Multimodal Power Outage Prediction for Rapid Disaster Response and Resource Allocation

TL;DR

This work proposes a novel visual spatiotemporal framework for predicting nighttime lights (NTL) and power outage severity and location before and after major hurricanes, and central to the solution is the Visual-Spatiotemporal Graph Neural Network (VST-GNN), optimizing spatiotemporal visual data utilization.

Abstract

Extreme weather events are increasingly common due to climate change, posing significant risks. To mitigate further damage, a shift towards renewable energy is imperative. Unfortunately, underrepresented communities that are most affected often receive infrastructure improvements last. We propose a novel visual spatiotemporal framework for predicting nighttime lights (NTL), power outage severity and location before and after major hurricanes. Central to our solution is the Visual-Spatiotemporal Graph Neural Network (VST-GNN), to learn spatial and temporal coherence from images. Our work brings awareness to underrepresented areas in urgent need of enhanced energy infrastructure, such as future photovoltaic (PV) deployment. By identifying the severity and localization of power outages, our initiative aims to raise awareness and prompt action from policymakers and community stakeholders. Ultimately, this effort seeks to empower regions with vulnerable energy infrastructure, enhancing resilience and reliability for at-risk communities.

Figures (4)

• Figure 1: Workflow of Visual-Spatiotemporal Graph Neural Network for NTL and power outage prediction. First, download images from NASA Black Marble. Images are downspampled by image encoder and projected onto lower-dimensional space with a linear layer. Following, we concatenate the image embedding with a time embedding which is used as input for the st-GNN. The output of the st-GNN will be a future $T$ embedding based on $S$ past graph signals, this future embedding is projected onto higher-dimensional space and upscaled using an image decoder to obtain pixel-level predictions of NTL and power outages.
• Figure 2: Visual representation of non-adaptive adjacency matrix (graph) and NTL for Florida counties. The graph is overlayed on top of a basemap provided by Esri. The NTL data is from Black Marble annual composite of 2022.
• Figure 3: Results from evaluating on a single hurricane at a time. Testing on H-Michael (Bay County) and H-Ian (Lee County) demonstrates power outage prediction effectiveness once the power outage has happened (after landfall) and accurately predicts areas where power outages occur. Subsequently, following the hurricane, the model is able to predict the areas whose power was recovered first. The last two rows shows a failure case, H-Idalia (Levy County), where the model correctly identifies the areas with most light but also generates incorrect nightlight patterns. Note, each test is out of distribution as the model was only trained on data from the other two hurricanes.
• Figure 4: Results from our case studies, evaluating on a single hurricane at a time. Each image from Figure \ref{['nightlights']} was processed to produce a Percent of Normal map. Consistent with Figure \ref{['nightlights']}, we see for H-Michael (Bay County) the predictions are sufficient to quantify the severity and localization of power outages. For H-Ian (Lee County), severity and recovery predictions are accurate but we see higher levels of noise in the output. The bottom two rows show our failure case, H-Idalia (Levy County) exhibiting similar levels of noise to the H-Ian case. The color represents the severity of the outage, ranging from 0% (red, severe outage) to 100% normal (green, no outage).