1. Introduction
The global energy landscape is undergoing a significant transformation, driven by the urgent need to mitigate climate change and facilitate the transition to sustainable energy systems. Renewable energy sources, such as solar photovoltaics and wind turbines, are increasingly being integrated into smart grids to reduce carbon emissions and dependence on fossil fuels. However, the intermittent nature of these resources poses challenges for grid reliability, stability, and cybersecurity, exposing to vulnerabilities such as False Data Injection Attacks (FDIA) and Denial of Service (DoS) attacks.
Improving the functionality and adaptability of smart grids is increasingly reliant on artificial intelligence (AI) and machine learning (ML). The paper briefly describes advanced models such as the Long Short-Term Memory (LSTM) network, the Random Forest (RF) model, and the autoencoder, among others, which have been applied in demand prediction, pattern detection, and anomaly detection in energy systems. Specifically, LSTM and Gated Recurrent Units (GRUs) do not suffer from the vanishing gradient problem, which supports their application in smart grid for tasks such as load-profile analysis and energy-routing optimization. Moreover, the Self-Organizing Maps (SOMs) can be employed to identify anomalies in consumer-load patterns and irregular sensor records, which may indicate equipment malfunctions, including defects or FDIA-related cyberattacks. These AI-based technologies enhance grid intelligence and adaptability, enabling real-time decision-making and effective use of resources to meet the demands of dynamic energy requirements.
In addition, the increasing digitalization of smart grids presents cybersecurity challenges, whereby smart meters (SMs), Phasor Measurement Units (PMUs), and Remote Terminal Units (RTUs) are vulnerable to cyber intrusions targeting their control. In this regard, this paper identifies and discusses the application of advanced solutions, such as the Cluster Partition–Fuzzy Broad Learning System (CP-FBLS) and multimodal deep learning, for the immediate and swift response to FDIA and DoS attacks in power-generation systems. The methods exploit neural network topologies as well as decentralized control and ensure effective resource management at high performance levels. The threat of the Advanced Persistent Threats (APTs), which can cripple grid functionality and cause cascading failures, is also addressed in this review study. This paper further notes that secure and locally resilient architectures are critical in safeguarding critical energy infrastructures against emerging cyberattacks, through the exploration of not only the frameworks but also the overall effects of integrating grid-level architectures into motion-control strategies that mitigate the degradation of resilient consensus.
Renewable energy intermittency and the resulting requirements to manage it pose significant demands for energy-storage systems (ESSs) and comprehensive policy frameworks necessary to support the sustainability of power systems. The paper also discusses hybrid energy-storage systems (HESSs), which consist of a combination of mechanical, electrochemical, and thermal storage technologies to improve grid stability and efficiency. The co-optimization models reviewed have reduced transmission expansion requirements by nearly 10%, thus cutting infrastructure costs and enhancing grid reliability. Moreover, the research highlights the role of policy instruments in overcoming the challenges of political opposition, financial limitations, and social acceptance. Combining econometric modeling, policy studies, and AI-driven optimization methods, the work offers a comprehensive framework for developing scalable, secure, and sustainable energy chains, capable of meeting the increasing global energy demand and ensuring environmental protection and system stability.
The smart-grid technologies, such as energy management systems (EMSs) and Advanced Metering Infrastructure (AMI), further enhance the efficiency of energy networks. EMS provides real-time visibility and control over grid operations as well as optimization, and AMI allows demand-side management based on consumption profiles derived in real time from consumption data. The paper also discusses energy routing, incorporating AI-based solutions and multi-agent systems (MASs) as approaches enabling decentralized control and peer-to-peer scheduling to enhance resilience. Nonetheless, these technologies face challenges such as heavy computational burdens and scalability limitations, underscoring the importance of future research into hybrid protocols that integrate metaheuristics, AI, and blockchain to create secure and flexible energy networks. By examining these developments in detail, this study contributes to the ongoing discourse on power system transformation, aiming to achieve sustainability, security, and efficiency in response to evolving energy needs.
This review aims to provide a holistic security and efficiency analysis for modern smart grids based on an extensive literature review, as presented in
Figure 1. While existing surveys often treat cybersecurity and renewable integration in isolation, this work synthesizes these domains to highlight their interplay. Furthermore, it addresses a significant gap in the literature by dedicating substantial focus to energy-routing protocols, a strategic decision-making layer essential for managing power flow in decentralized, renewable-heavy grids. By evaluating AI, metaheuristic, and multi-agent systems for routing optimization, this review offers a distinct perspective on achieving grid resilience, going beyond conventional discussions to address the core operational challenge of future energy networks.
3. Transmission and Distribution (T&D) System of Power Grid
In a traditional power grid, power flows in one direction from the utility through transmission and distribution systems to consumers. This one-way flow is a key component of traditional grid architecture, where centralized monitoring is essential to balance generation with consumer demand while respecting system limitations [
64]. Utility operators manage every stage of the process to ensure both reliability and financial sustainability. However, the rapid growth in electricity demand from new devices, notably electric vehicles (EVs), has introduced vulnerabilities in traditional grids, including losses, limited monitoring and communication, poor distribution routes, and a lack of advanced automation and digital technologies [
65,
66].
To address these limitations and make the grid system more reliable, the smart grid has been introduced. By integrating technologies like the Internet of Things (IoT), modern digital technology, and automation, the smart grid facilitates two-way communication and allows electricity to flow in both directions [
67]. These advancements enhance the monitoring of electricity transmission and distribution, support the seamless integration of electricity sources, and enable real-time monitoring and immediate response to disruptions, making the grid system more reliable and efficient. Smart-grid technology is better equipped to handle energy demands and operational issues in real time [
68].
However, smart grids have complex architecture, which makes it more challenging to maintain effective power quality when dealing with a variety of load types. It is more difficult to maintain voltage stability, control harmonic distortion, and prevent other quality concerns when the network faces issues caused by diverse devices and generating sources [
69].
Advanced, AI-powered grid interfaces are becoming essential tools for real-time control and adaptability [
70,
71]. AI-enhanced power quality-monitoring devices predict and mitigate future issues while also addressing current abnormalities. They achieve this by monitoring system conditions [
72], forecasting disruptions, and automatically modifying operating conditions. This proactive approach support future innovations like direct-current distribution networks [
73], which might further reduce power delivery issues in data centers, homes, and renewable energy agriculture. This enhances the functionality of connected equipment while encouraging safer, more efficient grid operation.
While integrating these IoT and modern digital techniques introduces challenges and safety risks like cyber threats, power outages, and blackouts [
74], an overview of the smart grid is presented in
Figure 2. To address these problems, much research has been conducted in recent years.
Utkarsh et al. [
75] introduced a Self-Organizing Map (SOM) method for resilience assessment and restoration control in a distribution system during severe events. The study proposes a temporal resilience index to support resource allocation and operational dispatch. The proposed SomRes framework detects weak states and promotes pre-emptive load shedding, thereby improving coordination of T&D system responses during contingencies.
A fuzzy SOM-based third-party approach for electricity distribution networks was presented by Ovidiu Ivanov. Under an affinity control policy, consumer classification through load-profile shapes was performed for the rapid development of a Typical Load Profiles (TLPs) database, which is essential for effective T&D planning. The approach demonstrated good strength and stability of segmentation under load.
Wu et al. [
76] proposed Gridtopo-GAN, a GAN-based method which uses topology-preserved node embeddings to correctly identify meshed and radial patterns in large-scale networks and performs effectively on both IEEE benchmark systems and real-world networks.
Yan et al. [
77] proposed UG-GAN, a model that synthesizes three-phase unbalanced distribution systems. By observing real network random walks, it emulates local structures and enriches the synthetic grids with realistic nodal demands and components. UG-GAN was tested on smart meter data to balance data privacy and structural realism.
Venkatraman et al. [
78] developed a CoTDS model with static/induction loads, reactive shunts, and DG inverters. The PI control and small-signal model serve as the foundation for analyzing mutual influences between the T&D systems. Simulations highlight the necessity of coordinated grid control under high DG penetration.
Mohammadabadi et al. [
79] introduced a GenAI-based communication model that employs pre-trained CNNs to generate synthetic time-series data. Local agents process the data, reducing transmission load and preserving privacy. This method supports scalable, distributed learning for tasks such as anomaly detection and energy optimization in smart grids.
Shankar et al. [
80] developed a deep learning OPA (dLOPA) model based on LSTM for smart-grid contingency analysis. It incorporates a pre-processing step that improves training efficiency, achieving 93% accuracy while maintaining low false-positive and false-negative rates. The dLOPA model does not target T&D systems specifically but contributes to the reliability of the grid.
Gonzalez et al. [
81] introduced an RNN approach for demand prediction in distribution network planning. By incorporating confidence intervals, the model successfully respects explicit transformer thermal boundaries and enhances planning when accounting for load variations and distributed generation.
Shrestha et al. [
58] proposed a federated learning LSTM-AE model for anomaly detection. The model achieved a 99% F1-score, maintained privacy, and outperformed conventional techniques.
Liu et al. [
82] introduced MDRAE, an attention-based autoencoder for missing data recovery in transmission and distribution systems, that outperformed comparison models and enhanced both data quality and grid stability.
Raghuvamsi et al. [
83] proposed a TCDAE for topology detection in distribution systems with missing measurements that outperformed other types of DAE variants.
Li et al. [
84] proposed GTCA, a mixed T&D contingency analysis application that integrates the Global Power Flow and DC models. The approach is more effective than traditional methods, especially for looped networks.
3.1. Important Factors of Transmission & Distribution (T&D) System
There are several important factors of the T&D system in the smart grid, which are presented in
Figure 3.
3.2. Advanced Metering Infrastructure (AMI)
AMI is an advanced digital system that automatically measures energy consumption in real time [
85]. In contrast, traditional metering relies on analog devices requiring manual readings. AMI enables two-way communication between consumers and prosumers, providing accurate consumption monitoring and billing based on exact energy usage. This technology also facilitates demand response programs and allows consumers to track their real-time energy usage, helping them to reduce consumption during peak hours and optimize costs [
86].
3.2.1. Communication Networks
Communication networks enable frequent status checks, the sharing of data, and two-way communication to enable accurate T&D data which, in consideration of smart-grid networks, requires a high level of security. With constant monitoring, prosumers can track grid status and identify potential problems that might occur in order to enhance distribution of energy properly. It is also remotely managed through automated control systems that operates electronic devices such as routers and electric vehicles. This enhances the efficiency of transfer data and communication channels between prosumers and consumers more efficiently [
87].
3.2.2. Distributed Energy Resources (DER)
Distributed energy resources (DER) include facilities located closer to consumption points than conventional generation facilities. These systems comprise small energy sources like solar arrays, batteries, wind generators, and electric cars that produce and store energy close to the households and enterprise premises. Implementing DER on the smart grid increases reliability and flexibility while reducing dependences on conventional power plants. Moreover, DER expands the utilization of green energy and reduces carbon emission. Moreover, self-sufficient energy systems, known as microgrids, can operate independently and control multiple processes with the smart grids [
88,
89].
3.3. Energy Management Systems (EMS)
An energy management system (EMS) is a computer-based system designed to monitor, manage, and optimize the operations of the smart power grid. It also assists utilities in generating, distributing, and managing electricity consumption in real time [
90]. The overall goal of EMS is to enhance grid stability through effective management of the power resources, energy supplies, and loss as well as enhance efficiency in general. Information collected by EMS in actual time varies in different sectors of the smart-grid tools in order to carry out functions such as forecasting about the energy consumption and regulation of the amount of needed power. Machine learning methods are also applied in EMS to increase system performance, enhance prediction accuracy, and enable autonomous systems to work [
91,
92].
Grid Security
Grid security aims to protect the smart-grid infrastructure from cyberattacks, which can disrupt operations and cause major losses, such as data theft. To prevent such disruptions, smart grids employ security systems that ensure secure and reliable operation [
93]. Cyberattacks primarily target smart grid-connected systems containing critical information. To prevent these attacks, grid security employs powerful encryption, secure authentication methods, and real-time monitoring for immediate threat identification and prevention. In addition, grids implement physical safeguards, such as fences, cameras, and guards, to prevent unauthorized physical access that could lead to equipment damage or cyber intrusions [
94,
95].
3.4. Detecting Methods
To ensure grid security and sustainability, several detection methods have been proposed in the literature, as summarized in
Table 8.
3.4.1. Optimization Models
Optimization models are mathematical and computational tools used to enhance the performance and efficiency of electricity transmission and distribution. These models allow smart-grid systems to make intelligent decisions and determine the optimal operational strategy in various scenarios. They also help reduce energy loss and improve real-time efficiency. Recent investigations demonstrate that these approaches are effective in both simulations and real-world situations [
115].
Barajas et al. [
116] introduced optimization models to enhance the generation, transmission, and distribution of electricity in a microscopic system using the energy hub concept. The first linear programming model addressed hourly operations, while a mixed-integer linear programming (MILP) model was applied for 24 h planning with 21,773 constraints. Results from Mexico’s national grid show that the integrated-cycle plant generates 50% of power, while the hydroelectric plant provides 18%. The proposed model achieved a 30% energy objective using the existing architecture, demonstrating its suitability for large-scale energy management.
