Resilience of Microgrids to Extreme Weather Events: A Bibliometric Analysis and Review of Control Strategies (2016–2025)

Abstract

The increasing frequency of high-impact, low-probability climate events has highlighted the limitations of conventional reliability criteria, including N-1 planning assumptions, and the need for dynamic resilience architectures in electrical systems. This article analyzes the evolution, trends, and technological challenges associated with microgrid resilience under extreme-weather disruptions. The study adopts a hybrid research design that combines quantitative bibliometric mapping of 283 Scopus-indexed article records for 2016–2025 using CiteSpace version 7 with a structured technical synthesis of the selected literature. The structural analysis identified nine thematic clusters and indicated a transition from service restoration and component-level recovery toward multi-energy energy-management systems, resilient distribution-system planning, and mobile restoration resources. The technical synthesis shows that modern resilience is increasingly associated with hierarchical control architectures: optimization and forecasting methods support tertiary-level energy management, while grid-forming inverters can provide primary-layer voltage references that support islanded operation and black-start sequences under appropriate design, protection, and validation conditions. The article concludes that future research should bridge stochastic planning and real-time physical operation through standardized interoperability frameworks, reproducible dynamic metrics, and experimentally validated control strategies for autonomous critical microgrid operation under extreme-weather conditions.

1. Introduction

The operational vulnerability of power systems stems, in part, from their legacy of deterministic planning based on “plausible” single-fault contingencies, as well as on lists of contingencies and extreme events formalized in planning standards such as TPL-001 [1,2]. In this article, N-1 refers to the conventional planning criterion according to which the system should withstand the loss of one credible component without unacceptable service interruption; by contrast, high-impact, low-probability (HILP) events can produce multiple correlated outages that violate this independence assumption. Recent evidence shows that extreme weather events impose common-mode faults and simultaneous damage to transmission and distribution assets, extending the duration and severity of outages [3,4], and requiring resilience approaches that capture impact-degradation-recovery phases [5], as well as specific metrics for HILP [6]. At the regulatory level, explicit recognition of the inadequacy of conventional frameworks in the face of concurrent failures caused by extreme heat/cold has led to requirements for performance analysis regarding severe weather events and their potential to cause cascading outages [7,8]. In operational terms, the North American Electric Reliability Corporation (NERC) reports that severe weather continues to pose the greatest threat to the Bulk Power System/Bulk Electric System (BPS/BES) and is associated with the most severe outage events observed in a recent reference year [9,10]. This mismatch between N-1 assumptions and HILP disturbances [11,12] facilitates propagation pathways where overloads, protective trips, and loss of operating margin amplify initial losses into cascading sequences.

Cases such as Superstorm Sandy in 2012 on the U.S. East Coast, which disrupted service to nearly 2 million customers in New Jersey [13]; severe snow and ice storms in the northeastern U.S. and eastern Canada in the early 2010s, causing widespread blackouts [14]; Supercyclone Phailin in India in 2013 [15]; Typhoon Haiyan in the Philippines in 2013, devastating the electrical infrastructure [16]; the extreme snowfall in Slovenia in 2014, affecting more than 100 power stations [17]; Tropical Cyclone Idai in Malawi, Mozambique, South Africa, and Zimbabwe in 2019, leaving more than one million people without electricity and triggering rolling blackouts in South Africa [17]; extreme heat in Australia in 2019, requiring forced blackouts to maintain operations; Typhoon No. 15 in Japan in 2019, causing power outages in nearly 935,000 households in the Kanto region [18]; Hurricane Elsa in Barbados and Saint Lucia in 2021, which caused power outages across the island of Barbados and in 90% of households in Saint Lucia [17]; and the intense windstorms in Europe between 2013 and 2016, causing multiple outages [19,20], illustrate how a single high-impact, low-probability weather event can induce simultaneous, large-scale degradations in generation availability, system operation, and supply, triggering emergency actions, cascading failures, and widespread service outages [21,22]. These events underscore the inadequacy of traditional N-1 planning approaches and highlight the pressing need to develop advanced resilience strategies capable of managing the complexity and uncertainty inherent in extreme weather events [23,24,25].

In this context, a microgrid is defined as a set of distributed loads and energy resources within defined electrical boundaries that functions as a controllable entity and can be connected or disconnected to operate in both grid-connected and islanded modes [26]. The shift from “service continuity” to “resilience” in microgrids depends on multi-layered hierarchical control architectures (primary, secondary, and tertiary), which simultaneously enable local stability, restoration of electrical variables, and coordination of exchanges and operational objectives with the main grid [27]. At the converter level, grid-forming (GFM) strategies implement a voltage-source behavior capable of constructing a waveform and maintaining voltage/frequency reference, enabling stable operation in weak grids and contributing to capabilities such as restarting after blackouts, with a well-documented technical roadmap for deployment in systems with high penetration of inverter-based resources [28,29]. In contrast, grid-following (GFL) strategies typically operate as current sources that synchronize against an external reference, so their ability to sustain islanded operation depends on the existence of a firm voltage source, a limitation highlighted in the anti-islanding and interconnection literature [29]. At the supervisory level, the energy management system (EMS) dispatches resources, balances power and energy, and coordinates constraints for autonomous or grid-connected operation, including seamless transitions between operating modes [30,31]. Microgrid controllers also include transition functions for planned and unplanned islanding, reconnection, and black-start, while black-start is defined as the ability to initiate self-generation and sequentially re-energize the microgrid as an island in the absence of a grid connection, relying on an auxiliary start-up source [29].

Existing reviews provide important but partial coverage of power-system resilience under extreme weather events. One strand characterizes meteorological impacts on transmission and distribution components, stochastic modeling challenges, and infrastructure-scale mitigation measures [32,33,34,35,36]. A second strand focuses on EMS-oriented optimization and decision-making in microgrids [37,38], whereas a third examines hierarchical control by comparing primary, secondary, and tertiary layers and their objectives [39]. However, these bodies of literature are often weakly connected: reviews on extreme events do not always address the control functions required for islanding, black-start, and service continuity, while EMS/control reviews frequently treat resilience as an operational objective without explicitly connecting it to HILP-induced correlated failures [40,41]. Therefore, a critical gap remains at the intersection of HILP threats, bibliometric evolution, and control strategies that enable microgrids to maintain or recover critical services after severe weather disruptions.

