Abstract
In recent years, extreme weather events frequently cause large-scale power outages, affecting millions of customers for extended duration. Resilience, the capability of withstanding, adapting to, and recovering from a large-scale disruption, has become a top priority for the power sector. However, a quantitative understanding of power grid resilience is still lacking. This study takes a data-driven approach by exploring real-world power outage data and weather records. We propose a spatio-temporal random process model that identifies key resilience capabilities and interdependence between service areas. Our analysis shows that weather only directly causes a small portion of outages; the majority result from planning vulnerabilities and the accumulation of weather effects, which then propagate through the grid.
Introduction
While "resilience" is a central theme in modern energy policy, quantifying it remains a significant challenge. Most existing research is either purely conceptual or focuses on isolated technical components, failing to capture the system-level dynamics of a complex power grid. Two major hurdles persist: the scarcity of high-resolution, real-world failure data and the inherent complexity of non-linear, cumulative weather impacts. This paper addresses these gaps by using a data-driven approach to measure four key aspects of resilience: planning vulnerability, maintenance sufficiency, criticality, and recoverability.
Conceptual Framework: The Resilience Trapezoid
To quantify resilience, the researchers developed the "Resilience Trapezoid" framework. This model conceptualizes a grid's response to an extreme event across five distinct states: prevention, degradation, degraded, restoration, and adaptation.
This framework allows for the identification of specific resilience capabilities, such as infrastructural resistance (the ability to withstand the initial hit) and operational recoverability (the speed and efficiency of restoration).
Methodology
The study proposes a spatio-temporal non-homogeneous multivariate Poisson process. To capture the non-linear relationship between weather and outages, a Deep Neural Network (DNN) is integrated to estimate the "base intensity" of outages.
A key innovation is the distinction between direct and indirect impacts. The model separates outages directly triggered by weather from "triggering effects," where a failure in one node cascades into neighboring areas. Furthermore, the model incorporates an "accumulation effect," using a discounted sum of recent weather data to account for how stress builds up on infrastructure over time before failure occurs.
Datasets
The analysis leverages high-resolution datasets covering three major events: the 2018 Nor'easters in Massachusetts, Hurricane Michael in Georgia, and Hurricane Isaias in the Carolinas.
- Outage Data: Quarter-hourly customer-level records from PowerOutage.us, encompassing over 1,100 geographical units and 9.6 million customers.
- Weather Data: High-Resolution Rapid Refresh (HRRR) data with 34 variables at a 3km resolution.
Spatio-Temporal Analysis and Key Insights
The model reveals several critical dynamics of grid failure. One of the most significant findings is the "Disruption Tolerance Capacity" (DTC). Outage ratios do not increase linearly with weather intensity; instead, they remain low until a cumulative stress threshold is reached, after which outages spike sharply in a sigmoid pattern.
As shown in the spatio-temporal maps, there is a notable "lag factor." Large-scale outages typically peak 12–36 hours after the peak of the extreme weather event. This confirms that resilience acts as a "time-buying" mechanism that eventually becomes overwhelmed by cumulative damage.
Furthermore, the study identifies "propagation hubs"—often mid-sized urban areas rather than major metropolitan centers—that serve as critical bottlenecks. Protecting these specific nodes could reduce total customer outages by approximately 45.5%. Similarly, hardening infrastructure in the top 12 most vulnerable units could reduce outages by nearly 50%.
Conclusion
This research demonstrates that massive blackouts are often the result of local failures propagating through critical hubs rather than direct weather hits alone. By identifying these hubs and quantifying the cumulative effects of weather, utilities can move beyond conceptual resilience to targeted, actionable interventions. Future work will aim to integrate physical grid topology and geographical features to further refine these predictive capabilities.
Summary updated