Tianqi et al. [
117] introduced an optimization model for smart-grid investment planning during T&D tariff reform. The proposed model integrates indices for planning efficiency and benefits, while also considering constraints from previous projects and the relative significance of different projects. A risk index model was additionally developed to address the uncertainty that can affect planning decisions. The results show that the proposed models outperform traditional models by effectively responding to changing user load profiles under peak–valley time-of-use pricing schemes and enhancing planning efficiency.
In their study, Ogbogu et al. [
118] used an optimal power-flow model to reduce investment costs, transmission losses, and maintenance requirements, as well as optimize power distribution. The proposed model is subject to constraints, such as generator power limits, bus voltages, transformer tap applications, and transmission line capability. These limitations are determined by the physical and operational principles of the power system.
3.4.2. Numerical Methods
Numerical models refer to mathematical tools that deal with physical systems by dexterously representing a system with numbers and an algorithm to simulate and analyze real-life problems. These models are critical for monitoring, planning, and optimization in the system of transmission and distribution. They are perfectly applicable in the calculation of solutions, more so in complex systems beyond the scope of analytical or manual calculation methods [
119].Bragin et al. [
120] proposed a surrogate Lagrangian relaxation (SLR) model to address the same issue in the smart grid by solving the coordinated operation of the entire T&D system. Such work aims to improve the efficiency of operationsand resource utilization. Convergence is compared with conventional methods like the sub-gradient model, where the results indicate that the model is effective and gives an added advantage to the previously available methods in enhancing the coordinated decision making.
A model where the electricity market equilibrium was presented as a transmission and distribution system in the work of Saukh et al. [
121], using electricity circuit theory, Kirchhoff’s laws, and a quadratic loss function. By using the measurement factor, this technique assists in making a more realistic simulation of the market and efficiently.
An OpenDSS model of the smart-grid planning using GIS data and CIM files was developed by Cordova et al. [
122]. This strategy can achieve better results of simulating and enhancing decision-making within the distribution system, thus supporting effective open-source smart-grid planning.
Moradi et al. [
123] proposed a coordinated mechanism of improving the performance of the T&D power system in smart-grid networks. The approach uses a mathematical step-by-step model, which considers grid risks that include the varying electricity demand, production, and behavior of electric vehicles. The proposed model is employed in battery storage, wind farms, switchable feeders, distributed generation and demand–response programs. There are also EV charging stations across different grid levels to cope with short-term uncertainty. The efficiency of the model was validated using with the IEEE RTS and test feeder of 33 nodes, validating the test experiment.
3.4.3. Artificial Intelligence Model (AI)
AI-based models are computer programs that use data for classification, prediction, and pattern recognition, encompassing a wide range of methods, such as Machine Learning (ML), Deep Learning (DL), data mining, evolutionary computation, and fuzzy logic [
124]. One of the most well-known subsets of AI is ML, which focuses on identifying patterns in large datasets [
125]. ML analyzes input features to develop a model that can be used for tasks like prediction, classification, and clustering.
There are three main categories of ML: Supervised Learning [
126], Unsupervised Learning [
127], and Reinforcement Learning (RL) [
128]. Supervised learning models are trained on labeled datasets with known input–output pairs. The aim of supervised learning is to build a model that can accurately map the input to output. Common models include linear regression, Support Vector Machines, and Random Forests.
In contrast, unsupervised learning is used for unlabeled data. The primary objective is to identify hidden patterns and structures within the data. This method is primarily used for clustering and pattern recognition with the help of models like k-means and autoencoders.
Reinforcement learning uses a goal-driven approach where agents learn by interacting with the environment to optimize long-term cumulative rewards based on prior experience. A key principle of this learning method is the Markov Decision Process (MDP), which helps the agent determine the most effective policy. RL also identifies and manages the trade-off between exploration (trying new actions to discover outcomes) and exploitation (using known-good actions).
3.4.4. Random Forest (RF)
RF is an ML model used for prediction, which enhances accuracy, performance, and model stability by integrating the outputs of multiple decision trees [
129]. RF commonly employs historical data from sensors, SCADA systems, and meteorological input to forecast a range of operational parameters. RF is also known for its ability to manage high-dimensional, nonlinear datasets that are typical of complex grid operations, and its resilience to overfitting, making it essential for applications such as load forecasting, fault detection, and optimizing power flow within smart grids [
130,
131].
3.4.5. Support Vector Machine (SVM)
SVM is a supervised learning method used for complex tasks such as classification and regression [
132]. To address these two functions, it is defined as either Support Vector Regression (SVR) for continuous value prediction or Support Vector Classification (SVC) for classification.
SVR is used to predict continuous values. The primary objective of SVR is to identify a hyperplane that accurately fits the training data, while minimizing the error for data points that fall outside a predefined margin of tolerance [
133,
134].
3.4.6. XGBoost
Extreme Gradient Boosting, or XGBoost, is a powerful technique for building predictive models. The process entails sequentially adding decision trees, where each new tree corrects the errors of its predecessors. It uses internal regularization controls to prevent over-complexity and thus reduces the risk of overfitting. Due to its speed and accuracy, XGBoost is a popular tool in a variety of fields. It is designed to process big data quickly and efficiently by performing several tasks in parallel. Because of these benefits, it is commonly used to predict energy consumption, identify equipment anomalies, and detect abnormal trends in modern power-grid systems for energy distribution and delivery [
135].
3.4.7. Artificial Neural Network (ANN)
ANNs are computational networks modeled after the human brain [
136] and are composed of hierarchies of interconnected neurons. Each neuron performs a weighted summation of its inputs, followed by an activation function [
137].
3.4.8. Long Short-Term Memory (LSTM)
Long Short-Term Memory (LSTM) networks are a variant of Recurrent Neural Networks (RNNs) designed to capture long-term dependencies to address the vanishing gradient problem. LSTMs also incorporate memory cells and a gating mechanism comprising an input gate, a forget gate, and an output gate to control the flow of information [
138].
LSTMs are deployed in smart-grid transmission and distribution systems, where they have been applied to load forecasting, anomaly detection, and identifying cyberattacks. This is because they can effectively model temporal dependencies within time-series data from sensors and meters. This capability enables earlier detection of errors or attacks, enhancing grid stability and security [
139,
140].
3.4.9. Convolutional Neural Network (CNN)
CNNs are a type of deep learning model designed to capture spatial and temporal features hierarchies from input data. In smart-grid transmission and distribution, CNNs are used to identify anomalies, cyberattacks, and system faults using real-time data streaming from devices such as Phasor Measurement Units (PMUs), Remote Terminal Units (RTUs), and smart meters. The building block of a CNN is the convolutional layer, which applies learned filters to the input data to extract informative features [
141].
In the context of the smart grid, CNNs are fundamental to many applications. They enhance cybersecurity by identifying anomalies resulting from infiltrations such as False Data Injection Attacks (FDIA), Denial of Service (DoS), and Load Redistribution Attacks. They also support in fault diagnosis by detecting and localizing line faults, transformer failures, and voltage instability. Finally, they enhance short-term load forecasting by learning temporal patterns from historical data [
142]. CNNs play an essential role in improving the efficiency, reliability, and robustness of smart-grid T&D networks, as they can automatically extract spatial–temporal dependencies.
3.4.10. Recurrent Neural Network (RNN)
Recurrent Neural Networks (RNNs) are deep learning models used to process time-series and sequential data. For smart-grid transmission and distribution systems, RNNs are useful in applications with time-dependent characteristics, such as load forecasting, anomaly detection, and state estimation. Unlike feedforward networks, RNNs do not maintain a separate representation for time steps; instead, information persists through repeating loops of weights [
140].
Due to their ability to model temporal dynamics, RNNs are suitable for forecasting future power load based on historical data. They can also detect patterns that indicate imminent failure or identify anomalies (e.g., cyberattacks like FDIA and DoS attacks) [
143]. For fault diagnosis, RNNs can process voltage and current waveforms to detect abnormalities by comparing them with normal behavioral trends. In cybersecurity, the RNN can examine communication or measurement sequences for unusual patterns.
RNN-based algorithms also have potential for load forecasting, as they can extract daily and seasonal variations, thus supporting utilities in generation and distribution planning. More advanced versions, such as LSTM and GRUs, overcome shortcomings caused by the vanishing gradients problem, making RNNs more useful for smart-grid applications [
144]. Overall, RNNs play an important role in enhancing the intelligence, adaptivity, and security of smart-grid operations.
3.4.11. Autoencoder Models
Autoencoders are unsupervised neural network architectures that learn compressed representation of the input data. They are consist of two parts: the encoder, which compresses the input into a latent-space representation, and the decoder, which reconstructs the input from this representation. The purpose of the autoencoder is to reconstruct the input as accurately as possible, with the latent space capturing the most salient features of the data [
145].
In smart-grid transmission and distribution systems, autoencoders can be used for anomaly detection, fault diagnosis, and data denoising. As an inherent feature of their learning process, autoencoders that have been trained on non-anomalous operational data are highly sensitive to deviations that result from False Data Injection Attacks (FDIA), Load Redistribution Attacks, and sensor failures [
146]. By examining the reconstruction error, the model can determine when a data point is not consistent with the expected norm. For instance, a sudden voltage collapse or abnormal load generation can be flagged as suspicious if the reconstruction loss exceeds a predefined threshold.
Autoencoders can also be employed for dimensionality reduction of high-dimensional observations from multiple smart-grid sensors, retaining the most relevant features for use as input for forecasting or classification. Furthermore, models such as sparse autoencoders, denoising autoencoders, and variational autoencoders (VAEs) enhance performance through improved feature extraction, noise robustness, and probabilistic interpretation [
147]. In this context, autoencoders represent a valuable tool for the reliable identification of disturbances and efficient representation of data, thus contributing to the security, resilience, and efficiency of smart-grid systems.
3.4.12. Self-Organizing Map (SOM)
The self-organizing map (SOM) is an unsupervised neural network model suggested by Teuvo Kohonen, known for clustering and visualizing high-dimensional data by projecting it into a lower-dimensional grid (usually 2D) [
148].
In smart-grid transmission and distribution systems, SOMs are especially beneficial for load profiling, anomaly detection, consumer behavior analysis, and fault-type detection. The SOM is consists of neurons in a 2D lattice, each of which is linked to a weight vector of the same size as the input [
149,
150]. During training, the SOM undergoes competitive learning, reorganizing itself so that similar input patterns will activate neurons that are nearby each other on the map.
In the context of the smart grid, SOMs are employed to identify patterns in consumer-load profiles to group similar consumption behavior and assist with demand-side management. They can also detect abnormal sensor or measurement data that could signal a failure or equipment defect. In addition, SOMs can help classify power quality events (PQEs) or cluster different parts of a network according to their operation [
151]. As SOMs can be trained in an unsupervised manner, they are well-suited for situations where data labeling is costly or impossible. Ultimately, SOMs enhance smart-grid operations, imparting the abilities of pattern discovery, anomaly detection, and dataset visualization.
3.4.13. Generative Adversarial Networks (GANs)
Generative Adversarial Networks (GANs), a family of deep learning models proposed in 2014 by Ian Goodfellow, operate on a generate-and-test basis. A GAN consists of two neural networks: a generator, tasked with producing synthetic data that appears realistic, and a discriminator, which acts as a classifier to determine whether the data is realistic or unrealistic [
152,
153]. The generator strives to produce more realistic data to deceive the discriminator, while the discriminator works to improve its classification accuracy. With each iteration, the generator gets better at deceiving the discriminator, and over time, it produces high-quality synthetic data that closely matches the real data distribution.
With regard to smart-grid transmission and distribution systems, GANs have several impactful applications. They can be used for data augmentation, particularly in datasets with imbalanced representation where incidents like cyberattacks (e.g., False Data Injection Attacks, DoS) are rare. By creating synthetic data of these rare incidents, GANs enhance the effectiveness of classifiers trained to detect them.
GANs are also used for anomaly detection, where they learn the characteristics of normal operational data and flag data points with high discriminator loss as anomalies. Additionally, GANs can enable privacy-preserving grid data sharing, enabling secure risk-sharing partnerships among utilities. In predictive maintenance, GANs can be used to model the effects of hypothetical fault conditions and overloads under future operating conditions [
154]. Other variants, such as Conditional GANs (CGANs) and Wasserstein GANs (WGANs), improve training stability and enable conditional generation based on labels like grid zones, time windows, or attack types. In general, GANs can make the grid smarter and more secure by promoting realistic simulation, data augmentation, and anomaly detection in complex T&D networks [
155].
3.5. Hybrid Models
The methodology in [
156] employs multivariable regression methods, including linear, polynomial, and exponential regressions, for predicting peak loads. When these models were applied to Jordan’s electricity grid, achieved an accuracy of nearly 90%, with a performance comparable to regular exponential regression techniques. As discussed in [
157], the aim is to examine issues of power consumption at the national level through active long-term demand forecasting for a deregulated electricity market. The study also underlines the importance of activities like energy procurement, infrastructure expansion, and contract management.
