Operational Roles of Artificial Intelligence in Energy Security: A Triangulated Review of Abstracts (2021–2025)

Abstract

The operational roles of artificial intelligence in energy security remain inconsistently defined across the scientific literature. To address this gap, the present review examines 165 peer-reviewed abstracts published between 2021 and 2025 using a triangulated methodology that combines trigram frequency analysis, manual qualitative coding, and semantic clustering with sentence embeddings. Eight core roles were identified: forecasting and prediction, optimisation of energy systems, renewable energy integration, monitoring and anomaly detection, grid management and stability, energy market operations/trading, cybersecurity, and infrastructure and resource planning. According to the results, the most frequently identified roles, based on the average distribution across all three methods, are forecasting and prediction, optimisation of energy systems, and energy market operations/trading. Roles such as cybersecurity and infrastructure and resource planning appear less frequently and are primarily detected through manual interpretation and semantic clustering. Trigram analysis alone failed to capture these functions due to terminological ambiguity or diffuse expression. However, correlation coefficients indicate high concordance between manual and semantic methods (Spearman’s ρ = 0.91), confirming the robustness of the classification. A structured typology of AI roles supports the development of more coherent analytical frameworks in energy research. Future research incorporating full texts, policy taxonomies, and real-world use cases may help integrate AI more effectively into energy security planning and decision support environments.

1. Introduction

Artificial intelligence (AI) integration into energy systems has become increasingly strategic amid climate change, geopolitical instability, and the decentralisation of infrastructures. As energy security expands to encompass resilience, cybersecurity, and adaptability, AI’s capacity to manage system complexity is drawing growing interest from the scientific community. While technical studies frequently examine AI applications in various fields, the specific operational functions that AI serves within the broader framework of energy security remain underexplored. For the purpose of this study, energy security—following the definition provided by the International Energy Agency (IEA) [1]—is understood as the uninterrupted availability of energy sources at an affordable cost. In this context, the concept is further extended to include the resilience of energy supply chains and the protection of critical digital infrastructure. It essentially ensures that energy needs are met reliably, sustainably, and without unacceptable disruption for consumers and businesses. Recent years have seen a surge in review articles at the intersection of AI and energy; however, their coverage varies widely in terms of focus, methodological scope, and disciplinary lens.

Numerous studies have explored the technical implementation of AI in widely understood energy management [2], including forecasting [3,4,5,6], optimisation [7], fault diagnostics, and predictive maintenance [8,9,10,11], as well as in control systems [12]. Some reviews offer algorithmic classifications [13] or provide comprehensive overviews of AI applications and their associated challenges [14]. A substantial portion of the literature is dedicated to maintenance and condition monitoring [9], covering aspects such as fault detection [8], fault location [11], fault categorisation, and their cause analysis [10].

The role of AI in energy security has also been investigated in cybersecurity [15,16,17,18,19,20] and supply chain optimisation [20,21], both of which are increasingly recognised as critical components of resilient energy systems. Beyond technical dimensions, broader societal and governance-oriented perspectives are emerging. These include AI predicting power consumption [22] and energy prices [23], assessing environmental performance and enhancing climate resilience [17]. Other contributions address power system adaptability in the face of extreme weather events [24] and sustainability [24,25]. There are also strategic perspectives on national security, climate security, and broader geopolitical issues [18,19].

Although numerous studies explore AI applications in energy systems, thematic fragmentation persists. Operational functions—such as forecasting, optimisation, or cybersecurity—are often referenced yet rarely classified in a structured or comparative manner. Technological innovation and geopolitical framing dominate prior studies, leaving the functional deployment of AI insufficiently articulated in terms of energy security objectives.

Focusing on this gap, the review aims to identify and categorise the operational roles of AI in the context of energy security, based on patterns emerging from empirical data. Rather than surveying the full spectrum of AI-enabled technologies, the analysis centres on how functional roles are positioned and expressed within scientific discourse. To this end, a corpus of 165 peer-reviewed journal articles’ abstracts from Scopus and Web of Science databases (2021 to 2025) formed the empirical basis of the study. The primary method involved trigram-based lexical mining of abstracts to extract high-frequency three-word expressions. To accurately identify the operational roles reflected in these expressions, categorisation was carried out in collaboration with a researcher. To enhance the robustness of the classification, a triangulation strategy was applied, incorporating two additional methods: (1) human interpretive coding of text excerpts and (2) unsupervised semantic clustering using Sentence-BERT and HDBSCAN. The integrative approach enabled the identification of both dominant functions and thematically relevant yet infrequently articulated roles.

The guiding research question is as follows:

What operational roles does artificial intelligence most frequently play in the context of energy security, based on scientific abstracts?

Answering this question supports a systematic synthesis of AI’s functional deployment in energy security discourse, highlighting underrepresented dimensions, particularly those involving strategic foresight or long-term resilience. The results are intended to inform future research trajectories, policy development, and the integration of AI within critical infrastructure systems.

Section 2 presents the dataset and methodological framework. Section 3 reports the findings, including intersections identified through triangulation. Section 4 discusses theoretical and practical implications and cross-sectoral relevance. Section 5 concludes with a synthesis of the study’s contributions.

2. Materials and Methods

The study applied a triangulated analytical process to examine how AI contributes to energy security in scientific research. The analysis involved two stages: constructing a thematically relevant dataset of 165 abstracts and applying three independent methods—trigram-based lexical frequency analysis, manual qualitative coding, and semantic clustering based on language embeddings. Figure 1 illustrates the whole methodological framework.

An auxiliary search of the grey literature (arXiv preprints) was conducted to assess the robustness and broader relevance of the identified operational roles.

