Grid Breaks

UC San Diego DSC 80 — Final Project
Author: Uzair Gheewala

Introduction

Major power outages disrupt lives and economies. The U.S. Department of Energy’s Major Power Outage database (2000 – 2016, ≈ 190 k events) records what happened (CAUSE.CATEGORY) and how bad it was (duration, customers affected, MW lost, geography).

During an initial scan we noticed one label – fuel-supply emergency – whose outages last months, unlike storms or equipment failures. That raised two questions:

Descriptive Do numeric outage characteristics naturally cluster in a way that mirrors the official cause codes?
Predictive Can we infer the cause at the moment an outage starts, using only information available in real time?

Data Cleaning and Exploratory Data Analysis

Cleaning steps

Dropped non-data header rows from the DOE Excel file.
Parsed start / restoration timestamps and computed duration_hours.
Replaced obviously bad durations (≤ 0 h) with NaN and median-imputed all numeric features inside modelling pipelines.
Renamed unwieldy columns:
POPDEN_RURAL, POPDEN_UC, … for clarity.

|   YEAR | U.S._STATE   | CAUSE.CATEGORY     |   duration_hours |   CUSTOMERS.AFFECTED |   DEMAND.LOSS.MW |
|-------:|:-------------|:-------------------|-----------------:|---------------------:|-----------------:|
|   2011 | Minnesota    | severe weather     |             3060 |                70000 |              nan |
|   2014 | Minnesota    | intentional attack |                1 |                  nan |              nan |
|   2010 | Minnesota    | severe weather     |             3000 |                70000 |              nan |
|   2012 | Minnesota    | severe weather     |             2550 |                68200 |              nan |
|   2015 | Minnesota    | severe weather     |             1740 |               250000 |              250 |

Univariate view

Fig 1. fuel-supply emergencies (red) tend to extend to greater extremes,

10 000 h (≈ 14 months).

Bivariate & aggregates

Fig 2. Durations by cause — the fuel-supply box literally towers above all others.

| CAUSE.CATEGORY                |   mean_duration |   median_customers |   count |
|:------------------------------|----------------:|-------------------:|--------:|
| fuel supply emergency         |         13484.0 |                0.0 |    38.0 |
| severe weather                |          3899.7 |           111555.0 |   741.0 |
| equipment failure             |          1850.6 |            51500.0 |    54.0 |
| public appeal                 |          1468.4 |                0.0 |    69.0 |
| system operability disruption |           747.1 |            69000.0 |   120.0 |
| intentional attack            |           521.9 |                0.0 |   332.0 |
| islanding                     |           200.5 |             2342.5 |    44.0 |

The table confirms that fuel-supply emergencies last on average ~13 000 h (≈ 1½ years) while other causes stay below 4 000 h.

Severe-weather events are most common but not the longest; fuel-supply emergencies have mean duration > 3000 h.

Assessment of Missingness

One high-missingness column – DEMAND.LOSS.MW (~45 % NaN) – reflects the fact that DOE only records MW loss when utilities supply it. The remaining model features have ≤30 % missingness, with location and calendar measures fully observed. These gaps are most likely MAR (measurement not always reported) rather than MCAR.

A qualitative review suggests CAUSE.CATEGORY itself can be NMAR: utilities may delay labelling politically sensitive intentional attacks, leaving the field blank until investigations conclude.

Fig 3. Fraction of NaNs per feature; we median-impute in pipelines.

Hypothesis Testing

We formally test the anecdotal observation:

Fuel-supply emergencies last longer than intentional attacks.

H₀ (null): Durations follow the same distribution.
H₁ (alt): Fuel-supply durations are stochastically greater (KS alternative="less").
Test statistic: one-sided two-sample Kolmogorov–Smirnov.
Significance: α = 0.05.

Fig 4. ECDF plot of duration show months-long tail for fuel-supply outages.

Result: KS = 0.65, p < 1 × 10⁻¹³ ⇒ reject H₀. Fuel-supply emergencies are dramatically longer (median ≈ 3900 h) than intentional attacks (≈ 30 h).

Framing a Prediction Problem

Task Predict CAUSE.CATEGORY at outage start (multiclass classification, 7 labels).

Why Early inference helps utilities dispatch the correct crews (e.g. security vs fuel-logistics).

Features known at time 0

kind	features
Real-time numeric	`duration_hours` (current), `CUSTOMERS.AFFECTED`, `DEMAND.LOSS.MW`
Geography	`POPDEN_RURAL`, `POPDEN_UC`
Calendar	`MONTH` (and engineered sin/cos)

Metric Macro-averaged F1-score — treats minority classes equally.

Baseline Model

StandardScaler ➜ LogisticRegression (multinomial) with median imputation.

Macro-F1 = 0.20 – the model all but ignores rare labels.

Final Model

Random-Forest (600 trees, class-weight balanced)

engineered logs & cyclic month features.
Best parameters: n_estimators = 600, max_depth = None
Macro-F1 = 0.53 (↑ 0.33 absolute).

Fig 6. Confusion matrix – most errors now occur among the three rarest labels.

Fairness Analysis

Question Does the model perform worse for outages in rural (counties ≥ median POPDEN_RURAL) versus urban locations?

H₀: Macro-F1 identical. H₁: Different.
Observed ΔF1 = -0.14 (rural lower).
Permutation p = 0.18 ⇒ fail to reject H₀.

Fig 7. Null distribution of ΔF1; observed gap (red) is not statistically significant.

While the point estimate hints at a rural performance dip, evidence is inconclusive given sample size. Future work could collect more rural events or incorporate grid topology to close any potential fairness gap.

Last updated: 2025-06-06