Extreme Weather & Major U.S. Power Outages

By Jack Q. Hu (jackqhu@umich.edu)

Website: jackqhu.github.io/power-outage-analysis

Introduction

This project examines comprehensive data, used in the article entitled “A Multi-Hazard Approach to Assess Severe Weather-Induced Major Power Outage Risks in the U.S.” (Mukherjee et al., 2018), on large-scale power outage events in continental U.S. from Jan 2000 to July 2016.

The data was compiled from multiple authoritative sources, including:

• U.S. Department of Energy (DOE)
• National Oceanic and Atmospheric Administration (NOAA)
• U.S. Department of Labor’s Bureau of Labor Statistics
• U.S. Census Bureau.

As defined by the DOE, a “major outage” is classified as one that either:

1. Impacted at least 50,000 customers, or
2. Caused an unplanned firm load loss of atleast 300MW

The scope of power outage dataset used for analysis includes information on outage:

1. Geographic locations
2. Duration of blackout
3. Climate patterns & anomolies
4. Land & environment characteristics
5. Energy consumption patterns
6. Demographic patterns
7. Socioeconomic characteristics

My analysis is focused on answering the question: What are the key climate and geographical factors that contribute to severe weather-induced power outages?

Goal of Project: To learn weather patterns & their relationships with grid reliability/vulnerability, and build predictive models on risk of blackouts.

This is done by identifying key climate and geographical factors contributing to severe weather-induced outages to help inform infrastructure planning and risk management.

This question is consequential as weather-related disruptions continue to rise due to climate change.

Electrical grid disruptions and blackouts are catastrophic, and severely impact vulnerable populations such as those residing in:

1. Food deserts
2. Low-income communities
3. Regions with hot summers and/or cold winters.

Understanding the weather-related blackout relationship can help energy companies better prepare and manage risks, inform public policy decisions on infrastructure, and help analyze vulnerability patterns across different states & contribute to higher accuracy local-level risk maps.

Preprocessing and Data Cleaning

1. Duration standardization:
   • Converted outage durations from minutes to hours for better interpretability.
   • Replaced zero values in duration and impact metrics (customers affected, demand loss) with NaN as they likely represent missing data
2. Feature engineering:
   • Combined multiple urban-related metrics into a single normalized factor
   • Standardized datetime fields to preserve consistency in temporal analysis

Exploratory Data Analysis

Severe Weather is Leading Cause of Outages

Cause	Count	Proportion
severe weather	763	49.7
intentional attack	418	27.2
system operability disruption	127	8.3
public appeal	69	4.5
equipment failure	60	3.9
fuel supply emergency	51	3.3
islanding	46	3.0

Geographic Distribution Heatmaps

States most affected by outages

U.S._STATE	Total Outages
California	198
Texas	122
Michigan	95
Washington	89
New York	70

Other Findings

Summary Statistics for Weather-Related Outages:

• Total number of weather-related outages: 763
• Average outage duration: 64.73 hours
• Median outage duration: 41.00 hours
• Average customers affected: 188,575

Univariate Analysis:

• The distribution of outage durations shows a right-skewed pattern, with most outages lasting less than 24 hours Severe weather is the predominant cause of major power outages, accounting for a significant portion of all events

Bivariate Analysis:

• There appears to be a positive correlation between climate anomaly levels and outage duration Different climate regions show varying susceptibility to weather-related outages, with some regions experiencing significantly more events

Interesting Aggregates:

• Climate regions with more extreme weather patterns tend to have both more frequent and longer-duration outages The relationship between urbanization and outage frequency/duration suggests that urban areas may be more resilient to weather-related outages

Missing Data / Imputation

• Some outage durations and customer impact numbers are missing, but the pattern of missingness appears to be random rather than systematic Climate and geographical data is generally complete, making it reliable for analysis

Predictive Problem Framing

Prediction Task Our supervised ML model will focus on forecasting durations of weather-related power outages (caused severe weather events).

This is framed as a regression problem (rather than classification), as we are predicting a continuous numerical value (outage duration) rather than categorizing into discrete groups.

Response Variable:

• Outage duration (in mins, will convert to hours in results section)

Prediction-Time Available Features:

• Climate characteristics (region, anomaly level)
• Geographic info (state, NERC region)
• Pop/urban characteristics
• Economic/infrastructure metrics
• Temporal features (year, month)

Evaluation Metric: Root Mean Square Error (RMSE) as loss function for the model.

• Large penalty for large training errors; reduces model’s underestimation of outage durations for inference
• Same unit as our prediction/target variable (duration in minutes); preserves interpretability

Baseline Supervised-Learning Regression Model

Our simple model is based on a pipeline with simple linear regression, one-hot-encodng, single imputation, and minimal numerical + categorical features:

• Climate region
• Anomaly level
• Year
• Urban factor

Performance metrics:

• RMSE: 100.3 hours
• R² Score: -0.055

The negative R² score indicates this simple model performs worse than predicting the mean duration, suggesting the need for a more sophisticated approach.

Grid-Search CV Supervised-Learning Regression Model

This predictor model incorporates additional features and uses a Random Forest Regressor with the following improvements:

1. Additional Features:
   • Monthly seasonality
   • Economic indicators (total price, sales)
   • Infrastructure metrics (total customers)
   • Geographic granularity
2. Model Enhancements:
   • Quantile transformation of numeric features
   • Hyperparameter tuning via grid search
   • Feature engineering for seasonality

The enhanced model achieved:

• RMSE: 72.6 hours
• R² Score: 0.31

This significant improvement over the baseline suggests that incorporating additional features and using a more sophisticated algorithm better captures the complex relationships in power outage patterns. The model performs particularly well in predicting outages caused by severe weather events, which represent the majority of cases in our dataset.