Generative AI for Sustainable Smart Environments: A Review of Energy Systems, Buildings, and User-Centric Decision-Making

Abstract

The rapid evolution of Generative Artificial Intelligence (GenAI) is reshaping the energy sector, enabling new levels of adaptability, efficiency, and user-centric interaction. This review systematically maps and critically evaluates the chosen literature across buildings, grids, and urban systems. Through major scientific databases and for the span of five years, from 2021 to 2025, the review aims to identify key application domains, synergies, and research gaps. The analysis on recent advancements illustrates how GenAI enhances energy forecasting, demand–response strategies, anomaly detection, and cyber-resilience in power networks, while also supporting predictive modeling and optimal control in distributed renewable integration. Within smart buildings, GenAI empowers autonomous agents and AI copilots to balance comfort with energy efficiency through adaptive environmental control and user preference modeling. At the grid level, generative models improve renewable generation forecasting, grid stability, and decision support for operators. A further emerging application lies in the generation of synthetic energy data, which supports model training, scenario simulation, and robust decision-making in data-scarce environments. In the broader context of smart cities, GenAI-driven digital twins, multi-agent systems, and conversational interfaces facilitate sustainable planning and energy-aware citizen engagement. A central theme across these applications is the alignment of technological solutions with human needs and sustainability objectives. Key challenges remain in uncertainty quantification, trustworthy deployment, and data governance, underscoring the need for secure, adaptive, and human-centered GenAI systems to drive the next generation of intelligent energy management. This review provides a comprehensive analysis to promote a better understanding of generative models as they are being applied in a variety of scenarios in the energy domain.

1. Introduction

The transition toward sustainable smart environments has become an integral part of global climate and energy strategies, since urban areas account for more than $65\%$ of the global energy consumption and more than $70\%$ of the associated carbon emissions [1]. Smart grids, intelligent buildings and decision-support systems are the core of this transformation, integrating renewable energy sources with advanced digital infrastructures. However, the complexity of these environments presents an array of challenges, from fluctuating renewable generation and nonlinear demand patterns, to interoperability issues of heterogeneous sources [2].

To address these challenges, a range of conventional approaches to optimization and control were employed and proved effective in narrow applications, but struggled to capture the full complexity of sustainable smart environments. In earlier energy systems, methods like statistical forecasting, optimization heuristics and support vector regression were employed for load and renewable forecasting [3,4,5,6,7,8] and over the past years reinforcement learning (RL) has improved building energy management, especially in HVAC systems [9,10].

Similarly, digital twin (DT) technologies expanded the analytical capacity of buildings and cities, creating real-time virtual replicas of physical assets and thus gathered momentum in predictive maintenance, energy efficiency and occupant comfort applications [11,12,13]. These methods, though efficient, remain constrained by data scarcity due to their large domain-specific data requirements, and often struggled when confronted with dynamic uncertain conditions.

Despite advancing in this domain, these approaches were largely predictive and reactive but lacked the creative and generative capabilities to propose novel strategies, fill data gaps or simulate complex “what-if” scenarios. Taking these shortcomings into account, there is an urgent need for alternatives that not only predict and optimize, but also generate adaptive, human-aligned solutions for sustainable environments.

Generative AI (GenAI) marks a paradigm shift by introducing models capable of not only analyzing but also creating new, plausible data, strategies and scenarios. It introduces transformative possibilities that extend far beyond the predictive and optimization boundaries of conventional AI, leveraging architectures such as Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs) and Large Language models (LLMs), and thus addressing the inherent complexity of sustainable smart environments [14,15,16,17].

One of the principal virtues of GenAI lies in scenario generation and synthesis. Unlike predictive models that extrapolate existing trends, GenAI can generate plausible future scenarios by modeling uncertainty and variability in renewable energy production, user behavior or climate impacts [18,19]. This capacity is particularly relevant for uncertainty modeling in the energy domain, where system behaviour is governed by stochastic weather patterns, fluctuating demand profiles, and heterogeneous distributed energy resources. Generative models learn the underlying joint probability distributions of these complex variables, allowing them to capture rare events, nonlinear interactions, and high-dimensional dependencies that traditional statistical approaches often fail to represent. By sampling from these learned distributions, generative methods produce multiple plausible futures rather than a single deterministic estimate, enabling probabilistic forecasting, risk-aware planning, and resilience assessment in modern energy systems.

Another strength of GenAI is data augmentation. Many sustainable smart environments face the challenge of incomplete, sparse or noisy datasets, originating from sensor malfunctions, limited pilot studies or heterogeneous system integrations. Therefore, GenAI can synthesize realistic datasets, thereby enhancing model training, improving generalization and accelerating the deployment of intelligent energy and building systems [14,20,21]. Furthermore, GenAI excels in multi-objective optimization [22,23], since smart environments must often reconcile competing goals, such as minimizing energy consumption, while maximizing occupant comfort or ensuring grid stability, while integrating intermittent renewable sources. Generative models can propose novel configurations of system parameters, building operations, energy flows, or comfort strategies, that conventional optimization might overlook. GenAI also facilitates interactive decision-making. Through natural language interaction, user preference modeling and adaptive personalization, GenAI systems can bridge the gap between technical optimization and human experience. With the advent of LLMs, users can articulate their needs in natural language, and AI systems can translate them into technical adjustments in real time, creating a more intuitive interface between people and infrastructure [16,17].

1.1. Literature Analysis Approach

This review paper adopts an extensive analysis of the emerging role of GenAI in sustainable smart environments. Using targeted searches across scientific databases, this review synthesizes literature from 2021 to 2025, across domains of energy systems, building optimization, digital twins and human-in-the-loop decision-making and studies are categorized according to the generative architecture employed (language models, diffusion models, GANs), the application context (smart buildings, smart grids, user-centric systems) and sustainability outcomes (energy efficiency, emissions reduction, comfort optimization). Our approach is methodical, ensuring each selected article receives thorough attention, and includes the following steps:

• Criteria for integrated papers: Peer-reviewed journal or conference articles, preprints with substantial methodological contributions, and studies explicitly applying or evaluating GenAI techniques within the energy or build environment domains.
• Related keywords: “Generative AI”, “Large Language Models”, “Digital Twins”, “Smart Buildings”, “Energy Grids”, “Sustainable Environments”, “User Preferences”, “Reinforcement Learning”, “Urban Energy Systems”.
• Article selection: Literature was identified through systematic searches in Scopus, Web of Science, IEEE Xplore and Google Scholar. Relevant studies were screened based on titles, abstracts, and full texts.
• Data extraction: Key information was extracted, including publication year, model type, application domain, methodology, and reported outcomes.
• Quality assessment: Articles were evaluated based on methodological rigor, clarity of problem definition, reproducibility, and relevance to GenAI applications.
• Data analysis: The extracted data were thematically categorized and analyzed to identify methodological patterns, technological trends, application domains, and research gaps.

Hence, this study provides a detailed discussion of GenAI’s role in sustainable energy environments, while explaining the challenges and limitations of conventional approaches to modern applications, thus guiding scholars to improve existing methodologies. To ensure methodological transparency and reproducibility, the literature search followed a structured and replicable strategy. Searches were conducted in Scopus, Web of Science, IEEE Xplore, ScienceDirect, and Google Scholar databases, covering the period 2021–2025, which corresponds to the emergence of generative AI applications in the energy and built-environment domains. Example search strings included combinations of: (“Generative AI” OR “Generative Artificial Intelligence” OR “GAN” OR “VAE” OR “diffusion model” OR “LLM” OR “large language model”) AND (“energy systems” OR “smart building” OR “smart grid” OR “digital twin” OR “user preference” OR “decision support”).

