Leveraging the Synergy of Digital Twins and Artificial Intelligence for Sustainable Power Grids: A Scoping Review

Abstract

As outlined by the International Energy Agency, 44% of carbon emissions in 2021 were attributed to electricity and heat generation. Under this critical scenario, the power industry has adopted technologies promoting sustainability in the form of smart grids, microgrids, and renewable energy. To overcome the technical challenges associated with these emerging approaches and to preserve the stability and reliability of the power system, integrating advanced digital technologies such as Digital Twins (DTs) and Artificial Intelligence (AI) is crucial. While existing research has explored DTs and AI in power systems separately, an overarching review of their combined, synergetic application in sustainable power systems is lacking. Hence, in this work, a comprehensive scoping review is conducted under the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR). The main results of this review analysed the breadth and relationships among power systems, DTs, and AI dynamics and presented an evolutionary timeline with three distinct periods of maturity. The prominent utilisation of deep learning, supervised learning, reinforcement learning, and swarm intelligence techniques was identified as mainly constrained to power system operations and maintenance functions, along with the potential for more sophisticated AI techniques in computer vision, natural language processing, and smart robotics. This review also discovered sustainability-related objectives addressed by AI-powered DTs in power systems, encompassing renewable energy integration and energy efficiency, while encouraging the investigation of more direct efforts on sustainable power systems.

1. Introduction

Global greenhouse gas emissions have continuously increased, resulting from unsustainable practices in the consumption of energy and other resources, land usage, exorbitant lifestyles, and production methods. The Intergovernmental Panel on Climate Change (IPCC) reports a deteriorating global warming trend, with the global surface temperature surpassing pre-industrial levels by 1.1 °C between 2011 and 2020, primarily due to anthropogenic greenhouse gas emissions. This exceeds the goal set by the Paris Agreement, which aims to limit the temperature rise to 1.5 °C above pre-industrial levels. Consequently, significant and swift changes have occurred in the atmosphere, oceans, polar regions, and ecosystems, resulting in observable impacts on various weather and climate extremes worldwide. These changes have given rise to widespread negative effects and consequent losses and damages to both nature and human communities [1].

The power and energy sector is a major contributor to this grave predicament. Despite worldwide efforts to reduce carbon emissions in the energy sector, electricity and heat production accounted for approximately 44% of global CO₂ emissions from fuel combustion in 2021. Specifically, coal-fired power plants contributed approximately 73% of these emissions. The CO₂ emissions attributed to electricity generation stem from the collective impact of multiple factors, including electricity output, generation efficiency, the proportion of fossil fuels in the overall energy mix, and the carbon intensity of fossil fuel-based generation [2]. In addition to the severe weather events and other negative consequences outlined earlier, the power and energy sector stands as the leading contributor to air pollution, affecting more than 90% of the global population. This pollution is associated with over 6 million premature deaths annually [3].

In response to this critical scenario, the International Energy Agency (IEA) recommends that global carbon emissions from electricity generation must be reduced to the net zero level by 2045, and advanced economies should achieve this target by 2035, to align with the objectives of the Paris Agreement. To accomplish this, the IEA has mandated strategies such as scaling up renewable energy sources, enhancing energy efficiency, switching to lower and zero-carbon fuels, using carbon capture technologies, and expanding electricity transmission and distribution grids to achieve the requirements of their net zero emissions scenario (NZE). Concurrently, the Climate Action Tracker (CAT) has also introduced a new set of benchmarks for power sector decarbonisation, including phasing out coal power plants and fossil gas expansion by 2040 globally, increasing global proportion of renewable energy sources to 93–98%, and achieving clean electricity generation by 2040 [4]. Furthermore, IEA has recognised current and future challenges for power grids in emerging markets and developing economies with respect to reliability, resilience, increased losses, affordability and access, rising demand and changing demand profiles, variable and distributed energy, and new loads and storage such as EVs [5]. Against this complex backdrop, it is apparent that the power industry is compelled to transition towards a sustainable, carbon-neutral environment while overcoming technical, financial, and social hurdles against energy efficiency, security, resilience, and affordability.

Hence, smart grids, microgrids, and Distributed Generation (DG) coupled with the increased adoption of Renewable Energy Sources (RESs) have inevitably emerged as major propellants in the transformation of the power industry. While their role in achieving carbon neutrality is well recognised, these concepts also facilitate the efficient generation, transmission, and distribution of electricity by leveraging advanced infrastructure and relocating loads closer to sources, thereby minimising losses. Concurrently, strategies such as Demand Response (DR) and Building Energy Management Systems (BEMSs) promote efficient energy consumption within these concepts. Moreover, they hold the potential to enhance energy security through a diversified generation mix and advanced energy storage and backup systems, ensuring uninterrupted access to energy sources. With respect to resilience, they mitigate reliance on vulnerable traditional sources, reduce single points of failure, enable rapid prediction and response to disruptions, minimise outage impact through automatic power re-routing, and sustain critical infrastructure operation independently during outages. However, these emerging trends introduce technical challenges that adversely impact the stability and reliability of the power grid, originating from the uncertainty, complexity, and volatility inherent in them. Maintaining frequency and voltage stability under volatile generation conditions, power quality issues due to the introduction of higher-order harmonics, reduced system inertia, and the compelling demand for robust communication infrastructure, big data management, predictive capabilities and advanced control strategies are some instances of such major obstacles.

Intelligent and advanced digital technologies can address these technical challenges while preserving the energy efficiency, reliability, stability, and environmental sustainability of the grid. In fact, IEA has outlined the following digital solutions to address the current and future challenges mentioned above for power grids: embedded sensing, automation and control, automated real-time optimisation, enhanced design and predictive monitoring, and distributed generation and demand response [5]. In this era of Industry 4.0 and Industry 5.0, leveraging Cyber-Physical System (CPS) technologies such as Internet-of-Things (IoT), cloud and edge computing, Big Data, Machine Learning (ML), Artificial Intelligence (AI), and Digital Twins (DTs) can significantly enhance the effectiveness of such digital solutions [6]. Moreover, these technologies can introduce a sustainable, clean outlook for the power industry and overcome the technical challenges mentioned above when deployed through smart grids [5]. DTs, in particular, as virtual representations of physical objects, processes, or systems, exhibit real-time monitoring, bidirectional communication, and control capabilities, making them instrumental in the transition of the power grid towards environmental sustainability. The integration of AI techniques with DTs presents a transformative opportunity to overcome technical and sustainability-related hurdles faced by current and future power grids, harnessing the power of predictive modelling, optimisation, advanced analytics, and data-driven decision-making. Hence, this research is focused on conducting a comprehensive scoping review to analyse the existing state of the synergetic use of DT and AI for the enhancement of sustainability in power grids. The contributions of this review are described in detail in Section 1.2.

1.1. Related Surveys and Reviews

In previous research, concepts of DT and AI were connected with various application domains such as industrial manufacturing, aviation, transportation, and power systems in the form of enabling technologies. Possible links among these paradigms are illustrated in Figure 1. For the scope of this review, the following sub-domains are considered to be included in the broad domain of power systems:

• Main power grid: generation, transmission, distribution sectors,
• Smart grids,
• Microgrids,
• Distributed Generation (DG) systems and Renewable Energy Sources (RESs)
• Demand side management: Demand Response (DR), energy storage and backup systems, Energy Management Systems (EMSs),
• Smart integrated energy systems,
• Individual power assets: generators, transformers, converters, inverters, measurement devices, protection devices, control devices, battery systems, transmission and distribution lines, etc.

Due to the fact that DTs were originally introduced within an industrial context during the early stages of Industry 4.0, an extensive number of reviews and surveys were conducted focussing on DTs in the manufacturing industry [7,8,9,10,11] that were subsequently extended to other domains such as healthcare and aviation [8]. This relationship is illustrated by link number 1 in Figure 1. In addition, numerous publications have reviewed the combined use of AI and DTs, where the efficiency and effectiveness of data-driven decision-making of DT solutions could be ameliorated through the improved knowledge representation, pattern recognition, predictive modelling, and analytical capabilities of AI ([12,13,14,15,16] are some of these reviews). Link number 1 combined with number 5 represents this relationship in Figure 1.

With respect to the power system domain, there are a fair amount of reviews and surveys where physics-driven DTs are used in their vanilla form without AI, which is illustrated by link number 3 in Figure 1. Sifat et al. [17] and Sleiti et al. [18] have focused on DTs for main electrical grids in their reviews, whereas Bazmohammadi et al. [19] and Kumari et al. [20] have presented their reviews in the subdomain of microgrid DTs. Moreover, numerous reviews and surveys have explored the demand-side DTs, encompassing energy management [21,22,23], demand side recommendation systems [24], energy storage [25], and Electric Vehicles (EVs) [26].

In addition, a large number of reviews and surveys have been conducted that focused on the application of AI in power grids, which is illustrated by link number 4 in Figure 1. Most of these reviews have addressed the use of AI in smart grids [27,28,29,30]. Ahmad et al. [31] have reviewed the application of AI for RESs and Demand Side Management (DSM). Antonopoulos et al. [32], Barrett and Haruna [33], and Farzaneh et al. [34] have presented their reviews in the subdomain of DSM, focussing on DR, energy storage, and building energy efficiency, respectively.

In Figure 1, link number 3 combined with number 5 represents AI-powered, data-driven DTs used in the previously stated sub-domains of power systems. Some surveys and reviews are available for the combined application of DTs and AI in power systems as well. Ghenai et al. [35] have explored the use of DTs and AI in power generation, storage, and energy consumption in buildings, transportation, and industrial applications. The review of Jafari et al. [36] is focused on DTs augmented by ML for smart grids, whereas Shen et al. [37] have presented their review encompassing power grids as well as power assets. Moreover, multiple reviews have addressed the use of AI for battery system DTs [38,39,40,41].

Table 1 summarises the publication timeline of the currently available reviews and surveys spanning the concepts, application domains, and associations illustrated in Figure 1. Note that only a selectively curated collection of reviews and surveys with a significant number of citations and notable unique contributions are included in this table.

Table 1. Reviews and surveys encompassing the application of Digital Twins (DTs) and Artificial Intelligence (AI) for power systems and other application domains: a timeline.

Category (Link from Figure 1)	2018	2019	2020	2021	2022	2023
DTs for Industrial and Other Application Domains (Link 1)	[11]	[8,9,10]	-	-	[7]	-
DTs and AI for Industrial and Other Application Domains (Links 1 and 5)	-	-	[13,14]	[12,16]	[15]	-
DTs for Power Systems (Link 3)	-	-	-	[19,24] [23,26]	[18,21] [22]	[17,20] [25]
AI for Power Systems (Link 4)	-	-	[27,28] [32,33]	[29,31] [34]	[42]	[30]
DTs and AI for Power Systems (Links 3 and 5)	-	-	[38]	[39]	[35,40]	[36,37,41]

Table 1. Reviews and surveys encompassing the application of Digital Twins (DTs) and Artificial Intelligence (AI) for power systems and other application domains: a timeline.

Category (Link from Figure 1)	2018	2019	2020	2021	2022	2023
DTs for Industrial and Other Application Domains (Link 1)	[11]	[8,9,10]	-	-	[7]	-
DTs and AI for Industrial and Other Application Domains (Links 1 and 5)	-	-	[13,14]	[12,16]	[15]	-
DTs for Power Systems (Link 3)	-	-	-	[19,24] [23,26]	[18,21] [22]	[17,20] [25]
AI for Power Systems (Link 4)	-	-	[27,28] [32,33]	[29,31] [34]	[42]	[30]
DTs and AI for Power Systems (Links 3 and 5)	-	-	[38]	[39]	[35,40]	[36,37,41]

1.2. Contributions

As described in Section 1.1, existing reviews and surveys have explored the application of DTs and AI for various sub-domains of power systems. However, to our knowledge, no comprehensive scoping review has been conducted to broadly identify key concepts and factors related to the synergetic application of DT and AI technologies, with a specific focus on driving sustainable performance of the sub-domains of power systems listed in Section 1.1. Therefore, this study can be considered the inaugural attempt at bridging this research gap by conducting a systematic scoping review to examine the research undertaken on this topic, identify and analyse the key concepts and characteristics of such research, and outline the knowledge gaps. The main contributions of this work can be summarised as follows:

• A comprehensive scoping review is conducted under the formal guidelines of Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR) statement for scoping reviews [43].
• This scoping review focuses on existing research in which DT technology is integrated with AI techniques for applications in electrical power systems. Existing research selected for this review harnessed the real-time information synchronisation and modelling capabilities of DTs as well as the core capabilities of AI such as knowledge representation, intelligent learning and cognition, data mining and pattern recognition, predictive modelling, data-driven optimisation, advanced analytics, and advanced perception capabilities (i.e., related to the processing of natural language, speech, and vision). In these research studies, the synergetic application of DT and AI technologies realised the ultimate objective of improving the data-driven decision-making process more effectively than when individually used, by complementing the capabilities of each technology.
• Selected research studies are classified under a novel classification framework covering the following areas of application in power grids: (1) elemental aspect: entire systems and individual power assets, (2) topological aspect: main power grids, smart grids, microgrids, systems with DG and RESs, and smart integrated energy systems, (3) sectoral aspect: generation, transmission, distribution, and demand side, and (4) functional aspect: planning, design and construction, operations, maintenance, and economics.
• Major characteristics and a variety of reviewed studies are discussed with respect to power systems, DT versions, and AI techniques.
• The evolution of DT and AI concepts under the previously stated scope of application is analysed individually. Moreover, three distinct evolutionary eras of the combined application of DTs and AI in power systems are identified and discussed.
• The environmental sustainability implications of the reviewed studies are outlined under three main objectives.
• Knowledge gaps, challenges, and future research directions under the topic are also discussed.

