Introduction

We utilize data which pertains to the major outages witnessed by different states in the continental U.S. during January 2000–July 2016. There are a total of 1534 total outages reported in the dataset and 57 columns of collected data. Further details regarding the raw data and its collection process can be found in the original published data article.

We were initially interested in exploring the question: How are different socioeconomic areas affected by power outages? (i.e. duration, severity, frequency, # of customers affected, cause, electricity price). However upon an exploration of the data we found that it would be difficult to isolate outages by socioeconomic area, this is because data is aggregated by state, which is too large of a lens through which to observe socioeconomic differences. From this we pivoted to exploring the ‘severity’ of an outage, in terms of its effect on medically vulnerable individuals. We understand that for individuals living off of devices such as ventilators or CPAP machines there is limited backup battery. Many individuals rely on refrigerated medication such as insulin and losing refrigeration power can cause these medicines to lose viability. Some elderly or heat-sensitive populations are at increased risk when HVAC systems are out. After several hours of power outage, medically vulnerable individuals may need to be transported to an area with power or to a medical center to alleviate these issues. How realistic is this possibility of relocation? Are those at risk of severe power outages also located in rural areas where the nearest medical facility may be a great distance away? Therefore we aim to explore the question: How do states with more urban or rural land area correlate with severity of power outage duration?

Table 1 describes the relevant columns we selected for our study.

Table 1. Power Outage Column Descriptions

Variable Name	Description
OUTAGE EVENTS INFORMATION
CAUSE.CATEGORY	Categories of all the events causing the major power outages
OUTAGE.DURATION	Duration of outage events (in minutes)
CUSTOMERS.AFFECTED	Number of customers affected by the power outage event
REGIONAL LAND-USE CHARACTERISTICS
AREAPCT_URBAN	Percentage of the land area of the U.S. state represented by the land area of the urban areas (in %)
AREAPCT_UC	Percentage of the land area of the U.S. state represented by the land area of the urban clusters (in %)
REGIONAL CLIMATE INFORMATION
CLIMATE.REGION	U.S. Climate regions as specified by National Centers for Environmental Information (nine climatically consistent regions in continental U.S.A.)

Note: “NA” in the data file indicates that data was not available.

Data Cleaning and Exploratory Data Analysis

After extracting it from the downloaded excel, we store and manipulate the data within a Pandas dataframe. From the master DataFrame we construct a new dataframe where we select our relevant columns and remove rows with null values. Data conversions are made to ensure numerical data can be used appropriately in subsequent steps. After cleaning our resulting DataFrame has 1051 outages for us to work with. We have created our own column AREAPCT_RURAL which is assumed to make up the remaining percentage of land area within the state. Here’s a few rows from our cleaned DataFrame:

POSTAL.CODE	OUTAGE.DURATION	AREAPCT_URBAN	AREAPCT_UC	CAUSE.CATEGORY	CUSTOMERS.AFFECTED	CLIMATE.REGION	AREAPCT_RURAL
MN	3060	2.14	0.6	severe weather	70000	East North Central	97.26
MN	3000	2.14	0.6	severe weather	70000	East North Central	97.26
MN	2550	2.14	0.6	severe weather	68200	East North Central	97.26
MN	1740	2.14	0.6	severe weather	250000	East North Central	97.26
MN	1860	2.14	0.6	severe weather	60000	East North Central	97.26

Univariate Analysis

We begin by investigating the distribution of the OUTAGE.DURATION column.

This distribution depiction isn’t very helpful. The data is clearly right-skewed but the scale of the x-axis bins are so large it is difficult to derive real meaning for our exploration of medical vulnerability. Let’s try to frame this on a more relevant scale. We have created a new categorical variable, OUTAGE.DURATION.DESCRIPTION which places the values of OUTAGE.DURATION in the following bins, these are our severity categories:

Short-term (0–2 hours): Minimal risk for most

Moderate (2–4 hours): Begins affecting medical equipment

Critical (4–8 hours): Medication spoilage risk, food safety, health risks escalate

Severe (8+ hours): Systemic risk to vulnerable populations, economic losses

Over half of our data is categorized as having a medically severe duration, which demonstrates a dangerous trend for those with higher medical risk. To learn more about these vulnerabilities, we need to understand the geographic landscape in which these severe outages occur.

