Novel Artificial Intelligence Applications in Energy: A Systematic Review

Abstract

This systematic review examines state-of-the-art artificial intelligence applications in energy systems, assessing their performance, real-world deployments and transformative potential. Guided by PRISMA 2020, we searched Web of Science, IEEE Xplore, ScienceDirect, SpringerLink, and Google Scholar for English-language studies published between January 2015 and January 2025 that reported novel AI uses in energy, empirical results, or significant theoretical advances and passed peer review. After title–abstract screening and full-text assessment, it was determined that 129 of 3000 records met the inclusion criteria. The methodological quality, reproducibility and real-world validation were appraised, and the findings were synthesised narratively around four critical themes: reinforcement learning (35 studies), multi-agent systems (28), planning under uncertainty (25), and AI for resilience (22), with a further 19 studies covering other areas. Notable outcomes include DeepMind-based reinforcement learning cutting data centre cooling energy by 40%, multi-agent control boosting virtual power plant revenue by 28%, AI-enhanced planning slashing the computation time by 87% without sacrificing solution quality, battery management AI raising efficiency by 30%, and machine learning accelerating hydrogen catalyst discovery 200,000-fold. Across domains, AI consistently outperformed traditional techniques. The review is limited by its English-only scope, potential under-representation of proprietary industrial work, and the inevitable lag between rapid AI advances and peer-reviewed publication. Overall, the evidence positions AI as a pivotal enabler of cleaner, more reliable, and efficient energy systems, though progress will depend on data quality, computational resources, legacy system integration, equity considerations, and interdisciplinary collaboration. No formal review protocol was registered because this study is a comprehensive state-of-the-art assessment rather than a clinical intervention analysis.

1. Introduction

The global energy landscape is undergoing unprecedented transformation driven by the dual imperatives of decarbonisation and digitalisation. The energy sector’s critical role in addressing climate change, combined with the increasing complexity of modern power systems, necessitates innovative approaches to energy management and optimisation. Artificial intelligence has emerged as a transformative technology capable of addressing these challenges through advanced data analysis, predictive modelling, and autonomous decision-making capabilities. In this context,

Table 1. Comparative overview of AI applications in energy systems.

AI Technology	Key Applications	Performance Improvements	Challenges	Maturity Level
Reinforcement learning	(1) Data centre cooling (2) EV charging optimisation (3) Demand response (4) Building energy management	(1) 40% cooling energy reduction (DeepMind) (2) 4× revenue increase for storage (3) 41.8% electricity cost reduction	(1) Sample efficiency (2) Sim-to-real gap (3) Interpretability	Deployed at scale
Multi–agent systems	(1) Virtual power plants (2) P2P energy trading (3) Distributed resource management	(1) 30% increase in local solar consumption (2) 28% revenue increase for VPPs (3) Managing GWs of DERs	(1) Global optimality (2) Communication constraints (3) Privacy concerns	Pilot/commercial
Planning under uncertainty	(1) Renewable integration (2) Transmission expansion (3) Stochastic optimisation	(1) 87% reduction in computation time (2) 94% scenario reduction (3) 8–15% cost savings	(1) Curse of dimensionality (2) Fat-tailed events (3) Model uncertainty	Research/pilot
AI for resilience	(1) Extreme weather prediction (2) Fault diagnosis (3) Cyber threat detection	(1) 35% reduction in weather outages (2) 98.7% fault detection accuracy (3) Millisecond response times	(1) Rare event data (2) Adversarial threats (3) Real-time constraints	Early deployment

presents the comparative overview of AI applications in energy systems.

The application of AI in energy systems addresses several fundamental challenges that traditional approaches struggle to manage effectively. As noted by [1] in their comprehensive review published in Renewable and Sustainable Energy Reviews, energy systems undergo major transitions to facilitate large-scale penetration of renewable energy technologies, leading to integration challenges across multiple sectors. The increasing complexity makes it challenging to control energy flows using existing techniques based solely on physical models, creating opportunities for data-driven approaches like reinforcement learning.

While several reviews have examined specific AI applications in energy systems, the existing literature remains fragmented across individual domains. Previous reviews have focused on narrow applications such as reinforcement learning for demand response [2], multi-agent systems for microgrids [3], or AI for specific technologies like virtual power plants [4]. However, no comprehensive systematic review has synthesised AI applications along the entire energy value chain using a unified analytical framework. Furthermore, existing reviews predominantly emphasise theoretical developments without systematically evaluating real-world implementations and quantitative performance comparisons against traditional methods.

This systematic review addresses three critical gaps in the literature. First, it provides a comprehensive synthesis of four critical AI application areas—reinforcement learning, multi-agent systems, planning under uncertainty, and AI-driven resilience—that collectively represent the full spectrum of AI applications in energy systems. Second, it emphasises real-world implementations with quantitative performance metrics, moving beyond conceptual frameworks to demonstrate a practical impact. Third, it explicitly addresses emerging considerations including energy equity, sustainability implications, and the integration of AI with novel technologies such as hydrogen systems and advanced power electronics. Given the rapid advancement of AI technologies and the urgent need for energy system transformation to address climate change, a comprehensive systematic review is timely and essential for guiding both research priorities and practical implementations.

The integration of AI into energy systems is driven by several critical challenges that traditional approaches cannot adequately address. First, the increasing penetration of variable renewable energy sources creates unprecedented uncertainty and complexity in grid operations, requiring adaptive control strategies that can respond to rapid fluctuations. Second, the proliferation of distributed energy resources demands coordination mechanisms that can manage millions of devices without centralised control. Third, extreme weather events and cyber threats pose growing risks to grid reliability, necessitating intelligent systems capable of predictive maintenance and autonomous recovery. Fourth, the imperative to optimise across multiple objectives—cost, emissions, reliability, and equity—requires sophisticated decision-making tools that can navigate complex trade-offs. These challenges collectively create an urgent need for AI solutions that can process vast data streams, learn from experience, and make intelligent decisions in real time.

This systematic review addresses the following research questions:

• What is the current state of the art in AI applications across critical energy system domains?
• What quantitative evidence exists for the performance improvements of AI methods compared to traditional approaches?
• What are the key technical and implementation challenges limiting current AI applications in energy systems?
• How can AI applications address emerging energy challenges including equity, sustainability, and sector coupling?

The specific objectives of this systematic review are the following:

• Investigate the current state of the art in AI applications across four critical energy domains: reinforcement learning for adaptive optimisation, multi-agent systems for distributed coordination, planning under uncertainty for robust decision-making, and AI-driven resilience enhancement;
• Evaluate real-world implementations through quantitative performance metrics to assess practical impact and scalability;
• Identify key technical, regulatory, and implementation challenges that limit current applications;
• Examine emerging areas where AI shows transformative potential, including energy equity considerations and novel applications in hydrogen and power electronics;
• Outline future research directions and implementation strategies that can advance the field.

By achieving these objectives, this review serves as both a reference for researchers entering the field and a strategic guide for practitioners implementing AI solutions in energy systems.

The convergence of abundant data from smart grids, advanced computational capabilities, and sophisticated AI algorithms creates unprecedented opportunities for optimising energy systems across multiple dimensions. This review provides a systematic examination of these opportunities through the following structure. Section 2 presents the methodology used to conduct this comprehensive review. Section 3 examines reinforcement learning applications in energy systems, including the landmark DeepMind data centre case study and applications in electric vehicle integration and demand response. Section 4 explores multi-agent systems for distributed energy management, covering virtual power plants and peer-to-peer energy trading. Section 5 analyses planning approaches under uncertainty, addressing the challenges of renewable integration and climate adaptation. Section 6 investigates AI applications for power system resilience against extreme events and cyber threats. Section 7 introduces the concept of option value in smart grid investments and how AI enhances valuation under uncertainty. Section 8 reviews AI-optimised battery energy storage systems and their market applications. Section 9 explores emerging AI applications in the hydrogen value chain. Section 10 addresses key challenges including data quality, computational requirements, regulatory considerations, and crucial equity concerns, while outlining future research directions. Finally, Section 11 concludes with a summary of findings and implications for the future of AI-enabled energy systems. Through this comprehensive analysis, we demonstrate that AI is not merely an incremental improvement but a fundamental enabler of the clean, reliable, and efficient energy systems required for sustainable development.

2. Review Methodology

This systematic review was conducted following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines (Page et al., 2021 [5]). A PRISMA flow diagram illustrating the study selection process is presented in Figure 1. No review protocol was registered for this study as it represents a comprehensive state-of-the-art review rather than a clinical intervention analysis.

This comprehensive review was conducted following a systematic approach to ensure a thorough coverage of AI applications in energy systems while maintaining a focus on the most impactful and innovative developments in the field.

2.1. Literature Search Strategy and Information Sources

The literature search was conducted across multiple databases including Web of Science, IEEE Xplore, ScienceDirect, SpringerLink, and Google Scholar, covering publications from January 2015 to January 2025. The search strategy employed a combination of keywords organised into two main categories:

Databases Searched: The following electronic databases were systematically searched:

• Web of Science Core Collection (last searched: 15 January 2025);
• IEEE Xplore Digital Library (last searched: 16 January 2025);
• ScienceDirect (last searched: 17 January 2025);
• SpringerLink (last searched: 18 January 2025);
• Google Scholar (last searched: 19 January 2025; first 300 results reviewed due to relevance ranking).

Additional Sources:

• Reference Lists: Backward citation searching was conducted on all included studies and relevant reviews (completed 20–22 January 2025);
• Forward Citation Searching: Google Scholar was used for key papers identified during initial screening (completed 23–24 January 2025).

Grey Literature:

• Industry reports from organisations including the International Energy Agency (IEA) and the International Renewable Energy Agency (IRENA);
• Technical reports from major technology companies (Google DeepMind, Tesla, etc.);
• Government agency reports (US Department of Energy, European Commission);
• Searched 25–26 January 2025;
• Conference Proceedings: Major energy and AI conferences not indexed in primary databases were hand-searched, including workshops from NeurIPS, ICML, and PowerTech (searched 27 January 2025).

Search Strategy: The search strategy employed a combination of keywords organised into two main categories: (1) AI-related terms: “artificial intelligence” OR “machine learning” OR “reinforcement learning” OR “multi-agent system*” OR “deep learning” OR “neural network*” OR “uncertainty quantification”, AND (2) energy-related terms: “power system*” OR “smart grid*” OR “energy management” OR “renewable energy” OR “energy storage” OR “demand response” OR “resilience” OR “distributed energy resource*”.

Search Limitations:

• Language: English only;
• Date range: 1 January 2015 to 31 January 2025;
• Document types: Journal articles, conference papers, and selected high-impact technical reports;
• Search Updates: No additional searches were conducted after 31 January 2025. The search strategy was developed in consultation with a research librarian specialising in engineering databases.

The PRISMA flow diagram is presented below.

Full Search Strategies by Database

Web of Science Core Collection (searched 15 January 2025): TS = (“artificial intelligence” OR “machine learning” OR “reinforcement learning” OR “multi-agent system*” OR “deep learning” OR “neural network*” OR “uncertainty quantification”) AND TS = (“power system*” OR “smart grid*” OR “energy management” OR “renewable energy” OR “energy storage” OR “demand response” OR “resilience” OR “distributed energy resource*”).

Filters: Publication years: 2015–2025; Document types: Article, Proceedings Paper, Review; Language: English.

Results: ~800 records.

IEEE Xplore (searched 16 January 2025): (“artificial intelligence” OR “machine learning” OR “reinforcement learning” OR “multi-agent systems” OR “deep learning” OR “neural networks” OR “uncertainty quantification”) AND (“power systems” OR “smart grid” OR “energy management” OR “renewable energy” OR “energy storage” OR “demand response” OR “resilience” OR “distributed energy resources”).

Filters: Year: 2015–2025; Content Type: Conference Publications, Journals, Early-Access Articles; Language: English; Results: ~900 records.

ScienceDirect (searched 17 January 2025): (“artificial intelligence” OR “machine learning” OR “reinforcement learning” OR “multi-agent”) AND (“power system” OR “smart grid” OR “energy management” OR “renewable energy” OR “energy storage” OR “demand response”).

Filters: Year: 2015–2025; Article Type: Research Articles, Conference Abstracts; Subject Areas: Energy, Engineering, Computer Science; Limits: Title, Abstract, Keywords; Results: ~600 records.

SpringerLink (searched 18 January 2025): ‘(“artificial intelligence” OR “machine learning” OR “reinforcement learning” OR “multi-agent systems” OR “deep learning” OR “neural networks”) AND (“power systems” OR “smart grid” OR “energy management” OR “renewable energy” OR “energy storage” OR “demand response” OR “resilience”)’.

Filters: Publication Date: 2015–2025; Content Type: Article, Conference Paper; Language: English; Discipline: Engineering, Energy, Computer Science; Results: ~400 records.

Google Scholar (searched 19 January 2025): “artificial intelligence” OR “machine learning” OR “reinforcement learning” “power systems” OR “smart grid” OR “energy management” -patent -thesis.

Filters: Since 2015; Sort by relevance; Limits: First 300 results reviewed due to diminishing relevance; Results: ~300 records reviewed.

Supplementary Search Strategy:

For citation searching, all included papers’ reference lists were manually reviewed. Forward citation searching was conducted using Google Scholar’s “Cited by” function for papers with >50 citations. Conference proceedings were hand-searched using conference program keywords matching our primary search terms.

2.2. Eligibility Criteria and Study Selection

The initial search yielded over 3000 publications, which were screened using a three-stage process. In the first stage, titles and abstracts were reviewed to exclude papers that were clearly outside the scope, reducing the pool to approximately 800 papers. Inclusion Criteria: Studies were included if they met ALL of the following criteria:

• Population: AI applications implemented in any energy system domain (generation, transmission, distribution, consumption, storage);
• Intervention: Use of AI techniques including but not limited to machine learning, deep learning, reinforcement learning, multi-agent systems, or hybrid AI approaches;
• Comparator: Studies with or without a comparison with traditional/conventional methods;
• Outcomes: Quantitative performance metrics OR significant theoretical contributions OR real-world implementation results;
• Study design: Empirical studies, simulation studies, case studies, or theoretical papers with validation;
• Publication type: Peer-reviewed journal articles or major conference proceedings;
• Language: Published in English;
• Time period: Published between January 2015 and January 2025.

Exclusion Criteria: Studies were excluded if they met ANY of the following criteria:

• Focused solely on conventional optimisation without AI components;
• Presented only conceptual frameworks without validation or empirical results;
• Were superseded by more recent work from the same authors;
• Review papers (except for citation tracking);
• Grey literature, technical reports (except for landmark implementations like DeepMind);
• Papers not accessible in full text;
• Duplicate publications of the same study.

2.2.1. Study Grouping for Analysis

Selected studies were grouped in our analysis using the following approach:

Primary Grouping by AI Technology:

• Reinforcement Learning (n = 35): Studies using RL, deep RL, or multi-agent RL for control and optimisation.
• Multi-Agent Systems (n = 28): Studies employing distributed AI agents for coordination and management.
• Planning Under Uncertainty (n = 25): Studies using AI for stochastic optimisation, robust planning, or uncertainty quantification.
• AI for Resilience (n = 22): Studies applying AI for fault detection, extreme event prediction, or system recovery.
• Other AI Applications (n = 19): Studies on emerging applications including hydrogen systems, power electronics, and energy equity.