[
158] highlight a deficiency in prior research regarding hybrid deep learning. They note that limited work has explored combine multiple deep learning structures for large-scale smart-grid load forecasting. According to their findings, combining models improves predictive accuracy and provides insights into the relationships between various models and complex data patterns. In [
159], a hybrid algorithm was proposed that combines evolutionary deep learning with LSTM and a genetic algorithm to optimize parameters as lag windows and the number of neurons. This mixed method proved more accurate and flexible than traditional forecasting methods.
In [
160], the authors examined two recurrent neural network sequence-to-sequence (S2S) implementations to predict building-level energy consumption using GRU and LSTM layers. They also compared these hybrid models with five conventional forecasting techniques: multiple linear regression, stochastic time-series analysis, exponential smoothing, Kalman filtering, and state-space modeling, using an identical dataset for the comparison. An LSTM-based approach proposed by Yong et al. [
161] was specifically designed to forecast the same-day energy demand of buildings.
3.6. Cyberattacks in Transmission and Distribution
Cyberattacks are malicious activities designed to damage networks, steal data, and disrupt normal processes in the smart grid. Their primary goal is to compromise transmission and distribution, potentially leading to blackouts and physical damage. Several types of cyberattacks target T&D systems, with each having unique features and potential outcomes. For instance, a Man-in-the-Middle (MiTM) attack targets communication between two devices or users. Similarly, Replay Attacks (RA) and Time Synchronization Attacks (TSA) can replay or alter measurement data and signals.
TSA and False Data Injection Attacks (FDIA) are designed to manipulate smart meter readings in order to manipulate state estimation and evade bad-data detection systems. Additionally, Denial of Service (DoS) attacks prevent information from reaching its intended destination. Countermeasures are introduced to detect these attacks, protect the grid, and prioritize cybersecurity efforts and resources. Some of these cyberattacks pose a greater threat than others due to their potential impact on grid stability and security [
162].
Table 9 provides a comparative analysis of several studies focusing on cyberattacks in T&D.
3.6.1. Man-in-the-Middle Attack (MiTM)
A Man-in-the-Middle (MiTM) attack occurs when an attacker intercepts communication between two parties or devices and impersonates one of them, making it seem as though the information is being transmitted normally [
173]. Kulkarni et al. [
174] investigated the potential security threats related to MiTM attacks in power systems, focusing on vulnerabilities in the Modbus TCP/IP protocol used for communication.
Fritz et al. [
175] developed a prototype of a MiTM attack on a smart-grid emulation platform. The authors also demonstrated a technique for compromising the integrity and validity of IEEE Synchrophasor protocol packets. Detecting packet interception is difficult because of the physical distance between the PMU and the Phasor Data Concentrator (PDC). Additionally, the PDC acquires data with minimal time for encryption, authentication, and integrity testing.
3.6.2. Replay Attack (RA) and Time-Delay Attack (TDA)
Accurate timing of control signals is essential for maintaining proper system operation. Time-Delay Attacks (TDAs) compromise this by introducing random delays in data transmission and reception, which can disrupt timing-dependent control components [
175]. In contrast, Replay Attacks (R) are used to capture sensor measurements over a set period and then replace real-time data with these previous measurements. Additionally, they may involve retransmitting past recorded control signals from the operator to the actuator [
176,
177]. These two attacks use legitimate past data to manipulate system behavior, resulting in a system state that differs from the intended scenario. This can lead to major outages by disrupting and damaging the power grid [
178].
3.6.3. False Data Injection Attack (FDIA)
Smart grids represent a radical departure from conventional electrical networks, aiming to promote interconnection among operators, suppliers, and users to create a tightly linked environment. This greater interconnection introduces significant challenges, particularly concerning consumer data privacy and the safe operation of the grid. A major concern is the vulnerability of real-time system operations, including generator synchrophasormeasurements, to cyber threats like FDIAs.
FDIAs pose a significant threat to smart-grid cybersecurity. First proposed by Liu et al. [
179], compromises the state estimation (SE) process at the control center by injecting false values into measurement data. Although bad-data detection (BDD) tools, which use the l2-norm, are employed to identify inconsistencies, it is possible to design attack strategies that manipulate these systems. By injecting errors stealthily, attackers can bypass the BDD tools.
The FDIA is a hidden attack due to the subtle modification of the attack vector in the measurement data. Attackers are typically modeled as either requiring knowledge of the grid topology to launch an attack (FC-attackers) relying solely on observed data without topology information. FDIA behavior is also affected by application-specific vulnerabilities. For example, open communication links make Wireless Sensor Networks (WSNs) more susceptible. In power systems, attacks are more difficult to execute because of the complexity of modeling real grid parameters. Attackers may conduct broad or targeted attacks depending on their objectives and the available system information. Crucially, these attacks may threaten the integrity, availability, or confidentiality of the system. One example is providing false meter readings, which can lead to incorrect system operation, service outages, and private customer data leakage, as recorded in the AMI [
180].
3.6.4. Load Redistribution Attack (LRA)
LRA is a type of FDIA attack proposed by Yuan et al. [
181] and collaborators, which aims to disrupt the operation of the power system by targeting the Economic Dispatch (ED) process. ED aims to minimize generation costs and optimize power distribution to meet load demands. When LRA manipulates the estimated system state, it can lead to incorrect ED decisions, resulting in grid instable and inefficiency.
LRA has two primary objectives: immediate effects or delayed effects. The immediate effect aims to significantly increase the operating cost right after the attack. In contrast, the delayed effect aims to gradually overload power lines, which can ultimately cause physical damage to the power grid.
3.6.5. Denial of Service (DoS)
In the smart grid, there are numerous measurement devices, such as smart meters (SMs), data aggregators, Phasor Measurement Units (PMUs), Remote Terminal Units (RTUs), Intelligent Electronic Devices (IEDs), and Programmable Logic Controllers (PLCs). These devices exhibit vulnerabilities that can be exploited for Denial of Service (DoS) attacks.
A DoS attack in a power system makes measurement data unavailable to both users and the control system, disrupting communication and causing network instability. The attack also hinders the smart-grid ability to capture events during its occurrence. A DoS attack can be caused by flooding a device or communication channel with high volumes of data or exploiting system weaknesses such as jamming or routing vulnerabilities [
182,
183,
184]. A puppet attack is a newer form of DoS attack that blocks communication in the Advanced Metering Infrastructure (AMI) network [
185]. The attacker selects normal nodes to send attack data. When those nodes receive the data, the attacker continues to send more, overloading the system and blocking communication.
4. Energy Utilization
The objectives of energy T&D models are to effectively transfer energy from generation plants to consumers. Energy utilization has become a critical concern due to the pressing need to reduce consumption, combat climate change, and enhance the sustainability of energy systems. While the oil and gas industry has long driven global economic growth, the increasing link between national economies and global energy demands necessitates maximizing energy efficiency to harmonize economic growth with environmental protection. Buildings, which account for about 34% of global energy consumption and carbon emissions, are a major source of energy use and therefore a priority for efficiency enhancement, as discussed in [
186]. This specific study, however, did not use a dedicated model to examine energy utilization. In contrast, studies by [
187,
188] used optimization models to quantify the impacts of energy use and its utilization.
Conventional building energy management strategies have traditionally been based on fixed power grids and conventional energy systems, which are often inefficient, costly, and environmentally harmful [
189,
190,
191,
192]. Over the last few years, there has increasingly shifted toward smarter energy management systems through the integration of renewable energy sources (RES) into built environments [
193,
194,
195,
196]. These systems, which include solar photovoltaic (PV) systems, wind turbines (WTs), and energy-storage systems (ESS), aim to reduce dependence on the grid, lower energy costs, and minimize building’s carbon footprints [
197,
198,
199].
The concept of energy utilization optimization in buildings encompasses a diverse range of strategies, including demand-side management (DSM), smart grids, energy storage, and real-time monitoring of consumption patterns [
200,
201,
202,
203,
204]. These studies highlight that optimization techniques such as genetic algorithms and demand-side strategies are pivotal in addressing the dynamic challenges of balancing energy supply and demand, managing uncertainty, and coordinating resources. The ability to track energy utilization in real time enables adaptive building operations to optimize energy use, minimize losses, and coordinate consumption with periods of high renewable energy production [
205,
206,
207]. These studies also emphasize the evolution of solar technologies, from first-generation silicon-based cells to advanced PV/T and quantum dot cells, aiming for improved efficiency and lower costs. Such developments, supported by policy incentives and innovation, are critical for maximizing solar energy utilization and achieving global decarbonization goals.
With increasing urgency in energy conservation efforts, the advantages of combining smart meters with building energy management systems (BEMS) are abundant [
191,
208]. These systems can automate energy optimization processes, respond to changes in energy generation, and enhance overall efficiency. Smart meters facilitate flexible pricing, enable effective peak demand management, and improve grid stability. They also empower occupants to monitor and control their energy usage. At the same time, rapid cost reductions in solar, wind, and battery-storage technologies, along with climate policies and carbon pricing, have accelerated a global shift toward electricity-dominated energy systems [
209,
210]. Recent research on wind energy emphasizes the optimization of turbine performance, control systems, and offshore infrastructure to enhance efficiency, reliability, and integration into future hydrogen economies and applications.
This section discusses the importance of smart meters in maximizing energy utilization in buildings, exploring how they increase energy efficiency, decrease operational expenses, and facilitate the integration of renewable sources. By investigating new developments and trends in smart meter technology, this section provides a comprehensive overview of how these systems transform energy use in the built environment and contribute to the broader goal of sustainability.
Figure 4 provides a complete overview of energy utilization in the smart grid.
4.1. Technologies for Energy Utilization in Smart Grid
The integration of smart grids has transformed energy systems by introducing advanced technologies designed to enhance efficiency, reliability, and sustainability in energy use. For example, policy-driven demand-side management in industrial sectors further improves energy efficiency and support broader sustainability goals [
68,
211]. Smart grids are characterized by their ability to provide real-time monitoring, data analytics, and automated control, which collectively enable better energy management and optimization. As discussed by [
212,
213], demand-side management plays a pivotal role in enhancing energy efficiency, reducing emissions, and facilitating smart-grid integration through optimized consumption strategies and the use of hybrid algorithms.
Several major technologies are key to the effective operation of smart grids [
214,
215,
216]. These include renewable energy integration, Advanced Metering Infrastructure (AMI), advances in energy-storage technologies, and demand-side management tools. The key technologies that enable optimal energy use within smart grids and how they affect efficient grid operation are presented in
Table 10.
4.1.1. Advanced Metering Infrastructure (AMI)
Smart grids rely on Advanced Metering Infrastructure (AMI), which is an integrated system of smart meters, communication networks, and data management platfoms. AMI systems gather real-time data on energy usage, voltage, and power quality for analysis and monitoring by utilities. AMI systems support self-configuration, algorithm integration, and scalable deployment, paving the way for intelligent energy market participation and system resilience [
215,
217].
These systems also enable utilities to monitor energy consumption in detail, providing valuable insights into usage patterns. AMI helps facilities gather real-time energy usage data, enabling instant access for both consumers and utilities. Furthermore, AMI facilitates two-way communication that supports remote meter control and troubleshooting for both consumers and utilities.
Table 11 presents a comparative analysis of several studies and their key findings.
4.1.2. Renewable Energy Integration
Smart grids are crucial for integrating renewable energy systems, such as solar photovoltaics (PV), wind turbines (WTs), and hydropower, to overcome the constraints of conventional fossil-fuel-based energy systems. Advanced optimization techniques and smart-grid technologies are key to managing renewable variability and ensuring reliable, efficient, and sustainable energy supply [
222,
223].
Integrating a large number of renewable energy sources into the grid introduces variability and uncertainty. Smart grids handle this through advanced forecasting, energy storage, and demand–response techniques. The integration of PVs and electric vehicles (EVs) requires coordinated operation models and simulation-based techniques to mitigate grid instability, optimize load scheduling, and reduce energy costs. In parallel, modern power systems leverage AI, machine learning, and dynamic modeling to enhance grid resilience, enabling real-time control and supporting the seamless integration of renewable energy sources [
224,
225]. Moreover, intelligent grids allow for the efficient use of distributed energy resources (DERs), including rooftop solar panels and small-scale wind turbines, to provide local energy. Smart grids also employ weather forecasting, data mining, and machine learning to predict the availability of renewable energy, thereby streamlining grid operations, especially during periods of fluctuating energy production [
226,
227].
Table 12 presents a comparative analysis of several studies, highlighting their limitations.
4.1.3. Energy-Storage Systems (ESS)
One of the key elements of smart grids is the energy-storage systems (ESS), which includes batteries and pumped hydro storage facilities. These systems store energy surpluses generated by renewable sources during periods of high production and release them during times of low generation to maintain the balance between supply and demand [
233]. Accurate load forecasting using advanced deep learning models, such as the IntDEM framework, enables optimized charging and discharging schedules for ESS in IoT-enabled smart grids [
234].
