optimization frameworks in future UBEM research.

Recent advances in multi-agent reinforcement learning (MARL) have opened up new avenues for managing complex, distributed energy systems in urban environments. By modeling buildings, devices or control zones as autonomous agents, MARL frameworks facilitate decentralized, adaptive coordination, enabling dynamic optimization of energy use, flexibility, and occupant comfort [81,82]. In the context of UBER, MARL has been adopted for tasks such as city-scale demand response, distributed HVAC control, and real-time coordination of energy storage and renewable generation assets [83,84,85]. Notably, international benchmark environments (e.g. CityLearn) and several large-scale pilot studies have demonstrated the ability of MARL agents to efficiently reduce peak loads, enhance the utilization of renewable energy, and guarantee fairness and privacy in multi-building settings [86]. Integrating digital twins with MARL frameworks allows control policies to be developed and validated safely in high-fidelity, data-rich virtual environments [87]. By linking real-time sensor data, simulation engines, and city-scale BIM/GIS, digital twins allow urban planners and policymakers to test various retrofit and energy management strategies before implementation, thereby minimizing operational risks. These platforms facilitate scenario analysis, investment evaluation, and adaptive policymaking at the district or city level, promoting more resilient and cost-effective urban retrofitting solutions [88]. Benchmark environments such as CityLearn [86] and high-fidelity digital twins (EnergyPlus, CitySim, Modelica [89]) provide safe platforms for training and evaluating RL agents, supporting the coordination of energy strategies across multiple buildings and real-world operating conditions.

These innovations are increasingly relevant as cities transition toward distributed, renewable-intensive grids and require scalable, autonomous, and flexible control. However, significant hurdles remain, including sample inefficiency, reward engineering complexity, interpretability, and ensuring operational safety in real deployments [38]. Ongoing research focuses on explainable RL, hybrid DRL-MPC (Model Predictive Control) integration, and robust multi-agent learning, which are expected to accelerate the deployment of intelligent, multi-objective control systems in practical UBER settings.

3.2.5. Comparative Analysis of ML Models

Evaluating the performance of ML models is essential for informed decision-making in UBER. Standard evaluation metrics include R-squared ($R^2$), Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE), which together provide insights into both predictive accuracy and model generalization. The mathematical definitions are as follows:

$$\begin{array}{cc}\hfill \mathrm{MAE}& ={\displaystyle \frac{1}{n}}\sum _{i=1}^n|y_i-{\widehat{y}}_i|\hfill \end{array}$$

(1)

$$\begin{array}{cc}\hfill \mathrm{MSE}& ={\displaystyle \frac{1}{n}}\sum _{i=1}^n{(y_i-{\widehat{y}}_i)}^2\hfill \end{array}$$

(2)

$$\begin{array}{cc}\hfill \mathrm{RMSE}& =\sqrt{{\displaystyle \frac{1}{n}}\sum _{i=1}^n{(y_i-{\widehat{y}}_i)}^2}\hfill \end{array}$$

(3)

$$\begin{array}{cc}\hfill \mathrm{MAPE}& ={\displaystyle \frac{100\%}{n}}\sum _{i=1}^n\left|{\displaystyle \frac{y_i-{\widehat{y}}_i}{y_i}}\right|\hfill \end{array}$$

(4)

$$\begin{array}{cc}\hfill R^2& =1-{\displaystyle \frac{{\sum }_{i=1}^n{(y_i-{\widehat{y}}_i)}^2}{{\sum }_{i=1}^n{(y_i-\overline{y})}^2}}\hfill \end{array}$$

(5)

where $y_i$ and ${\widehat{y}}_i$ denote the observed and predicted values, respectively, $\overline{y}$ is the sample mean, and n is the number of samples.

Table 3 presents a consolidated benchmarking of representative ML and DL models applied in UBER case studies, summarizing key performance metrics (e.g., $R^2$, RMSE, MAE, MAPE, F1) across diverse datasets, building types, and algorithmic paradigms. This comparative overview enables direct quantitative comparison between commonly used models such as tree ensembles (RF, XGB, LightGBM, GBDT), neural networks (MLP, DNN, LSTM, GRU, Transformer), and advanced graph-based or hybrid architectures (ST-GCN, GCN-LSTM, biLSTM, TFT), under a variety of real-world urban building scenarios.

Table 3. Performance metrics and benchmarking of representative ML/DL models in UBER case studies.

Ref.	ML/DL Model	Dataset	Performance Metric	Value
[45]	LASSO / Ridge / SVR / KNN / DNN / RF / GBDT / LightGBM / XGB	Urban buildings (multi-type)	R2	0.45 / 0.45 / 0.50 / 0.35 / 0.49 / 0.51 / 0.49 / 0.49 / 0.49
			RMSE	0.84 / 0.84 / 0.80 / 0.91 / 0.81 / 0.79 / 0.81 / 0.80 / 0.81
	RF (multi-family / commercial / mixed / public)	–	R2	0.39 / 0.58 / 0.45 / 0.25
			RMSE	0.64 / 0.82 / 0.69 / 1.23
[58]	XGB+Jaya	Residential buildings (heating/cooling)	RMSE MAE R2	0.381 / 0.9757 0.2781 / 0.612 0.998 / 0.989
[59]	XGB / DNN / RF	Green building projects	MAE	92.0 / 196.5 / 378.0
			RMSE	132.5 / 284.0 / 507.9
			MAPE	19.9 / 32.4 / 40.4
			$R^2$	96.0 / 91.0 / 87.0
			$R_{adj}^2$	95.9 / 90.9 / 86.8
[60]	SVR / RF / LightGBM / Random-LightGBM / Grid-LightGBM / CMA-ES-LightGBM	Residential buildings (heating)	RMSE	1.9231 / 0.7731 / 0.7162 / 0.4938 / 0.3610 / 0.4023 / 0.2714
			MAE	1.6624 / 0.6000 / 0.5634 / 0.3894 / 0.2562 / 0.2996 / 0.1416
			R2	0.8710 / 0.9835 / 0.9852 / 0.9934 / 0.9965 / 0.9957 / 0.9981
			MAPE	5.0079 / 1.8968 / 1.7176 / 1.2038 / 0.8025 / 0.9348 / 0.4699
	TPE-LightGBM	Residential buildings (cooling)	RMSE	0.1901
			MAE	0.1394
			R2	0.9924
			MAPE	2.3509
[68]	MLP / MLP + BBO / MLP + GA / MLP + PSO / MLP + ACO / MLP + ES / MLP + PBIL	Residential buildings	RMSD (heating, MLP + BBO)	2.82
			R2	0.920
			MAE	2.15
			RMSD (cooling, MLP + BBO)	3.18
			R2	0.887
			MAE	2.97
[69]	RNN (Rec, Dir, MIMO) / GRU (Rec, Dir, MIMO) / LSTM (Rec, Dir, MIMO)	Educational building (HK), 17,000 samples	MAE	185.2 / 90.0 / 119.5 / 185.0 / 75.3 / 92.9 / 180.2 / 81.0 / 83.4
			RMSE	242.4 / 123.6 / 196.7 / 245.1 / 111.9 / 157.2 / 265.9 / 118.2 / 151.7
			CV-RMSE (%)	34.6 / 17.6 / 28.1 / 35.0 / 16.0 / 22.4 / 38.0 / 16.9 / 21.6
[70]	LSTM	539 residential + 153 public buildings, monthly/yearly data	MAPE (Model 1)	0.41
			R2 (Model 1)	0.57
			MAPE (Model 5, residential/public)	0.093 / 0.194
			R2 (Model 5, residential/public)	0.975 / 0.99
[71]	LSTM / LSTM-CNN / CNN-LSTM / LSTM-Attention / CNN-Attention-LSTM / LSTM-Attention-CNN	60 buildings, 4 climate zones, Building Data Genome Project 2	RMSE (one-year data, CNN-LSTM vs. LSTM)	LSTM: base / CNN-LSTM: −2.9%
			RMSE (two-year data, LSTM-Attention vs. LSTM)	LSTM: base / LSTM-Attention: −5.6%
			RMSE (all improved LSTM models, after param tuning)	−6.2%   −29.2% (vs. vanilla LSTM)
			Computational cost	Improved models: higher, but stable with LSTM-Attention
[72]	Adaptive DNN (feature clustering + GA) vs. Standard DNN	Office building (UK)	MAPE (train/test, adaptive DNN)	2.87% / 6.12%
			MAPE (train/test, standard DNN)	3.81% / 6.95%
			Improvement (adaptive vs. standard)	−24.6% (train) / −11.9% (test)
[76]	Last hour / Avg / LR / MLP / XGB / GRU / ST-GCN	University campus (Atlanta)	RMSE (kWh)	43.81 / 50.32 / 34.49 / 25.20 / 22.52 / 28.58 / 18.56
			MAPE	16.67 / 14.04 / 30.47 / 7.85 / 6.93 / 7.72 / 5.21
[77]	GCN-LSTM (EDT) / LSTM / CNN / MLP	Educational building, multi-zone	MAE (5-step)	8.24 / 22.61 / 19.01 / 12.73
			CV(RMSE)	5.78 / 18.39 / 13.22 / 8.71
[90]	biLSTM / LSTM / Linear / MLP / Naive	UBEM simulation, census tract, Los Angeles County	nMAE (ELEC)	0.0311 / 0.1037 / 0.1555 / 1.1384 / 0.0500
			nRMSE (ELEC)	0.1031 / 0.3825 / 0.4356 / 2.2285 / 0.1205
[91]	TFT-Prob / TFT-Deter / Trans-Prob / Trans-Deter / RNN-Prob / RNN-Deter / LR	114 buildings (US, multi-climate)	RMSE (Heating, kWh)	71.51 / 69.76 / 114.73 / 97.50 / 131.03 / 79.41 / 121.86
			RMSE (Cooling, kWh)	109.54 / 115.62 / 185.77 / 131.69 / 150.85 / 125.60 / 181.91
			F1 (Heating)	0.80 / – / 0.57 / – / 0.52 / – / –
			F1 (Cooling)	0.86 / – / 0.77 / – / 0.68 / – / –
[92]	REED / Bi-GRU-attention / GRU / SimpleRNN / DNN / SVR / KNN	University campus, 8 buildings (S1)	R2	0.943 / 0.929 / 0.939 / 0.911 / 0.895 / 0.921 / 0.738
			RMSE	0.670 / 3.035 / 2.814 / 3.388 / 3.674 / 0.068 / 0.124
			MAPE	0.040 / 0.573 / 0.577 / 0.734 / 0.858 / 0.172 / 0.303
[93]	Developed / RFR	Residential, Uwall	R2	0.708 / 0.130
			RMSE	0.122 / 0.196
			MAE	0.101 / 0.151
			MAPE	0.148 / 0.225
	Developed / RFR	Non-residential, Uwall	R2	0.621 / 0.188
			RMSE	0.113 / 0.176
			MAE	0.089 / 0.133
			MAPE	0.152 / 0.229

