Multi-Scale Temporal Analysis for Failure Prediction in Energy Systems

TL;DR

This paper introduces a multi-scale temporal analysis approach using PMU data and machine learning to improve failure prediction in energy systems, especially under extreme weather conditions.

Contribution

It presents a novel multi-scale analysis method combined with feature selection and machine learning, outperforming single-scale models in failure prediction accuracy.

Findings

LightGBM achieved 0.896 precision in failure prediction.

Multi-scale analysis outperforms single-window models (0.841 accuracy).

Key features are identified across different temporal scales.

Abstract

Many existing models struggle to predict nonlinear behavior during extreme weather conditions. This study proposes a multi-scale temporal analysis for failure prediction in energy systems using PMU data. The model integrates multi-scale analysis with machine learning to capture both short-term and long-term behavior. PMU data lacks labeled states despite logged failure records, making it difficult to distinguish between normal and disturbance conditions. We address this through: (1) Extracting domain features from PMU time series data; (2) Applying multi-scale windows (30s, 60s, 180s) for pattern detection; (3) Using Recursive Feature Elimination to identify key features; (4) Training multiple machine learning models. Key contributions: Identifying significant features across multi-scale windows; Demonstrating LightGBM's superior performance (0.896 precision); Showing multi-scale analysis outperforms single-window models (0.841). Our work focuses on weather-related failures, with plans to extend to equipment failure and lightning events.