2. Background

This section is dedicated to outlining the background of the concepts discussed in this scoping review, with the aim of providing the readers with the necessary knowledge and directing them towards more focused, well-recognised resources for further reading. Initially, the evolution and definitions of DT technology are discussed, followed by a synopsis of the capabilities of AI and AI techniques relevant to power systems.

2.1. Digital Twins

Digital Twin (DT), a virtual representation of physical entities in cyber-space, can be considered a critical digital technology under the broad realm of CPS [6]. In fact, other CPS technologies such as big data, industrial IoT, distributed computing (i.e., cloud, fog, and edge computing), AI, and ML facilitate the integration, storage, seamless communication, and processing of diverse big data, while enabling analytics, predictive modelling, and intelligent decision making for DTs. When data are seamlessly transferred between the digital replica and the physical system, high-fidelity synchronisation and smooth operation can be achieved, while enhancing and facilitating planning, information exchange, learning, analysis, visualisation, and optimisation functions of the physical entities. Most importantly, DTs can serve as virtual platforms to observe, assess, and evaluate the current and future states of the physical entity, while visualising and suggesting optimum enhancements for the operation of the same. This capability of DTs is beyond the concepts of mere simulations or testing environments, due to the real-time, bidirectional communication, and intricate intelligence capabilities of DTs. Furthermore, full-fledged DTs have the capability to issue control commands followed by an intelligent, data-driven decision-making process, with the aim of optimising the actions and behaviour of their physical counterpart.

2.1.1. Evolution of the Digital Twin Concept

DT technology was initially conceptualised and introduced in the context of the manufacturing industry and the aerospace/aviation industry. Figure 2 represents the evolution of DTs as a concept from the 1970s to 2017, including important milestones such as the informal introduction of concepts, the formulation of formal definitions, and the publication of pioneering research articles. According to this evolutionary journey, three key stages of development can be identified with respect to the level of intelligence displayed by the DT in data-driven decision-making.

Stage 1: Prior to and during the early 2000s, DTs were passive digital representations of physical systems utilised for simulation and modelling [8,44]. Despite the fact that the virtual asset would continuously mirror the most recent representative attributes of the physical assets in a remote monitoring mode, such DT models cannot be considered as full-fledged DTs due to the lack of a seamless integration and real-time, two-way communication. Hence, it is prudent to define these virtual replicas as ‘Digital Mirrors’ [7,16,45] or ‘Digital Models’, as described later in this section. Nevertheless, these digital models have proven their expertise in addressing inefficiencies through real-time monitoring and proactive maintenance, especially in industrial applications.

Stage 2: During the late 2000s and early 2010s, the concept of Cognitive DTs was gradually unveiled [46,47,48]. This version of DTs can essentially mimic the human decision-making process under a specified set of rules and instructions to simulate known scenarios and optimise operations based on historical data by prescribing the most suitable action plan. On top of the predictive and prescriptive capabilities of cognitive DTs, the real-time, seamless connectivity between the virtual and physical entities was reinforced by the advent of big data and IoT technologies.

Stage 3: Starting from the 2010s, DTs were transformed to be more ‘intelligent’, by embedding self-learning and self-adapting capabilities. Such intelligent DTs can create and fathom novel knowledge and uncover hidden patterns without the input of structured instructions or rules to anticipate and respond to unknown scenarios. This stage of DTs is nourished by human–AI collaboration to become inherently resilient, where disruptions are anticipated, prevented, and managed proactively, while continuously optimising resource utilisation.

2.1.2. Definitions of Digital Twins

Consequent to the developmental stages of DTs outlined in Section 2.1.1, its definition has also been refined over time. As depicted in

Table 2. Comparison of digital twin definitions.

Number	Reference	Year	Definition
1	[46]	2010	“An integrated multi-physics, multi-scale, probabilistic simulation of a vehicle or system that uses the best available physical models, sensor updates, fleet history, etc., to mirror the life of its flying twin. It is ultra-realistic and may consider one or more important and interdependent vehicle systems”.
2	[49]	2013	“The coupled model of the real machine that operates in the cloud platform and simulates the health condition with an integrated knowledge from both data-driven analytical algorithms as well as other available physical knowledge”.
3	[50]	2016	“DT is a special simulation, built based on the expert knowledge and real data collected from the existing system, to realise a more accurate simulation in different scales of time and space”.
4	[48]	2017	“A set of virtual information constructs that fully describe a potential or actual physical manufactured product from the micro atomic level to the macro geometrical level. At its optimum, any information that could be obtained from inspecting a physical manufactured product can be obtained from its Digital Twin”.
5	[51]	2017	“Digital twinning is an approach to perform a real-time optimisation to a physical system using its digital copy”.
6	[52]	2018	“The digital twin is actually a living model of the physical asset or system, which continually adapts to operational changes based on the collected online data and information, and can forecast the future of the corresponding physical counterpart”.
7	[53]	2019	“A real mapping of all components in the product life cycle using physical data, virtual data and interaction data between them”.
8	[8]	2019	“DTs can be defined as (physical and/or virtual) machines or computer-based models that are simulating, emulating, mirroring, or ‘twinning’ the life of a physical entity, which may be an object, a process, a human, or a human-related feature”.
9	[54]	2019	“A Digital Twin is a virtual instance of a physical system (twin) that is continually updated with the latter’s performance, maintenance, and health status data throughout the physical system’s life cycle”.
10	[14]	2020	“A DT is a comprehensive software representation of an individual physical object. It includes the properties, conditions, and behaviour(s) of the real-life object through models and data. A DT is a set of realistic models that can simulate an object’s behaviour in the deployed environment. The DT represents and reflects its physical twin and remains its virtual counterpart across the object’s entire life cycle”.

, early definitions, such as definitions one, two, three and four, were limited to considering the DT entirely as a simulation that uses data obtained from physical models, sensors and other sources to replicate the behaviour of its real counterpart in a continuous manner. Such instances should be considered as ‘Digital Shadows’ due to the lack of bidirectional communication and interactions required to be defined as a DT, according to more comprehensive and accurate definitions introduced later in the developmental timeline. Although Definition 4, which was presented by Grieves, introduced the concept of ‘virtual’ and ‘physical’ counterparts more holistically, it also lacked the vital communication and controlling aspect of full-fledged DTs. Definitions 5 and 6 were constrained on the functional aspect of DTs, where the former emphasised real-time optimisation while the latter focused on forecasting the future states of the physical object. Moreover, Definition 7 can be considered as the inaugural attempt to include the three main components of a DT system: physical entity, virtual entity, and the interaction (i.e., communication link) between them, though the capabilities and functionalities of DTs were not included within the definition. Despite contributing to the progression of DTs, the limitations inherent in the definitions presented in Table 2 can be summarised as follows: they are constrained to mere simulations with continuous monitoring capability and they focus exclusively on either the component-level or functional-level perspectives, thereby not providing a holistic portrayal of full-fledged DTs.

While analysing the definitions of DTs, it is crucial to fathom the misconceptions of DTs to accurately identify the components and capabilities that must be embedded in DTs. Some common misinterpretations addressed by previous research are outlined in the following list.

• Digital Model: “A digital version of a pre-existing or planned physical object, to correctly define a digital model there is to be no automatic data exchange between the physical model and digital model. Examples of a digital model could be but not limited to plans for buildings, product designs and development” [13].
• Digital Shadow: “A digital representation of an object that has a one-way flow between the physical and digital object. A change in the state of the physical entity leads to a change in the DT, but not vice versa” [13].
• Digital Thread: “A communication framework that allows a connected data flow and integrated view of the asset’s data throughout its life cycle across traditionally siloed functional perspectives” [8].
• Product Avatar: “A digital counterpart, or a set of digital counterparts, of an ‘Intelligent’ or ‘Smart Products’, which have been developed to let any user or stakeholder access the attributes and services of the Smart Product during its whole life cycle” [8].

With this knowledge of the vital components, capabilities, and misinterpretations of DTs, the following definition by Mihai et al. [7] will be adopted to represent a full-fledged DT in the rest of this work. “Digital Twin is a self-adapting, self-regulating, self-monitoring, and self-diagnosing system-of-systems with the following properties: (1) it is characterised by a symbiotic relationship between a physical entity and its virtual representation, (2) its fidelity, rate of synchronisation, and choice of enabling technologies are tailored to its envisioned use cases, and (3) it supports services that add operational and business value to the physical entity” [7].

2.2. Artificial Intelligence

AI can be considered an integral part of the current digital landscape, which has transformed from a concept in science fiction to an everyday reality that encompasses a multitude of application domains including power systems. AI is a fusion of computer science, mathematics, and statistics focusing on the design and development of intelligent systems capable of emulating human behaviour in areas such as autonomous perception, cognition, reasoning, learning, decision-making, and action.

2.2.1. Capabilities of Artificial Intelligence

The capabilities of AI span a broad spectrum, encompassing key functionalities such as sensing, reasoning, assessing, inferring, predicting, and acting on data-driven decisions. The following list outlines these capabilities with respect to their utility in power systems.

• Perception and sensing: AI systems employ sensing mechanisms to perceive and interpret data from their environment. This involves utilising sensors such as cameras, microphones, IoT devices, or other input sources to acquire information in real-time. Moreover, the Natural Language Processing (NLP) and computer vision branches of AI stem from this capability, where sensory inputs such as speech, text, images, and videos are processed and analysed in an intelligent manner.
• Logical reasoning and inference: Logical reasoning and inference capabilities form the foundation for the decision-making process performed by AI systems. Logical and probabilistic reasoning techniques such as fuzzy logic, heuristic algorithms, Bayesian networks, and Markov models are used by AI to analyse information, draw meaningful conclusions, and make informed decisions. This is achieved by discovering hidden patterns and trends in the data, especially under ambiguous or uncertain conditions.
• Prediction of outcomes: Predicting involves forecasting future outcomes or trends based on historical data and probabilistic models. AI uses predictive analytics techniques, such as machine learning algorithms, to identify trends, detect anomalies, and anticipate future events.
• Performance of actions: This capability refers to the execution of actions or manipulation of the environment based on the outcomes of the logical reasoning, inference, and prediction processes. This involves controlling physical actuators, such as robotic arms or motors, or virtual interfaces to interact with digital environments. Acting capabilities enable AI to translate its decisions into tangible actions, enabling it to perform tasks autonomously.
• Evaluation of outcomes: AI systems can also assess the outcomes of actions taken by them to determine their effectiveness or appropriateness. AI employs feedback mechanisms to assess performance against predefined objectives or criteria, enabling it to refine its strategies and improve over time. Assessment capabilities are crucial to improving the efficiency, reliability, and adaptability of AI systems in dynamic environments. The application of Reinforcement Learning (RL) can be considered a classic example in which AI systems exhibit this capability.

2.2.2. Artificial Intelligence Techniques

The capabilities discussed above can be realised through multiple AI techniques nourished from probabilistic and data-driven concepts such as supervised learning, semi-supervised learning, unsupervised learning, Reinforcement Learning (RL), Deep Learning (DL), Monte Carlo methods, evolutionary algorithms, swarm intelligence, and expert systems. Figure 3 illustrates the main AI categories mapped to these capabilities, which are analysed with respect to power systems in this scoping review.

3. Methods

3.1. Research Design

For this scoping review, the process outlined in the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR) was followed [43]. The following research objectives are addressed in this review:

1.

To classify selected studies under a multi-faceted novel classification framework for electrical power systems as outlined in Section 1.2 and identify the main characteristics of the studies and use cases presented by them with respect to power systems;

2.

To identify and analyse the evolution of DT concepts applied in the studies, with respect to the evolutionary journey and definitions presented in Section 2.1;

3.

To identify and analyse the evolution of AI concepts utilised in the studies, with respect to the capabilities and techniques outlined in Section 2.2.

4.

To analyse the breadth and variety of studies with respect to the DT and AI techniques applied.

5.

To synthesise the relationships among power system concepts and use case categories, DT concepts, and AI techniques under the synergetic application of DT and AI in power systems.

6.

To discuss the environmental sustainability-related implications of the studies reviewed.

3.2. Identification of Relevant Studies

Studies with the potential to be included in this review were identified by conducting a comprehensive search in the following electronic databases: Scopus (i.e., Elsevier), Web of Science, and IEEE Xplore. This search was conducted on 26 June 2024 and the following preliminary criteria were adopted to initially identify relevant studies prior to the screening stage of the PRISMA-ScR process:

• Directly related to electrical power systems,
• Technological application is related to DTs,
• Technological application is related to AI,
• Full text available in English,
• Published in a scientific journal, magazine, book, book chapter, conference, or workshop and peer-reviewed (i.e., not partial content such as research abstracts, keynote presentations/speeches, dissertations, in-progress research, conference reviews, notes, letters, discussions, meeting abstracts, poster sessions, and blogs).

Since the evolution of the concepts of DT and AI should be identified as per the research objectives, no limit on the published year (e.g., including the most recent publications only) was imposed on potential studies as an eligibility criterion during the identification stage.

The electronic search strategy was constructed by combining keywords related to the previously-mentioned aspects of power systems, DT concepts and equivalent variations, and AI techniques. These keywords were combined in the search queries using the ‘AND’ and ‘OR’ operators so that the identified studies include the technological application of both DT and AI in the domain of electrical power systems. In Scopus, ‘Title-Abstract-Keywords’ were searched. ‘Topic’ was searched in Web of Science, which also includes title, abstract, and keywords. In IEEE Xplore, ‘All Metadata’ was adopted for the search as the most similar criterion available, which is equivalent to that of Scopus and Web of Science. ‘All Metadata’ includes abstract, index terms, and bibliographic citation data such as document title, publication title, and authors, etc. The electronic search query used in Scopus and Web of Science is included in Appendix A.

A total of 6818 studies were initially identified by adhering to these initial inclusion criteria and search strategies, where 3151 studies were obtained from Scopus, 2181 from Web of Science, and 1486 from IEEE Xplore. Out of these initially identified studies, 1677 duplicate studies were removed before proceeding to the screening stage of the PRISMA-ScR process.

