Towards Sustainable Buildings and Energy Communities: AI-Driven Transactive Energy, Smart Local Microgrids, and Life Cycle Integration

Abstract

The transition towards sustainable and low-carbon energy systems highlights the crucial role of buildings, microgrids, and local communities as key actors in enhancing resilience and achieving decarbonization targets. The application of artificial intelligence (AI) is of paramount importance as it enables accurate prediction, adaptive control, and optimization of distributed resources. This paper reviews recent advances in AI applications for transactive energy (TE) and dynamic energy management (DEM), focusing on their integration with building automation, microgrid coordination, and community energy exchanges. It also considers the emerging role of life cycle-based methods, such as life cycle assessment (LCA) and life cycle cost (LCC), in extending operational intelligence to long-term environmental and economic objectives. The analysis is based on a curated set of 97 publications identified through structured queries and thematic filtering. The findings indicate substantial advancement in methodological approaches, notably reinforcement learning (RL), hybrid model predictive control, federated and edge AI, and digital twin applications. However, this study also uncovers shortcomings in the integration and interoperability of sustainability. This paper contributes by consolidating fragmented research and proposing a multi-layered AI framework that aligns short-term performance with long-term resilience and sustainability.

1. Introduction

The accelerating transition towards sustainable and resilient energy systems is profoundly reshaping the design and operation of buildings, communities and distributed infrastructures. In the context of ongoing transformations within the energy sector, particularly with regard to power grids, the importance of local microgrids is increasing [1,2]. This phenomenon can be primarily attributed to the increasing adoption of renewable energy sources (RES), particularly their integration within the infrastructure of residential properties, commercial buildings, building complexes, and local communities. Recent advances in distributed renewable generation, energy storage, and digital infrastructures present significant opportunities to improve efficiency, flexibility, and resilience [3,4]. Concurrently, these advances introduce unprecedented complexities, thereby necessitating intelligent coordination across diverse spatial scales. In consequence, advanced control methodologies and algorithms have become increasingly significant in the organization of energy systems and their efficient utilization [5,6]. Therefore, the advent of Artificial intelligence (AI) has been identified as a pivotal catalyst for this transformation, providing data-driven instruments for forecasting, optimization, and adaptive decision-making that extend from individual devices to entire energy communities [7,8]. This progress results from the emergence of dynamic energy management procedures in recent years, as well as the increased participation of prosumers (individuals and communities) in transactive processes. These procedures require effective and dynamic responses to changes in tariffs, as well as demand and supply levels in local microgrids and the external energy supply system.

In this context, two complementary paradigms have gained particular prominence. The first is transactive energy (TE), which facilitates decentralized, market-based coordination, thereby allowing prosumers and microgrids to trade energy and services according to dynamic value signals [9,10,11]. The second is dynamic energy management (DEM), which focusses on real-time optimization of distributed resources, combining forecasting, control algorithms, and reinforcement learning for adaptive coordination. Recent advances in automation and communication technologies have served to reinforce both of these paradigms [12,13,14]. However, a significant proportion of research in this field continues to prioritize short-term operational objectives over long-term sustainability. In this area, the author has analyzed development directions and identified new organizational concepts for prosumer microgrids, in the context of the ability to support demand-side management (DSM) functions through standard building automation and control systems (BACS), aligning with ongoing research and engineering development [15,16,17]. The advent of sophisticated data processing methodologies and the integration of cloud-based solutions for analysis in subsequent years has guided the research and application trajectory towards ascertaining the viability of organizing energy management systems in homes and buildings using deep reinforcement learning (DRL) [7,18,19]. In parallel, research and analysis on the effective use of tools to support the functional optimization of BACS are being conducted, with a view to improving energy performance and increasing the level of building readiness for smart grid solutions, particularly in the context of RES and energy storage integration [20,21,22,23,24].

Furthermore, beyond operational and market-oriented approaches, a third less developed but increasingly critical dimension concerns the integration of life cycle-based methods—such as life cycle assessment (LCA) and life cycle cost (LCC)—with AI-enabled energy management. Traditionally treated as separate instruments for sustainability evaluation, LCA and LCC are now progressively linked to digital twins, predictive analytics, and building automation [25,26,27,28]. This integration offers the possibility of extending the scope of the TE and DEM frameworks, so that optimization encompasses not only short-term efficiency but also long-term environmental and economic performance. Including carbon footprint, embodied energy, and cost factors in energy management is essential if buildings and energy communities are to align with larger decarbonization trajectories and resilience targets [29,30,31,32,33,34]. In addition, emerging research has begun to extend the discussion of local microgrids toward more constrained and self-sufficient infrastructures, including Closed Ecological Systems (CES). Although this area remains peripheral compared to mainstream building and community applications, CES concepts—developed for space missions or isolated habitats—offer a unique testbed to study how AI-driven energy management, automation, and life cycle integration can operate under extreme sustainability requirements. Insights from such research may, in turn, enrich the development of terrestrial microgrids and energy communities, especially in contexts that require high levels of autonomy and resilience [32,33,34].

Together, the convergence of these domains defines the central scope of this review. Despite substantial progress in each area, the literature remains fragmented, with methodological advances often developed in isolation and with limited transferability across domains. Overcoming this fragmentation is essential to move beyond incremental efficiency gains toward effective Demand and Energy Management (DEM) and systemic sustainability transitions in buildings and energy communities.

In light of this background, the present paper develops a multi-layered AI framework that integrates operational intelligence (perception, prediction, and control) with strategic sustainability assessment (LCA/LCC). Unlike previous reviews that have treated these domains separately, this work proposes a cross-domain synthesis that emphasizes methodological interoperability between real-time energy management and long-term life cycle objectives. The framework positions AI not only as a control enabler but also as a linker between optimization, sustainability, and transactive coordination across system layers.

The paper therefore aims to verify several interrelated theses:

• AI-based energy management should evolve from algorithm-centric implementations toward multi-layer architectures that align operational performance with sustainability goals;
• The LCA/LCC and sustainability dimensions are key components of this architecture, allowing assessment of both operational and embodied energy impacts within a unified framework;
• Cross-domain synthesis between DEM, TE, and life cycle-oriented AI provides the methodological foundation for evaluating resilience, circularity, and resource efficiency.

Accordingly, the objectives of this review are threefold: (i) to synthesize the state of the art in AI applications for transactive and dynamic energy management (TE and DEM) across buildings, microgrids, and communities; (ii) to assess how life cycle-based approaches are incorporated into these AI frameworks; and (iii) to outline integrative research pathways toward sustainability-oriented, interoperable AI architectures. By consolidating these perspectives, this paper clarifies its distinct contribution among existing reviews and provides a coherent methodological direction to link operational intelligence with long-term sustainability evaluation.

The reminder of the paper is organized as follows. Section 2 provides a comprehensive overview of the methodology and systematic elements that were applied in the process of searching and screening the literature. The primary results of the review are described in Section 3, covering transactive energy, dynamic energy management, AI methodologies, and complementary life cycle perspectives. Section 4 provides a critical discussion, situating the findings within broader methodological and conceptual debates, and outlining a multi-layered framework for AI-driven sustainable energy systems. The final section, Section 5, highlights the original contributions of the paper, identifies research gaps, and suggests future directions for research.

2. Materials and Methods

To address the research objectives outlined in the Introduction, a structured procedure for literature identification and selection was adopted. Although this review paper follows the structure of a classical narrative review (IMRAD format), the process incorporated systematic review elements such as transparent queries, multi-stage filtering, and explicit inclusion/exclusion criteria, ensuring both rigor and thematic flexibility.

2.1. Literature Search Approach and Queries

Before defining the final query set, an initial thematic search was conducted in Google Scholar to map the overall research landscape. This exploratory step yielded several thousand records per thematic area (AI and Transactive/Peer-to-Peer Energy and Smart Local Energy Systems/Microgrids and Life Cycle Assessment/Life Cycle Cost and Buildings and Sustainable/Smart/Green Buildings and Energy Performance); however, nearly 80% of these items were not peer-reviewed and lacked formal identifiers (DOI or ISBN) or full-text accessibility. Consequently, the Google Scholar results were used solely to refine the search terms and thematic categories, while the final dataset was compiled exclusively from Web of Science (WoS) and Scopus to ensure scientific reliability, transparency, and reproducibility. Therefore, a structured literature search was carried out in the WoS Core Collection and Scopus, which were selected for their extensive coverage of high-impact journals and conference proceedings (e.g., IEEE, ACM, Elsevier conferences). The time frame under review was limited to 2015–2025, reflecting the period of rapid development of AI applications in energy systems. The database queries were conducted between 18 and 23 August 2025. The focus of the study was on research addressing AI-driven approaches to energy management in buildings and microgrids, including sustainability perspectives.

