A Review of Artificial Intelligence and Deep Learning Approaches for Resource Management in Smart Buildings

Abstract

This comprehensive review maps the fast-evolving landscape in which artificial intelligence (AI) and deep-learning (DL) techniques converge with the Internet of Things (IoT) to manage energy, comfort, and sustainability across smart environments. A PRISMA-guided search of four databases retrieved 1358 records; after applying inclusion criteria, 143 peer-reviewed studies published between January 2019 and April 2025 were analyzed. This review shows that AI-driven controllers—especially deep-reinforcement-learning agents—deliver median energy savings of 18–35% for HVAC and other major loads, consistently outperforming rule-based and model-predictive baselines. The evidence further reveals a rapid diversification of methods: graph-neural-network models now capture spatial interdependencies in dense sensor grids, federated-learning pilots address data-privacy constraints, and early integrations of large language models hint at natural-language analytics and control interfaces for heterogeneous IoT devices. Yet large-scale deployment remains hindered by fragmented and proprietary datasets, unresolved privacy and cybersecurity risks associated with continuous IoT telemetry, the growing carbon and compute footprints of ever-larger models, and poor interoperability among legacy equipment and modern edge nodes. The authors of researches therefore converges on several priorities: open, high-fidelity benchmarks that marry multivariate IoT sensor data with standardized metadata and occupant feedback; energy-aware, edge-optimized architectures that lower latency and power draw; privacy-centric learning frameworks that satisfy tightening regulations; hybrid physics-informed and explainable models that shorten commissioning time; and digital-twin platforms enriched by language-model reasoning to translate raw telemetry into actionable insights for facility managers and end users. Addressing these gaps will be pivotal to transforming isolated pilots into ubiquitous, trustworthy, and human-centered IoT ecosystems capable of delivering measurable gains in efficiency, resilience, and occupant wellbeing at scale.

1. Introduction

The modern world faces pressing challenges—climate change, rising energy demand, and the quest to enhance occupant comfort [1,2]. As buildings are among the largest consumers of energy and materials, they have become a focal point for effective mitigation strategies [3,4,5]. The smart-building paradigm replaces traditional, static control systems with adaptive environments that learn from and respond to changing conditions and user needs [6,7,8]. These intelligent buildings aim to optimize energy use, improve indoor environmental quality (IEQ), and ensure safety, comfort, and operational efficiency throughout their life cycles [9,10].

However, managing resources in smart buildings is inherently complex owing to the unpredictability of outdoor conditions, the variability of human behavior, and the intricacy of engineering subsystems such as HVAC, lighting, and power supply [11,12]. Rule-based or scenario-driven controllers rarely achieve optimal performance under such dynamic, uncertain circumstances [13].

To address these limitations, artificial intelligence (AI), machine learning (ML), and—most notably—deep learning (DL) have emerged as transformative enablers of truly intelligent buildings [14,15,16]. By processing the vast data streams generated by heterogeneous sensors (temperature, humidity, CO₂, motion, illuminance) [17,18,19], smart meters [20,21,22], IoT devices [23,24], and embedded HVAC equipment [6], AI/ML/DL models can uncover hidden patterns, anticipate future needs, and make decisions in real time [25], thereby shifting building operation from reactive to proactive management [8].

Coupling AI/ML with IoT infrastructure [3,4], big-data platforms [17,20], digital twins (DT) [26], and building information modelling (BIM) [27] lays the groundwork for fully autonomous, optimized building-management systems [28]. Emerging technologies such as 5G communication [16] and blockchain [29] further enhance performance, security, and decentralization in this domain [30].

Deep learning, in particular, has attracted intense interest [31]. Architectures such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), and long short-term memory (LSTM) networks excel at extracting information from complex temporal and spatial sensor data [4,5]. They have proved effective for energy-consumption forecasting [10,25,31], occupancy and activity detection [6,11], thermal-comfort maintenance [1,5], and equipment-fault diagnosis [13,26]. Advanced paradigms—including transfer learning [32,33] and (deep) reinforcement learning (RL/DRL) [5]—further mitigate challenges such as data scarcity in newly constructed buildings and the need for adaptive, autonomous control [7,12].

Figure 1 shows the architecture of the “Smart Building” system and its interaction with artificial intelligence. The components work together to collect and manage data and monitor and control the building’s systems. All data collected from the components is transmitted to the artificial intelligence module. The artificial intelligence analyzes the received information, creates intelligent models, and sends them back to the smart building system. Artificial intelligence optimizes the operation of the building systems, enhancing efficiency, security, and comfort.

The 2024 recast of the Energy Performance of Buildings Directive (EPBD) introduced the Smart Readiness Indicator (SRI), a harmonized metric that rates how effectively a building (i) optimizes energy performance, (ii) adapts services to occupants, and (iii) interacts with the grid. Recent reviews demonstrate that AI-enabled HVAC reinforcement learning can raise the heating-domain sub-index by up to 0.23 points, while occupant-centric comfort models account for 35% of the “Adaptation” score variance across 105 European case studies. Mapping such advances to the nine SRI technical domains therefore offers a policy-relevant lens for gauging technological maturity and investment priorities [34].

The opacity—or “black-box” nature—of AI and deep-learning controllers has emerged as an equally critical barrier: operators cannot verify decision logic, and occupants struggle to trust the outcomes. The EU AI Act 2024 classifies such building-control systems as high-risk and mandates demonstrable transparency [35], while both the EPBD Smart Readiness Indicator and ISO/IEC TR 24028:2021 echo the need for explainability at the building scale [36,37]. Consequently, developing explainable-AI (XAI) approaches is indispensable for ensuring regulatory compliance and stakeholder trust in smart-building deployments.

Despite the clear benefits of AI, previous reviews have covered only partial aspects of its application in smart buildings. Some focused on classical machine-learning pipelines for demand and comfort prediction but did not provide a comparative quantitative analysis across algorithm classes or account for the impact of AI workflows on embodied energy [38,39]. Other work addressed data processing and sensor integration but omitted algorithm-level distinctions or edge deployment scenarios [40,41]. Reviews dedicated to deep learning categorized models by architecture but did not include transformer-based or LLM-enabled strategies, nor did they examine interpretability [42,43], while reviews on reinforcement learning lacked coverage of trade-offs related to carbon footprint and the scalability of large models [44]. As a result, a holistic picture of the current state of the field remained incomplete.

Building on this, the present review aims to provide a comprehensive synthesis of the existing literature, clarify the current state of AI and deep-learning applications in smart buildings, and highlight unresolved issues. We ask, how and to what extent do AI techniques improve resource management, and what knowledge gaps remain? To answer these questions, we extend the search window through April 2025, include GPT-class automation scenarios in the analysis, and systematically couple energy-saving outcomes with their computational footprints. Although recent advances demonstrate clear benefits, several obstacles still impede large-scale deployment. Real-world implementation is hindered by limited or noisy data [3,13], persistent privacy and security risks [8,11], and the challenge of making heterogeneous systems and communication protocols work together seamlessly [25,31]. Furthermore, models trained for a specific building type or climate often fail to generalize, restricting broader applicability [27,28,33]. By mapping these findings onto the Smart Readiness Indicator (SRI), this study contributes a novel synthesis that complements—but substantively extends—the prior literature and offers a foundation for prioritizing future research.

Table 1. Roadmap for AI and deep learning-based resource management in smart buildings.

Phase/Area	Key Tasks	Main Challenges	Expected Outcomes and Impacts
Data Collection and Standardization	Creating and standardizing high-quality datasets from various sensors and IoT devices	Continuous monitoring and feedback Ongoing model and interface improvements Functional expansion	Accessible, unified datasets Data readiness for model training
AI/DL Model Development and Training	Training specialized deep learning models (CNN, RNN, LSTM, DRL, GNN)	Lack of labeled data High computational demands Data scarcity	Accurate energy consumption and occupancy forecasting Adaptive HVAC and lighting control
Integration with IoT and Digital Technologies	Processing complex temporal and spatial data	Compatibility across heterogeneous systems Cybersecurity and communication reliability	Robust, secure, and scalable platform Real-time data processing
Testing and Pilot Deployment	Implementing digital twins and BIM	Variability in building architectures and climates Integration with legacy systems	Matured solutions ready for commercial scaling
Scaling and Optimization	Utilizing 5G and blockchain to enhance security and performance	Maintaining stable operation across numerous sites Managing updates and system adaptation	Scalable, efficient, and resilient system Enhanced energy savings and user comfort

presents a detailed roadmap that structures this analysis. It sequentially describes the key stages—from data collection to full-scale deployment—defining the main tasks, challenges, and expected outcomes for each, thus serving as a guide for the reader through the subsequent discussion.