Inclusion criteria comprised peer-reviewed articles, conference papers, and preprints with explicit methodological or application-level use of GenAI in energy, buildings, or sustainability contexts. Exclusion criteria involved works focusing solely on traditional AI/ML without generative components, studies lacking reproducible methodology, and purely theoretical or opinion pieces.

The initial screening yielded many records, of which 72 studies met the inclusion criteria after full-text review. These were thematically categorized by application domain (buildings, grids, user-centric systems), generative architecture, and sustainability outcome, forming the core corpus analyzed in this review.

1.2. Paper Structure

The structure of the paper is organized to provide a clear, thematic, and chronological overview of the role of Generative AI in sustainable smart environments. Section 2 presents an overview of the GenAI theoretical foundations. Section 3 presents the core body of the analysis, structured around three major GenAI application domains. The first domain focuses on Smart Buildings and Smart Cities, further divided into four subcategories that examine applications related to building energy management, HVAC control and optimization, urban-scale digital twins, construction and operation optimization and integrated urban planning. The second domain addresses Energy Grids and Renewable Systems, also comprising four subcategories, which explore GenAI applications for grid stability and sustainability, data privacy and cybersecurity, forecasting and scenario generation and fault and anomaly detection and energy community management. The third domain centers on User Preferences and Decision-Making, highlighting human-in-the-loop frameworks, personalized recommendations, and adaptive control strategies that bridge technology with user behavior. Section 4 synthesizes insights across these domains, identifying major research tendencies and future directions, including methodological evolutions, thematic shifts, and interdisciplinary convergence. Finally, Section 5 concludes the paper by summarizing the key findings and outlining promising avenues for future research and development in the field.

2. Theoretical Foundations of Generative AI in Energy Systems

The evolution of GenAI in the energy sector reflects a shift from algorithmic experimentation to application-driven deployment. Initial studies primarily demonstrated the feasibility of using generative models to simulate or reconstruct energy-related data, focusing on problems such as load profile synthesis, renewable forecasting, or anomaly detection. These early works relied on controlled datasets and focused on improving accuracy and data quality, rather than on system-level integration and management. Progression towards more complex problems requires emphasis on robustness and uncertainty representation, as real-life physical systems are inherently defined by stochastic behaviors and dynamic variability. System-level efforts emphasize interpretability, physical consistency, and integration with optimization and control workflows.

The choice of a particular generative family, such as VAEs, GANs, diffusion, autoregressive and transformer models or hybrid approaches, is defined by the task objective, the data characteristics and the practical constraints. For example, VAEs are able to learn smooth probabilistic latent representations and are particularly useful where a compact, interpretable latent space and principled uncertainty estimates are required. In energy applications, VAEs have been used in hybrid forecasting and anomaly-detection workflows [24], where latent variables help denoise measurements and provide uncertainy-aware reconstructions.

GANs and time-series variants have been popular for creating realistic synthetic energy records, such as household or building consumption data, augmenting datasets, and simulating rare events for stress-testing algorithms. Time-variant GANs have been used to generate these data while preserving temporal dynamics and protecting privacy [20]. When sample realism should be prioritized, GANs are often preferred, as they produce the most realistic outputs. However, they do so at the cost of training instability and reduced capacity for calibrated likelihoods or direct uncertainty quantification.

Diffusion-based models, such as denoising diffusion probabilistic models or score-based models, have more recently been introduced to energy forecasting, as they combine high sample quality with explicit probabilistic sampling, suitable for scenario generation and forecasting of renewables or load residuals. Key energy-focused demonstrations show competitive performance against GANs/VAEs for probabilistic scenarios [25].

Autoregressive and transformer-based models adapted from Natural Language Processing (NLP) are well suited to energy time-series tasks that require modeling of long-range temporal dependencies, such as electric loads, prices and control sequencies. Transformer adaptations for load forecasting demonstrate improved accuracy and scalability relative to Recurrent Neural Network (RNN)-based approaches by capturing long context windows and enabling parallel training [26].

To incorporate realism, interpretability, and trustworthiness, hybrid approaches that integrate different generative model families have seen increasing adoption. By embedding domain knowledge and physics-based constraints within data-driven architectures, these approaches ensure safety, regulatory compliance, and adherence to operational limits that pure data-driven generators struggle with. They combine the strengths of each model family, leading to improved efficiency, robustness, and accuracy in addressing complex energy-related challenges [27].

3. Applications of GenAI Across Smart Buildings, Energy Grids, and User-Centric Systems

3.1. Smart Buildings and Smart Cities: Comfort, Management, and Design Optimization

Generative Artificial Intelligence (GenAI) has emerged as a transformative paradigm for the smart building and grid environments, influencing how buildings and cities are designed, operated, and managed. Applications range from indoor comfort optimization and energy modeling to AI copilots for building management, digital twins for resilience planning, and construction process optimization. This section provides an extended review of recent contributions, highlighting both opportunities and challenges in deploying GenAI for smart buildings and smart cities.

Overall, studies in this domain demonstrate that LLMs and generative co-pilots outperform conventional predictive tools by enabling semantic reasoning, conversational decision support, and automated model generation for building management. GAN- and VAE-based methods, while still valuable for simulation and data augmentation, remain less effective for real-time or context-aware control. Across works such as [28,29,30], a clear research gap persists in integrating generative reasoning with sensor-driven control loops, ensuring explainability, and benchmarking model performance under dynamic occupant and environmental conditions.

3.1.1. Thermal Comfort and Indoor Environment Control

A recent study has investigated the use of GenAI to assist building management with the core task of maintaining thermal comfort, where one line of work introduced a chatbot framework for thermal comfort monitoring based on LLMs with retrieval augmentation, which integrates Building Information Modeling (BIM), Internet of Things (IoT) sensor data, weather forecasts, and regulatory standards, allowing occupants or managers to query comfort levels in natural language, proactively identify anomalies, such as overheating or excessive humidity, and propose actionable measures to address indoor conditions. Despite its promise, limitations include reliance on a limited set of environmental variables, assumptions about sensor placement, and the use of a single open-source LLM, which may reduce accuracy. Future extensions could incorporate more contextual data such as occupancy and air quality, while fine-tuned models could enhance domain-specific reliability [28]. Complementary approaches emphasize autonomy in environmental control. By combining Microsoft’s Semantic Kernel with generative models, a framework has been demonstrated for agentic smart home systems capable of orchestrating devices such as HVAC, lighting, and air quality systems. Unlike rule-based automation, this architecture enables memory, planning, and contextual reasoning, allowing the system to learn occupant preferences and adapt dynamically. For instance, the agent can dim lights, adjust thermostats, or respond to emergencies such as fires, without explicit user input, translating high-level goals (“activate energy-saving mode”) into sequences of coordinated actions. Such advances signal a shift toward more proactive and personalized environmental control, though concerns about reliability, safety, and verification in real-world deployments remain open challenges [31].

Focusing on building energy management, Glaws et al. (2025) introduced BuildingsBench [32], a research platform designed to advance AI methodologies for short-term load forecasting (STLF). The proposed framework assists in the development and evaluation of AI models which generate accurate load profiles for individual buildings, without the need for extensive historical data. The platform provides an extensive dataset, with over 900,000 building simulation models, to better represent a significant portion of the U.S. building stock. AI models developed within BuildingsBench can assist grid operators to dynamically match energy supply to demand, while providing deeper insight into consumer behavior and energy consumption patterns. Furthermore, the platform has the ability to generate synthetic building models to support community-level energy transition planning. In the same study, researchers examine the performance of Graph Recurrent Attention Networks (GRANs) compared to classical graph generation algorithms to showcase how GenAI can be used to create new topologies for power system infrastructure. Additionally, Jurišević et al. explore GPT-based reasoning for energy management in kindergartens, showing how LLMs can interpret metering and comfort information to recommend operational adjustments. Their study highlights both the promise and practical limitations of such ‘digitainability’ workflows in real buildings [33].