In rapidly expanding fields, bibliometric analysis offers a quantitative mechanism for mapping intellectual structure, identifying emerging clusters, evolutionary trajectories, and burst nodes through co-citation, bibliographic-linkage, and keyword co-occurrence analyses, thereby reducing selection biases and enhancing methodological traceability [42]. CiteSpace is used in this study as a specialized bibliometric analysis and visualization tool, rather than as a methodology by itself, because it operationalizes science mapping through dynamic visualizations of bibliometric networks, thematic clusters, and evolutionary timelines derived from large databases [43]. The hybrid design of this article lies in the combination of bibliometric quantification with a structured technical synthesis, an approach recently adopted in studies of electrical-system and microgrid resilience to extreme weather events [44,45]. This design addresses the nexus between HILP threats and control topologies for islanding, black-start, and seamless recovery, while identifying critical gaps in operational resilience capabilities.

Therefore, this article presents a hybrid study that integrates quantitative bibliometric analysis using CiteSpace with a structured technical synthesis of the literature. The study is guided by three research questions: RQ1: How has scientific production on microgrid resilience under extreme weather evolved between 2016 and 2025? RQ2: Which institutions, countries, journals, keywords, and co-citation clusters structure the field? RQ3: How do the dominant bibliometric clusters relate to technical control strategies—EMS, stochastic planning, mobile restoration resources, reinforcement learning, and grid-forming/grid-following control—that enable islanding, black-start, and seamless recovery? By answering these questions, the article explicitly links HILP climate threats with hierarchical control topologies and operational functions that determine dynamic resilience and service continuity in microgrids.

2. Theoretical Framework

2.1. From Service Continuity to Dynamic Resilience

The conceptual transition from traditional reliability—focused on high-probability, low-impact events and evaluated under the N-1 criterion that assumes independent and statistically treatable failures in a quasi-stationary regime [46,47]—toward dynamic resilience in electrical systems and microgrids responds to the urgent need to address high-impact, low-probability weather events [46]. Dynamic resilience refers to the system’s ability to resist, absorb, adapt, and recover rapidly from correlated and simultaneous failures caused by hurricanes, heat waves, severe winter storms, or massive floods [48,49].

This difference is not merely semantic: N-1 assumptions of independent failures prove insufficient when HILP disturbances produce multiple correlated losses and cascade trajectories [24,47,50]. Critical events such as the extreme cold in Texas in February 2021 [51,52,53] and the annual Bulk Power System performance reports published by NERC show that a single severe weather event can induce simultaneous large-scale degradations in generation availability including common exposure failures triggering protective trips and loss of operating margin that amplify the initial contingencies [54]. To capture this phenomenon, recent studies propose multi-criteria metrics based on the Choquet integral and graph theory, which overcome the limitations of conventional static indices [55,56,57]. In this context, resilience can be mathematically formalized through the system-performance trajectory during the time window of the disruptive event, expressed as:

$$P_{\mathrm{resilience}}={\int }_{t_0}^{t_e}\left[R\right(t)-R_d(t\left)\right]\mathrm{d}t$$

This formulation, commonly associated with the resilience trapezoid, quantifies the area of performance loss rather than resilience as a positive stock. By subtracting the degraded performance R_d(t) from the nominal performance R(t), the integral evaluates both the magnitude and duration of the impact of the extreme event between the onset of the disturbance (t₀) and full recovery (t_e) [58,59]. Consequently, a smaller integral indicates lower service degradation or faster restoration, whereas a larger integral indicates deeper or longer performance loss.

Figure 1 illustrates the resilience trapezoid through the temporal evolution of a normalized performance function P(t) [5,60]. The initial degradation phase represents robustness and absorption capacity; a less steep decline indicates a better capacity to preserve service after the disturbance, whereas the minimum point represents the severity of operational degradation. The shaded area should be interpreted as loss of performance, calculated by integrating the difference between nominal and degraded performance. The recovery phase captures the speed with which the microgrid restores operating conditions, a process in which black-start capabilities of grid-forming inverters and control algorithms aligned with IEEE 2030.7 can reduce outage duration under suitable design and protection conditions [61]. Finally, the adaptation stage represents the possibility of reaching a more robust post-event state, moving beyond the static view of traditional reliability.

2.2. Hierarchical Control Architecture in Resilient Microgrids

The hierarchical architecture in Figure 2 organizes microgrid control into levels to support continuity in the face of extreme weather events [32,40]. At the tertiary level, the EMS manages proactive dispatch using optimization models that integrate forecasts to preserve critical energy reserves [31,62]. Under the Under the IEEE 2030.7 standard this level can support preventive isolation and logical reconfiguration of the grid in the face of severe disturbances [63,64,65].

The secondary level acts as a technical link to correct frequency and voltage deviations, ensuring accurate islanding detection following disconnection from the main grid [40,66]. Its role is vital for resilience by managing safe transitions and preventing unnecessary tripping of protection systems during the event’s impact [63,67,68]. Likewise, this level coordinates the resynchronization necessary for a safe return to normal operation once the environment has stabilized [63].

At the primary level, fast stability support is provided through inertial response emulation, local voltage/frequency regulation, and autonomous power sharing [66,69]. Here, GFM inverters are important because they can establish voltage and frequency references that support islanded operation and black-start sequences when protection, synchronization, and resource-adequacy requirements are satisfied [29,70]. This hierarchical structure can reduce initial performance loss, accelerate restoration, and support post-event adaptation, although the magnitude of these benefits depends on system design and validation context [71].

For post-disaster resilience, the feasibility of recovery depends partly on the inverter-control taxonomy in the primary layer, particularly the distinction between GFM and GFL strategies introduced above [72,73]. Figure 3 summarizes their functional differences and their relationship with a black-start sequence. GFM strategies emulate voltage-source behavior and can impose voltage and frequency references [29,74]. These characteristic supports operation in weak grids, autonomous supply of critical loads, and the initial energization of a critical bus after a blackout. By contrast, GFL strategies operate as controlled current sources that generally require a phase-locked loop to synchronize with an external voltage reference, either the main grid or a GFM unit. Consequently, GFL inverters are suitable for grid-connected operation but cannot normally sustain an islanded system without a firm voltage reference [29,74].