3.3. Screening of Studies

After identifying relevant studies and removing duplicates, 5141 studies were screened. Screening was performed based on the title, abstract, and author keywords since these attributes were identified to be more relevant for pre-full text retrieval screening.

For this screening process, the following inclusion and exclusion criteria were used as an extension to the preliminary inclusion criteria used in the study identification stage in Section 3.2.

Inclusion criteria:

1.

Directly related to electrical power systems, encompassing the following aspects: (1) main (i.e., national) electrical grids, smart grids, microgrids, distributed generation systems, renewable energy systems, smart integrated energy systems, and energy internet, (2) generation, transmission, distribution, and demand side sectors, and (3) power system assets (e.g., generators, converters, inverters, transformers, measurement devices, control devices, batteries, transmission/distribution lines).

2.

Technological application is related to DTs and equivalent variations of DTs such as digital shadows, virtual twins, digital replicas, and real-time monitoring and control systems.

3.

Technological application is related to the broad domain of AI, encompassing techniques such as ML, fuzzy logic, probabilistic reasoning, swarm intelligence, evolutionary algorithms, NLP, computer vision, expert systems, and smart robotics.

Exclusion criteria:

1.

Related to other application domains such as manufacturing, industrial, commerce, education, healthcare, aviation, other utilities (i.e., water, gas, heating), telecommunication, and construction.

2.

Technological application is not focusing on the utilisation of equivalent variations of DTs and AI (i.e., merely related to other technological applications such as big data, IoT, mathematical and statistical modelling, electronics, and conventional software engineering).

Due to the significantly large number of studies identified under the broad scope of this review, manual screening (i.e., fully performed by humans) was not feasible. Therefore, the automated assistance of a Large Language Model (LLM) was obtained for the screening process, which improved the speed and efficiency of manual screening [55]. GPT 3.5 Turbo LLM [56] of OpenAI was used in this process, which is discussed in Appendix B. With the help of this LLM, the first author (A.R.) performed the screening and then two additional authors (Y.S. and K.H.) independently reviewed and validated it to ensure accurate inclusion.

Finally, 325 studies were included in the review followed by the title-abstract-keyword-based screening and the final review based on the full text. These screening processes are discussed comprehensively in Section 4. The following variables were obtained for data charting as meta-data of the publications included in the full-text review: Digital Object Identifier (DOI), full names of authors, title of the study, published year, publication title (i.e., title of the journal or conference where the study was published in), document type (i.e., conference paper, journal article, review, book, book chapter, editorial), abstract, author keywords, index keywords, publisher, citation count, author affiliations, and source (i.e., Scopus, Web of Science, IEEE Xplore).

3.4. Data Extraction and Analysis

In the full-text review process, concepts relevant to power systems, DT, and AI were extracted. With respect to power systems, a novel classification framework is introduced in this research. Each of the 325 articles was classified under this proposed classification framework. As illustrated in Figure 4, this classification framework encompasses multi-faceted aspects of electrical power systems. These aspects are as follows:

• Elemental nature of the target use case: whether DT and AI technologies are applied for an entire system with interconnected assets and components or for an individual power asset.
• Topological nature of the target use case: whether DT and AI technologies are applied for a main electrical grid, smart grid, microgrid, integrated energy system, distributed generation system, or renewable energy system.
• Sectoral nature of the target use case: whether the use case is in the power generation sector, transmission sector, distribution sector, or demand side (i.e., demand response, energy management and efficiency, energy storage).
• Functional nature of the target use case: whether the use case is related to the stages of planning, design and construction, operations, or maintenance in the power system life cycle, or whether it is related to an economic aspect of the power system such as electricity markets, peer-to-peer trading, capital investments, etc.

In addition to these aspects, the following data were extracted with respect to power systems:

• If the DT and AI implementation is done for an individual power asset, the type of that specific asset such as transformer, transmission and distribution lines, power electronic device, EV, and battery.
• If the DT and AI implementation is related to the power generation sector, and the source of generation, such as wind power, solar photovoltaic (PV), thermal power, nuclear power, and hydropower.
• Use case category and exact use case. Use case categories were pre-defined under each functional aspect as follows: (1) planning: power system planning, (2) design and construction: power system designing, (3) operations: reliability (i.e., applications related to power system protection, outage prevention, and fault management), stability (i.e., applications related to power system status monitoring, operational control, and optimal scheduling), loss reduction, cyber security, demand response (DR), demand forecasting, energy management and efficiency, energy storage, generation forecasting, power system digitalisation, environmental sustainability, training and education, (4) maintenance: maintenance management, predictive maintenance, and (5) economic: electricity trading, other use cases on electricity markets.

Regarding AI-related concepts, the AI domain, main category, sub-category, and AI technique (i.e., algorithm) were identified and extracted for each of the selected studies. AI domain was predefined, inspired by the capabilities of AI listed in Section 2.2.1. These domains are as follows: reasoning and inference, optimisation, machine learning, perception, and action. The main AI categories and sub-categories were also pre-defined and mapped to each of these AI domains. This mapping among AI domains, main categories, and sub-categories is outlined in

Table 3. Mapping of AI domain, main category, and sub-category.

AI Domain	AI Main Category	AI Sub-Category
Reasoning and Inference	Fuzzy Logic	-
	Probabilistic Models	-
	Expert Systems	-
Optimisation	Evolutionary Algorithms	-
	Swarm Intelligence	-
	Bayesian Optimisation	-
Machine Learning	Supervised Learning	Classification
		Regression
		Ensemble Methods
	Unsupervised Learning	Clustering
		Association
		Dimensionality Reduction
	Reinforcement Learning	Value-based Reinforcement Learning
		Policy-based Reinforcement Learning
		Deep Reinforcement Learning
	Deep Learning	Feedforward Neural Networks (FNN)
		Convolution Neural Networks (CNN)
		Recurrent Neural Networks (RNN)
		Unsupervised Deep Learning
		Generative Models
		Graph Neural Networks (GNN)
		Learning Vector Quantization (LVQ)
		Attention Mechanisms
		Transformers
	Monte Carlo Methods	-
	Emotional Learning	-
Perception	Natural Language Processing (NLP)	-
	Computer Vision	Object Detection
		Image Classification and Segmentation
		Target Tracking
Action	Smart Robotics	Sensor Fusion Robotics
		Planning and Control of Robotics

. If a selected study applied multiple AI techniques belonging to multiple domains, main categories and sub-categories, the technique with the highest performance (i.e., in terms of F-score or accuracy) or the primarily applied technique was selected for the data extraction and analysis process. For instance, if a study includes the application of an Artificial Neural Network (ANN), where a Particle Swarm Optimisation (PSO) algorithm is used for the purpose of optimising the hyperparameters of the ANN model, that study was classified under ANN as the AI technique. Moreover, if a study advocates the application of AI or ML in general, not limited to specific techniques (i.e., surveys and reviews), such studies were classified under the labels ‘AI in General’ or ‘ML in General’.

With respect to the domain of DTs, each selected study was classified under ‘Digital Twin’, ‘Digital Shadow’, or ‘Supervisory Control and Data Acquisition (SCADA) system’. A study was tagged under the ‘Digital Twin’ label if the authors(s) of that study explicitly mentioned that their use case is related to an intelligent DT with bidirectional communication, simulation, and predictive capabilities. Other digital representations such as ‘Digital Shadow’ and ‘Supervisory Control and Data Acquisition (SCADA) system’ were also considered as DT versions with lower intelligence and maturity, where the virtual and physical counterparts are in real-time synchronisation. A study was classified under ‘Digital Shadow’ if bidirectional communication between the physical and virtual entities is absent, as discussed in Section 2.1.2. In contrast, studies focusing on AI-powered SCADA applications were tagged under the label ‘SCADA’, with the rationale that SCADA systems are equivalent to ‘Digital Twins’ with respect to real-time monitoring, controlling, and bidirectional communication capabilities. The main difference between SCADA systems and ‘Digital Twins’ is that the latter is equipped with predictive and simulation capabilities, whereas the former is limited to reactive and monitoring capabilities. Therefore, both the “Digital Shadow” and “SCADA” concepts were considered as DT versions of lower intelligence and maturity when compared with the “Digital Twin” concept.

Data extraction under these characteristics was performed by two authors (A.R. and K.H.) independently, followed by the resolving of disagreements through discussion. In such disagreements, the second and third authors (D.A. and Y.S.) served as the conflict resolvers and provided the final decision.

To address the research objectives, an interactive dashboard was built as the main method of data analysis [57]. This dashboard was constructed using the extracted dataset for the finally reviewed 325 papers, which consisted of the meta-data attributes outlined in Section 3.3, full text, and the identified characteristics and concepts related to power systems, DT, and AI. The interactive visualisations allowed us to explore relationships and insights by combining various characteristics and attributes of the dataset. Most importantly, this dashboard facilitated a historical analysis of selected studies, where the evolution and progression of DT concepts and AI techniques could be identified. This dashboard has been made publicly accessible through the link in the supplementary materials. Sample screen captures of several pages from the interactive dashboard are shown in Figure 5. For a clearer analysis of these visualisations, the provided link can be used.

4. Results

4.1. Search Results and Selection of Studies

The selection process of studies and the search results are illustrated in Figure 6. As discussed in Section 3.2, 6818 studies were initially identified using the electronic search strategy in three databases. From Scopus, Web of Science, and IEEE Xplore, 3151, 2181, and 1486 studies were retrieved, respectively. Then, 1677 duplicate studies were removed and the rest of the 5141 studies were subjected to the first phase of screening (i.e., screening based on Title-Abstract-Author Keywords), as comprehensively discussed in Section 3.3. From this first phase of screening, it was identified that the primary focus of 2160 studies was not in the application domain of electrical power systems. For example, studies that focus mainly on water, gas, heating, transportation, and aviation systems were eliminated. In addition, 2574 studies were not related to the application of both DT and AI technologies. They were either focusing on only one of these two technologies or not focusing on any of the two technologies at all. For example, technological studies employing conventional software engineering or purely physics-driven mathematical optimisation techniques were eliminated at this stage. Altogether, 4734 studies were removed due to non-compliance with the primary inclusion and exclusion criteria outlined in Section 3.3.

The resulting number of studies whose full text was sought was 407. Out of this number, full text of 382 studies were available and those were subjected to the second phase of screening (i.e., full text-based screening). Since the full text-based screening was more comprehensive compared to the first phase of screening, fifty-seven more studies were deemed to be irrelevant to the defined scope and research objectives. Out of these fifty-seven studies, twenty-five studies focused on non-electrical subsystems of power systems such as mechanical, thermodynamic, fluid dynamic, telecommunication, and chemical subsystems. Since the main focus of this review is to analyse the electrical subsystems of power systems, such studies were eliminated at this stage. Furthermore, in twenty and twelve studies, respectively, no significant application of AI or DT technologies could be identified in the full text. To be more precise, merely the importance of applying AI and DT techniques was mentioned in the title, abstract, and author keywords of such studies. Finally, research objectives were addressed based on the remaining 325 studies.

4.2. Overview of Results

In this section, the main results identified from the review process and the interactive dashboards are presented. Firstly, characteristics and dynamics related solely to power systems are explored, followed by those of the DT and AI concepts. Then, the relationships among power system characteristics, DTs, and AI are mapped together to analyse the synergetic application of DT and AI technologies in power systems. Finally, the implications of such studies on environmental sustainability are presented.

4.2.1. Power System Characteristics

Figure 7 illustrates the distribution of the reviewed studies in the various aspects of the power system classification framework proposed in this work. With respect to the top-most classification layer, elemental aspect, 75% of the studies focused on entire systems consisting of multiple and diverse interconnected components and power assets. DT and AI techniques were applied for individual power system assets and components in the rest of the studies. Regarding the topological aspect, nearly 52% of studies were focused on the main, national power system, whereas distributed generation [58] and renewable energy systems [59,60] ranked in second place with 19% of the studies. Also, twelve studies could be related to integrated energy systems [61] where electrical power served as a major contributor to gas and thermal power. Regarding the sectoral aspect, the distribution of the studies was more balanced than that of the other aspects. Studies encompassing the generation, transmission, and demand side sectors were more prominent in quantity. There were sixty-one studies that applied DT and AI techniques across multiple sectors (i.e., multiple combinations of the main sectors) as well. With respect to the functional aspect, the distribution of studies is significantly unbalanced, with more than 90% of studies belonging to a use case in the operation stage of the life cycle of the power system. The maintenance function accounted for approximately 6% of the studies, while very few studies focused on the combined application of DT and AI techniques for planning [62,63,64], design [65,66] and economic [67,68,69] functions. A more granular and thorough analysis of the distribution of studies across these power system aspects can be found on page five of the interactive dashboard.

Regarding the studies that focus on the power generation sector, Figure 8 shows the count of studies under various sources of power generation and the topological aspect to which the studies belong. According to Figure 8, it is apparent that the vast majority of studies focused on power generation used renewable energy sources such as solar PV [59,70,71,72], wind power [60,73,74], pumped hydro [75], and conventional hydropower [76]. Consequently, most of such studies belong to renewable energy systems under the topological aspect. Also, there were twelve studies focusing on nuclear power generation [77,78], while eleven studies comprised multiple renewable energy sources.

Another main characteristic of this review under the power system domain is the distribution of power assets. The count of studies reviewed that focused on individual power assets under the elemental aspect of the power system classification framework is illustrated in Figure 9. In addition, this figure depicts the power system sectors to which the assets belong. More than 20% of these studies addressed a use case on transformers [79,80,81]. The vast majority of those transformers are power transformers found in the transmission sector of the power system. Battery systems ranked in second place with thirteen studies out of eighty-one (i.e., 16%) [82,83], whereas power electronic converters and inverters were also featured in a significant number of studies (i.e., seventeen studies) [84,85]. The rest of the major assets covered by the reviewed studies were transmission line conductors [86], EVs [87], and wind turbines [74]. The majority of batteries and EVs belonged to studies on the demand side sector of the power system, which can be attributed to the fact that energy storage is a salient component of DSM.