The scope of this review was captured by four thematic areas. Queries were constructed in WoS using the Topic Search (TS) field (title, abstract, keywords), and equivalent TITLE-ABS-KEY queries were used in Scopus.

• AI + Transactive/Peer-to-Peer Energy WoS example: TS = (“artificial intelligence” OR “machine learning” OR “deep learning” OR “reinforcement learning” OR AI) AND TS = (“transactive energy” OR “peer-to-peer energy” OR “P2P energy”) AND PY = 2015–2025;
• AI + Smart Local Energy Systems/Microgrids WoS example: TS = (“artificial intelligence” OR “machine learning” OR “deep learning” OR “reinforcement learning” OR AI) AND TS = (“local energy system” OR “smart local energy system” OR “smart microgrid”) AND PY = 2015–2025;
• AI + Life Cycle Assessment/Life Cycle Cost + Buildings WoS example: TS = (“artificial intelligence” OR “machine learning” OR “deep learning” OR “reinforcement learning” OR AI) AND TS = (“life cycle assessment” OR “life cycle cost” OR “LCA” OR “LCC”) AND TS = (“building” OR “buildings”) AND PY = 2015–2025
• AI + Sustainable/Smart/Green Buildings and Energy Performance WoS example: TS = (“artificial intelligence” OR “machine learning” OR “deep learning”
  OR “reinforcement learning” OR AI) AND TS = (“sustainable building” OR “building energy performance”) AND PY = 2015–2025.

2.2. Initial Identification

The database search yielded a total of 2101 records (715 from WoS, 1386 from Scopus). The distribution of the thematic areas is presented in

Table 1. Initial search results (2015–2025, according to defined queries).

Set of Records	Thematic Area	Web of Science	Scopus	Total
1	AI + Transactive/Peer-to-Peer Energy	189	322	511
2	AI + Smart Local Energy Systems/Microgrids	53	122	175
3	AI + Life Cycle Assessment/Life Cycle Cost + Buildings	149	211	360
4	AI + Sustainable/Smart/Green Buildings and Energy Performance	324	713	1055
Total		715	1386	2101

.

2.3. Selection and Eligibility

The records were processed through a multi-stage filtering procedure, carried out on the basis of abstracts and keywords.

• Stage 1—Basic merging: Publications were retained only if they were present in both databases (WoS and Scopus), included a valid DOI, and had complete metadata (e.g., authorship information). This step reduced the dataset to 614 publications;
• Stage 2—Thematic filtering: Abstracts and keywords were selected for explicit relevance to energy management in buildings, leaving 306 publications;
• Stage 3—Content-based filtering: Works outside the technical scope of this review were excluded, such as purely economic market models, forecasting without EMS/building context or sustainability assessments without AI.

The rationale for the inclusion and exclusion criteria adopted is summarized in

Table 2. Inclusion and exclusion criteria applied in the literature screening.

Criterion	Included If…	Excluded If…
Source quality	Record indexed in WoS or Scopus, with complete metadata and DOI.	Record without DOI, missing authors, or incomplete metadata.
Topical scope	Explicit mention of energy management in buildings (including Heating, Ventilation, Air Condition (HVAC), lighting, microgrids, Energy Management Systems, Demand Side Response (EMS/DSR), building performance).	Focus exclusively on unrelated domains (e.g., mobility, large-scale grid operations).
AI relevance	AI techniques explicitly applied (Machine Learning ML, Deep Learning DL, Reinforcement Learning RL, etc.) to energy-related functions in buildings or local microgrids.	No AI component, or purely conceptual without technical application.
Application domain	EMS, DSM, DSR, predictive control, optimization, building energy performance, sustainability with AI.	Purely economic/market models (auctions, bidding, trading) without EMS/control aspects.
Forecasting role	Forecasting integration into EMS, DSM/DSR, or microgrid operation.	Standalone forecasting (photovoltaic PV, wind, price) without EMS/control context.
Sustainability assessment	AI applied to LCA/LCC in connexon with building energy management.	LCA/LCC without AI or without EMS/building application.

, while the numerical results of each selection stage are shown in

Table 3. Multi-stage selection and reduction in publications.

Stage	Set 1	Set 2	Set 3	Set 4	Total	% of Previous	% of Start
Initial identification (WoS + Scopus)	511	175	360	1055	2101	100%	100%
After merging (both databases, DOI, completeness)	173	48	100	293	614	29.2%	29.2%
After thematic filtering (EMS in buildings)	119	29	41	117	306	49.8%	14.6%
After content-based filtering (final set)	29	23	1	106	159	52.0%	7.6%

.

Following the filtration process, a total of 159 publications were retained for further analysis. This corpus forms the foundation for the ensuing analysis, which is structured in accordance with the IMRAD review format.

It should be noted that Set 3, which integrates AI with LCA and LCC in buildings, experienced the strongest reduction during the filtration process. Initially, 41 publications were retained after thematic selection (see Table 3), yet only one met the strict inclusion criteria. The majority of the studies addressed sustainability or life cycle analysis without explicit AI integration.

Despite this reduction, the LCA/LCC dimension continues to be of significant relevance to the objectives of this review. As discussed in the Introduction, long-term sustainability and life cycle performance are key pillars of future building energy systems, and there is a growing need for AI-driven methodologies to support predictive and data-informed decision-making. Consequently, Set 3 with all 41 positions was subjected to an additional qualitative assessment to capture ongoing developments at the intersection of AI, sustainability evaluation, and building energy performance.

This methodological extension serves two goals: (i) map the current state of research in LCA/LCC for buildings, even when AI is only implicit; and (ii) highlight research opportunities where AI can enhance traditional life cycle methods, enabling dynamic and integrated sustainability assessments. This treatment ensures that the review not only reflects established AI–energy management studies but also encompasses emerging sustainability-oriented research, reinforcing the originality and forward-looking scope of the contribution.

Building on this extended approach, all records that passed the previous filtering stages were next verified for full-text accessibility and detailed content evaluation. Among the 158 (159-1 from Set 3) publications from Sets 1, 2, and 4, 129 full texts were available through open-access sources or institutional library resources. For Set 3, representing the complementary sustainability-oriented track, 38 of 41 records were retrieved and analyzed within the qualitative assessment described above. Subsequently, a comprehensive full-text screening was conducted to evaluate the methodological depth, technical relevance, and integration of AI concepts. This process resulted in 78 studies from the core corpus and 19 from the LCA/LCC subset, yielding a total of 97 publications that constitute the final analytical foundation of this review, as summarized in

Table 4. Detailed full-text screening and final inclusion of reviewed publications.

Stage	Set 1 + 2 + 4 (TE, DEM, Sustainable Buildings)	Set 3	Total
Publications before full-text assessment	158	41	199
Full texts available	129	38	167
After detailed full-text screening and evaluation	78	19	97

.

The multi-stage literature selection procedure is described in the PRISMA flow diagram (Figure 1), which presents the complete process of identification, screening, and inclusion. The diagram summarizes the transition from the initial retrieval of 2101 records to the final inclusion of 97 studies, covering both the systematic filtering of the core datasets (Sets 1, 2, and 4) and the complementary qualitative reassessment applied to the sustainability-oriented Set 3 (LCA/LCC).

