EAGLE-I Outage Dataset Overview

• EAGLE-I Outage Dataset is a high-frequency, county-level outage reporting resource with 15-minute intervals, enabling real-time tracking of power disruptions.
• It integrates outage records with external sources like DOE OE-417 and NOAA weather data to enhance outage prediction and restoration time modeling.
• Advanced machine learning and clustering techniques applied to this dataset improve prediction accuracy and support comprehensive disaster analytics.

The EAGLE-I Outage Dataset is a high-frequency, county-level power outage reporting resource maintained by Oak Ridge National Laboratory for the U.S. Department of Energy. It provides granular, time-stamped records on the number of customers without power across U.S. counties, commonly at 15-minute intervals, and underpins a range of research in outage detection, resilience modeling, prediction, and knowledge graph construction. Its extensive coverage and temporal resolution make it a foundational asset for studies addressing both operational and analytical aspects of outage dynamics, particularly in the context of extreme weather and infrastructure interdependencies.

1. Dataset Structure and Collection

The EAGLE-I Outage Dataset consists of time series records for every county in its coverage area, most frequently sampled at 15-minute intervals. Each record contains:

• FIPS county code and name,
• High-resolution timestamp (often reported in UTC or local standard),
• Customer impact figures (number of customers without power).

Data acquisition occurs through direct utility feeds and aggregation from public and partner sources. The dataset is designed for high temporal fidelity, facilitating dynamic tracking of outage evolution and event timing at the sub-state level (Frakes et al., 30 Jul 2025).

In typical preprocessing pipelines, gaps of modest length are imputed via interpolation, and logical consistency checks are deployed to address misreporting or incomplete notification (Stanishevska, 4 Oct 2025). Synchronization with external datasets (e.g., DOE OE-417 event reports, NOAA weather records) is common for enhanced semantic richness and event attribution (Wang et al., 2021).

2. Temporal and Spatial Granularity

Temporal granularity in EAGLE-I is a key attribute: with 15-minute reporting intervals, it enables near-real-time outage monitoring, temporal clustering, and fine-grained statistical analysis of restoration processes. Spatial granularity is at the county level—a scale large enough to mask intra-county heterogeneity but highly suitable for state-level emergency response, contingency planning, and disaster mapping (Frakes et al., 30 Jul 2025).

Papers employing EAGLE-I for hurricane analysis often extract "peak" outage values within multi-day windows to correlate outage severity with exogenous influences such as critical infrastructure interdependencies (Bose et al., 13 Jul 2024). In contrast, studies on summer thunderstorms in Michigan use the dataset to capture sharp, event-driven spikes and facilitate event-centric forecasting at the hourly and day-ahead horizons (Stanishevska, 4 Oct 2025).

3. Analytical and Operational Applications

EAGLE-I has supported research in several domains:

• Outage Prediction: The dataset enables development of early-warning models (e.g., two-stage machine learning models combining logistic regression and LSTM) with routine retraining and feature-driven pipelines that focus on spatio-temporal precursors such as moisture advection, wind shifts, and convective signatures (Stanishevska, 4 Oct 2025).
• Restoration Time Modeling: Transfer learning-enhanced neural networks trained on clustered outage events utilize EAGLE-I to improve restoration time prediction, address class imbalance, and outperform global models in low-prior clusters (Wang et al., 2021).
• Resilience Assessment: By integrating EAGLE-I outage counts with infrastructure interdependency graphs (NAERM-IA), research rigorously quantifies the correlation between the spread of outages and k-hop network connectivity during hurricane events, revealing values often exceeding r = 0.6 for certain advisories (Bose et al., 13 Jul 2024).
• Knowledge Graph Construction: GeoOutageKG employs EAGLE-I records as the temporal backbone, correlating power loss timeseries with high-spatial-resolution nighttime light (NTL) satellite imagery and synthesized spatial outage maps for composite, multimodal analysis (Frakes et al., 30 Jul 2025).

4. Integration with Multimodal and Derived Datasets

Because EAGLE-I is county-aggregated, integration with datasets offering finer spatial resolution is standard. GeoOutageKG aligns each EAGLE-I OutageRecord with corresponding NTL images and derived OutageMaps using a semantic ontology (GeoOutageOnto). This integration provides a multiresolution lens for disaster analytics, uncovering spatially localized patterns within counties and supporting reasoning across disparate sources (Frakes et al., 30 Jul 2025).

For disaster mapping, differences in radiance between pre/post-event satellite images are linked to EAGLE-I outage records for validation, enabling the construction of event-specific outage severity maps. This integration enriches outage reporting by adding context from visual and geospatial data, facilitating advanced queries and cross-modal prediction (Frakes et al., 30 Jul 2025).

5. Methodological Implications and Model Adaptation

Machine learning frameworks utilizing EAGLE-I data employ dimensionality reduction (PCA), robust statistical modeling (Poisson Regression for discrete outage counts), and sequence architectures (Seq2Seq LSTM with Adam optimization) to address inherent data noise and variability (Das et al., 20 Sep 2025). These models demonstrate improved accuracy and amplitude sensitivity—particularly in peak-event windows—over simpler baselines, with reductions in cMASE (by 2–3% near peaks) and increased peak detection (F1 rises from 57.1% to 66.7% at ±48h windows) in real-world test scenarios (Stanishevska, 4 Oct 2025).

Clustering and embedding techniques (SDESC and t-SNE) support the decomposition of large, imbalanced EAGLE-I datasets into homogeneous subsets for transfer learning, ensuring high prediction fidelity across diverse outage event types (Wang et al., 2021).

6. Challenges, Limitations, and Future Prospects

Primary limitations include spatial aggregation (county-level resolution masks sub-county variation), possible biases or gaps in utility reporting, and epistemic uncertainty related to data completeness. Researchers mitigate these challenges by applying interpolation, using supplementary event verification, and bootstrapping for uncertainty quantification (Stanishevska, 4 Oct 2025).

A plausible implication is that future research will continue to broaden the dataset’s coverage, integrate additional modalities (e.g., meteorological satellite data, infrastructure mapping), expand ontology frameworks, and develop methodologies for sub-county resolution outages. There is active interest in leveraging the dataset for retrieval-augmented generation (RAG) in benchmarking LLMs and for improving real-time dashboard systems in grid operation and emergency response (Frakes et al., 30 Jul 2025).

7. Significance and Broader Impact

The EAGLE-I Outage Dataset underpins a diverse spectrum of research in outage modeling, infrastructure resilience, disaster analytics, restoration time forecasting, and knowledge graph construction. Its granular temporal structure and extensive coverage allow for detailed event-centric forecasting, high-resolution disaster mapping, and strategic planning. Further, by serving as a "ground truth" for multimodal, cross-domain analyses, it anchors complex methodologies that require interoperability with visual, spatiotemporal, and infrastructure datasets.

Overall, EAGLE-I is an essential reference for technical and operational studies, offering a robust platform for both data-driven model building and integrated analytic frameworks in the power systems and disaster resilience research communities.