Introduction

From 2000-2016, Purdue compiled a dataset documenting major power outages across the United States. Each row corresponds to an outage instance, including details like location, duration, cause category, and the affected NERC (North American Electric Reliability Corporation) region. Using these features, I will be investigating the following:

What characteristics are most strongly associated with various causes of major power outages?

My dataset has 1534 entries, and the following 10 features will be relevant to my analysis:

• YEAR: indicates the year when the outage event occurred [2000,2016].
• MONTH: indicates the month when the outage event occurred [1,12].
• U.S._STATE: indicates the U.S. state where the outage event occurred.
• NERC.REGION: the NERC region involved in the outage event.
• OUTAGE.START.DATE: indicates the day of the year when the outage event started.
• OUTAGE.START.TIME: indicates the time of the day when the outage event started.
• OUTAGE.RESTORATION.DATE: indicates the day of the year when power was restored to all customers.
• OUTAGE.RESTORATION.TIME: indicates the time of the day when power was restored to all customers.
• CAUSE.CATEGORY: categories of all the events causing the major power outages.
• OUTAGE.DURATION: duration of an outage event (in minutes).

Data Cleaning and Exploratory Data Analysis

Data Cleaning

To clean the data, I began by converting missing values from 'NA' to 'nan' representation for consistency across missing values. Then, I worked on conbining date & time columns into single Datetime columns. So, OUTAGE.START.DATE and OUTAGE.START.TIME are combined into OUTAGE.START, and OUTAGE.RESTORATION.DATE and OUTAGE.RESTORATION.TIME combine into OUTAGE.RESTORATION. I then chose to drop the original 4 columns from the dataset.

Then, I one hot encoded cause categories. We have seven possible cause categories, so an additional seven features were created (example column name: cause==severe weather). I kept the original column so we can one hot encode it again within our pipeline, but it’s valuable for EDA and will help us impute values.

Next, I looked at the feature NERC.REGION. After some initial research, I learned there are six main NERC regions across the United States. However, collecting value counts from this feature showed that my dataset had 14 possible values. So, I investigated the region names I did not recognize, and learned that most subregions that belong to one of the six main regions. I strategically mapped these subregions to the main regions; this way, I’ll be able to consider regional trends’ impact on the cause of an outage.

Lastly, I had to handle missing values in the dataset. I began by dropping entries where both OUTAGE.START and OUTAGE.RESTORATION were missing, this only accounted for 9 entries (0.587% of the original data) so dropping these entries poses no concern for data generalization. After this, there were 49 remaining entries with missing values in OUTAGE.DURATION and OUTAGE.RESTORATION. I imputed these values with the following methods:

1. Compute mean duration for each cause category.
2. Iterate over the entire dataset, and fill missing duration values with the entry’s cause category mean value.
3. Compute restoration time using OUTAGE.START and OUTAGE.DURATION. See imputation for a visualization describing how these imputed values modify our dataset’s statistics.

These data cleaning steps leave us with a dataset like this:

   YEAR  MONTH U.S._STATE      CAUSE.CATEGORY  OUTAGE.DURATION        OUTAGE.START  OUTAGE.RESTORATION  cause==severe weather  cause==intentional attack  cause==system operability disruption  cause==equipment failure  cause==public appeal  cause==fuel supply emergency  cause==islanding NERC.REGION
0  2011    7.0  Minnesota      severe weather           3060.0 2011-07-01 17:00:00 2011-07-03 20:00:00                      1                          0                                     0                         0                     0                             0                 0         MRO
1  2014    5.0  Minnesota  intentional attack              1.0 2014-05-11 18:38:00 2014-05-11 18:39:00                      0                          1                                     0                         0                     0                             0                 0         MRO
2  2010   10.0  Minnesota      severe weather           3000.0 2010-10-26 20:00:00 2010-10-28 22:00:00                      1                          0                                     0                         0                     0                             0                 0         MRO
3  2012    6.0  Minnesota      severe weather           2550.0 2012-06-19 04:30:00 2012-06-20 23:00:00                      1                          0                                     0                         0                     0                             0                 0         MRO
4  2015    7.0  Minnesota      severe weather           1740.0 2015-07-18 02:00:00 2015-07-19 07:00:00                      1                          0                                     0                         0                     0                             0                 0         MRO

Univariate Analysis

We see here that NERC regions WECC and RFC contribute almost 2/3 of the reported outages. We should investigate regional trends further, and determine if certain causes correspond to some region.