2. Methodology

2.1. Search Strategy and Selection Criteria

This review was conducted in accordance with PRISMA guidelines. We analyzed the English-language literature published from January 2010 to April 2025 that reports on artificial intelligence (AI) and deep-learning (DL) techniques for resource management in smart buildings. Searches across Scopus, Web of Science, IEEE Xplore and ScienceDirect targeted peer-reviewed studies that include quantitative evaluations of AI or DL within a smart-building context. Our strategy combined terms such as “smart buildings”, “artificial intelligence”, “deep learning”, “resource management”, “energy efficiency”, “HVAC control”, “deep reinforcement learning”, “graph neural networks” and “federated learning”, along with related synonyms, to ensure comprehensive coverage.

As shown in the PRISMA diagram in Figure 2, the initial search yielded 1358 unique records. After automatic de-duplication (74), 1284 titles and abstracts were screened. Of these, 1141 papers were excluded because they either lacked a quantitative, AI-based evaluation or addressed domains adjacent to, but not centered on, the review’s scope. The remaining 143 articles were examined in full against the inclusion criteria, such as the application of AI/ML techniques, quantitative performance reporting, and the study explicitly quantifying at least one SRI impact criterion or mapping its AI/DL solution to an SRI technical domain (Domain 1 Heating, Domain 9 EV-charging)). All 143 articles passed this stage and were included in the final synthesis. Although only 79 Scopus records (<0.07% of the entire “building energy performance” corpus) mention the SRI since 2018, the year-on-year growth rate is ≈24%, indicating rapid scholarly uptake of the indicator as a regulatory benchmark [38].

2.2. Data Extraction and Appraisal

Because the review is narrative rather than strictly systematic, formal risk-of-bias tools were not applied. Instead, each study was appraised qualitatively for real-world applicability to resource optimization, HVAC control, and occupant comfort; the robustness of its empirical evidence and evaluation metrics; its methodological contribution, such as novel architectures, hybrid schemes, or adaptive context-aware control; and its transparency in reporting, in line with PRISMA requirements. This structured appraisal strengthens the credibility of our synthesis.

A structured Microsoft Excel form was used for data extraction. The full Microsoft Excel form was added to Appendix A. Two reviewers independently extracted data, including title, authors, publication year, building context, AI technique, and key results. Any discrepancies were resolved by consensus to minimize error.

2.3. Data Synthesis and Analysis

Three main methods were used to analyze the corpus of 143 selected articles. First, to map the intellectual landscape, a bibliometric analysis using TF-IDF and k-means clustering was conducted. To map the corpus’s intellectual landscape, we first lemmatized the titles and author-supplied keywords in the 143 studies, removed stop words, and converted the remaining terms into term-frequency–inverse-document-frequency (TF-IDF) vectors. Projecting these vectors into a lower-dimensional space, we applied k-means clustering; the silhouette coefficient pointed to an optimal solution of seven clusters. These clusters represent distinct research strands: C0, ML-based demand prediction; C1, IoT platforms for smart buildings; C2, HVAC control strategies; C3, computer-vision and deep-learning monitoring; C4, BIM-driven digital twins; C5, security- and interoperability-focused architectures; and C6, transfer learning for climate adaptation. Figure 3 plots the centroids of the seven clusters in two dimensions, highlighting their compactness and relative separation and laying the groundwork for subsequent comparative analysis.

As the thematic map in the Figure 3 shows, we derived the clusters only after rigorously checking for potential bias at two levels. First, literature-selection bias was curtailed by cross-validating the initial Scopus search against Web of Science and manually excluding records outside the “AI-enabled smart-building” scope; inter-reviewer agreement on inclusion decisions reached Cohen’s κ = 0.87. Second, analytical bias stemming from algorithmic choices was quantified via 50-fold bootstrap resampling: in each fold, the entire pipeline was re-run, yielding a median Adjusted Rand Index of 0.79 and an average silhouette width of 0.45 ± 0.04, indicating that the emerging thematic structure remains stable under small corpus perturbations [45,46].

The corpus—titles and author keywords of 143 papers—was lower-cased, de-punctuated, lemmatized, and filtered through an extended English stop-list that preserved domain terms such as BIM, HVAC, and sensor. Bigrams like “smart building” and “energy prediction” were retained with a collocation detector. We then vectorized the text with TF-IDF implemented in scikit-learn v1.5.0 (Python 3.11) using sub-linear term-frequency scaling, ngram_range = (1, 2), min_df = 2, max_df = 0.8, and max_features = 10,000 [47]. Because TF-IDF produces a high-dimensional sparse matrix, we applied Latent Semantic Analysis—specifically TruncatedSVD with n_components = 2 and random_state = 42—to project documents onto two orthogonal latent dimensions that together explained 28% of the corpus variance while preserving interpretability [48].

Clustering was carried out in the full TF-IDF space (not on the reduced coordinates) using K-means (init = ‘k-means++’, n_init = 50, max_iter = 500, random_state = 42) [49,50]. The optimal number of clusters was chosen by maximizing average silhouette width over k = 2 … 12; the peak at k = 7 (silhouette = 0.46) balanced cohesion and separation and was corroborated by an elbow in total within-cluster SSE [45,46]. Bootstrap validation showed that 93% of documents stayed in the same cluster in at least 40 out of 50 runs, further attesting to the robustness of the seven-cluster solution. Cluster centroids were labelled post-hoc by inspecting the ten highest-loading TF-IDF terms per cluster, yielding the thematic descriptors displayed in Figure 3 (“ML-based demand prediction”, “IoT platforms for smart buildings”).

Second, the growth of publications by year was analyzed. As shown in Figure 4, publication activity rose steadily and climbed sharply in 2025, underscoring the field’s accelerating momentum.

Third, a comparative analysis of energy-saving outcomes was performed on the 79 studies where these data were quantitatively reported.

Table 2. Median energy savings by algorithm class across 79 quantitatively reported smart-building studies (2019–2025).

Algorithm Family	Median Energy Saving %	IQR %	k (Studies)	Typical Controlled Load
Rule-based (baseline)	7	5–10	8	HVAC, lighting
Classical ML (SVM, RF, GB)	12	7–18	18	HVAC, lighting
Supervised DL (CNN/RNN/LSTM)	17	10–24	20	HVAC, lighting, plug loads
Hybrid DL + physics	22	15–28	6	HVAC, domestic hot water
Deep RL (DQN, PPO, SAC)	26	19–34	17	HVAC set-points, demand response
LLM-enabled (surrogate + LLM, multi-agent)	31	25–37	10	HVAC orchestration, simulation tasks

summarizes the median energy savings across different algorithm classes. The analysis shows that deep-reinforcement-learning (DRL) and large-language-model (LLM) controllers delivered the highest median savings (26% and 31%, respectively) but also had the widest interquartile ranges, reflecting their sensitivity to data quality and implementation. Hybrid models (DL + physics) exhibited more stable results. It is noteworthy that only 38% of studies reported computational overheads; within this subset, LLM-based workflows incurred a 1.4-fold higher environmental cost per unit of energy saved compared with DRL.

3. Results

3.1. IoT Sensor Ecosystem and Digital Twin

Smart buildings increasingly rely on the Internet of Things (IoT) to elevate occupant comfort, safety, and energy efficiency. When an IoT platform is tightly coupled with the building-management system (BMS), it minimizes energy losses, streamlines day-to-day operations, and enables continuous, bidirectional communication between the fabric of the building, its operators, and its users [51]. Because it supplies fine-grained, real-time data, IoT is now regarded as a cornerstone technology for integrated strategies that boost both energy performance and indoor environmental quality (IEQ) [52,53]. Such data-driven ecosystems are already helping public facilities move from static timetables to responsive, evidence-based decision-making [54].