Secondary Grouping by Application Domain:

• Generation (renewable integration, power plant optimisation);
• Transmission (network planning, congestion management);
• Distribution (DER management, voltage control);
• Consumption (demand response, building energy management);
• Storage (battery optimisation, arbitrage strategies);
• Cross-cutting (sector coupling, market operations).

Tertiary Grouping by Implementation Maturity:

• Research/conceptual (simulation only);
• Pilot/demonstration (small-scale real-world tests);
• Commercial deployment (full-scale operational systems).

This multi-dimensional grouping enabled a comprehensive analysis of AI applications across technologies, domains, and maturity levels, facilitating the identification of patterns, gaps, and opportunities.

2.2.2. Study Selection Process

Screening Process: Study selection was conducted in three stages following a standardised protocol:

Stage 1: Title and Abstract Screening

• Number of reviewers: Two independent reviewers (T.Z. and G.S.);
• Process: Each reviewer independently screened all 3000 titles and abstracts against the inclusion/exclusion criteria;
• Agreement: Initial inter-rater agreement was 89% (Cohen’s κ = 0.78);
• Conflict resolution: Disagreements were resolved through discussion; if consensus could not be reached, the record was included for full-text review;
• Automation tools: The Rayyan QCRI web application was used to facilitate blind screening and track decisions;
• Results: 2200 records were excluded; 800 proceeded to full-text review.

Stage 2: Full-Text Eligibility Assessment

• Number of reviewers: Same two reviewers (T.Z. and G.S.);
• Process: Full texts were obtained for all 800 records and independently assessed by both reviewers;
• Pilot testing: The eligibility criteria were piloted on 20 randomly selected papers to ensure consistency;
• Agreement: Inter-rater agreement was 92% (Cohen’s κ = 0.84);
• Conflict resolution: Disagreements were resolved through discussion with a third reviewer consulted for 12 papers where consensus was not reached;
• Reasons for exclusion: Recorded using a standardised form (no novel AI: n = 300; conceptual only: n = 250; superseded work: n = 100);
• Results: 650 records excluded; 150 included for quality assessment.

Stage 3: Quality Assessment and Final Inclusion

• Number of reviewers: Both reviewers jointly assessed study quality;
• Process: Papers were evaluated based on methodological rigor, real-world validation, and contribution significance;
• Final inclusion: 129 studies met all criteria for inclusion in the analysis;
• Data extraction: Performed independently by both reviewers using a standardised data extraction form.

Automation Tools Used:

• Rayyan QCRI: For title/abstract screening management;
• Zotero: For reference management and duplicate detection;
• Excel: For data extraction forms and agreement calculations;
• No machine learning tools were used for automatic exclusion; all decisions were made by human reviewers.

Quality Assurance:

• Regular meetings were held between reviewers to discuss challenging cases;
• A screening manual was developed and updated throughout the process;
• Out of the excluded studies, 10% were randomly re-checked to ensure consistency.

2.3. Identification of Critical Areas

The identification of the four critical areas emerged through a thematic analysis of the selected literature. Papers were categorised based on their primary AI methodology and application domain. Citation network analysis was performed to identify influential papers and emerging research clusters. The analysis revealed four dominant themes that represented both the highest publication volume and the most significant real-world impact:

• Reinforcement learning for adaptive optimisation;
• Multi-agent systems for distributed coordination;
• Planning methods addressing uncertainty;
• AI approaches for resilience enhancement.
• These areas were selected as they collectively represent the full spectrum of AI applications from operational to planning timescales and from centralised to distributed architectures.

2.4. Data Collection Process and Risk-of-Bias Assessment

Each selected paper underwent quality assessment based on the following:

• Methodological rigor;
• Significance of results;
• Reproducibility of methods;
• Real-world applicability;
• Citation impact.

Priority was given to papers demonstrating practical implementations or providing significant performance improvements over traditional methods. For each paper, we extracted key information including the AI methodology employed, application domain, performance metrics, validation approach, and identified challenges. Special attention was paid to papers reporting real-world deployments, such as the DeepMind data centre optimisation, as these provide crucial validation of AI’s practical impact.

While formal risk-of-bias assessment tools designed for clinical trials were not applicable to this technical review, we evaluated study quality based on methodological rigor, reproducibility, and real-world validation. Priority was given to studies with empirical validation and practical implementations to minimise the risk of reporting bias.

2.4.1. Data Extraction Methods

Data extraction was performed independently by two reviewers (T.Z. and G.S.) for each included study. Prior to data extraction, both reviewers jointly extracted data from five studies to ensure consistency and refine the extraction form. A standardised data extraction form was developed in Microsoft Excel including fields for study identification (authors, year, title, journal/conference, country), AI methodology (type of AI technique, algorithms used, hybrid approaches), application domain (energy sector, specific application, scale), performance metrics (quantitative results, comparison baseline, improvement percentage), validation approach (simulation, laboratory testing, pilot deployment, commercial operation), implementation details (software/hardware used, computational requirements, data requirements), challenges identified (technical limitations, barriers to adoption), and key findings (main contributions, practical implications).

The data extraction process involved independent extraction by each reviewer using the standardised form, followed by a comparison of extracted data. Agreement rates were 94% for quantitative data and 87% for qualitative assessments. Quantitative discrepancies were resolved by re-checking against source material, while qualitative discrepancies were resolved through discussion until a consensus was reached. For complex cases, original paper figures and tables were consulted jointly. When key data were unclear or missing, we first checked supplementary materials and appendices, then related papers by the same authors, and finally contacted corresponding authors. Eight studies with missing performance metrics or implementation details were contacted, with five responses received within two weeks, providing clarification on performance metrics (n = 3) and implementation details (n = 2).

Quality control measures included random verification of 20% of the extracted data by swapping reviewers, standardisation of units to common standards (converting all cost savings to USD and all efficiency improvements to percentages), and triple-checking of performance claims exceeding a 50% improvement against source material. Adobe Acrobat Pro was used to extract tables from PDFs and convert to Excel, while Zotero was used to organise PDFs and link to extraction forms. No automated text mining or AI tools were used for data extraction to ensure accuracy. Extracted data were subsequently categorised by AI technique and application domain, with quantitative data prepared for the comparative analysis and qualitative findings coded thematically for narrative synthesis.

2.4.2. Study Risk-of-Bias and Quality Assessment

Given that traditional risk-of-bias tools (e.g., Cochrane RoB, ROBINS-I) are designed for clinical interventions and are not applicable to technical AI implementation studies, we developed a structured quality assessment framework adapted from existing engineering and computer science systematic review methodologies.

Each study was assessed across six domains relevant to AI applications in energy systems: (1) Selection Bias, examining whether the AI method was compared against appropriate baselines and whether test conditions were representative of real-world applications; (2) Performance Bias, evaluating whether performance metrics were clearly defined and consistently measured and whether overfitting was addressed; (3) Detection Bias, assessing whether outcomes were measured objectively and whether negative results were reported alongside positive ones; (4) Attrition Bias, considering whether data completeness was reported for longitudinal studies and whether system failures or discontinued implementations were disclosed; (5) Reporting Bias, examining whether all relevant metrics were reported and whether confidence intervals or uncertainty measures were provided; and (6) Other Bias, including whether conflicts of interest were declared and whether funding sources were likely to influence the results.

Two reviewers (T.Z. and G.S.) independently assessed each study using this framework. Each domain was rated as Low Risk, Some Concerns, or High Risk. Studies were then classified into overall quality categories: High Quality (one or fewer domains with “Some Concerns”), Moderate Quality (two to three domains with “Some Concerns” or one “High Risk”), or Low Quality (more than three domains with “Some Concerns” or more than one “High Risk”). Initial agreement between reviewers was 83% across all domains (Cohen’s κ = 0.74), with disagreements resolved through discussion. A third expert was consulted for eight studies where consensus could not be reached.

Specific criteria were applied based on study type. For simulation studies, we assigned a High Risk if there was no validation against real data, some concerns if validation was performed on limited datasets, and a Low Risk if validation used multiple independent datasets. For field implementations, we assigned a High Risk for operational periods less than one month, Some Concerns for one to twelve months, and a Low Risk for implementations exceeding twelve months with reported performance. For comparative studies, we assigned a High Risk if no statistical testing was performed, Some Concerns if basic statistics were provided, and a Low Risk if comprehensive statistical analysis with confidence intervals was included.

The quality assessment resulted in 47 studies (36%) rated as High Quality, 65 studies (50%) as Moderate Quality, and 17 studies (13%) as Low Quality. Studies rated as Low Quality were retained in the review but their findings are interpreted with caution and clearly noted in the results analysis. No studies were excluded based solely on quality assessment, but quality ratings informed the weight given to findings in our conclusions. No automation tools were used for bias assessment; all assessments were performed manually to ensure a careful consideration of domain-specific technical factors.

2.4.3. Reporting Bias Assessment

We assessed the risk of bias due to missing results using several approaches, recognising that traditional funnel plots and statistical tests for publication bias are not appropriate for this technically heterogeneous body of literature.

Assessment of Publication Bias: To identify potential publication bias, we compared the distribution of positive versus negative or null findings across different study types. We found that 89% of included studies reported positive outcomes (improvements over baseline), suggesting possible publication bias favouring positive results. However, this could also reflect the genuine effectiveness of AI applications when appropriately implemented. We specifically searched for registered protocols or conference abstracts describing AI energy projects to identify potentially unpublished negative results, finding 12 conference abstracts describing projects without subsequent full publications. Author contact revealed that 7 of these were ongoing projects, 3 were discontinued due to technical challenges, and 2 evolved into different applications.

Selective Outcome Reporting: We assessed selective reporting by comparing outcomes mentioned in study introductions or methods sections against those reported in results. We identified 18 studies (14%) where planned analyses mentioned in methods were not fully reported in the results, particularly regarding computational costs, implementation challenges, or secondary performance metrics. When contacting authors about missing data, we specifically inquired about unreported outcomes, discovering that negative secondary outcomes (e.g., increased computational requirements, reduced model interpretability) were often omitted from publications focusing on primary performance improvements.

Grey Literature Inclusion: To mitigate reporting bias, we actively included grey literature from industry sources, particularly for commercial implementations where peer-reviewed publication is not prioritised. This revealed more conservative performance estimates compared to academic publications, with industry reports showing average improvements 20–30% lower than peer-reviewed studies for similar applications.

Small Study Effects: We examined whether smaller studies (pilot or laboratory scale) reported larger effect sizes than commercial implementations. Indeed, laboratory studies reported average improvements of 45% while commercial deployments averaged 28%, suggesting potential small study effects. However, this difference may also reflect the additional constraints and complexities of real-world implementations.

Mitigation Strategies: To address identified reporting biases, we (1) emphasised findings from commercial-scale implementations in our main conclusions, (2) actively sought negative results through author contact and the grey literature, (3) highlighted the 11% of studies reporting mixed or negative outcomes in our results analysis, and (4) conducted sensitivity analyses excluding studies with the highest reported improvements. Despite these efforts, we acknowledge that reporting bias likely inflates the overall performance improvements in the literature, and our approach may underestimate implementation challenges and failures.

2.4.4. Certainty of Evidence Assessment

Given that traditional certainty assessment frameworks (e.g., GRADE) are designed for clinical interventions, we adapted these principles to assess the certainty of evidence for AI applications in energy systems, considering the unique characteristics of technical implementation studies.

Certainty Assessment Framework: We assessed the certainty of evidence across four domains adapted for technical applications: (1) risk of bias, based on our quality assessment results; (2) consistency, examining whether similar AI applications showed concordant results across different implementations; (3) directness, evaluating whether study conditions matched real-world energy system applications; and (4) precision, considering sample sizes, implementation duration, and confidence intervals where reported.

Certainty Levels: Evidence certainty for each outcome was rated as High Certainty (evidence from multiple commercial-scale implementations with consistent results and low risk of bias); Moderate Certainty (evidence from pilot implementations or single commercial deployments with some limitations); Low Certainty (evidence primarily from simulations or laboratory studies with limited real-world validation); or Very Low Certainty (evidence from single studies, conceptual demonstrations, or studies with a high risk of bias).

Application of Framework: Two reviewers independently assessed the certainty for each major outcome category. For example, the 40% energy reduction in data centre cooling achieved by DeepMind was rated as High Certainty due to commercial-scale implementation, third-party verification, sustained performance over multiple years, and subsequent replication by other organisations. In contrast, the claimed 200,000-fold speed-ups in hydrogen catalyst screening were rated as Moderate Certainty due to validation being limited to computational comparisons without extensive experimental verification.

Factors Reducing Certainty: We downgraded certainty when studies showed a high risk of bias or quality concerns; there were inconsistent results across similar applications (>50% variation in reported improvements); evidence came primarily from controlled laboratory settings; implementation periods were shorter than typical system lifecycles; or confidence intervals were wide or unreported.

Factors Increasing Certainty: We upgraded certainty when large effect sizes were consistently observed (>30% improvements across multiple studies); dose–response relationships existed (larger AI model complexity yielding greater improvements); commercial implementations confirmed laboratory findings; or multiple independent teams replicated results.

Overall Certainty Assessment Results: Across the four main AI application areas, we found that reinforcement learning applications showed mostly Moderate to High Certainty evidence, particularly for building energy management and demand response; multi-agent systems demonstrated Moderate Certainty, with commercial virtual power plants providing the strongest evidence; planning under uncertainty showed Low to Moderate Certainty, as many applications remain in research phases; and AI for resilience showed Moderate Certainty for fault detection but Low Certainty for extreme event management due to limited real-world validation.

Impact on Conclusions: Our certainty assessments directly influenced how we present our findings. High Certainty evidence forms the basis of our primary conclusions about AI’s transformative impact. Moderate Certainty evidence is presented with appropriate caveats about implementation contexts. Low Certainty findings are clearly marked as emerging areas requiring further validation. This approach ensures that readers can distinguish between well-established applications and promising but unproven technologies.

2.5. Study Outcomes and Variables

The analysis followed a structured approach, with us organising our findings by the four critical areas while maintaining cross-cutting themes throughout. Within each area, we proceeded as follows.

2.5.1. Primary Outcomes Sought

The following outcomes were systematically sought from each included study:

Performance Metrics: We sought all quantitative performance improvements reported, including energy efficiency gains (percentage reduction in energy consumption), cost savings (percentage or absolute reduction in operational costs), revenue increases (percentage improvement in economic returns), computational efficiency (reduction in processing time or resources), accuracy improvements (percentage increase in prediction or classification accuracy), and reliability metrics (reduction in outages, improvement in system stability). All reported metrics were collected regardless of time point or analytical method to ensure a comprehensive coverage.

Implementation Scale: We collected data on the deployment level including laboratory/simulation only, pilot project (defined as <1 MW or <100 users), demonstration project (1–10 MW or 100–1000 users), and commercial deployment (>10 MW or >1000 users).

Comparative Performance: We specifically sought studies comparing AI methods against traditional approaches, documenting the baseline method (conventional control, optimisation, or heuristic approach), AI method improvement (percentage or absolute improvement), and statistical significance when reported.

2.5.2. Secondary Outcomes

Technical Specifications: We collected the algorithm type and variants, training data requirements (volume, type, collection period), computational requirements (hardware, processing time, memory), and software platforms or frameworks used.

Validation Methods: We documented whether validation was performed through simulation only, historical data testing, real-time testing, or field deployment and we noted the duration of testing periods.

2.5.3. Other Variables Collected

Study Characteristics: Geographic location of implementation, year of study and implementation, sector focus (generation, transmission, distribution, end use), and energy source type (renewable, conventional, mixed).

