AI-Driven Early-Warning System for Weather-Related Faults in Power Distribution Networks

The increasing frequency and severity of weather-related faults in power distribution networks pose significant challenges to reliability and operational planning. Electric utilities are turning to AI-driven solutions to improve efficiency, sustainability, and resilience through proactive fault management. In this work, we propose an early-warning fault monitoring system leveraging meteorological data and historical distribution fault logs. We validate the proposed system with a real-world case study in collaboration with a local utility in a previously identified high-impact service area. For fault detection we evaluated different production-grade ensemble methods where LightGBM was selected for a balanced detection performance. Three machine learning models: Random Forest, XGBoost, and LightGBM, were each tuned via particle swarm optimization and compared on a test dataset. The LightGBM model achieved the best fault detection balance, outperforming the other candidates. To make the system applicable in the case study, predictions are exposed through a Flask-based web interface featuring an interactive Leaflet map, where users can toggle between the current week's and the following week's risk predictions, along with weather metrics forecasts. We also address common challenges in deploying smart grid applications such as data availability, open access records and extensive preprocessing requirements. The deployed tool provides actionable real-time risk insights, delivering actionable insights for proactive crew deployment and resource allocation.