Spatially Aware GNNs with Contrastive Learning Improve Power Outage Prediction During Extreme Weather

Researchers at UConn and Eversource Energy Center have developed Spatially Aware Hybrid Graph Neural Networks (SA-HGNN) with contrastive learning to predict power outages caused by extreme weather events — before they occur.

The Problem

Extreme weather (hurricanes, ice storms, severe storms) causes widespread power outages that:

• Halt industrial operations
• Damage critical infrastructure
• Disrupt economies across sectors
• Are worsening with climate change

The Innovation

Previous models ignored the spatial effects of extreme weather. SA-HGNN addresses this by:

1. Encoding spatial relationships of both static features (land cover, infrastructure) and dynamic event features (wind speed, precipitation)
2. Hybrid graph construction combining static and event-specific dynamics
3. Contrastive learning to improve feature representations
4. Pre-emptive forecasting before weather events occur

How It Works

The hybrid GNN processes both feature types through a unified spatial graph, capturing how weather affects different parts of the grid differently.

Real-World Deployment

Developed in partnership with Eversource Energy Center, the system is designed for actual utility operations — not just academic benchmarking.

Why It Matters

• Climate adaptation — As extreme weather increases, better prediction becomes critical
• Grid resilience — Pre-emptive crew deployment and resource allocation
• Economic impact — Faster restoration reduces business losses
• Public safety — Vulnerable populations (hospitals, nursing homes) can be prioritized