Bivariate Analysis

Now we explore the OUTAGE.DURATION.DESCRIPTION relative to the AREAPCT_... columns. Recall from our descriptions in Table 1 above that AREAPCT_... is the percentage of land area within each state that is classified as being Urban, Urban Cluster, or Rural.

We can extrapolate from these two graphs that almost all states have a greater proportion of rural areas compared to urban or urban clusters. Note: The one ‘state’ with 100% urban land area (and correspondingly 0% rural is D.C.)! We can see that the more concentrated categories, Moderate and Critical, are likely occurring in a smaller sample of states. Minimal outages however probably occur across all 50 states, regardless of the mix of urban vs rural land area in the state. Interestingly, Severe outages appear to be more concentrated than Minimal, despite the fact that Severe outages make up an overwhelming majority of our dataset. Let’s look at a breakdown of these categories by state to learn more:

Interesting Aggregates

Our extrapolations from above are correct! There are quite a few 0s in the Moderate and Critical rows. The Severe row definitely appears to have the most concentration happening within a few states, namely: Michigan, Texas, California, and Pennsylvania.

Imputation

We did not perform any missing value imputation. A majority of our dropped rows occurred due to null values in the CUSTOMERS.AFFECTED and CAUSE.CATEGORY columns. We felt that it would not be appropriate to imputate these columns as they are truly unique to each outage.

Framing a Prediction Problem

We seek to make a multiclass classification prediction problem. We are predicting on the column we have created earlier: OUTAGE.DURATION.DESCRIPTION. We can safely assume that all of the possible features we have pulled (in Table 1) are available at the beginning of an outage and therefore can be used to predict duration.

We chose F1 Weighted Average as a way to evaluate our model for this multi-class classification problem. Weighted average calculates the average of the per-class scores, but it weights each class’s contribution by the number of samples in that class. Our target variable (OUTAGE.DURATION.DESCRIPTION) is multi-class and imbalanced—some outage duration categories are far more frequent than others. A macro average would treat all classes equally, regardless of size, potentially giving too much influence to rare categories. In contrast, weighted average accounts for class frequency, giving more influence to the performance on more common classes. While it still includes all classes, weighted average provides a middle ground between favoring dominant classes (as micro average might) and overemphasizing rare ones (as macro average might). This ensures our model’s precision isn’t unfairly penalized for not predicting rare events.

Baseline Model

We decided to use a decision tree classifier for our baseline model. We used the features AREAPCT_URBAN, AREAPCT_RURAL (quantitative variables representing the percentage of area that is urban and rural) in order to find our target OUTAGE.DURATION.DESCRIPTION, which describes the severity of power outages (Critical, Minimal, Moderate, Severe). here were:

In short, we had:

• 2 quantitative features: AREAPCT_URBAN, AREAPCT_RURAL

• 0 ordinal features

• 0 nominal (categorical) features

Since both features are numeric and continuous, we applied standard scaling using StandardScaler to normalize their values before feeding them into the classifier. No encoding was needed since there were not any categorical variables. We split the data into training and test sets using an 80/20 ratio, and trained the decision tree with default hyperparameters (except setting random_state=42 for reproducibility). The pipeline had a preprocessing step as well. The model’s performance in terms of our chosen metric: F1 Weighted Average=61%.

Overall, while the weighted average may seem moderate, the model is not yet good because it suffers from significant class imbalance and poor performance on minority classes. This indicates the model is biased toward the dominant class (Severe), and may not generalize well in more balanced or real-world situations.

Final Model

To enhance the predictive performance of the model, we adjusted and added several features.

• CUSTOMERS.AFFECTED: We assumed that more customers affected would lead to a longer outage. We performed an exploratory analysis to test our hypothesis with the figure below. We found that this plot greatly resembles our original distribution of OUTAGE.DURATION.DESCRIPTION counts. Therefore we anticipate CUSTOMERS.AFFECTED to be a strong predictor for our model.