Taken together, Figure 3 establish the theoretical link between resilience metrics and control implementation. The trapezoid defines the performance loss that the system seeks to minimize, while the hierarchical architecture and the GFM/GFL taxonomy identify the technical mechanisms that affect each phase of the curve: tertiary EMS and stochastic scheduling reduce pre-event exposure, secondary control supports safe transitions and resynchronization, and primary GFM control enables voltage-reference formation and black-start during islanded restoration. This link is used in Section 4 and Section 5 to interpret the bibliometric clusters as operational resilience capabilities rather than only as publication topics.

3. Materials and Methods

3.1. Database

Data were collected from the Scopus database using the following search string: TITLE-ABS-KEY (microgrid* OR “micro grid” OR microrred*) AND TITLE-ABS-KEY (resilien* OR “grid resilience” OR “power system resilience”) AND TITLE-ABS-KEY (“extreme weather” OR “severe weather” OR hurricane* OR typhoon* OR cyclone* OR storm* OR flood* OR wildfire* OR heatwave* OR “heat wave” OR “cold snap” OR blizzard* OR blackout* OR outage*) AND TITLE-ABS-KEY (“hierarchical control” OR “distributed control” OR “multi-agent” OR “inverter control” OR “grid-forming” OR “droop control” OR “virtual inertia” OR EMS OR dispatch* OR schedul* OR MPC OR island* OR “adaptive protection” OR reconfigur* OR “self-healing” OR “black start” OR “load shedding”) AND NOT TITLE-ABS-KEY (“public policy” OR “energy policy” OR “policy framework” OR “regulatory framework” OR “market design” OR tariff*) AND PUBYEAR > 2015 AND PUBYEAR < 2026 AND SUBJAREA (ENER OR ENGI OR COMP OR MATH) AND DOCTYPE (ar). The final corpus consisted of 283 article records retrieved from Scopus on 25 February 2026 after topical relevance screening, duplicate removal, and exclusion of the three records indexed in 2015. Therefore, all bibliometric analyses, figures, annual counts, and interpretations are aligned with the January 2016–December 2025 study window.

Figure 4 shows a sustained upward trend in publication output during 2016–2025, increasing from 11 records in 2016 to 115 records in 2025. After excluding the three 2015 records to align the analysis with the stated decade, the descriptive linear fit is y = 11.64x + 0.71 (R² = 0.946) and the exponential fit is y = 12.77e^0.260x (R² = 0.952), where x = year − 2016. These curves are used only as descriptive summaries of past publication behavior rather than as forecasts, because 2025 indexing may still be incomplete and future growth can be affected by database coverage and publication delays.

3.2. CiteSpace Analysis Parameters and Techniques

Bibliometric analysis encompasses a set of quantitative methods for measuring, tracking, and analyzing scientific literature using statistical and computational techniques [75]. Knowledge maps generated through bibliometric analysis allow researchers to gain an intuitive and in-depth understanding of the current landscape of a disciplinary field, while rigorously capturing emerging trends [76]. CiteSpace was selected as the bibliometric analysis and visualization tool because it supports multivariate, temporal, and dynamic visualizations of citation networks [77]. Developed by Professor Chaomei Chen [78], the application offers advantages for cluster detection, research-frontier identification, burst analysis, and workflow reproducibility [79,80].

Unlike traditional literature reviews limited to descriptive syntheses of topics, publication volumes, and thematic domains, the bibliometric approach is useful for datasets such as this one (283 documents), as it enables metric mapping of knowledge and tracks the dynamic evolution of the field. Various bibliometric visualization platforms coexist, including VOSviewer version 1.6.20 [81], and Bibliometrix version 5.4.0. [82]; CiteSpace was selected because of its suitability for multivariate, temporal, and dynamic literature maps and its ability to detect focal points, turning points, and burst terms [43,82]. The study adopted CiteSpace version 6.2.R4. The workflow (Figure 5) consisted of: (i) exporting Scopus records in a format compatible with CiteSpace; (ii) screening the records for topical relevance to microgrid resilience, HILP weather disturbances, and control or restoration functions; (iii) importing and deduplicating the final corpus of 283 records; (iv) setting the time frame from January 2016 to December 2025 with annual time slices; (v) using title, abstract, author keywords, and indexed terms as terminology sources; (vi) prioritizing the top 50 items per slice; (vii) applying pruning to improve network readability; and (viii) constructing collaboration, co-occurrence, co-citation, timeline, and burst networks. These steps are reported to make the bibliometric analysis reproducible and to clarify that CiteSpace is the analytical tool within a broader hybrid design rather than the full methodology.

Table 1. Bibliometric screening and CiteSpace configuration used for reproducibility.

Component	Configuration or Decision	Purpose
Database	Scopus; final corpus of 283 article records retrieved on 25 February 2026 after topical screening, duplicate removal, and exclusion of 2015 records.	Clarifies the database source and removes the Scopus/Web of Science inconsistency
Time slicing	January 2016–December 2025; one-year slices	Ensures complete annual intervals for trend and cluster interpretation
Text fields	Title, abstract, author keywords, and indexed keywords	Defines the terminology sources used for co-occurrence and cluster labels
Selection criterion	Top 50 items per slice	Keeps networks comparable across years while controlling density
Networks generated	Institutions, authors, countries, journals, keywords, co-citations, timeline clusters, and bursts	Connects the bibliometric outputs with the research questions

summarizes the protocol parameters that support reproducibility of the bibliometric corpus and CiteSpace configuration.

In the Table 1 reports the database source, screening decisions, search fields, time slicing, and CiteSpace settings used to construct the analytical corpus, it is placed in the methodology rather than in the results. Its function is to document the reproducibility protocol and to distinguish the bibliometric mapping procedure from the subsequent interpretation of collaboration, co-occurrence, co-citation, timeline, and burst results.

4. Results

4.1. Institutional Analysis

The institutional collaboration network suggests a field structured around a limited number of bridging institutions and several productive but less connected clusters. Argonne National Laboratory has the highest betweenness centrality among the listed institutions (0.03), which indicates a comparatively important bridging position in the mapped collaboration network since 2016 (

Table 2. Co-occurrence and centrality analysis of the main research institutions.