ESS ensures grid stability by storing energy during low-demand periods and discharging it during peak demand, thus preventing blackouts and maintaining a constant supply. By discharging stored energy during peak load periods, ESS helps reduce the need for costly peaking power plants and reduce electricity prices [
235]. An integrated demand–response framework that combines medium-term forecasting and an optimized battery ESS can reduce costs and emissions while improving energy efficiency [
236]. Furthermore, ESS enables greater integration of intermittent renewable energy sources by compensating for fluctuations in generation, thereby ensuring a continuous and reliable energy supply.
The following
Table 13 compares several studies in the domain of energy utilization by highlighting their key findings and limitations.
4.1.4. Demand-Side Management (DSM)
Demand-side management (DSM) refers to strategies employed to influence consumers’ energy usage in order to improve efficiency and reduce grid stress. DSM systems enable utilities to actively manage consumption patterns through incentives, dynamic pricing, and direct control of appliances. For instance, one study used conceptual frameworks like intersectionality and social license to examine social barriers to DSM adoption, while another employed statistical and machine learning models to optimize DSM implementation through dynamic pricing and load-control strategies [
242,
243].
Smart grids enable dynamic pricing schemes that encourage the use of energy during off-peak periods, thereby reducing peak load demand and promoting a more balanced energy distribution [
244]. Furthermore, appliances equipped with smart technologies, such as smart thermostats, washing machines, and refrigerators, can be controlled remotely to optimize energy consumption based on grid conditions and pricing signals. The key findings and limitations of several previous studies are presented in
Table 14.
4.1.5. Artificial Intelligence and Machine Learning
Artificial intelligence (AI) and machine learning (ML) play an increasingly important role in smart-grid technology by enabling advanced analytics, real-time decision-making, and optimization. For example, an AI-driven framework using Deep Reinforcement Learning, blockchain, and federated learning has demonstrated enhance virtual power plant and smart-grid performance by optimizing renewable integration and ensuring secure, scalable operation [
250,
251].
ML algorithms can predict energy demand patterns, renewable energy generation, and potential faults in the grid, thereby allowing for proactive management and reduced downtime. Furthermore, AI-powered systems optimize energy dispatch from various resources, including renewable energy, storage, and conventional power plants, ensuring the most efficient and cost-effective energy mix. AI algorithms can analyze sensor data to detect anomalies, predict equipment failures, and initiate preventive maintenance, thereby improving grid reliability [
252].
A comparison of several studies highlights the importance of AI in the energy system, as presented in
Table 15.
4.2. Cyber Attacks on Energy Utilization in Smart Grids
The integration of advanced technologies in smart grids has clearly transformed energy management into a more efficient, sustainable, and adaptive system. Nevertheless, this revolution has introduced new vulnerabilities, especially with respect to the security of digital infrastructure. As detailed in a comprehensive review by [
258,
259], the rising threat of cyberattacks is a major challenge due to the widespread integration of IoT in modern energy systems. These studies examine smart-grid architecture, categorize cyber threats, and review past attack incidents. As smart grids rely on sophisticated networks of communication and data management and interconnected devices, they are prime targets for cybercriminals who exploit potential flaws.
The effects of such cyberattacks on energy systems can be devastating, leading to shortages of energy supply, misrepresentation of consumption data, loss of revenue, and threats to national security. This section discusses different types of cyber threats to smart grids, the severity of these attacks, and strategies for risk mitigation.
Figure 5 illustrates electricity consumption throughout the life cycle of an AI model. For example, when implementing AI models in the energy market, electricity is consumed during model development, training, and deployment. These values are detailed in a study [
253]. Furthermore, the energy sector continues to be heavily impacted by ransomware attacks, with critical infrastructure facing increasing vulnerabilities. According to a report by Sophos [
260], the primary root causes of these attacks were exploited vulnerabilities (35%), compromised credentials (32%), and phishing emails (27%), as shown in
Figure 6, while misconfiguration and other causes contributed marginally.
Figure 7 highlights the recovery methods, showing a strong reliance on backups, with the majority of organizations using backups to restore data after an attack. A smaller fraction reported paying ransom, recreating data, or using a combination of these methods.
Furthermore, an infection range analysis, shown in
Figure 8, revealed that in 30% of incidents, over 75% of the organization’s devices were impacted, demonstrating the devastating reach of such attacks. Approximately 22% of cases involved infection rates below 25%, suggesting some success in containment strategies.
4.3. Types of Cyberattacks on Energy Systems
Energy grids are typically targeted by cyberattacks that exploit the digital infrastructure supporting the generation, distribution, and consumption of energy. Such attacks vary in size, complexity, and purpose and can be classified into the following categories.
Denial of Service (DoS) Attacks
When energy management systems are attacked, an overload of traffic can cause a disruption in services. In the smart grid, a DoS attack can target communication networks, leading to delays or a complete shutdown of energy-distribution systems. Studies by [
66,
261] discuss cybersecurity in energy systems using advanced models. The first paper proposes a resilient consensus control strategy embedded with reduced-order dynamic gain filters to identify DoS attacks in multi-agent systems, thereby improving energy efficiency and resilience. The second paper presents a new DDoS attack model for electric vehicle charging stations that employs a time-variant Poisson process and Ornstein–Uhlenbeck dynamics, though it does not provide specific datasets or quantitative accuracy metrics.
4.4. False Data Injection Attack (FDIA)
In a False Data Injection Attack (FDIA), an adversary selectively manipulates measurement data or sends misleading information to the grid’s communication system. This is particularly dangerous in IoT-driven smart grids, where a large number of sensors and interconnected devices including AMI gateways, smart meters, distributed energy resources (DERs) controllers, smart inverters, Phasor Measurement Units (PMUs), Remote Terminal Units (RTUs), smart thermostats, and grid-connected sensors constantly monitor and transmit data. By vulnerabilities in these IoT devices and their messages, attackers can induce errors in state estimation, tamper with electricity consumption reports, conceal illegal activity, interfere with automatic control signals, or cause false alarms at the control center. Within the framework of smart grids, FDIAs have the potential to destabilize the grid, improperly trigger demand–response functions, change dynamic prices, and cause significant losses for both utilities and consumers.
The sophisticated nature of FDIAs makes them difficult to detect, as the injected data can closely mimic legitimate measurements and bypass common anomaly detection [
262]. The study by [
263] briefly highlights the importance and outcomes of FDIA attacks and the associated challenges. Researchers have addressed these challenges; for example, one study [
264] proposed a multimodal deep learning model, including a variational graph auto-encoder (VGAE), a temporal convolutional network (TCN), and a gated recurrent unit (GRU), for the identification and classification of FDIAs under varied power-system topologies. Tested on the IEEE-14 and IEEE-118 bus systems, the model achieved an accuracy of 91.2%. To reduce the impact of FDIAs, a CNN-based model incorporating noise elimination and state estimation was implemented in [
265]. In [
266,
267] researchers suggested a model based on variational mode decomposition (VMD), Fast Independent Component Analysis (FastICA), and an XGBoost classifier optimized with Particle Swarm Optimization (PSO). When tested on the IEEE-14 bus system, this model achieved an accuracy of 99.84%. The sophisticated nature of FDIAs makes them difficult to identify when attackers exploit flaws in communication networks. A common technique used to facilitate such data manipulation is the Man-in-the-Middle (MiTM) attack.
4.4.1. Man-in-the-Middle Attack (MiTM)
In a MiTM attack, an attacker secretly intercepts the traffic between two parties and may alter the data. By placing themselves in the communication process, they can sabotage the confidentiality and integrity of the information. MiTM attacks pose a significant threat to IoT-based smart grids, where data interactions must be secure and trustworthy to ensure reliable operations. To counter these threats, researchers have proposed several countermeasures. For example, Ref. [
268] proposed an improved intrusion detection system (IDS) that integrates a machine-learning-based detection, a Quick Random Disturbance Algorithm, and a feature-selection technique. The model was validated on the IEEE-39 and IEEE-118 bus systems with an accuracy rate of 99.99%. In a different study, Ref. [
269] proposed a BP-based hybrid deep learning model (AEXB) that integrates an autoencoder with XGBoost to detect and prevent MiTM attacks in real time. Similarly, in [
270], a hybrid deep learning algorithm combining a Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) was developed, tested on network traffic data, and achieved an accuracy of 97.85%.
4.4.2. Ransomware Attack
Besides MiTM and FDIA, ransomware attacks have emerged as a significant cybersecurity threat in smart grids [
271]. Ransomware refers to malicious software which encrypts vital system information to extort money from victims in exchange for data recovery. Ransomware’s rapid spread necessitates timely detection. To address these issues, various researchers have proposed diverse solutions. For example, Ref. [
272] presented a ransomware detection model based on a CNN, which converts binary executable files into 2D grayscale images. This approach achieved an accuracy of 96.22%. In [
273], the authors presented a twofold taxonomy of ransomware in cyber–physical systems (CPS), discussing infection vectors, targets, and objectives. They also analyzed real-life cases of ransomware against CPS to identify major security weaknesses. In a separate study, Ref. [
274] developed a deep-learning-based ransomware detection framework for SCADA-automated vehicle charging station. Appraising the performance of three deep learning models a deep neural network (DNN), a convolutional neural network (CNN), and a long short-term memory (LSTM) the authors reported that their approach had an average accuracy of 98%.
4.4.3. Advanced Persistent Threats (APTs)
These represent long-term and strategic attacks intended to gain unauthorized entry into critical energy infrastructure. The actors are typically sophisticated and/or state-sponsored. APTs can be difficult to detect, as they are covert and designed to accumulate intelligence or cause destruction over an extended period. APTs may tamper with energy grid operations, steal secrets, or disrupt renewable energy systems [
275,
276]. These studies also investigate the growing threat of APTs in energy systems, using the MITRE ATTACK framework to identify weaknesses in current mitigation strategies.
Table 16 presents a comparative analysis of several studies, identifying the key findings and limitations.
4.5. Potential Impact of Cyberattacks on Energy Utilization
The implications of cyberattacks on smart grids can have significant consequences, not just for energy distribution but also for public safety, financial markets, and the environment. These are some of the potential effects:
- 1.
Grid Disruptions and Blackouts: Cyberattacks on critical infrastructure can lead to widespread energy outages. A successful attack on the controllers of power stations or substations can result in blackouts, leaving millions of people without power and disrupting daily life. A coordinated attack on multiple grid nodes can make recovery efforts take days or even weeks.
- 2.
Financial Losses: Cyberattacks can result in significant financial losses for energy utilities and consumers. Downtime, system recovery costs, and legal liabilities can be extremely expensive for utilities. Data manipulation can lead to unnecessarily high bills for consumers, who may even face penalties for alleged over-utilization of energy due to tampered smart meters.
- 3.
Energy Efficiency and Sustainability Setbacks: Cyberattacks can sabotage energy-efficiency programs by disrupting smart data and compromising information in smart meters and grid control systems. For example, tampering with energy consumption data might lead to inefficient demand–response programs, which undermines the goals of energy use optimization and carbon reduction.
- 4.
Loss of Consumer Trust: A cybersecurity incident can erode public confidence in the energy system and the smart grid. When consumers believe their energy or personal information is at risk, they may be unwilling to accept smart meters or participate in energy-efficiency programs. This lack of public trust can make it significantly harder to implement smart-grid projects.
5. Energy Routing and Energy Internet
Energy Internet (EI) is a emerging innovation that enables peer-to-peer (P2P) energy trading and facilitates highly efficient energy distribution. According to [
286], EI can be understood as a smart-grid implementation in which energy flows over the Internet similarly to data packets. The study in [
287] highlights the advantages of EI, noting that it is open, robust, and reliable. While the concept of EI is widely discussed in China and the United States, analogous frameworks have been identified globally under different terminologies. For example, in Japan, the EI concept is commonly referred to as the Digital Grid.
The Digital Grid in Japan represents a fully decentralized energy system [
288], enabling energy transactions for P2P, stabilizing and reinforcing the grid, and supporting new on-demand energy markets. Routers within the Digital Grid facilitate the transfer of energy from point to point, analogous to the movement of data packets on the Internet. This innovation is transforming energy delivery and market dynamics, allowing for more versatile and efficient energy flow. The EI architecture can be structured into seven layers, as illustrated in
Figure 9, which are based on the Open Systems Interconnection (OSI) model.
The foundational layer is the physical layer, which consists of energy cells. The energy link layer interconnects these energy cells to enable P2P trading within an Energy Intranet. Energy Routers (ERs) play a central role in establishing and maintaining these connections. The communication layer ensures seamless information exchange among the system, stakeholders, and other entities. The business layer manages financial transactions, while the transmission layer facilitates the transfer of power between various Energy Intranets. Finally, the consumption layer regulates energy usage in the energy cells.