2.1. Search Strategy and Review Protocol

To identify the relevant literature on the application of AI in the context of energy security, a structured search was conducted using the Scopus and Web of Science databases, two of the most comprehensive sources of peer-reviewed research publications (as of April 2025). The search targeted original research articles published between 2021 and 2025, written in English, and indexed as journal articles (Scopus: DOCTYPE(ar); Web of Science: DT = Article). The time frame was selected to capture recent developments related to digital transformation, the impact of the COVID-19 pandemic, and growing concerns over geopolitical instability and infrastructure resilience. The review is based on the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) framework [26] to ensure transparency and reproducibility of the search, screening, and inclusion process.

2.2. Selection of Keywords

A single Boolean query was designed to capture publications at the intersection of artificial intelligence and energy security. The query combined general AI-related terms (“artificial intelligence”, “AI”, “machine learning”, “deep learning”) with the core concept of “energy security”, reflecting the thematic scope of the review. For Scopus, the string was as follows: TITLE-ABS-KEY (“artificial intelligence” OR “machine learning” OR “deep learning”) AND TITLE-ABS-KEY (“energy security”) AND PUBYEAR > 2020 AND PUBYEAR < 2025 AND DOCTYPE (ar). For Web of Science, it was as follows: TS = (“artificial intelligence” OR “AI” OR “machine learning” OR “deep learning”) AND TS = (“energy security”) AND PY = (2020–2025) AND DT = (Article). Such an approach ensured consistency in the literature retrieval and provided a coherent dataset for subsequent analysis.

2.3. Inclusion and Exclusion Criteria

Inclusion and exclusion criteria were applied at the query level using Boolean search filters in Scopus and Web of Science. The search was restricted to abstracts and keywords of peer-reviewed journal articles written in English and thematically aligned with the research questions. Publications that did not present original research—such as literature reviews, policy papers, and editorial commentaries—were excluded. Records focused solely on renewable energy forecasting or general smart grid optimisation, without a specific link to energy security, were also removed. The retrieved abstracts served as the basis for subsequent thematic classification and were excluded if they lacked methodological clarity or did not sufficiently address the scope of the review.

2.4. Preparation and Preprocessing

The preparation of records for analysis followed a structured multi-phase process: identification, deduplication, screening, and thematic classification.

In the identification phase, bibliographic records were exported from Scopus and Web of Science in RIS and CSV formats to ensure compatibility with both reference management software and text analysis environments. All records were imported into EndNote version 20, where initial deduplication was performed using the software’s built-in algorithm (based on DOI, title, author, and publication year).

Due to metadata inconsistencies (e.g., missing DOIs, non-standard punctuation, abbreviated names), many duplicates remained undetected. A manual review was therefore conducted, and as a post-validation step, the cleaned dataset was further processed in Python 3.11 [27] using fuzzy string matching (fuzz.token_set_ratio, threshold = 90%) from the fuzzy wuzzy package [28]. The procedure confirmed the completeness and the reliability of the manual deduplication. The final dataset consisted of 204 unique records.

Screening and thematic classification (relevance filtering) were conducted using a combination of automated parsing and manual expert coding. Complete abstracts and keyword fields (both author-supplied and index-supplied) were included, following the normalisation and merging of bibliographic metadata from Scopus and Web of Science. Screening extended beyond titles to ensure comprehensive semantic coverage.

Each record was classified into one of two mutually exclusive categories.

Relevant—publications addressing both artificial intelligence (AI)- and energy security-related topics functionally or systemically.

Not relevant—publications with no meaningful connection to either domain.

The classification strategy was based on controlled keyword matching and refined through a manual review. Two distinct keyword sets were used to identify semantic proximity to the two core domains:

-

Artificial intelligence-related terms:

“artificial intelligence”, “AI”, “AI-based”, “machine learning”, “deep learning”, “neural network”, “predictive model”, “data-driven”, “digital twin”, “smart system”, “forecast”, “algorithm”, “automated”, “classification”, “intelligent”, “cognitive computing”, “fuzzy logic”, “support vector”, “optimisation”, “decision support”, “data analytics”, “data mining”, “computational method”, “modeling”, “learning-based”.

-

Energy security–related terms:

“energy security”, “power grid”, “electricity”, “renewable energy”, “grid stability”, “energy transition”, “supply reliability”, “distribution network”, “resilience”, “storage”, “critical infrastructure”, “infrastructure”, “load forecasting”, “decarbonisation”, “climate”, “sustainability”, “resource”, “demand”, “bioeconomy”, “waste-to-energy”, “energy planning”, “smart energy system”, “system reliability”.

This dual set was designed to reflect domain-specific terminology (e.g., “AI”, “energy security”) with operational descriptors common in applied research (e.g., “data analytics”, “load forecasting”, “resilience”). Such a balanced approach enabled the accurate identification of publications that address the intersection of AI methods and energy-related challenges.

The keyword sets were designed to reflect both domain-specific terminology and the language of applied research. Automated classification was followed by a manual expert review, particularly for borderline cases where relevance was implicit or context-dependent.

To ensure interpretative consistency, the classification process was repeated by the same researcher after three weeks using the same decision rules. Reliability was assessed and the level of agreement was assessed using Cohen’s kappa coefficient and interpreted according to the Landis and Koch scale [29].

Text parsing and keyword matching were implemented in Python 3.11. Final relevance classification was conducted manually, based on an expert review. Abstracts and associated metadata were exported using the Subject Bibliography function in EndNote 20 and manually curated. The final set of 165 relevant records was saved as a structured dataset and used as input for subsequent trigram-based lexical analysis (see Supplementary Material S1_Dataset.xlsx). The full preprocessing script is provided as Supplementary Material S0_Preprocessing_Script.py to ensure transparency and replicability.

