Power Outage Severity Model

Created By William Czech (wczech@umich.edu)

Introduction

The dataset I chose contains information about power outages that occurred in the U.S from 2000-2016, and has 1534 rows which each represent a unique power outage. Each outage contains data on the duration of the outage, the cause of the outage, its location in the U.S, the date of the outage, and economic and population data about the location where the outage occurred. The question I am seeking to answer with this dataset is what features of a power outage contribute to its severity?

To measure the severity of an outage, I will use the duration of the outage using the OUTAGE.DURATION column, and will be only looking at cases involving severe weather and not targeted attacks or random equipment failures, as these cases are more difficult to predict. The features I will use to predict the duration of an outage are the region where the outage occurred (CLIMATE.REGION), details on what severe weather cause the outage (CAUSE.CATEGORY.DETAIL), the total number of customers in the state (TOTAL.CUSTOMERS) and the percentage of land in the state where the outage occured that is classified as an urban area (AREAPCT_URBAN).

The model I will build from this analysis will allow for better predictions on where more severe power outages in the U.S may occur, and could potentially help urban planners better protect areas of the U.S at greater risk for these outages.

Data Cleaning and Exploration

To clean my dataset, I merged two columns containing the start date and start time into a single timestamp, and completed the same process for the ending time of the outage. To make the duration of the outage a more simple unit, I converted the time of each outage from minutes to hours. Since my analysis is only focused on severe weather cases, I dropped all rows in my dataset that did not have severe weather as the cause of the outage. Additionally, I dropped rows that contained missing values with columns I was to perform analysis on with my model, and also dropped all values in the CAUSE.CATEGORY.DETAIL column that had less than 4 occurences to speed up the training of my model.

After this cleaning steps, this is the first few rows of my dataframe for the columns that I will analyze.

cause_detail	climate_region	area_urban	total_customers	outage_duration
heavy wind	East North Central	2.14	2.5869e+06	50
thunderstorm	East North Central	2.14	2.60681e+06	42.5
winter storm	East North Central	2.14	2.5869e+06	31
tornadoes	East North Central	2.14	2.5869e+06	49.5
thunderstorm	East North Central	2.14	2.47491e+06	66

Plots and Analysis

After cleaning my data, I developed some plots to get a better understanding of the relationships in the dataset.

The plot above shows that most outages caused by severe weather lasted between 12 and 86 hours, with the longest outage lasting over 800 hours. We are determining the severity of an outage based on the outage duration, so this plot depicts the distribution of power outage severity.

The histogram above shows that outages caused by mainly by thunderstorms, hurricanes, and winter storms. If we can estimate the severity of outages from these weather occurences, we can build an effective model to predict the severity of outages.

This histogram is interesting because it depicts that specific outage causes may lead to longer (more severe) outages. Flooding, hurricanes, and wildfires had the longest outages on average, likely due to the severity of these natural disasters.

This histograms shows that the Southwest region has a much longer average power outage duration compared to other U.S regions with 288 hours. For other regions, the distribution is more uniform with a range of 35 - 100 hours on average, with an outlier in the West North Central region. This plot illustrates which regions have the most severe outages on average, and indicates that the Southwest, South, and Northwest have the highest severity power outages.

Data Aggregations

start_month	flooding	heatwave	heavy wind	hurricanes	snow/ice	storm	thunderstorm	tornadoes	wildfire	wind storm	wind/rain	winter	winter storm
April	0	0	6	0	0	1	9	1	0	0	1	0	5
August	0	2	0	13	0	4	17	0	0	0	1	0	0
December	0	0	12	0	6	2	1	1	3	0	1	0	16
February	0	0	9	0	7	2	2	0	0	0	0	4	30
January	0	0	4	0	0	7	0	0	0	2	0	18	34
July	0	5	0	3	0	7	42	1	8	1	0	0	0
June	2	0	2	0	0	8	63	1	2	0	2	0	0
March	1	0	3	0	0	3	2	0	0	2	5	1	4
May	1	1	1	0	0	4	19	2	2	0	0	0	0
November	0	0	10	0	0	0	9	1	2	1	0	0	7
October	0	0	9	20	0	2	3	0	6	0	2	0	4
September	0	2	4	37	0	1	8	1	0	0	0	0	0

This pivot table is significant because it depicts what storms are the most common during specific months of the year. If we know what storms cause the most severe outages, knowing what months they are most likely to occur gives us an idea of when the most severe outages occur.

Imputed Values

For this dataset, I chose not to impute any values because I dropped rows for categorical columns with missing values instead of adding an impution technique. After these rows with missing categorical values were dropped, there were no more rows with missing numerical values, so there was nothing left to impute.

