As Smart Cities development grows, deploying advanced technologies, such as the Internet of Things (IoT), Cyber–Physical Systems, and particularly, Artificial Intelligence (AI), becomes imperative for efficiently managing energy resources. These technologies serve to coalesce elements of the energy life cycle. By integrating smart infrastructures, including renewable energy, electric vehicles, and smart grids, AI emerges as a keystone, improving various urban processes. Using the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) and the Scopus database, this study meticulously reviews the existing literature, focusing on AI technologies in four principal energy domains: generation, transmission, distribution, and consumption. Additionally, this paper shows the technological gaps when AI is implemented in Smart Cities. A total of 122 peer-reviewed articles are analyzed, and the findings indicate that AI technologies have led to remarkable advancements in each domain. For example, AI algorithms have been employed in energy generation to optimize resource allocation and predictive maintenance, especially in renewable energy. The role of AI in anomaly detection and grid stabilization is significant in transmission and distribution. Therefore, the review outlines trends, high-impact articles, and emerging keyword clusters, offering a comprehensive analytical lens through which the multifaceted applications of AI in Smart City energy sectors can be evaluated. The objective is to provide an extensive analytical framework that outlines the AI techniques currently deployed and elucidates their connected implications for sustainable development in urban energy. This synthesis is aimed at policymakers, urban planners, and researchers interested in leveraging the transformative potential of AI to advance the sustainability and efficiency of Smart City initiatives in the energy sector.

In developing distribution networks, the deployment of alternative generation sources is heavily motivated by the growing energy demand, as by environmental and political motives. Consequently, microgrids are implemented to coordinate the operation of these energy generation assets. Microgrids are systems that rely on power conversion technologies based on high-frequency switching devices to generate a stable distribution network. However, disrupting scenarios can occur in deployed systems, causing faults at the sub-component and the system level of microgrids where its identification is an economical and technological challenge. This paradigm can be addressed by having a digital twin of the low-level components to monitor and analyze their response and identify faults to take preventive or corrective actions. Nonetheless, accurate execution of digital twins of low-level components in traditional simulation systems is a difficult task to achieve due to the fast dynamics of the power converter devices, leading to inaccurate results and false identification of system faults. Therefore, this work proposes a fault identification framework for low-level components that includes the combination of Real-Time systems with the Digital Twin concept to guarantee the dynamic consistency of the low-level components. The proposed framework includes an offline trained Self Organized Map Neural Network in a hexagonal topology to identify such faults within a Real-Time system. As a case study, the proposed framework is applied to a three-phase two-level inverter connected to its digital model in a Real-Time simulator for open circuit faults identification.

The relevance of power electronics converters is undoubted and growing in many modern applications. Real-Time Simulation (RTS) is commonly used to test converters before deployment, assessing what-if scenarios that may pose risks or add costs. RTS can also run alongside the real system for online assessment through direct comparison. Most research in this area focuses on accelerating the execution of RTS to comply with smaller sampling times. As a result, Field-Programmable Gate Arrays (FPGAs) are typically preferred, while microcontrollers have been mostly overlooked under the premise of their inferior throughput. However, it has not been assessed how fast an RTS must actually be, beyond the usual suggestion of always accelerating its execution. Therefore, this study was intended to find a fair balance between computational requirements and RTS error, aiming to identify operating conditions applicable to digital systems. A case study on a commercial buck converter with an RTS in a microcontroller tested the modeling Integral Squared Error (ISE). The RTS was run at various execution frequencies while the input PWM frequency varied between 2 to 100 times faster; 6,000 randomized duty cycles (0.2 to 0.8) were tested, and their ISEs were statistically analyzed. Results show that smaller RTS time steps do not necessarily reduce error. The minimum ISE occurred when the RTS execution frequency was about 35 to 45 times higher than the PWM frequency, regardless of RTS frequency. This study challenges the idea that faster RTS always means lower error and demonstrates the potential of low-end digital devices for RTS.

This work introduces an online technique to detect variations at the milliohm-level of the equivalent internal resistance in a battery. This method is based on the voltage ripple generated by a DC-DC switching converter operating normally as its charger. It targets simple battery-associated hardware that measures only terminal voltage and output current, without additional excitation signals or specialized impedance instrumentation. An online dynamical model of the nominal converter–battery system replicates the baseline terminal voltage ripple generated under the same PWM input. The online calculated covariance matrix for the measured and model-predicted ripple components is used to reveal variations in the internal resistance. Experimental validation, using a lead-acid battery together with a series connected resistor network in the mΩ range to emulate an increase in internal resistance, showed smooth, monotonic dependencies between the proposed descriptors and resistance variations. This proof of concept suggests that converter-induced switching ripple could potentially support state-of-health estimation based on internal-resistance variations, using lightweight statistics compatible with typical charging hardware.