A prime showcase for IoT in buildings is heating, ventilation, and air-conditioning (HVAC), where equipment performance directly governs comfort and energy demand. Unlike conventional designs that respond slowly to complex, rapidly changing conditions, modern IoT-enabled solutions use sensors and actuators to deliver enhanced monitoring and control [55,56,57]. Continuous streams of temperature, humidity, and CO₂ data underpin model-based optimization routines. This allows for tight temperature bands, which curbs wasteful overheating, while demand-controlled ventilation that reacts to measured and predicted occupancy can cut energy use by 40–70 percent [51,54,58]. Throughout this process, maintaining comfort and meeting specific requirements, such as infection control in hospitals, remain co-equal goals [55,57].

Digital-twin technology and machine-learning (ML) methods push this optimization frontier further. Real-time sensor feeds populate BIM-based models to create live replicas of building systems, which exposes inefficiencies and incipient faults before they disrupt occupants and provides a foundation for predictive maintenance [55,56,59]. Deep-learning (DL) and reinforcement-learning (RL) algorithms extend these capabilities, delivering adaptive control and accurate occupancy prediction [52,55,60].

Despite their effectiveness, the adoption of these advanced methods faces two key challenges. The first is the interoperability barrier among heterogeneous devices, which is being addressed by semantic ontologies and hybrid architectures that blend ML with rich semantic models [61]. The second is the “black-box” nature of complex models, which can erode user trust. This is being solved by explainable AI (XAI), which translates model outputs into human-readable rationales, improving the perceived value of intelligent control solutions [62].

Table 3. Characteristics and performance metrics of IoT solutions in smart buildings.

Technology	Monitored Parameters	Application Area	Sample Size/Scale	Performance Metrics	Degree of Personalization	References
Wireless IoT Environmental Sensors	Temperature, Humidity, CO2, VOCs, Particulate Matter, Pressure	Real-time indoor air quality monitoring and HVAC optimization in smart buildings	3 office buildings (real + simulated)	Avg. 36.8 kW saved/h; up to 6.18 MWh/week	Low	[51]
LoRaWAN Multi-Sensor Networks	CO2, Temperature, Humidity, Occupancy, Motion, Door Status	Occupancy detection and adaptive HVAC control in office buildings for energy savings	10 rooms, RWTH Aachen University	TPR 95%, 0 FNs; LightGBM best F1-score	Medium	[54]
IoT Sensors with Edge Computing	Air Quality Parameters, Temperature, Occupancy	Privacy-preserving AI management of indoor air quality	Synthesis of 34 studies; 8–10 buildings reviewed	IAQ prediction accuracy 90–99% via federated learning	High	[58]
IoT-Enabled BIM Integration Sensors	Environmental Parameters, Energy Consumption	Energy management and refurbishment via combined BIM and IoT data for HVAC systems	2 annexes in Roccaruja Hotel (Italy)	PV: 13.5 kW; 1395 kWh/m2/yr; ML HVAC design	High	[59]
IoT Gateways and Actuators	Environmental Parameters, Equipment Status	Failsafe smart building management with continuous monitoring and backup control	Simulated failures over 24 h; real data from 3 buildings	883.2 kWh/day saved during sensor failure	High	[56]
IoT Devices with Machine Learning Algorithms	Energy Usage, Temperature, Occupancy Patterns	Predictive maintenance, anomaly detection, and adaptive HVAC system control	3 university buildings, 200+ sensors, >100 rooms	R2 = 0.996; CO2 MSE = 535; 10–15% energy saved	Low	[52]
Sensor Networks Integrated with Digital Twins	Temperature, Airflow, Pressure, Energy Data	Dynamic real-time HVAC system simulation and control for hospitals and educational facilities	4 hospital buildings; 7-room office layout simulated	HVAC design time ↓90%; evacuation <180 s; energy ↓31–35%	Medium	[60]

presents a comparative analysis of various IoT solutions. To establish a basis for this comparison, a conceptual unification of metrics was performed: primary parameters such as temperature, humidity, and CO₂ were grouped according to their functional purpose.

This analysis is qualitative in nature due to significant variability among the referenced studies. This variability manifests in several key aspects. Firstly, the sample sizes in the investigated works differ substantially, ranging from the number of deployed sensors and monitoring duration to the scale and type of surveyed buildings. Secondly, the methodologies employed, including sensor models, calibration protocols, and data processing algorithms, are also not standardized.

Despite these limitations, the systematization of data in the table enables a qualitative comparison and helps identify current trends in the application of IoT technologies.

The provided Table 3 clearly illustrates the evolution of Internet of Things (IoT) technologies within smart buildings, demonstrating a distinct transition from basic monitoring systems to complex and personalized management platforms. This progress can be traced through the increasing sophistication of the technology, the expansion of monitored parameters, and, consequently, the rising degree of personalization.

At the initial level are wireless IoT environmental sensors, which gather fundamental environmental data such as temperature, humidity, and CO₂. Their primary application is real-time air quality monitoring and general HVAC optimization, which corresponds to a low degree of personalization, as the system reacts to overall conditions rather than to specific users [51,54]. The next step in this evolution involves network technologies like LoRaWAN multi-sensor networks, which augment environmental data with information on occupancy, motion, and door status. This enables a shift from simple monitoring to adaptive HVAC control based on actual human presence, increasing the personalization level to medium [59].

A high degree of personalization is achieved through more advanced technological solutions. IoT sensors with edge computing allow for local data processing, ensuring privacy and rapid responses to user actions [58]. IoT-enabled BIM integration provides the system with detailed building-specific context, enabling energy management tailored to its unique characteristics [59]. Even IoT gateways and actuators contribute to high personalization by ensuring failsafe and continuous control adapted to the current equipment status [56].

The culmination of this trend can be seen in sensor networks integrated with digital twins, which combine multiple data streams (temperature, airflow, energy data) to create a dynamic, real-time model of the building. This allows for not just reacting to events but simulating and managing complex systems, which is especially critical for facilities like hospitals [60].

Interestingly, the application of IoT devices with machine learning algorithms is itself rated with a low degree of personalization [52]. This can be explained by the fact that, in this context, ML models are often used for system-wide tasks like predictive maintenance or anomaly detection in energy consumption, rather than for customizing an environment for an individual.

3.2. AI and Deep Learning Techniques

Artificial intelligence (AI) and deep learning (DL) have become indispensable for resource management in smart buildings, delivering measurable gains in efficiency and sustainability [63,64]. Current applications span a wide range of tasks, from high-precision occupancy detection and energy-use forecasting to fully adaptive HVAC control [65,66].

One of the key tasks for AI is prediction. For building-scale classification and regression tasks, classical machine-learning (ML) algorithms such as support-vector machines (SVM), k-nearest neighbors (KNN), and decision trees remain effective, while ensemble learners such as gradient boosting often push accuracy still further [20,67,68,69]. For analyzing time series, such as predicting energy demand, recurrent neural networks (RNNs)—especially long short-term memory (LSTM) units—are particularly effective [65,70]. Convolutional neural networks (CNNs), in turn, excel at extracting spatial patterns from camera feeds and heterogeneous sensor arrays [64,66].

Reinforcement learning (RL) and deep reinforcement learning (DRL) are rapidly gaining traction because they derive adaptive control policies under the dynamic, uncertain conditions typical of buildings [71]. Algorithms such as Q-learning and deep Q-networks (DQN) learn optimal HVAC schedules or electric-vehicle-charging strategies directly from interaction data, without an a priori model [64,72,73]. Integrating expert knowledge, such as physics-based models, into the DRL training loop can cut training time by up to 8.8× while keeping thermal and IAQ parameters within comfort bands [73]. For high-dimensional state spaces, hybrid systems that merge DRL with graph neural networks (GNNs) can optimize HVAC set-point recommendations [69,74].

AI and DL also underpin predictive-maintenance regimes for HVAC and other building plant [66,75]. Continuous sensor-data analysis flags incipient faults, enabling a shift from calendar-based to condition-based servicing. This Maintenance 4.0 (M4.0) paradigm weaves together AI, IoT, and big-data analytics to optimize technical upkeep [66]. Marrying AI with digital-twin (DT) technology, which is fueled by live IoT streams, further expands monitoring capabilities. It allows for rapid anomaly detection and event forecasting, such as when SVM and LSTM models embedded in a hospital DT forecast pedestrian traffic to adjust ventilation accordingly [55,75,76].

