Artificial Intelligence in Modern Power Systems: Recent Advances and Use Cases

1. Introduction

In recent years, artificial intelligence technology has been widely applied in power systems and energy management. For example, paper [1] proposes a fault location method for hybrid transmission lines. This method eliminates the phase “out-of-step” issue through a data synchronization algorithm and uses a physics-informed neural network (PINN) to learn the mapping from the input to the fault location. Paper [2] introduces a photovoltaic power prediction method based on a recurrent neural network. In this paper, the prediction model is updated every hour to ensure real-time performance. Paper [3] presents a Federated Hybrid Graph Attention Network with Two-step Optimization (FedHMGAT) to forecast electricity consumption. The hybrid architecture balances the demands of modeling periodic patterns and volatile peaks, and the two-step strategy addresses the “feature dilution dilemma”. Paper [4] proposes a deep spatio-temporal feature extraction network to assess the stability of the power system. An improved graph attention network and a residual bidirectional temporal convolutional network are combined as a feature extraction module, and the Kolmogorov-Arnold network is employed to construct the classification module.

The unstopping pace of other modern AI technologies provides a bigger imagination space for potential power system applications. For example, large language models (LLM) possess powerful natural language understanding capabilities, enabling the conversion of complex power system data into natural language and assisting operators in quick decision-making. Safe reinforcement learning has become a hot topic in decision-making-related problems for power systems. Computer vision, as one of the most important pillars in modern AI (artificial intelligence) development, has also demonstrated some novel applicability in the power system area, e.g., solar panel defect inspection or geographical information extraction. Leveraging graph machine learning techniques in the power system area is another trend, since the power grid has a natural graph structure, and graph convolutional neural networks can help extract certain dynamic features of the grid, e.g., for carbon emission modeling. On the other hand, traditional unsupervised machine learning methods, such as empirical mode decomposition (EMD) and spectral clustering, which do not require the output labels, are very suitable for certain power system problems, such as grid topology identification. Lastly, simple but classic deep learning methods like Long Short-Term Memory (LSTM) are still applicable in system operation and planning problems, e.g., combining time series forecasting with a carbon-pricing-oriented optimization framework.

In light of the above research interests from both industry and academic circles, this Special Issue “Artificial Intelligence in Energy Sector” aims to collect and publish recent progress made pertaining to either theoretical innovation or practical applications of cutting-edge AI methods in energy system-related areas (e.g., electrical power, carbon, gas, and heat). The topics include, but are not limited to, the following:

• AI methods for active distribution networks (outage detection, restoration, etc.);
• AI methods for power markets (trading, auction, mechanism design, etc.);
• AI methods for microgrid operation and control (island operation, protection, etc.);
• AI methods for power system dynamics (simulation, model reduction, etc.);
• AI methods for power system reliability analysis (Monte Carlo acceleration, etc.);
• AI hardware for power system applications (edge computing, embedded AI, etc.);
• AI for building energy optimization and control;
• AI for EV charging scheduling;
• Other topics involving novel AI progress in energy systems.

2. Summary of the Contributions

The Special Issue finally collects six papers as follows, and the detailed technical summary of each paper is presented below.

Large language models are a cutting-edge technology in the field of artificial intelligence. Based on the historical outage data from 14 U.S. states, in light of this, Ansarinejad [I] proposes an end-to-end data-to-prediction framework for outage duration. The framework integrates exploratory data analysis, predictive modeling, and an LLM-based interface. Through exploratory data analysis, key patterns and trends are identified, providing valuable insights for enhancing grid reliability. The study utilizes features such as planned duration, facility name, outage owner, priority, season, and equipment type and employs a random forest regression model to predict outage duration efficiently, demonstrating the practical value of integrating planning information and seasonal factors.

Furthermore, complex data analysis results are translated into natural language by the LLMs, significantly improving operators’ comprehension and decision-making efficiency. The authors suggest incorporating real-time sensor data to enable dynamic outage prediction, adopting more advanced time-series forecasting models, or enhancing the domain knowledge integration of LLMs in future work. Overall, the study establishes an intelligent outage management framework with robust scalability and practical application potential.

