Understanding Power Outages Across the U.S.

Name: Janice Rincon

Website Link: https://rinconjm.github.io/power-outage-analysis/

Introduction

In this project I examined a dataset of major power outages across the continental United States from January 2000 to July 2016. This comprehensive dataset includes 1,534 records detailing major outages, which are defined by the U.S. Department of Energy to be outages affecting at least 50,000 customers or resulting in a loss of 300 megawatts or greater. The dataset can be accessed from Purdue University’s Laboratory for Advancing Sustainable Critical Infrastructure.

Included in the dataset is information regarding the geographical location of the outages, regional climatic information, land-use characteristics, electricity consumption patterns and economic characteristics of the states affected by the outages.

The central question I am exploring is: what are are the characteristics and causes of major power outages? I aim to use data analysis techniques to examine temporal patterns in outage events, as well as patterns in demographic, and environmental factors such as climate region and anomaly levels. Building on this analysis, I will develop a predictive model to estimate the length(duration) of major power outage given a selection of main characteristics.

This analysis is important because it can help inform policy decisions and climate resilience plans by anticipating the severity of outages under regional conditions.

Introduction of Columns

The original dataset includes 1534 rows and 57 columns. For the sake of my analysis, I will focus only a subset of these. Below are the chosen columns and their descriptions:

Column Description
YEAR The year when a power outage occurred
MONTH The month when a power outage occurred
U.S._STATE The U.S. state where the outage took place
NERC.REGION Indicates the North American Electric Reliability Corporation (NERC) region associated with the outage event
CLIMATE.REGION U.S. climate regions, as defined by the National Centers for Environmental Information, where the power outage took place
ANOMALY.LEVEL Indicates the oceanic El Niño/La Niña Index (ONI) value, representing sea surface temperature (SST) anomalies in the Niño 3.4 region
CLIMATE.CATEGORY Categorization of climate episodes based on the corresponding year and ONI value: “Warm” (El Niño), “Cold” (La Niña), or “Normal”
OUTAGE.START.DATE Indicates the day of the year when the outage event began (as reported by the corresponding Utility in the region)
OUTAGE.START.TIME Indicates the time of day when the outage event began (as reported by the corresponding Utility in the region)
OUTAGE.RESTORATION.DATE Indicates the day of the year when power was restored to all affected customers (as reported by the corresponding Utility in the region)
OUTAGE.RESTORATION.TIME Indicates the time of the day when power was restored to all affected customers (as reported by the corresponding Utility in the region)
CAUSE.CATEGORY Categories of all primary causes of the power outage
OUTAGE.DURATION The duration of the outage event (in minutes)
DEMAND.LOSS.MW Amount of peak demand or total demand lost during an outage event (in Megawatt)
CUSTOMERS.AFFECTED The number of customers affected by the power outage
TOTAL.CUSTOMERS Annual number of total customers served in the U.S. state

Data Cleaning and EDA

Data Cleaning

The following steps were performed to clean the data:

  1. Resolved Formatting Issues:
    • During the conversion from .xlsx to .csv, the file contained several formatting issues, specifically, the first few rows included NaN values, and the actual column headers were located on the fifth row. These rows were removed and the correct headers were reassigned to the dataset.
  2. Keep Relevant Columns:
    • The columns that were not relevant to my research question where dropped from the data frame, keeping only the following columns of interest: YEAR, MONTH, U.S._STATE, NERC.REGION, CLIMATE.REGION, ANOMALY.LEVEL, CLIMATE.CATEGORY, OUTAGE.START.DATE, OUTAGE.START.TIME, OUTAGE.RESTORATION.DATE, OUTAGE.RESTORATION.TIME, CAUSE.CATEGORY, OUTAGE.DURATION, DEMAND.LOSS.MW, CUSTOMERS.AFFECTED, TOTAL.CUSTOMERS.
  3. Combinations:
    • The columns OUTAGE.START.DATE, OUTAGE.START.TIME, OUTAGE.RESTORATION.DATE and OUTAGE.RESTORATION.TIME were merged into two columns: OUTAGE.START and OUTAGE.RESTORATION, respectively. Both columns are of pd.Timestamp format.
    • The original columns were dropped from the data frame after merging
  4. Type Conversions:
    • Several quantitative columns were originally stored as strings. These were converted to the appropriate numeric type to be compatible with analysis and future modeling.
  5. Feature Engineering:
    • A new column, OUTAGE.DURATION.HRS, was created by converting outage duration from minutes to hours for easier interpretation.