Despite these gains, several hurdles remain. Protecting personal data is paramount, prompting interest in techniques such as federated learning and differential privacy, often combined with edge architectures [58,65,77]. Data-quality issues and the sheer volume of streaming information demand robust cleansing, curation, and real-time processing pipelines [77]. Interoperability also lags: standard protocols are still needed to knit together heterogeneous IoT devices and BMS platforms [42,78]. Finally, model interpretability and user trust pose ongoing challenges. Explainable-AI (XAI) tools such as SHAP are being developed to address this by translating model outputs into actionable insights [67]. However, scalable, resource-efficient XAI solutions capable of running on constrained edge devices are still in their infancy [20,42,55,58,62,67,68,69,70,71,72,73,74,75,76,77,78]. A stronger focus on occupant behavior, preferences, and feedback will be essential for truly human-centered smart-building systems [65,69].

3.3. LLM in Management of Smart-Building Resources

In an intelligent laboratory, continuous tracking of key environmental variables—temperature, pressure, CO₂ concentration, and overall air quality—is essential, as these factors influence both experimental outcomes and staff well-being [79,80,81]. Large language models (LLMs) provide powerful tools for analyzing quantitative sensor data alongside qualitative human feedback, yielding a more complete picture of micro-climate dynamics and improving the precision of environmental control. Recent empirical studies further substantiate the role of LLMs not only in analyzing environmental states, but also in directly controlling HVAC systems and supporting decision-making in real-world smart buildings [82,83].

Built on transformer architectures and trained on extensive multimodal datasets, LLMs can fuse heterogeneous data streams—sensor readings, textual logs, and subjective user inputs—to uncover subtle relationships between physical conditions and occupant comfort [81,84,85]. Such capability is critical in smart laboratories, where even minor fluctuations can compromise research validity or personnel performance [86]. In simulation settings, a pretrained ChatGPT 4.0 controller achieved a 16.8% reduction in HVAC energy use versus rule-based schedules, closely approaching the 24.1% savings of a fully trained DQN reinforcement-learning agent—despite requiring no additional training or calibration [82].

Through prompt engineering and retrieval-augmented generation (RAG), LLMs autonomously produce clear, structured reports for both technical staff and non-specialist users [85,87,88]. For example, by combining real-time CO₂ and temperature data, an LLM can recommend ventilation schedules and explain their physiological and cognitive benefits for occupants [80,89]. Automating this analysis reduces manual workload and speeds up decision-making. Pilot deployments support this premise: in a multi-zone residential building in Thessaloniki, Greece, a GPT-4-based recommendation system achieved an estimated 10% reduction in electricity use over two weeks by modulating HVAC usage in response to occupancy and photovoltaic generation cycles [83]. Notably, open-source models such as DeepSeek-Qwen-1.5B achieved similar semantic alignment scores at a fraction of GPT-4’s inference cost, making them attractive for high-frequency decision support.

LLMs also act as natural-language interfaces for smart-lab management, allowing users to specify environmental set-points conversationally [86,87,90]. When integrated with IoT sensors and actuators, they orchestrate HVAC, lighting, and safety subsystems to balance comfort and energy use in real time [81,90,91]. These capabilities can extend to multi-agent frameworks that adapt continuously to occupant behavior and evolving conditions [90]. In one such implementation, a local LLM assistant running on Raspberry Pi devices achieved an action-execution F1 score of 0.94 and autonomously maintained optimal thermal conditions with over 90% accuracy based on occupancy and comfort thresholds—without cloud reliance [83].

Evidence from analogous smart-environment deployments underscores the transformative potential of large language models (LLMs) [92]. LLMs already automate the creation of detailed EnergyPlus input files and help interpret complex simulation outputs, streamlining workflows that once required expert labor [80,81,84]. When paired with interpretable machine-learning techniques, they supply transparent rationales for HVAC-control actions, boosting operator trust and accelerating adoption [80,81]. They also translate natural-language commands into executable automation scripts for smart devices, delivering seamless, personalized environmental control [87,90]. In robotic platforms, LLMs act as high-level agents that plan and execute tasks autonomously in industrial and manufacturing settings [93]. Concepts such as “Follow-Me AI” extend this capability by deploying LLMs across edge and cloud layers to align computational resources dynamically with user needs, thereby enhancing comfort and sustainability [90].

Field evidence already sorts published LLM investigations into three maturity tiers. Deployed systems are still rare but instructive: Follow-Me AI [90] runs a lightweight GPT-variant on edge devices in a 120 m² office and reports an 8% reduction in HVAC electricity over a four-week A/B trial, while GPT-4 automation of EnergyPlus scripting in a laboratory retrofit cut model-setup time by 93% without loss of fidelity [80]. Research prototypes, exemplified by the GPT-4-assisted fault-diagnosis pipeline evaluated by Liu et al. [81], operate on high-resolution datasets and controlled testbeds yet remain unlinked to live BMS hardware. A broader set of papers sits at the conceptual stage, proposing speech-driven facility chatbots, federated fine-tuning, or digital-twin reasoning without quantitative field metrics.

Table 4. Applications of large language models (LLMs) for resource management in smart buildings.

Technology	Monitored Parameters	Application Area	Sample Size/Scale	Dataset Description (Age ± SD, %)	Provenance and Accessibility	Real-Time Applicability	Explainability/Transparency	References
GPT-4, GPT-4-0613	HVAC inputs, temperature	Automates HVAC simulation input generation	10 workflow tests + 3 case studies	EnergyPlus simulations; iUnit building, DOE office	Proprietary or API-accessible LLMs	Medium (simulation focused)	Limited (focus on automation)	[80]
GPT-3.5, GPT-4, open-source LLMs	HVAC states, temperature, energy consumption	Automated HVAC fault detection and optimization	Simulation + semi-real building logs	Technical logs, operational data, and descriptive textual inputs	Open access + proprietary models	Medium-high with system integration	Moderate with multimodal data integration	[81]
GPT-3.5	Outdoor air temp, zone temp, occupancy	Interpretable ML-based HVAC control	DOE small office (511 m2); 31-day simulation	Simulation only; 744 hourly steps + 1-month training	Published datasets + simulation	Medium, with interactive Q&A potential	Strong explainability using SHAP + LLM narratives	[80]
ML surrogate + fuzzy logic	Indoor CO2, temperature, humidity	Stepwise fuzzy-guided HVAC setpoint optimization	CFD simulation of 3.5 × 3.4 × 2.5 m3 room + past real validation	Indoor CFD simulations validated by chamber experiments	Published test chamber data + CFD	High-designed for fast online HVAC control	Moderate; fuzzy logic offers interpretability	[89]
GPT variants (unspecified)	Temperature, humidity, CO2, occupancy, lighting	Personalized, occupant-adaptive HVAC and environment control	Smart campus at Univ. of Oulu; dynamic real-time data	Live user location, device sensors, building sensors	Multi-agent system deployed on real buildings	High-multi-agent edge/cloud AI for low latency control	Moderate; AI agents provide adaptive control, limited explicit explanation	[90]

presents a comparative overview of recent studies employing large language models (LLMs) or hybrid AI approaches for HVAC system management. The table’s goal is to provide a basis for comparative analysis; however, this requires acknowledging the high methodological variability that characterizes the current state of the field and complicates direct comparisons.

An analysis of Table 4 shows that the application of large language models (LLMs) in HVAC management is a rapidly developing yet methodologically heterogeneous field, characterized by key trade-offs between model complexity, real-time applicability, and transparency.

A clear spectrum of model applications is observed, depending on their complexity and purpose. Powerful models, such as GPT-4, are used for complex offline tasks, including the automation of high-accuracy simulation input generation, but their real-time applicability is rated as “Medium” [80]. On the other hand, for tasks requiring rapid responses, either lighter models (GPT-3.5, open-source LLMs) are used for fault detection [81], or specialized hybrid approaches are employed. A system based on a surrogate model and fuzzy logic [89] and a multi-agent system [90] are specifically designed for online control and demonstrate “High” real-time applicability.

Comparing the performance of these systems is complicated by the lack of harmonized metrics. Performance is measured in entirely different ways: from “high accuracy” in simulations [80] and concrete “~35% energy savings” [89] to the qualitative metric of “high operator trust” [80]. This also highlights an important trade-off between automation and explainability. The system in [62], which purposefully uses GPT-3.5 to generate SHAP-based explanations, achieves “strong” transparency. Meanwhile, the most autonomous systems, such as in the study [90], provide adaptive control but have only “moderate” explainability.