3.1.2. GenAI Copilots and Agentic Interaction in Building Operations

Generative AI is also being positioned as an intelligent copilot for building operations. One noteworthy advancement is an AI copilot that protects privacy by integrating with IoT systems and generating semantic data models. The system shares metadata and uses the LLM to construct queries through SPARQL or Python code that are ran locally, rather than exposing raw building data to external models. This guarantees the security of private energy or occupancy data. The design is further strengthened by a multi-agent framework, which includes specialized agents for orchestration, debugging, and code generation. While the copilot automatically obtains and processes pertinent data, end users can engage with the building by posing natural language queries, such as asking for energy usage summaries or identifying rooms that surpass specific criteria [34]. By safeguarding privacy and reducing hallucination risks, this design lowers the barrier to accessing building analytics. However, deployment remains complex due to the need for semantic standardization and extensive prompt engineering. At a broader scale, similar copilot principles have been extended to urban environments, where GenAI systems act as intermediaries between citizens, city operators, and planners. A recent survey highlights how user-centric GenAI copilots harness IoT data and Urban Digital Twins (UDTs) to provide conversational interfaces for city services, energy management, and mobility. Retrieval-augmented generation is emphasized as a key method for grounding outputs in authoritative datasets, reducing hallucination risks while enabling transparent, auditable interaction. These copilots allow residents to query municipal data in natural language, operators to optimize urban systems, and planners to simulate policy impacts, thereby scaling the copilot concept from the building level to the city level [35]. This illustrates how GenAI copilots can evolve into multi-layered decision-support tools, enabling seamless human–AI collaboration across different levels of the built environment.

3.1.3. Construction Process and Operational Optimization Through GenAI

GenAI is increasingly helping to bridge the long-standing digital adoption gap within the construction sector. Text-based generative models are being used to automate project management processes, improve safety and risk assessment, and streamline documentation. Jurišević et al. (2025) assess the role of LLMs in public building energy management, identifying both the operational opportunities and institutional barriers of integrating LLMs into construction and facility workflows. Their findings emphasize the importance of domain adaptation and human oversight for reliable deployment [36]. Additionally, applications include generating and reviewing contracts, extracting insights from large building codes, and enhancing communication through natural language interfaces. Furthermore, prototypes like “BIM-GPT” have been utilized to more effectively query BIM databases, enabling practitioners to work with design models without requiring specialised technical expertise. Nevertheless, significant challenges remain. Current LLMs lack the specific domain knowledge required to ensure accuracy in areas like cost estimation or regulatory compliance, and construction processes are data-rich but extremely fragmented. Risks of hallucination or misinterpretation can have severe consequences, highlighting the need for fine-tuned models and human oversight [37]. On a larger scale, a variety of applications across the lifecycle are revealed by systematic assessments of GenAI in architecture, engineering, construction, and operations (AECO). These include BIM integration, structural engineering, construction management, operations, project briefing, early design creation, and even urban governance. Collectively, these studies demonstrate that GenAI has the potential to augment creativity, automate routine processes, and improve efficiency throughout the AECO domain. However, adoption remains limited and fragmented, with most applications still at prototype level and lacking rigorous validation. Progress is further hampered by the lack of uniform datasets and benchmarks. Interdisciplinary cooperation, empirical analysis, and frameworks to direct the responsible integration of GenAI in the construction environment are urgently needed [29].

To support early-stage creation of Building Information Modeling (BIM) models, researchers have proposed Text2BIM, a multi-agent LLM framework that translates natural language into native, editable 3D BIM models [38]. The core innovation involves representing a BIM model as imperative code scripts for a software API (e.g., Vectorworks). The framework coordinates multiple LLM agents to collaboratively generate this executable code, which then produces a model directly within the authoring tool. It integrates a rule-based model which provides iterative feedback to the LLM agents and allows them to autonomously resolve design conflicts and improve the quality of the trained models against domain-specific rules. Text2BIM was validated as a feasibility study and demonstrates a promising “modeling-by-chatting” approach, aiming to facilitate designers during the conceptual design phase, while ensuring the output allows further human refinement.

3.1.4. GenAI-Enhanced Digital Twins for Urban Planning and Resilience

GenAI is increasingly being used with digital twin platforms at the city level to improve disaster management and resilience. While generative models enhance digital twins by creating synthetic data and testing various situations, digital twins offer virtual versions of actual systems that are updated in real time with sensor data. This combination makes it possible to predict complicated urban occurrences like pandemics, fires, and floods with greater accuracy. Urban planners can stress-test infrastructures and find vulnerabilities before crises arise, for instance, by using GenAI-enhanced digital twins that can replicate innumerable catastrophic event scenarios [39]. Additionally, generative models increase simulation accuracy by producing high-resolution situations that closely resemble the intricacy of natural disasters. This makes it possible for emergency managers to better deploy resources, optimize evacuation plans, and virtually practice responses. However, there are still difficulties in managing computational cost, guaranteeing the validity and realism of synthetic scenarios, and preventing biases that could produce false results. Therefore, thorough validation, governance, and interdisciplinary cooperation are needed for integrating GenAI into crucial city planning tools.

In an effort to automate the complex and tedious process of building modeling, Jiang et al. (2024) [30] proposed the EnergyPlus-Large Language Model, or Eplus-LLM, a computing platform utilizing a Text-to-Text Transfer Transformer (T5) architecture capable of translating natural language descriptions from operators and modelers into high-fidelity, highly-precise building EnergyPlus models. This platform enhances usability by enabling AI-empowered communication and thorough handling of more flexible inputs, including unstructured text and grammatical variations. The platform’s effectiveness and stability was validated through 152 cases, where it achieved 100% accuracy in generating correct models, showcasing its potential to significantly reduce the effort in current modeling practices.

A research paper focusing on building retrofitting presents the technical development of LuminLab [40], an AI-powered platform designed to streamline the building retrofit process. The online tool integrates a custom AI chatbot with predictive energy modeling, allowing users to rapidly generate and discuss personalized retrofit plans through natural language dialogue. The platform aims to support the recursive, end-to-end retrofit journey, from initial conception to planning, contractor assignment, and monitoring, by de-siloing stakeholder knowledge and clarifying cost trade-offs. While not yet widely deployed, the authors position LuminLab as a pragmatic, stakeholder-informed prototype that offers a valuable roadmap for future AI-driven decision support systems in building design and retrofit.

3.2. Energy Grids and Renewable Systems: Forecasting, Stability and Cybersecurity

The integration of intermittent renewable energy sources (RESs), electric vehicles (EVs), and distributed energy resources (DERs) has introduced unprecedented volatility and complexity to power grid management. Challenges such as forecasting renewable generation, managing bi-directional power flows from Vehicle-to-Grid (V2G) systems, and optimizing real-time electricity trading require advanced analytical tools. Consequently, the grid sector has emerged as a significant domain for GenAI, which provides powerful solutions for scenario generation, predictive modeling, and the optimization of grid operations under uncertainty. Recent review studies underline the rapid adoption of artificial intelligence across renewable energy systems, covering forecasting, storage integration, smart-grid control, and decentralized asset management, while highlighting emerging opportunities for generative and hybrid AI frameworks [41]. Similarly, Henao et al. [42] provide a systematic overview of AI in power systems, emphasizing how intelligent algorithms can enhance grid resilience, fault detection, and decision support, yet also noting persistent challenges related to interpretability, fairness, and operational trust. Within energy systems, GANs and VAEs have proven particularly effective for renewable generation forecasting, anomaly detection, and cyber-physical resilience, as highlighted in [43,44,45]. Conversely, emerging diffusion and transformer-based models achieve higher fidelity in uncertainty quantification but require significant computational resources and large datasets. A key limitation identified across the literature is the lack of unified benchmarks and cross-domain validation, which hinders the transfer of generative models from controlled simulations to real-world grid operation. Future studies should prioritize hybrid GenAI architectures and standardized evaluation frameworks to ensure reliability and transparency in energy decision support.

