Quantile Regression for Power Grid Damage Prediction

Abstract

This article addresses the challenge of predicting electrical grid damage in the Latvian regions of Zemgale and Kurzeme using machine learning and quantile regression techniques. The study leverages a comprehensive dataset spanning 2009–2025, incorporating environmental, infrastructure, seasonal, and weather-related features. To handle the significant data imbalance, various spatial and temporal aggregation strategies are tested. Feature selection and aggregation are carefully tailored to preserve correlations between predictors and damage counts. The XGBoost algorithm, configured for quantile regression, is employed with extensive hyperparameter tuning and cross-validation. Results indicate that models trained at the maintenance-region scale (6-cell) achieve the best overall balance between accuracy and reliability, with a mean absolute error of 4.09 and an R2 of 0.53 for daily damage prediction. While the state region (2-cell) model excels in predicting high damage count events (30 or more), it also provides more precise minimum and maximum bounds, offering actionable uncertainty bounds for operational planning. The findings highlight the importance of spatial aggregation and feature engineering in improving predictive performance for rare-event forecasting in power grids. Several promising directions for further development have been identified, and the obtained results are sufficiently promising to support the application of the proposed methodology to broader geographical regions.