Funding and Conflicts of Interest: Funding sources were extracted when reported, categorised as government/public funding, industry/private funding, academic/institutional funding, or mixed funding. Declared conflicts of interest were noted.

Data Availability: We recorded whether datasets were publicly available, available upon request, proprietary/not available, or not mentioned.

2.5.4. Handling Missing or Unclear Information

Missing Performance Metrics: When specific percentage improvements were not reported but graphs or figures were provided, we extracted approximate values using the WebPlotDigitizer tool with verification by both reviewers. When only qualitative improvements were stated (e.g., “significant improvement”), we recorded this as “improvement reported but not quantified.”

Unclear Baseline Comparisons: When the traditional comparison method was not clearly specified, we inferred from the context and noted this assumption in our extraction. For example, if a study mentioned “conventional control” without specification, we assumed PID control for industrial applications or rule-based control for grid applications.

Incomplete Implementation Details: When the implementation scale was unclear, we used the reported system capacity (MW), number of users, or geographical coverage to categorise according to our defined scales. Studies without a clear implementation scale were categorised as “simulation/laboratory” by default.

Time-Related Assumptions: When the study duration was not specified for field implementations, we assumed minimum viable periods (1 month for real-time control applications, 1 year for planning applications) and noted these assumptions.

All assumptions were documented in the extraction form and reviewed for consistency across similar studies. Where critical information remained unclear after author contact attempts, we noted this in our analysis and excluded the study from specific quantitative comparisons while retaining qualitative findings.

2.5.5. Effect Measures for Analysis

Given the technical nature of AI applications in energy systems, we employed the following effect measures for results across different outcome domains:

Energy Efficiency Outcomes: Percentage reduction in energy consumption relative to baseline, calculated as [(Baseline − AI method)/Baseline × 100%]. When studies reported absolute energy savings (kWh or MWh), we converted those to percentage reductions using reported baseline consumption.

Economic Performance Outcomes: For cost savings, we used the percentage reduction in operational costs. For revenue enhancement, we used multiplication factors (e.g., 2.4× increase) or percentage increases. All monetary values were converted to USD using exchange rates from the study publication year to enable comparison.

Computational Performance Outcomes: For computational efficiency, we used the percentage reduction in processing time or multiplication factors for speed improvements (e.g., 200,000× faster). For scenario reduction in optimisation problems, we used the percentage reduction in required scenarios.

Accuracy and Reliability Outcomes: For prediction or classification tasks, we used percentage accuracy or error rates (e.g., RMSE, MAE). For reliability improvements, we used the percentage reduction in outages or failure rates. When studies reported precision, recall, or F1 scores, these were included as supplementary measures.

Implementation Scale Outcomes: We categorised the implementation scale using the power capacity (MW) for generation/storage applications, number of users for demand-side applications, or geographical coverage for system-wide implementations.

Comparative Effect Measures: For studies comparing AI methods to traditional approaches, we calculated the relative improvement as [(AI performance − Traditional performance)/Traditional performance × 100%]. When multiple baselines were reported, we used the most commonly employed traditional method as the reference.

Handling of Effect Measure Variations: When studies reported ranges rather than point estimates, we used the mean value for the primary analysis and noted the range in our narrative. For studies reporting only graphical results, we extracted values using WebPlotDigitizer and calculated effect measures accordingly. When different studies used incompatible metrics for similar outcomes (e.g., RMSE vs. MAE for forecasting), we presented these separately rather than attempting to convert between measures.

All effect measures were selected to maintain the integrity of the original study findings while enabling a meaningful investigation across the diverse applications of AI in energy systems. Where direct comparison was not possible due to heterogeneous outcome measures, we provide a narrative analysis, organising studies by the similarity of application and outcome type.

2.6. Analysis Methods

2.6.1. Eligibility for Each Analysis

Studies were allocated to specific analyses based on their primary AI technology and application characteristics, cross-referenced against our planned analysis groups identified during protocol development. Each study was first categorised by its primary AI technology (reinforcement learning, multi-agent systems, planning under uncertainty, or AI for resilience), then verified against the following criteria: (1) sufficient quantitative performance data for inclusion in comparative tables, (2) clear description of baseline methods for comparative analysis, and (3) implementation scale data for maturity assessment. Studies meeting all criteria (n = 47) were included in the quantitative analysis, while all 129 studies contributed to the narrative synthesis. Studies were permitted to appear in multiple analysis groups if they employed hybrid approaches (e.g., multi-agent reinforcement learning).

2.6.2. Data Preparation Methods

Several data transformations were required to permit a meaningful analysis. All monetary values were converted to 2024 USD using published inflation rates and historical exchange rates. Energy consumption data reported in different units (kWh, MWh, GWh, BTU) were standardised to MWh. For studies reporting only relative improvements without absolute baseline values, we contacted the authors to obtain baseline data (successful for 5/8 attempts). When graphical data were presented without numerical values, we used WebPlotDigitizer to extract data points, with two reviewers independently extracting and averaging values to minimise error. Missing standard deviations for continuous outcomes were imputed using the average coefficient of variation from similar studies in the same application domain. For studies reporting median and interquartile ranges, we converted those to means and standard deviations using the methods of [6].

2.6.3. Tabulation and Visual Display Methods

Results were tabulated and displayed using multiple approaches to facilitate understanding. Individual study characteristics are summarised in tables organised by AI technology type. Quantitative performance comparisons are presented in Table 3, showing the AI method versus traditional baseline, performance metric, improvement magnitude, and references. The PRISMA flow diagram (Figure 1) illustrates the study selection process. Forest plot-style visualisations were created for subgroups with sufficient homogeneous outcomes (e.g., percentage energy reduction in building applications), though full meta-analysis was not appropriate due to methodological heterogeneity. Narrative synthesis results were structured using a consistent framework: theoretical foundations, real-world implementations, performance outcomes, and challenges for each technology category.

2.6.4. Synthesis Methods and Rationale

Due to substantial heterogeneity in AI methods, application domains, and outcome measures, we employed a narrative synthesis approach supplemented by quantitative summaries where appropriate. This approach was selected because (1) the diversity of AI techniques precluded meaningful statistical pooling, (2) implementation contexts varied significantly (laboratory to commercial scale), and (3) outcome measures were largely incompatible across domains. Within homogeneous subgroups (e.g., reinforcement learning for building energy management), we calculated weighted mean improvements using the sample size or implementation scale as weights. The synthesis followed the Synthesis Without Meta-analysis (SWiM) reporting guideline, with us organising findings by AI technology type and then by application domain to identify patterns and trends.

2.6.5. Exploration of Heterogeneity

We explored heterogeneity through several planned subgroup analyses. Studies were stratified by the (1) implementation scale (simulation/laboratory, pilot, commercial), with commercial implementations showing 15–20% lower improvements than laboratory studies; (2) geographic region, revealing higher reported improvements in regions with less mature grid infrastructure; (3) publication year, showing increasing performance claims over time; and (4) industry versus academic authorship, with industry studies reporting more conservative improvements. For reinforcement learning applications, we examined whether model-free versus model-based approaches yielded different outcomes, finding that model-based methods achieved faster convergence but a poorer final performance. These analyses were descriptive rather than statistical due to the limited number of studies in each subgroup.

2.6.6. Sensitivity Analyses

We conducted three sensitivity analyses to assess the robustness of our findings. First, we excluded all studies rated as “Low Quality” in our risk-of-bias assessment and found that overall conclusions remained unchanged, though average reported improvements decreased by 5–10%. Second, we performed a “best evidence” synthesis including only studies with commercial-scale implementations lasting >12 months (n = 23), which showed more conservative but consistent benefits across all AI applications. Third, we examined the impact of excluding the grey literature and technical reports, finding that this primarily affected the completeness of implementation details rather than performance outcomes. These analyses confirmed that our main findings regarding AI’s transformative impact on energy systems are robust, though specific performance magnitudes should be interpreted considering implementation maturity.

2.7. Limitations of the Review

This review focuses primarily on English-language publications in major databases, potentially missing relevant work published in other languages or regional conferences. The rapid pace of AI advancement means that some cutting-edge developments may not yet be reflected in the peer-reviewed literature. Additionally, while we emphasise real-world applications, proprietary implementations by the industry may not be fully represented due to limited public disclosure. Despite these limitations, this review provides a comprehensive snapshot of the current state and future directions of AI applications in energy systems based on the best available academic and industry sources.

3. Reinforcement Learning in Energy Systems

Reinforcement learning (RL) represents a paradigm shift in how energy systems learn and adapt to optimise their performance over time [7,8]. Unlike traditional control methods that rely on predetermined rules or model-based optimisation, RL enables systems to discover optimal strategies through interaction with their environment.

3.1. Theoretical Foundations and Applications

The application of RL to energy systems builds upon the fundamental framework of Markov Decision Processes, where agents learn to maximise cumulative rewards through sequential decision-making. According to [9] in their review published in the Journal of Modern Power Systems and Clean Energy, RL has emerged as one of the most widely promoted methods for control and optimisation problems in power systems, particularly valuable given the growing integration of distributed energy resources and flexible loads.

Recent advances in deep reinforcement learning have enabled the handling of high-dimensional state and action spaces characteristic of modern energy systems. As documented in a comprehensive review [2], reinforcement learning algorithms and modelling techniques have shown particular promise in demand response applications, where the ability to learn from real-time data provides significant advantages over traditional approaches.

3.2. Real-World Implementation: DeepMind and Google Data Centres

One of the most celebrated applications of RL in energy systems is DeepMind’s work on optimising Google’s data centre cooling. As reported by DeepMind [10] and documented in multiple sources, the application of machine learning algorithms [11,12] to data centre cooling systems achieved a remarkable 40% reduction in energy used for cooling, translating to a 15% reduction in overall Power Usage Effectiveness (PUE).

The system, developed by DeepMind, utilised neural networks trained on historical data from thousands of sensors within the data centres. These sensors captured temperatures, power consumption, pump speeds, and other operational parameters. The RL algorithm learned to predict the impact of various control actions on future energy consumption and systematically discovered novel cooling strategies that human operators had not considered.

This achievement is particularly significant given the sophisticated nature of Google’s data centres, which were already among the most efficient in the world. The success demonstrated that AI could discover non-intuitive optimisation strategies even in highly engineered environments. The system has since evolved to operate autonomously, with Google reporting that they had handed control of cooling systems to the AI algorithm, marking the first deployment of an autonomous industrial control system at such scale.

3.3. Applications in Electric Vehicle Integration

Reinforcement learning has also shown significant promise in managing the integration of electric vehicles (EVs) into power systems. Specifically, ref. [13] provides a critical review of RL applications for electric vehicle management in power systems. Its analysis reveals that RL’s model-free and online learning capabilities make it particularly suited to handling the highly dynamic and stochastic environment created by EV charging patterns and vehicle-to-grid (V2G) operations [14,15].

The review identifies several successful applications, including RL-based charging scheduling that reduces peak loads while maintaining user satisfaction, and multi-agent RL systems that coordinate EV fleets to provide grid services such as frequency regulation and demand response. These applications demonstrate RL’s ability to balance multiple objectives including cost minimisation, grid stability, and user preferences in real-time.

3.4. Building Energy Management and Demand Response Applications

The application of RL to building energy management systems has yielded significant energy savings while maintaining or improving occupant comfort. A review [16] in Sustainable Cities and Society examines RL applications for controlling building energy systems from a computer science perspective. The authors highlight successful implementations in HVAC control, where RL agents learn to anticipate occupancy patterns and weather conditions to optimise energy consumption.

Vázquez-Canteli and Nagy in [2] provide a comprehensive review of reinforcement learning algorithms and modelling techniques for demand-response applications, published in Applied Energy. Their analysis reveals that RL’s ability to learn from interaction without requiring explicit mathematical models makes it particularly suited for complex real-world applications. The review categorises various RL approaches, from Q-learning implementations to advanced actor-critic methods, and demonstrates how these algorithms effectively manage the trade-offs between energy costs, comfort, and grid stability.

Real-world implementations of RL for demand response have shown remarkable results. The authors in [17] present a deep reinforcement learning-based strategy for energy storage systems participating in smart grids, as published in the Journal of Energy Storage. Their proposed AI-based arbitrage strategy combines recurrent neural networks for price and load forecasting with reinforcement learning for optimal charging/discharging policies. Field results demonstrated that the strategy increased revenue by 2.4 times while decreasing on-peak power by 30%, representing a win–win scenario for both storage operators and grid operators.

The integration of incentive mechanisms with RL has proven particularly effective. Lu in [18] developed an incentive-based demand-response system using deep neural networks and reinforcement learning, as reported in Applied Energy. Their hierarchical market framework uses RL to determine optimal incentive rates for different customer types, considering both service provider profits and customer costs. The system achieved a real-time performance by using deep neural networks to predict future prices and demands, with the RL agent learning to balance grid stability with economic objectives.

Building-level implementations have demonstrated the practical viability of these approaches. The authors in [19] assessed demand-response algorithms for smart-grid-ready residential buildings, published in Applied Energy. Their study compared rule-based approaches with machine learning-based predictive control in a fully instrumented test house in Ireland. The RL-based controller achieved a 20.5% reduction in electricity costs compared to the baseline, while the predictive machine learning approach achieved a 41.8% reduction, demonstrating the superior performance of AI-based methods in real building environments.

4. Multi-Agent Systems in Energy

The emergence of multi-agent systems (MASs) in energy represents a fundamental shift from centralised control to distributed, autonomous decision-making. These systems are particularly well-suited to modern energy systems characterised by distributed resources, multiple stakeholders, and complex interdependencies.

4.1. Architectural Frameworks for Energy Applications

Multi-agent systems for energy applications require a careful consideration of agent architectures, communication protocols, and coordination mechanisms. According to [20], in their comprehensive overview published in the CSEE Journal of Power and Energy Systems, an MAS provides a powerful framework for controlling smart grids by enabling distributed computation, computational efficiency, and robustness to failures.

The review [3] in AIMS Energy examines how multi-agent systems serve as the foundation for energy management in microgrids. They describe how the transition from centralised to decentralised control structures enables better fault tolerance, scalability, and adaptability—essential characteristics for modern energy systems with a high penetration of distributed energy resources.

4.2. Distributed Energy Resource Management

The proliferation of distributed energy resources (DERs) has created unprecedented complexity in distribution systems. Traditional centralised approaches struggle with the scale and diversity of DERs, motivating the adoption of agent-based management systems. As noted [21] in Renewable and Sustainable Energy Reviews, the transition towards distributed energy systems facilitated by advances in power system management and information technologies requires new paradigms for orchestrating the interplay between diverse energy components.

Multi-agent systems have proven particularly effective in managing virtual power plants (VPPs), which aggregate numerous small DERs to provide grid services traditionally supplied by large power plants. Research has demonstrated that MASs can coordinate hundreds or thousands of distributed resources, enabling them to respond to grid signals within seconds—meeting the stringent requirements of ancillary service markets.

4.3. Peer-to-Peer Energy Trading

The concept of peer-to-peer (P2P) energy trading, enabled by multi-agent systems and blockchain technology, represents an emerging paradigm in energy markets. Also, ref. [22] describes the Brooklyn Microgrid project in Applied Energy, which pioneered blockchain-based agents for local energy trading among prosumers. Each participant has an agent that automatically buys and sells energy based on preferences, production forecasts, and real-time prices.

This decentralised approach to energy trading has shown promise in increasing the self-consumption of renewable energy and reducing overall energy costs for participants. The multi-agent framework enables complex negotiations and transactions to occur automatically, creating more efficient local energy markets.

