Multi-Scale Temporal Analysis for Failure Prediction in Energy Systems

Many previous studies on power networks focus on identifying disturbances but they often miss how the system behaves across different time scales. Many existing models struggle to predict these systems’ complex, nonlinear behavior of these systems during extreme weather conditions. This study proposes a multi-scale system temporal analysis for failure prediction in energy systems, focusing on predicting failures in electric power networks using Phasor Measurement Unit (PMU) data. The model integrates multi-scale analysis with advanced machine learning techniques to capture both short-term and long-term behavior. By analyzing disturbances across multiple temporal scales, we aim to improve failure prediction accuracy in complex energy systems. One of the main challenges with PMU data is the lack of labeled states, despite the availability of disturbance times logged as failure records. This limitation makes distinguishing between normal, pre-disturbance, and post-disturbance conditions difficult. We apply multi-scale system dynamics modeling to address this issue and extract key predictive features from PMU data for failure prediction. First, multiple domain features were extracted from the PMU data from time series data. Second, a multi-scale windowing approach was applied, using window sizes of 30s, 60s, and 180s to capture both short-term and long-term system dynamics. This multi-scale analysis allowed the model to detect patterns across different time scales. Third, feature selection was performed using Recursive Feature Elimination (RFE) to reduce the dimensionality of the dataset and identify the most important features. We selected RFE as the best feature selection for its ability to systematically eliminate less important features, thus improving model interpretability and reducing complexity. This step was crucial for identifying the most relevant features for predicting disturbances. Fourth, these features were used to train multiple machine-learning models. Our main contributions include: 1. Identifying significant features across multi-scale windows. 2. Demonstrating LightGBM as the best-performing model, achieving 0.896 for precision with multi-scale windows. 3. Showing that multi-scale window sizes (30s, 60s, 180s) significantly improved model performance by capturing both short-term and long-term trends, outperforming single-window models (0.841 score). Our current work focuses on weather-related failure modes, with future research extending this approach to other failure modes, equipment failure, and lightning events.