Although this review follows the IMRAD format rather than a fully systematic review, it incorporates key quantitative elements to ensure transparency and replicability. The figures presented in Table 1, Table 2, Table 3 and Table 4 and summarized in the PRISMA-style diagram provide a quantitative overview of the literature screening process, reducing 2101 initial records to 97 included studies. These data also enable a basic trend interpretation: approximately 40% of the final corpus relates to reinforcement learning and hybrid MPC approaches, about 15% to federated, edge-AI, or digital twin frameworks, and below 10% to LCA/LCC studies. These present proportions illustrate the methodological focus of recent research and confirm the dominant role of operational intelligence methods over long-term sustainability integration. Such quantitative observations are intended to complement, rather than replace, the thematic synthesis that defines the scope of this structured narrative review. Moreover, to enhance transparency, the complete search strings, time frame, and inclusion/exclusion criteria are detailed in this section, while Table 1, Table 2, Table 3 and Table 4 and the PRISMA diagram summarize the entire selection process. The complete list of selected records and decisions is archived and available upon request from the author, as also indicated in the Data Availability Statement.

3. Results

The reviewed corpus includes 97 publications from major publishers, with Elsevier accounting for about 40%, followed by IEEE and MDPI (each around 20%). The remainder come from Springer, Wiley, Taylor & Francis, Frontiers, SAGE, Oxford University Press, AIP, and selected conference proceedings such as IBPSA and ACM. This distribution reflects the dominance of specialized journals in energy and building research (e.g., Applied Energy, Energy and Buildings, Journal of Building Engineering, Energies, Sustainability) and the growing visibility of AI studies in interdisciplinary outlets.

The reviewed works cover a broad range of AI applications in building and microgrid energy management, with most focusing on operational prediction and control—mainly short-term forecasting of demand, price, and renewable output, and the real-time optimization of distributed resources. Several comprehensive reviews define the methodological baseline for this field, classifying ML, DL, and RL approaches and defining their evolution from conventional statistical models to learning-based techniques [3,35,36].

Two main thematic clusters emerge. The first concerns TE, addressing peer-to-peer and community trading, hierarchical markets, and agent-based bidding. Representative studies propose bilevel market designs for fairness, DRL agents for EV charging, and reduced-order models for load bidding, as well as coordination via deterministic policy gradients [37,38,39]. These studies form the background for Section 3.1.

The second DEM cluster focuses on the near-real-time coordination of distributed resources, storage, and flexible loads under uncertainty. This research often combines forecasting and optimization, applying ML for online control, edge-AI for localized prediction, and digital twins for scenario-based demand estimation [35,40,41]. These works underpin Section 3.2.

Across both mentioned domains, AI methods show increasing methodological diversity. Classical ML remains common for anomaly detection and short-term load prediction, DL is used for sequential data analysis, and multi-agent RL has become the key paradigm for decentralized decision-making. Recent reviews and methodological studies summarize this evolution and highlight both the potential and the current limitations of AI in energy management [42,43,44]. These insights provide the rationale for Section 3.3, which integrates the AI methods across TE and DEM.

3.1. Research on Transactive Energy

Research on the TE has advanced rapidly last few years, emphasizing decentralized coordination in smart local energy systems. The reviewed works address TE from conceptual market designs to device-level implementations and AI-based optimization.

3.1.1. Concepts and Market Designs

In the literature, TE is defined as a set of control and market mechanisms that enable value-based coordination among distributed actors such as prosumers, aggregators, and DSOs on a community or feeder scale. Three dominant design families are identified:

• hierarchical or bilevel coordination of communities and markets;
• peer-to-peer (P2P) and community energy sharing;
• agent-based transactional control integrated into local markets.

Gholizadeh et al. [38] propose a fair-optimal bilevel TE framework for microgrid communities that accounts for user discomfort, demand-response rebound, and voltage/current constraints—an early example of equity-aware market design. Amasyali et al. [45] develop a distributed game-theoretic transactional model in which the DSO iteratively adjusts price vectors while aggregators respond with demand, achieving convergence that preserves privacy. Agent-based approaches extend this concept to P2P and community trading. Studies employ RL modified diagonalization to compare billing and mid-market mechanisms under different pricing regimes [46]. Yu et al. [47] simulate a residential community using a TE bidding scheme with model predictive control via mixed-integer linear programming (MPC/MILP), quantifying demand, import, and cost reductions while revealing optimal bidding and device-control strategies.

In general, TE research has evolved from conceptual market models toward applied community-scale demonstrations, laying the groundwork for practical implementation.

3.1.2. Implementations in Microgrids and Local Energy Systems

The operationalization of TE occurs through domain-specific controllers for flexible loads and distributed energy resources. Liu et al. [37] automated transactive HVAC control using RL within the Pacific Northwest National Laboratory’s Transactive Energy Simulation Platform (TESP) developed by Pacific Northwest National Laboratory, addressing limitations of continuous-state control and system heterogeneity in Q-learning. Sharma et al. [39] developed an EV bidding agent based on recurrent Proximal Policy Optimization (PPO) under a partially observable Markov decision process (POMDP), showing policy convergence on real price data and encoding user goals in bidding strategies.

At the market–physics interface, reduced-order models of aggregate loads, such as thermostatic populations, enable co-simulation with power-system solvers, linking market dynamics of TE to feeder constraints [48]. These studies reflect a shift from high-level market coordination toward device-level integration and grid-aware operation.

3.1.3. AI Methods for TE Coordination and Trading

According to the relevant technical literature in the field of TE, AI serves two complementary roles:

• policy learning for market agents;
• prediction and estimation supporting market clearing and control.

Multi-agent RL dominates policy learning for market interactions under uncertainty, including transactive EV bidding with PPO [39] and hybrid learning for multi-microgrid energy sharing among prosumer buildings [49]. Federated and distributed learning approaches address privacy-preserving coordination, aligning local learning with consensus-based optimization for scalable TE systems [50].

Recent reviews on energy storage-powered microgrid energy management [3] position TE as part of a wider methodological landscape that includes game theory, agent-based control, and robust optimization. AI thus emerges not as a supporting tool but as a core enabler of scalable and adaptive TE system design.

3.1.4. Evaluation Criteria: Welfare, Fairness, and Grid Constraints

Evaluation of TE performance increasingly extends beyond economic efficiency toward fairness, comfort, and technical feasibility under grid constraints. Gholizadeh et al. [38] demonstrate this shift with a bilevel community TE model that minimizes both energy costs and user dissatisfaction while maintaining feeder limits through fair and scheduled export restrictions. Zhou et al. [46] introduce replicable indices to benchmark P2P trading models across pricing environments. These studies illustrate a transition from purely cost-driven optimization to socio-technical evaluation, integrating welfare, equity, and reliability as key criteria.

3.1.5. Robustness, Security, and Resilience

A parallel research direction focuses on the robustness and cybersecurity of TE systems and their supporting platforms. Studies emphasize the need to detect anomalies and adversarial behaviors in transactional infrastructures, complementing market design and agent learning [51]. Furthermore, Liu et al. [37] highlight how RL-based HVAC agents enhance scalability and stability by managing continuous control and heterogeneous device behavior, overcoming Q-learning limitations. Their TESP-based experiments stress the importance of simplified state representations and tailored reward structures to maintain convergence as the number of agents grows.

Reduced-order models of aggregated bid loads not only enable integration with power-system solvers, but also enhance stability and resilience under cyber-physical uncertainties [48]. Collectively, these developments indicate that integrity, security, and robustness are becoming embedded components of TE research, even as most studies continue to focus on short-term operation.

3.1.6. Identified Gaps and Future Directions

Despite significant progress in short-horizon market clearing, agent learning, and community-scale trials, most TE research still prioritizes short-term operational optimization under fixed tariffs and asset portfolios. Long-term dimensions—such as asset degradation, investment strategies, seasonal reliability, and sustainability of the life cycle—remain largely underexplored. Yu et al. [47] note that environmental goals, including reducing greenhouse gases, are often treated as secondary objectives. Mutluri and Saxena [1] further identify the absence of strategic planning and long-term resilience as a structural gap, observing that blockchain and AI improve adaptability and security but not investment or sustainability planning.

In general, the literature reveals a short-term bias in TE studies, with limited integration of long-term economic, environmental, and infrastructure considerations—issues that remain central to future research agendas.

The main technical aspects and challenges discussed in this subsection are summarized in

Table 5. Technical synthesis of TE research. Key dimensions, methods and system-level challenges.

