A hybrid machine learning approach for real-time reliability evaluation of power systems

Abstract Reliable operation of modern power systems requires accurate state evaluation and efficient load management under dynamic and uncertain conditions. This study presents an AI-enhanced hybrid optimization framework that integrates DC power flow modeling, mixed-integer linear programming (MILP), and a Transformer-based architecture to dynamically optimize generator dispatch and key reliability metrics, including Loss of Load Probability (LOLP), Expected Energy Not Supplied (EENS), and Loss of Load Frequency (LOLF). The framework incorporates a self-attention mechanism to enhance stability prediction and support the integration of renewable energy sources. The proposed framework demonstrates superior performance on the IEEE RTS-96 and Saskatchewan Power Corporation (SPC) systems, achieving 93.7% prediction accuracy with the lowest RMSE and MAE among all tested models. It outperforms benchmark models such as Convolutional Neural Networks (CNN), Convolutional XGBoost (ConXGB), Convolutional Random Forest (ConRF), Physics-Informed Neural Networks (PINN), and Graph Neural Networks (GNN), while also reducing computational time by 60.5%, confirming its efficiency and suitability for real-time reliability assessment. Additionally, the proposed approach improves cost-reliability trade-offs in load curtailment decisions, offering a scalable and adaptive solution for modern power system reliability analysis.