3.2.1. Grid Stability and Sustainability

The expansion of RES and DER loads in the electrical grid increases fluctuation and grid instability, as aging and outdated infrastructure struggles to follow modern trends. Further, GenAI’s use itself can dramatically increase electricity demand, especially when difficult, complex tasks are involved. A study from Prodan et al. [46] aims to shed light on how data centers can be utilized for optimized use of the surplus of solar renewable energy in Australia, developing a proof-of-concept optimization model for co-location of renewable generation and GenAI data centers. The study claims that GenAI applications can also significantly contribute to the increase in the declining productivity of the country’s workforce.

A study from Bibri and Huang [47] positions the integration of GenAI and Artificial Intelligence of Things (AIoT) as a foundational paradigm for sustainable smart cities. Generative Artificial Intelligence of Things (GAIoT) is a unified framework which demonstrates that generative models, such as GANs, VAEs, transformers and diffusion models, enhance AIoT capabilities across the key domains of cybersecurity, anomaly detection, resource optimization and scenario generation. The authors’ study aims to showcase how operational advancements in each domain lead to an improved system, where environmental performance and resilience are key, enabling smarter and more sustainable urban ecosystems.

To effectively utilize valuable multi-modal data along with the spatial and temporal factors that affect renewable energy generation, Kang et al. [43] proposed an innovative cross-modal GAN (cGAN), a model that employs a spatio-temporal transformer architecture to integrate various data types, ranging from geographical graphs (GPS) to time-series power data. It outputs interpretable renewable generation scenarios through probability approximation, enabling the generation of an unlimited number of high-quality scenarios, while assessing the contribution of each data type. The model’s effectiveness is validated through experiments, where it achieves state-of-the-art performance on both NREL and real-world wind datasets.

A study from Mousavi et al. [48] investigates the context in which GenAI’s applications extend to solar energy resources. It highlights GenAI’s capacity in improving a multitude of domains, such as grid stability, the sizing or smart grid integration of components, and energy forecasting. It aims to validate GenAI’s efficiency in solving a variety of tasks and to further showcase how it enables improved decision-making when dealing with the complexities of the modern energy grid.

To better account for the inherent uncertainties in climate-impacted energy systems, Perera et al. [49] introduced a computational framework that aims to bridge the high-resolution climate modeling with building simulation and project future heating and cooling demands. The system utilizes a GAN, which resamples the data to generate a wide sample space of potential demand scenarios, thereby quantifying the link between climatic and operational uncertainty. Their analysis reveals that these compounded risks have a substantial economic impact, changing the present value of the energy system and threatening the grid with energy supply shortage, if ignored.

3.2.2. Data Privacy and Cybersecurity

To maintain reliability and improved operational capacity, the security of the grid’s digital infrastructure is equally critical to on-site operations. To ensure digital communication on modern smart grid components is secure, researchers have proposed a GenAI framework for anomaly detection in IEC61850-based digital substations [50]. They address the critical vulnerability of coordination protocols like GOOSE, as traditional systems struggle to protect against novel, zero-day attacks. Important contributions include the Advanced Adversarial Traffic Mutation (AATM) technique which draw on GenAI to generate balanced datasets of both normal and attack-scenario GOOSE messages. Additionally, the authors propose a GenAI-based Anomaly Detection System (ADS), which utilizes task-oriented dialogue for deep contextual understanding. This system utilizes the power of GenAI, surpassing traditional machine learning-based ADSs by identifying novel threats without a-priori knowledge of specific signatures, enhancing the grid’s resilience against cyberattacks that could disrupt the integration of renewables and other smart grid assets.

Another work by Farajzadeh-Zanjani et al. [51] also addresses fault diagnosis and cyber-attacks, utilizing a novel Generative Adversarial Dimensionality Reduction framework. Specifically, the authors developed two techniques, a supervised (GASDR) and an unsupervised (GAUDR) technique, both leveraging a GAN architecture to reduce existing data to low-dimensional spaces where data processing is more efficient, and data are highly separable. Affinity correlation and separability, two specific constraints integrated to this framework, both preserve feature relationships and maximize the distinction between existing classes directly into the GANs’ objective function. Therefore, the curse of dimensionality problem is efficiently solved while the framework provides accurate smart grid anomaly detection with more powerful feature spaces.

Data privacy and communication overhead in smart grids are additionally two significant problems when GenAI is incorporated. Mohammadabadi et al. [52] exploit GenAI for synthetic data generation and specifically utilize GenAI models that have been pre-trained, and fine-tune them to produce generated data that reflect real-world statistical distributions accurately, without the need to share sensitive or raw information from the system. This method fine-tunes GANs, VAEs and foundational models, and provides a robust foundation for federated learning, as the generated data preserves important patterns for effective model training, simultaneously preserving consumer privacy. Further, communication burden is reduced as only compact, fine-tuned model parameters need to be transmitted to a central server, avoiding the transmission of large amounts of raw sensor data. This makes the method particularly suitable for resource-constrained edge and IoT devices, offering a scalable and efficient pathway for secure, distributed grid analytics.

Addressing both grid stability and cybersecurity, Efatinasab et al. [44] introduce GAN-Stability. It utilizes GANs in an unconventional way, as the generator is not trained to produce near-realistic data. Rather, it aims to generate Out-of-Distribution (OOD) samples to identify when instability occurs. It is noteworthy that unstable data are not required during training, and the training dataset in general is limited. When compared to state-of-the-art models, the framework demonstrated similar or better performance. Regarding the issue of cybersecurity, the model integrates adversarial training to detect cyber-attacks as a form of instability, without needing a separate anomaly detection system.

3.2.3. Forecasting and Scenario Generation

Accurate forecasting and robust scenario generation are fundamental to managing the uncertainty of renewable energy, directly impacting grid planning, market operations, and stability. Deep learning has been widely incorporated in modern forecasting models, as their performance often depends on site-specific data, a requirement that is impractical for scaling up of operations. A study by Khan et al. [53] addresses this critical issue by introducing a hybrid transfer learning framework which combines VAEs with deep neural networks. Utilizing a dual-feature space, generalized knowledge is transferred from a pre-trained model, while fine-tuning on a small set of features from a target wind farm. This approach effectively bridges the gap between the need for high accuracy and the reality of data scarcity across diverse geographical locations, significantly reducing forecasting errors and minimizing the computational cost of retraining.

Addressing the issue of the increasing high-dimensional data generation due to smart metering and sensing devices, Kaur et al. [15] proposed a VAE bidirectional LSTM (VAE-BiLSTM) framework for dimensionality reduction and renewable energy forecasting. Their approach creates efficient, low-dimensional representations of the original, complex time-series energy data, and demonstrates improved forecasting accuracy when compared to benchmark models like standard VAEs, VAE-RNN and VAE-LSTM.

Another study utilizing a VAE-BiLSTM framework for electricity demand forecasting was conducted by Moradzadeh et al. [24]. They address the challenge of short-term electricity demand forecasting and aim to solve the problem of inefficient training when noisy and time-series data exist. The authors utilize the batch training technique to avoid overfitting over large datasets and to ensure accurate representations of the current data are used. The proposed framework is validated against conventional LSTM and Support Vector Regression (SVR) benchmarks, exploiting meteorological, historical and temporal data from the Tabriz city in Iran. The results indicate a significant improvement compared to these methods, showcasing better accuracy and higher correlation coefficient values.

