Power Outage Detection

Subteam 

• Prof. Mengjie Li

• Alejandro Aparcedo

• Christian Lopez

Power Outage Detection Framework

We use spatiotemporal graph neural networks to process satellite imagery. 

We use computer vision and spatiotemporal graph neural networks extract information from satellite imagery and generation predictions a few timesteps ahead (a few days). In short, we compress satellite imagery to vector representations that will be used by the ST-GNN; ST-GNN outputs are decompressed back to the original input shape.

Results

Figure 2: Next-day predictions of nighttime light (NTL) patterns for two specific hurricanes: Hurricane Michael in Bay County and Hurricane Ian in Lee County.  As seen immediately after landfall (1 & 2 days after), the model  struggles to accuracy predict when major outage (significant drops in NTL) will occur, once outages have happen predictions closely match ground truth ground truth NTL outage patterns.

Figure 3: Post-processing of next-day predictions of NTL patterns for Hurricane Michael in Bay County and Hurricane Ian in Lee County (Figure 2). Model predictions for Bay County reveals the slow recovery of the grid (3 & 4 days after landfall). Bottom rows reveals the recovery of the grid despite the predictions being noisy.

While the model may not predict initial outages well, its ability to match post-landfall NTL patterns could be valuable for estimating the extent of damage and tracking recovery progress in the immediate aftermath of a major weather event. As well as identifying areas with poor power outage resilience

Power Outage Risk Maps

We analyzed images during outages and calculated pixel-level vulnerability and exposure values. Using the risk equation Risk = Exposure × Vulnerability, we assessed the power outage risk at the pixel level​. 

Risk = Exposure × Vulnerability​

Exposure: The number of affected pixels.​

Vulnerability: The susceptibility of a pixel to power outages