Centrality	Count	Year	Institutions
0.03	7	2016	Argonne National Laboratory
0.02	2	2017	Xi’an Jiaotong University
0.01	12	2022	Islamic Azad University
0.01	4	2021	University of the Ryukyus
0.01	4	2021	University of Vaasa
0.01	4	2018	Southern Methodist University
0.01	3	2019	Illinois Institute of Technology
0.01	2	2022	University of Isfahan
0.01	2	2017	King Abdulaziz University
0.00	9	2020	Nanjing University of Science and Technology
0.00	6	2023	Department of Electrical and Electronics Engineering
0.00	6	2017	Pacific Northwest National Laboratory
0.00	6	2021	University of Tabriz
0.00	6	2016	Washington State University
0.00	5	2019	Aalborg University
0.00	5	2020	Hohai University
0.00	5	2016	Sharif University of Technology
0.00	5	2016	Idaho National Laboratory
0.00	5	2023	Tsinghua University

). Islamic Azad University has the highest publication count in the table (12 records) but lower centrality (0.01), suggesting that publication volume and network brokerage do not necessarily coincide. The visualization also indicates a geographically distributed but fragmented network, with visible institutional groups in the United States, China, and Iran. These results should be interpreted as bibliometric patterns of collaboration rather than direct measures of institutional research quality or operational influence (Figure 6).

4.2. Author Analysis

The co-authorship network and productivity indicators show that structural connectivity is not always proportional to publication volume. Chen, Chen appears as one of the most connected authors in the mapped network, with betweenness centrality of 0.02 since 2016, while other authors with larger publication counts, such as Cai, Sheng, have lower centrality (

Table 3. Analysis of publication frequency and centrality of the most influential authors.

Centrality	Count	Year	Authors
0.02	6	2016	Chen, Chen
0.02	4	2021	Shafie-khah, Miadreza
0.02	3	2017	Shahidehpour, Mohammad
0.02	3	2017	Bie, Zhaohong
0.02	2	2019	Amirioun, Mohammad Hassan
0.01	6	2016	Wang, Jianhui
0.00	9	2020	Cai, Sheng
0.00	7	2023	Xie, Yunyun
0.00	6	2020	Wu, Qiuwei
0.00	5	2016	Farzin, Hossein
0.00	4	2022	Hemmati, Reza
0.00	4	2021	Masrur, Hasan
0.00	4	2021	Senjyu, Tomonobu
0.00	3	2023	Strbac, Goran

). This pattern indicates a distinction between authors who function as bridges among subcommunities and authors whose output is concentrated within narrower collaboration groups. The interpretation remains bibliometric: the results identify collaboration structure and potential integration opportunities rather than ranking individual scientific quality (Figure 7).

4.3. Keyword Co-Occurrence Analysis

The keyword co-occurrence analysis indicates an intellectual structure that has evolved from operational response and restoration topics toward a broader design framework centered on microgrid resilience (

Table 4. Top 10 Keyword Co-occurrence.

Rank	Keyword	Count	Centrality	Year
1	Microgrid	160	0.16	2021
2	Microgrids	107	0.14	2020
3	Resilience	79	0.12	2018
4	Outages	69	0.20	2016
5	Disasters	53	0.07	2016
6	Distribution Systems	53	0.09	2016
7	Restoration	50	0.06	2017
8	Integer Programming	44	0.05	2016
9	Optimization	44	0.08	2017
10	Electric Power Transmission Networks	39	0.06	2017

). Although the term “Microgrid” dominates publication volume, the centrality of “Outages” (0.20) suggests that service-interruption mitigation connects several research lines in the corpus. Chronologically, the map points to a transition from technical and methodological foundations, such as distribution systems and integer programming, toward integrative resilience-oriented research. This pattern should be interpreted as evidence of thematic consolidation in the literature, not as direct proof that all proposed technologies have reached operational maturity [83].

4.4. Country-Based Analysis

The international collaboration network indicates a global structure in which the United States, Iran, and China occupy prominent but different roles. The United States contributes the largest number of records in the corpus, while Iran has the highest betweenness centrality (0.60), indicating a strong brokerage position in the mapped country network (

Table 5. Co-occurrence analysis of countries by document volume and intermediation centrality.

Serial No.	Documents	Centrality	Year	Keywords
1	97	0.40	2016	U.S.
2	66	0.12	2017	China
3	58	0.60	2016	Iran
4	25	0.16	2018	India
5	25	0.16	2018	United Kingdom
6	11	0.14	2017	Saudi Arabia
7	9	0.11	2019	Denmark
8	8	0.03	2018	Finland
9	8	0.10	2020	Australia
10	8	0.08	2020	Canada
11	8	0.14	2019	South Korea

). China ranks highly in publication volume but has comparatively lower centrality (0.12), suggesting a more concentrated collaboration pattern in this dataset. The presence of India, the United Kingdom, Australia, and Canada also points to geographic diversification. These findings should be read as bibliometric evidence of collaboration patterns rather than as direct measures of national technological capability (Figure 8).

4.5. Journal Analysis

The journal co-citation analysis reveals a specialized intellectual base concentrated in power-engineering, smart-grid, energy, and storage outlets. IEEE Transactions on Smart Grid has the highest citation count and betweenness centrality in the mapped set, indicating that it is a major source connecting research on smart microgrids and resilience. IEEE Access and Energy also have substantial centrality values, suggesting that they function as important dissemination channels for technical, economic, and energy-management studies (

Table 6. Co-citation analysis of journals.

Count	Centrality	Year	Cited Journals
159	0.28	2016	IEEE Transactions on Smart Grid
140	0.23	2019	IEEE ACCESS
108	0.21	2018	ENERGY
90	0.11	2017	ENERGY APPLICATIONS
81	0.18	2016	ENERGIES
69	0.04	2018	IEEE Transactions on Sustainable Energy
62	0.05	2018	IEEE Transactions on Smart Grid
53	0.04	2016	IEEE PROC.
52	0.02	2017	APPL ENERGY
51	0.02	2018	IEEE Trans. Power Syst.
46	0.04	2017	IEEE Trans. Power Syst.
41	0.01	2020	Int. J. Electr. Power Energy
39	0.08	2017	IEEE Power & Energy Magazine
33	0.04	2021	SUSTAINABILITY
32	0.01	2021	J. ENERGY STORAGE

). The emergence of Sustainability and the Journal of Energy Storage around 2021 suggests an expansion of the field toward storage solutions and sustainability-oriented frameworks. These findings describe the citation structure of the corpus and should not be interpreted as journal rankings beyond this dataset.

