Reliable Smart Grid Stability Prediction Using Calibrated LightGBM

This paper presents a reliability-focused approach for smart grid stability prediction using calibrated machine learning models. A LightGBM classifier is combined with probability calibration to improve confidence estimation and support more reliable decision-making. The study analyzes prediction behavior through reliability diagrams, risk-coverage analysis, confidence distributions, and accuracy-coverage trade-offs. Results show that calibration significantly improves probability quality by reducing Brier Score and Expected Calibration Error while maintaining prediction accuracy. The proposed framework helps distinguish reliable and uncertain predictions, making it suitable for practical smart grid applications where dependable and interpretable decisions are important.