Transient instability poses a critical challenge to the reliable operation of modern power systems, often leading to large-scale blackouts. Despite the success of data-driven Transient Stability Assessment (TSA), its practical implementation remains limited by challenges in processing high-speed real-time data streams and preserving data privacy. To address these limitations, this article develops a novel Federated Adaptive Random Forest (FedARF) method that integrates federated learning with the Adaptive Random Forest (ARF) model. The proposed decentralized framework incorporates concept drift adaptation mechanisms to accommodate the stochastic and dynamic characteristics of modern power systems. FedARF facilitates distributed knowledge aggregation learned from various heterogeneous local data sensors (clients) to predict and evaluate the TSA status with minimal communication overhead. Comprehensive experiments on the New England 39-Bus system, the IEEE 68-Bus system, and the large-scale ACTIVIgs 25k-Bus system demonstrate the efficiency of the proposed method with an overall accuracy of 99.65%. Compared to traditional centralized forecasting methods, and state-of-the-art models, the proposed approach not only maintains high prediction accuracy but also enhances data privacy preservation while substantially reducing communication bandwidth requirements.

Partitioning the power system into smaller, manageable units facilitates better grid monitoring and control, thereby improving the grid’s stability and reliability. However, large-scale power networks consist of thousands of nodes and edges, which complicates the process of learning appropriate node embeddings and aggregating information from neighboring nodes. By representing power grids as undirected weighted graphs, this study proposes a novel power network partitioning approach using Graph Neural Networks (GNN). The proposed model simplifies the clustering objective by focusing on a single balancing term, which reduces computational complexity while maintaining competitive clustering performance. The power network is represented as a graph where the proposed GNN uses the normalized graph Laplacian, which effectively captures the complex connectivity of the nodes, instead of the traditional adjacency matrix. Active power levels serve as nodal attributes, ensuring that clusters represent both the physical and operational characteristics of the network. This dual-focused approach promotes a partitioning that is topologically coherent and functionally homogeneous, vital for enhanced grid management. When applied to the IEEE 14, 39, and 118 bus systems, the proposed method has successfully delineated coherent clusters of buses, underlining its potential for improving power grid management. The simulation results confirm the method’s efficacy and applicability.

The integration of the large language model (LLM) technology in electric vehicles (EVs) has sparked a significant leap forward in the evolution of intelligent transportation. LLM technology enables real-time, context-aware communication, thereby elevating the safety and convenience of driving experiences. LLMs play a pivotal role in refining human-vehicle interactions, offering an intuitive and responsive interface for vehicle controls and navigation systems. In addition, LLMs contribute to the sustainable development of EV technology by optimizing energy consumption patterns and supporting the integration of EVs into smart grid systems. To this end, this paper aims to review the essential elements of LLM-based EVs to emphasize their current capabilities toward smart transportation and infrastructure services. This paper explores the multifaceted contributions of LLMs in enhancing functionality, user experience, and technological development of EVs. This tutorial also addresses the challenges and future prospects of LLM applications in EVs, emphasizing their potential to transform EVs into intelligent companions on the road and pave the way for a more sustainable and user-centered future for personal transportation.

The recent surge in distributed energy resources has made voltage fluctuations more complex and unpredictable. Consequently, traditional voltage control (VC) methods such as stochastic programming and robust optimization may struggle to manage rapid and significant fluctuations. Facing this challenge, this paper proposes an efficient dueling deep Q network (Dueling DQN)-based autonomous VC method. This study formulates the VC as a Markov decision process and develops an agent that learns optimal operational strategies to maintain voltage levels within safe limits, ensuring grid stability and reliability. The proposed agent operates within the power system environment, designed to mimic real-world grid conditions, including voltage variability and load fluctuations. The Dueling DQN model processes comprehensive observations, including production levels, loads, and voltage measurements, to predict action values that ensure effective VC. The Dueling DQN architecture, training process, and operational mechanisms based on VC are thoroughly detailed. Extensive case studies performed on the modified IEEE 14-bus system and a reduced IEEE 118-bus system and conducted over numerous episodes, demonstrate that the Dueling DQN agent consistently outperforms deep Q networks derivatives and deep deterministic policy gradient approach.