• CAUSE.CATEGORY: We observed earlier an interesting concentration of Severe outages within a few states, namely: Michigan, Texas, California, and Pennsylvania. These states have significantly different climates and we therefore chose to explore the overall distribution of OUTAGE.DURATION.DESCRIPTION versus CAUSE.CATEGORY to see if causes other than climate might be affecting these states. After observing the figure below we see that CAUSE.CATEGORY is overwhelmingly skewed toward severe weather events. Based on this we hypothesize that CAUSE.CATEGORY might be used in tandem with CLIMATE.REGION (described below).CAUSE.CATEGORY was included using one-hot encoding to retain categorical distinctions without creating ordinal bias.

• CLIMATE.REGION: Following up with our exploration of CAUSE.CATEGORY we decided to investigate CLIMATE.REGION. Different climate zones (e.g., Southeast vs. Northwest) are subject to distinct weather events—like hurricanes, snowstorms, or droughts—that can significantly impact both the frequency and severity of outages. Based on the variation in the figure below we hypothesized that CLIMATE.REGION would make a more generalized feature than POSTAL.CODE. The CLIMATE.REGION column was processed using OneHotEncoder within the pipeline, enabling the model to leverage it without introducing ordinal bias, while maintaining compatibility with cross-validation.

Numerical Feature Transformations:

• FunctionTransfomer: log_customers: We transformed the numeric variables using a log transformation (log1p). This stabilizes variance and reduces the skewness from extreme outliers (e.g., major outages affecting thousands of customers), enabling the model to better differentiate between typical and large-scale events without being overwhelmed.

• StandardScaler: We applied standardization to our numerical features using StandardScaler within our pipeline. This step ensures that all numeric input variables are transformed to have a mean of 0 and a standard deviation of 1, enabling the model to treat all features on a comparable scale.

Final Model Structure

We selected a Random Forest Classifier for its robustness to overfitting, ability to handle both numerical and categorical data, and strong performance on imbalanced classification problems when paired with class weighting. To fine-tune our model, we used GridSearchCV to explore different combinations of hyperparameters for the RandomForestClassifier. The grid search used 5-fold cross-validation to evaluate the performance of each combination. The varied parameters and resulting best parameter are found in Table 2 below.

Table 2. Hyperparameter Options for GridSearch and Resulting Best Parameter

Parameter	Description	Values	Best
n_estimators	Number of trees in the forest	50, 100, 200	100
max_depth	Maximum depth of the trees	5, 10, 20, None (unlimited)	10
min_samples_split	Minimum samples required to split a node	2, 5, 10	2

This configuration offered a good balance between model complexity and generalization. A moderate depth (max_depth = 10) helped prevent overfitting, while using 100 trees ensured stability in predictions without making the model too slow or heavy.

Compared to the Baseline Model (a simple Decision Tree), the tuned Random Forest showed notable improvements in the F1 Weighted Average, increasing from 61% to 72%. These results highlight better generalization across all duration types, not just the dominant Severe class. Overall results can be viewed below, note that the Critical group which were least represented in the overall data received zero classifications.

Conclusion

Our exploration of this dataset shows incredible risk for medically-vulnerable communities during a power-outage. We hope that infrastructure changes and more proficient communication (informed by more accurate predictive modes) to medically-vulnerable individuals can help to increase action in the event of a power outage, ensuring that individual’s are not forced to struggle through extended outages.

Acknowledgements

This project report is part of EECS 398-003 Practical Data Science portfolio homework by Professor Suraj Rampure at the University of Michigan College of Engineering.

1. Kosari S, Walker EJ, Anderson C, Peterson GM, Naunton M, Castillo Martinez E, Garg S, Thomas J. Power outages and refrigerated medicines: The need for better guidelines, awareness and planning. J Clin Pharm Ther. 2018 Oct;43(5):737-739. doi: 10.1111/jcpt.12716. Epub 2018 Jun 13. PMID: 29900564.