This benchmarking not only highlights the range of predictive accuracies achieved by different ML/DL methods in published UBER applications, but also reveals the sensitivity of model performance to data characteristics, use case, and task formulation. As seen, recent advances—particularly in gradient boosting, recurrent and attention-based neural networks, and graph neural architectures—consistently outperform traditional models on large, heterogeneous datasets, while hybrid or physics-informed approaches show promise for robust, interpretable, and transferable predictions. The summary thus provides both a reference for future method selection and a baseline for evaluating methodological progress in the field.

3.3. Representative Applications of AI-Driven Multi-Objective Optimization in Urban Building Energy Retrofit

Building on the preceding discussions of retrofit objectives, strategies, and advanced AI methods, this section highlights representative real-world applications of AI-driven MOO in UBER. Recent studies have demonstrated that integrating ML/DL/RL techniques with MOO frameworks enables tailored optimization of retrofit strategies—including high-performance envelope upgrades, HVAC and system improvements, and renewable energy deployment—across diverse building types and urban contexts [38,50]. Table 4 provides a systematic overview of these applications, detailing the district building types, retrofit interventions, objective functions, and AI/MOO methodologies employed in the literature.

Envelope upgrades remain a foundational strategy in demand-side retrofit, with AI-driven MOO frameworks increasingly applied to optimize insulation levels, window performance (e.g., U-values, SHGC, WWR), and infiltration control for maximum energy savings and occupant comfort. Studies have demonstrated that ML-based surrogate models and optimization algorithms can rapidly evaluate the impact of different envelope interventions across diverse building stocks and urban morphologies, enabling tailored, cost-effective retrofit solutions [9,38,39,45,50]. In particular, multi-objective formulations balance energy reduction, carbon mitigation, and life-cycle costs, supporting scenario-based decision-making in large-scale urban contexts.

System upgrades—particularly HVAC retrofits—represent a critical pathway for enhancing operational efficiency in UBER. Recent studies have prioritized the adoption of high-efficiency heat pumps and advanced chillers, with ML techniques widely applied to forecast retrofit savings, optimize control strategies, and support scenario analysis [52]. Integrating real-time, AI-driven intelligent control systems further enables predictive management and dynamic adaptation, improving energy flexibility and occupant comfort, especially at the district scale.

On the supply side, the deployment of renewable energy systems—most notably photovoltaics (PV)—is a cornerstone of net-zero and low-carbon retrofit strategies. Advanced ML methods are increasingly used for PV potential estimation, optimal siting, and grid-integration studies, accounting for urban morphology and climate variability [3,94]. District-scale simulation frameworks leveraging hybrid optimization algorithms (such as genetic algorithms and Monte Carlo methods) enable coordinated management of renewable generation and consumption, advancing urban energy systems toward net-zero emissions [55].

As summarized in Table 4, recent applications of AI-driven MOO in UBER increasingly integrate advanced analytical tools such as SHAP, TOPSIS, and spatial proximity analysis. SHAP explains the contribution of each input variable to model predictions and enhances the transparency of complex ML/DL models [44,45]; TOPSIS systematically ranks alternative retrofit solutions under multiple criteria, supporting stakeholder-driven decision-making [39,95]. These methods collectively facilitate transparent optimization, scenario analysis, and stakeholder engagement, aligning technical solutions with policy objectives in large-scale urban retrofit planning.

Building upon these advances in MOO and decision analysis, the integration of supervised learning models has become central to scaling and accelerating UBER applications. ML surrogates rapidly approximate simulation outcomes and enable tractable multi-objective searches. Supervised learning methods such as ANN, RF, XGB, and SVM are increasingly deployed as surrogate models within MOO frameworks, supporting rapid scenario evaluation, intervention selection, and outcome prediction across complex urban settings. Transformer-based surrogates [91] and BiLSTM models [90] have achieved substantial speed-ups in hourly energy forecasting. By coupling these surrogates with heuristic optimization (GA, PSO, NSGA-II, TOPSIS), recent studies have demonstrated rapid identification of Pareto-optimal solutions across energy, cost, carbon, and comfort metrics [39,96]. Araújo et al. [97] demonstrated a novel workflow in Lisbon, Portugal, integrating algorithmic design, UBEM simulation, and ML to enable rapid, automated evaluation of thousands of retrofit scenarios across large urban neighborhoods. In this case study, the proposed surrogate modeling approach was able to generate and simulate a wide range of building archetypes, producing accurate annual energy predictions orders of magnitude faster than conventional UBEM workflows. By substantially reducing computational time and expertise requirements, such integrated ML-surrogate pipelines have shown strong potential to support municipal policymaking, scenario-based planning, and citywide retrofit prioritization, making advanced energy analysis more accessible and scalable for urban practitioners. Reynolds et al. [98] applied an ANN–genetic algorithm (GA) approach to optimize retrofit scenarios for a mixed-use urban district, achieving up to 32% energy cost reduction and reducing computational time by an order of magnitude compared to brute-force methods. Similarly, Zhang et al. [38] leveraged an SVM–particle swarm optimization (PSO) model to identify optimal envelope and material retrofits for complex, non-standard urban buildings, demonstrating both robust performance and flexibility in high-dimensional decision spaces.

Ensemble models such as XGB and LightGBM have been increasingly adopted for rapid energy forecasting and scenario optimization across large urban building portfolios. For instance, Ali et al. [5] developed a scalable data-driven methodology that combines ensemble-based ML with end-use demand segregation to predict the energy performance of urban residential buildings at the city scale. Their framework incorporates archetype development, physics-based parametric modeling, and the generation of synthetic datasets for one million buildings, achieving 91% prediction accuracy—significantly outperforming traditional approaches. This level of precision enables stakeholders, including policymakers and urban planners, to design and implement informed, large-scale retrofit strategies that effectively reduce urban energy consumption and emissions.

At the urban block scale, AI-assisted MOO has enabled the systematic exploration and design of city block morphologies to maximize solar access and optimize volumetric configurations. Veisi et al. [99] employed Pareto front-based GA to analyze a vast design space of urban blocks—varying block type, scale, orientation, and grading—to identify optimal forms for solar radiation maximization. Leveraging an extensive simulation dataset of over 170,000 configurations, they trained an ANN as a surrogate model to emulate the simulation process, significantly reducing computational time while enabling detailed sensitivity analysis. Their results demonstrated, for the first time, that a 100m × 100 m city block represents an optimal configuration for solar potential under the studied conditions. More broadly, this work showcased how ANN–surrogate integration with multi-objective evolutionary optimization can support evidence-based urban morphology planning, allowing researchers and planners to efficiently evaluate and optimize urban energy performance at scale.