Focus Area	Main Technical Approach (with Performance Note)	Key Performance Aspect	Identified Limitation /Failure Point
Market Design and Coordination	Bilevel and agent-based market models with P2P trading; efficient in static and small-community setups, but sensitive to data latency	Economic welfare, fairness, feeder constraint compliance	Lack of interoperability and unified APIs between DSO–aggregator layers
Control and Implementation	RL/PPO agents for HVAC and EV control; high short-term efficiency (>10–15% cost reduction) but degraded convergence with >100 agents	Real-time control, local balancing	Scalability bottlenecks and unstable learning under device heterogeneity
AI for Market Learning	Multi-agent RL and federated learning; fast convergence in controlled settings but heavy communication overhead	Policy adaptation, privacy preservation	Low generalization under dynamic price and demand conditions
Evaluation and Robustness	Welfare/fairness indices and reduced-order co-simulation; effective for benchmark testing (within ±5% reproducibility)	Equity, resilience, and grid compliance	Limited validation under cyber-physical disturbances
Cross-cutting Constraints	Edge-AI and co-simulation (TESP); moderate computational efficiency, high modularity	Interoperability, scalability, cybersecurity	Fragmented architectures, DRL instability, weak anomaly detection latency control

, which consolidates the key findings on TE systems from a performance and implementation perspective.

3.2. Research on Dynamic Energy Management

While TE mechanisms (discussed in Section 3.1) focus on market-based coordination and value exchange, DEM addresses operational intelligence—how distributed and flexible resources are sensed, predicted, and controlled in real time. It centers on system adaptation to changing conditions rather than market interaction. In the reviewed literature, DEM emerges as the second key pillar of local energy system research, covering microgrid-scale EMS, building and home EMS, and hybrid forecasting–control frameworks.

3.2.1. Scope and Reference DEM Architectures

Research in DEM highlights reference architectures that link fast device-level actuation with supervisory scheduling and learning. Early frameworks, such as DEMs [12], applied adaptive dynamic programming for continuous microgrid optimization under uncertainty. Shakir and Biletskiy [52] proposed a home-oriented EMS that integrates sensing, forecasting, and DER scheduling under comfort constraints.

A large body of work now situates DEM within the convergence of AI, IoT, and edge computing as its technological foundation [53]. System-level analyses trace the evolution of microgrids from AC and DC configurations to hybrid and multi-energy systems centered on storage [3]. In this context, Mutluri and Saxena [1] discuss networked microgrids (NMGs) as scalable and resilient configurations with hierarchical primary–secondary–tertiary control. Complementary studies explore multi-agent EMS, where distributed agents coordinate DERs, storage, and flexible loads—applied not only to buildings but also to specialized sectors such as greenhouse energy management [54].

3.2.2. DSM/DSR and Flexible Asset Coordination with RL

A substantial body of DEM research focuses on DSM/DSR, emphasizing the real-time coordination of flexible assets through reinforcement learning RL. Iqbal and Mehran [55] showed that model-free RL can minimize microgrid operating costs under uncertainty of renewable and storage. Dridi et al. [56] compared classical Q-learning with deep recurrent agents, confirming the latter’s superiority under partial observability and typical temporal correlations of EMS operation. Darshi et al. [57] developed a decentralized EMS with multiple RL controllers operating in asset clusters, reducing the communication load while maintaining near-optimal control. Furthermore, Arwa and Folly [35] synthesized this progress, mapping key RL families—Q-learning, deep deterministic policy gradient (DDPG), PPO, and hierarchical RL—to DEM tasks and highlighting the need for safe and constrained RL formulations. Collectively, these studies show a clear evolution from tabular RL to deep and distributed agents, positioning RL-driven DSM/DSR as the operational core of modern DEM.

3.2.3. Forecast-Informed Control Loops

To enhance control robustness, forecasting is increasingly embedded within DEM loops. Lv et al. [40] implemented edge-based recurrent neural network (RNN) forecasters for short-term load and power prediction, reducing latency and cloud dependence. Bayer et al. [41] used digital twin simulations to test DEM strategies under high penetration of EV, while Sadrian Zadeh et al. [58] applied supervised learning for state estimation based on IoT, improving system observability and control reliability. At the distribution edge, Peiris et al. [59] used ML-based profiling to separate PV and EV load patterns, supporting flexibility allocation. These approaches indicate a shift from isolated control to integrated predict–decide–act architectures, where forecasting and control form a single adaptive feedback loop.

3.2.4. Control Strategies Beyond RL: MPC, Hybrid and Physics-Informed Tracks

Alongside RL, MPC remains a key method in DEM, especially in buildings and community-scale applications. Chen et al. [60] proposed a robust data-driven MPC that constructs uncertainty sets for weather and occupancy using clustering and density estimation, achieving adaptive comfort–cost trade-offs between HVAC, geothermal, PV, and storage systems. The practical efficiency of MPC depends on model calibration, co-simulation interoperability, and lifecycle maintenance, yet lack of standardization still limits its wider adoption [61]. Moreover, Aruta et al. [62] showed that ANN-assisted MPC reduces computation time while preserving comfort by using nonlinear autoregressive surrogates, achieving measurable energy savings compared to fixed setpoints.

At the methodological frontier, physics-informed ML (PIML) integrates domain knowledge into learning-based control. Ma et al. [63] identified four PIML pathways—inputs, loss functions, architectures, and ensembles—demonstrating improved interpretability of building models. Consequently, Qi et al. [64] combined ANN forecasting with metaheuristic optimization and event-triggered MPC, cutting computation frequency by up to 80% while maintaining comfort. Related studies validate this trend: ANN-enhanced MPC improves HVAC performance [65], adaptive setpoint MPC stabilizes multi-zone buildings [66], and comfort-aware MPCs embed PMV-based cost functions [67]. Event-triggered formulations further reduce computational load while preserving stability [68].

Long-term evaluations confirm continuous progress toward hybrid and scalable MPC. Reviews by Renganayagalu [61] and Aruta et al. [62] document a decade of field pilots, identifying persistent issues such as model complexity, integration cost, and non-standard BMS interfaces. District-level studies report 15–28% heating energy savings using hybrid gray-box and ANN/RNN MPC models [69]. In multi-carrier microgrids, MPC coordinates hydrogen, thermal, and battery storage to balance physical and market objectives [70]. Recent analyses highlight the convergence of MPC with model-based control (MBC), occupant-centric, and anomaly aware controls, as well as its integration into digital twins [71]. Collectively, these developments signal a transition from standalone MPC to hybrid and physics-informed pipelines that balance adaptability, safety, and computational efficiency. MPC thus emerges not as an alternative, but as a complementary pillar to RL in advanced DEM architectures.

The key technical approaches, performance outcomes, and remaining challenges discussed in this subsection are summarized in

Table 6. Technical synthesis of DEM research. Architectures, methods and operational challenges.

Focus Area	Main Technical Approach (with Performance Note)	Key Performance Aspect	Identified Limitation /Failure Point
Reference Architectures	Hierarchical EMS and multi-agent coordination; reliable on the microgrid scale, latency <1 s in supervisory layers	Real-time sensing, coordination, adaptability	Limited standardization of communication protocols; fragmented BMS/EMS interoperability
DSM/DSR with RL	Model-free, deep, and distributed RL controllers; achieve 10–20% cost reduction and stable convergence under moderate agent counts	Demand flexibility, cost efficiency	Scalability loss beyond ~100 agents; slow convergence under partial observability
Forecast-Informed Control	Edge-based RNN/LSTM forecasting and DT stress testing; latency reduced by 30–50% vs. cloud setups	Prediction accuracy, responsiveness	Dependence on high-quality data; limited fault tolerance at edge nodes
MPC and Hybrid Control	Robust MPC and ANN-assisted MPC; reaches 10–25% energy savings and improved comfort metrics	Comfort–cost balance, computational efficiency	High modeling complexity; lack of standardized calibration and co-simulation workflows
Physics-Informed and Metaheuristic Tracks	PIML and event-triggered MPC; 60–80% reduction in computation frequency, stable comfort levels	Interpretability, computational scalability	Limited validation in real deployments; integration cost with existing EMS
Crosscutting Challenges	Hybrid RL–MPC frameworks with edge AI and DT integration	Adaptability, safety, and lifecycle reliability	Interoperability gaps, high integration expenditures, weak cyber-resilience of decentralized controllers

, providing a concise overview of current advances in DEM research.

