View on GitHub

power-outages

Predicting U.S. power outage severity (2000–2016) using machine learning, achieving 80% accuracy with Random Forest. Submitted as Final Project for DSC 80 at UC San Diego.

Predicting the Severity of Power Outages

Introduction

This project analyzes major power outage events in the continental United States from January 2000 to July 2016, investigating which factors affect the duration and intensity of power outages, and whether we can preemptively predict and detect large-scale power outages.

The dataset contains 1,534 rows representing individual power outage events. The key columns relevant to this analysis include:

Data Cleaning and Exploratory Data Analysis

Data Cleaning

The raw dataset required significant preprocessing to make it suitable for analysis. The original Excel file contained metadata and headers that needed to be separated from the actual data.

Key cleaning steps:

  1. Column extraction: Removed metadata rows and extracted proper column names from row 4, with units of measurement from row 5
  2. Data subset selection: Retained 22 relevant columns from the original 57 columns, focusing on temporal, geographic, demographic, and outage characteristics
  3. Data type conversion: Converted OUTAGE.DURATION and CUSTOMERS.AFFECTED to numeric values, handling invalid entries as NaN
  4. Date parsing: Converted date columns to datetime objects for temporal analysis
  5. Missing value handling: Replaced ‘NA’ strings with proper NaN values, and filled missing CUSTOMERS.AFFECTED values with 0 (assuming unreported outages had minimal impact)

Feature Engineering:

Created additional derived features to enhance analysis:

Here’s the head of the cleaned DataFrame (limited to a few columns):

OBS YEAR MONTH U.S._STATE ANOMALY.LEVEL CUSTOMERS.AFFECTED OUTAGE.DURATION
1 2011 7 Minnesota -0.3 70000.0 3060
2 2014 5 Minnesota -0.1 68200.0 1
3 2010 10 Minnesota -1.5 70000.0 3000
4 2012 6 Minnesota -0.1 68200.0 2550
5 2015 7 Minnesota 1.2 250000.0 1740

Univariate Analysis

Summary Statistics:

The distribution of outage durations shows a right-skewed pattern with most outages lasting under 4,000 minutes, but with a long tail of extended outages. The median duration of 701 minutes indicates that half of all outages are resolved within a day’s time, while the mean being higher suggests some extremely long-duration events significantly impact the average.

Generally, this data seems to lack instances of small-scale outages, but does have a lot of missing/unclassified data.

Bivariate Analysis

Analysis of outage duration by cause category reveals significant differences between causes. Severe weather events tend to produce the longest outages with the highest variability, while intentional attacks typically result in shorter, more predictable durations. Equipment failures show moderate duration with less variability than weather events.

The relationship between customers affected and outage duration shows a weak positive correlation, suggesting that larger outages don’t necessarily last longer. This indicates that the scope and duration of outages are influenced by different factors - scope likely depends on grid interconnectedness and population density, while duration depends more on the nature of the damage and repair complexity.

I also plotted the seasonality to ascertain whether it made a difference in terms of outage distribution.

Interesting Aggregates

State-by-State Analysis (Top 10 by Average Customers Affected):

State Avg Customers Affected Avg Duration (min) Avg Population
Florida 282,939 4,095 18.1M
South Carolina 251,913 3,135 4.5M
Illinois 198,026 1,602 12.8M
District of Columbia 175,238 4,304 622K
Texas 165,227 2,705 25.2M

Florida leads in both average customers affected and duration, likely due to frequent severe weather events and hurricanes.

Cause Category Analysis:

Cause Category Mean Duration Median Duration Mean Customers Median Customers
Severe Weather 3,884 2,460 177,206 105,000
Fuel Supply Emergency 13,484 3,960 0.02 0
System Operability 729 215 137,941 25,000
Equipment Failure 1,817 221 50,968 0

This table reveals that severe weather causes both the most widespread outages (highest customer impact) and among the longest duration outages, making it the most significant outage category for overall grid reliability.

Assessment of Missingness

NMAR Analysis

I believe the column HURRICANE.NAMES is likely NMAR (Not Missing At Random). This column is 95.31% missing, and the missingness is directly related to the value itself - hurricane names are only recorded when the outage is actually caused by a named hurricane. The absence of a hurricane name is meaningful information indicating the outage was not hurricane-related.