Recent research, such as [
289], describes the EI as a transformative network system comprising three primary layers: the physical, information, and value layers. The physical layer interconnects diverse energy sources, including heat, electricity, gas, and cooling to facilitate demand–response and energy sharing among distributed energy resources (DERs). The information layer collects data from the physical layer, enabling real-time coordination and decentralized energy management. The value layer leverages this information to create new business models.
The authors in [
290,
291] presented an EI model as a decentralized process for trading energy, where resources can exchange energy without the intervention of a centralized operator. System stability is maintained through the operation of an Energy Internet Service Provider (E-ISP), which ensures secure and stable energy exchange by regulating transaction volumes in real-time and managing centralized resources when necessary. The EI architecture is designed to be structurally compatible with traditional Internet protocols, incorporating EI Cards (analogous to MAC addresses) and Energy IP addresses to enable mobility tracking across networks. Additionally, a dedicated energy transport layer protocol ensures reliable delivery of energy communications, while an application layer facilitates standardized energy communication exchanges.
Energy Routers (ERs) play a pivotal role in integration resources within the EI, by enabling the routing of energy across geographically dispersed resources. These devices manage power dispatch, facilitate information exchange, and perfrom transmission scheduling, thereby supporting a dynamic grid and flexible grid structure. ERs are essential for regulating energy flows along interconnected lines, ensuring both stability and operational efficiency of the EI. A typical ER consists of input–output ports that interface with energy sources, electrical loads, and other ERs, as well as a core module comprising power converters and a controller. The literature presents several ER architectures with a generalized structure, illustrated in
Figure 10.
Microgrids (MGs) play a vital role within the EI by integrating intelligent devices, storage systems, distributed energy sources, and communication networks to optimize the flow of energy. As the network architecture evolves, traditional routers are being enhanced to support energy-routing functionalities, enabling bidirectional energy transfer and advance routing operations. Energy-routing challenges include consumer matching, congestion mitigation, and failure prevention through transmission scheduling, and determination of optimal energy paths. To address these issues, routing algorithms must be embedded within ERs to distribute energy from a source to a destination. Optimized energy routing not only minimizes transmission losses and maximizes energy delivery but also reduces greenhouse gas emissions and fossil fuel reliance. Furthermore, efficient routing supports renewable energy integration, enhances grid reliability, and stabilizes electricity flow. P2P energy trading and decentralized energy markets heavily rely on these routing protocols, as they facilitate cleaner energy consumption and reduce reliance on fossil-fuel-based power plants.
This section is motivated by the existing gaps in the literature on energy-routing protocols. Existing reviews [
292,
293,
294,
295,
296,
297] exhibit notable limitations: they often fail to provide a systematic classification of routing methods based on their characteristics and functionalities, and lack comprehensive comparative analyses, and offer limited guidance for future research directions. Additionally, these studies do not thoroughly address the diverse methodologies and terminologies employed to tackle energy-routing challenges, which impedes the development of robust and innovative solutions. Energy routing in modern energy networks is a rapidly evolving area, presenting challenges such as dynamic operation, security, and computational efficiency. Therefore, there is a pressing need for an in-depth comparative framework. Such a framework should evaluate routing protocols in terms of their strengths limitations, and suitability across different application environments.
This section addresses these gaps by preenting a comprehensive discussion of energy routing protocols, with particular emphasis on artificial intelligence (AI), Multi-Agent Systems (MASs), and optimization models, including metaheuristics. It offeres an in-depth overview and a structured classification of existing methods, highlighting their strengths and limitations. Furthermore, the proposed framework integrates metaheuristic optimization for minimizing energy losses, AI-driven predictive routing for adaptive decision making, and MAS to enable decentralized control in dynamic, renewable energy networks. The paper also identifies critical challenges such as reducing energy losses, ensuring scalability, integrating renewable energy sources, and responding to fluctuating demand. The future research directions identified herein are intended to guide the development of efficient and sustainable energy routing, ultimately optimizing energy flow within intelligent and decentralized power networks.
5.1. Energy Network Architecture and Routing Constraints
The energy network architecture comprises two layers: the energy transmission layer and the communication layer, interconnected through multiple Energy Routers (ERs) [
298,
299,
300,
301,
302,
303,
304]. While the architecture of ERs remains largely consistent across existing studies [
305,
306,
307,
308], variations emerge in terms of implementation details and control design strategies.
Energy-routing algorithms, implemented within the ER’s routing controller as shown in
Figure 10, adapt their routing decisions dynamically based on information received from the communication layer. It is important to distinguish the energy-routing problem from the Optimal Energy Flow (OEF) problem. The OEF problem primarily addresses system-level optimization, such as determining generation levels and reduce transmission costs. In contrast, the energy-routing problem focuses on real-time energy routing within decentralized networks that involving ERs, distributed energy resources (DERs), and peer-to-peer (P2P) energy trading. It aims to ensure effective energy distribution by determining the best routes for source–load pairs while minimizing transmission losses and adhering to energy constraints.
The (EI) model can be represented as a graph, where of Energy Routers (ERs) represent the vertices and energy transmission lines from the edges, as shown in
Figure 11. The routing of energy should be subjected to the following requirements:
The total energy loss across all paths within a network,
, in Equation (
1), must be less than the total transmitted energy
. Each path consists of one or more ERs that transfer energy packet between the destination and source.
In Equation (
2), the transmitted energy
must not exceed the maximum capacity of the selected path. This capacity, denoted as
, is determined by the minimum ER interface capacity
, and the minimum energy line capacity,
.
In Equation (
3), the total energy transmitted through an energy line must not exceed its capacity,
.
In Equation (
4), the energy entering a given ER interface must not exceed its interface capacity,
.
Minimizing transmission losses
to identify optimal energy-routing paths is a central focus in the literature [
309,
310,
311,
312,
313,
314,
315,
316,
317,
318]. The losses in a DC lines connecting two ERs depend on the line resistance
, voltage
, prior energy
, and the transmitted energy
. These parameters are incorporated into Equation (
5) to compute
for the line
.
The energy loss within an ER (
i), as described in Equation (
6), depends on the electronic converter efficiency
of that ER.
The total energy loss along a transmission path between a consumer and a producer as expressed in Equation (
7), is the sum of all losses in the ERs and the connecting energy lines along that path.
5.2. Existing Research on Energy Routing
Numerous studies have investigated the challenges of energy routing in decentralized networks, aiming to enhance efficiency and reliability. These works propose a variety of routing algorithms designed to achieve objectives, such as minimizing energy loss, ensuring scalability and adaptability, and responding to dynamic network conditions. As illustrated in
Figure 12, key strategies include Multi-Agent (MA) architectures for decentralized decision-making, metaheuristic optimization techniques for determining optimal energy paths, and artificial intelligence (AI) methods for predictive and adaptive routing. Emerging approaches based on graph theory and game theory are also being explored to further advance energy routing, with comprehensive reviews of these methodologies addressed in separate publications.
5.2.1. AI-Based Approach in Energy Routing
The application of Artificial Intelligence (AI) models and algorithms has gained prominence in automating the complexities of energy management in modern energy systems [
319,
320,
321]. This section provides an in-depth analysis of AI-based computation, highligting its roles, advantages, limitations, and applicability in enhancing energy routing within decentralized energy networks. A comprehensive overview for researchers and practitioners is presented in
Table 17.
An example of this approach is a model-free Deep Reinforcement Learning (DRL) algorithm that optimizes energy management by controlling Energy Routers (ERs) and Distributed Generators (DGs) within sub-grids. The algorithm leverages real-time observations of the system state to make energy allocation decisions [
322]. While evaluating on a nine-node network revealed limitations, including high computational demands, the requirement for large dataset, and challenges in hyper-parameter tuning.
In studies by [
323,
324], neural network-based reinforcement learning methods, including actor–critic and Q-learning algorithms, are employed to enable adaptive energy routing and demand prediction in large-scale energy networks. While these frameworks improve efficiency and resilience, they need for large training datasets, and limited scalability under dynamic conditions.
AI techniques, including DRL and neural network-based approaches, enhance energy routing by enabling adaptive, efficient, and cost-effective management in decentralized networks. However, their high computational demands, dependence on large datasets, and challenges in hyper-parameter tuning underscore the need for further research to improve scalability and support practical deployment.
Table 17 presents a comparative analysis of selected studies that apply AI to energy routing.
5.2.2. Metaheuristic Approach in Energy Routing
Metaheuristic algorithms plays a crucial role in optimizing energy-routing protocols by improving path selection, resource allocation, and load balancing, thereby contributing to robust and efficient energy-distribution networks. Their key characteristics, limitations, applicability, and advantages are summarized in
Table 18.
Firefly Algorithm (FA) The Firefly Algorithm (FA), inspired by the bioluminescent behavior, is employed in [
325,
326] to address the subscriber-matching problem in energy networks. In [
325], FA is utilized to match consumers with producers based on Euclidean distance and energy price, aiming to minimize overall costs, transmission losses, and distances. The algorithm operates in three phases: (i) parameter initialization(energy quantities, pricing, positions), (ii) fitness evaluation considering demand, supply, losses, and price, and (iii) energy quantity updates to efficiently satisfy consumer requirements. Validated on a 17-node Energy Internet (EI) network, this approach offers a decentralized solution but overlooks critical factors such as Energy Router (ER) impacts, voltage drops, and link impedance, which could lead to increased losses in smart grids. The study in [
326] extends FA to optimize energy-efficient routing paths and is tested on an 11-node network with four consumers and seven producers. While both approaches show promise, they could benefit from incorporating dynamic real-time data, enhancing scalability for larger networks, and adopting more accurate loss models. Future research may focus on developing hybrid FA-based models for real-time congestion management and power flowoptimization.
Genetic Algorithms (GAs): Genetic algorithms (GAs) are widely applied in energy optimization, as demonstrated in [
327,
328]. In [
328], GA is employed for subscriber matching by encoding solutions as chromosomes that represent producer–consumer links, aiming to minimize energy losses and overall costs in a 17-node EI network. Similarly, Refs. [
327,
329] utilize GA for energy-efficient path selection, leveraging crossover and mutation operators to identify low-loss routing paths. These methods, validated on 7- and 17-node networks, reported a 7.75% efficiency improvement over Ant Colony Optimization (ACO). Furthermore, Ref. [
328] integrates GA with graph traversal and the U-NSGA-III algorithm for multi-objective optimization, balancing carbon emissions and energy losses using a Parent-Centric Crossover (PCX) to enhance exploration. This hybrid approach was tested on an 11-node network. Despite these advances, GA methods face limitations such as scalability challenges, reliance on centralized control, and limited adaptability to dynamic conditions like network failures or demand fluctuations.
FA and GA provide robust solutions for energy routing, with FA excelling in decentralized subscriber matching and GA offering flexibility for path selection and multi-objective optimization. However, both methods face scalability and adaptability challenges, highlighting the need for hybrid and distributed approaches to improve real-time performance in complex energy networks. Additionally, bio-inspired metaheuristic algorithms, such as Discrete–Artificial Bee Colony (D-ABC) and Particle Swarm Optimization (PSO), demonstrated significant potential for optimizing energy routing in smart grid (SG) environments by effectively addressing challenges related to pathfinding, energy allocation, and operational efficiency within the Energy Internet (EI).
Discrete–Artificial Bee Colony (D-ABC): The D-ABC algorithm, as proposed in [
309], enhances the traditional Artificial Bee Colony (ABC) method for efficient energy pathfinding within constrained-capacity EI networks. By introducing mutation and crossover operations during the employed and onlooker bee phases, D-ABC achieves faster convergence and mitigates suboptimal solutions through diversified search strategies. Experimental validated in smart-grid environments demonstrates its effectiveness in addressing complex routing challenges. However, the algorithm exhibits limitations, including sensitivity to parameter tuning, memory-intensive processes, and high computational requirements for large-scale networks. Future research could explore adaptive mechanisms and leverage advanced computational resources to enhance scalability and dynamic adaptability in evolving EI systems.
Particle Swarm Optimization (PSO): Particle Swarm Optimization (PSO) has been widely applied for energy routing in a smart-grid environment [
330,
331].In [
330], PSO facilitates optimal subscriber matching by selecting producers and allocating energy to meet consumer demand efficiently. Similarly, Ref. [
331] employs PSO within a multi-objective framework, optimizing energy paths based on metrics such as transmission latency, hop count, distance, and cost, with validation on a 10-node EI network. Furthermore, a discrete PSO variant, proposed in [
332] addresses discontinuous routing issues in a 7-ER network derived from a 30-bus system, improving energy efficiency and minimizing transmission losses. Although PSO offers rapid convergence and high accuracy, it suffers from premature convergence and limited adaptability under dynamic conditions. Future reserach could focus on hybrid approaches and adaptive strategies to enhance scalability and robustness in real-world EI scenarios.
D-ABC and PSO represent powerful bio-inspired approaches for energy routing, with D-ABC excelling in efficient pathfinding and PSO offering versatile optimization for subscriber matching and path selection. Nevertheless, both face challenges related to computational complexity, parameter sensitivity, and scalability in dynamic EI environments. Future research should focus on adaptive and hybrid approaches to enhance robustness, real-time performance, and applicability in large-scale, heterogeneous networks.