3.3. AI Methods and Techniques Applied in TE and DEM

This section outlines the AI methodologies applied in TE and DEM research, focusing on algorithmic families rather than functional mechanisms. It covers classical ML, deep learning, reinforcement and multi-agent learning, as well as emerging federated, edge, and DT frameworks. The methods are organized by functional layers, emphasizing their operational roles and development trends.

3.3.1. Perception and Prediction Layer: From Classical ML to Edge-AI and DT

In TE and DEM applications, short-term load, price forecasting, and state estimation dominate the perception layer. Classical ML algorithms such as Random Forest (RF), Support Vector Machine (SVM), and Extreme Learning Machine (ELM) remain widely used, while deep models such as LSTM, GRU, and hybrid CNN–RNN architectures deliver improved temporal accuracy. Modeling of prosumer behavior (PV/EV patterns) and occupancy detection further extend this layer [40,41,59,72]. DTs support scenario-based planning and operational optimization [41], while edge-AI reduces latency and communication overhead in local forecasting [40]. In buildings, vision-based occupancy detection and drone-assisted thermal imaging are increasingly used for control calibration and audit purposes [72,73]. A broader shift is visible—from model-centric approaches to data- and deployment-centric paradigms. This includes edge learning for rapid local inference [40], federated forecasting without centralized data aggregation [50], and integration of DT + AI for predictive control and explainable analytics [61,74,75].

In general, perception- and prediction-oriented research prioritizes short-term accuracy and local responsiveness, with a growing emphasis on federated and edge deployment. However, integration with LCA and long-range decision processes remains limited, marking a persistent gap between precise forecasting and strategic management.

3.3.2. Decision-Making and Control Layer (DEM—Oriented)

In DEM applications, RL has become the dominant control paradigm in EMS and HEMS systems, evolving from tabular Q-learning to advanced actor–critic and policy optimization algorithms. Arwa and Folly [35] trace this evolution from Q-learning to PPO and TRPO, focusing on transfer learning and prioritized re-enhanced experience.

Other studies apply DDPG and Soft Actor–Critic (SAC) for flexible resource aggregation [76], compare Deep Q-Network (DQN) and RNN/LSTM in EMS control [56], and employ deep RL in IoT-based microgrids [77]. Increasingly, RL is coupled with MPC, MILP, or metaheuristics to accelerate optimization and mitigate the “curse of dimensionality” [2]. Research thus advances from discrete-action Q-learning toward continuous, hybrid actor–critic control adapted to uncertain environments.

Despite progress, scalability, sample efficiency, and safe deployment remain major obstacles—especially in large heterogeneous microgrids. Hybrid RL–MPC strategies appear to be most promising in balancing adaptability and operational stability.

3.3.3. Market-Level Coordination Layer (TE—Oriented)

In TE systems, the main challenge lies in coordinating multiple agents under market rules and communication constraints. Advances include Bayesian multi-agent RL (MARL) resistant to communication failures [78], game-theoretic hierarchical control [45], and bilevel optimization integrating fairness in community microgrids [38]. Reduced-order models support scalable bidding and constraint handling [48]. The applications span EV agents [39] and HVAC transactional control [37], illustrating growing cross-domain use. Emerging studies integrate MARL with PPO/TRPO and transfer learning, supported by federated optimization and edge-AI for forecast-to-decision coupling.

Although these frameworks improve fairness, coordination, and efficiency, their robustness under imperfect communication and real-world pilot validation remains limited, particularly in community-scale and CES contexts.

3.3.4. Distributed and Secure AI Frameworks: From FL to Hybrid Optimization and DT

Recent advances show a strong convergence of AI techniques towards distributed and secure frameworks combining federated and edge learning, hybrid optimization, and DT analytics. These approaches address the cybersecurity, scalability, and privacy challenges identified in TE and DEM, embedding AI as a system-level capability rather than a standalone predictor or controller.

Federated learning (FL) preserves the locality of the data while enabling collaborative model training among distributed actors. Adaptive and clustering-enhanced FL improves multi-horizon forecasting and anomaly detection in community buildings [79]. Privacy-preserving IoT–blockchain systems integrate distributed ML with integrity control for P2P trading [80], while edge-AI reduces inference latency and communication load in microgrid EMS [40,53] and building-scale BMS integrated with DT and blockchain [81,82].

Cybersecurity also emerges as a critical frontier. AI-based intrusion detection systems (IDS) are increasingly applied to mitigate false-data injection and denial-of-service attacks [83,84]. Autoencoder-based anomaly detection enables lightweight, unsupervised edge monitoring [85], and virtualized training platforms strengthen operator resilience [86,87]. The “FMEA 2.0” methodology extends the use of ML to the evaluation of risk in smart microgrids, while federated sensing and cyber-range testbeds define the foundation of AI-centric security [88].

AI also accelerates optimization, embedding learning surrogates in classical methods. Neural networks and Gaussian processes serve as proxies for network and storage models, enabling faster co-optimization of ancillary services [3]. Hybrid combinations—RL + MPC/MILP, deep RL for resource aggregation [76], and metaheuristic-assisted EMS—enhance responsiveness [2,89]. At the building and microgrid scales, hybrid DL–metaheuristic frameworks (e.g., bi-directional LSTM/capsule network—CapsNet with hybrid gazelle and seagull optimization algorithm—HGSOA) improve forecasting and flexible-load scheduling [52,90].

Finally, DT-driven AI integrates BIM, IoT telemetry, simulation, Bayesian calibration, and XAI for real-time benchmarking and predictive control [61,73,74]. At urban and portfolio scales, Distributed Ledger Technology (DLT) and blockchain ensure traceable cross-lifecycle dataflows [81,90]. Combined with PIML, they embed physical constraints in learning, improving extrapolation and safety in critical infrastructures [63].

Collectively, these developments mark the emergence of distributed holistic AI ecosystems that unify learning, optimization, and cybersecurity. AI is no longer applied as an isolated forecasting or control tool, but as an embedded and trustworthy intelligence layer that enhances efficiency, privacy, resilience, and transparency in the TE and DEM domains. The cross-cutting

Table 7. Cross-cutting overview of AI methods and applications in TE, DEM, buildings, microgrids, energy communities, and CES.