To make this column MAR (Missing At Random), we would need additional data such as:

Missingness Dependency

I investigated whether the missingness of CAUSE.CATEGORY.DETAIL depends on other columns using permutation testing.

Testing CAUSE.CATEGORY dependency:

Testing OUTAGE_DAYOFWEEK dependency:

Hypothesis Testing

Research Question: Do severe weather outages last longer than equipment failure outages on average?

Null Hypothesis (H₀): The mean outage duration is the same for severe weather and equipment failure outages.

Alternative Hypothesis (H₁): Severe weather outages last longer, on average, than equipment failure outages.

Test Statistic: Difference in mean duration (Weather - Equipment)

Significance Level: α = 0.05

Results:

Conclusion: With a p-value much less than 0.05, we reject the null hypothesis. There is strong statistical evidence that severe weather outages last significantly longer than equipment failure outages on average. This makes practical sense as weather damage often affects larger areas and requires more complex repairs than localized equipment failures.

Framing a Prediction Problem

Prediction Problem: Predict whether a power outage will be “Short” (< 6 log-minutes duration) or “Long” (≥ 6 log-minutes duration) based on information available at the time the outage begins.

Problem Type: Binary Classification

Response Variable: DURATION_CLASS (derived from LOG_DURATION)

Evaluation Metric: Accuracy, with additional focus on precision and recall for both classes to ensure balanced performance.

Features Available at “Time of Prediction”:

This prediction task is valuable because early classification of outage severity can help utilities allocate appropriate resources and set realistic restoration expectations for customers and emergency services.

Baseline Model

Model Description: Decision Tree Classifier using 11 features to predict outage duration class.

Features Used:

Preprocessing:

Performance:

              precision  recall  f1-score  support
Long             0.80    0.78      0.79     170
Short            0.71    0.73      0.72     126
accuracy                           0.76     296

Assessment: The baseline model achieves 76% accuracy with reasonable precision and recall for both classes. The model performs slightly better on predicting long outages (precision 0.80) than short outages (precision 0.71), which could be valuable for emergency planning purposes. However, there’s room for improvement in the overall performance and class balance.

Final Model

Added Features and Rationale:

  1. Enhanced preprocessing for CUSTOMER_DENSITY: Applied log transformation to handle the skewed distribution of customer density ratios, which better captures the relationship between customer concentration and outage characteristics.

  2. Improved numeric preprocessing: Used StandardScaler for most numeric features to ensure equal contribution to tree-based decisions, while applying specialized log transformation for customer density.

  3. Advanced categorical handling: Maintained one-hot encoding but with improved imputation strategies.

Model Algorithm: Random Forest Classifier with hyperparameter tuning

Hyperparameter Optimization:

Best Hyperparameters:

Final Model Performance:

              precision  recall  f1-score  support
Long             0.82    0.82      0.82     170
Short            0.76    0.76      0.76     126
accuracy                           0.80     296

Improvement Over Baseline:

The Random Forest approach with proper hyperparameter tuning provides better generalization than the single Decision Tree, while the enhanced feature preprocessing better captures the underlying relationships in the data.

Fairness Analysis

Fairness Question: Does the model perform equally well for high-urbanization areas vs. low-urbanization areas?

Group Definition:

Evaluation Metric: Precision for “Long” outage predictions

Null Hypothesis (H₀): The model is fair across groups - precision scores for ‘Long’ outages are equal regardless of urbanization level.

Alternative Hypothesis (H₁): The model is unfair across groups - precision scores for ‘Long’ outages differ between high and low urbanization areas.

Test Statistic: Difference in precision scores (Group X - Group Y)

Significance Level: α = 0.05

Results:

Conclusion: With a p-value of 0.7018, which is much greater than 0.05, we fail to reject the null hypothesis. There is insufficient evidence to conclude that the model performs unfairly across different urbanization levels. The model appears to achieve fairness parity between high and low urbanization areas for predicting long-duration outages.

Conclusion

This analysis successfully developed a machine learning model capable of predicting power outage duration categories with 80% accuracy. Key findings include:

  1. Severe weather is the dominant factor in both outage frequency and severity
  2. Geographic and demographic factors significantly influence outage patterns
  3. The final Random Forest model demonstrates both good performance and fairness across different community types