Table 18.
Overview of energy-routing optimization methods and their characteristics.
Table 18.
Overview of energy-routing optimization methods and their characteristics.
References | Limitations | Applicability | Advantages |
---|
[325] | Operates under the belief that reducing Euclidean distance directly reduces losses. | Useful for producer identification in less complex, ideal smart-grid frameworks. | Optimally identifies the best-fit producers in smart-grid networks. |
[326] | Employs reduced-complexity models. | Improvement is necessary for real-world grid adaptation, along with upgrades to handle larger-scale and more complex environments. | Handle the problem of finding optimal energy paths and matching subscribers effectively. |
[327] | Based on centralized control and complete network visibility, yet it fails to address subscriber matching, scheduling, and has minimal capacity for managing congestion and system failures. | Capable of efficient energy routing in orderly networks, but lacks scalability without distributed system designs. | Enhances energy-routing efficiency by minimizing power losses, particularly across extended transmission distances, and shows strong adaptability in complex network environments. |
[328] | Centralized approach incurs heavy computational costs and needs enhanced responsiveness to real-time environmental changes. | Useful in energy-routing scenarios requiring multi-objective optimization, enabling a balance between sustainability and system performance. | Uses a multi-objective optimization framework that evaluates both power loss and carbon footprint, while introducing advanced exploration techniques to bypass suboptimal results. |
[309] | Requires considerable time for parameter optimization and large memory resources for population storage, limiting scalability in extensive networks. | Optimizes energy routing effectively in compact networks and supports energy-efficient path selection under EI capacity constraints. | Utilizes common node-based crossover and mutation to strengthen exploration and accelerate convergence in local optimization. |
[330] | Lacks scalability and adaptability, and involves substantial computational load and parameter optimization. | Designed for stable, medium-scale grids, but can be extended to larger, evolving grids using hybrid techniques. | Capable of optimizing different elements of energy routing within medium-scale grid systems. |
[331] | Early convergence may occur, hindering global optimality exploration, and evaluation steps contribute to high computational overhead. | Designed for decentralized energy coordination in dynamic systems with numerous energy sources and users, and is highly applicable to real-world contexts requiring multi-objective optimization and distributed control. | Maintains equilibrium between energy supply and demand by applying multi-objective optimization, relying on multiple criteria for effective decisions. Decentralized functionality enhances both scalability and adaptability in changing smart-grid environments. |
5.2.3. Multi-Agent Systems for Energy-Routing Optimization
Multi-Agent Systems (MAS) facilitate decentralized decision-making in energy networks, enabling autonomous agents to a manage energy flows, optimize routing paths, and allocate resources dynamically. By promoting distributed intelligence and local adaptability, MAS enhance efficiency, scalability, and resilience in modern energy management and routing applications.
A novel fully distributed peer-to-peer (P2P) control strategy of Networked Renewable Energy Resources (NRERs) is proposed in [
333]. The proposed approach integrates MAS with IoT technologies, featuring a dual-layer architecture comprising a primary droop control layer for local energy sharing and a secondary distributed diffusion algorithm layer for network-side optimization. Experimental validation on a modified IEEE 9-node feeder demonstrates a 52.34% reduction in energy losses. However, challenges such as communication latency, system complexity, and security vulnerabilities persist, necessitating advanced strategies for enhanced ER placement and loss minimization.
The study in [
334] introduces a MAS-based coordinated scheduling model for Active Distribution Networks (ADNs). The proposed model optimizes Distributed Generation (DG) outputs and minimizes operating costs through iterative bid adjustments and market clearing mechanisms. Simulation on a test comprising four suppliers, three buyers, and four ADNs demonstrates effective energy dispatch and improved system stability. However, the model primarily focuses on market coordination and does not comprehensively routing functions such as dynamic path discovery.
The work in [
335] presents a MAS-based distributed electricity trading framework enabling P2P energy sharing among prosumers. The system leverages coalition formation techniques alongside blockchain to ensure secure and transparent transactions, thereby optimizing trading costs and improving scalability. However, the framework does not address energy-routing functions such as transmission-path discovery, highlighting an opportunity for future research to extend its applicability to multi-energy router networks.
References [
336,
337] explore the integration of MAS with reinforcement learning (RL) for energy management. In [
336], a Multi-Agent Reinforcement Learning (MARL) protocol employing Q-learning optimizes energy routing across a 9-ER network, effectively minimizing losses during system failures. Similarly [
337] applies Multi-Agent Deep Reinforcement Learning (MADRL) to a 16-node multi-microgrid system, using Markov games to enhance cost efficiency and network stability. However, this framework relies on switches instead of ERs, limiting its direct applicability to fully decentralized EI networks.
MAS enhances energy routing by enabling decentralized control, facilitating P2P energy sharing, ADN scheduling, and RL-based optimization to improve efficiency and resilience. Despite these advantages, challenges such as communication latency, limited scalability, and incomplete routing functionalities highlight the need for future research on adaptive ER strategies and hybrid RL-MAS approaches.
5.3. Characteristics and Key Functions of Energy-Routing Protocol
Designing a robust energy-routing protocol for the Energy Internet (EI) requires addressing key characteristics and core functionalities to ensure efficient, secure, and adaptable energy distribution in dynamic energy networks.
Energy-Routing Schemes: Energy-routing approaches can be categorized asdistributed, centralized, and semi-centralized approaches. Distributed routing allows independent path selection, which reduces losses but may not achieve global optimization. Centralized routing enforces network constraints and minimizes transmission losses but depends on a central controller, both strategies employing a central controller, potentially reducing reliability and privacy. Semi-centralized routing combines both strategies, employing a central controller for constraint management while leveraging distributed algorithms to select efficient paths, thereby balancing computational efficiency and flexibility.
Algorithm Complexity and Computation Time: Algorithm complexity, measured in terms of memory usage and execution time, depends on network size and the employed techniques. More specialized algorithms typically increase complexity, potentially slowing computations and hindering real-time performance in large-scale EI networks.
Energy-Routing Constraints: Routing is constrained by factors such as limited energy supply, Energy Router (ER) capabilities, and link capacities. These constraints require overflow prevention and secure energy delivery, directly affecting path selection and overall system stability.
Congestion Management: Rising energy demand causes congestion, which delays delivery and increases losses. Effective management requires assessing link capacities and employing multi-packet routing to meet consumer needs through multiple sources or paths, thereby mitigating delays and reducing losses.
Failure Handling and Topology Changes: Protocols must adapt dynamically adapt to ER or link failures and topology changes to ensure reliable energy delivery, thereby maintaining network resilience during disruptions.
Security: Security measures including authentication, data integrity, and confidentiality safeguard energy networks against cyber threats, ensuring the secure transmission of energy packets.
Scenarios: Protocols must accommodate diverse scenarios including single-source single-load (SSSL), single-source multi-load (SSML), multi-source multi-load (MSML), and multi-source single-load (MSSL) ensuring flexible operation across varying source and load configurations.
Energy-Routing Functions: The functions of energy routing can be summarized into three main types:
- 1.
Energy-Efficient Path (EEP): EEP aims to minimize energy losses during transmission by accounting factors such as voltage drops, congestion, and link impedance. Optimal paths must respect ER and link capacities to prevent failures and ensure efficient energy delivery.
- 2.
Subscriber Matching (SM): SM enables demand-driven peer-to-peer (P2P) energy trading by pairing consumers with suppliers based on delivery time, energy quantity, and price. ERs perform SM to identify optimal suppliers, supporting both one-to-one and one-to-many configurations for efficient fulfillment of critical load demands.
- 3.
Transmission Scheduling (TS): TS addresses network congestion by scheduling optimal transmission paths according to energy demand and source availability. It mitigates risks associated with bidirectional flows, voltage fluctuations, and intermittent renewable generation, ensuring stable and efficient energy distribution.
5.4. Discussion and Future Directions for Energy-Routing Protocols
Metaheuristics, AI, and MAS architectures provide complementary strengths for energy routing. Metaheuristics excel at optimizing stable networks, AI facilitates dynamic adaptation, and MAS delivers decentralized resilience. Hybrid approaches, augmented with data-efficient AI techniques and security measures such as blockchain, can overcome scalability and computational challenges, fostering sustainable and efficient energy networks.
Table 19 presents a concise comparison of energy-routing protocols, highlighting their strengths, limitations, and ideal applications in decentralized energy systems.
Metaheuristic algorithms excel at optimizing energy paths and minimizing transmission losses, making them highly effective for stable networks. However, their high computational demands and sensitivity to parameter settings limit real-time deployment. Integrating reinforcement learning for adaptive parameter tuning, hybridizing with machine learning models, and leveraging GPU-based parallel processing can significantly improve convergence speed, accuracy, and scalability in dynamic energy networks.
AI methods, such as Graph Neural Networks and network flow optimization, enable adaptive, real-time decision-making in dynamic energy networks by predicting and responding to changing conditions. However, their reliance on large datasets and high computational costs poses challenges, which can be mitigated through edge computing and data-efficient algorithms such as few-shot learning.
MAS facilitate decentralized decision-making, providing scalability and flexibility for dynamic topologies and effective congestion management. Coordination challenges in large networks can be mitigated through agent clustering and reinforcement learning, enhancing overall autonomy.
The integration of AI and metaheuristics introduces challenges such as data scarcity and high computational complexity. Energy routing can become more robust and adaptive through efficient data-driven approaches, including federated learning, transfer learning, and decentralized AI systems.
Hybrid protocols that combine metaheuristics, AI, and MAS offer resilient and scalable energy routing. AI can predict demand fluctuations, MAS frameworks enhance decentralized control, and blockchain ensures security. Future research should focus on integrating these approaches to optimize adaptability, security, and sustainability in decentralized energy networks.
5.5. Real-World SCADA Testbed Using Protocol-Specific Traffic: A Case Study
In the literature discussed above, we have extensively analyzed various DL/ML, and hybrid models for detecting cyberattacks a on smart grid. However, a major limitation in existing smart-grid cybersecurity research is that most studies rely on datasets not collected from real-world SCADA devices. Many of these datasets also lack integration with smart grid-specific communication protocols, such as Modbus. Additionally, evaluations are often performed on balanced datasets, which do not accurately reflect the operational condition of real-world smart grids. In real SG infrastructure, a large number of heterogeneous SCADA devices continuously communicating using SCADA-specific protocols. Furthermore, under normal operating conditions, routine communication between devices are far more frequent than malicious activities, such as cyberattacks.
To address these challenges, a recent study conducted by [
338] proposed a model that closely replicates real-world smart-grid SCADA deployments. In this study, the authors implemented a fully simulated yet realistic SG network topology, where heterogeneous SCADA devices including distributed energy resources (DERs), DER controllers, smart meters, RTU, field sensors, g gateways, and aggregator platforms communicate using the Modbus protocol, one of the most widely used protocols in SG SCADA system.
Figure 13 illustrates the communication among these SG SCADA devices and the deployment scenarios for a potential cyberattack. In this study, the authors collected raw data from this network topology in PCAP format, capturing real-time packet-level exchanges between these devices. From the PCAP files, they extracted 29 Modbus-specific features, including packet-level metrics (such as TCP flags, source/destination IPs, and window sizes) and flow-based features (e.g., flow duration, inter-arrival times, packet entropy, retransmission counts). The resulting Modbus DDoS attack (MDDA) dataset provides one of the most protocol-relevant resources currently available for smart-grid cybersecurity research.
To evaluate model performance under realistic conditions, the authors applied several ML and DL techniques, with a particular focus on a sparse autoencoder (SAE) model. The experimental setup also incorporated Gaussian noise into the network traffic and varied the intensity and durations of attacks to emulate the unpredictability of real-world SG environments.
A key feature of this study is that the authors preserved the natural class imbalance in the dataset, with 662,750 normal samples and 233,891 DDoS samples, rather than applying oversampling or balancing techniques. This approach accurately reflects the actual operational behavior of smart-grid systems, where benign traffic is far more prevalent than attack traffic. The dataset is divided into four training and testing sets to maintain temporal consistency, and the sample distribution is presented in
Table 20 of the study.
This real-world case study demonstrates that Modbus-specific attack detection can be effectively achieved using an SAE model, attaining up to 76% detection accuracy, which surpasses baseline models such as Random Forest, CNN, and GRU, as presented in
Table 21.
Moreover, it underscores the importance of evaluating models on realistic, protocol-aware, and imbalanced dataset, rather than on generic benchmark datasets like CICIDS2017 or NSL-KDD, which fail to capture the communication structures and traffic characteristics of smart-grid environments.
6. Future Research Directions
6.1. Energy Generation
The comprehensive analysis of factors impacting energy generation, as discussed in the preceding sections, underscores significant advancements in resource assessment, technological efficiency, environmental considerations, policy frameworks, grid integration, geopolitical influences, and cybersecurity. Nevertheless, several gaps and opportunities remain for future research aimed at enhancing the sustainability, reliability, and security of power-generation systems. Below, we outline key directions for future investigation, building on the findings of the referenced studies, to address these challenges and foster innovation in the energy sector.