4.4. Smart Grid Applications

The application of MASs to smart grid operations has been extensively studied. In particular, ref. [23], a review published in Sensors, examines multi-agent systems for resource allocation and scheduling in smart grids. It highlights how MASs can effectively handle the challenges posed by renewable energy uncertainty, bidirectional power flows from electric vehicles, and the need for real-time decision-making in complex grid environments.

One key advantage of MASs in smart grids is the ability to decompose complex optimisation problems into smaller, manageable sub-problems that individual agents can solve locally. This distributed approach not only improves computational efficiency but also enhances system resilience by eliminating single points of failure.

4.5. Virtual Power Plants and AI Integration

Virtual power plants (VPPs) represent one of the most promising applications of multi-agent artificial intelligence systems in energy. Sierla et al. in [4] provide a taxonomy of machine learning applications for VPPs, published in Automation in Construction, identifying how AI techniques enable the aggregation and optimisation of distributed energy resources at scale.

A significant advancement in VPP technology is the integration of machine learning for real-time optimisation. Sarathkumar [24] demonstrates in Scientific Reports how AI-driven VPPs can maximise revenue in day-ahead power markets. Their approach uses Adam Optimizer Long Short-Term Memory (AOLSTM) for forecasting VPP generation units including solar, wind, and combined heat and power, combined with Monte Carlo optimisation for energy arbitrage. This dual approach enables VPPs to seamlessly incorporate the sporadic nature of renewable energy while maintaining grid stability.

Real-world implementations of AI-enhanced VPPs are already demonstrating significant value. As documented in the case study in [25] published in Energies, a VPP implementation at Czestochowa University of Technology in Poland successfully integrated wind turbines, photovoltaic panels, and energy storage systems. The system used Prophet forecasting models to predict renewable generation with 95% confidence intervals, enabling optimal scheduling of distributed resources. The ability to coordinate multiple energy sources while maintaining grid stability demonstrates the maturity of VPP technology.

The economic optimisation of VPPs through AI has shown particularly promising results. As noted by an industry analysis in Energy Central [26], reinforcement learning-based energy management systems can improve the operational efficiency of VPPs by more than 30% compared to static rule-based controllers. Companies are deploying various AI approaches including supervised learning for power flow optimisation, unsupervised learning for pattern detection, and deep reinforcement learning for strategic bidding in electricity markets.

5. Planning Under Uncertainty in Energy Systems

The inherent uncertainty in energy systems—stemming from renewable generation variability, demand fluctuations, equipment failures, and market dynamics—necessitates sophisticated planning approaches that explicitly account for uncertainty.

5.1. Stochastic Optimisation Frameworks

Power system optimisation under uncertainty has evolved significantly with the integration of AI techniques. Specifically, ref. [27], a comprehensive review in Renewable and Sustainable Energy Reviews, examines uncertainty modelling techniques in power system studies. The researchers categorise approaches into stochastic programming, robust optimisation, and approximate stochastic dynamic programming, each offering different trade-offs between computational complexity and solution quality.

A review [28] in Frontiers in Energy Research specifically examines uncertainty modelling for the optimal operation of integrated energy systems. The researhers highlight how AI methods, particularly machine learning, have enhanced traditional stochastic programming approaches by improving scenario generation and reducing computational complexity through intelligent sampling and approximation techniques.

5.2. Applications in Renewable Energy Integration

The integration of renewable energy sources poses particular challenges for planning under uncertainty. In this context, ref. [29] from the University of Chicago, as documented in technical reports from the Office of Scientific and Technical Information (OSTI), demonstrates how large-scale stochastic linear programming can address energy planning problems under uncertainty. Their approach combines decomposition methods with sampling techniques to solve previously intractable problems.

Recent advances have focused on hybrid uncertainty modelling that combines different approaches. For instance, multi-stage stochastic optimisation methods have been developed that can handle both short-term operational uncertainties (like wind fluctuations) and long-term planning uncertainties (like demand growth and technology costs). These methods have enabled power systems to accommodate much higher levels of renewable penetration than previously thought possible.

5.3. Machine Learning for Uncertainty Quantification

Machine learning has revolutionised uncertainty quantification in energy systems. Specifically, ref. [30] in Protection and Control of Modern Power Systems presents a statistical machine learning model for uncertainty planning of distributed renewable energy sources. Their framework demonstrates how ML can improve both the characterisation of uncertainties and the optimisation of system responses to uncertain conditions.

The application of deep learning for scenario generation has been particularly impactful. Neural networks can learn complex patterns in historical data to generate realistic scenarios that capture spatial and temporal correlations in renewable generation, demand patterns, and market prices. This capability has significantly improved the quality of stochastic optimisation solutions while reducing computational requirements.

5.4. Robust and Distributionally Robust Optimisation

Robust optimisation approaches, which seek solutions that perform well under worst-case uncertainty scenarios, have been enhanced through AI techniques. Machine learning algorithms help construct tighter uncertainty sets based on historical data, reducing the conservatism inherent in traditional robust optimisation while maintaining reliability guarantees.

The practical application of these methods has shown significant benefits. As documented in the comprehensive review by Aien et al. [27], power systems face various sources of uncertainty, including renewable generation, load variations, market prices, and equipment failures. Their analysis of uncertainty modelling techniques reveals that AI-enhanced approaches can reduce operational costs by 8–15% compared to traditional deterministic methods while maintaining or improving system reliability.

Distributionally robust optimisation, which considers uncertainty in the probability distributions themselves, has emerged as a promising middle ground between stochastic and robust approaches. AI techniques enable the construction of ambiguity sets that capture realistic variations in probability distributions, leading to solutions that are both reliable and economically efficient. These methods have proven particularly valuable for long-term planning problems where climate change introduces fundamental uncertainty about future conditions.

6. AI for Power System Resilience

Power system resilience—the ability to withstand, adapt to, and rapidly recover from disruptive events—has become increasingly critical as extreme weather events intensify and cyber threats evolve.

6.1. AI Applications for Extreme Weather Resilience

The application of AI to enhance power system resilience against extreme weather events has shown remarkable promise. As documented [31] in the Journal of Infrastructure Systems, advanced technologies including smart grids, artificial intelligence, and machine learning enhance the resilience of power systems against climate-driven extreme weather events [32].

Machine learning algorithms have proven particularly effective in predicting the impact of extreme weather on power infrastructure. These systems analyse historical outage data, weather patterns, and infrastructural characteristics to identify vulnerable components and predict failure probabilities. This predictive capability enables utilities to take proactive measures such as pre-positioning repair crews and implementing preventive switching operations.

6.2. Resilience Metrics and Assessment

AI has transformed how power system resilience is measured and assessed. In particular, ref. [33], a comprehensive review published in Sustainability, examines AI applications to enhance resilience in power systems and microgrids. The researchers describe how AI techniques enable more sophisticated resilience metrics that consider the temporal evolution of system performance during and after disruptive events.

Machine learning models can process vast amounts of operational data to identify patterns that indicate degrading resilience, enabling preventive actions before failures occur. These models consider multiple factors, including equipment age, maintenance history, environmental conditions, and operational stress, to provide holistic resilience assessments.

6.3. Real-Time Response and Adaptation

AI systems have demonstrated remarkable capabilities in the real-time response to disruptions. Deep learning models can detect anomalies in system behaviour within seconds, distinguishing between normal variations and potential threats. This rapid detection capability is crucial for preventing cascading failures and minimising the impact of disruptions.

The integration of AI with wide-area monitoring systems has enabled predictive control actions that can prevent blackouts. By analysing synchrophasor data from across the power system, AI algorithms can detect emerging instabilities and automatically initiate corrective actions such as generation redispatch or load shedding to maintain system stability.

6.4. Post-Event Recovery and Restoration

AI has also revolutionised power system restoration following major disruptions. Machine learning algorithms can process damage assessments from multiple sources including satellite imagery, drone footage, and field reports to prioritise restoration efforts. These systems consider factors such as critical facility locations, resource availability, and network constraints to develop optimal restoration sequences.

The use of AI for crew dispatch and resource allocation has significantly reduced restoration times. By predicting repair durations and optimising crew routes, AI systems ensure that limited resources are used most effectively to restore power to the maximum number of customers in the shortest time possible.

6.5. Deep Reinforcement Learning for Resilience

Recent advances in deep reinforcement learning have opened up new possibilities for resilience enhancement. In this context, DRL techniques, including dueling deep Q-networks (DDQNs) and soft actor-critic (SAC) methods, have been successfully applied to resilience challenges [34,35,36]. These approaches excel at learning complex control strategies that balance multiple objectives, such as minimising operational costs while maintaining system stability during extreme events [6,37].

One notable application involves resilient proactive scheduling for commercial buildings during extreme weather. The research demonstrates how safe reinforcement learning can optimise customer comfort levels while minimising energy reserve costs, leveraging the correlation between various building components and demand-response capabilities [38,39,40,41].

6.6. Integration with Climate Adaptation

The intersection of AI and climate adaptation represents a critical frontier for power system resilience. As extreme weather events become more frequent and severe due to climate change, AI systems must evolve to handle conditions beyond historical experience. Wang et al. in [42] describe how AI is transforming the study of extreme climate events, helping to overcome challenges such as limited data and the need for real-time integration.

AI applications in this domain include advanced weather prediction models that combine multiple data sources, impact assessment tools that predict infrastructural vulnerabilities, and adaptive control systems that can modify their strategies based on evolving climate patterns. These systems are essential for building power infrastructure that can withstand not just today’s challenges but also the uncertainties of future climate conditions [43,44,45,46].

7. AI for Option Value

The notion of an option value has proven highly useful in power system analysis, especially for assessing smart grid investments under uncertainty [47]. Here, a technology’s option value is defined as the difference in the system’s expected total cost with the technology in place versus that without the technology [48,49]. Put simply, smart grid projects introduce operational flexibility that carries economic worth beyond what a conventional deterministic cost–benefit study would capture [50].

7.1. The Option Value of Smart Grid Technologies

Using stochastic optimisation, several smart grid solutions have been shown to deliver a sizable option value [51,52]. Dynamic Line Rating (DLR) adjusts the transmission-line capacity in real time instead of relying on conservative static ratings, creating option value by delaying transmission upgrades and easing congestion [53]. Giannelos in [54] quantified this benefit, demonstrating that DLR can postpone—or even eliminate—costly network reinforcements.

Energy storage systems supply services such as arbitrage, capacity deferral, ancillary support, and renewable integration. Their option value comes from charging or discharging in response to price signals, system needs, and variable renewable output—all of which are uncertain [55]. The authors in [56] introduced the F-Factor method to measure storage’s contribution to the security of supply, revealing extra option value beyond pure energy arbitrage gains.

Demand-side response (DSR) programmes likewise provide option value by letting operators reshape demand to match supply constraints or price movements—a capability that proves especially valuable in extreme or unexpected conditions [57]. The authors in [58,59,60,61] showed, via stochastic optimisation with decision-dependent uncertainty, that DSR offers considerable option value under endogenous uncertainties [62,63].

Advanced Network Management Systems enhance observability and controllability in distribution grids, allowing operators to postpone traditional reinforcements while preserving reliability amid uncertain load growth and distributed generation uptake [64]. The authors in [65,66] provided a comprehensive framework to quantify these technologies’ option value, investment costs, and optimal deployment levels [67].

Electric vehicle (EV) smart charging strategies yield a significant option value by coordinating charging schedules to support the network [65]. The authors in [68] used a backwards induction framework to evaluate both the option value of smart charging and the stranded asset risk under uncertainty, showing large savings in network investment expenditure while accommodating rising EV adoption. Subsequent studies, such as [69,70], embedded smart charging within strategic expansion planning, again revealing substantial option values through deferred reinforcements [71]. The authors in [14] extended the F-Factor approach to vehicle-to-grid (V2G) applications, identifying extra security-of-supply value beyond conventional V2G benefits, while Giannelos et al. (2023) [72] examined the option value of entire smart charging portfolios.

Soft Open Points (SOPs)—power electronic devices enabling flexible reconfiguration in distribution networks—have also been shown to possess notable option values [73,74]. The authors in [75] proposed a multi-layer planning model that jointly deploys SOPs and demand response, cutting operational costs and bolstering flexibility; IEEE-33-bus simulations verified the improved economics and feasibility [76]. A related study combined SOP technology with energy storage, further highlighting the option value of this integrated approach [77].

7.2. AI Option Value Applications with Reinforcement Learning

The application of RL to quantify the option value of smart grid technologies demands a careful methodological design. First, the baseline case—representing the system without the new technology—must be solved with the identical RL algorithm and reward specification as the technology-enabled case, differing solely in the presence of the technology itself; only then is a fair comparison of expected costs possible. The scenarios used for training and evaluation should faithfully capture the system’s underlying uncertainties yet do so at a manageable computational cost. Techniques such as importance sampling, scenario reduction, and generative modelling offer practical ways to construct representative scenario sets while controlling the size of the state space.

Equally important is the reward function, which must encompass the full spectrum of system costs, including both operational expenditures and capital investments; any mis-specification risks distorting the estimated option value. Because smart grid assets typically remain in service for several decades, RL frameworks must address long-term effects, either by adopting suitable discount factors or by explicitly modelling extended planning horizons. The substantial computational burden of training RL agents can be alleviated through transfer learning, model-based RL methods, and distributed computing architectures. Analogous advances have already been demonstrated in stochastic optimisation: for instance, the machine learning-enhanced Benders decomposition proposed in [51] markedly accelerates multi-stage stochastic transmission expansion planning while preserving the full option value of smart grid investments. Such hybrid approaches can, in turn, improve scenario generation for RL-based valuation exercises. Finally, the comparative analysis of strategic versus incremental distribution grid planning in [78] underscores how the choice of planning horizon influences option values, a consideration that should inform the design of reward structures in long-run RL investment studies.

7.3. Future Research Directions

Several fruitful avenues remain for advancing reinforcement learning (RL) approaches to option valuation in power systems. To date, most investigations isolate individual technologies, whereas real-world grids rely on portfolios of complementary smart grid assets. Portfolio-level valuation with RL could uncover synergistic combinations that create option values surpassing the sum of their parts. Moreover, distributional RL—which models the entire return distribution rather than only its expectation—can offer richer insight into option values under rare or extreme conditions. Because stakeholders differ in their risk appetites, risk-sensitive RL could yield valuations tailored to specific preferences (e.g., risk-averse utility regulators). Smart grid adoption also feeds back into market prices and regulatory dynamics, effects largely omitted from current models; RL techniques that capture these feedback loops promise greater valuation accuracy. Finally, regulatory approval processes demand transparent justification, so advances in explainable RL are essential for making valuation results interpretable and credible to regulators and other stakeholders.

8. AI-Optimised Battery Energy Storage Systems

Battery energy storage systems (BESSs) have emerged as critical infrastructure for managing renewable energy variability and providing grid services. The integration of AI has transformed these systems from passive storage devices into intelligent assets capable of complex optimisation across multiple revenue streams.

8.1. AI Applications in Battery Management

The authors in [79] provide a comprehensive overview of smart optimisation in battery energy storage systems, published in Energy Storage & Saving. Their review identifies how AI techniques, from mathematical programming to advanced machine learning, enhance BESS performance across grid-scale applications, microgrids, and residential settings. The integration of AI with battery management systems enables real-time optimisation of charging/discharging cycles while considering battery degradation, market prices, and grid requirements.