AI Method	TE (Trading, Markets)	DEM (DSM/DSR, Control)	Buildings (BEMS/HEMS)	Microgrids	Energy Communities	Closed Ecological Systems
Classical ML (SVM, RF, K-Nearest Neighbors—kNN, Extreme Learning Machine—ELM)	Price and demand forecasting; bidding profiles [48,91,92]	Load prediction, Non-Intrusive Load Monitoring (NILM), anomaly detection [59,93,94,95,96]	Energy Use Intensity (EUI) benchmarking, HVAC classification [47,92,94,96]	DER pattern recognition and energy quantification methods [59,92]	Local demand/supply modeling [49,52]	Resource forecasting and pattern recognition for life-support loops [1,54]
Deep Learning (LSTM, GRU, CNN, Regional CNN, Autoencoders)	Short-term price signals, prosumer response [97,98,99]	HVAC load prediction, IAQ/IEQ modeling [40,41,60,75,90]	Occupancy detection, CO2 prediction, drone thermal imagery [60,72,75,79]	Recurrent EMS controllers [40,56]	Net demand forecast, VPP integration [47,50,79]	Prediction of environmental variables (temperature, humidity, CO2) in greenhouses and space habitats [54,71]
Reinforcement Learning (Q, DQN, A2C/A3C, PPO, DDPG, SAC)	Transactive bidding, EV scheduling [39,45,76]	LowEx control, EMS with storage [2,55,77,100]	Smart HVAC dynamic control, comfort-aware policies [60]	Adaptive EMS, ancillary services [35,76,77,89,101]	MARL for distributed DR and pricing coordination [45,50,78]	Adaptive control of life-support subsystems, water/air recycling optimization [54,102]
Multi-Agent AI and Game Theory	Cooperative/competitive market negotiation, fairness [38,45,78,80]	Hierarchical DSM/DSR coordination [57,103]	Occupant centric decision and behavior prediction [104]	Coordination in networked microgrids, resilience enhancement [1,3,103]	Federated MARL for transactive energy communities, blockchain-based TE [50,79,80]	Multi-agent control of food–energy–water loops in habitats [54,105]
Hybrid AI (MPC+ML, Metaheuristics+ML, Surrogates)	Surrogate MILP/MINLP for bidding optimization [3,101,106]	RL+MPC and RL+MILP for EMS responsiveness [2,76,101]	HEMS scheduling with LSTM+GA; DSM via BLSTM/CapsNet+HGSOA [52,90]	ANN/GP surrogates for ESS and multi-energy scheduling [3,57,101]	Consensus + FL for distributed optimization [50,57]	MPC+RL for CES climate /energy management [71,73]
Federated and Edge AI	FL-assisted distributed trading and coordination [50,57,99]	Edge-AI for real-time EMS [40,53]	Adaptive FL for building forecasting; privacy-by-design automation [79,82]	Edge-enabled EMS with IoT integration [50,53]	FL-assisted aggregation and consensus building [50,57,79]	Edge/federated AI to preserve privacy and autonomy in CES habitats [73,81]
Digital Twins and Blockchain Integration	DT-enabled TE forecasting, auditing, blockchain-secured trades [80,81]	DT+AI for EMS and predictive resilience [2,61,74]	BIM/IoT+DT for performance gap reduction [61,73,74]	DT/Blockchain/Building Management System (BMS) for microgrids [81,107]	DT frameworks for resilience in NMGs and TECs [1,57]	DTs of bioregenerative CES habitats [63,71,73]
Physics-Informed and Interpretable ML (PIML, Explainable AI—XAI)	Trust metrics, explainable bidding and optimization [63]	PIML for EMS stability and reliability [63]	Bayesian calibration, explainability in building energy management DTs [61,74]	XAI-based anomaly detection and IDS in MGs [2,85]	Explainability in federated trading optimization [9,74]	PIML/XAI for lifecycle resilience in CES habitats [63,82]
Cybersecurity and Risk Aware AI	Blockchain-secured TE markets and EV transactive flows [2,39,80]	IDS with ML frameworks for EMS; homomorphic encryption for anomaly detection [2,51,83,84,85]	Risk-aware ML in building automation [82,88]	FMEA 2.0 for MG risk assessment; operator cyber-range training [86,87,88]	Cybersecurity in federated TE and IoT environments [50,84,86,87]	Predictive anomaly detection and ML-based maintenance in CES loops [63,81]

provides additional contextual information on these methods, illustrating their applications in buildings, microgrids, energy communities, and CES.

It should be noted that the inclusion of CES in Table 7 extends beyond the dominant scope of the existing literature. CES represents an emerging field where energy management, life-support, and resource recycling are tightly integrated. Despite the limited academic focus on AI for TE and DEM in this context, related domains—such as space habitat engineering, controlled environment agriculture, and bioregenerative life-support systems—are advancing rapidly through industrial and space initiatives. The CES-oriented applications and trajectories outlined in the table are therefore authorial extrapolations, derived from analytical insights and AI-assisted SCOPUS mapping. This frame of reference CES as a promising frontier where methods validated in buildings, microgrids, and energy communities can be adapted to self-sustaining environments.

Although the reviewed studies vary in scope and scale, several quantitative indicators can be identified in AI-based energy management applications.

Table 8. Quantitative comparison of key AI-based energy management approaches.

Methods	Reported Energy Savings [%]	CO2 Reduction [%]	Cost Savings [%]	Control Horizon	Key References
Model Predictive Control (MPC)	10–25	8–20	12–22	Medium-term: 15–60 min predictive window	[51,57,63,66]
Reinforcement Learning (RL)	15–35	10–28	20–30	Short-term: 1–15 min adaptive control steps	[71,74,78,84]
Federated Learning (FL-RL)	12–27	10–25	18–29	Adaptive: 5–20 min (local) with periodic global update 1–4 h	[90,91,92]
Hybrid AI (RL + MPC/DL)	20–40	15–30	25–35	Variable horizon: local 1–10 min with global 30–120 min layer	[94,96,97,98]

summarizes the typical performance ranges reported in the literature for selected control strategies, including RL, MPC, and hybrid AI approaches. The metrics reported encompass a range of factors, including energy cost reduction, comfort deviation, emission savings, and computational requirements.

In the reviewed publications, MPC remains the most widely validated control strategy, typically achieving 10–25% energy cost savings under real-time optimization constraints. RL-based methods demonstrate greater potential (up to 35%) but exhibit higher computational and convergence variability. Federated and hybrid models introduce scalability and privacy advantages, though their quantitative validation is still limited to small-scale testbeds. In the author’s opinion, while quantitative benchmarking across AI-based EMS studies remains difficult due to heterogeneous objectives and datasets, the collected evidence indicates comparable or higher efficiency of learning-based approaches (particularly RL and hybrid models) relative to classical MPC.

3.4. Complementary AI and Life Cycle Perspectives for Sustainable Buildings

This subsection extends the review by integrating the LCA, LCC, and long-term sustainability perspectives into the discussion of AI for building energy systems. As outlined in Section 2, this topic is strategically relevant since life cycle performance is a key dimension of future energy communities and transactive microgrids. The rationale for treating it separately is twofold. First, it maps the state of the art in applying LCA/LCC to building and community-scale energy management, even where explicit AI integration is limited. Second, it highlights the gap where AI can complement traditional life cycle tools through dynamic, predictive and data-driven assessments that go beyond static evaluation. This lens introduces a long-horizon systemic context that complements the operational focus of earlier sections. Although prior subsections emphasized real-time optimization, the perspective presented here shifts toward strategic decision frameworks, where AI and life cycle methodologies jointly support sustainability in buildings and local energy systems.

3.4.1. AI for Dynamic and Predictive LCA/LCC in Building Energy Systems

The transition from static to predictive LCA was first identified by Zheng and Yan [26], who noted limited integration between LCA and digital/AI workflows. Sharif and Hammad [108] addressed this gap using ANN-based surrogates to approximate the renovation of LCA/LCC, showing their potential for large-scale decision support. Amini Toosi et al. [109] embedded ML in Life Cycle Sustainability Assessment (LCSA) pipelines to capture trade-offs dynamically, while related surrogate models for hybrid HVAC/PV systems balanced operational flexibility with embodied costs [110].

At the community level, Elomari et al. [111] combined ML with multi-objective optimization (MOO) and multi-criteria decision-making (MCDM) to integrate LCA/LCC into renewable energy governance. Abokersh et al. [112] improved this approach through ANN-based optimization of solar district heating. Collectively, these works illustrate how AI-enhanced LCA/LCC provides a predictive layer for transactive energy exchanges, linking short-term operation with long-term sustainability. Additional research on embodied and operational emissions in HVAC design [113] and optimization in extreme climates further evidences the growing fusion of life cycle and performance modeling [114].

3.4.2. Retrofit and Building-Integrated PV: AI-Enabled Life Cycle Optimization

Retrofit and Building-Integrated Photovoltaics (BIPV) are domains in which AI and life cycle methods converge at the building-to-grid interface. Sharif et al. [115] pioneered the use of deep generative learning with variational autoencoders to generate retrofit scenarios for evaluation of LCA/LCC, expanding the design options for building owners. Imalka et al. [116] applied ANN surrogates for BIPV optimization, treating the life cycle cost as an explicit objective alongside the energy yield. Li et al. [117] developed an autonomous BIPV deployment framework that combines 3D scanning, solar analysis, and LCC evaluation.

These methods reduce environmental and economic burdens while linking building-level assets with local microgrids. AI ensures that retrofit and PV investment decisions align with transactive energy strategies, establishing the technical and economic foundation for the participation of prosumers. Further studies on structural optimization with evolutionary algorithms and ensemble learning [28,118] confirm the growing fusion of AI and life cycle thinking in retrofit design.

3.4.3. Community and District Energy Systems: Life Cycle Anchors for AI-Driven Microgrids

At the community scale, AI-enhanced life cycle methods support the evolution of renewable energy communities and district heating networks. Elomari et al. [111] proposed a MCDM–ML framework integrating LCA for governance and design optimization in renewable energy communities. Abokersh et al. [112] coupled ANN with sensitivity analysis to manage uncertainty in solar district heating, maintaining life cycle consistency under variable operation. Their follow-up study [119] embedded ANN–MCDM in near-zero and passive building communities, showing that high renewable shares can remain economically feasible over the life cycle.