An emerging frontier is the explicit inclusion of future climate scenarios in retrofit optimization. Micro-genetic optimization embedded in physics-driven models has enabled long-term, climate-resilient strategy development, with case studies in Abu Dhabi, Changsha, and Toronto reporting up to 80% savings and robust adaptation to extreme events [100,101,102]. For instance, a recent engineering-scale study in Toronto, Canada, demonstrates the application of micro-genetic optimization integrated with urban physics simulation for future-ready residential retrofits [102]. By coupling AI-based multi-variable optimization with climate scenario modeling, the research identified retrofit solutions that reduce annual gas use by over 80% and electricity demand by more than 60%, with payback periods under 20 years. The approach revealed that optimal strategies shift with projected climate conditions, underscoring the need for scenario-adaptive, cost-effective retrofit planning in municipal decarbonization efforts. These advances underscore the growing need for scenario-aware, adaptive frameworks that can anticipate both present and future urban energy challenges.

Notably, XAI methods such as SHAP and LIME have been integrated with surrogate models (e.g., RFs, gradient boosting, deep neural networks), providing both prediction accuracy and transparency for early-stage decision-making [45,103]. These techniques allow practitioners and policymakers to diagnose the impact of retrofit measures at both building and district scales, supporting evidence-based urban energy planning.

Table 4. Representative case studies of AI-driven multi-objective optimization frameworks in Urban Building Energy Retrofit.

Ref.	Authors	Building Type	Retrofitting Strategies	Objective Function	Method
[50]	Li et al.	Office, Commercial	Building envelope	Energy consumption	ML + MOO + SHAP (BR/ETR/RFR/GBR/ADB/XGB + GA + SHAP))
[45]	Li et al.	Mixed-use	Building envelope	Energy consumption	ML + SHAP (Lasso/Ridge/SVM/KNN/DNN/RF/GBDT/LightGBM/XGB + SHAP)
[34]	Tao et al.	Commercial	Building morphology	FIPV application	ML + SHAP + MOO (RF + SHAP + NSGA-II)
[39]	Luo et al.	Campus	Envelope characteristics	Energy consumption	ML + MOO + TOPSIS (ANN + NSGA-II + TOPSIS)
			Mechanical system set points	Carbon emission
			HVAC system set points	LCC
[5]	Ali et al.	Residential	Construction characteristics	Energy consumption	ML + SHAP (XGB/LGBM/GB/HGB/RF/NN/DT/LR/KNN/SVM + SHAP)
			Internal gains
			Occupancy density
			Heating or cooling systems
[104]	Yu et al.	Residential	Courtyard	Energy consumption	ML + MOO (LightGBM + GA)
				Daylight
[46]	Hey et al.	Residential	Retrofit strategies	Energy consumption	ML + MOO (ANN/DNN + GA)
				Energy
[40]	Thrampoulidis et al.	Residential	Building envelope	GHG emission	ML (ANN)
			Energy system selection	Life cycle cost
			PV capacity
			HVAC systems
[99]	Veisi et al.	Urban block	Urban morphology	Solar radiation	ML + MOO (ANN + NSGA-II)
				Volume
[38]	Zhang et al.	Residential	Retrofit strategies	Retrofit emissions	ML + MOO + TOPSIS (ANN + GA + TOPSIS)
				Retrofit costs
[9]	Zygmunt et al.	Residential	Building envelope	Energy consumption	ML (ANN)
			HVAC systems
[37]	Wenninger and Wiethe	Residential	Energy source type	Energy consumption	ML (ANN/D-vine copula quantile regression/Extreme GB/RF/SVF)
			Insulation
			Living space
[7]	Thrampoulidis et al.	Residential	Building envelope	Carbon emission	ML + MOO + TOPSIS (ANN + NSGA-II + TOPSIS)
			Energy systems	Total cost
[1]	Nutkiewicz et al.	Mixed-use	Window retrofit	Energy consumption	DL (LSTM/RNN)
			Switch to LED bulbs
[36]	Li and Yao	Mixed-use	Building envelope	Energy consumption	ML (SVR/RF/GB/ANN))
			HVAC system heating efficiency
			Ventilation method
			Internal gains
[3]	Ali et al.	Residential	Building envelope	EUI saving	ML (DR/NN/DT/RT/GB/KNN)
			Heating efficiency	Carbon
			Main space heating fuel	Retrofit cost
			Main water heating fuel
			Hot water efficiency
			Low energy lighting percent
			Solar hot water heating
			Solar PV

Note: ML = Machine Learning; DL = Deep Learning; MOO = Multi-objective Optimization; SHAP = SHapley Additive exPlanations; TOPSIS = Technique for Order Preference by Similarity to Ideal Solution; FIPV = Façade-integrated Photovoltaics; GHG = Greenhouse Gas; LCC = Life Cycle Cost; EUI = Energy Use Intensity; RF = Random Forest; GB = Gradient Boosting; DNN = Deep Neural Network; LR = Linear Regression; NSGA-II = Non-dominated Sorting Genetic Algorithm II; ANN = Artificial Neural Network; XXX = algorithm unspecified. The Method column shows the combination of ML/DL algorithms and optimization or explanation tools. For example, “ML + MOO (RF + NSGA-II)” means a ML model (RF) is used as a surrogate to accelerate the MOO process (NSGA-II algorithm).

Table 4. Representative case studies of AI-driven multi-objective optimization frameworks in Urban Building Energy Retrofit.

Ref.	Authors	Building Type	Retrofitting Strategies	Objective Function	Method
[50]	Li et al.	Office, Commercial	Building envelope	Energy consumption	ML + MOO + SHAP (BR/ETR/RFR/GBR/ADB/XGB + GA + SHAP))
[45]	Li et al.	Mixed-use	Building envelope	Energy consumption	ML + SHAP (Lasso/Ridge/SVM/KNN/DNN/RF/GBDT/LightGBM/XGB + SHAP)
[34]	Tao et al.	Commercial	Building morphology	FIPV application	ML + SHAP + MOO (RF + SHAP + NSGA-II)
[39]	Luo et al.	Campus	Envelope characteristics	Energy consumption	ML + MOO + TOPSIS (ANN + NSGA-II + TOPSIS)
			Mechanical system set points	Carbon emission
			HVAC system set points	LCC
[5]	Ali et al.	Residential	Construction characteristics	Energy consumption	ML + SHAP (XGB/LGBM/GB/HGB/RF/NN/DT/LR/KNN/SVM + SHAP)
			Internal gains
			Occupancy density
			Heating or cooling systems
[104]	Yu et al.	Residential	Courtyard	Energy consumption	ML + MOO (LightGBM + GA)
				Daylight
[46]	Hey et al.	Residential	Retrofit strategies	Energy consumption	ML + MOO (ANN/DNN + GA)
				Energy
[40]	Thrampoulidis et al.	Residential	Building envelope	GHG emission	ML (ANN)
			Energy system selection	Life cycle cost
			PV capacity
			HVAC systems
[99]	Veisi et al.	Urban block	Urban morphology	Solar radiation	ML + MOO (ANN + NSGA-II)
				Volume
[38]	Zhang et al.	Residential	Retrofit strategies	Retrofit emissions	ML + MOO + TOPSIS (ANN + GA + TOPSIS)
				Retrofit costs
[9]	Zygmunt et al.	Residential	Building envelope	Energy consumption	ML (ANN)
			HVAC systems
[37]	Wenninger and Wiethe	Residential	Energy source type	Energy consumption	ML (ANN/D-vine copula quantile regression/Extreme GB/RF/SVF)
			Insulation
			Living space
[7]	Thrampoulidis et al.	Residential	Building envelope	Carbon emission	ML + MOO + TOPSIS (ANN + NSGA-II + TOPSIS)
			Energy systems	Total cost
[1]	Nutkiewicz et al.	Mixed-use	Window retrofit	Energy consumption	DL (LSTM/RNN)
			Switch to LED bulbs
[36]	Li and Yao	Mixed-use	Building envelope	Energy consumption	ML (SVR/RF/GB/ANN))
			HVAC system heating efficiency
			Ventilation method
			Internal gains
[3]	Ali et al.	Residential	Building envelope	EUI saving	ML (DR/NN/DT/RT/GB/KNN)
			Heating efficiency	Carbon
			Main space heating fuel	Retrofit cost
			Main water heating fuel
			Hot water efficiency
			Low energy lighting percent
			Solar hot water heating
			Solar PV

Note: ML = Machine Learning; DL = Deep Learning; MOO = Multi-objective Optimization; SHAP = SHapley Additive exPlanations; TOPSIS = Technique for Order Preference by Similarity to Ideal Solution; FIPV = Façade-integrated Photovoltaics; GHG = Greenhouse Gas; LCC = Life Cycle Cost; EUI = Energy Use Intensity; RF = Random Forest; GB = Gradient Boosting; DNN = Deep Neural Network; LR = Linear Regression; NSGA-II = Non-dominated Sorting Genetic Algorithm II; ANN = Artificial Neural Network; XXX = algorithm unspecified. The Method column shows the combination of ML/DL algorithms and optimization or explanation tools. For example, “ML + MOO (RF + NSGA-II)” means a ML model (RF) is used as a surrogate to accelerate the MOO process (NSGA-II algorithm).