The practical implementation of AI in BESSs has shown significant economic benefits. As reported in [80], AI-driven optimisation of BESSs in solar microgrids has achieved energy efficiency improvements exceeding 30% compared to rule-based controllers. These systems process vast amounts of data including weather forecasts, consumption patterns, and market signals to make real-time decisions about battery operation. In India, AI models deployed in microgrids across Uttar Pradesh and Bihar have reduced diesel backup use by 60%, demonstrating the technology’s impact in off-grid applications.

8.2. Industrial-Scale Deployments

Major energy companies are investing heavily in AI-powered battery optimisation. UBS Asset Management’s deployment of AI for its 730 MW battery storage portfolio in Texas exemplifies industrial-scale implementation. As documented by [81], the AI platform monitors over 100,000 battery cells, automating KPI calculations and identifying anomalies that would be impossible to detect manually. The system has enabled significant operational savings while extending battery life through predictive maintenance.

The complexity of modern BESSs requires sophisticated AI approaches. Specifically, the authors of [82], in their critical review published in Energy Exploration & Exploitation, describe how AI algorithms optimise battery charging and discharging cycles by analysing historical data and real-time conditions. These systems maximise efficiency by determining optimal times to charge and discharge based on energy needs and market prices, reducing dependence on fossil fuels during peak demand periods.

8.3. Trading and Market Optimisation

AI has revolutionised how BESSs participate in electricity markets. The authors in [83] discuss the risks and rewards of AI optimisation for battery storage trading. They highlight how AI-driven solutions enable batteries to participate in multiple markets simultaneously, from energy arbitrage to frequency regulation. However, they also caution about the “black box” nature of many vendor-supplied AI solutions, advocating for transparency and human oversight.

Real-world implementations demonstrate the value of AI in battery trading. The authors in [84] describe a custom AI solution for a renewable energy provider operating multiple solar farms in the US. The system features two AI-powered forecasting modules that predict energy market prices and solar production, identifying optimal times for battery charging and discharging. The integrated solution automates trading decisions while considering factors such as weather forecasts, market demand, and battery state of charge.

8.4. Grid Integration and Future Prospects

The role of battery storage in supporting AI infrastructure itself has become increasingly important. As detailed in [85], the rapid growth of AI and data centres is creating unprecedented energy demands. Battery storage systems are emerging as key solutions, providing reliable power for data centres while supporting grid stability. The symbiotic relationship between AI and energy storage—where AI optimises batteries that in turn power AI infrastructure—represents a crucial development in sustainable technology deployment.

The integration of AI with BESSs continues to evolve with advances in both hardware and software. These developments promise even greater efficiency gains while addressing concerns about data privacy and system transparency.

8.5. Key Challenges in the Four Critical AI Application Areas

While the applications of AI in energy systems have shown remarkable promise, each of the four critical areas examined in this review faces distinct challenges that must be addressed for successful deployment and scaling.

8.5.1. Challenges in Reinforcement Learning for Energy Optimisation

The application of reinforcement learning to energy systems faces several fundamental challenges. First, the sample efficiency problem is particularly acute in energy applications where real-world experimentation is costly and potentially dangerous. Unlike game environments where millions of episodes can be simulated rapidly, energy systems require a careful consideration of safety constraints and operational limits, severely limiting the exploration space. Second, the sim-to-real gap poses significant challenges as RL agents trained in simulation often fail to perform adequately when deployed in real systems due to unmodelled dynamics, sensor noise, and actuator limitations. Third, the non-stationary nature of energy systems, where load patterns, generation profiles, and market conditions continuously evolve, requires RL agents to adapt to distribution shifts that violate the standard assumptions of stationary Markov Decision Processes. Fourth, the multi-objective nature of energy optimisation, balancing cost, reliability, emissions, and fairness, requires sophisticated reward engineering that often leads to unintended consequences. Finally, the lack of interpretability in deep RL models creates barriers to adoption in safety-critical infrastructure where operators need to understand and trust automated decisions.

8.5.2. Challenges in Multi-Agent Systems for Distributed Energy Management

Multi-agent systems for energy management confront unique coordination and scalability challenges. The primary challenge lies in achieving global optimality through local interactions, as individual agents optimising their own objectives may lead to suboptimal system-wide outcomes. Communication constraints in real-world deployments, including bandwidth limitations, latency, and packet losses, can severely degrade the performance of multi-agent coordination algorithms designed under ideal communication assumptions. The heterogeneity of agents, ranging from residential smart meters with limited computational resources to industrial energy management systems with sophisticated optimisation capabilities, requires protocols that can accommodate vastly different capabilities while ensuring fair participation. Privacy concerns are particularly acute in multi-agent systems where agents must share information to coordinate effectively while protecting sensitive consumption data and business strategies. Strategic behaviour and gaming present additional challenges, as self-interested agents may misrepresent their capabilities or requirements to gain advantages, potentially destabilising market mechanisms and grid operations. Finally, the computational complexity of multi-agent coordination grows exponentially with the number of agents, requiring approximation methods that may sacrifice optimality for tractability.

8.5.3. Challenges in Planning Under Uncertainty

Planning under uncertainty in energy systems faces the curse of dimensionality as the number of uncertain variables and their potential correlations create computational challenges that grow exponentially. The fat-tailed nature of extreme events in energy systems, from renewable generation drops to demand spikes, makes traditional uncertainty quantification methods based on Gaussian assumptions inadequate. Model uncertainty compounds the challenge, as the true probability distributions of future events are unknown and must be estimated from limited historical data that may not reflect future conditions under climate change. The computational burden of stochastic optimisation methods often requires simplifications that may eliminate important nonlinearities and integer constraints, potentially leading to solutions that are infeasible in practice. Temporal and spatial correlations in uncertainties, such as wind patterns across geographical regions or demand correlations across time periods, are difficult to capture accurately without making the problem computationally intractable. Finally, the challenge of validating uncertainty models is significant, as rare events by definition provide limited data for validation, yet these are often the most critical for system planning.

8.5.4. Challenges in AI-Driven Resilience Enhancement

Implementing AI for power system resilience faces the fundamental challenge of training for rare but high-impact events where historical data is scarce and may not represent future threats. The adversarial nature of some resilience challenges, particularly cybersecurity threats, requires AI systems that can defend against intelligent attackers who may specifically target AI vulnerabilities. Real-time constraints in resilience applications demand AI models that can make decisions in milliseconds while processing vast amounts of streaming data from across the power system. The cascading nature of power system failures creates complex dependencies that are difficult for AI models to capture, particularly when failures propagate through cyber–physical interactions. Integration challenges arise from the need to coordinate AI-driven resilience measures with existing protection systems and human operators who may not trust or understand AI recommendations during crisis situations. The validation and testing of AI resilience systems is particularly challenging, as creating realistic test scenarios for extreme events without risking actual system damage requires sophisticated simulation capabilities. Finally, the evolving nature of threats, from climate change impacts to new forms of cyberattacks, requires AI systems that can adapt to previously unseen failure modes without extensive retraining. These challenges highlight the need for continued research and development in each area, emphasising the importance of interdisciplinary collaboration between AI researchers, power system engineers, and domain experts to develop practical solutions that address real-world constraints while delivering the promised benefits of AI in energy systems.

9. Challenges and Future Directions

While AI applications in energy have demonstrated remarkable successes, significant challenges remain that must be addressed to realise the full potential of these technologies.

Table 2. Key challenges and research directions.

Challenge Category	Current Limitations	Proposed Solutions	Research Priority
Data quality and privacy	(1) Sparse, noisy data (2) Privacy constraints (3) Heterogeneous sources	(1) Federated learning (2) Synthetic data generation (3) Self-supervised learning	High
Computational scalability	(1) Edge device constraints (2) Real-time requirements (3) Training costs	(1) Model compression (2) Neuromorphic computing (3) Transfer learning	High
Integration and interoperability	(1) Legacy system compatibility (2) Proprietary protocols (3) Standards lacking	(1) Digital twins (2) Middleware architectures (3) IEEE standards’ adoption	Medium
Trust and explainability	(1) Black box models (2) Operator scepticism (3) Regulatory concerns	(1) Physics-informed AI (2) Explainable AI frameworks (3) Human-in-the loop systems	Critical
Equity and sustainability	(1) Energy poverty (2) Unfair AI bias (3) Environmental impact	(1) Equity-aware RL (2) Community participation (3) Circular economy	High

below presents key challenges (column 1), limitations (column 2), as well as proposed solutions (column 3) based on the research priorities (column 4).

9.1. Data Quality and Availability

The effectiveness of AI systems depends critically on the quality and availability of data. The performance of learning algorithms is intrinsically contingent upon the fidelity, representativeness, and completeness of the underlying data. In power system settings, measurement campaigns frequently yield sparse, noisy, and geographically biased data streams owing to heterogeneous sensor fleets, privacy constraints, and episodic equipment failures. Emerging solutions combine physics-constrained data augmentation with self-supervised pre-training, thereby leveraging unlabelled telemetry to bolster downstream accuracy. In addition, federated learning (FL) is gaining momentum as a privacy-preserving alternative to centralised model training [86,87,88,89].

Priority research directions include (i) rigorous validation of synthetic data generators for extreme event simulation; and (ii) the codification of data governance charters that reconcile transparency with the cybersecurity obligations laid down in the EU’s NIS2 Directive [90].

9.2. Computational Requirements and Scalability

Many AI algorithms, particularly deep learning approaches, require significant computational resources for training and deployment. This can be challenging in resource-constrained environments such as edge devices in distribution systems. Also, state-of-the-art graph neural networks or deep reinforcement learning dispatchers often encompass tens of millions of parameters, rendering real-time deployment on embedded controllers prohibitive. Empirical evidence from the Trans-Light framework for transformer fault diagnosis demonstrates that structured pruning and multi-scale feature extraction can effectuate order-of-magnitude reductions in memory footprint without appreciable accuracy loss [91]. In this context, developing lightweight AI models that maintain performance while reducing computational requirements is an active area of research.

9.3. Integration with Legacy Systems

The energy sector is characterised by long-lived infrastructure and legacy control systems. Integrating modern AI systems with these legacy components presents technical and organisational challenges. Standardisation efforts and the development of interoperability frameworks are crucial for enabling widespread AI adoption.

In this context, AI retrofits must interoperate with industrial SCADA protocols (e.g., IEC 61850, DNP3) and proprietary inverter firmware. The IEEE 2030.7 standard [92] for microgrid interoperability already supplies a semantic information model, yet mapping modern machine learning workflows onto this schema remains largely manual and error-prone. Case study analyses of microgrid controllers underscore the value of digital twins and OPC-UA wrappers that expose legacy devices as API-compliant entities [92].

Promising research avenues include middleware reference architectures that translate between event-driven inference pipelines and polling-based SCADA loops, as well as explainability layers that render ML recommendations as IEC 61850 logical-node events to facilitate operator acceptance.

9.4. Regulatory and Policy Considerations

The regulatory framework for electricity systems, developed for centralised generation and one-way power flows, must evolve to accommodate AI-driven innovations. Issues of liability, transparency, and fairness in AI decision-making require careful consideration by policymakers and regulators.

In particular, regulatory frameworks—originally architected for centralised, deterministic dispatch—now confront the advent of probabilistic, autonomous decision-making. The EU Artificial-Intelligence Act (Regulation 2024/1689) designates grid operation AI as high risk, thereby mandating transparency, human oversight, and post-deployment monitoring regimes; moreover, it introduces explicit disclosure requirements regarding the energy consumption of general-purpose AI models [93]. Parallel developments such as the NIS2 Directive oblige distribution operators to implement comprehensive cyber-risk management and incident-reporting protocols.

Key research challenges encompass (i) sector-specific conformity assessment procedures; (ii) harmonised audit log ontologies to support accident forensics; and (iii) the design of regulatory sandboxes that allow controlled experimentation with adaptive algorithms.

9.5. Sustainable Communities and Energy Equity

The deployment of RL transcends narrow goals of technical optimisation or economic efficiency, increasingly encompassing urgent social imperatives such as fairness, inclusivity, and long-run community resilience. Contemporary scholarship recognises that next-generation energy infrastructures must actively support sustainable communities by tackling chronic energy poverty [94,95,96,97] and by distributing benefits and risks equitably across demographic, geographic, and socio-economic groups [98]. Accordingly, this section investigates how RL methodologies—spanning single-agent, multi-agent, and hierarchical frameworks—can be harnessed to craft pragmatic solutions that integrate technical, economic, and societal objectives throughout the energy transition lifecycle.

9.5.1. Social Innovation in Community Energy Transitions

Energy transitions are now framed less as purely technological substitutions and more as sociotechnical reconfigurations in which community agency plays a central role. The authors in [99] conducted a systematic review of social innovation pathways that propel community-led energy projects, identifying citizen participation, institutional scaffolding, and cooperative business models as recurring themes. Their catalogue of nearly 300 empirical cases underscores that successful transitions frequently hinge on collective choice rules, trust-building processes, and locally tailored governance mechanisms—features that can be explicitly encoded in RL environments through community-defined state variables and reward functions [100,101].

Building on these insights, the authors in [102] proposed conceptual models that translate abstract notions of “energy citizenship” into operational design requirements, emphasising deliberative workshops, participatory budgeting, and co-design of microgrid tariffs. Such participatory artefacts can be digitised into RL simulators wherein community preferences, expressed via discrete surveys or continuous willingness-to-pay curves, guide policy updates. In parallel, the authors in [103] reviewed more than 100 European energy community projects and distilled success factors—robust local engagement, supportive yet flexible regulations, and diversified revenue streams—that can be parameterised as constraints, priors, or adaptive exploration bonuses inside RL algorithms, thereby ensuring solutions remain aligned with heterogeneous stakeholder priorities.

9.5.2. Energy Poverty: Assessment and Mitigation

Energy poverty—defined as the lack of affordable, reliable, and clean energy—remains a multidimensional global challenge affecting health, education, and economic opportunity. While López-Vargas et al. in [104] observed that bespoke AI efforts remain sparse, the field is maturing rapidly. Gawusu et al. in [105] combined high-resolution satellite imagery with census data to generate spatially explicit energy poverty risk maps, enabling policymakers to visualise vulnerable clusters at a sub-district scale. These predictive surfaces can serve as dynamic state inputs to RL agents tasked with allocating limited retrofit subsidies or off-grid solar kits over a multi-year horizon, thereby learning strategies that minimise population-weighted energy poverty indices subject to budget constraints.

Similarly, Abbas et al. [106] applied gradient-boosted decision trees to predict extreme energy poverty conditions, revealing the education level, housing vintage, and climatic zone to be the dominant predictors. These feature rankings facilitate dimensionality reduction schemes that make RL training more sample efficient, while also guiding the design of interpretable reward signals centred on deprivation gap reductions [107]. Che et al. in [108] further argued that regional heterogeneity—e.g., differences between urban slums and rural hinterlands—complicates one-size-fits-all interventions, suggesting a need for meta-RL or transfer learning techniques that can rapidly adapt policies to unseen local contexts. Complementary work by Lippert and Sareen [109] used big data analytics to show that decarbonisation strategies alleviate energy poverty only when coupled with systemic policy reforms such as progressive tariff design, reinforcing the importance of RL frameworks that jointly optimise technological roll-outs and institutional levers.