A similar framework has additionally been applied for forecasting large-scale renewable electricity demand by Kim et al. [54] to a South Korea study. Through the VAE, the model generates synthetic data to address data scarcity, and the BiLSTM network is responsible for temporal forecasting. The authors have further incorporated Recurrent Neural Networks (RNNs) to allow for previous output data to be used as inputs and therefore enhance forecasting accuracy. The methodology presented outperforms multiple benchmarks, including Gated Recurrent Units (GRUs), LSTM, ANN and SVR, in multiple metrics.

To enhance short-term wind power forecasting, Harrou et al. [55] proposed a deep learning methodology that incorporates a Self-Attentive VAE (SA-VAE). In its core, both the encoder and decoder of the VAE integrate self-attention mechanisms, allowing the model to both capture the nonlinear patterns in time-series data and dynamically identify and prioritize the most important features for prediction. This methodology is validated against eight deep learning models such as LSTMs, GRUs and other traditional VAEs in comprehensive benchmarks, demonstrating better performance and outperforming all compared state-of-the-art methods.

On the subject of scenario generation for wind and solar power, Dias [56] proposed a system that combines a VAE with a Radial Basis Function (RBF) kernel. Their approach aims at creating a well-defined latent space from the VAE, which in turn will be used as a measure of similarity to select the most similar historical profiles from the generated scenarios. The system was tested and validated on the Brazilian power system, where it demonstrated better performance in capturing the temporal, spatial and correlation characteristics of real-world data, compared to a conventional VAE.

Capel et al. [25] introduced a diffusion-model approach to scenario-based forecasting of load, PV, and wind timeseries, using the GEFCom 2014 dataset. They compare the denoising diffusion probabilistic model (DDPM) with GANs, VAEs and normalizing flows, demonstrating superior performance in several quality and value metrics. Their work is a notable contribution for adapting a high-fidelity generative model class into an energy-forecasting context, addressing data-scarcity and scenario-generation needs. The authors highlight practical limitations such as longer training and sampling times and the need for improved parameterization of the reverse process.

L’Heureux et al. [26] adapt the transformer architecture to the task of electrical load forecasting across 20 different data streams, forecasting horizons and sliding-window lengths. The authors introduce modules for contextual features and N-space transformation to account for the numerical and structural differences between language data and energy time-series. The results indicate the adapted transformer model outperforms state-of-the-art sequence-to-sequence RNN methods, and further enables parrallelised training and inference for larger consumer sets. The study advances the generative/control-oriented paradigm by showing that transformer-based sequence models can bring long-horizon and context-rich forecasting capabilities to energy data.

3.2.4. Fault and Anomaly Detection

Proactive fault and anomaly detection in grid components is essential for optimizing maintenance procedures, preventing system downtime, and ensuring its overall stability and resilience. High-dimensional data in smart grids are also a challenge, especially during the data preprocessing and feature extraction stage. To enable more accurate and efficient fault and anomaly detection, systems provide a cleaner, lower-dimensional representation of the data. To tackle this challenge on small and noisy wind turbine datasets, Zhang and Yang [57] proposed a Long Short-Term Memory (LSTM)-based VAE Wasserstein GAN, which captures spatiotemporal features and uses the Wasserstein distance to improve the learning of complex, high-dimensional data distributions. The method enhances robustness, employing a reconstruction error indicator with a threshold set via kernel density estimation. Multiple comparative experiments on a wind farm in Northeast China were conducted, indicating the lowest Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) for the proposed approach.

A work from Li et al. [45] proposed a Sequential Conditional Variational Autoencoder (SCVAE) to address a key limitation in photovoltaic anomaly detection, which is the modeling of the dynamic, sequential impact of environmental factors on power generation. The authors developed a novel data processing pipeline to construct training datasets from 30 real-world PV sites across China, containing normal samples from unsupervised SCVAE model training. The results indicate that the SCVAE framework outperforms state-of-the-art unsupervised anomaly detection models, demonstrates strong scaling capabilities, and simultaneously enables a more targeted diagnosis approach, after identifying and analyzing its latent variables.

Another research work from Dai [58] utilized VAEs to detect load anomalies in smart grid energy data. The proposed model successfully learns normal load patterns and identifies variations and anomalies through analyzing reconstruction errors. The model additionally utilizes SHapley Additive exPlanations (SHAP) analysis to enhance interpretability and to identify the most influential features driving anomaly detection, demonstrating the method’s practical utility for real-world grid monitoring applications.

Addressing unsupervised anomaly detection in power grids, Guha et al. [59] aim to tackle cases where training data include anomalies. They propose a robust LSTM-VAE, designed to tolerate a moderate amount of data with anomalies during training. They demonstrate enhanced parameterization and an improved loss function, providing a more practical and robust solution for real-world grid monitoring, where perfectly clean data is often unavailable.

A study by Aldegheishem et al. [60] focuses on solving the Electricity Theft Detection (ETD) problem in smart grids, which is important both from a financial and an operational standpoint, and proposed two models to tackle it. The first model is a hybrid model named SMOTEENN-AlexNet-LGB model (SALM), which synthesises minority class samples and uses the SMOTEENN sampling technique to effectively balance the dataset. It further uses AlexNet for feature extraction and Light Gradient Boosting (LGB) for classification. The second proposed model employs a Conditional Wasserstein GAN (CWGAN-GP) to produce realistic consumption data from customers with suspicious behaviour. Both models outperform current techniques when validated on a real-world dataset from the State Grid Corporation of China (SGCS), with GAN-NETBoost obtaining high precision, recall, and F1-score metrics.

3.3. User Preferences and Decision-Making in Sustainable Environments

Generative AI (GenAI) plays a pivotal role in modeling user preferences and facilitating decision-making processes within sustainable smart environments. By leveraging advanced techniques such as LLMs, Retrieval-Augmented Generation (RAG), and generative adversarial networks (GANs), GenAI enables personalized, context-aware systems that align energy management with individual user needs, behaviors, and sustainability goals. This section reviews recent literature on GenAI applications in user preference modeling, behavior simulation, and decision support, highlighting how these technologies enhance energy efficiency in smart homes, buildings, and cities while addressing challenges like privacy, hallucination risks, and computational demands [61].

A key application of GenAI lies in delivering tailored energy recommendations that optimize consumption based on user preferences and real-time data. For instance, the Generative Usage-based Insights for Device Efficiency (GUIDE) framework, a hybrid AI system, integrates device recognition, energy profiling, and user-centric recommendations to promote energy-efficient behaviors in smart homes [27]. GUIDE employs GenAI to analyze appliance usage patterns from smart meter data, generating personalized suggestions such as adjusting thermostat settings or scheduling high-energy devices during off-peak hours. The system’s prescriptive analytics component uses LLMs to provide actionable insights, demonstrating a 15–20% reduction in household energy use in simulated scenarios. However, challenges include the need for accurate device disaggregation and user adoption, which GUIDE addresses through user-friendly interfaces and gamification elements.

Similarly, the Artificial Intelligence-driven IoT Recommender Energy System (ATI-RENDEE) focuses on enhancing energy management in smart houses by combining IoT sensors with GenAI for personalized recommendations [62]. ATIRENDEE processes data from appliances, environmental sensors, and user inputs to model preferences, such as preferred temperature ranges or lighting levels, and suggests optimizations like automated shading during peak solar hours. Evaluations show improved energy savings of up to 25% compared to rule-based systems, with GenAI enabling adaptive learning from user feedback. Limitations noted include dependency on high-quality sensor data and potential biases in preference modeling if training datasets lack diversity.

GenAI also facilitates simulation of personalized datasets for smart home activities, as explored in a study using modified versions of the Simulacra framework [63]. This approach employs GenAI to generate synthetic datasets of activities of daily living (ADL) and appliance usage, tailored to individual user profiles. By incorporating user preferences (e.g., energy-conscious behaviors like preferring low-power modes), the model creates realistic scenarios for testing energy management algorithms. Findings indicate that GenAI-produced datasets closely mimic real-world patterns, with accuracy rates exceeding 90% in predicting energy consumption, aiding in the development of user-centric systems. Challenges involve ensuring ethical data generation to avoid reinforcing stereotypes in user behavior models.