The efficient and secure operation of power grids is essential for ensuring reliable electricity supply and supporting the integration of renewable energy sources. Yet, the landscape is marred by burgeoning adversarial attacks, particularly targeting power systems employing cutting-edge deep reinforcement learning (DRL) methodologies. This study proposes a proximal policy optimization (PPO) agent against a randomized adversarial opponent aiming to disrupt grid operations. The performance of the PPO agent is assessed across various power grid environments alongside several baseline agents, including the do-nothing agent, the random agent, the topology greedy agent, and the power line switch agent with adversarial training. Over multiple epochs of adversarial training, the average rewards, number of steps to resolution, and computational time are recorded. The simulation results on the IEEE 14-bus system and the reduced IEEE 118-bus system demonstrate a nuanced supremacy and applicability of the PPO algorithm compared to heuristic and randomized approaches. The main contributions of this paper include 1) Introducing an optimized PPO algorithm assessed using two IEEE bus system environments; and 2) Applying an adversarial-training-based DRL to improve the robustness of PPO alorthim’s policies in the electrical grid environment.

Photovoltaic (PV) power is progressively being subsumed into power grids. Consequently, reliable PV power forecasting (PVPF) has become essential to avoid ramp events that can adversely affect the operations of integrated power systems. This article presents a deep-learning-based algorithm for PVPF. The gated recurrent units (GRU) network was implemented to predict the non-linear spatiotemporal correlations of the weather data, leading to higher reliability of the PV stations. Experimental results obtained from actual testing demonstrate the validity of the GRU networks for accurate PVPF, contributing to the efficient operation and management of smart grids and renewable energy systems. The conducted case study shows that the proposed model outperforms bidirectional long short term memory (BiLSTM) and long short term memory (LSTM) models in terms of computation power, root-mean-square error, and mean absolute error metrics.

Condition monitoring of electrical Rotating Machines (RM) serves in structural changes detection during machine’s operation. However, the frequent fault occurrence reduces the RM remaining useful life and accelerates their deterioration. Therefore, this paper proposes an effective multi-fault classification system for the faults in electric rotating machines. The proposed method employs an Artificial Neural Network (ANN) and Synthetic Minority Over-sampling (SMOTE) technique for automatically detecting rotating machines failures. This model's efficacy stems from the use of the relief feature selection approach to identify the most affecting features and improve the model's performance. A case study analysis uses the Machinery Fault Dataset (MAFAULDA) to test the models' performance. Simulation results are obtained to demonstrate that the proposed paradigm provides outstanding performance based on a fair assessment using the MAFAULDA dataset and shows that the proposed model has a high potential to detect rotating machine state.

Accurate Short-Term Load Forecasting (STLF) has been considered a topic of extreme importance for efficient energy management, reliable energy transactions, and economic operation dispatch in smart grids. However, the continuous instability of the load demand essentially due to the high volatility of weather conditions and customers' demand behavior dramatically affects the STLF accuracy. In order to overcome this problem, five effective Deep Learning (DL) techniques are proposed for multivariate time series STLF based on Convolutional Neural Network (CNN), Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), and stacked Auto-Encoder (AE). These DL based techniques are consolidated to build stacked Bidirectional GRU (BiGRU), Convolutional LSTM (ConvLSTM), stacked Bidirectional LSTM-AE (BiLSTM-AE), hybrid CNN-LSTM-AE (CNN-LSTM), and LSTM-AE (LSTM-AE) techniques. Simulation studies are conducted to demonstrate the performance superiority of BiLSTM-AE compared to the other DL models. The main contributions of this paper include 1) integrating a variety of deep neural networks for STLF; 2) employing time series as a benchmark to compare between heterogeneous DL architectures; 3) conducting the analyses on real data set.

Temperature forecasting based on meteorological data is the key stage for an accurate estimation of PV power production and demand-side management leading to better grid stability. Typically, weather forecasting is the prediction of weather parameters for seconds until months ahead based on the historical weather database. Thus, researchers create several approaches to maximize the accuracy of these predictions and increase the period of estimation. This paper proposes a new medium and long-term temperature forecasting approach based on Multi Inputs Single Output (MISO) model base on empirical equations. The parameters of the proposed model are computed using a Recursive Least Squares (RLS) method. Using a set of real meteorological data, simulation results are presented to show the high accuracy of the proposed temperature forecasting approach.

Keyframe extraction is one of the basic procedures relating to video retrieval and summary. It consists on presenting an abstract of the video with the most representative frames. This paper presents an efficient keyframe extraction approach based on local description and graph modularity clustering. The first step is to generate a set of candidate keyframes using a windowing rule in order to reduce the data to be examined. After that, detect interest points in these set of images. Then compute repeatability between each two images belonging to the candidate set and stocks these values in a matrix that we called repeatability matrix. Finally, the repeatability matrix is modelled by an oriented graph and we will select keyframes using graph modularity clustering principle. The experiments showed that this method succeeds in extracting keyframes while preserving the salient content of the video. Further, we found good values in term of precision, PSNR and compression rate.