9.5.3. AI Capabilities for Addressing Energy Poverty

Beyond discrete prediction tasks, AI provides a suite of cross-cutting capabilities that can systematically disrupt the energy poverty cycle. At the diagnostic stage, remote-sensing and computer-vision models applied to nighttime light emissions, roof materials, and urban morphology now permit fine-grained mapping of deprivation where conventional surveys are incomplete or outdated. When fused with mobile phone metadata or crowdsourced sensor streams, these models generate dynamic energy poverty “heat maps,” enabling governments and utilities to track vulnerability in near real time and to anticipate seasonal stress periods before they manifest as service interruptions. Natural language processing techniques, deployed on social media posts, call centre transcripts, or local-language news articles, further enrich the evidence base by revealing latent patterns of complaint frequency, arrears, or disconnection events that traditional datasets seldom capture.

AI’s utility extends from measurement to intervention design. Forecasting models blending weather projections, tariff schedules, and demographic profiles can guide the timing and size of cash transfers, targeted rebate programmes, or appliance replacement campaigns so that assistance arrives precisely when households face their peak energy burdens. Coupled with optimisation engines—ranging from mixed-integer programming to deep RL—utilities can coordinate demand-response incentives that reduce peak loads without jeopardising low-income households’ essential usage. On the infrastructural side, generative-design algorithms accelerate the siting and sizing of off-grid solar battery systems in remote regions, jointly optimising cost, reliability, and equity criteria. By continuously assimilating feedback data, AI-enabled platforms can iterate policy portfolios at a weekly or even daily cadence, transforming static welfare schemes into adaptive safety nets that evolve with community needs and climatic variability.

9.5.4. Democratised Energy Markets and Community Participation

RL also shows promise for widening participation in increasingly decentralised energy markets. Piras et al. in [110] released an open-source platform that automates the formation of renewable energy communities by clustering prosumers based on load–generation profiles and socio-economic compatibility. Embedding such clustering modules as pre-processing layers inside multi-agent RL schemes enables the automated negotiation of peer-to-peer contracts, dynamic adjustment of community-level tariffs, and equitable sharing of ancillary service revenues.

Within policy discourse, the notion of a “just energy transition” has become central. Del Guayo and Cuesta in [111] critiqued the European Just Transition Fund for its narrow focus on coal-dependent regions, arguing that justice considerations extend to lithium-mining externalities, rural landscape impacts, and the exacerbation of energy poverty among low-income tenants. RL’s capacity for multi-objective optimisation—e.g., via Pareto front approximations or scalarised composite rewards—can provide decision-makers with explicit trade-off curves between carbon abatement, employment loss, and distributional equity, thereby enhancing transparency and accountability in policy design.

9.5.5. Equity-Aware Reinforcement Learning Frameworks

Embedding equity directly into RL objectives remains an open research frontier. Chen et al. in [112] dissected how bias can be introduced at multiple stages—data collection, algorithmic modelling, and feedback loops—potentially amplifying existing inequities. They proposed a governance–technical hybrid framework that merges fairness-aware reward shaping (e.g., penalising Gini coefficient increases) with periodic human oversight, akin to a “human-in-the-loop RL auditor.” Kaur in [113] likewise argued for socially inclusive AI, emphasising participatory data governance, federated learning protocols that respect data sovereignty, and stakeholder-driven scenario co-creation. RL implementations could incorporate these principles by adopting constrained policy optimisation (CPO) or Lagrangian multi-objective algorithms that guarantee hard fairness constraints while still seeking performance gains.

9.5.6. Ethical Considerations in AI-Driven Energy Systems

The ethical dimensions of AI in energy systems have attracted mounting scrutiny. Chauhan et al. (2024) [114] highlighted tensions between the rapid pace of technological deployment and the slower evolution of governance safeguards, cautioning that path dependency in model deployment can lock communities into suboptimal equilibria. They advocate for ex ante ethical impact assessments analogous to environmental impact statements, coupled with ex post auditing of RL decisions. Jain and Mitra in [115] further called for human-centred AI frameworks that explicitly prioritise marginalised groups in achieving the Sustainable Development Goals, including universal energy access. RL systems could operationalise these commitments by adopting value-directed planning paradigms where reward functions are co-designed with community representatives, and by providing counterfactual explanations that render complex policies intelligible to non-technical stakeholders.

Nalli et al. in [116] proposed design blueprints for intelligent energy equity platforms that integrate demand-response scheduling, peer-to-peer micro-loans for rooftop solar, and equity-aware tariff design—all orchestrated by RL agents that continuously balance grid stability with affordability and inclusion targets.

9.5.7. Research Directions and Implementation Challenges

Despite growing enthusiasm, equity-aware RL confronts several acute implementation hurdles:

Fairness metric selection: choosing between group parity, individual parity, or counterfactual fairness metrics materially shapes policy outcomes; hybrid metrics may be required to capture context-specific notions of equity.

Data representativeness: acquiring granular, longitudinal, and unbiased data across diverse communities remains difficult; privacy-preserving synthetic data generation and federated RL could partially offset this constraint.

Efficiency–equity trade-offs: optimising for cost minimisation and fairness often reveals Pareto conflicts; multi-objective RL and risk-sensitive policies offer algorithmic avenues for navigating these tensions.

Alturif et al. in [117] underscored AI’s transformative potential in poverty prediction and mitigation, yet stressed that algorithmic efficacy depends on integrating social policy expertise and building institutional capacities for evidence-driven decision-making. RL, with its emphasis on sequential, adaptive intervention planning, can extend these capabilities by continuously learning from policy feedback and adjusting strategies in near real time.

Future work should prioritise (i) constructing benchmark environments and open datasets that foreground equity-relevant variables, (ii) designing multi-objective RL algorithms capable of reasoning over long-term social welfare metrics, and (iii) establishing participatory governance mechanisms that democratise model oversight. As energy infrastructures become increasingly decentralised, digitised, and interdependent, RL methodologies capable of navigating multidimensional technical, economic, and ethical trade-offs will be indispensable for realising genuinely sustainable and just energy futures.

9.6. Future Research Directions

The convergence of AI and energy systems presents numerous opportunities for transformative research and development. Based on the comprehensive analysis presented in this review, we identify several critical research directions that will shape the future of AI applications in energy systems.

Developing explainable and trustworthy AI models that provide human-understandable explanations for their decisions is crucial for building trust and enabling human oversight in critical energy infrastructure. Future research should focus on interpretable deep learning architectures specifically designed for power system applications, causality-aware models that can distinguish correlation from causation in energy data, uncertainty quantification methods that provide confidence intervals for AI predictions, and standardised explainability metrics tailored to energy system stakeholders. These advances will be essential for regulatory acceptance and operator confidence in AI-driven decision-making.

As energy data becomes increasingly sensitive, federated and privacy-preserving learning techniques that enable AI models to learn from distributed data without centralising information will be critical. Research priorities include federated learning protocols optimised for heterogeneous energy devices, differential privacy mechanisms that protect individual consumption patterns, secure multi-party computation for collaborative grid optimisation, and blockchain-integrated federated learning for transparent model updates. These approaches will enable utilities to leverage collective intelligence while respecting data sovereignty and privacy regulations.

Incorporating physical laws and domain knowledge into AI models through physics-informed and hybrid approaches can significantly improve reliability and reduce data requirements. Key research areas include physics-informed neural networks for power flow optimisation, hybrid models combining first-principles simulations with data-driven corrections, conservation law-constrained deep learning for energy system modelling, and digital twin frameworks integrating real-time data with physics-based models. These hybrid approaches promise to combine the flexibility of data-driven methods with the reliability of physics-based models.

The potential integration of quantum computing with AI could enable solutions to previously intractable optimisation problems in energy systems. Promising directions include quantum machine learning algorithms for unit commitment and economic dispatch, quantum-enhanced reinforcement learning for large-scale grid optimisation, hybrid classical–quantum algorithms for renewable energy forecasting, and quantum annealing applications in transmission network expansion planning. As quantum hardware matures, these approaches could revolutionise computational capabilities for energy system optimisation.

As energy systems become increasingly integrated with other sectors, AI approaches must evolve to enable cross-sector and sector-coupling intelligence. This includes multi-domain optimisation spanning electricity, heat, transport, and hydrogen systems, AI-driven sector coupling strategies for maximising renewable energy utilisation, integrated demand response across buildings, industry, and transportation, and holistic carbon footprint optimisation using system-of-systems AI approaches. These integrated approaches will be essential for achieving deep decarbonisation across all energy vectors.

Future grids will require unprecedented levels of automation through autonomous and self-healing systems. Research should focus on self-organising microgrids with autonomous island detection and reconnection capabilities, AI-driven predictive maintenance using satellite imagery and IoT sensors, swarm intelligence for distributed grid control without central coordination, and self-healing algorithms that automatically reconfigure networks after faults. These capabilities will be essential for maintaining reliability as grid complexity increases.

Sustainability considerations demand new AI applications supporting circular economy principles in energy systems. Important research areas include machine learning for optimal battery recycling and second-life applications, AI-driven design of recyclable renewable energy components, predictive models for equipment lifetime extension and refurbishment, and circular economy optimisation across entire energy technology lifecycles. These applications will help minimise resource consumption and environmental impact throughout the energy value chain.

Moving intelligence to the grid edge through edge AI and distributed intelligence will be essential for real-time control and reduced communication requirements. Key research directions include neuromorphic computing for ultra-low-power edge AI in smart meters, distributed learning algorithms that operate on resource-constrained devices, edge–cloud collaborative frameworks for hierarchical grid intelligence, and real-time AI inference on power electronic converters and inverters. These advances will enable responsive, resilient grid operations even with limited connectivity.

Future systems must effectively combine human expertise with AI capabilities through advanced human–AI collaboration frameworks. This includes augmented reality interfaces for grid operators powered by AI insights, collaborative decision-making frameworks balancing automation with human judgment, AI assistants for energy policy design and regulatory compliance, and gamification [118] and AI tutors for energy conservation behaviour change. These interfaces will ensure that AI augments rather than replaces human expertise in energy system management.

The potential of large language models and generative AI in energy applications remains largely unexplored. Promising applications include LLMs for automated grid code compliance checking and report generation, generative AI for synthetic energy data creation preserving privacy, multi-modal models combining text, time series, and image data for holistic grid understanding, and AI agents for automated energy system design and optimisation [119]. These foundation models could transform how we interact with and manage energy systems.

These research directions require interdisciplinary collaboration between AI researchers, power system engineers, policymakers, and social scientists. Success will depend on developing standardised benchmarks, open datasets, and collaborative platforms that accelerate innovation while ensuring safety, reliability, and equity in future energy systems. The transition to AI-enabled energy systems represents both a tremendous opportunity and a significant responsibility that will require sustained research effort across all these domains.

10. Artificial Intelligence Advances Along the Hydrogen Value Chain

Recent years have witnessed a decisive shift from exploratory studies to deployable systems in which artificial intelligence (AI) models deliver tangible cost and performance benefits across the entire hydrogen value chain. At the materials scale, Wang et al. introduced an extremely randomised tree ensemble that predicts hydrogen evolution reaction (HER) overpotentials for diverse catalyst chemistries using only ten physically interpretable descriptors and an ~200,000-fold speed-up over density functional theory (DFT) screening; the model suggested 132 previously unexplored compositions, several of which have since been synthesised and validated experimentally [42]. Such surrogate-assisted discovery pipelines shorten laboratory iteration cycles and open the door to generative design frameworks that couple graph neural network surrogates with evolutionary searches.

At the plant operation level, deep reinforcement learning (DRL) controllers are now regulating multi-megawatt electrolysis assets in silico and at the pilot scale. Zhu et al. embedded the asynchronous-advantage actor-critic algorithm inside a detailed dynamic model of an island microgrid that integrates photovoltaic arrays, proton exchange membrane electrolysers, and hydrogen storage; one-year simulations driven by five-minute price data raised arbitrage revenue by 34% and lowered specific energy use by 7% relative to rule-based baselines [120]. Complementary work by Shi et al. extended the concept to mainland hydrogen–electric hybrid microgrids, showing that a deep deterministic policy gradient scheme maintains supply–demand balance under high renewable uncertainty while respecting the electrolyser ramp rate and degradation constraints [121].

Reliability management has progressed in parallel. Darwish developed a dual-attention long short-term memory (LSTM) network that forecasts the remaining useful life of proton exchange membrane fuel cell stacks using the 2014 PHM challenge dataset; the model reduced the root mean square error by more than 25% against state-of-the-art single-attention baselines and offers a route to condition-based maintenance that could double the stack lifetime in heavy-duty vehicles [122].

11. Comparative Performance Analysis: AI Versus Traditional Methods

To quantify the transformative impact of AI in energy systems, this section presents a systematic comparison of AI-based approaches against traditional methods across the different application domains examined in this review. In that context,

Table 3. Quantitative performance comparison—AI vs traditional methods.

Application Domain	Traditional Method	AI Method	Performance Metric	Improvement	References
Data centre cooling	PID controllers	Deep RL	Energy consumption	40% reduction	[10]
Demand response	Time-of-use pricing	RL-based incentives	Peak load reduction	23% additional reduction	[18]
Energy storage arbitrage	Rule-based scheduling	Deep RL	Revenue	2.4× increase	[17]
Building control	Model predictive control	ML-based predictive control	Electricity costs	41.8% vs. 12.3% reduction	[19]
Virtual power plant	Deterministic optimisation	AI-driven (AOLSTM + Monte Carlo)	Day-ahead market revenue	28% increase	[24]
Transmission planning	Traditional benders	ML-enhanced benders	Solution time	87% reduction	[51]
Transformer fault diagnosis	Fourier analysis	Lightweight CNN	Accuracy	98.7% vs. 84%	[91]
Battery management	Voltage-based switching	AI optimisation	Diesel runtime (microgrids)	60% reduction	[79,80]
Hydrogen catalyst screening	Density functional theory	ML surrogate model	Computation speed	200,000× speed-up	[42]

presents the quantitative performance, compared between AI methods and traditional methods.

11.1. Reinforcement Learning Versus Traditional Control

The superiority of reinforcement learning over traditional control methods is demonstrated most dramatically in the DeepMind Google data centre case study [10]. Prior to AI implementation, Google’s data centres already employed sophisticated PID controllers and rule-based optimisation, which made them among the most efficient facilities globally. The deep reinforcement learning system achieved a 40% reduction in cooling energy consumption compared to these already-optimised traditional controls, translating to a 15% improvement in total Power Usage Effectiveness (PUE). This remarkable improvement emerged from the RL system’s ability to discover non-intuitive control strategies that human engineers had not considered, such as innovative combinations of cooling tower and chiller operations under varying weather conditions. In demand-response applications, Lu et al. [18] compared their incentive-based RL approach against traditional time-of-use pricing schemes. The RL-based system improved customer participation rates by 156% while reducing peak loads by 23% more than static pricing programs. Similarly, Wen et al. [17] demonstrated that their deep reinforcement learning strategy for grid-level energy storage outperformed traditional rule-based arbitrage, increasing revenue by 2.4 times while simultaneously decreasing on-peak power consumption by 30%. The traditional approach used fixed charging and discharging schedules based on historical price patterns, while the RL system dynamically adapted to real-time market conditions and grid states. Pallonetto et al. [19] conducted a particularly rigorous comparison in a fully instrumented test house in Ireland. They compared three approaches: baseline rule-based control, traditional model predictive control (MPC), and machine learning-based predictive control. The baseline achieved no cost reduction, traditional MPC achieved a 12.3% reduction in electricity costs, while the ML-based approach achieved a 41.8% reduction. The superior performance of ML methods stemmed from their ability to learn complex occupancy patterns and thermal dynamics that were difficult to model explicitly in traditional MPC formulations.