The cleaned DataFrame contains 1,534 rows and 15 columns. Below are the first few rows of the processed dataset:

YEAR MONTH U.S._STATE NERC.REGION CLIMATE.REGION ANOMALY.LEVEL CLIMATE.CATEGORY CAUSE.CATEGORY OUTAGE.START OUTAGE.RESTORATION OUTAGE.DURATION OUTAGE.DURATION.HRS DEMAND.LOSS.MW CUSTOMERS.AFFECTED TOTAL.CUSTOMERS HOUR HOUR_LABEL
2011 7 Minnesota MRO East North Central -0.3 normal severe weather 2011-07-01 17:00:00 2011-07-03 20:00:00 3060 51 70000 2595696 17 5 PM
2014 5 Minnesota MRO East North Central -0.1 normal intentional attack 2014-05-11 18:38:00 2014-05-11 18:39:00 1 0.0166667 2640737 18 6 PM
2010 10 Minnesota MRO East North Central -1.5 cold severe weather 2010-10-26 20:00:00 2010-10-28 22:00:00 3000 50 70000 2586905 20 8 PM
2012 6 Minnesota MRO East North Central -0.1 normal severe weather 2012-06-19 04:30:00 2012-06-20 23:00:00 2550 42.5 68200 2606813 4 4 AM
2015 7 Minnesota MRO East North Central 1.2 warm severe weather 2015-07-18 02:00:00 2015-07-19 07:00:00 1740 29 250 250000 2673531 2 2 AM

Univariate Analysis

Distribution of Severity

The histogram shows that most power outages lasted less than 100 hours, with a very steep drop-off in frequency as the duration increases. The distribution is heavily skewed right which indicates the presence of extreme values, with some outages lasting over 1,000 hours. This suggests that while most power outages are relatively short, the average and variance of overall duration is impacted by a few severe events.

Temporal Patterns

This bar chart shows that major power outages were most frequent during the summer months, peaking in June and July, with over 180 incidents each. This suggests a seasonal trend, most likely influenced by the increased demand in electricity and weather events (heatwaves, storms, etc.) In contrast, November and September had the fewest outages, with under 100 incidents. Overall, outages were most common between June and August which suggests that summer presents a greater risk to infrastructure.

Interestingly enough, when visualizing the distribution of outages based on the time of day, it was found that the majority of outages occurred between 1pm and 4pm, peaking around 3pm. Morning to early evening (7am-6pm) consistently shows high outage frequency, indicating that most outages occur during active business hours, not overnight.

Geographic Patterns

Outage Count by State Avg. Outage Duration by State

The map on the left, Outage by State shows that California experienced the highest number of outages. This is likely due to factors such as wildfire risk and population size. Texas, along with parts of the Midwest and the Northeast show a higher outage frequency. However, the South as well as states located in the center of the country reported relatively few major outages, potentially due to lower population density.

The map on the right, Average Outage Duration by State, shows that states in the Midwest and Northeast, particularly Michigan and West Virginia, show the longest average durations. Interestingly enough, California shows only moderate average durations, despite having the most reported outages.

Cause Analysis

This chart shows that severe weather was the leading cause of major power outages in the U.S. between 2000 and 2016, accounting for nearly 800 incidents. The second most common cause was intentional attacks, with over 400 incidents. All other causes occurred significantly less, with less than 130 incidents each.

Bivariate Analysis

The line chart shows how major power outages in the U.S. varied by cause from 2000 to 2016. Severe weather was the most dominant and consistent cause, continuously increasing from 2003 and 2012, and peaking around 2011. Intentional attacks spiked in 2011, briefly surpassing all other causes before declining, possibly due to isolated high-impact incidents. Before 2010, however, it showed a quantity of 0 perhaps suggesting a lack of recording.

Other causes (such as equipment failure, fuel supply emergencies, and system operability disruptions) were relatively low and stable. Overall, the chart highlights that natural and intentional disruptions were the primary causes of major outages.

The scatter plot visualizes the relationship between power outage duration (in hours) and the number of customers affected. Most outages last less than 500 hours, and the majority of customers affected is 500k. Interestingly enough, there does not appear to be a clear linear correlation between the two variables, suggesting that longer outages do not necessarily affect more customers.

Interesting Aggregates

The table below shows the average outage duration by hour of day the outage occurred and cause category. Severe Weather appears to have the most even distribution, with no severe spikes. In contrast, there appear to be large values for equipment failure and fuel supply emergencies at odd hours. This suggests that the cause of an outage may influence not just how long it lasts, but when it tends to occur.

