Predicting Extreme Weather-Induced Outages in Distribution Networks Using Graph-Based Neural Network Models

Power grid infrastructure continues to advance, yet vulnerabilities persist due to climate change and extreme weather events, which can induce extensive power outages throughout the grid. To promote efficient and timely grid evolution, accurate modeling of these outages is crucial for understanding system-level impacts and further enabling grid operators to prepare effectively. While traditional approaches rely on explicit functions relating weather conditions to physical component damage, the data necessarily required to determine such functions are often difficult to acquire, even for utility companies, and artificial intelligence (AI) is a key approach to overcome these limitations. In this paper, a Graph Attention Recurrent Neural Network (GATRNN) model is proposed to integrate multiple distribution networks, their topologies and physical attributes, and time-series weather data to predict the probability of outages across the networks under extreme weather events. The GATRNN is validated through experimental results on both synthetic and real-world events, showcasing its ability to implicitly learn the relationship between weather and physical damages of power system components as well as their impact. Furthermore, the proposed model and its tools are made open source, ensuring accessibility and transparency for further research and applications.