Finally, under the characteristics of power systems exhibited by the reviewed studies,

Table 4. Power system functions and use case categories of reviewed studies.

Functional Aspect	Use Case Category	Count of Studies
Planning	Power System Planning	3
Design and Construction	Power System Designing	2
Operations	Reliability (i.e., power system protection, outage prevention, and fault management)	72
	Stability (i.e., power system status monitoring, operational control, and optimal scheduling)	110
	Loss Reduction	8
	Cyber Security	12
	Demand Response	10
	Demand Forecasting	7
	Energy Management	20
	Energy Storage	5
	Generation Forecasting	23
	Digitalisation	10
	Environmental Sustainability	13
	Training and Education	1
	Multiple Use Cases	4
Maintenance	Maintenance Management	15
	Predictive Maintenance	7
Economic	Electricity Trading	1
	Other Use Cases on Electricity Markets	2
Total Studies Reviewed	-	325

represents the mapping between the functional aspect of power systems and the use case category addressed by the studies. According to this table, it is apparent that the vast majority of studies (i.e., 295 studies) addressed a use case under the ‘operations’ functional aspect of a power system. Out of such studies, use cases aimed at enhancing and maintaining the power system stability dominated with 110 studies (i.e., 34%). These studies addressed use cases encompassing real-time status monitoring [88,89,90], controlling of operations [70,84], and optimisation of power grid dispatching [61,91]. In contrast, there were 72 studies (i.e., 22%) aimed at improving the reliability and resilience of power systems, addressing use cases in the fields of power system protection [92,93], prevention of outages and blackouts [94,95,96], and prevention and management of faults [74,79]. Twelve studies were also reviewed that focused on cyber security aspects of integral electrical subsystems of the power system [97,98]. Moreover, DSM use cases such as DR [99,100,101], demand forecasting [102], energy management and efficiency enhancement [103,104,105], and energy storage [40,106] were included under the operations function of the power system. One study addressed a use case in the category of training and education, where a virtualised experiential learning platform was developed to train personnel of smart grids [107]. In contrast, ten and thirteen studies focused directly on use cases related to digitalisation [108,109] and environmental sustainability [110,111].

Under the maintenance function, seven studies included use cases on predictive maintenance [77,80], whereas the rest of the fifteen studies focused on general maintenance management tasks such as life cycle management of electrical equipment [112], maintenance scheduling [113], preventive maintenance [114], and lifetime prediction [115,116]. In contrast to the operations and maintenance functions of the power system, some qualitatively interesting use cases could be identified in the planning, design, and economic functions through this review, despite the quantitatively low number of studies that addressed these functions. Under the planning function, a power grid infrastructure project planning and management platform was introduced in [63] using a DT enhanced by a PSO algorithm and an ANN, whereas a DT-based solar PV potential assessment methodology was proposed in [64]. With respect to design, a methodology for intelligent design and construction coordination of nuclear power plants has been proposed in [65]. Moreover, in [66], a futuristic approach for the designing of HVDC (High-Voltage Direct Current) cable accessories is presented. Regarding the economic functions of the power system, an innovative DT approach for peer-to-peer energy trading cost minimisation is introduced in [67] employing PSO as the AI technique, whereas a K-means clustering method combined with a DT is presented in [68] for the prioritisation of decision-making experts and investments.

Pages five to seven of the interactive dashboard contain visualisations enabling a more insightful analysis of the power system characteristics outlined in this subsection.

4.2.2. Digital Twin Characteristics

Figure 10 illustrates the share of the DT versions defined in Section 3.4. According to this figure, it is apparent that the vast majority of the reviewed studies (i.e., more than 85%) enclosed an application of the most matured version of DT, which is defined as the ‘Digital Twin’. In contrast, nearly similar counts of studies focused on ‘Digital Shadow’ and ‘SCADA’ implementations that are enhanced by AI techniques.

The evolution of these DT versions is represented in Figure 11, through which the gradual progression of intelligence and other capabilities of the DT concept can be clearly identified. According to this figure, it is apparent that ‘SCADA’ implementations dominated the field of AI-powered digital representations in power systems until 2005. As discussed in Section 3.4, these implementations were not blessed with the predictive and simulation capabilities of DTs, but they had a real-time monitoring capability [92,117] along with bidirectional communication and controlling capabilities [95], which were essential for power system applications during that era. ‘Digital Shadows’ emerged in this field around 2006, which had real-time monitoring and simulation capabilities [97,118], minor predictive capabilities [119], and lower intelligence compared to ‘Digital Twins’. However, these implementations lacked the bidirectional communication inherent to ‘Digital Twins’. Since then and until around 2019, both ‘Digital Shadow’ and ‘SCADA’ implementations dominated the field. In 2019, ‘Digital Twins’ with a high level of intelligence, data-driven decision-making capabilities, predictive and prescriptive analytics, bidirectional communication capabilities, and simulation capabilities were introduced with the assistance of AI techniques [66,88,103,120,121,122].

This evolutionary journey significantly reflects the three key stages of development of the intelligence and data-driven decision-making capability of DTs, which are discussed in Section 2.1.1. With reference to these stages, the ‘SCADA’ DT version can be interpreted as the passive digital representation with real-time monitoring capabilities, which were dominant in the first stage. For instance, an expert system coupled with a real-time monitoring SCADA implementation was presented for transmission system alarm processing and protection control in [92]. Also, in [117], a fuzzy-logic and SCADA-based transmission system alarm processing and fault diagnosis system was introduced, while Zhang et al. [95] presented a blackout mitigation scheme for the transmission grid using a Support Vector Machine (SVM) as the AI technique and smart protective relays governed by a SCADA system. In contrast, the era of ‘Digital Shadows’ resembles the second development stage of DTs, where cognitive DTs with the ability to make data-driven decisions under known, predictable scenarios were unveiled. For instance, a test bed with real-time monitoring ability, advanced visualisations, and ML-based analytics is proposed for smart grid cyber security in [97]. Moreover, a real-time digital simulator (RTDS) for wind power prediction using Deep Learning (DL) is presented in [119].

Finally, the ‘Digital Twin’ version defined in this work can be connected to the third stage, which started after 2010. During this stage, the highest form of intelligence started to emerge with self-learning and self-adapting capabilities on top of the ability to discover hidden patterns in unknown scenarios [120,121,122]. Visualisations related to DT characteristics can be found on pages eight and nine of the interactive dashboard tool.

4.2.3. Artificial Intelligence Characteristics

A mind map illustrating the distribution of AI domains and the main categories defined in Section 3.4 can be found in Figure 12. According to this diagram, more than 68% of the studies stemmed from the ML domain, while 14% of the studies were related to the AI optimisation domain. In the ML domain, which is related to AI capabilities such as learning and cognition, prediction of outcomes, and evaluation of outcomes, the majority of studies belong to the categories of deep learning (DL) [61,73,79] and supervised learning [74,88,106]. Furthermore, a significant number of studies employed Reinforcement Learning (RL) as the main AI category [100,123,124]. Only a few studies included applications built using unsupervised learning [77,85,103], emotional learning [125], and Monte Carlo methods [101,126] for power system DTs.

With respect to the AI domain of optimisation, a major portion of the studies (i.e., forty-six studies) belonged to the category of swarm intelligence [70,127,128,129,130], while ten studies used evolutionary algorithms [131,132]. Of the studies reviewed in this work, sixteen studies originated from the logical reasoning and inference capability of AI, including expert systems [92,133], fuzzy logic [117], and probabilistic models [83,113] as AI techniques used for power system DT augmentation. Most importantly, a small number of studies promoted the application of computer vision [118,134,135] and NLP [136] in the perception domain of AI, along with smart robotic applications [137,138,139] in the action domain of AI.

Trends related to the distribution of the main AI categories and sub-categories are represented in Figure 13. With respect to the number of studies, Recurrent Neural Networks (RNN), regression, and swarm intelligence seem to be the leading sub-categories. These sub-categories belong to the main categories of DL, supervised learning, and swarm intelligence, respectively. This trend is explainable since the majority of use cases in power system DTs are focused on time-series forecasting and optimisation, where these categories excel. In contrast, the applications of the main categories such as computer vision, NLP, and smart robotic techniques were very limited, whereas RL, unsupervised learning, and evolutionary algorithms achieved moderate success as main categories in terms of count of studies.

Deep Learning: With respect to RNNs, the vanilla architecture of RNNs [102], Long Short-Term Memory (LSTM) [82], Gated Recurrent Unit (GRU) [98], and their variations (i.e., bidirectional LSTM, simplified LSTM, circuit LSTM, and GRU combined with attention mechanism) could be observed as prominent techniques related to the scope of this review. In addition, Convolutional Neural Networks (CNN) have also been applied in multiple studies [73,104]. Feed-forward deep neural networks such as Multi-Layer Perceptron (MLP) have also been applied in a significant number of studies [61,140]. In contrast, the application of subcategories such as generative neural networks [79,81], transformers [136], and Graph Neural Networks (GNNs) [141] were scarce under the considered topic.

Supervised Learning: Under the subcategory of regression, shallow ANNs [71,84,88], Nonlinear AutoRegressive eXogenous (NARX) network [106], and Extreme Learning Machine (ELM) [116] can be considered as the main techniques applied in the reviewed studies. Classification techniques have also been used in a significant portion of the studies, with Support Vector Machine (SVM) as the leading technique [95,142]. Logistic regression, decision trees, Naive Bayes, and k-nearest neighbour were also applied as classification techniques in the reviewed studies. Moreover, under the main category of supervised learning, ensemble methods such as random forest [86,143] and Extreme Gradient Boosting (XGBoost) [74,89], ranked after regression and classification subcategories.

Swarm Intelligence and Evolutionary Algorithms: With respect to swarm intelligence techniques, PSO and its variations have been leading in power system DTs [70,129], whereas a wide variety of other techniques such as the Whale Optimisation Algorithm (WOA) [127], Elephant Herding Optimisation Algorithm (EHOA) [128], and honeybee mating optimisation algorithm [130] were also utilised. Evolutionary algorithm-based optimisation techniques have also been applied for multiple studies, where the genetic algorithm and its variations dominated [131,132].

Reinforcement Learning: With respect to RL, deep RL was utilised in a large number of studies, with techniques such as Deep Q-network (DQN) [123], Deep Deterministic Policy Gradient (DDPG), and Twin-Delayed Deep Deterministic (TD3) policy gradient. In contrast, policy-based [144] and value-based [100,124] RL subcategories have also been applied in a considerable number of studies.

Probabilistic Models: According to Figure 13, probabilistic models stemming from the AI domain of reasoning and inference ranked after evolutionary algorithms. The main probabilistic techniques applied in the studies were the Extended Kalman Filter (EKF) [83] and the Markov model [113].

Clustering: Stemming from the main category of unsupervised learning, clustering was placed after probabilistic models in this figure with seven studies. K-means [77,103] and Self-Organising Maps (SOM) [85] are the leading techniques observed in clustering.

Computer Vision, NLP, and Smart Robotics: Despite the extremely low number of studies, an interesting observation that can be deduced from Figure 13 is the application of computer vision techniques in the subcategories of image classification [118] (for windmills), object detection [134] (for substation operation and maintenance), and target tracking [135] (for transmission grid equipment maintenance). In contrast, a Generative Pre-trained Transformer (GPT) technique can be found in the main category of NLP, which was employed for a chatbot [136]. Another unique trend is the use of smart robotics in the subcategories of sensor fusion robots [138] and robotic control [137,139]. Sun et al. [138] employed AI-powered smart robotics for power plant patrolling, environmental variable detection, equipment monitoring, and abnormal alarm detection tasks within a DT environment. In contrast, Shen et al. [137] and Park et al. [139] applied deep RL for robotic automation and control services.

The evolutionary timeline of the previously discussed main AI categories and major milestones is depicted in Figure 14. According to this chart, a dramatically rising trend of using AI for power systems can be observed, especially after the era of 2015–2017. This can be related to the advent of the exact term ‘Digital Twin’ in 2017, which is discussed in Section 2.1.1. Also, in line with the evolution of DT versions discussed in Section 4.2.2, it can be observed that expert systems and supervised learning were the major AI contributors for primal versions of DTs (such as ‘SCADA’ DT version) prior to the 2000s. Between 2000 and 2010, fuzzy logic and DL (i.e., mainly Multi-Layer Perceptrons) started to emerge, resembling the introduction of advanced logical reasoning and primitive predictive capabilities witnessed in ‘Digital Shadow’ implementations. Then, post-2015, swarm intelligence and Monte Carlo methods started to appear in power system DTs. There are three main milestones identified for the period after 2020, which marked the advent of highly intelligent, adaptive ‘Digital Twins’. Initially, starting in 2021, RL techniques flourished. This trend enabled the application of self-adapting DTs in power systems, which can optimise and alter their behaviour under unknown dynamic scenarios. Then, around mid-2020s, DL methods such as RNN and CNN dominated the field under consideration, along with evolutionary algorithms, smart robotics, as well as probabilistic models. This milestone further enhanced the capability of ‘Digital Twins’ to perform actions and make accurate data-driven decisions under complex system dynamics. Finally, since 2023, perception techniques such as computer vision and NLP have started to emerge, enabling the expansion of perception and sensing capabilities of DTs with advanced intelligence.

4.2.4. Outcomes of the Synergetic Application of DT and AI in Power Systems

In addition to the results presented under the evolution of DT and AI in the scope of this review, Figure 15 represents the three major eras identified with respect to the distinct evolutionary trends exhibited by DT versions, AI categories, as well as power system use case categories. This figure clearly illustrates the gradual development of intelligence and the maturity of DT versions, along with the evolution of AI techniques.