The table demonstrates a clear shift from using LLMs as an auxiliary offline tool toward their integration into complex, real-time control systems. However, for the field to advance, a standardization of evaluation methodologies and increased transparency in reporting, especially regarding the scale of experiments, are necessary.

In

Table 5. Evidence–maturity–sustainability matrix for AI- and IoT-based resource anagement platforms in smart buildings.

References	Platform/Algorithm	Study-Level Evidence Grade	TRL	Median Energy Saving %	Carbon Cost per kWh	Estimated Pay-Back
[42,58,62,76,77,78,79,80]	GPT-4, GPT-4-0613	B (1–2 studies, demonstration pilots)	TRL 6–7	Up to 35%	0.08 USD/kWh	1–2 years
[62,78,79,80,81]	GPT-3.5, GPT-4, open-source LLMs	B	TRL 6	25–30%	0.07 USD/kWh	2–3 years
[62,79,80,81,82,83,84,85,86,87,88,89,90,91,92]	GPT-3.5	Outdoor air temp, zone temp, occupancy	TRL 5	~20%	0.09 USD/kWh	3–5 years
[42,62,78,79,80,81,82,83,84,85,86,87,88,89]	ML surrogate + fuzzy logic	C (single pilot)	TRL 5–6	~35%	0.05 USD/kWh	2 years
[42,55,58,62,76,77,78,79,80,81,82,83,84,85,86,87,88,89,90]	GPT variants (unspecified)	B	TRL 6–7	Up to 40%	0.06 USD/kWh	1–2 years

we presented a comprehensive matrix of AI and IoT-based resource management platforms in smart buildings, evaluating their evidence level, technology readiness, energy savings, carbon cost, and payback periods. Higher maturity platforms like GPT-4 variants achieve up to 40% energy savings with payback times of 1–2 years and moderate carbon costs, while lower maturity or single-pilot studies report more modest savings and longer paybacks. Hybrid approaches combining ML surrogates and fuzzy logic deliver competitive savings with lower carbon impact. This matrix highlights the balance between technological maturity, sustainability, and economic viability, underscoring progress toward scalable, effective smart building solutions and the need for expanded validation and deployment.

The relationship between technology maturity, energy-saving performance, and explainability, identified in the tabular analysis, is visually represented in Figure 5.

Figure 5 provides a comparative overview of five AI-driven HVAC control platforms, each plotted according to its median reported energy savings (%) and corresponding technology readiness level (TRL). The color scale reflects the degree of explainability, ranging from strong (green) to limited (red), based on the use of interpretable methods such as SHAP values or fuzzy logic narratives. Real-world implementations using multi-agent GPT variants [90] report energy savings of up to 40% at TRL 6–7, albeit with moderate transparency. Simulation-focused solutions, such as GPT-4-0613 [80], demonstrate comparable energy performance but offer limited explainability. Hybrid systems incorporating fuzzy logic [89] and interpretable machine learning [80] represent a compromise between performance and trust. This visualization underscores key trade-offs between technological maturity, transparency, and energy impact in current LLM-enabled building control frameworks.

Comparative analysis of these applications highlights three consistent findings. First, while LLMs eliminate the prolonged training phase required by traditional reinforcement-learning methods, they still underperform DQN in energy savings under comparable conditions [82]. Second, semantic performance plateaus quickly with model size: open models between 1 and 2 billion parameters reach ≥80% of GPT-4’s alignment at an order of magnitude lower cost and are better suited to embedded or low-power settings [83]. Third, most published deployments are short in duration (3–14 days) and fail to report comfort or indoor air quality violations, undermining confidence in the broader applicability of their energy-saving claims.

Deploying LLMs in smart laboratories, however, introduces domain-specific challenges. Effective use demands fine-tuning and semantic enrichment to handle specialized terminology, diverse sensor modalities and bespoke experimental protocols [79,91]. Secure integration within IoT ecosystems calls for robust data-governance frameworks that safeguard laboratory and user information [79,94]. Real-time interoperability with existing control systems hinges on standardized protocols and ultra-low-latency operation—both active research areas [86,87]. Running large LLMs on edge devices or in latency-sensitive environments further requires model-compression and hardware-aware optimization techniques to meet computational and energy constraints [81,84].

Future work is likely to fuse LLMs with multimodal data streams—images, audio and high-frequency sensor signals—to extend smart-laboratory functionality [81,91]. Agent-based workflows, interpretable-AI frameworks and collaborative federated-learning schemes should improve model robustness and foster user trust [73,80,91]. Integrating digital-twin technology with LLM-driven natural-language interfaces promises new avenues for virtual experimentation and remote laboratory operation [86].

3.4. Smart Readiness Alignment

The Smart Readiness Indicator (SRI), introduced in the European Parliament Directive 2018/844/EU (EPBD), represents a unified framework for assessing the integration of information and communication technologies and automation in buildings. Its purpose is to improve energy efficiency, energy flexibility, and user interaction [95]. At present, most SRI assessments rely on qualitative approaches (Method B), which are heavily dependent on expert interpretation. However, the SRI methodology is evolving toward quantitative evaluation based on actual performance data (Method C), which creates a solid foundation for integrating more advanced technologies [96].

Although artificial intelligence and deep learning (AI/DL) methods are not yet part of the official SRI methodology, this shift toward data-driven approaches opens up significant potential for their implementation across several key domains. In the area of energy flexibility and demand response participation, which are core to the SRI, deep reinforcement learning (DRL) can provide adaptive control of HVAC systems and optimize energy consumption in real time [97].

In addition to this, smart services like predictive maintenance and fault detection can be significantly enhanced using graph neural networks (GNNs) and predictive analytics, which can identify hidden patterns in building system performance [98]. The capabilities of GNNs also extend to modeling the interdependencies between buildings and the urban energy grid, which is especially relevant in the transition to smart districts and cities. Finally, the future implementation of Method C creates a direct pathway for applying explainable AI (XAI), which can improve the transparency and explainability of automated decisions for building operators and occupants. These potential alignments between SRI domains and AI/DL technologies are summarized in

Table 6. Potential application of AI/DL technologies in SRI domains.

SRI Domain and Effect	AI/DL Technology	Potential KPI	SRI Method
Heating, Cooling, Flexibility	DRL	Peak reduction (kW), energy savings (kWh)	C
Monitoring and Control	GNN, XAI	Fault prediction accuracy, explainability	B/C
Grid Interaction	DRL, Forecasting	Demand response participation (%)	B/C
Occupant Comfort and Engagement	NLP, XAI	Occupant satisfaction index	B
Urban-Level Integration (Smart City)	GNN	Load shifting optimization	C/District SRI

.

Although DRL, GNN, and XAI are not yet part of the standard SRI assessment procedure, the evolving methodology and the emphasis on automated, data-driven approaches make them the logical next step in the development of smart readiness evaluation. The integration of these technologies will allow SRI assessments to become more objective, scalable, and directly linked to actual building performance, as recommended in several recent publications [97,99].

4. Discussion

4.1. Security and Data Quality Challenges

Despite their immense potential, the integration of IoT, machine learning, and AI in smart buildings faces several critical challenges that hinder reliable and efficient real-world implementation. One of the primary challenges is handling unreliable sensor data, which can lead to incorrect system decisions. Errors may arise from device malfunctions or attacks like False Data Injection, and detecting them is difficult due to high data volume and the reliance on basic aggregation techniques such as calculating averages or minimum values [88]. Capturing human behavior and environmental context adds an additional layer of complexity [20].

In addition to these issues, poor data quality poses a major obstacle in smart buildings, often stemming from calibration errors, transmission issues, environmental conditions, human mistakes, and outdated or non-standardized technologies. Cybersecurity threats and poor IT system design in smart buildings further degrade data integrity and system reliability. Inadequate protection exposes critical components such as HVAC, access control, elevator systems, and even parking management to potential cyberattacks. As a result, datasets commonly include anomalies like duplicates, outliers, and missing values, which distort analyses of energy use, occupancy, and air quality. Consequently, automation systems suffer, reducing occupant comfort and building efficiency. Ultimately, low-quality data damages stakeholder trust, and impedes progress toward sustainability goals and regulatory compliance [100,101].

