Phase 2 transitions from the controlled nonlinear surrogate used in Phase 1 to a full real AC Power Flow (AC PF) computation using the pandapower library on the IEEE 118-bus network. This represents the first deployment of inZORi's Frozen Elite strategy on a real grid physics simulator rather than a mathematical approximation. The operational regime is intentionally challenging: base loads are scaled to 1. 75x nominal (near voltage-collapse boundary), the Newton-Raphson iteration budget is constrained to nrₘax=5, and N-1 topology shocks occur every 300 steps in Stage 3 — a frequency roughly 3x higher than Phase 1. Under these conditions, static heuristic baselines converge at approximately 52% in Stage 3. The inZORi Multi-Genome Frozen Elite selector — an adaptive, O (1) -cost runtime mechanism choosing among six pre-evolved genomes based on contextual signals (season, recent convergence, N-1 status) — achieves 69. 6% S3 convergence, a +16. 8 percentage points (pp) improvement over Baseline B, with CI95 non-overlap confirmed across 20 independent seeds. Recovery time after N-1 line restoration: inZORi recovers in 1. 64 steps versus ~6. 7 steps for baselines (~4x faster). Key results (Real AC PF | IEEE 118-bus | 1. 75x load | N-1 every 300 steps | 50k x 20 seeds): - S3 convergence (inZORi Frozen Elite): 69. 63% (CI95 69. 33, 69. 92) - S3 convergence (Baseline A): 52. 08% (CI95 51. 51, 52. 65) - S3 convergence (Baseline B): 52. 84% (CI95 52. 16, 53. 51) - Improvement vs Baseline B: +16. 8 pp (CI95 non-overlap confirmed) - N-1 recovery: 1. 64 steps (inZORi) vs 6. 80 (Baseline A), 6. 61 (Baseline B) — ~4x faster - Runtime: O (1) genome selection, no online evolution at inference time This is Phase 2 of a two-phase study. Phase 1 (nonlinear surrogate, 99. 1% S3 convergence) is published at: https: //doi. org/10. 5281/zenodo. 18716837
Dumitru Novic (Fri,) studied this question.