In another example, VayuBuddy, an LLM-powered chatbot, provides personalized air quality and energy insights for smart home users [35]. By integrating IoT sensor data with user health preferences (e.g., sensitivity to pollutants), it recommends actions like activating air purifiers or adjusting HVAC settings to balance comfort and energy use. This user-centric approach reduces energy waste by 10-15% in pilot studies, though hallucination risks in LLM outputs necessitate RAG integration for factual accuracy.

Expanding further, integrations of GenAI with IoT devices emphasize enhanced user experiences through preference-driven adaptations [64]. For instance, systems that combine GANs with IoT frameworks generate adaptive recommendations, such as optimizing lighting based on user mood or activity patterns inferred from historical data. Privacy-preserving techniques are crucial here, with reviews highlighting differential privacy and federated learning to protect user preference data during model training [65].

GenAI excels in modeling user behavior to inform energy-efficient decisions, often through multi-modal data processing. The Follow-Me AI concept enhances user interactions in smart environments by using GenAI to predict and adapt to behaviors, optimizing energy and resource use [66]. In a smart campus case study, LLMs process sensor data and user preferences (e.g., quiet spaces with optimal lighting) to adjust room settings and network configurations proactively. This results in energy savings of 20–30% by reducing unnecessary heating or lighting in unoccupied areas. Key methods include model compression for edge deployment and semantic slicing for efficient data handling, but challenges like fluctuating Quality of Experience (QoE) due to mobile users require predictive orchestration.

A systematic review of user modeling for behavioral modeling in smart environments emphasizes AI’s role in creating preference-based profiles [61]. While not exclusively GenAI-focused, it discusses how generative techniques can simulate behaviors for energy optimization, such as predicting occupancy to control HVAC systems. Integration with user consent mechanisms ensures privacy, aligning behaviors with preferences like eco-friendly modes. Findings highlight improved system responsiveness, but data scarcity in diverse user groups poses a limitation, addressable via synthetic data generation with GANs.

In smart cities, GenAI models user behaviors for sustainable planning, as seen in Urban Generative Intelligence (UGI) frameworks [35]. UGI uses LLMs with urban knowledge graphs to simulate mobility patterns based on preferences (e.g., preferring low-emission routes), supporting decisions like converting parking to bike lanes. Evaluations show enhanced traffic flow and reduced emissions by 12–18%, with challenges in synthetic data fidelity mitigated by validation protocols.

Further expansions include anomaly detection in grids, where GenAI generates synthetic threats to train models, tying into user safety preferences for reliable energy supply [50]. This unified framework achieves 97.9% precision, supporting decision-making in critical infrastructure by identifying zero-day attacks.

GenAI-powered decision support systems (DSSs) incorporate user preferences to drive sustainable outcomes in energy systems. The LLM-SUC framework employs LLM agents to balance high-renewable energy systems under uncertainty [67]. By generating scenario trees based on wind forecast errors and user-defined policies (e.g., prioritizing renewables), it reduces operational costs by 1.1–2.7% and load curtailment by 26.3%. User-centric elements include configurable interfaces for preference alignment, though direct preference integration is limited, focusing more on system-level decisions.

In building energy applications, LLMs serve as DSS by processing multimodal data for fault detection and retrofit recommendations [68]. For instance, GPT models analyze operational data to suggest energy-efficient measures tailored to user preferences, such as code-compliant designs, achieving modeling effort reductions of over 95%. Interpretability is enhanced via Shapley values, fostering user trust in decisions.

For broader sustainable transitions, RAG-LLM systems like the Energy Chatbot provide multi-source decision support for small and medium enterprises (SMEs) [61]. By grounding responses in verified datasets, it offers personalized insights on sustainable energy initiatives (SEIs), such as subsidy eligibility based on user profiles. Pilot results indicate improved adoption rates of green practices, with energy savings of 15–25%, though challenges include computational costs and bias mitigation.

A review of LLMs in energy systems research underscores their potential for user-centric DSS [61]. LLMs optimize smart grids by modeling preferences in demand–response programs, enabling personalized pricing recommendations that reduce peak loads by 10–20%. Future directions include hybrid models combining LLMs with GANs for better uncertainty handling.

Additionally, platforms like LuminLab use AI chatbots for recursive retrofit journeys, generating personalized plans through natural language and predictive modeling [40]. This stakeholder-informed approach clarifies cost trade-offs, supporting decisions from conception to monitoring.

EPlus-LLM automates building energy models from natural language, handling user preferences for high-fidelity simulations [30]. Text2BIM generates BIM models via multi-agent LLMs, incorporating preferences for energy analysis [38].

Across these applications, common challenges include hallucination risks in LLMs, addressed by RAG and HITL approaches; data privacy, mitigated through consent-based frameworks; and computational efficiency, improved via edge computing. Future work should focus on interdisciplinary benchmarks for GenAI in user-centric systems, emphasizing ethical AI to ensure equitable preference modeling [69].

Beyond these system-level implementations, an emerging research frontier involves integrating GenAI into human-in-the-loop (HITL) decision architectures. In such designs, the generative model not only predicts or recommends actions but also continuously learns from user feedback, forming an adaptive preference model that evolves with behavioral change. Interactive dialogue systems, for example, allow occupants to refine or override automated controls using natural language, with GenAI subsequently updating its belief about the user’s comfort tolerance, risk attitude, and willingness to trade cost for sustainability. This dynamic modeling of subjective preferences transforms static optimization into a co-adaptive process where human cognition and machine inference jointly drive decisions.

Recent studies extend this concept by embedding cognitive behavioral modeling within energy agents, enabling GenAI to represent latent decision drivers such as habits, emotions, and social influence. Probabilistic diffusion models and variational autoencoders have been employed to capture stochastic preference evolution, improving the prediction of discretionary loads like appliance usage or thermostat adjustments. Such modeling frameworks are crucial for demand-side management schemes, where user intent, comfort inertia, and contextual cues strongly affect real-world adoption.

Another growing area concerns preference uncertainty and multi-objective optimization. When user intentions are ambiguous, GenAI systems can sample a distribution of plausible behaviors, generating ensemble recommendations that balance comfort, cost, and carbon footprint. This probabilistic reasoning has been coupled with reinforcement learning (RL) controllers in hybrid setups, where the generative layer supplies stochastic user scenarios to train robust policies under diverse behavioral conditions. Similarly, Bayesian preference elicitation approaches, supported by LLMs acting as conversational surveyors, allow users to express soft constraints (e.g., “I prefer cooler nights unless prices surge”) that are transformed into quantitative reward functions for control agents [70].

The governance and ethics dimension is equally vital. Preference models inherently handle sensitive behavioral and demographic information; thus, frameworks for privacy-preserving federated learning and differentially private fine-tuning are being explored to ensure compliance with data-protection principles [65,71]. Transparency mechanisms, such as explainable recommendation rationales and Shapley-based interpretability layers, enhance user trust and accountability. Furthermore, fairness auditing of generative outputs ensures that automated recommendations do not disproportionately benefit specific socioeconomic groups or behavioral archetypes, a crucial consideration for equitable energy transitions.

Evaluation methodologies for GenAI-driven user modeling are still maturing. Current best practices suggest combining quantitative metrics, like Mean Absolute Preference Error (MAPE), comfort satisfaction indices, and energy savings percentages, with qualitative measures such as perceived transparency, trust, and intervention acceptance rate. Multi-agent simulation environments (e.g., BuildingsBench [72]) provide reproducible testbeds for benchmarking preference-learning algorithms under varying climatic, cultural, and occupancy scenarios.