Here, we investigate seasonal trends. Observe that the most outages occur in the summer months (Jun-Aug), with additional peaks in the winter months (Dec-Feb). This tells us that there is a definite relationship between the month of the outage and the cause of the outage, and we should create features to clearly show this relationship.

Bivariate Analysis

Embed at least one plotly plot that displays the relationship between two columns. Include a 1-2 sentence explanation about your plot, making sure to describe and interpret any trends present and how they answer your initial question. (Your notebook will likely have more visualizations than your website, and that’s fine. Feel free to embed more than one bivariate visualization in your website if you’d like, but make sure that each embedded plot is accompanied by a description.)

Here, we see that the cause category ‘Sever Weather’ is the most common category for each month, excluding March. This category dominates the visual, but I also notice other trends: the category ‘Public Appeal’ is most likely to occur in the summer months, and outages caused by Intentional Attacks are more common in the first half of the year.

Interesting Aggregates

Here, we gain a few more valuable insights. The region WECC is most likely to face outages due to Intentional Attacks, while region RFC is dominated by Severe Weather outages.

Imputation

As mentioned in the data cleaning section, I chose to impute missing values for OUTAGE.DURATION using the mean duration values from outages with the same cause category. This table describes summary statistics for OUTAGE.DURATION before and after mean imputation.

                 Before Imputation  After Imputation    Change
count            1476.00            1525.00             49.00
mean             2625.40            2693.66             68.26
std              5942.48            5931.29             -11.20
min                 0.00               0.00             0.00
25%               102.25             108.00             5.75
50%               701.00             728.87             27.87
75%              2880.00            3000.00             120.00
max            108653.00          108653.00             0.00

These results are expected: an increased sample size since we filled missing values, a slightly increased mean indicating that the imputed values are a bit larger than the original values, and a decrease in standard deviation tells us that our data is slightly less variable now.

Framing a Prediction Problem

My prediction problem is as follows:

Predict the Cause Category of A Major Power Outage

This is a multiclass classification, with the response variable CAUSE.CATEGORY. When conducting EDA in steps 1-2, I was intrigued by some values I saw in our cause category feature. I plan to use outage OUTAGE.DURATION, OUTAGE.MONTH, and NERC.REGION to predict it’s corresponding cause category. To measure performance, I will be using the f1-macro score metric. My dataset has class imbalance, and this metric will treat all classes with equal importance.

Baseline Model

For my baseline model, I chose to predict CAUSE.CATEGORY using two features:

• OUTAGE.DURATION (quantitative): the duration of the outage in minutes
• NERC.REGION (nominal): the NERC region where the outage occurred

The baseline model’s pipeline consists of:

1. OneHotEncoder for NERC.REGION feature (we drop the first category to avoid multicollinearity)
2. RandomForestClassifier with default parameters

This model achieved an f1-macro score of 0.297, which is our baseline performance. The model scores low, which is expected given limited features and the difficulty of multiclass classifications. The model clearly needs improvement, so we’ll use feature engineering and model tuning to create our final model.

Final Model

In order to improve our score from the baseline model, I made a few enhancements:

Feature Engineering

1. I used sine and cosine transformations on MONTH in hopes of capturing the seasonal trends we discovered during EDA.
2. I standardized OUTAGE.DURATION
3. Used the same one hot encoding for NERC.REGION

Model Improvements

1. I used a ColumnTransformer for my preprocessing due to the extra steps
2. I added parameters to the RandomForestClassifier (I used grid search to find these optimal values):
   • n_estimators=200 for more stable predictions
   • max_depth=12 to help with overfitting
   • class_weight=balanced to help with class imbalance

Updated Performance

My final model achieved an f1-macro score of 0.389, which representts a 31% improvement from our baseline! This model stronly captures seasonal patterns and offers better generalization with hyperparameter tuning.