Advancing outage prediction modeling: Incorporating circuit spatial relationships with graph neural networks and lidar-derived tree risk

Storms significantly challenge electric utilities by causing power distribution disruptions. Outage prediction is crucial to utility preparedness, but the development of predictive models is complex due to the wide range of factors influencing storm outages. In this paper, we introduce new features related to outage prediction modeling focusing on the proximity of trees to power lines and vegetation management programs (tree trimming and removal), emphasizing power lines susceptible to trees. In addition, we address the imbalance in the dataset associated with the rarity of extreme weather events (e.g. hurricanes) by employing an event severity classification. Our dataset includes Connecticut outage events (2005–2023), combining weather, infrastructure, vegetation, and environmental data at the circuit level. We analyzed various machine learning models and selected three top-performing models representing different categories: ensemble methods, neural networks, and graph neural networks. Using Leave-One-Storm-Out Cross-Validation (LOSO-CV), Graph Convolutional Networks (GCN) outperformed others in accuracy and exhibited an R-squared of 0.94, a Mean Absolute Percentage Error of 50%, and a Nash Sutcliffe Efficiency of 0.88. The model accurately predicted circuit-level outages for Hurricane Irene (2011), a challenging event with nearly 16,000 outages in Connecticut. The GCN model’s performance was comprehensively evaluated across various spatial scales. Permutation analysis was performed to assess feature contributions, providing insights into the model's reliance on specific predictors. By effectively capturing spatial relationships and contextual data, the GCN model enhances outage prediction accuracy, offering utility companies a valuable tool for improving storm preparedness and resiliency planning.