AI-driven consistency-aware early warning framework for smart grid stability monitoring

Abstract

The increasing penetration of renewable energy sources and distributed generation has made modern power grids more dynamic and more vulnerable to instability. This increased operational uncertainty makes timely stability monitoring essential for reliable grid operation. Most existing studies emphasize either predictive accuracy or classification performance. These methods cannot be used as early-warning systems to detect grid instability. This study aims to solve the problem of identifying grid destabilization at an early stage while maintaining acceptable predictive performance. In this study, we propose a methodology that combines regression- and classification-based stability predictions, along with consistency and early-warning analyses. A publicly available grid stability dataset was used to evaluate a combined regression–classification framework integrating prediction, consistency analysis, and early-warning assessment. Four feature selection methods and six machine learning models are evaluated for predicting the continuous stability margin and generating early instability alerts. Model performance is assessed using regression metrics, global consistency, early stability margin consistency, and early warning rate. The results show that ensemble models achieve high predictive accuracy, with XGBoost attaining an R2 of approximately 0.94. However, models with moderate accuracy demonstrate stronger early-warning capability, with Early Stability Margin Consistency values exceeding 0.95 and Early Warning Rates reaching up to 0.69. The proposed framework supports proactive grid monitoring by enabling reliable stability prediction together with early instability detection, contributing to more resilient and intelligent power grid operation.