The first era (i.e., 1990 to 2005) was characterised by the ‘SCADA’ version of DTs, coupled mainly with expert systems, fuzzy logic and supervised learning techniques. With respect to the DT capabilities, these ‘SCADA’ implementations excelled in real-time monitoring and controlling functionalities, powered by bidirectional communication. These DTs were connected with various power system sensors and actuators to practically realise their capabilities. However, they were not equipped with intelligent predictive and simulation capabilities, thus being limited to reactive, passive digital representations. With respect to AI capabilities, these DTs were predominantly restricted to logical reasoning and inference, mainly targeted at providing diagnosis and decision support. For instance, Bernard and Burocher [92] introduced an expert system integrated with the SCADA system of a regional transmission control centre, which could continuously analyse alarm messages and diagnose the origin and consequences of transmission network faults. A similar transmission system fault diagnosis use case can be found in [117], where an abductive fuzzy knowledge-based system was coupled with the SCADA alarm processing unit. Furthermore, it can be observed that prior to 2005, the prominent use case category was reliability enhancement [92,133], encompassing use cases such as power system protection, outage prevention, and fault detection. This can be attributed to the fact that ensuring reliability and continuous supply was the primary concern for power grid operations during that time.

Between 2006 and 2019, the second distinct era can be identified, where ‘Digital Shadows’ augmented by DL, supervised learning, swarm intelligence, and Monte Carlo methods emerged as the prominent DT version. Owing to the application of these AI techniques, ‘Digital Shadows’ were equipped with predictive capabilities and more advanced learning, cognition, and optimisation capabilities when compared to ‘SCADA’ DTs. They also had real-time monitoring and simulation capabilities, despite lacking bidirectional communication and control capabilities. For example, a swarm-intelligence-based optimisation scheme is proposed in [145], where a real-time monitoring system consisting of smart meters is utilised for distribution network operations. Also, Oyewumi et al. [97] presented a test bed with real-time monitoring ability, advanced visualisations, and ML-based analytics for smart grid cyber security. During this era, on top of reliability-based use cases, power system stability management and operational optimisation [88] use cases started to emerge. Moreover, use cases in cyber security [97], as well as power grid loss reduction [146], were introduced in the era of ‘Digital Shadows’, marking a shift towards improving power system operational stability, security, and efficiency, in addition to merely maintaining reliability.

With the advent of more intelligent and advanced ‘Digital Twins’ post-2019, the third and final era can be identified. These ‘Digital Twins’ exhibited bidirectional communication, control, as well as simulation capabilities, thus serving as full-fledged versions of DTs. During this era, comparatively new and more advanced AI techniques such as RL, RNN, CNN, smart robotics, computer vision, and NLP accentuated the capabilities of ‘Digital Twins’, by providing advanced predictive and prescriptive analytical capabilities, perception-related capabilities, and smart decision-enabling capabilities. Moreover, the ability of intelligent DTs to discover and analyse hidden patterns under uncertain, dynamic conditions was significantly enhanced by self-learning and self-adapting capabilities [120,121,122]. Concurrent with the advancement of intelligence, the use case categories also expanded to cover a diverse range during the era of ‘Digital Twins’. DSM-related applications such as DR [99,100,101], consumer energy management [103,104,105], energy storage [40,106], load forecasting [102], as well as generation forecasting [72,143], sustainability [110,111] and digitalisation [108,109] started to emerge during this era. This is an interesting observation, which exhibits how the focus of power system operations broadened to include use cases in DSM, digitalisation, and sustainability while ensuring stability and reliability as primary requirements.

Most importantly, these three eras, marking the evolution of the synergetic application of DT and AI technologies, can also be used as a directive for future research on the same topic. It advocates that future research on power systems, where DT and AI techniques are applied jointly, must harness the logical reasoning, inference, decision support, bidirectional communication, and real-time monitoring capabilities exhibited by the era of ‘SCADA’. On top of that, the predictive learning and optimisation capabilities of the ‘Digital Shadow’ era must be embedded into future research, along with advanced visualisation and simulation functionalities. Moreover, drawing inspiration from the third era of ‘Digital Twins’, future research must embrace self-adapting, self-regulating, self-monitoring, and self-diagnosing capabilities, utilising advanced AI techniques such as RL, DL, computer vision, NLP, and smart robotics. Concerning the variety of use cases, this evolutionary road map indicates that future research on the synergetic application of DT and AI must focus on under-explored areas such as power system planning and design, environmental sustainability and carbon emission reduction, digitalisation, training and development of personnel, economic use cases and electricity markets, maintenance, and DSM.

More insightful and granular analysis of the relationships between power system, DT, and AI characteristics can be performed through pages seven to nine and thirteen to fifteen of the interactive dashboard tool. To complement such an analysis, Figure 16 is provided in this work, which depicts the mapping between the power system topological aspect, sectoral aspect, DT version, main AI category, sub-AI category, and AI technique. A more interactive visual of the same figure can be found on page fifteen of the dashboard tool. The main trends and patterns that can be observed from this visual are listed as follows:

• Under main electrical grids, transmission sector applications dominated in terms of the count of studies, with ‘Digital Twin’ implementations complemented by DL techniques such as RNN and CNN.
• In renewable energy systems and DG systems, generation sector use cases with ‘Digital Twin’ versions augmented by DL techniques such as RNN and feed-forward neural networks were dominant.
• Demand side use-cases contributed to the majority of microgrid and smart grid-related studies, where swarm intelligence techniques were applied primarily under ‘Digital Twin’ versions.
• In integrated energy systems, use cases applied across multiple sectors such as generation, transmission, distribution, and demand side dominated in terms of count of studies. Such use cases mainly employed deep RL techniques.
• Scarce but unique use cases involving computer vision, NLP, and smart robotic techniques originated primarily from the main electrical grid transmission sector, smart grid demand side sector, and main electrical grid generation sector, respectively.
• The majority of the ‘Digital Twin’ implementations employed techniques from ML (i.e., 187 out of 277 studies) and optimisation (i.e., 45 out of 277 studies) AI domains. The vast majority of ‘Digital Shadow’ versions applied ML techniques (i.e., seventeen out of twenty-one studies), whereas ‘SCADA’ versions mainly used techniques from ML (i.e., eighteen out of twenty-seven studies) as well as reasoning and inference (i.e., six out of twenty-seven studies) AI domains.

4.2.5. Environmental Sustainability Implications

Use cases aimed at improving environmental sustainability could be identified under the explicit use case category of ‘environmental sustainability’ under the ‘operations’ functional aspect of power systems, as well as other use case categories such as renewable energy-related use cases.

Table 5. Distribution of studies across environmental sustainability-related objective and publication year.

Environmental Sustainability Objective	2021	2022	2023	2024
Increase renewable energy share for power generation	[60,70,71,74]	[61,73,82,89]	[72,143,147]	[148,149,150,151,152,153,154,155,156]
Improve energy efficiency in demand side	[103,157]	[105,110,111,158]	[159,160]	[156,161]
Improve energy efficiency in generation, transmission, and distribution sectors	-	[140,162,163]	[164]	-

provides a summary of the most influential studies distributed across the main objectives of such use cases. As depicted in this table, three main objectives are identified for the studies primarily focused on addressing environmental sustainability-related challenges, applying DT and AI techniques. Of these three objectives, increasing the renewable energy share in the power generation sector is the most prominent one, which is followed by enhancing energy efficiency at the demand side as well as the supply side of the power system.

With respect to the first objective in Table 5, the expansion of renewable energy share, there is a multitude of studies reviewed in this research. However, only the most significant studies and studies directly focusing on environmental sustainability as the use case category are listed in the above table. Most of such studies were published in 2024, which is evident from Table 5. Out of those studies, Wang et al. [148] proposed a ‘Digital Twin’ for renewable energy generation prediction in a sustainable microgrid, employing evolutionary algorithms such as the AI technique. Similarly, Zhong and Li [152] presented a methodology for solar PV output prediction and operations optimisation in sustainable cities using the lion optimisation algorithm in a ‘Digital Twin’ environment. In addition, a methodology for financing net-zero renewable energy integration in smart cities is presented in [149], where a novel load-balancing approach is employed using the time-varying artificial bee colony optimiser as the AI technique. Moreover, novel methods for dispatch optimisation in renewable-dominated microgrids are presented in [150,151,155], where evolutionary algorithms, RNNs, and swarm intelligence were used, respectively. Chen et al. [154] presented a renewable energy output forecasting methodology for smart cities using GRU, while the optimised dispatching of renewable sources was realised through an ant colony optimisation algorithm.

The second objective driving environmental sustainability through DT and AI techniques is demand-side energy efficiency improvement. In references [110,159], improving energy efficiency and promoting carbon-neutral consumption patterns for data centres is the main focus, owing to the fact that data centres are one of the most energy-intensive consumer categories in the present world. For [159], a deep RL technique was applied through a ‘Digital Twin’ implementation. Dulaimi et al. [111] introduced a DT-based energy hub for sustainable smart cities, aimed at building energy management and energy-saving initiatives. In [156], the modified flower pollination algorithm (MFPA) was utilised in a ‘Digital Twin’ for short-term load forecasting, which is directed at promoting efficient energy consumption for net-zero emissions in smart cities. Another smart city-based DT implementation promoting building energy efficiency can be found in [161]. Also, Cicirelli et al. [157] proposed a ‘Digital Twin’ enhanced by the Deep Q Network for balancing thermal comfort constraints when minimising energy consumption in heating, ventilation, and air conditioning systems. A DT and AI-based energy efficiency management framework is proposed in [158] for smart campuses. Furthermore, in references [103,105], a ‘Digital Twin’ for net-zero energy buildings is presented, where unsupervised learning techniques were employed for initiatives improving energy efficiency.

Finally, a few studies addressing energy efficiency and carbon neutrality in the non-demand side sectors of the conventional power grids are also listed under the third objective of Table 5. In [140], a low-carbon electrical equipment management strategy is proposed by utilising an MLP network from DL. Xie et al. [163] proposed a five-dimensional structure model of a DT for the power distribution network, promoting carbon neutrality. Also, in [164], improving energy efficiency in the generation, transmission, and distribution aspects of microgrids is discussed.

5. Discussion

With respect to the synergetic application of DT and AI technologies for power systems, the breadth and variety of power system aspects, AI techniques, and DT implementations were explored in the form of a systematic scoping review in this work. According to the results of this review, the following research gaps, challenges, opportunities and future research directions could be inferred. Also, the role of DT and AI in overcoming technical challenges associated with transformative power grid concepts such as smart grids, microgrids, DG, and RES is discussed in this section.

5.1. Research Gaps

Regarding the power system dimension of this review, a major research gap could be identified with the limited focus on functional categories beyond the operations stage of the power system life cycle. Despite the critical role played by plans, designs, and economic functions in the stability, reliability, efficiency, affordability, and long-term sustainability of power systems, merely a fraction of 10% of studies addressed these areas in the form of AI-powered DTs.The application of AI and DTs could be transformative in these rarely considered functions, especially in power system designing and planning stages where the capabilities of proactive and multi-faceted analysis must be coupled with 3D visualisation and simulation features. For instance, other application domains such as product design and development [165], smart building design [166], and smart city design [167] utilised the assistance of both DT and AI technologies, owing to their excellent capabilities in data-driven decision making and immersive visualisations, respectively. Nevertheless, the highly skewed focus on operations and maintenance functions in AI-powered power system DTs suggests an over-reliance on routine operations, at the expense of enhancing the long-term sustainable performance.

Even under the operations function in power systems, more than 60% of research reviewed in this work focused on power system reliability and stability-related use cases such as protection and outage prevention, detection of anomalies and faults, status monitoring, operational control, and optimisation of dispatching. Regarding the main electrical grid, applications of AI-powered DTs for loss reduction, power quality enhancement (e.g., management of harmonics, voltage and current regulation), power flow analysis, transient stability analysis, and frequency regulation were limited, whereas these topics were covered significantly under other technological research aspects. Moreover, DSM use cases, including energy management, DR, and energy storage, were marginal as per the results of this review.

Another research gap identified with respect to power system characteristics is the under-representation of studies focusing on individual power system assets and components. Even out of the 25% of studies proposing AI-powered DTs for individual power assets, the vast majority focused on power transformers and batteries. Consequently, the coverage of other vital assets such as power electronic converters and inverters, line conductors, protective switchgear, measurement devices, generator units, wind turbines, and solar panels was very low. Also, only one study presented a DT for electric motors, which was augmented by AI [168], whereas electric motors and drives are key components in power systems.

Some notable research gaps could be identified with respect to the narrow or limited use of AI techniques within this review. One such prominent gap is the under-utilisation of AI domains beyond ML and optimisation. While ML techniques, particularly supervised learning, DL and RL, have dominated the field, other AI domains such as logical reasoning and inference, perception, and action remain under-explored. Though this trend might be traced to the fact that the majority of use cases in power system DTs require learning, cognition, forecasting, prediction, and optimisation capabilities, there is an abundance of more advanced AI techniques with these capabilities beyond the predominantly used ones. For instance, sophisticated and increasingly prominent DL techniques such as generative models, transformers, and GNNs were not utilised satisfactorily in the reviewed research. These techniques have exhibited promising results in DT research of other application domains, mainly due to their advanced, more nuanced predictive and analytical capabilities. Specifically, generative models such as Generative Adversarial Networks (GAN) and auto-encoders have been applied for time-series forecasting use cases in power systems with a commendable performance. In fact, GANs have been employed for accurate forecasting of renewable power generation under an interpretable as well as controllable setting. This way, volatile, stochastic characteristics of renewable generation could be captured without model-based, statistical assumptions [169,170]. In contrast, GNNs have proven their ability to capture temporal as well as spatial dynamics of power system applications, through a variety of flavours such as graph-CNNs, graph-RNNs, graph-RL, graph generative models, and graph attention networks. When combined with conventional DL methods such as CNN, RNN, and attention mechanisms, the unique and inherent ability of GNNs to detect hidden patterns in highly complex, spatiotemporal data could be enhanced. This capability has been extended to use cases without DTs, where the graphical network structure is relevant (i.e., power flow calculation and fault detection), as well as where such a physical graph structure is irrelevant (i.e., time-series prediction and scenario generation) [171]. Nevertheless, the depth and breadth of research conducted on GNN-powered DTs has been minimal so far. Simultaneously, very few unsupervised learning techniques were found in the reviewed research, which was also mainly focused on K-means clustering. No use case applied association rules, dimensionality reduction techniques or more sophisticated clustering methods beyond K-means and SOM, which would have added a significant impact on use cases such as entity profiling for electricity markets and load profiling for DSM measures.