In general, AI methods evolve from predictive assessment to decision support for market- and community-level structures. The literature consistently indicates that transactive energy performance is inseparable from life cycle outcomes, as long-term costs and impacts determine both the stability and credibility of local energy exchanges. Broader comparative reviews of sustainability systems [120] and policy-oriented studies on AI for net-zero transitions [31] reinforce the role of life cycle approaches as decision anchors for future energy communities.

3.4.4. Data, Interoperability, and Governance Challenges

For AI-enabled LCA to effectively support transactive energy and microgrid systems, robust and secure data infrastructures are essential. Potrč Obreht et al. [121] examined the integration of BIM and LCA in construction workflows, showing that without standardization and user awareness, automated life cycle processes remain fragmented. At the operational edge, Sun et al. [122] proposed ML-based intrusion detection rules to bridge the gap between trained models and real network conditions, emphasizing the need to protect life cycle data as a valuable market asset.

Despite growing adoption, interoperability and cybersecurity gaps still limit the reliability of AI-driven LCA to support local transactive exchanges. These studies highlight the need to integrate technical progress with secure governance frameworks to ensure trustworthiness in life cycle integration in smart microgrids. Complementary work on the assessment of green buildings using neural networks [123] and AI for sustainable buildings [31] further underlines the role of interoperability and structured knowledge in advancing life cycle–based decision-making.

The main AI approaches, integration mechanisms, and remaining technical challenges discussed in this subsection are summarized in

Table 9. Technical synthesis of AI–Life Cycle integration in sustainable building and energy community research.

Focus Area	Main AI Approach (with Performance Note)	Life cycle Integration Aspect	Key Limitation /Technical Challenge
Dynamic and Predictive LCA/LCC	ANN and ML surrogates for renovation and hybrid HVAC/PV systems; 10–20% improvement in accuracy of cost–impact estimation	Coupling of operational data with long-term embodied impacts	Limited interoperability between LCA tools and AI pipelines; scarce dynamic datasets
Retrofit and BIPV Optimization	Generative DL and ANN-based design models; multi-objective optimization with LCC functions	Integration of retrofit/BIPV decisions with transactive microgrid participation	High computational cost of generative modeling; dependence on reliable 3D and solar data
Community and District Energy Systems	ML–MCDM and ANN–sensitivity analysis; stable convergence under variable operating conditions	Linking LCA/LCC with renewable community governance and market structures	Incomplete coupling between life cycle objectives and short-term trading mechanisms
Data Infrastructure and Governance	BIM–LCA automation and ML-based IDS; low-latency detection and model update cycles	Secure management of life cycle data and model traceability	Lack of standardized workflows; weak cybersecurity and interoperability across platforms
Crosscutting Trends	Hybrid AI–LCA frameworks using DT, DLT, and PIML; moderate scalability and improved interpretability	Holistic sustainability integration across buildings, microgrids, and CES	Fragmented governance; need for unified data ontologies and validation protocols

, providing a concise overview of how life cycle methodologies intersect with AI in sustainable building and energy community research.

4. Discussion

In this section, the author interprets the results presented in Section 3 situating them within a broader context of conceptual and methodological debates. The discussion emphasizes the convergence of AI methods in different domains, the identification of the main research gaps, and the potential of these insights to facilitate the development of integrated frameworks for sustainable energy systems.

4.1. Integrative View of AI in TE, DEM, and Life Cycle Perspectives

The existing literature, as reviewed in Section 3, demonstrates that AI contributes to smart local energy systems at multiple, interdependent levels. Rather than being limited to isolated functions, AI applications can be grouped into a layered structure that links perception and prediction, operational control, market coordination, and long-term sustainability. This perspective is summarized in Figure 2, which illustrates the main layers and the information flows that connect them.

As illustrated in Figure 2, the lower layers provide the data and predictions essential for higher-level decision-making. Successive layers then translate this information into control actions, market interactions, and long-term planning. It is important to note that the framework is not purely hierarchical; feedback loops are evident, for example, when sustainability objectives impose constraints on operational strategies, or when market signals influence the scope and accuracy of prediction models. The bidirectional link between the control and market layers is of particular relevance, as it indicates that operational flexibility and trading mechanisms must evolve in parallel. This is an area where existing studies remain fragmented. In a similar manner, the downward arrows from sustainability to market and control emphasize the challenge of embedding long-term objectives, such as resilience or life cycle costs, into short-term optimization. These interactions indicate the presence of research gaps and establish the foundation for the subsequent discussion in Section 4.2, where methodological challenges and structural limitations are examined in greater detail. In this manner, the layered perspective emphasizes the significance of integration: The potential of AI in energy systems is not achieved through the utilization of individual algorithms; rather, it is realized through the coordination of these algorithms across different functional horizons.

4.2. Conceptual Gaps and Methodological Challenges

Although Section 3.2–3.4 documented substantial methodological progress in transactive energy, dynamic energy management, and AI-enabled life cycle integration, a closer synthesis reveals recurring blind spots and unresolved challenges. These limitations are not limited to a single research stream but emerge in multiple thematic areas, suggesting structural constraints in the current research landscape. To provide a structured overview, the main gaps and challenges identified in the reviewed literature are summarized in

Table 10. Research gaps and methodological challenges in the integration of TE, DEM, and LCA/LCC based on AI.

Area	Observed Focus in the Literature	Identified Gap/Challenge	Future Direction
Transactive Energy (TE)	Short-term market clearing (minutes–day-ahead), MARL-based bidding, bilevel fairness models	Weak coupling with grid reliability, seasonal variability, and long-term investment decisions; resilience under cyber-physical uncertainty underexplored [38,39,51]	Extend TE frameworks with multi-horizon optimization, AI-enhanced resilience metrics, and integration of environmental objectives
Dynamic Energy Management (DEM)	RL-based demand response, hybrid MPC for HVAC and microgrids, edge-AI forecasts	Scalability and sample efficiency of RL not solved; safe deployment in heterogeneous real-world systems largely missing; interoperability with legacy BMS limited [55,56,57,60,61]	Development of standardized DEM platforms combining robustness RL/MPC hybrids with edge computing and safe RL formulations
AI Methodologies	Strong innovation in RL, DL, federated/edge AI, emerging DT applications	Fragmentation across methods; limited explainability and trust; lack of integration into layered, interoperable frameworks [35,40,61,63,74]	Move towards multi-layered AI architectures that integrate perception, control, market, and sustainability with explainability by-design
Life cycle Integration (LCA/LCC)	Surrogate models for retrofit/BIPV, conceptual links to community energy	Lack of dynamic, predictive LCA coupled to EMS; minimal integration with operational control; uncertainty treatment and data standardization weak [26,108,109,110,111,113,117]	Embed predictive LCA/LCC in EMS workflows; couple AI-based control with embodied/operational impact models; improve interoperability of data and signals
Cross-domain (Buildings → Microgrids → CES)	Building EMS well studied; microgrids emerging; CES nearly absent	Limited research on transferability across scales and domains; no holistic studies linking building-level AI with CES-like survival-critical contexts [1,54,71,73]	Use CES as a frontier testbed to stress-test AI for resilience, closed-loop resource management, and long-horizon sustainability
Cybersecurity and Privacy	Early works on federated learning, blockchain, IDS for microgrids	Limited robustness against adversarial attacks; weak integration of cybersecurity into control loops; privacy preserved mainly in lab-scale pilots [51,81,82,83,84]	Advance privacy by-design AI in EMS/TE; validate adversarial robustness in pilots; integrate AI-based intrusion detection with control frameworks

.