4.2. Integration and System Constraints

The deployment of smart systems is often cost-prohibitive and typically requires replacing legacy infrastructure. Supporting thousands of devices that operate under different standards is a significant challenge. Limited interoperability among different devices and protocols complicates integration, hindering the development of cohesive, efficient smart building solutions. IoT devices generate large volumes of heterogeneous data in diverse formats, structures, and semantic meanings, and without careful handling, this complexity can lead to semantic misinterpretations or degraded system performance. To address this, middleware must harmonize various data formats and communication protocols to enable interoperability [20,24,102]. In this context, standardized schemas and adaptable middleware architectures become essential for building scalable and interoperable systems. A practical constraint arises from the limited computational resources of IoT devices, which often lack the power and memory needed to run complex ML and LLM models [103].

Current building management systems offer limited adaptability, relying on predefined rules rather than learning from performance data. This restricts their ability to dynamically adjust to changing conditions or optimize for long-term sustainability. Research into self-supervised and reinforcement learning methods is actively addressing these adaptability issues [100]. Another critical concern is energy consumption. Running AI and LLM models significantly increases power usage, which conflicts with the energy-efficiency goals of smart building management. Approaches like FastML (fast machine learning) aim to reduce power requirements, but balancing system performance with sustainability remains a significant challenge [100].

4.3. Privacy, Ethics and Regulatory Requirements

Among the pressing issues in AI and machine learning is the vulnerability to data poisoning attacks. These attacks refer to intentionally manipulating the training data of an AI model to disrupt its decision-making processes. When training data is gathered from external or unverified sources, attackers may introduce manipulated data samples. This manipulation leads to incorrect outputs and faulty decision-making, and in some cases, Trojan attacks are created by injecting malicious data to embed hidden vulnerabilities [104]. In addition to security concerns, privacy risks are a major challenge in smart buildings. Many smart devices, such as meters and health monitors, collect sensitive personal information that is often transmitted to cloud services or external providers without sufficient safeguards. This lack of privacy protection can expose occupants to data breaches and misuse. Moreover, smart building systems face scalability issues as the number of connected devices grows, straining existing infrastructure. Addressing these issues through unified standards is crucial for advancing smart building energy management systems (BEMS) [24].

The pervasive data collection intrinsic to IoT ecosystems raises complex privacy and data-protection issues, especially under stringent frameworks such as the European Union’s General Data Protection Regulation (GDPR). The adoption of IoT within smart buildings has resulted in the generation and processing of vast quantities of sensitive personal data. Data originates from heterogeneous sources, including sensors monitoring occupancy, energy consumption, environmental parameters, and even behavioral patterns. By design, these systems require data protection measures that not only secure communication channels and data storage but also adhere to the regulatory mandates of frameworks like the GDPR. The GDPR establishes privacy by design and by default as fundamental principles, thereby requiring system architects to integrate robust privacy safeguards into the core architecture of smart building platforms [105]. The regulatory landscape also demands clear communication regarding data practices and rights, creating the need for user interfaces that explain in simple terms how personal data is collected, processed, and stored. This transparency is critical for enhancing trust in smart building systems. Privacy impact assessments (PIAs) also serve as a mechanism for measuring inherent risks and guiding system modifications to better protect personal data [106]. The integration of privacy by design principles with advanced technical safeguards forms the cornerstone of modern, GDPR-compliant smart building architectures.

4.4. Energy-Performance Trade-Offs

Buildings are major drivers of global energy demand and greenhouse-gas emissions. Smart-building technologies mitigate this footprint by combining real-time data, predictive analytics, and demand-response control to align consumption with grid conditions, integrate on-site renewables, and optimize day-to-day operations—without sacrificing occupant comfort. Photovoltaic arrays, battery storage, and advanced building-energy-management systems (BEMS) further raise efficiency, while the EU Energy Performance of Buildings Directive (EPBD) underpins retrofit programs that move the stock towards nearly zero-energy-building (NZEB) performance targets [107].

The May 2024 recast of the Energy Performance of Buildings Directive will make the Smart Readiness Indicator (SRI) mandatory for large non-residential buildings from June 2027 [108].

Ontology-driven workflows are already cutting audit time: Kourgiozou et al. automatically derived campus-scale SRI scores from Display Energy Certificate records, eliminating most manual checks [109].

On the control side, Wang et al. showed that a DRL HVAC agent in a 10,000 m² office cut annual energy use by ≈26% and peak demand by ≈10%, gains that directly raise the SRI ‘energy-saving’ and ‘flexibility’ subscores [110].

Yet no open dataset is tagged with the 56 SRI smart-service functions, and black-box DRL policies remain hard to audit—underscoring the need for explainable, SRI-aware benchmarks before the 2027 compliance deadline [108].

The opacity (“black-box” nature) of AI- and deep-learning-based controllers creates significant risks and hampers their adoption in building energy management systems (BEMS) and HVAC, eroding the trust of operators and occupants. The EU Artificial Intelligence Act therefore classifies such controllers as high-risk AI systems and explicitly mandates transparency and explainability [35]. The recast Energy Performance of Buildings Directive—EPBD 2024/1275 introduces the Smart Readiness Indicator (SRI), likewise rewarding solutions whose logic is understandable to building users [36]. Similar requirements appear in the international ISO/IEC TR 24028:2021 standard on trustworthy AI [37]. Empirical studies already show that explainable AI can raise building-energy efficiency by 15–30% without compromising thermal comfort [111]. Consequently, integrating XAI is becoming a prerequisite both for regulatory compliance and for large-scale diffusion of smart-building technologies.

Sustainability planning increasingly spans the entire building life cycle. Life-cycle-assessment (LCA) tools now inform design teams about the embodied carbon in construction materials and processes, enabling lower-impact choices from ground-breaking through operation to end-of-life [112].

Yet post-retrofit studies reveal that HVAC performance can drift. One investigation recorded substantial energy savings after envelope air-sealing and controls upgrades, only to observe a gradual erosion linked to uncontrolled plug loads and sub-optimal control logic [113]. Similar work shows that air-source heat pumps and chillers lose efficiency over time unless continuous maintenance and adaptive controls are in place [114]. Data-driven system-identification techniques—autoregressive models and recursive least-squares algorithms—can surface subtle faults that leave temperatures unaffected but steadily degrade performance [115]. To preserve predictive-control accuracy, operational pipelines therefore need concept-drift detection that flags slow shifts in equipment behavior and triggers recalibration or retraining [116].

4.5. The Role of Digital Twins and Advanced Models

Blending AI, IoT, edge computing, and digital twins enables intelligent automation and optimization in smart building energy management. A three-dimensional data model, integrated with IoT sensors and building information modeling (BIM), supports real-time monitoring and scenario analysis for energy interventions. Edge computing ensures cost-effective, reliable data processing, reducing latency and dependence on centralized systems. AI enhances the digital twin’s predictive capacity, enabling it to learn from past events, optimize outputs, and make autonomous decisions. This integrated architecture promotes virtuous energy use while preserving indoor comfort, laying the groundwork for advanced, responsive smart-city ecosystems [117]. Digital twins in construction are defined as dynamic digital models maintained throughout a project’s lifecycle, enabled by a bidirectional link with the physical asset through Internet of Things (IoT), Cyber-Physical Systems (CPS), artificial intelligence (AI), and sensors. This integration allows real-time enrichment with semantic data, supporting accurate representation, continuous monitoring, and feedback. Practitioners highlight their role in enhancing information transparency and enabling predictive simulations and what-if analyses for proactive energy and safety management [118]. Digital twins extend beyond traditional BIM by enabling real-time interaction between a building’s indoor environment and its virtual counterpart. Unlike BIM, which serves mainly for design and construction, digital twins continuously monitor operational conditions and support predictive maintenance through real-time sensor data. Their development integrates multiple technologies—3D CAD modeling, wireless sensor networks (WSNs), machine learning, and data analytics—forming a cross-disciplinary framework. A practical example includes a digital twin of an office facade using WSNs to track light, temperature, and humidity, demonstrating benefits such as improved energy efficiency, proactive maintenance, and actionable insights for future building design [119].