Despite very few interesting use cases on computer vision, NLP, and smart robotics, the representation of these AI categories has also been insufficient, in contrast to the drastically rising research and industry attention received by these in the present world. Specifically during the Industry 4.0 and 5.0 era that we are currently living in, these AI categories have proven to be versatile and transformative, though their performance must be improved continuously with respect to interpretability, accountability, and data governance aspects that are critical for power systems. The only NLP application reviewed in this work was an interactive chatbot implemented for providing demand-side energy management recommendations, whereas NLP techniques have the capability to address a wide variety of use cases beyond that. Use cases on speech recognition and synthesis, machine translation, text summarisation, text classification, information extraction and retrieval, topic modelling, and sentiment analysis were not explored through the reviewed research. Under computer vision, the two use cases reviewed in this work addressed object detection and target tracking, whereas image classification and segmentation could have also contributed significantly to power asset management, vision-based fault and anomaly detection, and inspection use cases in power systems when coupled with the immersive visualisation capabilities of DTs.

Furthermore, only perception, learning, and optimisation are not sufficient for the data-driven decision-making process in DTs. Executing actions and enabling actions is where the real impact of data-driven decisions lies. Smart robotics is a major AI category that enables this fulfilment of decision-making when coupled with DTs, which has been proven through numerous research, especially in the manufacturing industry [12]. Smart, AI-driven robotics may not be relevant for power system use cases as much as for industrial use cases. However, it can be harnessed to handle complex, yet critical use cases in power systems, especially where the safety of humans must be considered and where the physical workload of humans can be reduced. For instance, the research on smart robotics coupled with DTs in power systems had proven their versatility under power plant patrolling [138], equipment replacement under hazardous conditions [139], and equipment sorting [137], which could have been extended to other use cases as well.

Furthermore, the reviewed research has not sufficiently explored the synergetic power of DT and AI with respect to the self-adapting, self-regulating, self-monitoring, and self-diagnosing capabilities of intelligent DTs, as discussed in Section 2.1.2. The vast majority of research focused merely on the predictive, analytical, and simulation capabilities of AI-powered DTs, neglecting these transformative intelligent capabilities. For instance, self-diagnosing and self-healing capabilities (i.e., capable of detecting, locating, and analysing faults and then executing appropriate corrective action with minimum human intervention) of AI-powered DTs could have been highly important to maintain power system reliability and resilience [172]. Also, regarding the implications on environmental sustainability unveiled in this review, it is apparent that the majority of studies focused on contributing to a carbon-neutral, sustainable setting in the power system through renewable energy integration and energy efficiency under DSM. No studies have harnessed the power of both DT and AI for other, more direct sustainability objectives such as the reduction in GHG emissions in thermal power generation, the application of low-carbon equipment in power grids, or the reduction in waste generation in power systems. However, previous research such as [173,174] advocates that AI-powered DTs could be used to monitor, estimate, and reduce greenhouse gas emissions in a more direct manner.

5.2. Challenges

Apart from the research gaps, various challenges related to the topic could be discovered, which range from technological and mathematical aspects to social and regulatory aspects. Mitigation strategies for these challenges are also discussed in this section to encourage and guide future researchers in implementing AI-augmented power system DTs.

5.2.1. Technological Challenges Related to the Application of Digital Twins

Specifically related to DTs, one major challenge would be the seamless and scalable integration across different sectors of the power system. A full-fledged DT itself is a seamless integration of layered modules for data acquisition, transmission, processing, and analysis. Therefore, significant technical challenges would be encountered when integrating such DTs with power systems, which are also a complicated, dynamic interaction of numerous elements. It could even be complicated to accurately model the behaviour of a single, individual power asset governed by inherently complex technical dynamics such as regulating electrical parameters, maintaining stable operations, and managing anomalies under highly uncertain, volatile environments. Hence, the seamless, real-time integration of entire systems where multiple assets are interconnected, could be even more cumbersome. While the recent advancements in the fields of IoT and telecommunication could assist in managing data collection and transmission challenges of DTs, interoperability and scalability are pressing issues in these regards. In fact, power systems employ a diverse range of communication protocols under a broad spectrum of supplementary hardware and software systems such as SCADA, wide area monitoring systems, and Enterprise Resource Planning (ERP) software. Even presently, the majority of these systems are of a legacy nature, which are used together with modern systems. Hence, integrating DTs to power systems would require interoperable and unified communication standards with industry-wide recognition. Moreover, most of the power system DTs have been restricted to prototype implementations in contrast to production-grade versions, where scalability will be another major challenge. While CPS technologies such as cloud computing, fog computing, and big data management techniques could alleviate data processing and analysis-related challenges, more complex technical challenges could arise with the gigantic volume, diversity, and velocity of power system data streams. These factors do not only point out technical challenges but also economic challenges due to the resource-intensive requirements to plan, design, develop, and manage power system DTs.

Moreover, the secure operation of power system DTs could be a major challenge, though a few studies reviewed in this research addressed issues related to cyber security [97,98]. Recent advancements in the field of cyber security such as blockchain could pave the path to managing such issues, especially mitigating and managing cyber attacks in digital systems and ensuring secure transactions in energy trading and other electricity market-related applications [175,176]. Another technical hurdle will be the application of AI-powered DTs across multiple sectors of the electrical power system since the studies reviewed in this work were primarily focusing on a single sectoral aspect. In such scenarios, diverse power engineering principles and technical aspects of multiple sectors such as generation, transmission, and demand side must be considered conjointly.

5.2.2. Technological Challenges Related to the Application of Artificial Intelligence

With respect to the application of AI in power system DTs, interpretability is a major concern. Most of the sophisticated, more advanced versions of AI like DL inherently lack the interpretability and accountability required by critical power system use cases such as transmission grid control, system dispatching, and outage management. Despite their merits in handling complex data and uncovering hidden insights, such black-box models are often rejected by power system experts and professionals, mainly due to the need to explain the rationale behind a decision under critical conditions. Furthermore, accuracy, precision, as well as the speed of decision making is of utmost importance to power systems, since all other sectors of the economy such as healthcare, education, commerce, and transport are reliant on the stable and reliable operations of the power system. Also, ensuring that AI models do not exhibit bias in their decision-making processes could be challenging, which could lead to unequal access to resources or services. Therefore, managing the fairness, controllability, accountability, and transparency of complex AI models could be another major hurdle, which extends beyond a mere technical challenge.

5.2.3. Technological Challenges Related to the Application of NLP, Computer Vision, and Smart Robotics

In addition to the previously discussed general challenges related to the application of AI in power system DTs, there are several specific challenges affecting the utilisation of NLP, computer vision, and smart robotics. Data integration and standardisation is one such primary challenge in implementing advanced AI technologies like NLP, computer vision, and smart robotics into power system DTs, which involve the generation of vast amounts of data from numerous sources under diverse formats and information standards. Inherently, NLP and computer vision belong to the ‘perception’ (i.e., sensing) domain of AI, as discussed in Section 4.2.3. Therefore, power system DTs augmented by NLP and computer vision models must manage the integration, storage, pre-processing, cleaning, and processing of unstructured data in the form of text, speech, images, and videos. In contrast, despite belonging to the ’action’ domain of AI, smart robotics applications in power system DTs also involve the integration of various sensors to obtain spatial measurements for movement control and to measure environmental variables such as temperature and lighting levels [138]. Hence, the challenge stems from maintaining consistency across different formats of these data types, while integrating, storing and processing those. This becomes more complicated when multi-modal data streams obtained from these AI techniques must be integrated and standardised for unified processing and analysis.

In addition to that, real-time data processing requirements associated with NLP, computer vision, and smart robotics are computationally demanding. For instance, most NLP use cases in other application domains require high processing power to derive insights from large amounts of textual and speech data. This requirement is even more stringent in power system DTs, where real-time or near real-time extraction of inferences and predictions must be performed, especially for reliability and stability management use cases like fault detection and dispatch optimisation. The same applies to computer vision use cases, dspecially for physical asset monitoring using drones or stationary cameras [135]. Also, smart robotics deployed for maintenance and inspection [138] use cases generate high-bandwidth sensor data, which require excellent processing capacity and hardware infrastructure. These processing requirements become more complicated when the data integration and processing must happen in parallel to coordinating with the other components and modules of the DT. The requirement for robust communication infrastructure is another challenge for the application of NLP, computer vision, and smart robotics in power system DTs. Latency and bandwidth constraints can be impactful for data transmission in power grids with RES and DG due to their geographically dispersed nature. This is intolerable in real-time computer vision applications deployed for surveillance and monitoring and in smart robotics applications where input/output data streams used for sensing and controlling must be transmitted under low latency and high quality. Specifically, if these computer vision and smart robotics applications are deployed in remote, inaccessible, and hazardous locations such as nuclear reactors and rural power plants, bidirectional data transmission can be even more challenging.

Despite the dynamic growth exhibited by the domains of NLP, computer vision, and smart robotics recently, the accuracy and precision of such AI models can be another significant challenge. Since most of the power system applications are mission-critical, where erroneous performance could lead to complete blackouts disrupting other sectors of the economy as well, model accuracy is of paramount importance. Also, due to the unavoidable complexity and volatility of grid conditions under RES and other factors, the risk of generating inaccurate predictions or misinterpretations through power system DTs can be higher than other application domains. For example, if NLP models are used for processing maintenance logs or operational reports in power plants and substations, accurately interpreting domain-specific language can pose a critical challenge, especially if not trained on sufficiently comprehensive datasets. Moreover, computer vision models deployed in power system DTs might struggle with high environmental variability under demanding outdoor conditions. Should these models be implemented for damage detection or asset failure prediction, inaccuracies could lead to unnecessary maintenance actions or missed opportunities for preventive measures. The same applies to AI-powered smart robotics, where mechanical damage could occur or critical tasks would not be performed under inaccurate perception or decision-making. Additionally, these advanced AI techniques can pose challenges to the security and data privacy of power systems, owing to their increased reliance on large-scale data acquisition and processing in cyber-physical and cloud environments. Similar to inaccurate performance, lack of cyber security measures can also cause the collapse of national power grids as well as entire economies. Most importantly, NLP and computer vision models processing sensitive operational data on power generation and transmission are susceptible to malicious attacks. Also, these vulnerabilities can affect the data privacy of consumers or prosumers connected to power grids if the breached data contains sensitive, personal data. Hence, ensuring the accuracy, security, and data privacy of these power system DTs is a pressing requirement.

Another major challenge discovered through the NLP, computer vision, and smart robotics research reviewed in this work is the requirement for interdisciplinary collaboration. These advanced AI techniques must be implemented in power system DTs with the active collaboration of NLP and computer vision experts, linguistic experts, IoT engineers, AI engineers, robotics and automation engineers, and also power engineers. However, this kind of collaboration will be challenging due to interdisciplinary knowledge gaps that will hinder effective implementation. All of the domains relevant to building a power system DT, including AI, NLP, computer vision, robotics, IoT, software engineering, and power engineering, are highly sophisticated and specialised fields, resulting in these interdisciplinary knowledge gaps. For example, the NLP experts will not always completely understand the power system domain knowledge and business requirements provided by power engineers. Conversely, power engineers will not always fathom the computational requirements and limitations of building an NLP model, leading to potential conflicts and unacceptable systems. Therefore, it is vital to properly communicate and understand these interdisciplinary requirements while collaborating on a project to develop a power system DT integrating NLP, computer vision, and smart robotics.

5.2.4. Mathematical Challenges

The mathematical challenges associated with AI-powered DTs primarily stem from their computational complexity, scale, and dynamic nature. Inherently, power systems contain variables with high dimensionality, especially under DG and RES systems. Thus, incorporating these variables into the AI models embedded in DTs creates a computationally challenging situation. However, with the help of AI techniques such as Principal Component Analysis (PCA) and autoencoders, the dimensionality of the problem can be reduced without losing critical information. In addition to that, the complex and high-dimensional nature of power systems poses scalability issues when the AI-powered DTs must be scaled up to the level of an entire power grid, from prototype and singular implementations. Also, performing optimisation tasks in this situation can be computationally expensive, which can be mitigated with the use of distributed and parallel processing. Another strategy would be to embrace hybrid modelling techniques (i.e., grey box modelling), which can balance the benefits of data-driven modelling and mathematical modelling, ensuring that physical variables and constraints are taken into account while improving computational costs. In contrast to purely data-driven (i.e., black box) models or physical (i.e., white box) models, these hybrid models can be highly effective and efficient in handling the complexity inherent to power system modelling. For instance, Manfren et al. [177] combined physical interpretations and data-driven predictions to provide interpretable insights into building energy performance using a DT.