4.6. Cluster Identification and Analysis (Knowledge Domain)

The document co-citation map and cluster metrics reveal a differentiated scientific structure, supported by Q-modularity of 0.5121 and an average silhouette of 0.7393, indicating an interpretable division of knowledge areas. The dominant cluster (#0), with 43 records and a median year of 2017, focuses on outage-resilient energy-management frameworks and represents one of the dense cores of the mapped field. The cluster structure suggests an evolution from service-restoration topics toward networked microgrids, mobile resources, and techno-economic resilience analyses. Overall, these results indicate a thematic transition in the literature; they do not by themselves demonstrate field deployment or operational effectiveness (Figure 9).

Figure 9. Visualization map of clusters in the document co-citation network. Source: authors’ own CiteSpace visualization from the corrected Scopus corpus. The colored regions represent algorithmically detected co-citation clusters; the figure supports the identification of thematic groupings, while the detailed interpretation relies on cluster metrics reported in Table 7.

Figure 9. Visualization map of clusters in the document co-citation network. Source: authors’ own CiteSpace visualization from the corrected Scopus corpus. The colored regions represent algorithmically detected co-citation clusters; the figure supports the identification of thematic groupings, while the detailed interpretation relies on cluster metrics reported in Table 7.

Table 7. Statistical characteristics and thematic tags of the main identified clusters.

N	Size	Silhouette	Mean	Label (LSI/LLR)
0	43	0.653	2017	Tie Line; Outage-Resilient Energy Management Framework; Extreme Event; Economic Performance
1	41	0.773	2012	Resilient Distribution System; Supply Resilience; Distribution System Restoration; Resilience Enhancement
2	40	0.824	2013	Post-Event Electric Taxis; Hierarchical Scheduling Framework; Resilience Enhancement; Renewable-Based Microgrid
3	40	0.629	2007	Reconfigurable Networked Microgrid; Hierarchical Outage Management; Enhancing Power System Resilience
4	29	0.792	2002	Resilient Operation; Critical Infrastructure; Distribution System; Networked Microgrid Scheduling
5	28	0.584	2011	Resilient Microgrid; Stochastic Extreme Event; Backup Generator; Islanded Mode; Proactive Management
6	25	0.664	2015	Resilience Enhancement; Robust Planning Model; Distribution Grid Resilience; Natural Disaster
7	25	0.817	2016	Techno-Economic Analysis; Movable Energy Resource; Enhanced Resilience; Robust MPC-Based Microgrid
8	13	0.926	1991	Service Restoration; Extreme Weather Event; Self-Healing Resilient Operation; Active Distribution Network

Table 7. Statistical characteristics and thematic tags of the main identified clusters.

N	Size	Silhouette	Mean	Label (LSI/LLR)
0	43	0.653	2017	Tie Line; Outage-Resilient Energy Management Framework; Extreme Event; Economic Performance
1	41	0.773	2012	Resilient Distribution System; Supply Resilience; Distribution System Restoration; Resilience Enhancement
2	40	0.824	2013	Post-Event Electric Taxis; Hierarchical Scheduling Framework; Resilience Enhancement; Renewable-Based Microgrid
3	40	0.629	2007	Reconfigurable Networked Microgrid; Hierarchical Outage Management; Enhancing Power System Resilience
4	29	0.792	2002	Resilient Operation; Critical Infrastructure; Distribution System; Networked Microgrid Scheduling
5	28	0.584	2011	Resilient Microgrid; Stochastic Extreme Event; Backup Generator; Islanded Mode; Proactive Management
6	25	0.664	2015	Resilience Enhancement; Robust Planning Model; Distribution Grid Resilience; Natural Disaster
7	25	0.817	2016	Techno-Economic Analysis; Movable Energy Resource; Enhanced Resilience; Robust MPC-Based Microgrid
8	13	0.926	1991	Service Restoration; Extreme Weather Event; Self-Healing Resilient Operation; Active Distribution Network

4.7. Analysis of the Evolutionary Trajectory of Research

The visualization of the cluster timeline suggests an intellectual evolution from service-restoration fundamentals toward more interconnected and proactive resilience themes. Clusters #8 (service restoration) and #3 (reconfigurable networked microgrid) show concentrated activity between 2016 and 2018, whereas clusters #0 (tie line) and #1 (resilient distribution system) remain visible in later time slices and include recent nodes toward 2024–2025. Cluster #2 (post-event electric taxis/mobile resources) indicates growing attention to mobile and flexible resources for disaster recovery. This longitudinal pattern supports a literature-based interpretation of thematic shift; it should not be read as causal evidence that the corresponding technologies are already deployed at scale.

Figure 10 should be read from left to right as a temporal map of thematic persistence. Each horizontal line corresponds to a co-citation cluster, nodes indicate cited documents positioned by year, and larger or warmer-colored nodes represent stronger citation influence or burst behavior. Therefore, the figure is not a causal diagram; it is a bibliometric visualization that helps identify when each topic became active and how long it remained relevant within the corpus.

The temporal distribution of research intensity by clusters complements the previous timeline by summarizing how strongly each cluster appears across the study period. Brighter or denser regions indicate periods in which publications and citations are concentrated within a given thematic line, whereas faded regions represent declining activity. Foundational clusters such as #8 (service restoration) and #9 (heuristic optimization) have lower recent intensity after their historical peaks, while clusters #0 (tie line) and #1 (resilient distribution system) maintain substantial activity during the most recent decade. Cluster #2 (post-event electric taxis/mobile resources) exhibits a recent increase, suggesting growing interest in mobile and flexible resources for disaster recovery. Thus, Figure 11 should be interpreted as complementary evidence of a shift from isolated restoration techniques toward resilient and interconnected distribution systems.

In Figure 11, the vertical ordering follows the cluster labels produced by CiteSpace, while the horizontal dimension represents time. The intensity bands indicate the relative concentration of activity within each cluster. This visualization supports the interpretation that resilient distribution systems, tie-line/interconnection topics, and mobile post-event resources have become more visible in the recent literature, whereas earlier service-restoration topics are less dominant in the latest slices. Because this figure is visually dense, it should be read together with Table 7 and Figure 10 rather than as a stand-alone evidentiary source.