3.4. ML-Enabled Urban Building Energy Modeling Platforms

The backbone of UBER research and deployment lies in a suite of UBEM platforms, which support both top-down and bottom-up simulation paradigms (see Figure 5). Notable examples include EnergyPlus [105], often integrated with OpenStudio or URBANopt [106] for batch simulation and scenario analysis at the city scale; CitySim [107] and CityBES [108], which facilitate archetype-based and rapid district-scale modeling; as well as TEASER [109], SimStadt [110], and ENVI-met [111] for archetype generation and microclimate analysis.

Figure 5. Categories of AI-driven UBEM methods.

Recent advances have enabled deeper integration of ML into UBEM workflows. Surrogate modeling, where algorithms such as ANN, SVR, and XGB are trained on simulation outputs, is now widely used to accelerate MOO and parameter sweeps [50]. Unsupervised learning techniques—including clustering and dimensionality reduction—facilitate the classification of building archetypes, preprocessing of large-scale input data, and standardization of simulation parameters [35,63]. Meanwhile, the use of deep reinforcement learning agents to interact with dynamic simulation environments, such as EnergyPlus and CitySim, has enabled closed-loop optimization for HVAC, storage, and demand-response strategies [12,86,112]. The development of open-source APIs and standardized data schemas (e.g., GeoJSON, IFC, CityGML) further enhances interoperability, cloud deployment, and collaborative research [106,113].

These platforms collectively offer the rigor of physics-based modeling and high-resolution urban context simulation, while ML integration enables scalable optimization, uncertainty quantification, and robust decision support for both policy and practice [39,40]. However, challenges remain—including high data requirements, computational expense for citywide studies, model calibration inconsistencies, and limited transferability of archetypes or control strategies across diverse regions. While hybrid and surrogate approaches alleviate some of these constraints, ongoing research is needed to address issues of generalization, data harmonization, and empirical validation across cities and climates [76,114].

Table 5 summarizes the key features of widely used UBEM simulation platforms, including their modeling engines, typical application scales, and integration strengths for ML-driven retrofit optimization. This overview provides a reference for understanding the technical underpinnings and future research priorities in hybrid UBEM–AI modeling.

Table 5. Common physics-based UBEM simulation platforms for UBER.

Platform	Modeling Engine	Key Features	Application Scale	Reference
EnergyPlus	Dynamic simulation engine	Detailed HVAC modeling, high temporal resolution, widely used and open-source	Building / Urban (with wrappers)	[105]
CitySim	Simplified thermal model	Fast computation, built-in solar radiation and daylight modeling	Block / District	[107]
URBANopt	OpenStudio + EnergyPlus	Urban-scale energy modeling, PV and battery integration, district systems	District / City	[106]
CityBES	EnergyPlus + archetypes	Web-based tool for U.S. cities, auto-generation of models from GIS	City	[108]
TEASER	Modelica interface	Archetype generation, open-source, focused on German stock modeling	Building / Neighborhood	[109]
SimStadt	Energy ADE + EnergyPlus	Bulk simulation from GIS, supports CityGML, developed for German cities	Urban / City	[110]
ENVI-met	CFD-based microclimate model	Urban heat island and outdoor comfort simulation, less focused on energy use	Street / Neighborhood	[111]

Recent advances in AI-enhanced UBER have driven three major shifts: the evolution from single-objective to integrated, multi-objective, and district-scale planning; the proliferation of ML surrogates combined with advanced optimization algorithms for tractable exploration of vast retrofit design spaces; and the increasing prominence of XAI for stakeholder engagement (Table 6). DL methods, such as CNNs for urban morphology and GNNs for district-level interactions, now enable fine-grained spatial and topological modeling [11,92,115]. These approaches, often paired with high-resolution data streams (e.g., LiDAR, satellite imagery), have improved predictive accuracy and adaptability. Nonetheless, challenges persist, including data heterogeneity, limited transferability across urban contexts, and the need for scalable, modular toolchains that support collaboration and reproducibility. Figure 6 and Table 7 illustrates the interplay between AI/ML methods and their application domains in UBER, with supervised learning dominating forecasting, DL and GNNs advancing spatiotemporal modeling, and optimization/decision models supporting multi-objective planning. The relatively limited use of unsupervised learning underscores opportunities for future methodological synthesis and innovation.

Table 6. Comparison of common machine learning methods in urban building retrofit modeling.

Method	Strengths	Limitations	Typical Applications	References
LR	Interpretable; fast computation	Poor for nonlinear relationships	Baseline modeling, trend analysis	[5,116]
RF	Robust; variable importance	Less interpretable, prone to overfitting with noisy data	Energy/load prediction, SHAP interpretation	[6,34]
GBM	High accuracy; handles missing/nonlinear data	Hyperparameter tuning; less interpretable	Optimization, variable ranking, surrogate modeling	[5,20,50]
SVM/SVR	Generalizable; good with small samples	Sensitive to data scaling; less scalable	Classification; regression for demand prediction	[45,54]
KNN	Simple; non-parametric	Inefficient for large datasets; local-only view	Clustering, occupant behavior patterns	[51,93]
ANN/DNN	Captures complex nonlinearities	Requires large data and compute; black-box	Multi-objective modeling, retrofit scenario testing	[7,38,39]
CNN	Learns spatial/morphological patterns	High computational cost; image-like inputs required	Urban morphology, PV potential mapping	[11,34,92]
LSTM	Good for temporal dynamics; handles long-term memory	Slower training; sensitive to noise	Time-series forecasting of energy/carbon trends	[1,8]
GNN	Captures graph-based spatial dependencies	Requires structured data (adjacency matrices)	District-level simulation; building network modeling	[1,13]
Hybrid (ML + MOO, ML + SHAP, DL + XAI)	Balances accuracy + interpretability; handles trade-offs	Implementation complexity; tuning required	Strategy ranking, decision support, multi-objective analysis	[20,39,50]

Figure 6. Mapping of AI and ML methods to key tasks in UBER.

Table 7. Key contributions of ML-Enhanced UBER optimization.

No.	Ref.	Author(s)	Year	Publication	Building Type	Method	ML/DL
1	[50]	Li et al.	2025	Building and Environment	Office, Commercial	ML + MOO + SHAP	BR, ETR, RFR, GBR, ADB, XGB
2	[11]	Geng et al.	2025	Applied Energy	Urban	DL	CNN
3	[34]	Tao et al.	2024	Sustainable Cities and Society	Commercial	ML + SHAP	RF
4	[39]	Luo et al.	2024	Building Simulation	Campus	ANN + MOO + TOPSIS	ANN
5	[45]	Li et al.	2024	Journal of Building Engineering	Residential, Public, Commercial	ML + SHAP	Lasso, Ridge, SVR, KNN, DNN, RF, GBDT, LightGBM, XGB
6	[92]	Chen et al.	2024	Energy and Buildings	Residential	DL	SimpleRNN, GRU, CNN
7	[6]	Alvarez-Sanz et al.	2024	Journal of Building Engineering	Residential	ML + SHAP	RF, XGB, Extra Trees
8	[5]	Ali et al.	2024	Energy and Buildings	Residential	ML + SHAP	XGB, LightGBM, GB, RF, NN, DT, LR, KNN, SVM
9	[117]	Zhang et al.	2023	Energy	Residential	ML + SHAP	LightGBM, XGB, RF, SVR
10	[104]	Yu et al.	2023	Buildings	Urban	ML + MOO	LightGBM
11	[40]	Thrampoulidis et al.	2023	Applied Energy	Residential	ANN	ANN
12	[62]	Nyawa et al.	2023	Annals of Operations Research	Residential	ML	DT, RF, GBoost, AdaBoost (ADB), LR, NB, Support Vector Classifier (SVC), ANN
13	[118]	Lu et al.	2023	Energy and Buildings	-	DL	SOLOv2 instance segmentation
14	[116]	Lu et al.	2023	Journal of Energy Engineering	Residential	ML	LR, BP
15	[46]	Hey et al.	2023	Journal of Building Performance Simulation	Residential	ML/DL + MOO	ANN/DNN
16	[119]	Gao and Yang	2023	IEEE Access	Residential	ML	SVM, RF, DT, GB, K-means
17	[38]	Zhang et al.	2022	Buildings	Residential	ML + MOO + TOPSIS	ANN
18	[93]	Wang et al.	2022	Building and Environment	Residential	ML	K-means, RF
19	[99]	Veisi et al.	2022	Sustainable Cities and Society	Urban	ML + MOO	ANN
20	[120]	Sun et al.	2022	Energy and Buildings	-	ML + SHAP	DCNN
21	[51]	Lan et al.	2022	Sustainable Cities and Society	Mixed-use	ML	ANN, RF, KNN, LDA, NB
22	[121]	Chen et al.	2022	Applied Soft Computing	-	ML + MOO	K-means clustering
23	[9]	Zygmunt et al.	2021	Energies	Residential	ML	ANN
24	[122]	Zhang et al.	2021	Building and Environment	Residential	ML	ANN
25	[123]	Wurm et al.	2021	ISPRS International Journal of Geo-Information	-	ML	DL
26	[37]	Wenninger and Wiethe	2021	Business Information Systems Engineering	Residential	ML	ANN, D-vine copula, XGB, RF, SVR
27	[7]	Thrampoulidis et al.	2021	Applied Energy	Residential	ML + MOO	ANN

4. Emerging Trends in AI for Urban Building Energy Retrofit

4.1. Spatial Analytics and Urban Morphology Modeling

Urban morphology—including building typologies, spatial layout, and environmental context—plays a pivotal role in shaping energy demand, retrofit effectiveness, and long-term resilience at the district scale. Recent advances in ML have greatly enhanced the quantitative analysis of morphological features, offering scalable, data-driven pathways to prioritize and implement urban retrofit strategies, from archetype classification to generative design and explainable decision support.