2.5. Identifying AI Operational Roles in Energy Security

To answer the research question “What operational roles does artificial intelligence most frequently play in the context of energy security, based on scientific abstracts?”, a triangulated analytical framework was implemented. Three complementary methods were applied independently to the same dataset of 165 scientific abstracts:

• Supervised extraction of high-frequency trigrams;
• Manual coding of meaningful excerpts;
• Embedding-based clustering using contextual language models.

This mixed-method design enabled both frequency-based and context-sensitive identification of AI roles. For selected analytical procedures, the level of agreement was assessed using Cohen’s kappa coefficient and interpreted according to the Landis and Koch (1977) scale.

2.5.1. Lexical Role Identification (Trigram Analysis)

The first method applied systematic text preprocessing and trigram extraction. After cleaning the abstracts (standardising case, removing special characters and stopwords, and excluding domain-specific high-frequency terms such as “AI” and “energy security”), trigrams were generated using a CountVectorizer. A total of 2144 unique trigrams were extracted, from which the top 200 by frequency were selected for analysis (see Supplementary Material S1_Dataset.xlsx).

A preliminary set of operational roles was defined based on an initial sample of abstracts and refined iteratively in consultation with the lead researcher. For each role, a set of associated keywords and representative phrases was compiled. These were adjusted to reflect the specific functions observed in the corpus.

Importantly, abstracts and trigrams were not assigned to roles in a one-to-one fashion. Each abstract or trigram could be associated with multiple roles if it contained keywords relevant to several categories. This many-to-many assignment structure allowed the method to capture overlapping or compound AI functionalities within a single contribution.

To quantify the relative prominence of each role, an importance score (W) was computed as follows:

W = O + α⋅U

Where the following were applied:

-

O: total number of trigram occurrences linked to a role.

-

U: number of unique trigrams linked to a role.

-

α: weight for diversity (set to 3).

This systematic approach enabled the identification and quantification of the most frequent operational roles of AI in energy security, as reflected in the analysed scientific abstracts.

2.5.2. Interpretive Role Coding (Manual)

The second method involved manual qualitative coding using NVivo ver. 12 [30]. From the same dataset, 219 functionally meaningful text segments were selected and coded by a single researcher. Coding focused on identifying the function of AI in context rather than matching keywords.

Each excerpt could be assigned to one or more roles. Role definitions emerged inductively from the data and were not based on pre-existing taxonomies.

The coded segments and assigned categories are available in Supplementary Material S3_Codes.xlsx.

2.5.3. Semantic Role Clustering (Embeddings)

The third method employed unsupervised clustering of 216 excerpts using sentence embeddings (all-MiniLM-L6-v2 model). Embeddings were projected via UMAP and clustered using HDBSCAN. Clusters were interpreted inductively by examining the content of representative excerpts.

Role labels emerged from dominant patterns of expression and action within each cluster (e.g., prediction, optimisation, anomaly detection). A multilabel assignment process was employed: excerpts were compared against all cluster centroids using cosine similarity and could be assigned to multiple roles if they exceeded a similarity threshold.

To ensure semantic coherence, cluster labels were derived inductively, with interpretation guided by the presence of key verbs and nouns indicative of AI functions (e.g., predict, monitor, optimise). To ensure consistency, a protocol was followed in assigning cluster labels, combining verb–object patterns, role-specific keywords, and cosine similarity to representative role vectors (see Supplementary Material S5). This linguistic focus helped anchor role identification in the functional language used across abstracts. Intermediate results are provided in Supplementary Materials S4_Embedding_Clustering.xlsx and S5_Operational_Roles_Multilabel.xlsx. Full excerpt texts grouped by an assigned role appear in S6_Operational_Roles_per_Excerpt.csv.

2.5.4. Triangulation and Statistical Concordance

To enable methodological triangulation, operational roles of AI in energy security were identified using three complementary approaches:

-

Lexical categorisation of frequent trigrams, with role labels refined through iterative validation (Colab notebook);

-

Manual coding of 219 excerpts, with inductively derived role labels based on contextual interpretation;

-

Semantic clustering of the same excerpts using Sentence-BERT embeddings and HDBSCAN, yielding emergent role categories without predefined labels.

A unified typology was developed through iterative alignment of role labels across these methods. To assess consistency between role prominence across approaches, two statistical measures were applied.

-

Spearman’s rank-order correlation, to evaluate concordance in the ranking of role importance across methods.

-

Pearson’s correlation coefficient, to assess the linear association between the number of excerpts or expressions assigned to each role.

High agreement between methods confirmed the robustness of the identified roles. The triangulated approach provided a more comprehensive understanding of the AI operational roles in energy security by integrating scalable lexical analysis with context-sensitive manual coding and semantically informed clustering (

Table 1. Summary of methodological approaches and outputs used for triangulating operational roles of AI in energy security.

Method	Description	Unit of Analysis	Output	Evaluated Metric
Trigram Frequency Analysis (2.5.1)	NLP-based extraction of three-word phrases with role assignment	165 abstracts	Role frequencies (based on trigrams)	Importance score (W) based on unique trigrams (U)
Manual Coding (2.5.2)	Interpretive categorisation of functional excerpts using NVivo, based on the researcher’s contextual understanding	219 excerpts	Single operational role per excerpt	Intracoder reliability (κ = 0.95)
Embedding Clustering (2.5.3)	Sentence-BERT clustering with UMAP + HDBSCAN, inductive labelling	219 excerpts	Operational role categories from clusters	Cluster coherence
Triangulation (2.5.4)	Cross-method alignment and statistical validation	Operational roles across methods	Concordant typology	Spearman’s ρ and Pearson’s r

).

The three methods offered complementary insights into the functional roles of AI in energy security. While trigram analysis prioritised lexical patterns, manual coding emphasised semantic interpretation, and embedding-based clustering uncovered latent thematic structures. Triangulation of results ensured a robust and convergent categorisation scheme.