Looking ahead, research should converge on: (i) standardized datasets and ontologies for preference representation, (ii) multimodal feedback fusion integrating speech, gesture, and contextual data, (iii) hybrid symbolic–neural models to encode ethical or regulatory constraints, and (iv) lightweight GenAI agents deployable at the edge for real-time personalization. These directions underscore that aligning GenAI with human decision-making is not only a technical challenge but also a socio-technical opportunity enabling sustainable, explainable, and user-empowering energy systems of the future.

4. Result Analysis: Strengths and Limitations of GenAI Compared to Traditional Methods

4.1. Strengths of GenAI (Compared to Traditional Methods)

Across the reviewed literature, GenAI demonstrates superior performance in tasks requiring contextual reasoning, multimodal integration, and adaptive decision-making in smart buildings and energy systems. Several studies show that LLM-powered systems offer more accurate, interpretable, and user-aligned recommendations compared to traditional ML. For instance, ref. [17] demonstrates that an LLM-based semantic comparison framework provides more context-aware, consistent, and personalized recommendations than classical rule-based or regression-driven approaches. Likewise, ref. [33] evaluatess GPT-style reasoning for public-building energy use and show that LLMs outperform deterministic baselines in handling complex, multi-objective building states. Enhanced interaction and interpretability are also frequently reported advantages. In the context of smart-home interfaces, ref. [34] shows that an LLM-driven copilot enhances human–AI collaboration by providing explanations, natural-language feedback, and adaptive decision guidance, capabilities that traditional ML does not offer. Ref. [29] similarly indicates that GenAI enables more intuitive AEC workflows through interactive text-based modeling. GenAI also excels in scenario generation, uncertainty modeling, and data augmentation, which improves downstream control and forecasting. For example, Digital Twin approaches [12,13] benefit from GenAI-enabled synthesis of occupant behavior and predictive maintenance insights, leading to more robust operation compared with classical sensor-only or rule-driven systems. In renewable generation, ref. [14] utilizes VAE-GAN hybrids to generate realistic household energy scenarios, improving RL-based smart-home control. At the urban scale, ref. [32] highlights that GenAI-assisted energy-system design allows multi-criteria reasoning (cost, emissions, demand) that cannot be matched by classical algorithms. Finally, several reviews emphasize GenAI’s cross-domain generalization: [16,37,47] all identify that GenAI integrates textual, operational, and environmental data in ways that traditional statistical or RL-based methods cannot. Collectively, the reviewed publications indicate that GenAI provides higher interpretability, richer contextual understanding, increased robustness, and improved user engagement, frequently surpassing traditional ML and rule-based strategies in complex smart-building and energy-management tasks.

4.2. Limitations and Risks of GenAI

Despite these advantages, the reviewed literature also identifies several limitations, risks, and unresolved challenges. A recurring issue is the potential for algorithmic bias, especially when models are trained on imbalanced or non-representative datasets. For example, BuildingSage [34] notes that LLM-based building copilots may produce recommendations that favor certain building categories or occupant profiles due to skewed training data, creating risks for energy fairness. Ref. [16] likewise emphasizes that biased GenAI systems can lead to unequal access to energy-saving recommendations in smart-city environments. The digital divide is another major concern. Studies similar to [47] stress that GenAI requires extensive sensor infrastructures, edge/cloud connectivity, and computational capacity, conditions not always available in low-resource buildings or municipalities. This creates a disparity where advanced GenAI benefits are accessible only to well-equipped environments, while traditional lightweight methods remain more deployable in smaller or underserved settings. GenAI’s performance also degrades significantly in data-scarce scenarios, despite large-scale pretraining. Ref. [36] reports that LLM-based energy-management tools exhibit hallucinations, unstable outputs, and incorrect reasoning when fine-tuned on small building datasets. Ref. [36] confirms that insufficient domain data lead to misaligned recommendations and weak generalization. In contrast, classical approaches, such as regression models [5,6], VAE/BiLSTM hybrids [24], and forecasting mechanisms based on structured features [3,4], remain robust with relatively small, clean datasets. Several authors also highlight the computational and sustainability implications of GenAI. Ref. [46] showcases GenAI-driven data centers and LLM computation impose a significantly higher energy burden compared with traditional ML pipelines. Refs. [37,42] further argue that GenAI models require substantially more computations during training and inference, raising concerns about the contradiction between sustainability goals and the heavy computational footprint of LLMs. This stands in contrast with lightweight rule-based or statistical models, such as those used in early smart-building systems [60] or traditional control structures [9,10], which remain far more energy-efficient. Finally, privacy and security risks are widely discussed in GenAI-based systems. Ref. [65] reviews privacy-preserving GenAI techniques and identifies leakage risks, model inversion threats, and the difficulty of securing multi-modal building data. Ref. [34] warns that LLM-based building copilots broaden the attack surface, requiring stronger safeguards than conventional ML. Overall, while GenAI significantly expands the capabilities of energy and building management systems, the literature consistently points to bias, inequality, sustainability burdens, instability under data scarcity, and security/privacy concerns, all of which must be resolved before large-scale adoption.

5. Research Tendencies and Future Directions

The research landscape for the application of generative artificial intelligence in smart buildings, energy grids and user-centered decision making has changed visibly over the last decade (Figure 1). The first studies, which were mainly published before 2022, were characterized by data-driven approaches to forecasting, anomaly detection and descriptive analysis. Generation methods were either lacking or limited to supporting functions, such as data synthesis or support for traditional optimization frameworks. These projects typically dealt with single domain issues, focusing on either energy performance at the building level or energy efficiency in isolation, relying mainly on rules-based control and classical machine learning techniques. On the contrary, more recent research since 2023 shows a significant methodological and thematic change. GenAI and big language models are now at the core of integrated frameworks that combine predictive, decision-support and proactive control. This shift reflects the wider shift from isolated, focused technical applications to a more systemic and holistic approach that covers multiple areas. Researchers are increasingly exploring interoperability between buildings, power grids, renewable resources and user behavior, using generative models for context-based thinking, scenario generation and real-time adaptation.

Methodologically, the shift from purely supervised learning pipelines to hybrid approaches is taking place, combining generative models with reinforcement learning, optimization algorithms, and digital twins. These hybrid frameworks allow for adaptive and pro-active control strategies that go beyond static forecasting. Thematic work is moving towards the use of multimodal data sources, combining sensor data, textual descriptions, weather forecasts and operational data to support more informed and dynamic decision-making. A growing body of recent literature also points to a gradual shift from studies based on simulation to pilot projects in real-world settings, such as smart grids and energy grids. This emerging trend shows that GenAI applications are reaching maturity levels that allow them to be tested in operational environments, although still on a limited scale. Despite these promising developments, there are still some research gaps. Many of the GenAI frameworks have not been tested under realistic conditions and problems remain in terms of data quality, standardization, interoperability, computational power and clarity. Addressing these issues will be crucial to move from conceptual innovation to reliable, scalable and reliable deployment. Overall, the current research trend is towards the development of scalable, real-time, context-aware and human-centric intelligent systems that can support decision making across the various layers of urban and energy infrastructures, presented in

Table 1. Summary of generative AI applications in smart infrastructure and energy systems.

