A Novel Framework for Detecting Disaster-Induced Power Outages Globally via Satellite Observations

Power systems are vital to our society, yet they remain vulnerable to widespread damage from natural disasters, often resulting in large-scale outages. Satellite-captured nighttime lights offer a global-scale opportunity to map daily electricity use and detect power loss. Prior research has primarily relied on aggregating nightlight data over several days during disasters to compare with recorded power outages. However, the interpretation of daily, short-term nightlight variations to determine outages remains limited, hindering efforts to use this global dataset for timely outage mapping. Leveraging unique outage data, this study introduces a novel interpretation of nightlight intensity loss from buildings during outages to propose a two-stage framework that uses nightlight to predict power outages. First, a classification model is used to identify regions where the fraction of customers experiencing outages exceeds 0.05, thereby mitigating noise and cloud interference that could hinder the identification of extreme outages. Second, a regression model is used to predict outage fractions for high-outage cases to map out actual events utilizing our developed interpretation of nightlight loss. We train and validate our model using historical blackouts during Hurricanes Irma (2017), Michael (2018), Ida (2021), and Ian (2022), each of which affected over a million customers. We show the model’s generalizability using out-of-sample data from Hurricane Fiona (2022) in Puerto Rico. We also demonstrate the transferability of the framework to other disasters by tracking outages after earthquakes in Turkey (February 2023) and Myanmar (March 2025). The framework utilizes open-source, hazard-independent inputs to predict power outages worldwide, caused by various disasters, including earthquakes or ongoing conflicts, and even in regions without automated outage reporting.