The synthesis presented in Table 10 indicates a persistent focus on short-term optimization tasks within studies, with long-term performance and sustainability objectives receiving limited consideration. In the field of transactive energy, research has historically placed significant emphasis on market-clearing efficiency and agent-based bidding. However, the degree to which reliability at the distribution-level and environmental criteria are integrated remains limited. In the context of dynamic energy management, the efficacy of RL and hybrid MPC has been demonstrated in simulation studies. However, concerns regarding scalability, safe deployment, and interoperability with existing building management systems persist. Advances in methodology are similarly dispersed; a variety of AI methodologies—including deep learning, federated learning, and digital twins—are utilized in isolation, without the presence of a unifying framework that would facilitate the comparison of these methodologies across a range of applications. Life cycle integration is a particularly underrepresented field, with only a few studies to date that attempt to integrate predictive LCA/LCC into EMS workflows. It is widely acknowledged that cybersecurity and privacy-preserving mechanisms are of paramount importance. However, their implementation remains limited to laboratory-scale demonstrations, with inadequate validation under real-world conditions. These observations indicate the need for research that moves beyond isolated algorithmic innovations towards systemic approaches that integrate operational intelligence, market coordination, sustainability, and resilience.

4.3. Cross-Domain Insights: From Buildings to Microgrids to CES

The discussion of research gaps in Section 4.2 emphasized that many challenges—such as limited scalability, weak integration of sustainability objectives, and fragmented methodological approaches—are not confined to a single application area. However, as the scope of the analysis is expanded to encompass different domains, these characteristics become more evident. Building on these insights and on the findings of Section 3, it becomes evident that AI methods evolve along a continuum that spans building-level management, community-scale microgrids, and, as an emerging frontier CES. As illustrated in Figure 3, the complexity of the systems and the required time horizons increase gradually when moving from operational building control, through community coordination, to survival-critical environments.

As presented in Figure 3, the building domain continues to be the most mature, with AI methods predominantly used for short-term operational tasks such as forecasting, comfort management, and anomaly detection. At the microgrid or community level, these same approaches must be adapted to coordinate heterogeneous actors and distributed resources, introducing additional uncertainty and the need for negotiation mechanisms. However, extension of this logic to CES results in the escalation of challenges: It is imperative that AI is not only capable of managing energy but also integrating life-support and recycling functions, where reliability and resilience are critical. The diagonal trajectory delineated in the figure serves to emphasize two key concepts: first, the transferability of methods and, second, the progressive amplification of challenges. It is evident that algorithms which are effective in controlled building environments are likely to confront issues of scalability and robustness in communities. These issues assume an existential dimension in the context of CES. This framework positions cross-domain transfer not simply as a matter of applying existing tools in new settings, but as a research agenda that demands rethinking integration, resilience, and long-term sustainability across scales.

4.4. Towards a Multi-Layered AI Framework for Sustainable Energy Systems

The synthesis of research gaps in Section 4.2 shows that studies on AI for energy systems remain fragmented, with methodological progress often confined to isolated domains such as prediction, control, or market coordination. To address this fragmentation, this review introduces a multi-layered AI framework that integrates these domains within a single conceptual structure composed of four complementary layers: (i) perception and prediction, (ii) control and optimization, (iii) market and transactive coordination, and (iv) sustainability anchoring. The framework connects the operational, market and environmental dimensions of AI-based energy management and provides a reference for cross-domain synthesis. Its structure and main components are presented in

Table 11. Quantitative and qualitative evaluation of the proposed multi-layered AI framework for sustainable energy systems.

Layer	Trends Observed in the Literature	Proposed Extensions (Framework Contribution)	Indicative Performance and Evaluation Metrics
Perception and Prediction	Widespread use of ML/DL for short-term forecasting (loads, prices, anomalies); early adoption of edge and federated approaches; DT mostly at experimental stage	Develop unified, scalable pipelines combining edge/federated AI and digital twins for real-time, privacy-preserving, and explainable prediction	Typical latency < 1 s; forecast MAE 5–8%; expected energy optimization gain 5–10%; improved data privacy and traceability
Control and Optimization	RL and MPC-hybrids show strong potential but remain validated mainly in simulations; limited safety guarantees and poor interoperability with legacy BMS	Advance robust RL/MPC formulations with built-in safety, interoperability standards, and deployment in real-world pilots at building and community scales	Control horizon 5 min–1 h; achievable energy savings 10–20%; cost savings 8–15%; stability > 95% under disturbances
Market and Coordination	MARL, game-theoretic models, and blockchain used in conceptual or lab-scale TE studies; DEM–TE coupling still fragmented	Establish integrated control–market architectures that embed fairness, resilience, and transparency, enabling deployment in energy communities and scalable TE platforms	Transaction time < 2 s; fairness index 0.9–1.0; market cost reduction 5–12%; consensus success > 95% under normal connectivity
Sustainability and Life cycle	Very limited works embedding LCA/LCC into EMS; mostly conceptual or surrogate models without operational integration	Embed predictive LCA/LCC in EMS workflows; couple AI-based control with embodied/operational impact models; improve interoperability of data and signals	Predictive LCA accuracy ±10%; lifecycle cost reduction 5–10%; emission reduction 8–15%; low data interoperability risk

.

In contrast to earlier classifications that prioritized algorithmic typologies or subsystem functions, the proposed framework emphasizes functional interoperability and methodological consistency between layers. It integrates short-term operational intelligence with long-term sustainability assessment. Its originality lies in explicitly embedding LCA/LCC-based evaluation and resilience criteria as inherent parts of AI-driven decision-making. In this sense, AI is conceived not only as an optimization tool, but as a mediating layer between control performance, market dynamics, and sustainability outcomes.

As summarized in Table 11, the framework is open and adaptable to evolving technologies, standards, and domains. Although the review focuses on building and microgrid applications, the same structure extends naturally to isolated infrastructures such as space habitats or terrestrial islands networks, where energy, life-support and recycling loops must be managed together. Although the framework is presented in a conceptual manner, its structure has been designed to serve as a foundation for subsequent validation studies in smart building and home energy contexts, and AI-enabled energy management experiments currently planned by the author’s research group. Furthermore, the objective is to stimulate further discussion and comparative analyses by other research teams, thereby providing a reference architecture for interdisciplinary studies and forthcoming conference and journal publications.

By organizing AI contributions along interoperable layers, the framework provides a distinct conceptual synthesis of the field and outlines directions for future research, including predictive LCA integration, cross-layer data exchange, and explainable control mechanisms. For engineers and practitioners, it offers a practical reference for combining efficiency, transparency, and sustainability in intelligent energy systems. In the author’s view, the multi-layered AI framework represents the original conceptual contribution of this review and forms a foundation for further digital-twin-based validation and implementation.

5. Conclusions

This review has examined the application of AI to TE, dynamic energy management, and life cycle-oriented approaches within smart local energy systems. The study analyzed 97 publications, highlighting both methodological advances and persistent gaps. The results demonstrate that, while forecasting, RL, and market-based coordination are becoming increasingly sophisticated in the domains of building and microgrids [40,55], their integration with long-term sustainability objectives remains limited [26,117].

The original contribution of this review has two complementary dimensions:

• A multi-layered AI framework (original conceptual synthesis): This integrates perception and prediction, control and optimization, market coordination, and sustainability anchoring. It addresses the methodological disconnect between operational intelligence and life cycle-based evaluation, positioning AI as a systemic enabler of sustainable infrastructures. The framework is designed to be open, interoperable, and adaptive to emerging technologies and standards.
• A cross-domain perspective (extension of scope): This demonstrates how AI methods validated in buildings and microgrids can inform critical domains such as Closed Ecological Systems (CESs) and islanded energy networks, where energy, life support and recycling loops must be co-managed. This perspective positions sustainability and resource circularity as intrinsic dimensions of AI-driven energy management.

Collectively, these contributions form a roadmap that links intelligent buildings, adaptive communities, and extreme-environment systems under a shared sustainability vision. Based on the results and analyses provided in this paper, it is suggested that future research should proceed in four directions: (i) toward interoperable and explainable AI stacks combining edge, federated and hybrid RL–MPC architectures; (ii) toward the stronger integration of market and control layers in energy communities; (iii) toward embedding predictive LCA and resilience in EMS workflows; and (iv) toward validating the framework in constrained contexts such as CES.

The next step is operational validation in real-world pilots, enabling the transition of AI from algorithmic innovation to systemic implementation. This evolution requires close collaboration among research, industry, and policy actors to align efficiency, resilience, and long-term sustainability across scales and domains.