4.8. Analysis of Resilience to Bursts

The citation-burst analysis for keywords indicates a shift in research attention from technical and structural foundations toward advanced operational management and dynamic system control. Initially, the field was strongly associated with the concept of the microgrid and related distribution-system foundations. Between 2018 and 2022, bursts in terms such as “storms,” “hurricanes,” and “natural disasters” indicate increased attention to resilience under climate threats [18,84]. In the most recent period (2023–2025), terms such as “electric loads,” “load modeling,” and “electric power system control” [85,86] suggest a growing focus on smart control and load dispatch. These burst patterns identify changes in research emphasis within the literature; they do not directly validate operational performance (Figure 12).

5. Discussion

The evolution of microgrids has shifted from static planning and restoration models toward dynamic systems supported by optimization, artificial intelligence, flexible resources, and hierarchical control. This discussion critically connects the bibliometric clusters identified in Section 4 with the technical mechanisms reviewed in the literature, emphasizing where evidence is mature, where claims remain mainly simulation-based, and where experimental validation is still limited.

5.1. Energy Management Systems (EMS) and Multi-Energy Optimization

This thematic area corresponds to Cluster #0. It represents a technical evolution from Clusters #5 and #1, focused on the integration of multiple energy vectors to improve the system’s overall efficiency and resilience to uncertainties. Systems based on fuzzy logic have demonstrated significant improvements, achieving an 80% reduction in electricity grid consumption, along with decreases of 7.4% in gas boiler consumption and 5.4% in electric boiler consumption [87], an 11.4% decrease in the maximum power drawn from the grid, and notable reductions in the ramp rates of the power profile in electro-thermal microgrids [88]. These results confirm the superiority of low-complexity fuzzy controllers over traditional methods, minimizing fluctuations without requiring extensive forecasts or precise models of the system dynamics [89]. These advances are complemented by energy management platforms that incorporate forecasting, human–machine interfaces, and data analysis, validated in real-world test environments such as the UCLA Smart Grid Energy Research Center and the Korea Institute of Energy Research [90], and extended to multi-objective approaches with hybrid storage (batteries and hydrogen) to balance demand, costs, and component lifespan [91].

Recent literature has expanded the scope of energy management systems toward multi-energy architectures that integrate thermal, electrical, and gas vectors [87]. The implementation of two-level energy management systems—advanced daily scheduling and real-time scheduling—has made it possible to maintain load consumption while improving the demand-side management mechanism based on load shifting techniques [92,93,94] implemented with low-complexity fuzzy logic controllers that achieve reductions of 80% in electricity grid consumption, 7.4% in gas boilers, and 5.4% in electric boilers, along with an 11.4% decrease in peak power consumption [88,89,95]. The implementation of model predictive control and rolling horizon predictive control has made it possible to compensate for changes in load and generation throughout the day, achieving an 18.7% reduction in operating costs compared to conventional dispatch methods, with additional improvements in efficiency and uncertainty management [37].

The integration of artificial intelligence into energy management systems represents an emerging frontier of research. Methodologies that combine multiple renewable energy systems with AI have reported reductions in carbon emissions while optimizing distributed-resource use [96]. The implementation of energy management systems based on high-performance embedded processors has enabled real-time optimization via hybrid Modbus-TCP/IP communication, providing a centralized management solution for distributed energy resources and storage [97]. Nevertheless, the reviewed evidence also shows an important limitation: many AI-based EMS approaches remain validated through simulation or laboratory prototypes, so their scalability, cybersecurity, and behavior under communication degradation during HILP events require further validation.

5.2. Robust and Stochastic Planning Under Uncertainty

This thematic area corresponds to Cluster #5 and Cluster #6. The literature distinguishes between stochastic approaches, which use probability distributions and scenarios to model uncertain variables such as load and renewable generation, minimizing the expected cost of current and future decisions [98], and robust or chance-constrained approaches, which guarantee operability under the worst-case scenario or within specific confidence limits, controlling operational risks without requiring precise distributions [38,99,100]. Although traditional robust optimization generates conservative and costly solutions by hedging against all possible uncertain realizations [101], methods such as robust programming with well-defined uncertainty sets have emerged as promising for analyzing multiple worst-case scenarios, providing feasible solutions that are less computationally intensive than stochastic methods [98].

Joint distributed robust programming with opportunity constraints has made it possible to directly control system reliability at a predefined level and determine the optimal cost, overcoming the limitations of traditional robust optimization, which generates excessively conservative and costly solutions by considering the worst-case scenario [99]. Robust optimization has received significant attention as a modeling framework for optimization under parameter uncertainty, seeking the allocation and dispatch of generation resources to immunize against all possible uncertain situations [101].

Stochastic programming models uncertainty using scenarios and probabilities, minimizing the current cost plus the expected future cost, which generates less conservative and computationally more intensive solutions than traditional robust optimization, which is less computationally intensive and immunizes against the worst-case scenario [98,101]. Opportunity-constrained programming, a stochastic branch, enables multi-objective models in isolated or reconfigurable microgrids with renewables, integrating storage, demand response, and user experience through stochastic scenarios and joint constraints that control risks at predefined levels [99,100,102].

5.3. Restoration, Reconfiguration, and Mobile Resources

This thematic area corresponds to Cluster #2 and Cluster #3. The main innovation in this field is the transition from static restoration strategies to the use of mobile resources, such as transportable energy storage systems and mobile energy resources, which allow for greater resilience compared to fixed storage due to their spatio-temporal flexibility. These systems, when coordinated with grid reconfigurations and microgrid scheduling, minimize load shedding during extended outages [103]. A spatio-temporal model of TESS, related to transmission networks and distribution systems, has been proposed to represent the difference between TESS and fixed ESS in terms of flexibility and cost reduction [104]. Two-stage optimization models have optimized investments in mobile units and their rerouting to form dynamic microgrids, avoiding load shedding expected due to disasters [103]. Furthermore, rolling optimization of mobile storage fleets coordinates local stationary resources for integrated restoration [105], while adaptive formation of multi-microgrids incorporates optimally positioned MERs in the face of extreme risks such as natural disasters [106], and coordination with dynamic reconfiguration enhances seismic resilience [107].

