Violation capture by review budget
VALIDATION POPULATIONTheoretical ceilingGBT risk rankingPre-outage loading
POWER SYSTEMS × MACHINE LEARNING
Screen the outage.
Prioritize the review.
A PySpark workflow evaluated on a synthetic IEEE 118-bus DC benchmark. Scenario-grouped validation measures ranking quality. Policy frozen. Final test unevaluated.
Explore the measured results01 / SCREENING EVIDENCE
What gets reviewed first?
Model comparison
PR-AUCModel metrics shown here belong only to the selected evidence source.
THE OPERATING QUESTION
How much review can a screen defer while retaining modeled violations?
Review counts describe this synthetic benchmark, not engineering time or financial savings.
Local validation gains over the strongest tested simple baseline
Here, pp means percentage points. GBT captures 95.07% of modeled violations at 20% review versus 78.78% for loading. At 10%, capture rises from 54.33% to 78.92%.
| Decision metric | Loading | GBT | Change |
|---|
Uncertainty. At 20% review the capture gain is +16.29 pp, with a 95% paired scenario-bootstrap interval of +15.75 to +16.72 pp. At 10%, the interval is +23.23 to +25.96 pp. Mean per-scenario recall at 20% is 79.39% for loading versus 93.80% for GBT, a +14.41 pp gain with interval +13.85 to +14.97 pp. Based on 1,000 scenario resamples. Intervals do not correct selection optimism or calibration overlap. PR-AUC uncertainty was omitted because Spark metric equivalence was not established.
Interpretation. Pooled capture ranks all validation cases together. Mean scenario recall gives each positive scenario equal weight after ranking within that scenario. These are different review policies.
Caution. At 20% pooled review GBT still misses 875 positives. Under within-scenario review, 44.05% of positive scenarios still contain a miss. Loading was selected on these validation rows from nine descending single features and loading times sensitivity. Gaps are descriptive and subject to selection optimism, with bootstrap uncertainty reported separately.
At 10% review, GBT reaches 91.44% of the nominal pooled capture ceiling versus 62.95% for loading. The chart uses the realized review count for its ceiling. Points are measured, lines guide the eye. These synthetic DC validation results do not measure real-grid benefits.
02 / METHOD, WITH GUARDRAILS
Keep related scenarios together.
Each observation pairs an operating scenario with a candidate outage. Splitting distinct scenario IDs first prevents related outage rows from crossing training, validation and test partitions.
- ValidateSchema, finite values, binary targets and scenario–outage integrity.
- Group & compareLR, RF and GBT. Metadata never enters the nine-feature vector.
- Choose & freezeTraining-only grouped search and validation analysis. Policy frozen. Final test unevaluated.
03 / HONEST BOUNDARIES
A screening prototype.
Engineering judgment stays essential.
Private project dataCanonical datasets are not distributed. Authorized users configure external local inputs; public material contains aggregates and methodology.
Generator not recoveredThe original physics-based generator is missing. ML workflow reproduction is supported with authorized inputs; full CSV regeneration is not claimed.
No utility safety claimSynthetic scenarios, modeled thermal limits and simplified DC physics do not replace AC contingency analysis or human validation.