Classifying building stock into representative archetypes is essential for city-scale retrofit planning and scenario analysis. Recent advances have shifted from manual surveys and fragmented energy datasets toward scalable, data-driven methodologies. For example, Ali et al. [5] proposed a hybrid ML framework for the Irish residential building stock, integrating ensemble models with physics-based parametric simulation. Their approach automated the creation of a synthetic dataset for one million dwellings by parameterizing 19 key variables across four main residential archetypes. By employing end-use demand segregation and ensemble learning, their model achieved 91% predictive accuracy—substantially outperforming traditional methods—and enabled rapid urban-scale assessment of heating, lighting, equipment, and renewable end-uses. Choi et al. [67] developed a symbolic hierarchical clustering approach leveraging change-point model parameters to capture building energy usage signatures at the city scale. Their method automatically derives energy performance signatures that encode key envelope and operational characteristics as symbolic data, enabling robust clustering of commercial buildings based on both technical and behavioral attributes. Applied to over 1000 buildings in Gangwon, South Korea, this framework distinguished five EPS archetypes, separating high- and low-performance groups, and further revealed behavioral insights, such as the tendency of some occupants in older, poorly insulated buildings to delay heating and thus moderate energy use. Visualizing these clusters across the city enables planners to quickly identify retrofit priorities and tailor strategies to both technical deficiencies and occupant practices. The approach also illustrates the value of open-source, city-scale energy data in supporting nuanced, data-driven urban energy management. Recent work by Nutkiewicz et al. further demonstrates the value of integrating data-driven and simulation-based approaches for large-scale retrofit planning. Their DUE-S framework combines physics-based UBEM with transfer learning to model retrofit scenarios across 52 buildings in a Californian city, enabling rapid evaluation of design alternatives and capturing neighborhood-scale retrofit impacts that traditional ML or simulation methods alone may miss [124]. Building on this, their later study leverages open-access socioeconomic and satellite data to construct archetypal UBEMs and an “urban context vector,” capturing heat island and equity issues at city scale [125]. This work highlights how advanced archetype extraction, when paired with rich contextual data and interpretable simulation models, can inform equitable, climate-adaptive retrofit policy in complex urban environments. Integration of ML and GIS, as detailed by Iseri et al., has standardized archetype development pipelines for large-scale urban applications, allowing for rapid scenario generation directly from remotely sensed or cadastral datasets [63]. DL frameworks—especially CNNs and GANs—have further enabled the extraction of facade and volumetric features such as window ratios and external insulation from images [38,74], now routinely reaching F1-scores above 95% in segmentation tasks. These advances are accelerating the shift toward fully automated, citywide building stock characterization, reducing survey costs and enhancing the scalability of UBER strategies.

ML-driven spatial analytics have revolutionized urban energy mapping, enabling stakeholders to identify, compare, and prioritize retrofits across thousands of buildings or neighborhoods. Early studies, such as Ghiassi and Mahdavi [126] and De Jaeger et al. [64], applied multivariate cluster analysis and geospatial clustering to support scalable citywide modeling, with recent work by Khajedehi et al. [127] demonstrating an 80% reduction in simulation time for Italian districts. At the national scale, Eggimann et al. [65] and Ali et al. [3] combined GIS mapping and DL for energy assessment of millions of buildings, generating high-resolution retrofit priority maps. More recent advances leverage heterogeneous data sources—including remote sensing, smart meters, and crowd-sourced urban data—to further enhance prediction accuracy and spatial coverage [6,120]. In parallel, the integration of DL, particularly CNNs and, increasingly, Transformer-based models, enables rapid, high-resolution mapping of solar potential and spatially explicit retrofit opportunities [11]. At the district level, GNNs and spatiotemporal models have emerged as powerful tools for capturing complex building-to-building interactions and network effects [115]. Notably, Wang et al. [114] demonstrated the successful deployment of a GNN-based urban energy mapping pipeline in a major Chinese city, achieving both high predictive accuracy and operational scalability, with outputs now informing district-level retrofit policy. These advances illustrate the transition from algorithm development to real-world deployment, enabling cities to dynamically update retrofit priorities in response to changing climate, energy prices, and policy mandates.

The convergence of generative and explainable AI is reshaping urban form optimization for energy-efficient and resilient cities. Deep generative models, particularly GANs, are enabling the creation of context-aware, energy-optimized urban layouts by embedding site constraints and performance objectives directly into the design process. Jiang et al. [128] proposed a site-embedded GAN (ESGAN) framework, a generative model incorporating spatial constraints for early-stage urban design, that encodes site constraints and spatial attributes, enabling context-aware, energy-efficient urban layouts in early design stages. Gan et al. [75] introduced UDGAN, which merges stylistic generation with energy-driven morphological optimization, drastically reducing design iteration times and facilitating integration of performance feedback during concept exploration. Further, GAN-based surrogate models enable rapid estimation of spatial performance indicators such as solar radiation, wind, and thermal comfort [129]. This facilitates MOO of urban forms in real time, empowering planners to test dozens or hundreds of scenarios before committing to a design.

While GAN-based generative models for urban form are predominantly at the research and digital prototyping stage, the broader class of AI-driven urban analytics—including graph neural networks (GNNs) for urban building energy modeling—has seen increasing application in real-world engineering contexts. For instance, GNN-based UBEM methods have been successfully deployed for large-scale urban energy prediction, district heating management, and dynamic load forecasting in several pilot projects [76,130,131]. Garg et al. [131] validated UBEM models across over 247,000 buildings in Chicago, demonstrating that careful model calibration enables city-scale predictions to closely match metered data, despite persistent challenges at finer scales. Similarly, Liu et al. [132] and Halaccli et al. [115] have shown that incorporating physically meaningful attributes and spatial dependencies into GNNs can substantially improve energy modeling accuracy and robustness in complex, real-world urban settings. Recent projects in Europe have also reported the use of GNN-UBEM methods in operational digital twin platforms. For example, the EU-funded SPHERE project has integrated ML-enhanced UBEM, including graph-based models, into city-scale digital twins for monitoring, scenario simulation, and evidence-based retrofit planning in cities such as Helsinki and Valencia [133]. Furthermore, digital design platforms such as Autodesk Generative Design and Sidewalk Labs’ Delve have started to incorporate GAN-inspired algorithms for early-stage massing and solar potential analysis, used by planning consultants in actual project bidding and feasibility studies [134].

The need for transparent and interpretable analytics has become paramount as AI-driven models are increasingly used for planning and policy decisions. XAI tools such as SHAP [135] and Local Interpretable Model-Agnostic Explanations (LIME) [136] now make it possible to decode the “black box,” helping stakeholders understand which spatial features most influence energy demand or retrofit suitability. Li et al. [45] combined XAI with spatial proximity analysis and Delaunay triangulation, systematically identifying and quantifying urban morphological factors that drive energy consumption. This supports explicit, data-driven zoning and prioritization at the district or city block scale. Yang et al. [136] applied LIME to interpret ML models for wind power. Still, the framework translates directly to urban energy modeling, allowing policymakers to quantify input feature trustworthiness and improve risk-informed decision-making. Recent trends highlight the convergence of spatial analytics with dynamic data streams, such as IoT sensor networks and climate data, fostering adaptive, data-rich environments for urban retrofit planning. The integration of computer vision, graph neural networks, and explainable AI is enabling new forms of real-time, cross-case learning and rapid scenario adjustment. However, scaling these models to data-sparse or unique urban contexts, harmonizing multi-source datasets, and benchmarking across cities remain major challenges. Future work should prioritize robust model transfer, modular open-source platforms, and cross-city validation, driving toward scalable and resilient AI-powered spatial analytics in urban energy retrofitting.