Mobile energy resources, such as mobile emergency generators, transportable energy storage systems, and electric vehicles, represent a significant advance in the resilient restoration of distribution service, offering greater spatiotemporal flexibility than fixed resources to support dispersed critical loads during extensive outages caused by disasters [108]. These flexible and transportable resources can be quickly and easily integrated into distribution systems in the face of sustained damage, causing prolonged outages, moving from staging locations to failure sites via trucks or transport routes [107]. Two-stage optimization models, solved using progressive algorithms such as progressive hedging, optimize investments in mobile units in the first stage and their redeployment in the second to form dynamic microgrids, coordinated with network reconfiguration, avoiding expected load shedding and significantly reducing outages and total costs in extreme seismic or climatic scenarios [109].

The adaptive formation of multi-microgrids as part of the critical service restoration strategy has emerged as a promising approach for extreme conditions such as natural disasters and physical and cyberattacks [106]. This strategy comprises microgrid formation steps and load switching sequences, where the MF forms multiple microgrids with minimal overall scale and radial or loop topologies, including mobile emergency resources optimally positioned to address potential risks under extreme conditions, verified on the IEEE 132-node system via simulation in MATLAB Simulink [106]. Rolling optimization of mobile energy storage fleets coordinates local stationary resources for an integrated restoration of the distribution system, minimizing total costs (outage, generation, and operation) through two-stage stochastic programming that accounts for damage to transmission and distribution networks, demonstrating the effectiveness of MESS mobility in tests with Sioux Falls and Singapore systems connected to 33-node networks [105]. Similar two-stage models optimize the pre-positioning and re-routing of mobile units to form dynamic microgrids, avoiding expected load shedding and reducing outages in IEEE 33/123 systems [110]. These advances represent a fundamental transition from static restoration approaches to dynamic and adaptive strategies, overcoming the limitations of fixed ESS through spatio-temporal flexibility and cost reduction through sharing among microgrids [104].

5.4. Reinforcement Learning and Multi-Agent Control

This thematic area corresponds to Cluster #3. The rise in reinforcement learning addresses the need to manage the complexity, renewable uncertainties, and real-time dynamics of modern microgrid systems, overcoming the limitations of exponential computation time, reliance on accurate forecasts, and the conservatism of traditional mixed-integer linear programming methods [111,112]. These results, validated in seasonal simulations under uncertainty, confirm the superior resilience, operational adaptability, and economic efficiency of PPO and RL compared to conventional approaches [113,114].

Deep reinforcement learning has become a prominent research direction for microgrid control and management, offering potential advantages over traditional methods in adaptive decision-making. A recent systematic review categorizes selected articles and summarizes the main findings, methodologies, and contributions in this area [115]. DRL algorithms, including off-policy reinforcement learning, inverse reinforcement learning, goal-conditioned reinforcement learning, and multi-agent reinforcement learning, are used to address selected control and management challenges in microgrids [115]. Multi-agent systems based on deep reinforcement learning can allow intelligent agents to learn from the microgrid environment and select actions that maximize cumulative rewards, although these results remain strongly dependent on model assumptions, training data, and validation settings [116].

Multi-agent reinforcement learning has emerged as a promising framework for power-system control tasks involving coordination among multiple agents. In frequency regulation, each generator can be treated as an individual agent making its own generation decisions, while the frequency dynamics are determined jointly by all power injections [117]. The PPO algorithm simplifies policy optimization by using a clipped objective function that prevents large parameter updates and eliminates the incentive to move the probability ratio outside the interval [1 − ε, 1 + ε] [118]. These advances suggest that reinforcement learning may support reliability-centered energy management in microgrids, but challenges related to computational constraints, generalization, explainability, safety guarantees, and cyber-physical vulnerabilities remain significant [119].

5.5. Black-Start, GFM Control, and Self-Healing

This thematic area corresponds to Cluster #8 (black-start, 13 documents, median year 1991, silhouette 0.926). Although the topic has deep historical roots, it has been renewed through its connection to Cluster #4 and the literature on GFM inverters for self-healing functions. GFM inverters are increasingly discussed as resources for future energy systems, particularly for establishing and restarting microgrids following blackouts [120]. Under suitable design conditions, GFM inverters can allow microgrids to operate independently of grid power and can provide fast startup, voltage reference, and frequency support for critical loads [120].

Autonomous control architectures for grid-interactive inverters can allow devices to detect reconnection to the utility grid without relying exclusively on communications and to transition from grid-forming mode to grid-following mode [121]. When a microgrid is ready to reconnect to the utility grid, a synchronization relay verifies synchronization, assuming that the inverters have restored voltage and frequency to nominal levels [121]. Prior to reconnection, a transformer with an automatic tap changer can correct minor voltage discrepancies on both sides of the relay [121]. These advances are promising for self-healing distribution networks, but their practical impact depends on protection coordination, interoperability, and field or hardware-in-the-loop validation.

The concept of island identification extends traditional island detection for distribution networks with fleets of GFM inverters [122]. When part of a distribution system forms an island, GFM inverters may remain online to form a microgrid and maintain local frequency and voltage without a complete blackout phase [122]. For this to happen, the inverters must identify island boundaries and establish secondary control with the correct set of devices located in the same island [122]. Recent work on autonomous restoration of interconnected microgrids using distributed and inverter-based resources indicates substantial potential, but demonstrated operational capability still requires robust validation under degraded communications and realistic damage scenarios [123].

5.6. Control Strategies and Climate Challenges

This thematic area corresponds to the synthesis of Clusters #0, #1, and #4, reflecting the need for integrated hierarchical control to address extreme weather events. Coordination between the tertiary layer, which manages economic and logistical decisions for mobile resources, and the primary and secondary layers, which maintain voltage/frequency stability and resynchronization, is central to resilience-oriented operation. The main unresolved challenge is coordination under degraded communication conditions, where future microgrids should be able to preserve critical functions with limited external support rather than assuming uninterrupted communications.