Guo [II] proposes a clustering-fusion-based method for used-feeder topology identification in low-voltage distribution networks. This study utilizes time-series data collected from smart meters to construct a modular data-driven framework to achieve high-accuracy topology identification. In the data preprocessing stage, the framework integrates wavelet-based noise reduction, PCA for dimensionality reduction, LSTM for temporal feature extraction, and DeepSVDD for anomaly detection, effectively improving data quality and feature representation. In the topology identification stage, seven clustering algorithms (KMeans, MeanShift, Hierarchical Clustering, Spectral Clustering, Self-Organizing Map, DBSCAN, and Density Peaks Clustering) are employed for identification. The results are integrated through a KMeans-based label fusion strategy. Experiments on three real-world distribution networks demonstrate that the method outperforms single clustering algorithms and voting-based fusion methods in terms of identification accuracy, verifying its effectiveness. The authors suggest extending the framework to multi-phase distribution systems or graph-based dynamic modeling in future work. They also propose utilizing longer time-series data and more advanced signal processing techniques to further improve performance. In summary, this study develops a modular framework that identifies user-feeder topology in low-voltage distribution networks.

To address the challenges of scalability and profitability in Renewable Energy Communities (RECs), Gonçalves [III] proposes a Machine Learning as a Service (MLaaS) framework. In this framework, each REC device is equipped with a customized reinforcement learning agent and an electricity price forecasting model to support decision-making. The system is based on an infrastructure compliant with MLOps standards, supporting parallel training pipelines and autoscalable inference services. Experiments were conducted using the open-source simulator PyMGrid, and the results demonstrate that the framework can significantly reduce costs, for example, lowering operational costs by up to 96.41%. The authors note that to protect tenant privacy, the agents are currently trained independently, and future work plans to incorporate federated learning for collaborative learning. Additionally, they suggest employing more advanced reinforcement learning algorithms and forecasting models to improve performance. Overall, this study constructs a scalable and cost-effective intelligent energy management platform, MLaaS, promoting sustainable and efficient energy management practices.

Wang [IV] proposes a novel distribution system planning method based on a carbon pricing optimization mechanism. This method aims to address the challenges arising from the large-scale integration of renewable energy and electric vehicle charging stations into distribution networks. First, the study combines complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and multi-channel LSTM networks to accurately predict wind and photovoltaic generation output based on features such as climate conditions. Then, based on the behavioral characteristics of electric vehicle users, the study uses queuing theory and probabilistic prediction models to model the loads of fast-charging and slow-charging users, respectively. Furthermore, an optimized carbon trading pricing mechanism is introduced for carbon emission cost accounting. This mechanism enables carbon pricing to guide the deployment of renewable energy and the optimization of charging stations. This research provides a systematic framework for new distribution system planning based on carbon perception, data-driven approaches, and collaborative optimization. Case studies based on the IEEE 33-bus system verify the economic efficiency and carbon reduction benefits of the proposed method.

To clearly characterize the dynamic carbon emission status of the power system, Fang [V] proposes a data-driven method. Based on the IEEE 118-bus system and one year of real operational data, the study first constructs a nodal marginal emission factor system. Then it simulates power system dispatch using DC optimal power flow. By combining time-series load, renewable energy generation, and environmental variables, a graph convolutional neural network (GCN) is employed to fit the spatial and temporal relationships between nodal marginal emission and other features. This method can accurately identify the marginal emission characteristics of the power system and avoid time-consuming real-time optimal power flow computation. Compared with other models (such as LSTM, ARIMA, and LASSO), GCN achieves lower normalized root mean square error (nRMSE) and normalized mean absolute error (nMAE). The study shows that an excessive tendency to reduce marginal emission may compromise the economic operation of the power system. In summary, this study proposes a framework for accurately identifying marginal emission characteristics and utilizes a GCN to achieve a precise estimation of the marginal emission factor.

Noise barriers along highways occupy large areas of land, but they have relatively low land-use efficiency. Installing photovoltaic (PV) panels on these noise barriers can enable renewable energy generation without additional land occupation. Tavares [VI] proposes a YOLO-based object detection algorithm to automatically identify the location of noise barriers along highways and estimate the potential power generation capacity after PV installation. Similarly to other computer vision techniques, the model’s performance highly depends on the diversity of the dataset and the consistency of annotations, indicating that manual data labeling will be one of the main challenges for the model. Nevertheless, when the dataset is sufficient, the model can accurately detect noise barriers and provide valuable guidance for PV system deployment. The study demonstrates the application prospects of computer vision in augmenting the use of renewable energy.