The first few rows of the aggregate is shown below:

HOUR equipment failure fuel supply emergency intentional attack islanding public appeal severe weather system operability disruption
0 14.75 1810.88 4.5 3.25 nan 71.99 53
1 nan 0.02 12.07 17.25 nan 102.05 19.62
2 nan nan 21.67 3.22 nan 47.81 5.59
3 6.66 nan 6.34 0.58 nan 58.28 15.52
4 nan 72 14.59 0.83 nan 80.5 13.6
5 11.95 48 17.08 0.25 nan 62.98 28.28
6 445.28 nan 3.84 nan 78 106.3 9.13

Assessment of Missingness

NMAR Analysis

Of all the columns that have missing data, the column most likely to be NMAR is OUTAGE.LOST.MW. When demand loss is very low, utilities may consider it negligible and choose to not report it. On the other hand, when demand loss is very high, utilities may choose not to report to as to not damage their reputation or for political reasons.

Missingness Dependency

In order to test missingness dependency, I will focus on the column of DEMAND.LOSS.MW which has a significant/non-trivial missingness. This will be tested against the columns CAUSE.CATEGORY and HOUR.

Dependency on CAUSE.CATEGORY

The distribution of CAUSE.CATEGORY is plotted when DEMAND.LOSS.MW is missing and not missing.

The observed TVD was 0.18, with a p-value of 0.0. Given this result, we reject the null hypothesis in favor of the alternative. This suggests that the mechanism behind missing values in DEMAND.LOSS.MW is not random, and is likely dependent on CAUSE.CATEGORY.

The empirical distribution of TVD values across 500 permutations is shown below:

Dependency on HOUR

The distribution of HOUR is plotted when DEMAND.LOSS.MW is missing and not missing.

The observed TVD was 0.1, with a p-value of 0.302. Given this result, we do not reject the null hypothesis in favor of the alternative. This suggests that the distribution of HOUR is not significantly different when DEMAND.LOSS.MW is missing/not missing. In other words, missingness in DEMAND.LOSS.MW does not appear to depend on the time of day when the outage occurred.

The empirical distribution of TVD values across 500 permutations is shown below:

Hypothesis Testing

To better understand how the cause of an outage relates to its severity, we compare the average duration of power outages caused by severe weather versus those caused by equipment failure. This comparison helps identify whether weather-related events tend to be more disruptive in terms of recovery time.

Results

A permutation test with 10,000 simulations was conducted in order to generate an empirical distribution of the group mean differences under the null hypothesis. The resulting p-value was 0.0001, which indicates strong evidence against the null hypothesis. At a significance level of α = 0.05, we reject the null hypothesis. There is a statistically significant difference in average outage duration between severe weather and equipment failure outages. The evidence suggests that on average, power outages caused by severe weather tend to last longer than those caused by equipment failure.

Problem Identification

My model will aims to predict the duration of a major power outage, making this a regression problem rather than a classification task. The target variable, OUTAGE.DURATION.HRS is continuous and represents the total number of hours an outage lasted. To evaluate this model, I use both Root Mean Squared Error(RMSE) and R^2 score. RMSE measures the average prediction errors in hours, while R-squared measures how well the model explains the variance in outage duration. At prediction time, we assume access to information such as the state, climate region, NERC region, anomaly level, month, hour, total customers affected, and cause category. These features provide a picture of the conditions surrounding each outage.

Baseline Model

The base regression model is a Random Forest Reggressor which uses the features CLIMATE.REGION, CAUSE.CATEGORY, and MONTH in order to predict the length of a power outage.

To prepare the data for modeling, One Hot Encoding was applied to the nominal features (CLIMATE.REGION and CAUSE.CATEGORY) to convert them into machine-readable format.

The model achieved an R-squared score of 0.21 and a RMSE of 116.35 hours. This means the model explains about 21% of the variance in outage duration, with an average error of over 115 hours.

While the model does capture some relationship between the features and the target, the low R-squared score and high RMSE shows that it is not very effective. There is significant room for improvement! Some features not captured in the current set, such as population, time of day, may provide a more comprehensive “image” of pre-outage zconditions, resulting in better model performance.

Final Model

For the final model, HOUR, CLIMATE.CATEGORY, TOTAL.CUSTOMERS, were added as features.

GridCVSearch was performed and the best hyperparameters were found to be:

With these, the model was able to improve performance and achieve an R-squared score of 0.301 and a RMSE of 109.17 hours. Compared to the baseline model, R-squared improved by approximately 43.33% and RMSE lowered by 6.17%.

Fairness Analysis

In order to the final model for fairness, two groups were created based on the year recorded for the power outage: 2000-2009 and 2010-2016. “Year” was selected as a metric to create the groups because it was not included as a feature for the model and because it can test if the model generalizes equally well across time. The year 2010 was chosen specifically as there was a spike in the frequency of reports in 2011 as well as the fact that it represents a new decade. The metric used to evaluate fairness will be RMSE.

A permutation test with 10,000 repetitions was performed. The p-value for this test was found to be 0.8981, significantly higher than the conventional significance level of α = 0.05. Therefore, we fail to reject the null hypothesis which suggests that the model performs equally for both groups.