AI-based simulation models identify patterns and inform decisions on building activities, layout, and functionalities, enhancing both performance and user experience. Digital twins serve as real-time digital replicas of physical structures, replicating and monitoring building behavior. IoT facilitates intelligent sensing and data collection, supporting informed and adaptive decisions. This holistic system enables continuous evaluation of design alternatives and performance metrics, fostering efficient, user-centered building environments [120]. AI-driven digital twins, enhanced by virtual reality (VR), are evolving toward autonomy through structured capability levels—from static virtual representations to self-controlling systems. Integrating AI with Internet of Things (IoT) sensors, such as Philips Hue and Disruptive Technologies, enables real-time monitoring and predictive control. Hybrid modelling approaches, including data-driven modelling (DDM) and physics-based modelling (PBM), support functions like indoor temperature and sun position forecasting. The concept of big data cybernetics further strengthens autonomous operation, using hybrid analysis and modelling (HAM) to guide the system toward optimal performance [121]. The integration of AI, Artificial Intelligence of Things (AIoT), and urban digital twins (UDTs) enhances data-driven planning in sustainable smart cities. AIoT embeds AI capabilities directly into Internet of Things (IoT) devices, enabling real-time sensing, analysis, and decision-making. This synergy allows UDTs to run AI-driven simulations for scenario evaluation and resource optimization, supporting more accurate environmental assessments and proactive urban sustainability strategies [122]. In addition, AI integration within Cyber-Physical Systems (CPS) is advancing toward real-time digital twins capable of mirroring and interacting with physical systems. Edge AI supports distributed intelligence and human-computer interaction, enhancing resilience at both technical and organizational levels. The shift toward cognitive, self-adaptive, decentralized CPS—where agents use cloud services to predict, optimize, and self-adjust—illustrates a growing convergence of AI, Internet of Things (IoT), and edge computing. While rooted in Industry 4.0, these developments are increasingly applicable to smart cities and other real-world systems [123].

4.6. Real-World Implementation of AI and Machine Learning

Research indicates that while generative AI has been explored for HVAC control in simulations, real-world validation remains scarce. Studies demonstrated the potential of LLMs like GPT-4 and ChatGPT in virtual simulations but did not incorporate real-world complexities, leaving a significant empirical gap in building automation [124]. Previous research on energy-saving strategies in BMS, including reinforcement learning and hybrid HVAC models, faces scalability limitations across diverse environments. While LLMs offer promising solutions, there is a critical gap in integrating semantic models with LLMs for BMS. Additionally, even fine-tuned ML models within a multi-agent framework fall short of benchmarks due to inadequate data preprocessing and direct use of raw, inconsistent data, which challenges effective training and evaluation [125]. The experimental setup of LLMs is restricted to specific instance types, limiting generalizability, and ongoing AI advancements require continuous model evaluation to keep findings relevant [126]. They have fundamental limitations in performing real-world physical tasks, as they lack grounding in reality, common sense, true reasoning, and planning abilities. Relying solely on LLMs for integration with physical control systems can lead to incorrect actions. While reinforcement learning methods may enhance real-world integration, they face challenges like poor sample efficiency, limited generalization, and the costly, complex design of reward functions [127]. Current works of deep learning reinforcement are limited by reliance on simulated environments with deterministic rules and repeatability. Real-world tasks like smart building automation are complex, are rarely deterministic due to stochastic factors like people and weather, and involve continuous actions such as temperature setpoints. Significant progress is still needed to apply DRL to real-world applications, especially given the challenge of naive discretization of continuous action spaces, which becomes intractable due to the curse of dimensionality [128]. A major barrier to AI integration is the lack of large-scale, annotated datasets needed for training deep learning models. Additional limitations include scalability and interoperability issues due to proprietary protocols and vendor competition, lack of real-time data monitoring, and broader challenges such as legal, regulatory, security, privacy, and market competition obstacles [129]. To address these empirical gaps, a structured research roadmap has been developed and is presented in Figure 6, outlining key milestones and future directions necessary for advancing the integration of LLMs and DRL in real-world building automation systems.

4.7. Methodological Challenges and Research Quality

Our ROBIS (risk of bias in systematic reviews) appraisal highlighted marked variation in methodological quality. Among the 143 publications examined, 81 met all core criteria and were judged “low risk of bias.” In contrast, 43 showed at least one major shortcoming and were classified “high risk of bias”. A further 65 lacked sufficient methodological detail to permit a confident verdict and were therefore labelled “unclear.” Finally, 21 studies were marked “not applicable” for specific ROBIS domains because those domains did not pertain to the study’s design or scope. (Note: counts exceed 143 because a single study can fall into more than one ROBIS category across different domains.) This finding requires careful consideration, as it directly influences the interpretation of this review’s overall conclusions. The presence of studies with a high or unclear risk of bias warrants a cautious interpretation of the aggregated evidence, particularly regarding the performance claims of novel AI/DL approaches. The conclusions drawn from research with “No” or “Unclear” ratings may overestimate the true effect size or lack the generalizability suggested by their authors. The lack of transparency in these papers—a key reason for an “Unclear” rating—also hinders the scientific community’s ability to replicate and build upon the work, thereby slowing down tangible progress. Therefore, while the trends identified in this review are prominent, the robustness of the evidence for their real-world efficacy is not uniform. The high-risk-of-bias studies temper the enthusiasm for these emerging technologies, highlighting that many applications may still be in a proof-of-concept phase rather than being mature, validated solutions. This underscores a critical need within the field for more methodologically sound research characterized by transparent reporting, robust validation methods, and adherence to established scientific practices. Future research should prioritize not only innovation but also the methodological quality necessary to build a truly trustworthy and reliable evidence base for AI in smart buildings.

These differences highlight the need for cautious interpretation of the current evidence base, particularly when comparing findings across diverse AI methodologies, datasets, and smart-building platforms. A recurring limitation observed in several studies was the lack of transparent reporting on dataset composition and provenance. In the absence of clear documentation on data sources, demographic characteristics, and environmental collection conditions, it becomes difficult to assess potential biases or to replicate results across different operational contexts.

5. Limitations

Despite the enormous potential of integrating IoT, machine learning, and artificial intelligence into smart buildings, this process faces a range of critical challenges that hinder reliable and effective implementation in real-world scenarios [23]. One of the primary difficulties lies in handling unreliable sensor data, which can lead to incorrect system decisions [100]. These errors may stem from device malfunctions or cyberattacks such as False Data Injection, which are particularly difficult to detect given the sheer volume of data and the reliance on basic aggregation techniques [83,100]. Moreover, incorporating human behavior and contextual environmental factors introduces additional layers of complexity [17].

Compounding these technical hurdles is the high cost of deployment, which often necessitates replacing outdated infrastructure [24,57]. Smart buildings must support thousands of devices operating under diverse standards, creating significant compatibility issues [61]. The lack of interoperability between devices and communication protocols continues to obstruct the development of integrated and efficient solutions [23,61]. Additionally, IoT ecosystems generate vast amounts of heterogeneous data in varying formats and structures, which, without proper preprocessing, can lead to semantic inconsistencies and decreased system performance. Addressing this requires middleware platforms capable of harmonizing data formats and protocols to ensure interoperability and scalability [20,24,102]. In this regard, standardized data schemas and adaptive middleware architectures become essential [61]. However, these efforts are often constrained by the limited computational capabilities of edge devices, which typically lack sufficient processing power and memory to support complex ML models and large language models (LLMs) [103].

Security concerns further complicate the adoption of intelligent systems. Cybersecurity threats, combined with poorly designed IT infrastructure, reduce data integrity and undermine system reliability [101]. Insufficient protection exposes critical building components—such as HVAC systems, access control, elevators, and even parking systems—to potential attacks [23,101]. The result is often the presence of anomalies in datasets, including duplicates, outliers, and missing values, which distort energy analytics, occupancy detection, and air quality monitoring [100]. These flaws degrade automation outcomes, reduce occupant comfort, and ultimately impair building efficiency. In turn, poor data quality erodes stakeholder trust and impedes the fulfillment of sustainability objectives and regulatory compliance [100,101].

Equally pressing is the issue of data poisoning in AI and ML applications. These attacks involve the intentional manipulation of training data to compromise model performance and decision-making capabilities [104]. When data are sourced from external or unverified inputs, adversaries can inject falsified examples that result in faulty predictions or even embed Trojan-like vulnerabilities. Such risks emphasize the need for secure data pipelines and robust validation mechanisms during model training.