2.6. Auxiliary Correspondence of Operational Roles in Preprints

To assess the consistency of operational roles beyond peer-reviewed abstracts, an auxiliary search of preprint abstracts was conducted using the arXiv.org repository. The platform was selected for its relevance to artificial intelligence and technical disciplines. The search took place in July 2025, applying the same Boolean expression as in the main analysis: “energy security” AND (“artificial intelligence” OR “machine learning” OR “deep learning”), limited to the years 2020–2025, with cross-listed papers excluded.

The initial query returned 119,174 entries, reflecting broad keyword logic and limited phrase indexing on the platform. Even with cross-listed content excluded and relevance-based sorting applied, the results remained diffuse, which hindered the systematic review and necessitated further refinement of the query. Narrower combinations such as “energy security” + “artificial intelligence” were used, resulting in 79 documents selected for manual screening. Based on thematic proximity, 39 preprints were chosen for full-text analysis.

Two researchers independently coded all documents using the operational role typology from the main study. Intercoder agreement was assessed with Cohen’s kappa, and inconsistencies were resolved through discussion. The goal was to verify whether additional operational roles appeared in the grey literature that were not captured in the peer-reviewed corpus.

3. Results

3.1. Literature Identification and Selection

This section presents the outcomes of each phase of the literature screening process, aligned with the PRISMA framework: identification and deduplication, screening and eligibility assessment, and inclusion.

3.1.1. Identification and Deduplication

A total of 293 publications were retrieved from Scopus (n = 152) and Web of Science (n = 141) covering the years 2021–2025. After removing 90 duplicate entries based on DOI matching and exact title-year similarity (52 entries automatically removed, 38 manually), 203 unique records remained for abstract screening.

3.1.2. Screening and Eligibility Assessment

The abstracts of all 203 records were assessed using a combination of keyword matching and manual thematic evaluation. In total, 38 records were excluded due to insufficient thematic alignment. Examples of papers considered relevant include studies on licensing documentation for advanced reactors [31]; an advanced neural network-based model for predicting court decisions on child custody [32]; a GIS-based design of water quality monitoring networks [33]; a linguistic analysis of climate discourse [34]; multilingual hate speech detection using AI [35]; optical fibre vibration signal recognition [36]; public discourses and governmental interventions concerning carbon neutrality goals [37]; AI used in the context of geoscientific data ecosystems and sustainability platforms [38]; legal text basic element identification [39]; a deep age-invariant fingerprint segmentation system [40]; vehicle classification and license plate recognition [41]; asbestos–cement roofs’ deterioration states [42]; bleached coral detection with deep learning [43]; the posthuman abstract: AI, dronology, and becoming alien [44]; a deep learning approach to investigating clandestine laboratories using a GC-QEPAS sensor [45]; enhancing river health monitoring [46]; and mussel culture monitoring [47]. Particular attention was given to ambiguous or borderline cases, where AI- and security energy-related concepts appeared in non-obvious configurations, e.g., in [48,49], or where energy security was mentioned without functional elaboration, leading to exclusion [38,46]. As a result, 165 publications were classified as relevant to the intersection of artificial intelligence and energy security and retained for further analysis.

3.1.3. Inclusion—Final Dataset

The 165 relevant publications constituted the final dataset used in the trigram-based thematic analysis described in subsequent sections. A summary of the inclusion process is presented in

Table 2. Screening and inclusion summary.

Stage	Count
Records retrieved from Scopus	152
Records retrieved from Web of Science	141
Total records retrieved	293
Duplicates removed automatically	52
Duplicates removed manually	38
Unique records after deduplication	203
Records excluded after screening	38
Final records included	165

.

To ensure the consistency of the manual selection process, intracoder reliability was assessed by repeating the relevance screening and thematic categorisation after a three-week interval. The resulting Cohen’s kappa coefficient of 0.95 indicates a very high level of agreement, confirming the interpretative stability of the classification decisions.

In addition to documenting the final dataset, the temporal distribution of publications provides further insight into the evolving academic interest in the intersection of artificial intelligence and energy security. As shown in

Table 3. Annual number of AI and energy security publications (2021–2025).

Year	Publications
2021	10
2022	14
2023	37
2024	64
2025	40
Total	165

, the number of relevant articles increased consistently from 2021 to 2024, reaching a peak of 66 publications. Although the 2025 count is lower, this likely reflects incomplete indexing at the time of data collection. Overall, the upward trend confirms a growing research focus on AI-driven approaches within energy security discourse over the analysed period.

3.2. Lexical Categorisation Results (Trigram Analysis)

The first method addressed in this triangulated study involved lexical frequency analysis based on trigram patterns extracted from scientific abstracts. Recurrent phrases were identified and associated with the preliminary operational roles of artificial intelligence in the energy security domain. To ensure that lexical categories captured functionally meaningful distinctions, role labels were developed inductively and refined through iterative consultation with a researcher familiar with the domain. The resulting categorisation served as the basis for comparing lexical prominence across roles in the corpus.

The highest importance score (W = 415) was attributed to energy market operations/trading, supported by 53 unique trigrams (U) and 256 total occurrences (O). Frequently recurring expressions included energy load forecasting, carbon price prediction, and grid operation, reflecting a strong research emphasis on market responsiveness and real-time balancing.

Energy storage (W = 410, U = 52, O = 254) and optimisation of energy systems/operations (W = 402, U = 51, O = 249) also ranked high in lexical prominence, featuring diverse terminology across abstracts.

By contrast, renewable energy integration (W = 49, U = 5) and forecasting/predicting (W = 13, U = 2) appeared less frequently but remained identifiable through specific trigrams such as energy load forecasting. Their lower scores suggest narrower or more context-specific lexical usage.