Another major mathematical challenge can be attributed to the stochastic nature of RES such as solar and wind. Under these stochastic variations of RES, the degree of uncertainty and probability that must be managed by the AI and DT models is unavoidably increased. Stochastic optimisation techniques such as stochastic programming, chance-constrained optimisation, and robust optimisation can manage uncertainty in power system data streams. Also, probabilistic AI models such as Bayesian optimisation, Markov models, and Kalman filters can assist in this regard [113]. Moreover, non-linear systems are highly common in power system use cases, especially in system stability and reliability analysis. These non-linear properties are introduced by the complex dynamics that arise from interactions between generation, transmission, distribution, and consumption sectors of the power system. Hence, handling this non-linear behaviour of power system assets such as electrical machines and transformers, as well as entire networks, can be mathematically challenging for AI-powered DTs. Specifically, solving non-linear optimisation problems using meta-heuristic algorithms such as PSO or genetic algorithms can be a critical challenge. However, if non-linear system modelling methods like Lyapunov methods are embedded into power system DTs, this issue can be alleviated. In addition, multi-objective optimisation requirements can also pose significant mathematical challenges for AI-augmented DTs. Specifically, DSM use cases in sustainable power grids require the satisfaction of multiple, conflicting objectives such as minimising operational costs of consumers while maximising their comfort and utility. Also, objectives such as maximising social welfare, maximising combined profit of the consumer and the electricity supplier, and minimising GHG emissions can be included in these optimisation problems. Therefore, mathematical methodologies with the ability to compute accurate as well as efficient solutions for multi-objective optimisation problems must be incorporated into power system DTs.

Numerous mathematical challenges are associated with free and commercial software libraries used for simulation and modelling, optimisation, controlling, and ML functions of AI-powered DTs in power systems. Specifically, commercial and proprietary software packages consisting of pre-built modules and algorithms lack flexibility, hindering the ability to cater to the unique needs of complex and evolving power systems. Also, these proprietary software tools have limited interoperability with open-source tools and platforms offering custom AI/ML models that are capable of satisfying the evolving requirements of power system DTs. Furthermore, most of these commercial software libraries offer black-box algorithms that are not accessible or transparent to the users. As discussed previously, these closed models can be problematic in critical power system applications, where the rationale behind decisions made by AI/ML models must be understood and justified. Another key challenge associated with proprietary software is the use of outdated algorithms that may not be optimised for the latest advancements in AI. While these algorithms might perform well for traditional tasks such as power flow analysis or state estimation, they may lack the adaptability required for more advanced AI-driven tasks such as predictive maintenance and demand forecasting. Despite the rapid progress of the AI/ML domain, commercial libraries are comparatively underperforming in utilising novel techniques like DRL and GNNs. Moreover, these software tools are not well-equipped to handle the high-dimensionality, complexity, and stochastic nature of power systems, especially under the volatile and probabilistic operating conditions discussed above.

In contrast, free software libraries and repositories also have certain limitations. Specifically, open-source code versioning and hosting platforms such as GitHub, Bitbucket, Launchpad, and SourceForge [178] can lead to issues related to the quality and reliability of code. This can be attributed to the large number of diverse contributors engaged with such platforms. While the diversity and broad engagement result in faster adaptation of advanced AI/ML techniques and enhanced flexibility, quality-related issues can detrimentally affect the stringent requirements of reliable power systems. In addition to that, free software libraries, particularly open-source AI tools available on the previously-mentioned platforms, may lead to scalability issues under limited memory and computational efficiency. This presents a significant challenge for solving large optimisation problems such as economic dispatch, unit commitment, or load balancing, where huge amounts of data must be processed in real-time. Also, the AI/ML algorithms available in such free software libraries are not well-suited to handle non-linear and complex dynamics stemming from interactions between components such as generators, loads, storage devices, and renewable energy sources. Most of the open-source AI libraries are developed for general applications and may not provide specialised algorithms capable of accurately capturing these non-linear dynamics, resulting in poor accuracy and precision. For instance, simple approximation and linearisation techniques utilised by the basic AI/ML models offered by these libraries may not perform well in accurately simulating transient stability events, voltage fluctuations, and frequency control dynamics. Also, the numerical stability and convergence of the algorithms available under free software tools can be sub-optimal. Hence, such tools will not perform well for use cases such as power flow analysis, state estimation, or dynamic stability simulations. To overcome these challenges associated with free as well as commercial software libraries, developers of power system DTs can adhere to several strategies. Combining free libraries with specialised or custom-developed commercial tools tailored to the unique requirements of power systems can be a solution. This hybrid approach can assist in harnessing the advantages of both types of software tools while minimising their mathematical limitations.

5.2.5. Challenges Related to Scalability and Adaptability

AI-augmented power system DTs must remain scalable and adaptable by addressing challenges related to the fast-paced evolution in technology, regulatory environments, and market dynamics of the current world. With respect to the technological dimension of this evolution, multiple challenges can impact the long-term viability of AI and DT technologies. First, the evolving nature of AI techniques themselves would pose a challenge. As evident by the evolutionary timeline analysed in Section 4.2.4, the world of AI is moving towards advanced techniques including GNNs, NLP, computer vision, smart robotics, and also multi-modal techniques. Furthermore, novel computational technologies such as quantum computing are on the rise. It would be complicated to ensure that DTs can adapt to and incorporate these innovative, newer techniques without requiring complete redesigns or disruptions to daily operations. Therefore, the power industry must avoid designing overly specialised, rigid AI-powered DTs, so that continuous updates to technological methods as well as load and supply conditions can be incorporated. Also, adopting a modular architecture within power system DTs can assist in overcoming the challenges related to technological scalability and adaptability. This way, each module can be designated with distinct tasks such as data ingestion, optimisation, and prediction, allowing them to evolve and scale independently without affecting the entire system.

Furthermore, data fusion for DTs is becoming more challenging under the explosion of big data, where a large number of IoT devices, sensors, and software systems must be connected to a single DT, resulting in an exponential rise of power system data that must be integrated, stored, processed, and analysed. Moreover, when the power system dynamics become more complex, especially under the rising trends like smart grids, microgrids, DG, RES systems, prosumers, and virtual power plants, the complexity of DT and AI modelling also gets scaled up. In this situation, the requirement for complex modelling and simulation techniques that can capture the behaviour of distributed and interconnected subsystems becomes more critical. Multi-agent and hierarchical DTs can be effective in addressing this requirement when blended with cloud and edge computing technologies appropriately. For instance, when the power system DT must be scaled up to include multiple layers or tiers (i.e., national grid, distribution system aggregators, local transformers, households, and individual appliances in households), cloud environments can be used for large-scale processing related to the central controller, whereas edge computing can be used for local processing associated with individual households [179]. This hybrid approach enables real-time processing at the edge, while centralised control and deeper analysis can be performed in the cloud. Also, as discussed previously under technical challenges related to the DT dimension, power system DTs must harness standardisation and interoperability between different hardware and software systems, as well as between modern and legacy systems in the power grid to ensure long-term adaptability. Moreover, under the inherently volatile and uncertain conditions prevalent in power systems, AI models embedded in DTs must be able to adapt and evolve under previously known and unknown conditions, without having to be retrained from scratch. RL techniques such as Q learning can be beneficial in this regard [123,179].

Not only the technological dimension but also the regulatory environments affecting power system DTs are rapidly changing. Within the current context, these regulatory requirements have been shifted from addressing preliminary issues such as market regulation, data protection and industrial relationships, towards more pressing concerns such as sustainability, grid resilience, and cyber security. Therefore, when planning, designing, and developing power system DTs, complying with the evolving regulatory requirements can be a demanding task. This can be overcome by continuously monitoring compliance with evolving regulations and flagging areas where adjustments are required. In fact, AI models can be used to automate this process. Also, data security and privacy must be considered from the initial planning and design stages of these AI-powered DTs, employing novel technologies such as blockchain for secure transactions and federated learning for localised data processing. Simultaneous to the regulatory changes, market conditions of the power industry tend to update frequently. For instance, new trends, such as dynamic pricing models for DR, demand bidding in wholesale markets, peer-to-peer energy trading, and blockchain transactions, are emerging [67,180]. Under these dynamic market conditions, the role of power system DTs will be challenging. However, AI can be a saviour for such instances, where adaptive algorithms can be used to compute optimal market strategies and trading prices under variable economic and environmental conditions, as well as demand fluctuations. RL models and dynamic optimisation methods are some solutions offering this adaptive capability. Most importantly, the self-monitoring, self-regulating, self-diagnosing, and self-adapting capabilities proposed by Mihai et al. [7] can be transformative for power system DTs in surviving and thriving under dynamic technological, regulatory, and market conditions.

5.2.6. Ethical and Regulatory Challenges

Though not comprehensively explored in this study, challenges related to the regulatory and policy environment and political factors must also taken into consideration when developing power system DTs. Despite the fast-paced advancements occurring in the fields of DT and AI, the political and regulatory dynamics have not been able to keep up with that. Consequently, various challenges would hinder the development and deployment of AI-powered DTs in power systems, owing to the insufficient attention and support provided by regulatory and political stakeholders.

Moreover, ethical and technical concerns encompassing data governance, data security, data privacy, and trustworthiness of data are also challenging areas of concern for AI-powered DTs in power systems. This can be attributed to the fact that power systems generate vast amounts of data that are mission-critical for the operations of entire nations and which must be governed and secured with robust measures. Therefore, utmost care and responsibility must be applied when AI models are deployed through power system DTs, which can be vulnerable to these challenges. Particularly in smart grids, microgrids, and DSM systems, large amounts of consumer data are collected. These data can contain energy consumption measurements, demand patterns, customer preferences, and also personally identifiable information like names, driver’s license numbers, credit card numbers, and phone numbers. The integration of AI and DTs in these power grid sub-domains introduces additional layers of complexity due to the real-time and continuous flow of data between physical and virtual counterparts. Therefore, power system DTs are highly susceptible to data privacy violations, which must be prevented and minimised through the implementation of personal data protection strategies such as anonymisation and pseudonymisation. With respect to the ethical dimension of data privacy, informed consent must be obtained from consumers regarding the data that are being collected and how it will be used. Also, complying with legislation imposed by the government on personal data protection and clarifying the ownership of data can also be beneficial to handle data privacy-related technical and ethical concerns in this field.

Moreover, as discussed above with respect to NLP, computer vision, and smart robotics, data security must be ensured when other AI techniques are deployed through power system DTs. Attention must be concentrated towards implementing state-of-the-art cyber security measures and regularly assessing system vulnerabilities for data breaches and malicious attacks. Authentication and access control can be another solution for mitigating data security issues in power system DTs. Regarding ethics, operators of AI-powered DTs must ensure that there is traceability and accountability in the data security-related decision-making processes.

The reliability and trustworthiness of AI and DT models implemented in power systems is another challenging ethical consideration, due to the mission-critical nature of power systems. Therefore, as discussed above under the technical challenges of AI, the outputs generated by the AI models must be explainable and transparent, thus enhancing the reliability of the data-driven decisions made with the assistance of the AI-powered DT. Also, ensuring the accuracy of the AI-powered DT models by maintaining desirable data quality and model adaptability can further support overcoming reliability-related challenges in this field. Additionally, human supervision and intervention must be facilitated for these automated models implemented in power systems. Despite the transformative contribution of AI and DT technologies in assisting the forecasting, pattern identification, decision-making and analysis tasks in power systems, it is prudent not to overly rely on the merits of these automated technologies under mission-critical scenarios.

Ironically, although the focus of this study is to review the application of AI-powered DTs in power systems and how they may contribute to environmental sustainability, these DTs might create their own environmental footprints. This is also a major ethical concern affecting the widespread application of AI and DT in power systems. Demanding electrical energy consumption and cooling requirements is one of the primary causes for this challenging situation. When the power system DTs and AI models discussed in the reviewed research are commercially implemented, digital infrastructure requirements such as data centres with high computational power, seamless bidirectional communication, large storage spaces, and strong cyber security measures will also be increased. Consequently, energy consumption for this digital infrastructure will be increased, which might counteract the originally intended sustainability goals of AI-powered DTs in power systems. This challenging situation can be managed by employing energy-efficient AI-powered DT systems. In contrast, there can be long-term societal impacts originating from loss of employment opportunities and job displacements due to the highly automated systems. Nevertheless, training the existing human workforce in supervision tasks and tasks that mandatorily require human expertise and skills, can be a solution to this concern.

Hence, it can be summarised that, while the primary cause of challenges in this field can be attributed to technical integration and data-related factors, ethical, social, and political aspects would also affect the growth of AI-powered DTs in power systems.

5.3. Role of Digital Twins and Artificial Intelligence in Overcoming Technical Challenges of Power Grids

According to IEA, power grids in emerging markets and developing economies are facing challenges with respect to reliability and resilience, increased losses, affordability and access, rising demand and changing demand profiles, and new loads and storage such as EVs [5]. Under this setting, utilisation of concepts such as smart grids, microgrids, DG, and RES have risen in popularity due to their ability to prevail over these challenges, as discussed in Section 1 of this work. Nevertheless, these novel power grid technologies have introduced another set of technical challenges themselves, which adversely impact the stability and reliability of the power grid. Some instances of such major challenges include maintaining frequency and voltage stability under volatile generation conditions; power quality issues due to the introduction of higher-order harmonics and reduced system inertia; and demanding requirements for robust communication infrastructure, big data management, predictive capabilities and advanced control strategies. These challenges primarily stem from the uncertainty, complexity, and volatility of novel power grid technologies. The synergetic application of DT and AI can serve as a promising solution for such technical challenges, on top of promoting environmental sustainability.