4.2. ML-Based Energy Forecasting Under Future Climates

Accurate forecasting of building energy demand in the context of a changing climate is imperative for the development of robust, future-proof retrofit strategies. Conventional energy assessments, predicated on historical climate norms, frequently underestimate the impact of intensifying heat waves, evolving humidity profiles, and extreme weather events. Recent advancements in artificial intelligence, particularly in the domains of DL and hybrid time-series modeling, have facilitated the development of dynamic, high-resolution forecasting methodologies. These methodologies are capable of adapting to both gradual climatic shifts and abrupt anomalies.

A foundational challenge in UBER is modeling how annual, seasonal, and sub-daily energy patterns will evolve. ML-based time-series models capture these dynamics, supporting retrofit decisions for HVAC sizing, envelope upgrades, and passive strategies. Classical approaches—such as SVR, RF, and GBM—have proven effective for short-term or event-driven prediction, modeling nonlinear responses to weather fluctuations and operational changes [137,138]. However, these models may struggle to extrapolate under new climate regimes or with highly autocorrelated data, highlighting the need for advanced temporal architectures. Recurrent Neural Networks (RNNs), especially long short-term memory (LSTM) networks, address this gap by retaining long-term temporal dependencies. Koschwitz et al. [139] demonstrated that NARX-RNN models improved district heating forecasts relative to SVR and ARIMA, especially in irregular demand periods. Fan et al. [69] systematically compared inference strategies for deep RNNs using operational building datasets, highlighting the importance of architecture selection in capturing real-world load fluctuations. Luo et al. [72] introduced a modular GA-DNN workflow, where weather profile clustering and sub-model optimization improved weekly prediction accuracy for commercial buildings.

For medium- and long-term forecasting, DL models such as LSTM and Gated Recurrent Units (GRU) have become the methods of choice, excelling in capturing seasonal cycles and long-range load transitions [140,141]. Adding attention mechanisms, as in transformer-based models, further enhances interpretability and the ability to capture non-stationary dependencies. Dai et al. [91] introduced CityTFT, a Temporal Fusion Transformer capable of predicting heating and cooling loads across diverse US cities, achieving a 240-fold speedup over physics-based models with competitive accuracy in unseen climate scenarios. Similarly, CNN-LSTM hybrids have been used to extract local temporal features while retaining long-memory effects, boosting performance in noisy, sensor-driven environments [142].

Spatial-temporal DL is also gaining traction: Hu et al. [76] integrated graph convolutional networks (GCN) with LSTM to simultaneously model spatial adjacency and temporal load variation, outperforming LSTM alone, particularly under sparse occupancy data. generative adversarial networks (GANs) have emerged for rapid surrogate modeling of environmental metrics; for example, Huang et al. [129] used GANs to generate real-time solar, wind, and comfort indicators, enabling multi-scenario evaluation of over 170 urban forms.

A hallmark of cutting-edge UBER research is the integration of future climate scenarios into ML pipelines. Global Climate Model outputs, such as those from CMIP6, are downscaled via statistical or dynamic methods and used to retrain or adapt ML models for future weather [143]. Chen et al. [101] simulated energy use for 59,000 buildings in Changsha under 2050 and 2080 scenarios, revealing that retrofits could reduce demand by 46.55% despite a 4.56% baseline increase under warming. Aliabadi et al. [102] combined micro-genetic optimization with climate-aware physics modeling for Toronto, achieving over 80% gas and 60% electricity savings across projected decades.

Accurate forecasting of building energy demand under changing climate conditions is essential for robust, future-proof retrofit strategies. Beyond static predictions, the most advanced ML frameworks now explicitly address model adaptation to new and unforeseen climate scenarios, leveraging techniques such as transfer learning, domain adaptation (DA), and ensemble simulation. Recent studies have applied transfer learning and domain adaptation to enable ML models trained under historical climate data to generalize to future, altered weather patterns or to transfer between cities with different climate regimes. For example, Sheng et al. leveraged transfer learning to adapt a multimodal neural network trained on data-rich urban regions for energy prediction in data-scarce cities, significantly improving accuracy and robustness in cross-climate application [144]. Gao et al. provide a representative engineering case for transfer learning in building energy applications. By developing and deploying a transfer learning-based multilayer perceptron (TL-MLP) model for thermal comfort prediction across multiple cities in the same climate zone, they demonstrated that models trained with multi-city data can robustly address data scarcity and significantly improve predictive accuracy [145]. Xuereb Conti et al. proposed a hybrid physics-based domain adaptation approach, combining lumped-parameter state-space models with subspace-based domain adaptation (SDA) to transfer knowledge across building systems with differing physical parameters, even under limited labeled data [146]. Gao et al. applied the Adversarial Discriminative Domain Adaptation (ADDA) algorithm for solar radiation prediction across Japanese cities, achieving a 14% improvement in predictive accuracy under zero-label transfer conditions [147]. Such advances are highly relevant for generalizable urban energy forecasting and climate-adaptive retrofit planning in cities facing limited labeled data or rapidly evolving environmental conditions.

Key evaluation metrics—MAE, RMSE, MAPE, and scenario robustness indicators—are increasingly complemented by model interpretability scores (e.g., SHAP values) and domain-relevant KPIs (e.g., avoided carbon, cost savings). State-of-the-art ML forecasting tools now support climate-aligned equipment sizing, carbon scenario assessment, and operational resilience planning. Municipalities are beginning to simulate the energy performance of classrooms or health facilities with next-generation HVAC or envelope systems under 2050+ climate files to optimize for future risk and investment [148,149].

Despite progress, significant challenges persist. Model generalization remains limited across climate zones, building typologies, and operational contexts, with most high-performing models still tailored to case-specific datasets. Data scarcity hinders transferability, especially for long-term, high-frequency urban energy records. Another frontier is the seamless integration of ML-based forecasting within digital twin platforms for continuous calibration, anomaly detection, and adaptive scenario planning. Looking ahead, future research must accelerate the fusion of explainable ML, multi-source climate projections, and real-time data pipelines. Robust, interpretable, and modular forecasting frameworks will be critical for operationalizing climate-resilient retrofit planning at both building and urban scales. Advances in federated learning and privacy-preserving ML could enable more cities to benefit from collective intelligence while protecting sensitive energy and occupant data. ML-based energy forecasting under future climates is evolving from static extrapolation to dynamic, adaptive, and scenario-driven intelligence, reshaping how we future-proof urban built environments.

4.3. Spatiotemporal Coupling for District-Level Planning

Urban buildings seldom operate as isolated entities; instead, they form tightly interwoven systems where shared infrastructure, environmental exposure, and temporal occupancy synchronization drive complex energy demand patterns. Traditional modeling frameworks—often treating spatial and temporal factors separately—have limited capacity to capture the real-world dynamics of urban districts. In response, spatiotemporal ML models have emerged as a transformative approach, coupling graph-based spatial representations with advanced temporal sequence learning to decode multidimensional urban interactions.

Hybrid DL architectures, especially those combining CNNs for spatial feature extraction with LSTM or GRU for sequence learning, are now widely applied in district-level load forecasting. Kim and Cho [150] developed a CNN-LSTM model for predicting household electricity consumption based on appliance-level data, achieving superior accuracy and RMSE compared to traditional ML. Such hybrid models excel at extracting complex correlations among appliances and capturing temporal fluctuations. Attention mechanisms have further improved these frameworks. Yu et al. [104] compared five model variants for district cooling/heating loads and found that attention-enhanced LSTM delivered the best 24-hour-ahead accuracy. At the same time, XGBoost performed better for very short-term forecasts. Visualization of attention weights enabled interpretable tracking of critical time lags and event-driven loads, enhancing practical trustworthiness. More recently, Vontzos et al. [77] extended the hybrid paradigm by integrating GCNs with LSTM, allowing the model to encode both spatial adjacency (e.g., building proximity, shared HVAC networks) and temporal evolution in a multi-zone educational building. The resulting GCN-LSTM model outperformed all tested baselines on MAE and MSE, underscoring the value of authentic dual-domain learning.

GNNs represent a frontier in spatial modeling by encoding buildings as nodes and their physical or operational connections as edges—capturing adjacency, infrastructure sharing, and urban heat feedback [76]. When coupled with recurrent networks (LSTM/GRU), these frameworks model how inter-building dependencies and feedback evolve. Alhussein et al. [142] presented a GNN-LSTM approach that learned how heat waves propagate and amplify across urban districts, enabling scenario-aware, adaptive interventions. Cheng et al. [130] advanced the field with a Spatiotemporal Graph Convolutional Network (STGCN), successfully forecasting campus-scale energy use using heat transfer, solar adjacency, and occupancy-driven effects. Their model significantly surpassed physics-based and conventional ML baselines in predictive accuracy and interpretability, illustrating the transformative potential of spatiotemporal graphs in energy planning.

