Nighttime Lights Data as an Indicator of Electricity Outage Vulnerability: Case Study of Winter Storm Uri

Abstract Power outage data aggregated at a specific administrative level, such as census tract or census block, is not publicly available. This creates barriers to understanding spatial distribution of outage vulnerability, resulting in challenges in outage preparedness and disaster response. We work to better understand spatial power outage vulnerability by deriving simulated values through implementation of maximum entropy models, which represent likelihood of outage occurrence at the census tract level in Houston, TX. We develop a model utilizing nighttime light data to produce power outage likelihood values at the census block level. Nighttime lights outages during Winter Storm Uri (February 11–18, 2021) are used as inputs in the MaxEnt machine learning model. Energy transmission, generation, hospitals, emergency services and tree coverage data are used as predictors within MaxEnt. Results show that the model performs relatively well with a mean area under the curve of 0.758 (a common threshold for model evaluation is 0.70). Power line density, tree coverage and proximity to schools are the most influential variables in power outage vulnerability (contributing percentages are 73.0%, 7.1% and 5.2%, respectively). Utilizing MaxEnt prediction, we generate likelihood of outage occurrence values between 0 and 1 for each census block. Our work provides a novel methodology for nighttime lights processing and new applications for MaxEnt. These results provide insight into which census tracts are the most vulnerable to power outage during extreme weather events.