Framing a Prediction Problem

The problem we want to solve with a model is: How can we predict the duration of a power outage using information on climate, outage cause, urban population statistics, and the number of customers potentially affected? From this model, we can consider outages we predict to last longer as more severe. This problem will be addressed using a regression model, as the duration of an outage is a numerical value.

At the time of predicting this model, we would have access to the information described in the problem question because this data is all static information that is not affected by the result of the power outage. Therefore, using it to predict the duration of the outage is valid in the context of this problem.

I will be using the accuracy metric to evaluate my model as the predicted value is numeric, so metrics such as recall or precision would not apply in this case.

Baseline Model

# cause_detail pipeline
cause_detail = Pipeline([
     ('impute', SimpleImputer(strategy='constant', fill_value='missing')),
     ('one_hot', OneHotEncoder(drop='first', handle_unknown='ignore')),
])

# climate_region pipeline
climate_region = Pipeline([
    ('impute', SimpleImputer(strategy='constant', fill_value='missing')),
    ('one_hot', OneHotEncoder(drop='first', handle_unknown='ignore'))
    
])

# column transformer
column_transformer = make_column_transformer(
    (cause_detail, ['cause_detail']),
    (climate_region, ['climate_region'])
)
# best_est pipeline w/ Regression

pipe = Pipeline([
    ('preprocessor',column_transformer),
    ('linreg', LinearRegression())
])

hyperparams = {}

# GridSearch CV Searcher
searcher = GridSearchCV(
    pipe,
    param_grid = hyperparams,
    scoring='neg_mean_squared_error',
    cv=10,
    error_score='raise'
)

searcher.fit(X_train,y_train)
searcher

My baseline model contained 2 nomial features, cause_detail (what caused the severe weather power outage), and climate_region (climate region of the U.S where the outage occured). Since both these features are categorical, I used a OneHotEncoding on both features to convert the categorical values into numeric ones. I tested the mean squared error of my model using the X_test data to make predictions and y_test to evaluate the squared difference. I found the mse on the test data was 6986.82, which means the average difference between two values was ~83.5, or ~83.5 hours difference in outage duration. Since the difference was very significant, I would not consider by baseline model to be good.

Final Model

# cause_detail pipeline
cause_detail_final = Pipeline([
     ('impute', SimpleImputer(strategy='constant', fill_value='missing')),
     ('one_hot', OneHotEncoder(drop='first', handle_unknown='ignore')),
])

# climate_region pipeline
climate_region_final = Pipeline([
    ('impute', SimpleImputer(strategy='constant', fill_value='missing')),
    ('one_hot', OneHotEncoder(drop='first', handle_unknown='ignore'))
    
])

# area_urban pipeline
area_urban_final = Pipeline([
    ('impute', SimpleImputer()),
    ('std_scaler', StandardScaler())
])

# start_month pipeline
total_customers_final = Pipeline([
    ('impute', SimpleImputer()),
    ('std_scaler', StandardScaler())
])

# column transformer
column_transformer_final = make_column_transformer(
    (cause_detail_final, ['cause_detail']),
    (climate_region_final, ['climate_region']),
    (area_urban_final, ['area_urban']),
    (total_customers_final, ['total_customers'])
)

# best_est pipeline w/ Regression
pipe_final = Pipeline([
    ('preprocessor',column_transformer_final),
    ('ridge', Ridge())
])

hyprm_final = {
    'ridge__alpha': [2**9,2**-5,2**-4,2**-3,2**-2,2**-1,1,2,2**2,2**3,2**4,2**5,2**6,2**7,2**8]
}

# GridSearch CV Searcher
searcher_final = GridSearchCV(
    pipe_final,
    param_grid = hyprm_final,
    scoring='neg_mean_squared_error',
    cv=10,
    error_score='raise'
)

searcher_final.fit(X_train,y_train)
#searcher_final

On top of the 2 features in my baseline model, I added the percentage of urban land area in the state that outage occured (area_urban) and the total number of customers in the state (total_customers). Both of these features were numerical, and I added a standard scaler to both of them to prevent their values from skewing the predictions during Ridge regression. I believe these two features improved model performance because states with large urban areas and a large number of customers that needed electricity would have a much larger power grid. As a result of a larger power grid, weather causing outages in these areas would last longer because power had to be restored to a larger number of people.

The algorithm I chose for my final model was Ridge regression. I had used linear regression in my previous model and had tried polynomial features to better model the features that were one-hot encoded, but I did not see much improvement in my model. For my final model, I chose Ridge regression because I knew there was a large initial disparity in the coefficient sizes between the numerical and categorical features, and wanted to add a factor that punished large coefficients. Using Ridge, the best value for lambda I recieved was 4, and the mse on my final model was 6513.97. Which represents an improvement of ~3 hours in duration prediction for a power outage.