The effectiveness of spatiotemporal ML models can be dramatically improved by fusing diverse data streams—weather forecasts, future climate projections (e.g., CMIP6), sensor data (temperature, CO₂, occupancy), mobility traces, and high-resolution satellite imagery. Encoder-decoder architectures (CNNs + LSTMs) have proven adept at integrating such heterogeneous data. Liu et al. [151] introduced a dynamically engineered multi-modal feature learning pipeline for forecasting office cooling. Their model integrated weather inputs with building control profiles and compared DL (LSTM/GRU) and AutoML solutions, consistently achieving superior accuracy and stability. The integration of ML with urban digital twins—virtual representations of urban districts that combine real-time sensor streams, historical records, and predictive simulations—is expected to revolutionize retrofit planning and operational control. Such digital twins enable continual calibration, uncertainty quantification, and rapid scenario testing, supporting resilient, adaptive city management.

The practical impact of spatiotemporal ML has been demonstrated in real-world retrofit prioritization and system-level scenario planning. Wang et al. [70] evaluated over 690 buildings and showed that while individual building forecasts using LSTM reached $R^2=0.57$, aggregation at the district scale improved accuracy to $R^2=0.95$, demonstrating error-canceling and the power of topology-aware models. Hu et al. [76] applied a solar adjacency-aware STGCN to a university campus and achieved a remarkable MAPE of 5%, particularly improving predictions for buildings with high interconnectivity. This highlights how “network effects” and spatial topology can be leveraged for robust urban-scale forecasting. On the city level, Pasichnyi et al. [35] developed a UBEM-driven approach for Stockholm, evaluating three retrofit packages for post-war multifamily buildings, which led to potential annual savings of 334 GWh and 19.6 kt-CO₂. This district-level approach supports strategic resource allocation for maximal impact. Li et al. [50] demonstrated that incorporating microclimate effects (Urban Heat Island, UHI) using Urban Weather Generator (UWG) adjustments can avoid systematic overestimation of retrofit savings by up to 25%, with even higher discrepancies (up to 60%) in commercial districts. This underscores the importance of UHI-aware modeling for dense urban regions. Key engineering evaluation metrics now include spatiotemporal resolution, predictive accuracy (MAE, RMSE, MAPE), model generalizability, and scenario robustness (across climate, policy, or occupancy changes). These indicators inform operational and strategic decisions for city planners and energy managers.

5. Challenges and Research Directions

While the integration of AI and ML has revolutionized UBER research—enabling unprecedented advances in prediction, optimization, and adaptive management—translating these innovations into city-scale, resilient practice remains fraught with technical, methodological, and operational challenges. This section systematically examines outstanding barriers and emerging opportunities spanning four core domains: (1) advancement and generalization of AI/ML algorithms, including hybrid and interpretable models; (2) climate-aware, high-resolution modeling under data and computational constraints; (3) platform interoperability and deployment at scale; and (4) data infrastructure, stakeholder engagement, and real-world validation.

Addressing these challenges will require holistic, cross-disciplinary collaboration—bridging data science, urban engineering, climate science, and policy—supported by open, modular tools and rigorous empirical validation. The following subsections synthesize current limitations, highlight priority research directions, and offer a roadmap toward robust, equitable, and future-proof AI-enhanced urban energy retrofitting.

Figure 7 provides a visual synthesis of research saturation and future methodological opportunities across the AI-driven UBER landscape, serving as a roadmap for prioritizing research and innovation. The clustered heatmap highlights areas of intense activity—such as RF and XGBoost for energy forecasting—and underexplored opportunities, including GNNs and Transformers for district-level and spatiotemporal modeling.

Figure 7. Clustered heatmap summarizing the research saturation and future methodological potential of AI/ML methods across key retrofit-related tasks in urban building energy retrofit (UBER). Rows represent task domains and subdomains, columns are grouped by algorithm type, with two metrics: research saturation (1–4) and potential (1–5). The matrix reveals both areas of intense activity (e.g., RF, XGBoost in energy forecasting) and underexplored opportunities (e.g., GNN, Transformer in district-level and spatiotemporal modeling), providing a roadmap for future methodological innovation.

5.1. Advancing Algorithmic Diversity and Hybrid Modeling

While ML has enabled real-time monitoring and adaptive control for post-retrofit operations such as HVAC scheduling and IEQ management, its deployment remains heavily reliant on conventional models like RFs and ANNs, which often fail to generalize across diverse building types, occupancy patterns, and climate scenarios.

Recent advances in DL—such as CNN, RNN, GNN, and Transformer-based models—offer superior capabilities in capturing nonlinear, high-dimensional, and spatiotemporal dependencies [92]. Despite this promise, adoption in UBER remains limited due to sparse labeled data, high computational costs, and poor interoperability with physics-based models [40,55]. Hybrid modeling frameworks that fuse data-driven surrogates with UBEM engines, and emerging techniques such as transfer learning, explainable XAI, and temporal fusion transformers, represent promising directions for achieving robust, interpretable, and scalable retrofit planning.

To address the challenge of generalizability, especially in small-sample or cross-city contexts, recent studies have begun exploring transfer learning, federated learning, and active learning paradigms. For example, Sheng et al. [144] developed a multimodal neural network and transfer learning framework that leverages both tabular and visual data, enabling accurate energy consumption prediction even in data-scarce regions. Their approach achieved up to 63.6% error reduction compared to traditional mono-modal models and significantly improved generalizability across cities. By incorporating explainable AI for model interpretation, they further identified key physical attributes—such as floor and wall insulation—as critical drivers, offering actionable guidance for data-driven retrofit and policy decisions in diverse urban contexts. Similarly, federated learning frameworks enable collaborative model training across multiple cities or building owners, enhancing model robustness without sharing sensitive data. Active learning approaches prioritize the most informative samples for labeling, further reduce annotation costs, and accelerate model adaptation to novel urban typologies [152].

5.2. Climate-Aware Modeling and Resolution Gaps

The necessity to adapt retrofit strategies to future climate variability is becoming increasingly apparent. However, the majority of current UBER workflows still rely on Typical Meteorological Year (TMY) datasets, which fail to capture nonstationary trends, urban microclimates, and extreme weather events [32,153]. Phenomena such as Urban Heat Island (UHI), prolonged heatwaves, and humidity anomalies are often excluded, leading to systematic underestimation of thermal loads and misaligned design interventions [50]. Spatial and temporal mismatches among input datasets are equally limiting. While microclimate simulations can achieve sub-hourly and sub-block granularity, geographic information systems (GIS), LiDAR, and satellite imagery often vary in resolution and format, impeding consistent geometric, radiative, and occupancy alignment. These resolution discrepancies compromise the fidelity of ML-enhanced predictions and impede the representation of intra-urban heterogeneity, particularly in informal settlements or mixed-use zones. A promising direction lies in the fusion of multi-source climate data, including ground observations, meteorological reanalysis (e.g., ERA5, MERRA-2), remote sensing (e.g., MODIS, Landsat), and even crowdsourced sensor networks [154,155]. Such fusion enables high-resolution mapping of microclimatic variables and local extremes, supporting risk-aware and site-specific retrofit design. Integrating these diverse sources requires standardized, scalable data assimilation frameworks and careful quality control, which are still emerging research frontiers.

The computational burden of climate-aware UBEM at urban scale often results in simplifications or data aggregation that erase local nuances. Most prevailing workflows do not adequately support standardized integration of downscaled climate projections (e.g., CMIP6, CORDEX) with ML-enhanced retrofit decision-making. The absence of interoperable pipelines inhibits model transferability, reproducibility, and robust cross-site benchmarking. Another emerging need is the development of dynamic, adaptive models capable of responding to evolving climate signals and unforeseen extreme events. Techniques such as online model calibration, sequential Bayesian updating, and weather-triggered scenario simulation enable continuous adaptation and improved robustness for retrofit planning [156]. For example, transfer learning can recalibrate urban energy models as new climate data or unprecedented heatwaves occur, thereby improving model relevance under nonstationary conditions.

Recent studies also highlight the potential of physics-informed ML and generative models (e.g., GANs, VAEs) to synthesize localized extreme event scenarios and improve sample efficiency for rare but impactful climate conditions [157]. These approaches support the design of resilient retrofit strategies that can withstand a broader range of future uncertainties.

In summary, addressing climate-aware modeling and resolution gaps will require advances in multi-source data fusion, dynamic self-adaptive ML pipelines, and robust uncertainty quantification frameworks. Future research should prioritize the development of lightweight microclimate proxies, standardized data assimilation workflows, and scenario-based decision support tools, all integrated within open and interoperable urban simulation ecosystems. Such innovations will enable climate-resilient and future-proof urban retrofits at scale.

5.3. Platform Interoperability and Deployment Barriers

The computational and data infrastructure supporting ML-assisted UBER is deeply fragmented, with simulation tools (e.g., EnergyPlus, URBANopt, CityBES, SimStadt) and ML frameworks (e.g., TensorFlow, PyTorch) often developed in silos and lacking standardized interfaces for cross-platform communication [1,108]. This isolation hampers model integration, collaboration, reproducibility, and especially scalability from pilot projects to citywide applications.