Considering advanced resource assessment and forecasting, the studies by [
1,
2,
3,
4,
5] demonstrate significant progress in leveraging GIS-based frameworks, Earth System Models (ESMs), and Artificial Neural Networks (ANNs) for resource assessment and forecasting. However, the growing variability of renewable resources due to climate change necessitates further refinement of these models. Future research should prioritize the development of high-resolution, real-time forecasting models that integrate multi-source data, such as satellite imagery, IoT sensor networks, and advanced climate models, to improve the accuracy of solar and wind resource predictions. Furthermore, expanding the open-access dataset initiatives, as illustrated in [
5], to encompass diverse geographical regions and extended temporal scales would enhance model validation and generalizability. Research should also investigate hybrid forecasting models that combine physics-based approaches with machine learning, addressing the limitations of standalone models, as suggested by [
4].
Enhancing technological efficiency advancements in machine learning and digital-twin technologies, as discussed by [
6,
7,
8,
9,
10], have significantly enhanced the efficiency of energy-generation systems. Future research should prioritize the developing adaptive, real-time optimization algorithms capable of dynamically adjusting operational parameters in response to fluctuating environmental conditions and grid demands. For example, extending the digital-twin framework proposed by [
9] to other renewable technologies, such as solar farms or hybrid systems, could facilitate more and accurate performance monitoring and predictive maintenance. Moreover, integrating physics-informed neural networks (PINNs), as demonstrated in [
10], with generative AI models could further improve forecasting accuracy and system resilience, particularly for offshore and remote installations.
Considering climate-resilient energy systems, the studies discussed by [
9,
11,
12,
13] emphasize the vulnerability of energy infrastructure to climate-induced disruptions, such as droughts and extreme weather events. Future research should prioritize developing climate-resilient energy infrastructure by integrating long-term climate projections into system design and operation. This includes creating adaptive-capacity models that account for multi-day weather events, as highlighted by [
12], and investigating modular, decentralized energy systems to mitigate the impact of regional climate variability. Additionally, research should examine the feasibility of hybrid renewable systems that combine solar, wind, and hydropower to buffer against climate-induced resource shortages, building on the hybrid system insights from [
3].
To address challenges and implement policy and regulatory innovations, the studies discussed by [
14,
17,
18,
19] emphasize the critical role of integrated policy frameworks in accelerating renewable energy adoption. Future research should investigate dynamic policy models that adapt to evolving technological and geopolitical landscapes. For instance, developing simulation-based policy evaluation tools that combine econometric models, as demonstrated by [
14], with real-time energy market data could assist policymakers in designing more responsive and effective regulations. Furthermore, research should aim on overcome barriers to policy implementation, such as political resistance and public acceptance, by leveraging insights from [
18] and exploring innovative engagement strategies, including gamified policy communication platforms.
The integration of hybrid energy-storage systems (HESS) and Long-Duration Energy Storage (LDES), as discussed by [
22,
25], is crucial for ensuring grid stability. Future research should optimizing HESS configurations using advanced control algorithms that account for real-time grid dynamics and renewable intermittency. Building on [
23], research should investigate co-optimization models that integrate storage, generation, and transmission planning at a global scale, especially in regions with underdeveloped grid infrastructure. Additionally, the weather-driven charging algorithm proposed by [
26] could be extended to include predictive maintenance strategies, thereby enhancing battery longevity and grid reliability in systems with high renewable penetration.
To address the geopolitical and social dynamics issues that influence grid and storage optimization, the studies discussed by [
28,
29,
30,
31] highlight the complex interplay of geopolitical risks and social acceptance in energy transitions. Future research should develop predictive models that quantify the impact of geopolitical events on energy supply chains, building on the econometric approach of [
28]. Additionally, leveraging machine learning for real-time sentiment analysis, as demonstrated by [
30], could be expanded to other social media platforms and regions to better understand and address public opposition to renewable projects. Research should also investigate game-theoretic models, as in [
31], to design cooperative international frameworks that balance geopolitical competition with sustainable energy goals.
To ensure the security and stability of energy-generation networks, cybersecurity studies discussed by [
34,
35,
36,
37,
38] and others highlight the growing sophistication of ML and DL techniques for detecting cyber threats. Nevertheless, challenges such as adversarial attacks and scalability persist. Future research should prioritize development of adversarially resilient models, building on [
38], by incorporating Generative Adversarial Networks (GANs) to simulate and counter evasive attacks, as suggested by [
43]. Additionally, research should investigate federated learning frameworks, as in [
58], to enable privacy-preserving, decentralized intrusion detection systems for distributed energy networks. Real-time, scalable detection systems that integrate streaming data, as proposed by [
45], should also be emphasized to address evolving cyber threats in smart grids.
A critical gap in the current literature is the lack of interdisciplinary approaches that integrate resource assessment, technological advancements, climate resilience, policy design, grid optimization, and cybersecurity. Future research should prioritize developing holistic frameworks that combine these dimensions. For example, integrating the GIS-based site selection models of [
1] with the cybersecurity frameworks of [
34] could facilitate the design of secure, resource-efficient renewable energy systems. Similarly, combining the climate resilience strategies of [
11] with the policy insights of [
14] could guide the development of adaptive, climate-resilient energy policies. Such integrated approaches would ensure that energy systems are sustainable, secure, and resilient to both environmental and geopolitical challenges.
Table 22 summarizes the future research directions relevant to energy generation.
6.2. Future Research Directions—T&D
The evolution of smart-grid transmission and distribution (T&D) systems, as detailed in the preceding sections, highlights significant advancements through the integration of IoT, AI, and advanced automation. Nevertheless, the complexity of smart grids introduces challenges such as cybersecurity risks, power quality issues, and the need for scalable, privacy-preserving models. Based on a critical analysis of the current state of research, several directions emerge for future investigation to enhance the reliability, security, and efficiency of T&D systems.
First, the integration of distributed energy resources (DERs) and microgrids presents opportunities for improving grid resilience and reducing reliance on traditional power plants [
88,
89]. Future research should prioritize developing scalable optimization models for DER integration that account for dynamic load profiles and the intermittent nature of renewable energy sources. Specifically, enhancing the coordination between DERs and energy management systems (EMS) using AI-driven predictive models could improve real-time decision-making and energy distribution efficiency [
90,
91]. For example, extending the work of Moradi-Sepahvand [
123] on integrated T&D models to include adaptive forecasting for EV charging and renewable energy fluctuations could mitigate short-term unpredictability.
Second, cybersecurity remains a critical concern due to the increasing prevalence of cyberattacks such as False Data Injection Attacks (FDIA), Load Redistribution Attacks (LRA), and Denial of Service (DoS) attacks [
162,
179,
181]. While models like DQDN, POMDP, and Q-learning have shown promise in detecting and mitigating these threats [
163,
167,
168], future research should prioritize developing hybrid AI models that integrate supervised, unsupervised, and reinforcement learning to detect sophisticated, multi-vector attacks. Additionally, investigating privacy-preserving techniques, including federated learning [
58] and GAN-based synthetic data generation [
79], could facilitate secure data sharing among utilities while safeguarding consumer privacy.
Third, the application of advanced AI models, such as Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and Generative Adversarial Networks (GANs), has demonstrated significant potential in anomaly detection, load forecasting, and fault diagnosis [
138,
141,
152]. Nevertheless, the scalability of these models in large-scale, real-world T&D systems remains underexplored. Future research should examine the computational efficiency and robustness of these models under noisy or incomplete data conditions, building on the work of Raghuvamsi [
83] and Liu [
82]. Furthermore, integrating quantum computing with AI models could improve the processing of high-dimensional datasets, enabling faster and more accurate predictions for complex grid operations.
Fourth, the development of direct-current (DC) distribution networks, as proposed by Javid [
73], presents a promising approach for reducing power delivery losses in data centers, residential buildings, and renewable energy applications. Future research could focus on designing hybrid AC-DC smart grids with AI-powered control interfaces to optimize power flow and minimize harmonic distortions. This direction could build on the findings of Venkatraman [
78] regarding coordinated T&D control under high distributed generation (DG) penetration.
Fifth, the role of advanced metering infrastructure (AMI) and communication networks in enabling two-way communication and real-time monitoring is well-established [
85,
87]. Nevertheless, the resilience of these systems against cyber–physical attacks, such as Man-in-the-Middle (MiTM) and Time-Delay Attacks (TDA), requires further investigation [
174,
175]. Future research should focus on adaptive encryption and authentication protocols tailored to the low-latency requirements of synchrophasor protocols and AMI networks. Additionally, leveraging blockchain technology for secure data transmission in smart grids could further enhance grid security and transparency.
Finally, the use of hybrid models combine deep learning with traditional methods, as explored by Almalaq [
158,
159], demonstrates significant potential for improving forecasting accuracy and operational efficiency. Future studies should investigate the integration of hybrid models with real-time optimization techniques, such as those proposed by Barajas-Villarruel [
116] and Tianqi [
117], to address multi-objective optimization challenges in T&D systems. These challenges include balancing energy efficiency, cost reduction, and environmental sustainability under varying operational constraints.
The following
Table 23 summarizes these future research directions, highlighting the key areas, proposed approaches, and relevant references from the current work.
6.3. Future Research Directions—Energy Utilization
The exploration of energy utilization within smart grids, as detailed in the preceding section, highlights significant advancements in technologies such as Advanced Metering Infrastructure (AMI), renewable energy integration, energy-storage systems (ESSs), Demand-side management (DSM), artificial intelligence (AI) and machine learning (ML). Nevertheless, the paper also identifies critical gaps and challenges, particularly in scalability, real-world validation, cybersecurity, and the integration of emerging technologies. These gaps present opportunities for future research to improve the efficiency, reliability, and security of smart-grid systems, supporting global sustainability objectives [
186,
187,
188].
One prominent area for future research is the scalability and practical deployment of the proposed models and frameworks. For example, studies such as [
228,
232] rely heavily on simulation-based results (e.g., IEEE 33-bus and IEEE-118 bus systems), which lack real-world validation. Future research should prioritize conducting large-scale, real-world pilot studies to assess the applicability of optimization models for renewable energy integration and ESS. These studies could address the computational complexity and data-quality challenges noted in [
230,
232], ensuring that models remain robust across diverse grid conditions and geographical regions [
231].
Cybersecurity remains as critical concern, with thehighlighting vulnerabilities to Denial of Service (DoS), False Data Injection Attacks (FDIA), Man-in-the-Middle (MiTM), ransomware, and Advanced Persistent Threats (APTs) [
258,
259]. While models such as those in [
264,
266,
268] demonstrate promising accuracy (e.g., 91.2–99.99%) in detecting cyberattacks, their reliance on specific datasets or simulated environments limits generalizability. Future research should prioritize developing universal, adaptive cybersecurity frameworks capable of operating across heterogeneous grid systems while incorporating real-time, privacy-preserving techniques, as suggested by [
280]. Additionally, exploring blockchain-based solutions for secure data transmission in smart grids could mitigate MiTM and FDIA risks, building on the preliminary work of [
250,
284].
The integration of AI and ML in smart grids, as discussed in [
250,
253,
256], offers considerable potential but faces challenges related to data scarcity, model explainability, and the energy footprint of AI systems. Future research should focus on developing lightweight, energy-efficient AI algorithms that reduce computational demands while maintaining high accuracy in forecasting and anomaly detection [
254,
255]. Hybrid AI-IoT frameworks, as proposed in [
255], could be extended to incorporate federated learning for decentralized data processing, addressing privacy concerns highligted in [
253]. Furthermore, research should examine the environmental impact of AI-driven energy systems, particularly the energy consumption associated with model training and deployment, as illustrated in
Figure 5 [
253].
The variability and intermittency of renewable energy sources, as highlighted in [
222,
223], remain significant barriers to large-scale integration. Future research should focus on developing advanced forecasting models that leverage real-time weather data and machine learning to enhance the accuracy of renewable energy-generation predictions [
226,
227]. Additionally, the high capital costs and lifespan limitations of ESS, as discussed in [
233,
241], underscore the need for research into cost-effective, sustainable battery technologies, such as next-generation lithium-ion alternatives or solid-state batteries. Studies could also investigate the integration of ESS with hybrid renewable systems to optimize load scheduling and strengthen grid stability [
236].
Demand-side management (DSM) strategies, as explored in [
245,
248], demonstrate significant potential for reducing peak loads and emissions but face challenges related to consumer adoption and regional applicability. Future research should focus on developing user-centric DSM frameworks that incorporate behavioral economics and gamification to encourage consumer participation [
242,
243]. Additionally, cross-regional studies could address the generalizability issues highlighted in [
245], ensuring that DSM models remain adaptable to diverse economic and regulatory environments.
This section highlights the role of policy incentives and carbon pricing in accelerating the adoption of smart-grid technologies [
209,
210]. Future research should examine the impact of dynamic policy frameworks on technology adoption, particularly in regions with underdeveloped regulatory systems, as emphasized in [
245]. Economic analyses could also investigate the long-term cost-benefit trade-offs of deploying AMI, ESS, and renewable integration systems, addressing the high deployment costs and economic barriers identified in [
218,
240].