The growing penetration of weather-dependent renewable energy sources introduces new challenges regarding system security, reliability, flexibility, and sustainability [124]. Microgrids, as decentralized energy systems, offer resilience and flexibility but require intelligent control strategies to balance variable generation and demand [115,125]. The inherent stochastic nature of wind and solar generation creates significant challenges for the reliability of modern energy systems, requiring control frameworks that can adapt in real time to changing conditions [126]. These challenges have been exacerbated by the increase in extreme weather events that can cause extensive physical damage to distribution infrastructure, requiring restoration strategies that consider both repair logistics and the coordination of mobile resources [127,128].

The integration of multiple control layers—from tertiary energy management to primary inverter control—represents an important research frontier in microgrid control. Two-layer EMS designs, with an upper energy-management layer and a lower control layer, can minimize daily operating costs while improving local self-consumption of renewable energy resources [129]. Model predictive control and rolling-horizon predictive control can compensate for changes in load and generation throughout the day [129], while complementary fuzzy-logic and decentralized EMS approaches support demand-side coordination and renewable self-consumption [130,131]. However, persistent gaps remain in coordinating these layers under degraded communication conditions, and further decentralized control strategies are required for autonomous operation during disasters.

The expanded synthesis also incorporates recent contributions on mobile restoration resources [132] and reinforcement-learning applications in power systems [133], which align the reference list with the representative evidence base summarized in

Table 8. Systematic Table of Findings and Gaps.

Thematic Area	Cluster(s)	Representative Reviewed Works	Main Finding	Critical Gap
Multi-energy EMS and demand-side management	#0, #1, #5	[87,88,89,90,91,92,93,94,95,129,130,131]	Fuzzy logic, MPC, rolling-horizon scheduling, and multi-vector EMS reduce grid dependence and smooth electro-thermal demand profiles.	Real-time scalability, cyber-secure communication, and validation during degraded communication conditions.
Robust, stochastic, and chance-constrained planning	#5, #6, #1	[38,98,99,100,101,102]	Robust and stochastic formulations manage renewable/load uncertainty and define resilience-aware operating margins.	Conservatism, scenario dependence, and limited coupling with physical damage models for HILP events.
Mobile restoration resources and post-event flexibility	#2, #3, #7	[103,104,105,106,107,108,109,110,111,132]	TESS, MESS, mobile emergency generators, and electric vehicles provide spatio-temporal flexibility for critical-load restoration.	Integration with transportation damage, road accessibility, repair logistics, and crew dispatch.
Reconfiguration and dynamic microgrid formation	#1, #3, #4	[61,103,106,107,123,127,128]	Adaptive formation and reconfiguration support islanded service areas and reduce load shedding after disturbances.	Protection coordination, radial/loop topology constraints, and validation under cascading multi-asset damage.
Reinforcement learning and multi-agent control	#0, #3, #4	[73,112,113,114,115,116,117,118,119,126,133]	RL and multi-agent methods support adaptive EMS, selective power-system applications, and distributed decision-making.	Operational stability guarantees, explainability, safety constraints, and transfer from simulation to real-time operation.
Grid-forming/GFL control and black-start	#8, #4	[28,29,70,72,74,120,121,122,123]	GFM inverters can establish voltage/frequency references and support islanded black-start sequences, while GFL units require an external reference.	Commercial-scale hardware-in-the-loop validation, protection coordination, and interoperability among heterogeneous inverter fleets.
Hierarchical control and interoperability standards	#0, #1, #4, #6	[40,63,64,65,66,129]	Primary, secondary, and tertiary control layers link stability, restoration, EMS optimization, and resynchronization functions.	Standardized data exchange, degraded-network coordination, and unified metrics across control layers.
Resilience metrics and HILP quantification	#6, #8	[5,55,56,57,58,59,60,71]	Performance-loss metrics, trapezoids, graph approaches, and dynamic resilience criteria quantify degradation and recovery.	Consistent normalization, comparability across studies, and integration with control-oriented performance indicators.

.

6. Conclusions

This review identifies a progressively consolidating research framework on microgrid resilience, based on the combined evidence from bibliometric mapping and structured technical synthesis. The results indicate a literature-level transition from predominantly reactive restoration approaches toward more proactive control, planning, and coordination architectures under high uncertainty. Because the methodology combines bibliometric patterns with literature-based interpretation, the conclusions below distinguish between mapped research trends, simulation-based evidence, laboratory or hardware-in-the-loop validation, and demonstrate operational capability.

First, the reviewed literature suggests that resilience to HILP events cannot be addressed only through passive physical redundancy. The dominant clusters point to growing interest in the orchestration of distributed energy resources as flexible operational assets. Studies on V2G-enabled electric vehicle fleets, transportable energy storage systems (TESS), mobile energy storage systems (MESS), and emergency generators indicate potential spatiotemporal flexibility for critical-load restoration. However, the evidence is still heterogeneous and often simulation-based; therefore, these resources should be described as promising enablers of dynamic microgrid formation rather than as universally demonstrated solutions.

Second, the operational viability of dynamic microgrids is strongly linked to modernization of the primary control layer. The reviewed technical literature indicates that GFM inverter technology can support self-healing functions by establishing voltage and frequency references for islanded operation and black-start sequences. Nevertheless, GFM control should be regarded as a key enabling technology rather than an automatically sufficient solution, because practical deployment also requires protection coordination, synchronization procedures, resource adequacy, interoperability, and validation under realistic disturbance conditions.

Third, the computational complexity of managing reconfigurable topologies has encouraged research on deep reinforcement learning and multi-agent reinforcement learning in the EMS layer. The literature indicates that these techniques can support adaptive scheduling, distributed decision-making, and selected real-time control tasks. However, claims of superiority over conventional mathematical programming remain context-dependent, and wider deployment requires stability guarantees, explainability, cyber-security assessment, and transfer from simulation to real-time or field operation.

Despite these advances, the study identifies limitations that define a research agenda for 2026–2030. Future work should develop standardized interoperability frameworks that enable coordinated operation among multi-owner microgrids during climate emergencies. Empirical validation using commercial-scale hardware-in-the-loop platforms, real-time digital simulation, and field demonstrations is needed to test seamless transitions between grid-connected, islanded, black-start, and reconnection modes. Finally, resilience modeling should integrate dynamic metrics that capture cyber-physical interdependence and equity-relevant service continuity, ensuring that algorithmic robustness is connected to measurable protection of critical social services.