Skip to content

Latest commit

 

History

History
38 lines (22 loc) · 2.29 KB

52-screening-summary.md

File metadata and controls

38 lines (22 loc) · 2.29 KB

52: Summary of the screening approaches

Date: 2026-11-16 · Plan phase: 13

Goal

One document naming the best deterministic and the best learned screen, with their operating points, what each needs and where each stops.

What was done

  • docs/SCREENING.md. No new computation: every number is quoted from the step that produces it (14, 18, 20, 21, 23 to 27, 29, 30, 40, 42, 47 to 51).

Commands

None. The numbers come from the scripts named in each quoted step.

Results

Screen Operating point
Best deterministic 2 × 2 admittance scan of the plant, least-squares fit, closed on the network at the 32 ends of its data's ranges; phasor-domain model beside it for the slow band 0 unstable cases cleared in 24,659; clears 57.8 % of doubly-fed cases and 87.8 % / 70.6 % of converters, with a 1 % scan and the network to within 10 % (step 24)
Best learned Hand calculation plus a learned correction, ensemble of five, fallback on the spread, range check 0.72 % of cases to the solver, 0 of 1,376 unstable cases called stable, in the range (step 40); 0.9 % and 3.2 ms per answer in the service (step 42)
  • The physics of the deterministic screen: the resonance is a series resonance. At the frequency where the reactances cancel, the rotor resistance and the current controller's gain, divided by a negative slip, leave a negative resistance against the network's positive one. A test on the sum of impedances sees it; a test on the loop gain does not.
  • Phase 13 did not improve the learned screen. Physics residuals as detectors (step 48) did no better than the spread, active learning (step 50) was cheaper but missed cases, and the converter (step 49) and multi-plant (step 51) versions did not beat their deterministic counterparts as screens.

Disagreements

  1. The two screens are not scored on the same cases. Step 24 used a random quarter of the cases of both plant types with uncertain network data; step 40 used the T1 test fold with the network exact. The table does not pair them.

Decisions

  • The deterministic scan screen is the reference screen. The learned screen is for throughput inside its range, with the solver behind it.
  • Phase 13 is closed.

Next

Open: validation rung 5 (recorded events) and a comparison with an electromagnetic-transient tool.