Table 24 summarizes these future research directions for energy utilization.
6.4. Future Research Directions—Energy Routing
The exploration of energy-routing protocols in the Energy Internet (EI) highlights their significant potential to enhance energy distribution efficiency, scalability, and sustainability. However, as highlighted in [
292,
293,
294,
295,
296,
297], the existing literature often lacks comprehensive classifications, in-depth comparative analyses, and clear guidance for future research. Building on insights from discussed AI-based, metaheuristic, and Multi-Agent System (MAS) approaches, several critical research directions emerge to address challenges related to dynamic operation, scalability, security, and computational efficiency in decentralized energy networks.
A key research direction involves developing hybrid protocols that integrate metaheuristics, AI, and MAS to leverage their complementary strengths. Metaheuristic algorithms, such as the Firefly Algorithm (FA) [
325,
326] and genetic algorithms (GAs) [
326,
327,
328], are effective for optimizing stable networks but often face challenges in real-time adaptability and scalability. Combining these with AI techniques, such as Deep Reinforcement Learning (DRL) [
322,
323,
324], can enhance dynamic adaptability by predicting demand fluctuations and network changes. Meanwhile, MAS approaches, as demonstrated in [
333,
334,
335,
336,
337], facilitating decentralized decision-making, improving resilience and scalability. Hybrid frameworks could employ AI for predictive routing, MAS for autonomous coordination, and metaheuristics for optimal path selection, effectively addressing limitations such as high computational demands and parameter sensitivity.
Scalability remains a critical challenge, particularly for AI-based approaches require extensive datasets and high computational resources [
322,
323,
324], as well as for metaheuristics such as Discrete–Artificial Bee Colony (D-ABC) [
309] and Particle Swarm Optimization (PSO) [
330,
331,
332]. Future research should focus on incorportaing edge computing and data-efficient learning algorithms, including few-shot learning or federated learning, to reduce computational overhead and enable real-time performance in large-scale EI networks. Furthermore, leveraging GPU-based parallel processing can significantly accelerate the convergence of metaheuristics, such as D-ABC [
309], thereby enhancing their applicability in dynamic resource-constrained environments.
Security remains a critical concern, as decentralized energy networks are increasingly vulnerable to cyberattacks and data breaches [
333]. Blockchain technology, as demonstrated in [
335], offers a promising solution for ensuring secure and transparent P2P energy trading. Future work should prioritize integrating blockchain with MAS and AI-based frameworks to provide robust authentication, data integrity, and confidentiality in energy-routing protocols. Additionally, the development of lightweight cryptographic schemes optimized for resource-constrained Energy Routers (ERs) is essential to maintain high security standards without introducing significant computational overhead.
Dynamic network conditions, such as Energy Router (ER) or link failures and frequsent topology changes, necessitate routing protocols capable of real-time adaptability [
336,
337]. Future research should focus on developing adaptive ER strategies that integrate MAS with Reinforcement Learning (RL) to effectively manage congestion, mitigate failures, and address renewable energy intermittency. For example, extending MARL approaches, as proposed in [
336], to multi-ER environments could significantly enhance path discovery and minimize energy losses. Furthermore, incorporating Graph Neural Networks (GNNs) to model the EI as a graph as suggested in [
313,
318], can enable predictive and topology-aware routing, thereby improving the networks’ resilience and adaptability under dynamic conditions.
The integration of renewable energy sources introduces significant challenges due to their intermittent and unpredictable nature [
289,
290]. Future research should focus on designing routing protocols that optimize energy flow from distributed energy resources (DERs) by leveraging AI-based demand forecasting in conjunction with metaheuristic path optimization. For instance, extending the multi-objective optimization framework proposed in [
328] to explicitly account for renewable energy variability could help balance carbon emissions, minimize energy losses, and maintain grid stability. Moreover, enhancing hybrid PSO variants [
330,
331] to prioritize renewable sources in real-time routing decisions would further improve energy utilization and sustainability in dynamic Energy Internet environments.
Congestion management and transmission scheduling (TS) are essential for ensuring efficient and reliable energy delivery within the Energy Internet. Future research should prioritize the development of advanced TS algorithms that utilize AI-driven models to anticipate congestion and dynamically allocate transmission paths, as proposed in [
331]. Integrating these approaches with MAS-based distributed control strategies [
334] could significantly reduce delays and transmission losses, particular under high-demand conditions. Furthermore, optimizing TS mechanisms for diverse operational scenarios including SSSL, MSML, and MSSL will enhance system resilience and adaptability in large-scale decentralized energy networks.
To address the absence of standardized methodologies highlighted in [
292,
293], future research should focus on developing a unified framework for energy-routing protocols. This framework should define standardized components such as EI Cards, Energy IP addresses, and transport layer protocols, as suggested in [
290], to ensure compatibility with existing Internet architectures. Establishing such standards would enable seamless interoperability across diverse Energy Internet deployments, including large-scale initiatives like the Digital Grid in Japan [
288], thereby promoting scalability, reliability, and global integration of decentralized energy systems.
In summary, future research should prioritize the development of hybrid energy-routing protocols that integrate AI, metaheuristics optimization, and MAS. These protocols should leverage enabling technologies such as edge computing, blockchain, and GNNs to ensure scalability, security, and adaptability in dynamic environments.
Table 25 provides an overview of these research directions, outlining their objectives, challenges, and potential approaches.
6.5. Future Research Directions for Quantum-Based AI Models
Several AI, ML, DL, and hybrid models have been applied to address smart-grid challenges. These include Long Short-Term Memory (LSTM), Random Forest (RF), autoencoders, Artificial Neural Networks (ANNs), Generative Adversarial Networks (GANs), and hybrid models such as XGBoost, CatBoost, LightGBM, and Cluster Partition–Fuzzy Broad Learning System (CP-FBLS). While these models demonstrate strong performance in applications such as False Data Injection Attack (FDIA) detection, demand forecasting, and energy routing, they also exhibit theoretical and practical limitations that restrict their scalability, adaptability, and efficiency in complex and dynamic smart-grid environments.
For example, LSTM and Gated Recurrent Units (GRUs) are widely used for load-profile analysis and energy-routing optimization due to their capability to process sequential data while mitigating vanishing gradient issues [
161]. However, these models are computationally intensive, requiring substantial training time and resources, particularly when applied to large-scale smart-grid datasets with high temporal resolution. Their performance is further constrained by sensitivity to noisy or incomplete data, which is prevalent in real-world grid operations [
4]. Moreover, LSTM-based models often struggle with capturing long-term dependencies in highly variable renewable energy data, thereby reducing forecasting accuracy under extreme climatic conditions [
2].
Random Forest (RF) models are frequently applied to anomaly detection and demand forecasting in smart grids due to their robustness against overfitting and ability to handle high-dimensional datasets [
3]. However, RF lacks inherent mechanisms for modeling sequential dependencies, which restricts its capability to capture complex temporal patterns in time-series data. Consequently, its effectiveness diminishes in dynamic smart-grid environments where real-time adaptability is critical [
6].
Autoencoders have demonstrated high accuracy (up to 99.99%) in detecting FDIA attacks by effectively modeling normal operational patterns [
339]. However, their performance is highly sensitive to hyperparameter tuning and the quality of training data. They often fail to generalize across diverse attack scenarios, particularly against sophisticated Advanced Persistent Threats (APTs) that exploit subtle data manipulations [
275]. Additionally, their reliance on unsupervised learning increases susceptibility to false positives in noisy grid environments.
ANNs are widely employed for hybrid renewable energy forecasting and energy-routing optimization [
3,
324]. Despite their predictive capabilities, ANNs require large labeled datasets for effective training a constraint in smart-grid applications due to data privacy concerns and limited availability of real-world operational data [
5]. GANs, have emerged as a powerful tool for generating synthetic data to enhance cybersecurity training in the smart grids. However, they face significant challenges such as training instability and mode collapse, which hinder their ability to capture the full diversity of attack patterns [
340].
Hybrid models such as XGBoost, CatBoost, and LightGBM have demonstrated superior predictive performance in smart-grid applications, achieving notably low forecast errors (e.g., 8.02% in [
6]). Similarly, CP-FBLS exhibits strong performance in detecting FDIA attacks [
162]. Despite these advantages, these models are computationally intensive, demanding substantial resources for both training and real-time deployment. Moreover, their inherent black-box nature compromises interpretability, posing a significant barrier for grid operators who require explainable decision-making processes [
283]. Additionally, CP-FBLS’s dependency on fuzzy logic introduces sensitivity to parameter tuning, potentially reducing robustness in dynamic and uncertain grid environments.
MAS-based frameworks have been widely adopted for decentralized energy routing and control in smart grids [
334,
335] However, they encounter significant scalability challenges in large-scale enviornments due to high communication overhead and increased coordination complexity. Furthermore, these systems remain vulnerable to cyberattacks that exploit inter-agent communication channels, posing security risks [
338].
Collectively, these limitations computational complexity, sensitivity to data quality, limited generalization capability, and lack of interpretability underscore the necessity for innovative approaches that enhance scalability, robustness, and efficiency of AI models in smart-grid applications.
Quantum–AI integration presents a promising avenue to overcome the inherent limitations of classical AI, ML, DL, and hybrid models. By leveraging fundamental quantum principles such as superposition, entanglement, and quantum parallelism, quantum computing enables significantly accelerated computations, making it particularly suitable for complex optimization and pattern-recognition tasks in smart grids [
341].
Compared to classical models such as LSTM, ANNs, and hybrid models (e.g., [
6,
161]), which are computationally intensive, particularly for large-scale grid datasets, quantum algorithms offer significant speed advantages. Algorithms like the Quantum Approximate Optimization Algorithm (QAOA) and Variational Quantum Eigensolver (VQE) exploit quantum parallelism to accelerate complex tasks [
342]. For example, QAOA can enhance ANN training by efficiently solving combinatorial optimization problems such as hyperparameter tuning, at speeds exponentially faster than classical gradient-based methods [
343]. This capability facilitates real-time energy routing and FDIA attack detection in dynamic and large-scale grid environments.
Specifically, autoencoders and GANs face challenges in generalizing across diverse attack scenarios and handling noisy data [
339,
340]. Quantum Neural Networks (QNNs), by leveraging quantum entanglement, can more effectively model complex, high-dimensional data distributions [
344]. For instance, QNNs can enhance FDIA detection by learning subtle patterns in noisy SCADA datasets, thereby improving robustness against APTs [
275]. Additionally, quantum-enhanced feature-selection techniques can reduce input dimensionality, mitigating the negative effects of incomplete or low-quality datasets [
345].
Furthermore, the black-box nature of hybrid models such as CP-FBLS and XGBoost limits their interpretability [
283]. Quantum-inspired algorithms, including quantum kernel methods, can generate probabilistic outputs compatible with explainable AI frameworks, providing grid operators with clearer insights into decision-making processes [
343]. For example, quantum kernel-based classifiers can be integrated with MAS to deliver interpretable energy-routing decisions, thereby enhancing trust in decentralized control systems [
334].
Furthermore, MAS-based energy-routing protocols encounter scalability challenges due to high communication overhead [
335]. Quantum communication techniques, such as Quantum Key Distribution (QKD), can secure agent interactions against cyberattacks, while quantum optimization algorithms can enhance coordination efficiency in large-scale grids [
342]. For instance, quantum-enhanced reinforcement learning (QRL) can optimize Multi-Agent energy routing by efficiently solving high-dimensional Markov Decision Processes, outperforming classical methods [
336].
Current quantum hardware, including noisy intermediate-scale quantum (NISQ) devices, suffers from limited qubit counts and high error rates, making the implementation of QNNs or QAOA computationally intensive for large-scale smart-grid applications [
342]. These limitations constrain their immediate deployment in resource-constrained environments. Additionally, quantum computers require specialized infrastructure, such as cryogenic cooling, which increases deployment costs relative to classical systems [
343], posing a barrier to widespread adoption in smart grids. Quantum algorithms for smart-grid applications remain in the early stages of development, with limited real-world validation compared to established classical models like LSTM or RF [
3,
161,
345]. Furthermore, implementing Quantum–AI solutions necessitates expertise in quantum computing, which is currently scarce among grid operators and researchers [
346].
Advancements in quantum hardware and algorithms are expected to address these challenges in the near future. By 2030, improvements in qubit coherence times and error correction are projected to enable stable NISQ devices capable of supporting large-scale smart-grid optimization tasks [
342]. Cloud-based quantum computing platforms will reduce infrastructure costs, making Quantum–AI solutions more accessible to utilities [
347]. Furthermore, hybrid quantum–classical frameworks will facilitate incremental adoption by bridging current classical models with emerging quantum systems [
345]. As quantum algorithm libraries mature, their integration with smart-grid applications will become increasingly seamless, allowing quantum-enhanced AI models to operate effectively for real-time forecasting, cybersecurity, and energy routing.