Year	Sector	Application	Method	Key Findings	Study
2024	Smart Building and Smart Cities	Building Management and Environmental Control	LLMs and RAG	LLM-based comfort monitoring	[28]
2025	Smart Building and Smart Cities	Device Orchestration for Agentic Smart Home Systems	LLMs and Semantic Kernel	Agentic smart home control	[31]
2025	Smart Building and Smart Cities	Building Energy Management	STLF and GRANs	Generative STLF platform	[32]
2024	Smart Building and Smart Cities	GenAI Privacy-Preserving Copilot	GenAI Copilots and LLMs	Privacy-preserving building copilot	[34]
2025	Smart Building and Smart Cities, User Preferences and Decision Making	GenAI Copilots and Conversational Interfaces	GenAI Copilots and LLMs	User-centric GenAI survey	[35]
2024	Smart Building and Smart Cities	Project and Building Management Processes	Text-based models	LLMs aid construction management	[37]
2025	Smart Building and Smart Cities	Architecture, Engineering, Construction and Operations	Review	GenAI in AEC review	[29]
2025	Smart Building and Smart Cities	Early-Stage Creation of BIM models	LLMs	LLM generates BIM models	[38]
2024	Smart Building and Smart Cities	City-Level Digital Twins	GenAI	GenAI for urban resilience	[39]
2024	Smart Building and Smart Cities	Building Modeling	LLMs and T5	LLM automates energy modeling	[30]
2024	Smart Building and Smart Cities	Building Retrofitting	LLMs	AI-driven retrofit platform	[40]
2025	Energy Grids and Renewable Systems	Data Centres for Optimized Use of Renewable Energy	Review	GenAI for renewable integration	[46]
2025	Energy Grids and Renewable Systems	Sustainable Smart Cities	GAIoT and GANs, VAEs, Transformers, Diffusion Models	GAIoT enhances sustainability	[47]
2024	Energy Grids and Renewable Systems	Multi-modal Data and Renewable Energy Generation	Cross-modal GAN	Cross-modal GAN for scenarios	[43]
2025	Energy Grids and Renewable Systems	Solar Energy Resources	Review	GenAI enhances solar systems	[48]
2022	Energy Grids and Renewable Systems	Climate-Impacted Energy Systems	GANs	GAN simulates climate uncertainty	[49]
2025	Energy Grids and Renewable Systems	Anomaly Detection and Cybersecurity in IEC61850-based digital Substations	Conditional-GAN and Adversarial Traffic Mutation	Synth data for grid security	[50]
2021	Energy Grids and Renewable Systems	Fault diagnosis and Cybersecurity	GANs	GAN detects faults and attacks	[51]
2024	Energy Grids and Renewable Systems	Data Privacy in Smart Grids	GANs, VAEs and Foundational Models	GenAI improves distributed learning	[52]
2025	Energy Grids and Renewable Systems	Grid Stability and Cybersecurity	GANs	GAN aids grid stability	[44]
2022	Energy Grids and Renewable Systems	Forecasting and Scenario Generation for RES Uncertainty	VAEs and Deep Neural Networks	VAE+TL improves wind forecast	[53]
2021	Energy Grids and Renewable Systems	Renewable Energy Forecasting and Dimensionality Reduction	VAE and BiLSTM	VAE for feature reduction	[15]
2022	Energy Grids and Renewable Systems	Electricity Demand Forecasting	VAE-BiLSTM	VAE-BiLSTM improves load forecast	[24]
2024	Energy Grids and Renewable Systems	Renewable Electricity Demand Forecasting	VAE-BiLSTM	VAE-BiLSTM predicts renewable demand	[54]
2024	Energy Grids and Renewable Systems	Short-Term Wind Power Forecasting	Variational auto-encoder with BiLSTM	Self-attentive VAE improves wind forecast	[55]
2023	Energy Grids and Renewable Systems	Scenario Generation for Wind and Solar Power	Self-attentive VAEs	VAE generates renewable scenarios	[56]
2023	Energy Grids and Renewable Systems	Scenario-based Load Forecasting	Diffusion model	Diffusion for probabilistic forecasting	[25]
2022	Energy Grids and Renewable Systems	Electrical Load Forecasting	Transformer-based architecture	Transformer for load forecasting	[26]
2023	Energy Grids and Renewable Systems	Fault and Anomaly Detection in Small and Noisy Wind Turbine Datasets	VAE with radial basis function kernels	LSTM-VAE-GAN detects anomalies	[57]
2024	Energy Grids and Renewable Systems	Photovoltaic Anomaly Detection	Conditional VAE	CVAE detects PV anomalies	[45]
2025	Energy Grids and Renewable Systems	Load Anomaly Detection	VAE for load profiles	VAE detects load anomalies	[58]
2023	Energy Grids and Renewable Systems	Fault and Anomaly Detection in Smart Grids	LSTM–VAE	LSTM-VAE detects grid anomalies	[59]
2021	Energy Grids and Renewable Systems	Electricity Theft Detection	Enhanced NN	GAN aids theft detection	[60]
2025	User Preferences and Decision-Making	User preference modeling and behaviour simulation	AI + LLM integration	LLMs inform climate strategy	[61]
2025	User Preferences and Decision-Making	Consumption Optimization and Energy Recommendations	Hybrid GenAI–optimization	LLM guides efficient appliance use	[27]
2026	User Preferences and Decision-Making	Energy-efficient management in smart houses	AI-driven IoT recommender system	IoT recommender saves energy	[62]
2025	User Preferences and Decision-Making	Personalized smart-home activity dataset generation	Generative AI simulation	GenAI simulates home activities	[63]
2023	User Preferences and Decision-Making	Enhanced user interaction with IoT devices	Generative AI integration	GenAI enhances IoT experience	[64]
2025	User Preferences and Decision-Making	Privacy-preserving control and personalization	Generative AI with privacy-preserving techniques	Privacy-preserving GenAI review	[65]
2025	User Preferences and Decision-Making	Energy-efficient user interaction and adaptation	Follow-me AI (context-aware generative system)	Context-aware AI saves energy	[66]
2025	User Preferences and Decision-Making	Balancing and optimization of energy systems	LLM agents	LLM agents manage energy	[67]
2025	User Preferences and Decision-Making	Building energy analysis and optimization	LLMs	LLMs aid building energy management	[68]
2025	User Preferences and Decision-Making	Environmental and regulatory implications of GenAI	Theoretical review (Oxford Handbook)	GenAI sustainability policy review	[69]
2024	User Preferences and Decision-Making	Human preference elicitation for energy-aware systems	Bayesian inference with language models	LLM elicits human preferences	[70]
2025	User Preferences and Decision-Making	Privacy-preserving control and personalization	Generative AI with privacy-preserving techniques	NIST GenAI risk framework	[71]
2025	User Preferences and Decision-Making	Renewable energy forecasting and system optimization	Generative AI optimization framework	GenAI aids renewable optimization	[72]

, and the critical issues that accompany this development are expected to be addressed in the future.

6. Conclusions and Future Work

This review analyzed the rapid emergence of Generative Artificial Intelligence (GenAI) across smart buildings, energy grids, and user-centric systems. The findings indicate that GenAI is shifting the paradigm from predictive analytics toward adaptive and creative intelligence, capable of generating operational strategies, synthetic data, and human-aligned control actions. Large Language Models (LLMs) enable semantic reasoning and natural language interaction, while GANs and VAEs strengthen forecasting, anomaly detection, and scenario generation for energy and building applications.

Across all domains, GenAI demonstrates tangible benefits in resilience, personalization, and efficiency, yet several barriers persist, including interoperability, computational demand, explainability, and limited real-world validation. The majority of current work remains simulation-based, highlighting the need for scalable frameworks and standardized evaluation.

Future research should focus on (i) hybrid GenAI–reinforcement learning architectures for adaptive control, (ii) standardized datasets and open benchmarks for comparative validation, (iii) trustworthy and explainable GenAI methods to improve interpretability and fairness, (iv) energy-efficient edge–cloud integration for real-time deployment, and (v) human-in-the-loop approaches aligning AI decisions with behavioral and policy constraints. Addressing these priorities will enable GenAI to evolve from isolated prototypes into a foundational technology for sustainable, intelligent, and human-centric energy environments.