Marginal categories included cybersecurity (W = 6, U = 1) and monitoring/anomaly detection (W = 6, U = 1), each associated with a single vague trigram—security issues related and parameter monitoring technology, respectively—providing insufficient thematic clarity for robust classification. Grid management and stability, as well as the planning of energy resources/infrastructure, recorded zero trigram matches, indicating a complete absence of lexical signals in the dataset.

Table 4. Summary of identified AI operational roles in energy security based on meaningful trigram and abstract analysis.

Operational Role	Exemplary Trigrams (up to 3)	Unique Trigrams (U)	Total Occurrences (O)	Importance Score (W)	Example Publications
Energy market operations/trading	[“energy load forecasting”, “carbon price prediction”, “grid operation”]	53	256	415	[23,50,51,52,53,54,55,56,57,58,59]
Energy storage	[“energy load forecasting”, “carbon price prediction”, “grid operation”]	52	254	410	[60,61,62,63,64]
Optimisation of energy systems/operations	[“grid operation”, “system management”, “energy management”]	51	249	402	[62,65,66,67,68,69,70,71]
Renewable energy integration	[“grid operation”, “system management”, “energy management”]	5	34	49	[70,72,73]
Forecasting/predicting	[“energy load forecasting”, “carbon price prediction”]	2	7	13	[23,52,55,56,57,58,59,74]
Cybersecurity	[“security issues related”]	1	3	6	[75]
Monitoring/anomaly detection	[“parameter monitoring technology”]	1	3	6	[76]
Grid management and stability	[“grid operation”, “system management”, “energy management”]	0	0	0
Planning of energy resources/infrastructure	[“grid operation”, “system management”, “energy management”]	0	0	0

presents a summary of the eight operational roles of AI identified through trigram analysis, including exemplary trigrams, the number of unique trigram patterns (U), total occurrences across the corpus (O), and a composite importance score (W) reflecting lexical salience.

The distribution of importance scores highlights a prevailing focus on system performance, optimisation, and market-oriented control. In contrast, areas related to long-term planning, infrastructure resilience, and digital threat mitigation are weakly represented in abstract-level lexical patterns.

Lexical absence, however, should not be interpreted as conceptual exclusion. Some operational roles may be discussed using ambiguous or inconsistent terminology, particularly in concise or abstract narratives. For instance, the word “security” appeared frequently, but primarily in the context of energy security as a policy objective, rather than digital protection. Disentangling general framing from technically grounded operational references remains essential when interpreting lexical trends in AI-related energy research.

3.3. Interpretive Coding Results (Manual Analysis)

The second method applied in this study involved manual qualitative coding of 219 excerpts, selected to represent the most frequently reported applications of artificial intelligence in energy security. Operational roles were derived inductively and categorised based on their thematic and functional coherence, as reflected in the excerpted content. The resulting classification includes seven distinct roles, summarised in

Table 5. Operational role categories of AI in energy security derived from manual qualitative coding of 219 excerpts.

No.	Operational Role	No. of Excerpts	Representative Excerpt
1	Forecasting and prediction	48	“model for real-world photovoltaic power generation forecasting” [68], “a statistical model that forecasts the power generation at the PV cell plant” [77], “consumption forecasting of oil and coal” [78], “forecasting natural gas consumption” [6], “energy security prediction and early warning mechanism” [79].
2	Optimisation of energy systems	42	“proximal policy optimisation algorithm is figured out” [70], “optimise energy production and distribution” [80], “optimising gasification processes” [66].
3	Renewable energy integration	33	“intelligent integration of new energy storage with the source, grid, and load” [64], “in-depth thermodynamic analysis and optimisation of an integrated renewable energy system” [66,81].
4	Monitoring and anomaly detection	29	“instability monitoring and adaptive error correction” [82], “gas drainage parameter monitoring technology” [76], “fault diagnosis” [83].
5	Grid management and stability	24	“maintaining grid stability” [84], “overcomes the impact of theft attacks on the smart grid” [85].
6	Energy market operations	23	“a strategic framework for managing energy trade complexity” [86], “building industry with optimal tradeoff strategies between energy consumption and thermal comfort of built environment” [87].
7	Cybersecurity	20	“implementing layered AI-based cybersecurity measures to defend smart energy systems” [88], “overcomes the impact of theft attacks on the smart grid” [85].

, along with representative quotations.

The most frequently identified roles were forecasting and prediction, represented in 48 excerpts, followed by the optimisation of energy systems (42 excerpts) and the integration of renewable energy (33 excerpts). Other recurring roles included monitoring and anomaly detection (29), grid management and stability (24), and energy market operations (23). Cybersecurity was the least represented category, with 20 excerpts coded under this theme.

The overall distribution indicates a strong emphasis on technical and performance-oriented applications of AI, particularly those related to prediction, optimisation, and integration of renewables. In contrast, system protection and resilience-enhancing roles, such as cybersecurity and grid stability, received comparatively less attention in the reviewed abstracts.

The number of excerpts reflects unique coded instances per category. While individual excerpts may have touched on multiple aspects of AI use, each was assigned to its most salient operational function.

To achieve reliability, the selection of excerpts and categorisation was repeated after three weeks, and the level of agreement was assessed using Cohen’s kappa (κ = 0.78), which indicates substantial agreement according to the Landis and Koch [29] scale.

3.4. Semantic Clustering Results (Embedding Analysis)

The semantic clustering procedure resulted in eight distinct operational roles assigned across 219 excerpts. The most prevalent category was forecasting and prediction (32 excerpts), followed by energy market operations (24) and renewable energy integration (19). These three accounted for the majority of the AI use cases identified in the corpus, emphasising the centrality of anticipatory and integrative functions in energy systems.

