The management of decentralized energy resources and smart grids needs novel data-driven low-latency applications and services to improve resilience and responsiveness and ensure closer to real-time control. However, the large-scale integration of Internet of Things (IoT) devices has led to the generation of significant amounts of data at the edge of the grid, posing challenges for the traditional cloud-based smart-grid architectures to meet the stringent latency and response time requirements of emerging applications. In this paper, we delve into the energy grid and computational distribution architectures, including edge–fog–cloud models, computational orchestration, and smart-grid frameworks to support the design and offloading of grid applications across the computational continuum. Key factors influencing the offloading process, such as network performance, data and Artificial Intelligence (AI) processes, computational requirements, application-specific factors, and energy efficiency, are analyzed considering the smart-grid operational requirements. We conduct a comprehensive overview of the current research landscape to support decision-making regarding offloading strategies from cloud to fog or edge. The focus is on metaheuristics for identifying near-optimal solutions and reinforcement learning for adaptively optimizing the process. A macro perspective on determining when and what to offload in the smart grid is provided for the next-generation AI applications, offering an overview of the features and trade-offs for selecting between federated learning and edge AI solutions. Finally, the work contributes to a comprehensive understanding of edge offloading in smart grids, providing a Strengths, Weaknesses, Opportunities, and Threats (SWOT) analysis to support cost–benefit analysis in decision-making regarding offloading strategies.

The Digital Twins (DTs) offer promising solutions for smart grid challenges related to the optimal operation, management, and control of energy assets, for safe and reliable distribution of energy. These challenges are more pressing nowadays than ever due to the large-scale adoption of distributed renewable resources at the edge of the grid. DTs are leveraging on technologies such as Internet of Things (IoT), big data analytics, machine learning, and cloud computing, to analyze data from different energy sensors, view and verify the status of physical energy assets and extract useful information to predict and optimize the asset’s performance. In this paper, we will provide an overview of the DTs application domains in the smart grid while analyzing existing the state-of-the-art literature. We have focused on the following application domains: energy asset modeling, fault and security diagnosis, operational optimization, and business models. Most of the relevant literature approaches found are published in the last 3 years showing that the domain of DTs application in smart grid is hot and gradually developing. Anyway, there is no unified view on the DTs implementation and integration with energy management processes, thus, much work still needs to be done to understand and automatize the smart grid management.

As IoT metering devices become increasingly prevalent, the smart energy grid encounters challenges associated with the transmission of large volumes of data affecting the latency of control services and the secure delivery of energy. Offloading computational work towards the edge is a viable option; however, effectively coordinating service execution on edge nodes presents significant challenges due to the vast search space making it difficult to identify optimal decisions within a limited timeframe. In this research paper, we utilize the whale optimization algorithm to decide and select the optimal edge nodes for executing services’ computational tasks. We employ a directed acyclic graph to model dependencies among computational nodes, data network links, smart grid energy assets, and energy network organization, thereby facilitating more efficient navigation within the decision space to identify the optimal solution. The offloading decision variables are represented as a binary vector, which is evaluated using a fitness function considering round-trip time and the correlation between edge-task computational resources. To effectively explore offloading strategies and prevent convergence to suboptimal solutions, we adapt the feedback mechanisms, an inertia weight coefficient, and a nonlinear convergence factor. The evaluation results are promising, demonstrating that the proposed solution can effectively consider both energy and data network constraints while enduring faster decision-making for optimization, with notable improvements in response time and a low average execution time of approximately 0.03 s per iteration. Additionally, on complex computational infrastructures modeled, our solution shows strong features in terms of diversity, fitness evolution, and execution time.

The AI-driven applications and large data generated by IoT devices connected to large-scale utility infrastructures pose significant operational challenges, including increased latency, communication overhead, and computational imbalances. Addressing these is essential to shift the workloads from the cloud to the edge and across the entire computing continuum. However, to achieve this, significant challenges must still be addressed, particularly in decision making to manage the trade-offs associated with workload offloading. In this paper, we propose a task-offloading solution using Reinforcement Learning (RL) to dynamically balance workloads and reduce overloads. We have chosen the Deep Q-Learning algorithm and adapted it to our workload offloading problem. The reward system considers the node’s computational state and type to increase the utilization of the computational resources while minimizing latency and bandwidth utilization. A knowledge graph model of the computing continuum infrastructure is used to address environment modeling challenges and facilitate RL. The learning agent’s performance was evaluated using different hyperparameter configurations and varying episode lengths or knowledge graph model sizes. Results show that for a better learning experience, a low, steady learning rate and a large buffer size are important. Additionally, it offers strong convergence features, with relevant workload tasks and node pairs identified after each learning episode. It also demonstrates good scalability, as the number of offloading pairs and actions increases with the size of the knowledge graph and the episode count.