A core barrier is the absence of unified toolchains and open-access APIs. Many workflows depend on bespoke scripts or ad hoc coupling strategies, which are hard to extend or maintain. For example, although OpenStudio-URBANopt and CityBES support plug-in architectures and open APIs, their interoperability with other urban platforms or BIM/GIS standards (such as IDM/IFC) is limited [106,108]. International efforts are underway to advance cross-platform compatibility and data sharing, such as the European Union’s drive for standardized urban energy platforms and recent open-data mandates in China, the US, and Japan. However, adoption remains patchy across regions and sectors.

Recent advances in digital twins, federated learning, and cloud-edge computing architectures are promising for enabling scalable, privacy-aware, and real-time urban energy management [97]. European “smart city” demonstration projects increasingly deploy digital twins and edge-cloud solutions for real-time urban analytics and adaptive control. This illustrates industrial uptake and the policy/technology co-evolution trend.

To address these deployment barriers, future research should prioritize the development of modular, API-compatible toolkits built on standardized data schemas, interoperable simulation engines, and lightweight wrappers for model communication. Emphasizing platform standardization, shared benchmarking datasets, and open-source collaborative environments will be essential for transitioning from academic prototypes to resilient, scalable UBER solutions. Additionally, coordinated policy support and international alignment are critical to accelerate industry adoption and ensure equitable access to digital infrastructure in urban energy retrofits.

5.4. Data, Decision Support, and Real-World Validation

The advancement of algorithms alone is insufficient to achieve a scalable and trustworthy AI-enhanced UBER. Robust data infrastructure, stakeholder-oriented decision tools, and comprehensive empirical validation are all crucial for bridging the gap between digital innovation and real-world transformation. Despite progress in simulation-based workflows, critical gaps persist in data quality, cross-sector communication, and practical field-testing. Recent progress in IoT sensors, smart meters, Building Management Systems (BMS), and high-resolution remote sensing (e.g., satellite and LiDAR imagery) now enables the collection of fine-grained, dynamic datasets for individual buildings and entire urban districts. Integrating heterogeneous data sources, including crowdsourced measurements, historical archives, and real-time weather feeds, can substantially improve the spatial and temporal granularity of energy models [11,47]. However, many urban areas remain under-instrumented, especially in developing regions or informal settlements. Data gaps, asynchronous streams, and noisy measurements present further challenges, necessitating the use of advanced data imputation and uncertainty quantification strategies, such as Bayesian updating, multiple imputation, and Monte Carlo simulation. These techniques enable closed-loop calibration and robust performance monitoring, enhancing the adaptability and reliability of AI-driven UBER solutions.

Effective translation of technical insights into actionable planning requires AI-powered Decision Support Systems (DSS) that address the needs of multiple stakeholders—policymakers, building owners, urban planners, and residents. Modern DSS platforms increasingly integrate ML-based forecasting with MOO, providing transparent visualization of trade-offs among energy use, cost, emissions, occupant comfort, and social equity [45,158]. Interactive dashboards, participatory interfaces, and scenario-based analytics—often underpinned by digital twin frameworks—are becoming essential for climate-resilient urban design and adaptive management. Incorporating XAI and sensitivity analysis is critical to foster trust, facilitate regulatory approval, and support risk-aware investment, particularly in high-stakes contexts such as urban policy or infrastructure funding.

Despite advances in surrogate models and simulation engines, large-scale empirical validation remains limited. Most existing studies are based on single-city or laboratory-scale case studies, constraining the generalizability of findings. There is an urgent need for longitudinal field trials, cross-climate demonstration projects, and multi-site behavioral impact studies to anchor ML-driven retrofit strategies in operational realities. Adopting transfer learning and federated learning strategies has shown promise for leveraging models trained in data-rich cities to support retrofit planning in smaller or under-instrumented urban areas. Combined with open data initiatives, cross-urban and cross-climate benchmarking can further improve model transferability and foster global best practice sharing.

5.5. Limitations and Critical Barriers

Despite rapid methodological progress, several fundamental limitations constrain the real-world impact of AI-enhanced UBER workflows.

First, the risk of overfitting—particularly for DL models trained on limited, noisy, or highly imbalanced urban datasets—remains a critical concern, often leading to optimistic in-sample accuracy but poor generalizability across diverse building typologies, climates, or operational regimes [159]. Model complexity must therefore be balanced with regularization, cross-validation, and, where possible, external benchmarking.

Second, practical deployment at city scale is challenged by the deployment gap: many published models are validated only on simulated or laboratory datasets, with limited evidence from large-scale, real-world field trials or operational energy data [160]. Bridging this gap will require the integration of empirical monitoring, participatory calibration, and continual model updating—moving beyond proof-of-concept toward robust, replicable, and policy-relevant solutions.

Third, the computational and data requirements of high-fidelity simulation and ML models pose significant scalability barriers, particularly for resource-constrained municipalities or developing regions. Simplified or surrogate models may be necessary, but such reductions often come at the cost of lost accuracy, reduced explainability, or untested assumptions [32]. Thus, the trade-off between model fidelity, computational cost, and actionable insights must be carefully evaluated.

Last but not least, ongoing challenges include data heterogeneity, proprietary data silos, and privacy concerns—each of which can limit both the reproducibility and transferability of AI-driven retrofit frameworks [161,162,163]. Advancing toward open, interoperable, and federated data platforms, while respecting data privacy and governance requirements, is essential for building trust and fostering widespread adoption. As AI-driven systems become more prevalent in urban retrofit decision-making, several ethical considerations must be addressed to ensure trustworthy and equitable outcomes [164,165]. First, the risk of algorithmic bias arises when models are trained on data that may not adequately represent all socioeconomic groups, building typologies, or climate contexts, potentially resulting in unfair allocation of retrofit resources or benefits. Second, issues of fairness and transparency are critical when multi-objective optimization frameworks inform policy or investment decisions that affect diverse stakeholders. Without explicit mechanisms for auditability, model decisions may remain opaque, hindering accountability and stakeholder trust. Third, the increasing use of high-resolution urban data (e.g., smart meters, IoT, mobility traces) raises significant privacy concerns, especially in jurisdictions with strict data protection regulations. Emerging technical solutions such as federated learning, differential privacy, and privacy-preserving analytics require institutional support and careful design to ensure compliance. Ultimately, advancing AI-enabled UBER will require technical innovation and interdisciplinary collaboration to establish ethical standards, transparent governance, and robust safeguards that protect the rights and interests of all urban residents.

In sum, while AI-enabled UBER holds substantial promise, realizing its full potential demands critical attention to these methodological, practical, and institutional limitations. Future research must prioritize robust, cross-domain validation, transparent reporting of model uncertainty, and the development of adaptive, scalable frameworks that can thrive under real-world constraints. Addressing these limitations will be essential to unlock the full transformative potential of AI-powered urban retrofitting.

6. Conclusions

This review comprehensively synthesizes advances, trends, and challenges in applying AI—including ML, DL, and heuristic optimization—to UBER. By systematically mapping retrofit objectives and strategies across building and district scales, and benchmarking state-of-the-art AI methods against traditional workflows, we demonstrate how data-driven and hybrid modeling approaches transform energy forecasting, multi-objective decision-making, and climate-adaptive planning.

Key advances include integrating surrogate modeling and spatiotemporal DL for rapid, high-fidelity prediction; the application of XAI for transparency and stakeholder trust; and emerging generative and graph-based models for urban-scale optimization. Despite these breakthroughs, persistent challenges remain—including data heterogeneity, limited generalizability across regions, fragmented computational ecosystems, and the absence of holistic frameworks uniting technical, economic, and policy objectives.

To address these barriers, we recommend prioritizing the development of open, interoperable, and uncertainty-aware frameworks that integrate physics-based simulation, high-resolution data streams (from IoT, BMS, and remote sensing), and participatory decision-support tools. Scaling validated AI strategies will require rigorous empirical validation, standardized datasets, and deeper engagement with policymakers and industry.

The accelerating evolution of generative models, digital twins, and climate-resilient urban analytics will further reshape the methodological landscape. Interdisciplinary collaboration and policy–technology alignment will be essential to translate AI innovation into resilient, equitable, and net-zero urban environments worldwide.

These advances offer policymakers and city planners actionable pathways to evidence-based, transparent, and participatory urban energy planning. AI-driven, multi-objective optimization can systematically prioritize retrofit investments in complex trade-offs, balancing energy savings, cost, carbon reduction, and occupant well-being. Adopting explainable AI and digital twins enhances stakeholder engagement, policy transparency, and real-time monitoring, while scenario-based analytics support adaptive, risk-informed decision-making in the face of climate uncertainty. Institutionalizing these capabilities will enable cities to efficiently implement high-impact interventions, equitably allocate resources, and ensure the long-term resilience of urban energy systems.

AI-driven UBER marks a transformative frontier for sustainable cities—enabling smarter, faster, and more inclusive retrofit pathways. Next-generation AI-augmented systems can become a cornerstone of digital, climate-adaptive urban governance by closing the gap between technical promise and real-world impact.