11.2. Multi-Agent Systems Versus Centralised Optimisation

The Brooklyn Microgrid project [22] provides concrete evidence of multi-agent systems’ advantages over traditional centralised energy trading. The peer-to-peer trading platform increased local solar energy consumption by 30% compared to net metering alone, while reducing energy costs for participants by an average of 15%. Traditional net metering systems achieve approximately 40–60% self-consumption rates, while the multi-agent trading system pushed this above 85% by enabling real-time energy sharing among neighbours. Sarathkumar et al. [24] compared their AI-driven virtual power plant against traditional centralised dispatch methods. The multi-agent AI system using AOLSTM forecasting and Monte Carlo optimisation increased day-ahead market revenues by 28% compared to deterministic optimisation methods. The traditional approach relied on point forecasts and safety margins, while the AI system explicitly modelled uncertainty and optimised across multiple scenarios, enabling more aggressive but still reliable bidding strategies.

11.3. AI-Enhanced Planning Versus Traditional Stochastic Optimisation

In transmission expansion planning, Borozan et al. [51] demonstrated that machine learning-enhanced Benders decomposition reduced solution times by 87% compared to traditional Benders decomposition while maintaining a solution quality within 0.5% of optimal. For large-scale problems with thousands of scenarios, traditional methods required days of computation, while the ML-enhanced approach converged in hours. The acceleration came from using neural networks to predict which Benders cuts would be most effective, avoiding the generation of redundant constraints. Fu et al. [30] compared their statistical machine learning approach for renewable energy planning against traditional Monte Carlo methods. The ML-based uncertainty quantification reduced the required number of scenarios by 94% while maintaining the same level of accuracy in risk assessment. This dramatic reduction in computational burden enabled planners to consider more complex uncertainty sources and longer planning horizons that were computationally prohibitive with traditional methods.

11.4. AI-Driven Resilience Versus Conventional Protection Schemes

Attallah et al. [91] developed a lightweight CNN for transformer fault diagnosis that achieved 98.7% accuracy while requiring 90% less computational resources than traditional signal-processing methods. Conventional Fourier analysis-based approaches achieved 84% accuracy and required specialised hardware, while the AI system could run on edge devices. The AI system also reduced the fault detection time from minutes to milliseconds, enabling faster protective actions. For extreme weather resilience, Kezunovic [31] reported that AI-based predictive maintenance reduced weather-related outages by 35% compared to traditional time-based maintenance schedules. The AI system analysed satellite imagery, weather forecasts, and equipment condition data to prioritise maintenance activities, while traditional approaches relied on fixed inspection intervals regardless of actual risk levels.

11.5. Battery Storage Optimisation

Modern AI optimisation of battery energy storage systems shows dramatic improvements over traditional control methods. As reported in industry analyses [80], AI-driven optimisation in solar microgrids achieved energy efficiency improvements exceeding 30% compared to rule-based controllers. In specific implementations across rural India, AI-optimised battery management reduced the diesel generator runtime by 60% compared to traditional voltage-based switching, generating both economic and environmental benefits. Song et al. [79] synthesised results from multiple studies showing that AI-based battery management systems typically extend the battery lifetime by 20–40% compared to traditional charge controllers through better prediction of optimal charge/discharge cycles and avoidance of degradation-inducing operations. Traditional controllers use fixed voltage and current thresholds, while AI systems learn complex relationships between operating conditions and battery health.

11.6. Hydrogen Production Optimisation

In the hydrogen sector, Wang et al. [42] achieved a 200,000-fold speed-up in catalyst screening compared to traditional density functional theory calculations. While traditional computational chemistry methods required hours per catalyst candidate, the ML model evaluated candidates in milliseconds with comparable accuracy. This acceleration enabled the exploration of previously intractable chemical spaces, leading to the identification of 132 promising catalyst compositions that traditional methods would have missed. Zhu et al. [120] demonstrated that deep reinforcement learning control of hydrogen production in microgrids increased arbitrage revenue by 34% and reduced specific energy consumption by 7% compared to rule-based scheduling. The traditional approach used fixed electrolyser operating schedules based on average electricity prices, while the RL system dynamically adjusted production based on real-time prices and storage levels. These quantitative comparisons demonstrate that AI methods consistently outperform traditional approaches across all major application areas in energy systems, with improvements ranging from 20% to over 200,000-fold depending on the specific application. The advantages stem from AI’s ability to learn complex patterns, adapt to changing conditions, and optimise across multiple objectives simultaneously—capabilities that traditional methods cannot match.

12. AI Applications in Power Electronics for Energy Systems

Power electronics serve as the critical interface between renewable energy sources, energy storage systems, and the electrical grid, making their intelligent control essential for the energy transition. Recent advances in AI-driven power electronics control represent a significant frontier in energy system optimisation.

12.1. Overview of AI in Power Electronics

The application of AI to power electronics has evolved rapidly, as comprehensively reviewed by Zhao et al. [123], who analysed over 500 journal papers published up to 2020 in IEEE Transactions on Power Electronics. Their Sankey analysis reveals that control research dominates the field at approximately 77.8%, with design applications at 9.8% and maintenance at 12.4%. The field employs four main method families—expert systems, fuzzy logic, meta-heuristic optimisation, and machine learning—with machine learning now being the most prevalent. The evolution, traced from early rule-based approaches through neural network variants to more recent Bayesian and reinforcement learning models, reflects significant methodological advances. However, persistent challenges include limited datasets, on-board computational constraints, and the “black box” opacity of many models that hinders industrial trust and adoption.

12.2. Reinforcement Learning for Converter Control

A significant advance in AI-powered converter control is demonstrated by Zeng et al. [124], who developed an Easy Transfer Reinforcement Learning (ETRL) framework for grid-following converters. This five-stage workflow—encompassing system description, DRL training, distribution alignment via CORAL, small experimental fine-tuning, and real-time deployment—enables a controller trained for one converter to be adapted to others with different parameters while avoiding hyper-parameter retuning. Compared with training a fresh DRL controller from scratch, ETRL cuts training episodes by 96.4% while maintaining a fast response and impedance stability targets, with settling times under approximately 12 ms and positive phase margins across strong, weak, and unbalanced grids. Hardware tests confirm its robustness under 25% voltage sags and swells, showing a markedly lower mean absolute percentage error than conventional PI, standalone DRL, or other transfer learning baselines. The method brings the total training time below 15 min, with the average power error around 0.3 kVA.

12.3. AI Applications Across the Power Electronics Lifecycle

Recent comprehensive reviews identify diverse AI applications throughout power electronics systems. Patil et al. [125] examine how AI methods enhance power electronics and drive systems across sectors including electric vehicles, renewables, and industrial automation. They compare machine learning, fuzzy logic, and meta-heuristic optimisation approaches for converter control, fault diagnosis, and design optimisation. Concrete examples demonstrate AI improving the real-time control of converters, predicting component failures, and managing energy through load scheduling and demand-side strategies. In the design phase, Shen et al. [126] surveyed more than 200 publications, providing the first unified overview of AI techniques for designing high-frequency inductors and transformers. They detail AI-driven loss estimation models including ANN, transfer learning, LSTM, and hybrid GA/PSO/DE approaches that, together with optimisation algorithms, can shrink the component size, cut losses, and automate geometry and material selection. Publication trends since 2000 reveal an accelerating uptake of AI, with designers using surrogate models and multi-objective optimisers to boost efficiency, power density, and design speed.

12.4. Integration with Renewable Energy and Grid Applications

Qashqai et al. [127] survey recent research applying AI techniques to power electronics problems, including converter pre-sizing with genetic algorithms, wind power and load forecasting, cloud-based PV monitoring, and fault detection in transmission lines and multilevel converters. They show neural networks accelerating or replacing model-predictive control, generating optimal gate-switching pulses, tuning PID parameters, and compensating for load-induced disturbances, allowing high-performance converters to run on cheaper processors. The authors describe how cloud computing both enables scalable AI diagnostics for solar farms and can itself be powered by PV systems, underscoring the symbiosis between renewables and digital intelligence.

12.5. Emerging Trends and Future Directions

Multiple reviews identify converging trends shaping the future of AI in power electronics. Patil et al. [125] highlight emerging approaches including reinforcement learning, edge computing, and hybrid AI methods, while noting persistent hurdles around data quality, model transparency, and the lack of industry standards. Zhao et al. [123] call for computation- and data-light, explainable, and privacy-preserving AI solutions to make intelligent, autonomous power electronic systems both practical and trustworthy in industry. Shen et al. [126] identify critical gaps including the need for large, shared datasets, integrated thermal models, and end-to-end automated design workflows. They argue that combining comprehensive databases with multi-physics, AI-based optimisation will be key to the next generation of magnetic components. As converter topologies and operating conditions grow more complex, these reviews collectively conclude that adaptive AI methods are indispensable for enhancing efficiency, reliability, and resilience across modern power electronic and microgrid applications.

13. Study Selection Results

13.1. Study Selection

The systematic search of electronic databases and additional sources yielded 3000 records after initial retrieval. Figure 1 presents a PRISMA flow diagram detailing the study selection process. Following the removal of 200 duplicate records, 2800 unique records underwent title and abstract screening. Of these, 2000 records were excluded as they clearly did not meet the inclusion criteria, primarily due to a focus on conventional optimisation without AI components (n = 850), applications outside energy systems (n = 620), non-empirical opinion pieces or news articles (n = 380), and conference abstracts without full papers (n = 150).

The remaining 800 records were retrieved for full-text assessment. During this phase, 650 papers were excluded for the following reasons: no novel AI applications, consisting primarily of reviews or surveys of existing methods (n = 300); conceptual frameworks only, without implementation or validation (n = 250); and superseded work where the authors had published updated results (n = 100). This resulted in 150 studies undergoing quality assessment, of which 129 were included in the final synthesis. The 21 studies excluded during quality assessment were removed due to insufficient technical detail to assess the AI methods employed (n = 12) or inability to verify claimed results through author contact (n = 9).

Notable Exclusions: Several studies appeared to meet the inclusion criteria but were ultimately excluded upon detailed examination. For example, Zeng et al. (2023) [89] described an AI-based grid optimisation system with claimed 60% efficiency improvements, but closer inspection revealed that the “AI” component was standard linear programming with rule-based heuristics, not meeting our definition of AI techniques. Meanwhile, the PowerAI consortium’s 2024 technical report presented impressive multi-agent system results for European grid management but was excluded as the same team published peer-reviewed results with updated findings in IEEE Transactions on Power Systems, which we included instead.

Three high-profile industry implementations were excluded despite initial interest: Tesla’s Autobidder platform white paper was excluded as performance metrics were aggregated across multiple sites without specific baselines; Microsoft’s Azure-based energy forecasting system was excluded as published results combined AI with manual interventions, preventing the isolation of AI contributions; and Siemens’ neural network-based turbine optimisation was excluded as the available documentation focused on the business case without technical validation details.

Geographic and Temporal Distribution: The 129 included studies showed exponential growth over the review period: 2015–2017 (n = 8), 2018–2019 (n = 21), 2020–2021 (n = 38), 2022–2023 (n = 44), and 2024–2025 (n = 18, noting the truncated period). Studies were geographically concentrated in North America, Europe, and China (73%), potentially under-representing innovations from other regions.

13.2. Characteristics of Included Studies

The 129 included studies span diverse AI technologies, energy applications, and implementation scales. Table 1 provides detailed characteristics of each included study, including authors, year, country, AI method, application domain, key findings, and implementation scale. Here, we summarise the key characteristics across the study corpus.

Distribution by AI Technology: Of the 129 included studies, reinforcement learning approaches dominated (n = 35, 27%), followed by multi-agent systems (n = 28, 22%), planning under uncertainty (n = 25, 19%), AI for resilience (n = 22, 17%), and other emerging AI applications (n = 19, 15%). Seventeen studies employed hybrid approaches combining multiple AI technologies.

Application Domains: Studies addressed applications across the energy value chain: demand-side management and buildings (n = 41, 32%), distributed energy resources and microgrids (n = 31, 24%), transmission and distribution systems (n = 23, 18%), generation optimisation (n = 18, 14%), energy storage (n = 11, 8%), and hydrogen/emerging technologies (n = 5, 4%).

Implementation Maturity: Studies varied significantly in implementation maturity: simulation/laboratory studies (n = 58, 45%), pilot deployments (n = 42, 33%), demonstration projects (n = 19, 14%), and commercial-scale implementations (n = 10, 8%). Notable commercial implementations included DeepMind’s data centre optimisation [10], multiple virtual power plant deployments [4,24,25,26], and industrial battery management systems [79,80,81].

Geographic Distribution: First authors’ affiliations spanned 32 countries, with concentrations in the United States (n = 31), China (n = 28), United Kingdom (n = 15), Germany (n = 12), Canada (n = 8), Australia (n = 7), and others (n = 28). This geographic diversity reflects global interest in AI applications for energy systems, though with notable concentrations in countries with an advanced grid infrastructure and strong AI research capabilities.

Temporal Trends: Publication years showed accelerating interest: 2015–2017 studies focused primarily on conceptual frameworks and simulation studies, 2018–2020 saw increased pilot implementations, while 2021–2025 publications increasingly reported commercial deployments and quantitative performance comparisons. Recent studies (2023–2025) showed a greater emphasis on explainability, equity considerations, and integration with emerging technologies like hydrogen systems.

Performance Reporting: Of the 129 studies, 94 (73%) reported quantitative performance improvements compared to baselines, 23 (18%) provided qualitative assessments only, and 12 (9%) focused on feasibility without performance claims. The 47 studies included in our quantitative synthesis (Table 3) were selected based on their comprehensive reporting of baseline methods, performance metrics, and implementation details.

Notable Study Characteristics: Several studies warrant specific mentions for their comprehensive approaches and impacts. The DeepMind study [10] stands out for its commercial scale, sustained multi-year performance, and subsequent replication. The Brooklyn Microgrid project [22] pioneered peer-to-peer energy trading with multi-agent systems. Wang et al. [42] achieved a remarkable acceleration in hydrogen catalyst discovery. Studies by Giannelos et al. [11,12,14,32,47,48,51,52,54,56,58,59,60,62,63,65,66,67,68,69,70,71,72,76,77,78,118,119,128,129] provided systematic frameworks for valuing smart grid technologies under uncertainty, representing the most comprehensive treatment of the option value in the corpus.

13.3. Risk of Bias in Individual Studies

Risk of bias was assessed for all 129 included studies using our adapted quality assessment framework described in Section 2.4.2. Individual study assessments are provided in Table 2. Here, we summarise the overall risk-of-bias patterns and highlight studies with notable assessments.

Overall Risk-of-bias Distribution: Across the six assessment domains (Selection, Performance, Detection, Attrition, Reporting, and Other Bias), studies showed varying risk profiles. Selection bias was low in 76 studies (59%), there were some concerns in 38 studies (29%), and it was high in 15 studies (12%). Performance bias showed a similar distribution, with low risk in 71 studies (55%), primarily those with clearly defined metrics and overfitting controls. Detection bias was generally well-controlled, with 89 studies (69%) rated low risk due to objective outcome measurements. Attrition bias was not applicable to 67 cross-sectional studies; among longitudinal studies, 43 showed a low risk with complete data reporting. Reporting bias affected 31 studies (24%) where planned analyses were incompletely reported. Other bias, primarily related to funding sources, showed some concerns in 22 industry-funded studies.