Mid-frequency roles included infrastructure and resource planning (15), optimisation of energy systems (13), and monitoring and anomaly detection (7). Fewer excerpts were associated with cybersecurity (7) and grid management and operations (5), reflecting more specialised but still recurrent themes.

The distribution of roles confirms the functional diversity of AI in the energy domain, ranging from predictive analytics to real-time system oversight.

Table 6. Operational roles of AI derived from Sentence-BERT embedding clustering (n = 219 excerpts).

No.	Operational Role	No. of Excerpts	Representative Excerpt
1	Forecasting and prediction	32	“wind speed forecasting” [89], “prediction of carbon price” [59], “crude oil price forecasting” [52], “enhancing electricity demand forecasting and optimising power grids to reduce energy losses” [90].
2	Energy market operations	24	“the impact analysis of wind energy on electricity prices” [91], “prediction of carbon price” [59].”
3	Renewable energy integration	19	“a novel framework for enhancing renewable energy systems” [92].
4	Infrastructure and resource planning	15	“development of ‘the new infrastructure” [93].
5	Optimisation of energy systems	13	“enhancing electricity demand forecasting and optimising power grids to reduce energy losses” [90], “optimise renewable energy deployment” [92].
6	Monitoring and anomaly detection	7	“resilience of power grid operations” [94].
7	Cybersecurity	7	“a secure energy trading mechanism based on blockchain technology” [95], “overcomes the impact of theft attacks on the smart grid” [85].
8	Grid management and operations	5	“aids in grid operation” [77].

presents the complete role typology with the number of associated excerpts and representative examples for each role.

Each excerpt could be associated with multiple roles based on its semantic proximity to cluster centroids. The resulting frequencies reflect all valid assignments, not mutually exclusive categories. The overall distribution closely aligns with the interpretive coding results, reinforcing the consistency of operational role categories across independent analytical procedures.

To assess intracoder reliability, the coding procedure was repeated by the same researcher after a three-week interval using the original decision rules. Cohen’s kappa was calculated at κ = 0.87, indicating a high level of interpretative consistency across coding rounds.

3.5. Triangulated Role Frequencies and Statistical Concordance

To consolidate findings from the three analytic procedures, all operational roles were aligned into a shared typology encompassing eight core categories. Each method contributed a distinct frequency profile, reflecting its analytical emphasis.

Table 7. Frequency of operational roles identified across all three methods.

Operational Role	Trigram Analysis (U)	Manual Coding	Semantic Clustering
Forecasting and prediction	2	48	32
Optimisation of energy systems	51	42	13
Renewable energy integration	5	33	19
Monitoring and anomaly detection	1	29	7
Grid management and stability	–	24	5
Energy market operations/trading	53	23	24
Cybersecurity	1	20	7
Planning of energy infrastructure	–	–	15

Note: U = number of unique meaningful trigrams assigned to the role.

presents the distribution of roles identified through trigram analysis, manual coding, and semantic clustering.

According to Table 7, optimisation of energy systems was highly represented in trigram analysis (51) and manual coding (42), though less so in semantic clustering (13). Energy market operations/trading showed similarly high trigram frequency (53) but lower manual coding (23), indicating strong lexical presence but limited contextual elaboration. Forecasting and prediction, despite a low trigram count (2), ranked highest in manual coding (48) and scored high in clustering (32), underscoring its thematic importance.

Renewable energy integration and monitoring and anomaly detection were more prominently identified via manual coding (33 and 29, respectively) and semantic clustering (19 and 7) but showed low trigram frequency (5 and 1). Grid management and stability were absent in trigram results (0) yet recognised through manual coding (24) and clustering (5). Cybersecurity remained marginal across all methods (1, 20, 7). Planning of energy infrastructure was absent in the trigram and manual results but appeared in the semantic clustering (15), suggesting implicit textual expression.

To facilitate visual comparison of the role frequencies across analytical methods, a heatmap is presented in Figure 2. The visualisation highlights shared areas of emphasis, such as forecasting and prediction and optimisation of energy systems, as well as methodological divergence for roles like cybersecurity and infrastructure planning.

In the chart, the number of role occurrences identified through trigram analysis, manual coding, and semantic clustering is shown. The visual contrast highlights convergence for core roles and discrepancies for less lexically explicit categories.

As shown in both Table 7 and Figure 2, specific operational roles were exclusive to individual methods. For example, infrastructure and resource planning appeared only in semantic clustering, likely due to the conceptual complexity and lexical variability of associated expressions, which limited their detection through trigram analysis or manual coding.

To assess methodological agreement, correlations were computed between the frequency vectors for each pair of methods (n = 8 categories). Spearman’s rank-order correlation measured concordance in the ranking of role importance, while Pearson’s r assessed linear agreement in absolute frequencies.

Table 8. Correlation coefficients between classification methods.

Comparison Method Pair	Spearman ρ	p-Value	Pearson r	p-Value
Manual Coding vs. Trigram	0.68	0.07	0.73	0.04
Manual Coding vs. Clustering	0.91	0.002	0.88	0.004
Trigram vs. Clustering	0.66	0.08	0.57	0.13

Note: Correlation computed on frequency values across 8 operational roles.

summarises the results.

The strongest agreement emerged between manual coding and semantic clustering, both in terms of the rank order (ρ = 0.91, p = 0.002) and linear association (r = 0.88, p = 0.004), suggesting that both methods captured similar semantic structures, albeit through different mechanisms. Trigram analysis showed moderate correlation with manual coding (ρ = 0.68, p = 0.07) and weaker alignment with clustering. The discrepancy reflects its reliance on lexical surface features, which may not fully capture deeper functional meanings. Although the category definitions were conceptually aligned, frequency-level concordance remained only moderate to weak, indicating that the perceived role salience is sensitive to the method of extraction.

Triangulated evaluation across all three methods strengthened the construct validity by confirming recurring operational roles despite analytical divergence.