In parallel with security threats, smart buildings raise serious concerns regarding privacy and ethical data use [105]. A growing number of intelligent devices, such as smart meters and health monitors, collect sensitive personal information, often transmitting it to cloud-based services or third-party vendors without adequate safeguards [105,106]. The pervasive data collection characteristic of IoT ecosystems introduces complex privacy and compliance challenges, particularly considering regulations such as the EU’s General Data Protection Regulation (GDPR) [105]. These regulations emphasize “privacy by design and by default,” mandating that system architects embed privacy protections into the core infrastructure of smart building platforms. Furthermore, conducting privacy impact assessments (PIAs) becomes essential for identifying inherent risks and guiding design decisions to enhance personal data protection [106].

Another significant concern is the trade-off between energy consumption and system performance. Most existing building management systems lack adaptive capabilities, relying instead on predefined control rules rather than learning from performance data [73]. This limits their ability to dynamically respond to changing environmental or operational conditions. Moreover, running AI and LLM workloads considerably increases energy usage, which runs counter to the goals of energy-efficient building design [81]. While methods such as FastML aim to mitigate energy demands, finding a balance between computational performance and long-term sustainability remains a challenging task [100].

Considering that buildings are among the largest consumers of energy and producers of greenhouse gas emissions [7], sustainability strategies increasingly focus on the entire building life cycle—from construction through operation to decommissioning [112]. Policy initiatives like the Energy Performance of Buildings Directive (EPBD) encourage modernization toward nearly zero-energy buildings (NZEB) [93]. Tools such as Life Cycle Assessment (LCA) are being used more frequently by architects and developers to quantify the carbon footprint of materials and processes, thereby supporting more sustainable design choices [112].

Nevertheless, this review also identified several limitations that should be considered when interpreting the findings. Notably, there was considerable variation in the methodological rigor of the reviewed studies, requiring cautious interpretation of reported performance outcomes for AI- and DL-based approaches. Furthermore, several constraints limit the generalizability of the conclusions. First, the scope of this review was restricted to the English-language literature, which introduces a potential selection bias and may exclude relevant findings published in other languages. Second, this study is conceptual rather than empirical—it synthesizes the existing literature without conducting meta-analyses or original experiments, which restricts the ability to draw quantitative conclusions. Third, a recurring issue noted across multiple studies is the lack of transparency regarding the composition and provenance of datasets used in AI research [25]. The scarcity of publicly available, standardized, and high-quality datasets poses a serious barrier to reproducibility and meaningful performance benchmarking. Finally, much of the research, especially in the emerging field of LLM applications, remains dependent on proprietary pilot projects and closed-source APIs—such as GPT-4—highlighted in study [80]. This reliance hinders independent validation and slows progress in open scientific inquiry.

6. Conclusions

This review shows that artificial intelligence methods—especially deep learning (DL) and deep-reinforcement-learning (DRL) controllers—have advanced from proof-of-concept to demonstrably effective tools for smart-building resource management. Across 143 peer-reviewed studies published between 2019 and April 2025, researchers report median HVAC energy savings of 18–35%, with DRL consistently topping performance benchmarks. The field is also diversifying: graph neural networks (GNNs) now capture spatial dependencies in building data, pilot federated-learning schemes address privacy constraints, and early integrations of large language models (LLMs) suggest natural-language interfaces that could democratize analytics and control.

Yet large-scale deployment remains hindered by interlocking barriers. Training data sets are still scarce, fragmented, and often proprietary, limiting model transferability across climates and building archetypes. Privacy and cybersecurity concerns persist because sensor streams may carry personally identifiable information, but few studies implement rigorous threat models or privacy-preserving techniques. Meanwhile, the computational and carbon costs of ever-larger models clash with the sustainability goals these systems aim to support, and the opacity of “black-box” algorithms—coupled with heterogeneous legacy equipment—undermines stakeholder trust and slows adoption.

Shifting from opaque, black-box optimizers toward explainability-by-design is already a legal and ethical baseline: the EU Artificial Intelligence Act classifies building-control AI as high-risk and mandates demonstrable transparency, while both the EPBD Smart Readiness Indicator and ISO/IEC TR 24028:2021 echo the same demand [35,37]. Embedding XAI modules directly into digital twins and operator chat-interfaces powered by large language models can preserve the reported 15–30% energy gains while making control actions auditable and human-overridable.

The literature points to several converging priorities that can unlock the full potential of AI-driven smart buildings. A significant barrier hindering progress in this field is the scarcity of standardized, high-quality datasets. Therefore, a crucial future research direction is the creation of open, high-fidelity benchmark datasets. These datasets must go beyond raw sensor readings to include standardized metadata, diverse building typologies and climatic conditions, and, most importantly, direct occupant feedback. The reliance on fragmented, proprietary data severely limits the generalizability and reproducibility of findings. Establishing a common, publicly available benchmark would catalyze progress by enabling transparent and fair comparison of different AI models. For the industry, such benchmarks would de-risk the adoption of new technologies by providing a clearer business case and validated performance expectations, thereby accelerating the development of more effective and commercially viable smart building solutions.

The escalating computational and carbon footprint of increasingly complex AI models necessitates a research focus on energy-aware and edge-optimized architectures. The current paradigm of streaming vast amounts of IoT data to centralized cloud servers is not only energy-intensive but also introduces significant latency and privacy vulnerabilities. Future work should prioritize the development of lightweight DL models and leverage federated learning to process data locally on edge devices. This approach minimizes data transmission, reduces power consumption, and enhances data privacy. Scientifically, this will spur innovation in novel algorithms and model compression techniques that are both powerful and efficient. From an industrial perspective-optimized solutions will pave the way for more scalable, resilient, and cost-effective smart building systems that can deliver real-time control and responsiveness while contributing to sustainability goals. Given that smart buildings collect and analyze granular data on occupant behavior, a pressing need exists for research into privacy-centric learning frameworks and robust cybersecurity. The unresolved privacy and security risks associated with continuous IoT telemetry are major impediments to large-scale deployment. Future research must advance privacy-preserving AI techniques, such as federated learning and differential privacy, to ensure compliance with stringent regulations and build occupant trust. The success and social acceptance of smart building technologies are contingent on the occupants’ trust, which can be easily eroded by security breaches. Scientifically, this will lead to a new generation of AI systems that are “private by design.” For the industry, the ability to offer verifiably secure and privacy-respecting solutions is not just a competitive advantage but a fundamental prerequisite for market access and widespread adoption.

The “black-box” nature of many advanced AI controllers is a critical barrier to their adoption, as operators are often hesitant to trust systems whose decision-making processes are opaque. Consequently, a vital research avenue is the development of hybrid physics-informed and explainable AI (XAI) models. These hybrid approaches integrate the predictive power of data-driven models with the interpretability of physics-based models, providing a clearer understanding of the system’s logic. Incorporating XAI is indispensable for debugging, ensuring safety, and gaining the trust of facility managers and end-users. Such models can also shorten commissioning times by embedding existing physical knowledge of the building, thereby reducing the dependency on vast amounts of training data. The impact of this research would be transformative, enabling facility managers to work synergistically with AI, leading to enhanced operational efficiency and fostering the development of AI that moves beyond mere prediction towards transparent and causal reasoning.

The integration of digital-twin platforms with LLMs represents a transformative frontier for smart building management. While digital twins offer a comprehensive virtual model of a building, interpreting the vast telemetry data they produce remains a challenge for human operators. Future research should focus on leveraging LLMs to create intuitive, natural-language interfaces that translate this raw data into actionable insights for facility managers. Early research indicates that LLMs can streamline complex simulation workflows and create more intuitive control interfaces. This trajectory could revolutionize human–building interaction, evolving from complex dashboards to conversational AI that allows for personalized environmental control through simple commands. The review shows that current applications of technologies like LLMs are mostly demonstration pilots with TRL between 5 and 7, indicating they are not yet commercially widespread. Similarly, DRL shows promise, its formal integration into standard assessment procedures like the Smart Readiness Indicator (SRI) is recognized as a future step.

If these directions are pursued with an emphasis on open data, responsible computing, and human-centered design, AI and DL can move from isolated pilots to ubiquitous, adaptive and trustworthy smart-building platforms that enhance energy efficiency, indoor-environmental quality and cost effectiveness at scale.