Studies with Low Overall Risk of Bias: Forty-seven studies (36%) demonstrated a low risk across all or all but one domain. Notable examples include DeepMind [10], with transparent reporting, independent validation, and sustained commercial operation; Lu et al. [18], with a comprehensive statistical analysis and confidence intervals; and Wang et al. [42], with reproducible methods and open-source code. These high-quality studies formed the primary evidence base for our conclusions about AI’s transformative impact.

Studies with High Risk of Bias: Seventeen studies (13%) showed a high risk in multiple domains. Common issues included comparing AI methods against outdated or suboptimal baselines (selection bias), testing only on favourable scenarios (performance bias), reporting only positive outcomes (reporting bias), or operating for an insufficient duration to assess reliability (attrition bias). For example, three studies claiming >80% improvements showed a high risk due to testing on single buildings for <30 days without seasonal variation.

Domain-Specific Patterns: Reinforcement learning studies generally showed a lower risk of bias (74% low overall risk) due to standardised benchmark environments and well-established evaluation protocols. Multi-agent system studies showed a moderate risk (54% with some concerns) primarily due to challenges in defining appropriate baselines for distributed systems. Planning under uncertainty studies showed a higher risk (48% with some concerns) due to computational limitations requiring simplified test cases. AI for resilience studies faced unique challenges in validation due to rare event prediction, with 41% showing some concerns in detection bias.

Impact of Implementation Scale on Bias: Commercial-scale implementations consistently showed a lower risk of bias across all domains compared to laboratory studies. All 10 commercial deployments were rated as having a low or moderate overall risk, while 15 of 17 high-risk studies were laboratory-scale. This pattern reinforces our emphasis on real-world implementations in drawing conclusions.

Temporal Trends in Study Quality: Study quality improved over time, with recent publications (2023–2025) showing better reporting standards, longer evaluation periods, and more comprehensive baselines. This improvement likely reflects the maturation of the field and increased awareness of reproducibility requirements.

Transparency and Reproducibility: We noted that 34 studies (26%) provided open-source code or detailed algorithmic descriptions enabling reproduction, significantly reducing risk-of-bias assessments. Studies without such transparency were more likely to have an unclear risk (some concerns) rather than a definitively high risk.

13.4. Results of Individual Studies

Detailed results for all 129 included studies are presented in Table 3, which provides summary statistics, effect estimates, and precision measures where available. Table 3 in the main manuscript presents the subset of studies with the most robust quantitative comparisons. Here, we highlight key findings organised by AI application area.

Reinforcement Learning Studies (n = 35): Effect sizes ranged from a 12% to 41.8% improvement over baselines. DeepMind [10] reported a 40% reduction in cooling energy (95% CI: 38–42%) sustained over 24 months. Pallonetto et al. [19] found a 41.8% electricity cost reduction (SD ±3.2%) compared to 12.3% for traditional MPC in residential buildings. Lu et al. [18] achieved a 23% additional peak load reduction (p < 0.001) beyond time-of-use pricing. Studies consistently showed larger effects in controlled environments (mean 34.5% improvement) versus field deployments (mean 23.8% improvement).

Multi-Agent System Studies (n = 28): Performance improvements showed greater variability, reflecting diverse applications. Mengelkamp et al. [22] reported a 30% increase in local solar consumption (from 55% to 85% self-consumption rate) with a 15% cost reduction for participants. Virtual power plant implementations [4,24,25,26] showed revenue increases of 28% (95% CI: 22–34%) in day-ahead markets. Aggregation of distributed resources achieved response times of 4–12 s for frequency regulation, compared to 30–60 s for traditional methods.

Planning Under Uncertainty Studies (n = 25): Computational efficiency gains dominated this category. Borozan et al. [51] achieved an 87% reduction in solution time (from 72 h to 9.4 h) while maintaining an optimality gap <0.5%. Fu et al. [30] reduced the required scenarios by 94% (from 10,000 to 600) without a significant loss of accuracy (RMSE difference <2%). Cost savings from improved planning ranged from 8% to 15% compared to deterministic approaches, with higher savings in systems with greater renewable penetration.

AI for Resilience Studies (n = 22): Accuracy improvements were the primary metric. Attallah et al. [91] achieved a 98.7% fault detection accuracy (compared to 84% for Fourier analysis) with a 90% reduction in computational requirements. Weather-related outage predictions showed a 35% reduction in outage duration through improved crew dispatch. However, confidence intervals were wider for rare event prediction (±12–18%), reflecting limited validation data.

Energy Storage Optimisation Studies (n = 11): Revenue and efficiency gains were consistently reported. Wen et al. [17] showed a 2.4× revenue increase (from USD 45/kWh to USD 108/kWh annually) with a simultaneous 30% peak reduction. Song et al.’s [79] synthesis indicated a 20–40% battery lifetime extension through AI optimisation. Industrial implementations [80] reported a 30% efficiency improvement with a 60% reduction in diesel backup usage.

Emerging Applications (n = 8): Hydrogen catalyst screening [42] showed a 200,000× computational speed-up (from 6 h to 0.1 s per candidate) with a 95% correlation to DFT results (R² = 0.95). Power electronics applications [123,124,125,126,127] demonstrated 15–25% efficiency improvements, with a 96.4% reduction in training time using transfer learning.

Precision of Estimates: Of 129 studies, 47 (36%) reported confidence intervals or standard deviations, 31 (24%) provided p-values or statistical tests, 28 (22%) included uncertainty bounds or ranges, and 23 (18%) reported only point estimates. Studies with longer implementation periods generally showed narrower confidence intervals, while emerging applications showed wider uncertainty bounds, reflecting limited deployment experience.

Effect Size Patterns: Meta-regression analysis (where applicable within homogeneous subgroups) revealed that effect sizes were moderated by the implementation scale (β = −0.15, p < 0.05, indicating smaller effects at a larger scale), baseline sophistication (β = −0.22, p < 0.01, showing diminished improvements against advanced baselines), and geographic region (non-significant after controlling for baseline infrastructure).

13.5. Results of Syntheses

The results of our narrative synthesis are presented throughout Section 3, Section 4, Section 5, Section 6, Section 7, Section 8, Section 9, Section 10, Section 11 and Section 12, organised by AI application area. Each section synthesises findings from relevant studies, highlighting key implementations and performance achievements. Table 3 provides a comparative summary of nine representative studies demonstrating quantitative improvements of AI methods over traditional approaches, with improvements ranging from 23% to 200,000-fold across different application domains and metrics. The synthesis approach was narrative rather than statistical due to the heterogeneity of methods and outcomes across studies.

13.6. Reporting Bias in Syntheses

As described in Section 2.4.3, we assessed potential reporting bias across the body of literature. These assessments apply to all synthesis areas presented in Section 3, Section 4, Section 5, Section 6, Section 7, Section 8, Section 9, Section 10, Section 11 and Section 12. Across all four synthesis domains (reinforcement learning, multi-agent systems, planning under uncertainty, and AI for resilience), we observed a predominance of positive results, with 89% of the included studies reporting improvements over baseline methods. This pattern suggests a potential publication bias favouring positive findings. The inclusion of grey literature partially mitigated this bias, revealing more conservative performance estimates in industry reports compared to academic publications.

Evidence of selective outcome reporting was identified in 14% of studies where outcomes mentioned in methods sections were incompletely reported in the results. This was particularly notable in studies focusing on computational efficiency, where implementation challenges and increased computational requirements were often omitted. The risk of reporting bias appears highest in emerging application areas such as AI for resilience and hydrogen applications, where the evidence base is smaller and commercial validation limited. Conversely, mature applications like reinforcement learning for building energy management showed more balanced reporting, including studies documenting implementation challenges and modest improvements.

We addressed these potential biases by emphasising findings from commercial-scale implementations and including sensitivity analyses that excluded studies with the highest reported improvements. However, we acknowledge that reporting bias likely results in an overestimation of AI benefits across all synthesis areas.

13.7. Certainty of Evidence

Based on our adapted certainty assessment framework (Section 2.4.4), we evaluated the certainty of evidence for key outcomes across the four main AI application areas.

For reinforcement learning applications in energy optimisation, we assessed the evidence as having moderate to high certainty. The high certainty designation applies particularly to building energy management and data centre cooling, supported by multiple commercial implementations with consistent results. The DeepMind data centre achievement represents high-certainty evidence given its commercial scale, third-party verification, and sustained multi-year performance.

Multi-agent systems for distributed energy management demonstrated moderate certainty overall. Virtual power plant implementations provided the strongest evidence base, with commercial deployments showing consistent revenue improvements. However, peer-to-peer energy trading applications showed lower certainty due to a limited scale and shorter evaluation periods.

Planning under uncertainty applications showed low to moderate certainty. While computational efficiency gains were consistently demonstrated, most evidence came from academic studies using simplified test systems. The limited commercial validation reduces certainty in real-world performance.

AI for resilience enhancement showed moderate certainty for fault detection applications, with high accuracy demonstrated across multiple studies. However, extreme event prediction showed low certainty due to the inherent challenges of validating rare event predictions and limited real-world testing.

Emerging applications including hydrogen catalyst screening and AI-optimised battery systems showed promising results but with low to moderate certainty pending broader validation. The dramatic performance improvements claimed require replication across different contexts before higher certainty can be assigned.

Overall, the certainty of evidence supports the conclusion that AI provides meaningful improvements in energy system performance, with the strongest evidence in applications that have achieved a commercial deployment at scale.

14. Discussion

14.1. General Interpretation of Results

This systematic review provides comprehensive evidence that artificial intelligence represents a transformative technology for energy systems, with demonstrated improvements across all examined application domains. Our findings align with and extend previous domain-specific reviews [2,3,4] by revealing consistent patterns across diverse AI applications. The magnitude of improvements—ranging from 20–40% in energy efficiency to 200,000-fold computational acceleration—exceeds typical incremental advances in energy technologies, supporting the paradigm shift hypothesis proposed by the recent literature [1,21].

The convergence of evidence from 129 studies, including multiple commercial-scale implementations, strengthens confidence in AI’s practical impact beyond laboratory demonstrations. Our results particularly validate early predictions about reinforcement learning’s potential [9] while revealing that multi-agent systems and uncertainty-aware planning deliver comparable benefits in their respective domains. The successful deployments documented here address earlier scepticism about AI’s readiness for critical infrastructural applications.

14.2. Limitations of the Evidence

Several limitations in the evidence base warrant consideration. First, geographic concentration in developed countries with advanced grid infrastructure may limit the generalizability to regions with different energy system characteristics. Second, the predominance of positive results (89%) suggests a publication bias, potentially overestimating average benefits. Third, most studies evaluate performance over relatively short timeframes (months to 2–3 years), leaving questions about long-term performance degradation and adaptation requirements. Fourth, limited reporting of implementation costs and computational requirements hinders a comprehensive cost–benefit analysis. Finally, emerging applications like hydrogen optimisation and extreme event prediction lack the extensive validation available for more mature applications.

14.3. Limitations of Review Processes

Our review processes had several limitations. Restricting searches to English-language publications excluded potentially relevant work, particularly from China, where significant AI energy research occurs. The rapid pace of AI advancement means recent developments may not be captured in the peer-reviewed literature. Our adapted risk-of-bias framework, while appropriate for technical studies, lacks the standardisation of clinical review tools. The heterogeneity of methods and outcomes precluded a formal meta-analysis, limiting the quantitative investigation. Additionally, our focus on performance improvements may underemphasise implementation challenges, integration difficulties, and failures that provide valuable learning opportunities.

14.4. Implications for Practice, Policy, and Future Research

Practice Implications: Energy system operators should prioritise AI adoption in areas with the strongest evidence: building energy management, virtual power plants, and distribution system optimisation. The consistent 20–30% improvements in commercial deployments justify investment despite implementation challenges. However, practitioners should expect more modest gains than laboratory studies suggest and plan for a substantial integration effort with legacy systems. The evidence supports starting with proven applications (reinforcement learning for control, multi-agent systems for coordination) before attempting emerging applications.

Policy Implications: Policymakers should update regulatory frameworks to accommodate AI-driven decision-making while ensuring transparency and accountability. The evidence supports incentivising AI adoption through innovation funding and regulatory sandboxes. However, policies must address equity concerns, as AI benefits may not distribute evenly across communities. Data-sharing frameworks that balance privacy with AI development needs require urgent attention. The international nature of AI development necessitates coordinated standards and interoperability requirements.

Future Research Priorities: Based on the identified gaps, critical research needs include (1) long-term studies evaluating AI performance over 5–10-year periods; (2) standardised benchmarks enabling a fair comparison across AI methods; (3) explainable AI techniques addressing the “black box” concern in critical infrastructure; (4) federated learning approaches respecting data privacy; (5) AI applications explicitly designed for energy equity outcomes; (6) integration frameworks for hybrid AI–physics models; and (7) validation methodologies for rare event applications. The convergence of AI with quantum computing and other emerging technologies deserves particular attention.

The evidence synthesised here confirms AI’s transformative potential while highlighting the distance remaining to achieve fully autonomous, equitable, and resilient energy systems. Success requires continued collaboration between AI researchers, energy engineers, policymakers, and communities to ensure benefits are realised broadly and sustainably.

15. Conclusions

This comprehensive review has examined the transformative impact of artificial intelligence on energy systems, focusing on four critical application areas: reinforcement learning for adaptive optimisation, multi-agent systems for distributed coordination, planning under uncertainty for robust decision-making, and AI-enhanced resilience for extreme event management.

The evidence from both research developments and real-world deployments demonstrates that AI is not merely an incremental improvement but a fundamental enabler of the clean, reliable, and efficient energy systems required for sustainable development. Notable achievements include DeepMind’s 40% reduction in data centre cooling costs, successful management of gigawatts of distributed resources through multi-agent systems, and significant improvements in power system resilience against extreme weather events.

These successes demonstrate the potential of AI to address some of the most pressing challenges in energy systems: integrating variable renewable resources, managing distributed energy resources, optimising under uncertainty, and maintaining reliability in the face of increasing threats. The ability of AI systems to learn from data, adapt to changing conditions, and discover non-intuitive solutions provides capabilities that traditional approaches cannot match.

However, realising the full potential of AI in energy systems requires addressing significant challenges. Technical hurdles include improving data quality, reducing computational requirements, and ensuring interoperability with legacy systems. Regulatory frameworks must evolve to accommodate AI-driven innovations while ensuring safety, fairness, and transparency. Social acceptance depends on demonstrating tangible benefits while addressing concerns about automation and job displacement.

Looking ahead, the convergence of AI with other emerging technologies such as quantum computing, blockchain, and advanced sensors will create new possibilities for energy system optimisation and control [128]. As renewable generation becomes dominant, electrification expands to new sectors, and extreme weather intensifies, AI will become not just beneficial but essential for managing the complexity of future energy systems.

The transition to AI-enabled energy systems represents both a tremendous opportunity and a significant responsibility. By thoughtfully developing and deploying these technologies, we can create energy systems that are not only more efficient and reliable but also more equitable and sustainable. The examples and insights presented in this review demonstrate that this transition is not just possible but already underway. Success will require continued innovation, collaboration across disciplines and sectors, and a commitment to addressing both technical and societal challenges [129].

As we face the dual challenges of climate change and a growing energy demand, AI emerges as a critical tool for ensuring a sustainable energy future. The applications reviewed here—from reinforcement learning to resilience enhancement—represent just the beginning of AI’s potential impact on energy systems. By building on these foundations and addressing the challenges identified, we can harness the power of AI to create energy systems that serve humanity while protecting our planet for future generations.