3.6. Correspondence of Operational Roles in Preprints

The supplementary screening of grey literature abstracts confirmed the stability of operational roles across publication types. Among the 39 preprints selected for full-text analysis, 20 were unpublished and 19 had already appeared in indexed journals or conference proceedings. The distribution of roles in this corpus aligned closely with the categories derived from peer-reviewed abstracts.

Intercoder agreement was substantial (Cohen’s κ = 0.78), confirming the reliability of role assignments across sources.

All the reviewed abstracts corresponded to previously identified operational roles, such as forecasting and prediction [96,97] and optimisation of energy systems [98,99]. It should be noted that for renewable energy integration [100] and monitoring and anomaly detection [101], the only examples retrieved through the grey literature search had already been published in indexed venues, thus reinforcing the thematic convergence across formal and informal sources. Additional categories such as management and stability [102] and energy market operations [103] are clearly represented and align with the core typology. By contrast, planning of energy infrastructure appears predominantly in combination with cybersecurity-related themes [104]. No additional role categories emerged. This consistency suggests that the proposed typology not only captures the main operational functions evident in peer-reviewed abstracts but also reflects broader patterns in informal or less formally structured scientific outputs. However, roles that were underrepresented in abstract-based lexical analysis—such as cybersecurity—were also infrequently addressed in the repository, reinforcing the notion that some strategic functions may require richer textual contexts to be fully articulated.

3.7. Limitations

The methodological framework employed in this study combined lexical, interpretive, and semantic approaches to identify the operational roles of AI in energy security. While triangulation enhanced the robustness of findings, several limitations should be noted.

First, the trigram-based analysis relied on the 200 most frequent lexical patterns extracted from preprocessed abstracts. This frequency-driven method prioritised commonly recurring expressions and may have overlooked low-frequency but functionally important roles. The decision to exclude the grey literature, including preprints and industry reports, was made to maintain methodological consistency and focus on peer-reviewed evidence. However, this may have limited the inclusion of cutting-edge or practice-oriented insights. Also, the trigram-based lexical analysis prioritised high-frequency patterns, which may have limited the detection of less frequent but functionally important operational roles. Future studies could explore alternative NLP techniques such as topic modelling (e.g., LDA) or full-text embedding approaches to enhance the detection of marginal functions. Moreover, the exclusion of studies that addressed topics such as smart grid optimisation or renewable energy forecasting—without explicitly referencing energy security—may have led to the omission of marginal but relevant cases. Future research should consider conducting sensitivity analyses to assess the impact of such exclusions on role identification.

Second, both the manual coding and semantic clustering were performed on a subset of 219 text excerpts, sampled from abstracts rather than full texts. This choice limited the depth of the contextual information available for interpretive analysis.

Third, thematic convergence across methods was conceptually strong but statistically mixed. Manual coding and semantic clustering demonstrated high alignment (Spearman’s ρ = 0.91), whereas lexical frequency analysis showed a weaker correlation with both methods (ρ = 0.68 and ρ = 0.66, respectively). These discrepancies reflect fundamental differences in how role salience is determined: frequency-based methods emphasise surface repetition, while interpretive and semantic approaches prioritise contextual meaning.

Fourth, the analysis was limited to peer-reviewed journal articles indexed in two major databases (Scopus and Web of Science) and published within five years (2021–2025). Despite careful preprocessing, some degree of lexical ambiguity and thematic spillover may persist, particularly in abstract-based analyses. Roles such as cybersecurity and planning of energy infrastructure, which often require more detailed exposition, may therefore be underrepresented due to the limited scope of abstract. Due to the timing of data collection (April 2025), indexing for the most recent year was incomplete, which may have resulted in the underrepresentation of emerging research trends.

Regarding the specific period, it should be stressed that indexing of 2025 publications may not have been complete at the time of data collection, yet the search included all available records from Scopus, Web of Science, and preprint repositories, ensuring coverage of the most up-to-date research accessible at that point.

4. Discussion

The triangulated analysis identified eight core operational roles of artificial intelligence in energy security: forecasting and prediction, optimisation of energy systems, renewable energy integration, monitoring and anomaly detection, grid management and stability, energy market operations/trading, cybersecurity, and infrastructure and resource planning. While several roles were consistently captured across all methods, others were less uniformly represented. Forecasting and prediction, optimisation of energy systems, and energy market operations/trading were frequently identified across all three analytical procedures, suggesting that these roles are widely established in current research discourse. In contrast, planning of energy infrastructure was identified exclusively through semantic clustering, with no supporting evidence in trigram analysis or manual coding. Roles such as cybersecurity and grid management and stability were also underrepresented in the lexical analysis, despite being captured through interpretive and semantic techniques. These findings suggest that surface-level frequency approaches are insufficient for capturing roles that rely on semantically diffuse or technically complex language. The roles associated with long-term planning, strategic coordination, or cross-sectoral interdependence are particularly prone to underrepresentation in analyses based solely on abstract-level text.

Each method brought unique advantages. Trigram analysis effectively highlighted roles with consistent lexical patterns, such as energy market operations/trading and optimisation of energy systems. Manual coding enabled the identification of roles based on functional content rather than phrasing, with strong representation of forecasting and prediction, monitoring and anomaly detection, and grid management and stability. Semantic clustering, by contrast, was most effective in revealing planning of energy infrastructure and other roles embedded in less formalised or more variable linguistic structures. The strong correlation between manual coding and semantic clustering (Spearman’s ρ = 0.91) demonstrates the value of context-aware methods in identifying functionally relevant roles. The limited convergence between trigram analysis and the other two approaches (ρ = 0.68, p = 0.07) reinforces that the methods capture distinct representational layers. As such, they should be regarded as complementary, not as validating the same underlying construct.

