Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Graph Attention Network for Predicting Duration of Large-Scale Power Outages Induced by Natural Disasters

Chenghao Duan, Chuanyi Ji

TL;DR

This work tackles predicting the duration of large-scale power outages caused by natural disasters by leveraging spatial structure in outage data. It introduces BiGAT, a bimodal Graph Attention Network with an unsupervised pre-training stage that assigns county nodes to one of two embedding clusters, followed by semi-supervised, attention-based learning to classify outage duration into three categories: short ($<2$ days), medium ($2$-$6$ days), and long ($>6$ days). Using real data from four Atlantic hurricanes (2017–2020) across 501 counties and 11.6 million customers, BiGAT outperforms XGBoost, Random Forest, GCN, and GAT by 2–15% in overall and class-wise metrics, achieving about 93% average accuracy. The model’s ability to incorporate spatial dependence and heterogeneity renders it a promising tool for resilience planning and emergency response, though generalization to unseen events remains an area for future work, including multi-event training and feature importance analyses.

Abstract

Natural disasters such as hurricanes, wildfires, and winter storms have induced large-scale power outages in the U.S., resulting in tremendous economic and societal impacts. Accurately predicting power outage recovery and impact is key to resilience of power grid. Recent advances in machine learning offer viable frameworks for estimating power outage duration from geospatial and weather data. However, three major challenges are inherent to the task in a real world setting: spatial dependency of the data, spatial heterogeneity of the impact, and moderate event data. We propose a novel approach to estimate the duration of severe weather-induced power outages through Graph Attention Networks (GAT). Our network uses a simple structure from unsupervised pre-training, followed by semi-supervised learning. We use field data from four major hurricanes affecting $501$ counties in eight Southeastern U.S. states. The model exhibits an excellent performance ($>93\%$ accuracy) and outperforms the existing methods XGBoost, Random Forest, GCN and simple GAT by $2\% - 15\%$ in both the overall performance and class-wise accuracy.

Graph Attention Network for Predicting Duration of Large-Scale Power Outages Induced by Natural Disasters

TL;DR

This work tackles predicting the duration of large-scale power outages caused by natural disasters by leveraging spatial structure in outage data. It introduces BiGAT, a bimodal Graph Attention Network with an unsupervised pre-training stage that assigns county nodes to one of two embedding clusters, followed by semi-supervised, attention-based learning to classify outage duration into three categories: short ( days), medium (- days), and long ( days). Using real data from four Atlantic hurricanes (2017–2020) across 501 counties and 11.6 million customers, BiGAT outperforms XGBoost, Random Forest, GCN, and GAT by 2–15% in overall and class-wise metrics, achieving about 93% average accuracy. The model’s ability to incorporate spatial dependence and heterogeneity renders it a promising tool for resilience planning and emergency response, though generalization to unseen events remains an area for future work, including multi-event training and feature importance analyses.

Abstract

Natural disasters such as hurricanes, wildfires, and winter storms have induced large-scale power outages in the U.S., resulting in tremendous economic and societal impacts. Accurately predicting power outage recovery and impact is key to resilience of power grid. Recent advances in machine learning offer viable frameworks for estimating power outage duration from geospatial and weather data. However, three major challenges are inherent to the task in a real world setting: spatial dependency of the data, spatial heterogeneity of the impact, and moderate event data. We propose a novel approach to estimate the duration of severe weather-induced power outages through Graph Attention Networks (GAT). Our network uses a simple structure from unsupervised pre-training, followed by semi-supervised learning. We use field data from four major hurricanes affecting counties in eight Southeastern U.S. states. The model exhibits an excellent performance ( accuracy) and outperforms the existing methods XGBoost, Random Forest, GCN and simple GAT by in both the overall performance and class-wise accuracy.

Paper Structure

This paper contains 14 sections, 6 equations, 4 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Method
  3. Variables
  4. Data
  5. Spatial correlation and heterogeneity
  6. Problem formulation
  7. Bimodal GAT
  8. Pre-training for node embedding
  9. Overall BiGAT structure
  10. Experimental setting for training and testing
  11. Results
  12. Performance metrics
  13. Model performance
  14. Conclusion

Figures (4)

  • Figure 1: K-means clustering plot for four hurricane events: (a) Florence, (b) Irma, (c) Laura, (d) Michael. $k=2$ is the number of clusters. Red area: significantly impacted counties. Yellow area: less impacted counties. Dark lines: boundaries of wind swaths of 64 knots, 50 knots, and 34 knots counted from inside out.
  • Figure 2:
  • Figure 3: BiGAT versus non-graph models - performance comparative analysis: (a) Florence (b) Irma (c) Laura (d) Michael. Each subplot presents three performance metrics: classification accuracy, macro F1, and balanced accuracy. Models are XGBoost, Random Forest, and BiGAT.
  • Figure 4: Graph models performance comparative analysis: (a) Florence (b) Irma (c) Laura (d) Michael. Each subplot presents three performance metrics: classification accuracy, macro F1, and balanced accuracy. Models are GCN, GAT, BiGAT undirected, and BiGAT.