Despite the immense contribution towards reducing GHG emissions in power generation, the inherent variability of RES such as wind and solar PV can result in frequent fluctuations of grid frequency and voltage. The real-time monitoring and controlling capability of DTs, combined with AI-based predictive algorithms, can assist in proactively regulating these fluctuations. For instance, in [70], a ‘Digital Twin’ with bidirectional communication and control capabilities is proposed for reactive power control and voltage regulation of solar PV. This DT can estimate real-time states under the lack of in-field measurements, while a PSO-based algorithm is tasked with computing the optimal reactive power set point. Moreover, Kharlamova et al. [82] presented a DT of a BESS for frequency regulation in a power grid dominated by renewable energy. This DT is augmented with AI algorithms such as LSTM and GRU to predict the battery state-of-charge required for the frequency regulation process. In summary, the DT proposed in [82] is aimed at minimising the disruptions caused by equipment faults and cyber attacks, while ensuring the reliability and predictability of BESS providing frequency control services. Moreover, the predictive capability of AI in intelligent DTs assists frequency and voltage regulation by accurately forecasting dynamic variations of generation output [72]. Most importantly, intelligent DTs with advanced simulation capabilities serve as a platform for virtually testing and implementing advanced voltage and frequency control strategies, without risk to actual physical infrastructure.

In addition to voltage and frequency fluctuations, the integration of inverter-based renewable energy systems introduces higher-order current and voltage harmonics, detrimentally affecting power quality standards. Owing to the continuous, real-time monitoring capabilities of DTs, these harmonics can be detected and analysed in real-time, with the assistance of predictive AI algorithms. For example, an online voltage harmonic monitoring scheme is proposed in [181], where a probabilistic neural network is employed in a DT environment. Intelligent, AI-powered DTs can also provide a comprehensive digital assessment of the impact of harmonics on power quality. Moreover, the optimisation capability of these intelligent DTs can be used for designing adaptive filtering techniques and computing system adjustments to mitigate harmonic distortions. Then, the simulation capability of DTs can be utilised to analyse the long-term effects of these corrective actions, enabling grid operators to identify optimal solutions for improving power quality under high penetration of RES.

Another major obstacle associated with DG and RES is the reduced system inertia. This challenge is inevitable under new power generation sources such as solar PV, where the mechanical inertia of large rotating masses present in conventional sources is replaced by power electronic-based renewable energy systems. The distributed and isolated nature of the power supply under DG further deteriorates this issue. This situation leads to the impairment of the entire power system inertia, resulting in more frequent frequency deviations under poor stability. Specifically, intelligent DTs can help in managing poor inertia conditions with the use of AI by deploying fast-responding algorithms that predict and correct frequency deviations. Additionally, DTs can serve as a simulation test bed, which enables the design and testing of virtual inertia strategies. These control strategies can dynamically adjust the system response based on real-time grid conditions, mitigating the risks posed by poor inertia.

With the advent of smart grids, microgrids, DG, and RES, the demand for robust communication infrastructure, big data management, predictive capabilities, and advanced control strategies has been more compelling than ever before. To be more concise, the transition towards more complex, distributed, and variable energy sources under these power grid concepts inevitably requires a secure, robust, seamless communication infrastructure. This can be facilitated by DTs with advanced communication capabilities. Specifically, under the distributed nature of connected agents or physical entities, the federated communication and control architecture of DTs can be effective. For instance, a multi-layer DT framework is designed in [179] to replicate actual household energy consumption through a household digital twin (HDT) connected to an energy production digital twin (EDT). This AI-powered DT implementation comprises an EDT serving as the central controller, a local transformer DT in the middle layer, and the HDTs and appliance DTs in the lower layer. Moreover, it is equipped with bidirectional communication between the distinct layers. For instance, household-specific energy consumption data are communicated to the central EDT through transformer DTs, whereas electricity tariff data are communicated in the opposite direction. This interconnected and federated DT architecture also ensures data privacy of households by communicating hourly aggregated consumption to the EDT without disclosing house-specific data. Simultaneously, optimised energy management actions are computed at HDTs and communicated to the physical controllers to actuate the operation of appliances. Also, these novel grid technologies employ various sensors generating a vast amount of data in real-time. Hence, DTs with IoT and big data management capabilities can assist in handling data streams with high volume, velocity, and variety. Moreover, with the help of IoT, control and automation engineering, optimisation techniques, and smart robotics, intelligent DTs can cater to the advanced control requirements related to the actuators connected to them. Moreover, the requirement for predictive and prescriptive analytics in smart grids can be addressed through AI techniques such as DL and supervised learning, as evidenced by studies such as [96,100,102,103].

5.4. Opportunities and Future Research Directions

Building upon the identified research gaps and challenges, several promising future research directions to advance and broaden the application of AI-powered DTs in power systems could be summarised as follows:

• Future research could advance beyond the current focus on operational and maintenance functions to explore the potential of AI-powered DTs in the planning, design, and economic aspects of power systems. Integrating AI with DTs during the early stages of the power system life cycle, such as planning and design, could offer significant benefits. Drawing inspiration from other fields such as smart city design, smart building design, and product design where AI and DTs are used for planning and design, power system research could benefit from similar methodologies to achieve long-term sustainability. Moreover, use case categories such as DR, energy storage, frequency regulation, power quality assurance, transient stability analysis, power flow analysis, and training and education could unveil interesting opportunities for future research in AI-augmented power system DTs.
• Given the under-representation of research focused on individual power system assets, future studies could develop AI-powered DTs for a wider range of critical power components. Extending the application of intelligent DTs to under-explored assets like power electronic converters, inverters, protective switchgear, measurement devices, line conductors, generators, and motors could improve the operational efficiency and reliability of these components.
• As discussed in Section 5.3, synergetic application of DT and AI can serve as a solution for numerous technical challenges associated with new power grid concepts such as smart grids, microgrids, DG, and increased penetration of RES. Also, the studies reviewed in this work had not sufficiently addressed these challenges through AI-powered DTs, though many studies present situations in which AI or DT technology is used separately [182,183]. Hence, future research could be stemmed from the combined utilisation of DT and AI to mitigate and manage power quality issues, frequency and voltage fluctuations, and reduced system inertia.
• Under-utilised AI domains beyond ML and optimisation could be investigated for power system DTs. Even in the ML domain, clustering, association rules, dimensionality reduction techniques, and Monte Carlo methods were under-explored. In addition, it would be transformative to investigate the application of generative models and GNNs, which could offer enhanced predictive accuracy and better handling of complex spatio-temporal data.
• Furthermore, computer vision, NLP, and smart robotics could enhance the advanced sensory and action capabilities required for power system DTs. Especially with the recent boost in interest towards LLMs and their multi-modal capabilities, use cases in power systems could benefit from such newer, advanced AI techniques. With respect to NLP, applications in speech recognition and synthesis, machine translation, text summarisation, text classification, information extraction and retrieval, and topic modelling can be implemented in power system DTs through future research. These applications shall be effective for the management of maintenance and operational records, predictive maintenance, root cause analysis, and fault analysis. Also, chatbots and AI-powered search engines can be developed with the use of advanced NLP techniques like Retrieval Augmented Generation (RAG), in order to serve as an interactive knowledge base assisting power system operators in planning and operations as well as for training and educational purposes. These interactive knowledge platforms and conversational agents can be extended to the operations of electricity markets and demand side operations, providing assistance to consumers and prosumers of the power grid. Adnan et al. [184] proposed a customer service chatbot for government sector organisations, which can be extended to DTs implemented in electricity utilities as well. Also, NLP-based sentiment analysis techniques can be utilised for DTs tasked with electricity market assessments. Moreover, the computer vision domain also provides promising future research paths for power system DTs. Automated drone-based inspection systems can be applied to detect physical damages to inaccessible or difficult-to-access power system assets such as transmission lines, switching equipment, transformers, wind turbines, and solar panels. A methodology comprising a DL model and a drone-based transmission line inspection system can be found in [185], which can be expanded to include a DT with advanced simulation and visualisation capabilities. Also, these applications can be extended to real-time anomaly detection and hazard detection DTs, especially in high-voltage environments, through image/video classification and segmentation techniques [186]. Computer vision techniques can be applied for power system DTs employed in planning and design functions as well. For instance, satellite or aerial images can be analysed to understand how environmental factors like vegetation growth, water sources, weather patterns, as well as urban constructions can impact power system infrastructure. This kind of analysis can be highly beneficial for the planning, design, and construction of power generation, transmission, and distribution infrastructure. Furthermore, smart robotics would ensure that unique human intelligence and other skills could be used for applications that mandatorily require that level of sophisticated precision, while improving the health and safety of power system personnel. For example, future research can originate from AI-powered robots utilised for maintenance and inspection tasks in power plants and substations. This kind of use case can be especially value-adding for operations in extreme environments such as offshore wind farms and nuclear power plants, where adaptive robots could perform inspections and repairs under difficult weather and physical conditions.
• Investigating the self-adapting, self-diagnosing, and self-healing capabilities of power system DTs could be another major future research direction, which can be realised with the recent advancements in the field of AI. Research should focus on creating DTs capable of autonomously managing faults, optimising system performance in real-time, and executing corrective actions with minimal human intervention. This transforms DTs from passive monitoring tools into proactive agents capable of maintaining system stability and reliability under dynamic, uncertain conditions.
• While the integration of renewable energy sources was a principal focus in the reviewed literature, future research could explore more direct applications of AI-powered DTs for environmental sustainability. This could include monitoring, predicting, and reducing greenhouse gas emissions in thermal power generation and other sectors of power systems, managing the use of low-carbon equipment and technologies, and improving the environmental footprint of the entire power system. AI-powered DTs could also play a crucial role in managing the life cycle impacts of power system components, from production to disposal, thereby contributing to broader sustainability goals.
• Addressing the challenges of integrating AI-powered DTs across different sectors of the power system will require interdisciplinary collaboration and the development of unified frameworks. Future research should focus on creating interoperable standards that facilitate the seamless integration of DTs with existing power system infrastructures, including legacy systems.
• To overcome the challenges associated with the deployment of advanced AI techniques in critical power system applications, future research should prioritise the development of interpretable and transparent AI models. This involves creating methods that provide clear explanations for AI-driven decisions, particularly in high-stakes scenarios such as grid control and outage management.
• The development of AI-powered DTs for power systems will require supportive policy and regulatory frameworks. Future research should explore the implications of emerging AI and DT technologies on regulatory practices and identify pathways for aligning technological advancements with existing regulatory standards within power systems. Collaboration between researchers, policymakers, and industry stakeholders will be essential to create an environment that fosters innovation while ensuring the security, stability, and reliability of power systems.

By pursuing these future directions, both academia and industry can address the current limitations and challenges, while leveraging the full potential of the synergetic application of AI and DTs in sustainable power systems.

6. Conclusions

This systematic scoping review explored the synergetic application of DT and AI technologies within power systems, presenting both qualitative and quantitative results with respect to the breadth and variety of power system use cases and AI techniques used. Also, the relationships among the three main dimensions covered under the review, power systems, DT, and AI, were critically evaluated through the novel classification framework proposed for power system concepts (i.e., elemental, topological, sectoral, and functional aspects) as well as AI domains, AI main categories, AI sub-categories, and DT versions (i.e., ‘SCADA’, ‘Digital Shadow’, ‘Digital Twin’). Moreover, the evolutionary journey of DT and AI concepts in power systems was investigated independently as well as jointly, by presenting a progression timeline including three distinct periods of maturity in terms of the level of intelligence and the diversity of use cases addressed. It was identified that the ‘Digital Twin’ version of DTs with advanced intelligence, simulation, and bidirectional communication capabilities has been dominating the third era spanning from 2019 to 2024, complemented by relatively new AI techniques such as RNN, CNN, NLP, computer vision, and smart robotics. Along this timeline, the focus of use cases has shifted from conventional topics like system reliability and stability towards areas such as DR, energy management, energy storage, sustainability, and digitalisation. Regarding sustainability implications, three main objectives covered by the reviewed literature were identified and analysed. The majority of reviewed studies focusing on environmental sustainability were limited to increasing the share of renewable energy generation and improving demand-side energy efficiency.

Finally, the research gaps, challenges, and future research directions were synthesised, which exhibit that the full potential of the synergetic application of DTs and AI in power systems is yet to be realised. Under-utilisation of AI techniques, such as generative models, GNNs, NLP, computer vision, and smart robotics, insufficient focus on power system planning, design, and economic functions, and under-representation of individual power system assets were identified as main research gaps. Moreover, it could be inferred that the reviewed research had not sufficiently explored the synergetic power of DT and AI, with respect to the self-adapting, self-regulating, self-monitoring, and self-diagnosing capabilities. Technological challenges related to system integration and interoperability, real-time modelling of complex dynamics, big data management, and model interpretability were outlined. Also, specific technological challenges such as data integration, high computational and communication requirements, and model accuracy, which are affecting NLP, computer vision, and smart robotics applications, were discussed. The high dimensionality and stochastic nature of systems, non-linear variables, and multi-objective optimisation requirements were outlined under mathematical challenges. Additionally, challenges related to the scalability and adaptability of AI-augmented power system DTs were explored, along with ethical and regulatory challenges spanning data privacy, security, and trustworthiness. Furthermore, the role of DT and AI in overcoming technical challenges of smart grids, microgrids, DG, and RES were discussed, specifically focusing on frequency and voltage regulation, power quality management, and improving poor system inertia. Based on the results of the review and identified research gaps and challenges, promising future research directions could also be identified. Specifically, broadening the utilisation of novel AI techniques such as GNNs, NLP, computer vision, and smart robotics was discussed, along with investigating the self-adapting, self-diagnosing, and self-healing capabilities of intelligent power system DTs. Moreover, it could be inferred that DTs and AI could contribute directly as well as indirectly towards environmental sustainability, during an era where the power and energy industry stands as a major culprit for carbon emissions and resulting climate change effects. DTs and AI indirectly fuelled a carbon-neutral power system by improving energy efficiency and supporting renewable energy integration, while alleviating the technical challenges associated with novel digitalisation and renewable energy-related trends. This review pointed out that the synergetic application of DT and AI could contribute to sustainable power systems more directly, by addressing the reduction in carbon emissions and waste management under a circular economy setting. Should these challenges and gaps be managed prudently, the synergetic application of DT and AI technologies could be the next transformative trend in the domain of electrical power systems.