The overall pattern highlights a broader limitation in the literature: roles that are strategically important but less standardised linguistically—such as cybersecurity or planning of energy infrastructure—tend to receive less attention, not due to their irrelevance, but because they are more difficult to isolate through standard keyword-based methods. Abstracts often prioritise technical outcomes, omitting references to system-level functions, uncertainty management, or long-term coordination. Incorporating full-text content, particularly the methodology and discussion sections, or even extending analysis to white papers and regulatory documents may improve detection of these less visible yet operationally critical roles.

The auxiliary analysis of preprint abstracts confirmed that the identified operational roles are not limited to peer-reviewed sources. Forecasting-, optimisation-, and market-related functions appeared consistently across the grey literature, with no additional role categories emerging. This thematic convergence reinforces the stability of the proposed typology across publication types. Future research could expand this line of inquiry by examining whether less prominent roles—such as cybersecurity or infrastructure planning—are more explicitly addressed in full-text preprints or other unstructured sources.

It should be noted that the auxiliary analysis could incorporate insights from white papers and policy documents. These documents offer important strategic perspectives on the operational relevance of AI in energy systems. Further examination of this material may also clarify the context of underrepresented roles, including cybersecurity and infrastructure planning, which tend to require more extensive textual coverage to be fully articulated. For instance, the Fraunhofer-TenneT report Generative Artificial Intelligence in the Energy Sector (2024) [105] outlines real-world applications of generative AI for grid forecasting, asset management, and stakeholder communication, aligning with operational roles such as forecasting and planning. The IEA report Energy and AI (2025) [106] highlights predictive maintenance, grid optimisation, and demand-side flexibility as domains where AI enhances system security and resilience. The RAND report The Use of AI for Improving Energy Security offers a quantitative mapping of AI applications in the electricity sector, identifying roles in forecasting, grid stability, and anomaly detection. The U.S. Department of Energy strategy document [107] emphasises AI’s contribution to sustainability, long-term infrastructure planning, and cybersecurity. At the EU level, the AI Act (2024) [108] establishes risk governance principles applicable to critical infrastructure sectors, indirectly supporting roles related to reliability and control. While the European Commission’s AI White Paper (2020) [109] remains a strategic vision without regulatory force, it introduced the early framing of trustworthy AI in energy contexts. Industry-level perspectives, such as the World Economic Forum report [110], discuss AI for grid decarbonisation and operational coordination. The Linux Foundation’s open-source initiative (2025) [111] advocates for shared AI tools in grid integration, supporting transparency and scalability—key attributes for the operational roles related to optimisation and anomaly detection. To fully capture the strategic dimensions emerging from policy and industry sources, a dedicated framework for their systematic classification would be required as an avenue for further research.

The issue of the underrepresentation of cross-disciplinary and strategic AI functionalities in energy security research calls for both methodological and conceptual advancements. The operational roles should be explicitly tied to tangible system challenges, such as blackout prevention, cyberattack mitigation, and the planning of secure infrastructure. Using ontology-based mappings to connect AI methodologies with security-critical functions can facilitate the development of more precise and resilient systems. Such an approach would be highly valuable for further studies.

A valuable future direction would be to examine how the operational roles of AI in energy security have evolved over time, particularly in response to global disruptions or technological shifts. However, such a temporal perspective would require a different methodological design. The current approach is cross-sectional and based on aggregated representations—trigram patterns, manually coded segments, and semantic clusters—which are not directly linked to publication years. Temporal segmentation could introduce statistical noise due to small sample sizes and inconsistent terminology across years. A robust diachronic analysis would likely require full-text datasets with structured time-aligned annotations or longitudinal topic modelling to trace role evolution with sufficient granularity. Such an approach would represent a new methodological pathway for future research.

Practical implications could concern the better alignment of identified operational roles of AI with established policy and regulatory frameworks, such as the European Union’s REPowerEU plan [112], the NIS2 Directive on network and information security [113], and the International Energy Agency’s Energy Sector Resilience Indicators [114], which would help ensure that AI applications address real-world energy security needs, rather than isolated technical objectives. A more system-oriented research perspective is needed, one that considers regulatory constraints, long-term infrastructure planning, and the role of AI in scenario-based risk assessment.

Finally, insights from other safety-critical sectors may help identify trajectories for the formalisation and broader adoption of operational roles. Given the growing use of AI in domains such as transportation [115,116,117], a comparative review across the security and safety landscape could further strengthen the applicability and generalisability of the proposed typology.

5. Conclusions

The study examined the operational roles of artificial intelligence in 165 scientific peer-reviewed abstracts published between 2021 and 2025. A triangulated methodology, combining trigram-based lexical analysis, manual qualitative coding, and semantic clustering, was applied.

The main conclusions are as follows:

• AI has gained increasing visibility in energy security research, reflecting its expanding role in managing complexity and supporting the functions of energy systems.
• Across all three methods, the most frequently identified roles involve functions such as forecasting and prediction, optimisation of energy systems, energy market operations/trading, and renewable energy integration/trading, indicating a strong emphasis on operational performance and system management.
• Roles related to infrastructure planning, grid management and stability, and cybersecurity appeared less prominently in the abstract-level data, particularly in lexical analyses, suggesting that such roles may be expressed in more varied or implicit ways.
• Methodological triangulation confirmed the consistency of the core role categories, while also demonstrating that the prominence of specific roles varies depending on the analytical lens. Lexical frequency methods emphasised standardised phrasing, whereas semantic and interpretive approaches captured more context-dependent functions.
• The current literature describes AI primarily through its operational capabilities. Future research may benefit from engaging with full texts and applying structured taxonomies to capture a broader spectrum of AI applications in energy security, including those related to resilience, planning, and cross-sectoral coordination.