Enhanced power grid reliability through 14-day machine learning predictions for proactive fault management

Climate change goals and evolving energy demands continue to place pressure on power systems, requiring robust and predictive operational capabilities. This paper addresses the challenge of forecasting feeder-level disturbances within a 14-day window — an interval identified by grid operators as highly beneficial for proactive fault management. The main focus is on prioritizing high precision in early fault detection, which is particularly well-suited to environments where unnecessary false alarms can be disruptive and expensive. Our approach applies machine learning techniques to a comprehensive proprietary dataset, spanning over a decade of aggregated power grid data, including fault registry entries, power quality metrics, consumption patterns, and operational measurements. Results demonstrate precision levels exceeding 90% for certain feeders with dense measurements, although unpredictable outages remain a recognized limitation. We also explore an extended 30-day window, which can improve recall and F1-scores, reflecting a trade-off between early warnings and prediction stability. Overall, this data-driven methodology-guided by domain insight-offers a practical means to bolster grid reliability and operational efficiency.