inZORi PF-Δ Phase 2 — Full Real AC Power Flow

Abstract

Phase 2 transitions from the controlled nonlinear surrogate (Phase 1) to full real AC Power Flow using pandapower on IEEE 118-bus — the first deployment of inZORi's Frozen Elite strategy on real grid physics. The operational regime is intentionally challenging: loads scaled to 1.75× nominal (near voltage-collapse boundary), Newton-Raphson budget constrained to nr_max=5, and N-1 topology shocks every 300 steps in S3 (3× more frequent than Phase 1). Static baselines converge at ~52% in S3. The inZORi Multi-Genome Frozen Elite selector achieves 69.6% S3 convergence (+16.8pp) over Baseline B, with N-1 recovery in 1.64 steps vs 6.7 steps for baselines (~4× faster). Validated across 20 independent seeds with CI95 non-overlap.

69.6%

inZORi Frozen Elite S3
CI95 [69.33%, 69.92%]

52.1%

Baseline A S3 convergence
(warm-start tracking)

+16.8pp

Advantage vs Baseline B
CI95 non-overlap confirmed

4×

Faster N-1 recovery
1.64 vs 6.7 steps

1.75×

Load scaling applied
Near voltage-collapse regime

Seeds validated
50,000 steps each

1. What Changed from Phase 1

Aspect	Phase 1 (surrogate)	Phase 2 (real AC PF)
Physics engine	Nonlinear surrogate (polynomial)	pandapower full AC Newton-Raphson
Network	118-bus surrogate topology	IEEE 118-bus real admittance matrix
Load level	1.0× nominal	1.75× nominal (near-collapse)
NR budget	`nr_max=8`	`nr_max=5` (tighter)
N-1 frequency	Every 900 steps (S3)	Every 300 steps (S3) — 3× more frequent
inZORi strategy	FULL online evolution	Frozen Elite (offline-evolved, O(1) inference)
Baseline S3 conv.	~84%	~52% (real AC physics harder)
inZORi S3 conv.	99.1%	69.6%
Advantage (pp)	+15.2pp	+16.8pp

2. Frozen Elite Pool — 6 Genomes

The Frozen Elite approach evolves a pool of diverse genomes offline (~10 generations on the 118-bus task). At runtime, no evolution occurs — O(1) contextual selection from the pre-computed pool:

#	Role	Profile
0	Balanced	Equilibrium between memory and exploration
1	Aggressive	Higher step scale, faster convergence attempt
2	Conservative	Low risk, stable in near-nominal conditions
3	High-memory	Strong warm-start bias, minimal exploration
4	N-1 recovery specialist	High DC-init jump tendency, activated on line outage
5	Diversifier	High step diversity, stress scenario coverage

Genome #4 (N-1 specialist) is automatically selected upon N-1 shock detection, providing systematic recovery via DC power flow warm-start fallback. Exact genome parameters are proprietary.

3. Results

Fig. 1 — S3 convergence rate with CI95 error bars. inZORi Frozen Elite: 69.6% vs ~52% for both baselines (+16.8pp). Non-overlapping CI95 confirms statistical significance. IEEE 118-bus · pandapower · 1.75× load · nr_max=5 · 20 seeds × 50k steps.

Fig. 2 — Steps from line restoration to first successful AC PF convergence after N-1 outage. inZORi: 1.64 steps vs Baseline A: 6.80, Baseline B: 6.61 — approximately 4× faster recovery. N-1 events every 300 steps in S3.

Fig. 3 — Three-panel dashboard: S3 convergence, mean NR iterations, and recovery steps. Across all metrics, inZORi Frozen Elite is favorable: highest convergence, acceptable iteration cost, lowest recovery time.

Fig. 4 — Research evolution Phase 1 → Phase 2. The +16.8pp advantage in Phase 2 is comparable to Phase 1's +15.2pp, confirming algorithmic robustness across fundamentally different physics engines.

Numerical Summary

Strategy	S3 Conv.	CI95	NR Iter (mean)	Recovery (steps)
Baseline A (warm-start)	52.08%	[51.51, 52.65]	1.81	6.80
Baseline B (periodic reset)	52.84%	[52.16, 53.51]	1.84	6.61
inZORi Frozen Elite	69.63%	[69.33, 69.92]	2.38	1.64
Advantage vs Baseline B	+16.8pp ✓	CI non-overlap ✓	+0.54 iter	~4× faster

Note: inZORi's higher NR iteration count reflects DC-init retry usage — additional computation that directly enables higher convergence, i.e., the extra iterations are productive.

4. Claims & Limitations

CLAIMS: Pre-evolved frozen genome pool achieves +16.8pp convergence improvement (CI95 non-overlap) over static baselines in real AC PF under 1.75× stressed load, N-1 every 300 steps
CLAIMS: Recovery post-N-1 is ~4× faster via DC-init fallback encoded in genome parameters — no explicit topology awareness in selector logic
CLAIMS: Consistent advantage (+15.2pp Phase 1 → +16.8pp Phase 2) across fundamentally different physics engines confirms algorithmic robustness
DOES NOT CLAIM: 69.6% absolute convergence is operationally "good" — it reflects an extremely stressed regime not typical of real SCADA operation
DOES NOT CLAIM: Optimality of the 6-genome pool — additional evolution rounds would likely improve results
DOES NOT CLAIM: Generalization to other network sizes or topologies without further experiments

5. Reproducibility

# Prerequisites
pip install pandapower numpy scipy matplotlib

# Run Phase 2 full experiment (12 cores recommended, ~2h)
cd inzori/problems/zor_pf_real_118/
python3 run_improved_full.py

# Key parameters:
# N_SEEDS=20, N_STEPS=50000, N_WORKERS=12
# LOAD_SCALE_BASE=1.75, NR_MAX_DEFAULT=5, N1_INTERVAL=300

# Interactive demo (nr_max=3, immediate differentiation)
python3 demo_interactive.py
      