PaperlayerView exact revision on arXiv

Source-linked AI summary

Physics-Informed but Not Physics-Consistent: Error Geometry and Subspace Projection for Neural AC Power Flow

Changhun Kim, Timon Conrad, Redwanul Karim, Karan Pahlajani, Julian Oelhaf, David Riebesel, Tomás Arias-Vergara, Andreas Maier, Johann Jäger, Siming Bayer

arXiv:2610.05959v1cs.LGcs.AI

TL;DR

Neural AC power-flow surrogates can predict voltages accurately while still producing power-balance violations, raising questions about prediction-error geometry. The paper analyzes these errors and applies calibrated solution-subspace projection; across four surrogates, it reduces residuals and voltage-magnitude error without retraining.

  • Problem

    Accurate voltage predictions do not necessarily imply physical consistency, and the paper asks which properties of prediction errors govern their AC residual impact.

    Source: §1 · p. 1

  • Method

    The authors analyze calibrated error components relative to a training-solution SVD subspace and propose post-hoc calibrated solution-subspace projection.

    Source: §1 · pp. 1–2, §2.2 · p. 2

  • Results

    Across four diverse surrogates, projection suppresses off-subspace errors and reduces voltage-magnitude error and power-balance residuals without retraining.

    Source: §4 · p. 4

  • Takeaways & Limitations

    The findings identify error direction, not only magnitude, as relevant to neural AC power-flow surrogate evaluation and design.

    Source: §4 · p. 4

    CSP is not a projection onto the nonlinear AC-feasible set and therefore does not guarantee exact feasibility.

    Source: §2.2 · p. 2

Abstract

from arXiv · show

Recent neural power-flow solvers, including emerging foundation models, achieve accurate voltage predictions, yet such accuracy does not necessarily imply physically consistent solutions. Even small complex voltage errors can yield large AC power-balance residuals. We study this accuracy-consistency gap across PIGNN-GC, GridSFM, gridfm-graphkit, and LUMINA on realistic 2224-bus Great Britain network (GBnetwork) scenarios, with cross-grid evaluation of GridSFM over 31 systems. Using a singular value decomposition (SVD) basis fitted to training AC power-flow solutions, we find that neural prediction errors contain substantial components outside the dominant solution subspace. To address this mismatch, calibrated solution-subspace projection (CSP) suppresses off-subspace prediction components after train-only bias calibration, reducing Mean PB by 67.0%, 37.8%, 40.5%, and 68.9% for PIGNN-GC, GridSFM, gridfm-graphkit, and LUMINA, respectively, relative to calibrated predictions, while improving voltage-magnitude accuracy in all four models. These results identify output-error geometry as an important factor in physics-consistent neural AC power flow. Code: https://github.com/Kimchangheon/neural-acpf-error-geometry

1. INTRODUCTION

Neural AC power-flow surrogates can predict voltages accurately without satisfying physical power-balance constraints. This section motivates studying how prediction-error direction relates to residuals and introduces a training-subspace correction.

1. INTRODUCTION

  • AC power flow underpins steady-state analysis, congestion management, contingency studies, optimal power flow, and state estimation.Neural surrogates have been explored to address the need for computational efficiency and physical consistency.Source: §1 · p. 1
  • Small voltage errors can still produce non-negligible power-balance violations because physics-informed objectives do not enforce exact nonlinear AC-equation satisfaction.The paper calls this discrepancy the accuracy–consistency gap.Source: §1 · p. 1
  • Similarly sized voltage errors can induce different AC residuals depending on their orientation across buses.The authors hypothesize that error direction matters beyond error magnitude.Source: §1 · p. 1
  • The study decomposes four surrogates’ calibrated errors into in- and off-subspace components using an SVD basis fitted to training AC-PF solutions.Fixed-norm interventions and local Jacobian analysis assess whether off-subspace errors are more residual-sensitive.Source: §1 · p. 2
  • Calibrated solution-subspace projection removes systematic bias and suppresses off-subspace predictions without retraining.The authors evaluate it on four frozen models on the 2224-bus GBnetwork and assess generality on 30 additional systems.Source: §1 · p. 2

2. ERROR GEOMETRY AND CSP

The method calibrates voltage predictions using training-only bias estimates, then projects them onto dominant directions learned from training power-flow solutions. It analyzes how error direction affects AC residuals and notes that projection trades removed prediction error against valid reference variation.

2.1. AC power-balance consistency and calibration

  • Power-balance residuals are recomputed from predicted voltages using the network admittance model, with specified bus quantities and voltage references restored before scoring.Mean and Max PB aggregate bus–scenario pairs on a common 100-MVA per-unit base.Source: §2.1 · p. 2

2.2. Error geometry and training solution subspace projection

  • The solution basis stacks centered voltage magnitudes and circularly centered angles, retaining the leading k state-space directions per channel; k = 16 means 16 per channel.Train-estimated mean bias is removed before geometric analysis and frozen at test time.Source: §2.2 · p. 2, §2.1 · p. 2
  • Equal-norm voltage-state errors can produce different AC residuals by orientation under the local residual Jacobian.The error is partitioned into in-subspace and off-subspace components after removing train-estimated bias.Source: §2.2 · p. 2
  • CSP removes estimated bias and projects calibrated predictions onto dominant training-solution directions, keeping the training mean and projection fixed at test time.It can discard valid reference variation as well as prediction error.Source: §2.2 · p. 2
  • State accuracy improves when removed off-subspace prediction-error energy exceeds the reference variation discarded by projection.The training-solution mean and projection remain fixed at test time.Source: §2.2 · p. 2
  • CSP is a post-hoc correction, not a projection onto the nonlinear AC-feasible set, so it does not guarantee exact feasibility.Its physical-consistency improvement depends on removed off-subspace components being residual-sensitive under the local AC Jacobian.Source: §2.2 · p. 2

3. EXPERIMENTS AND RESULTS

The experiments test error-subspace geometry and calibrated projection on four frozen surrogates using GBnetwork, then assess GridSFM across 31 systems and distribution shifts. CSP's effects are examined in terms of physical consistency, voltage accuracy, and Newton refinement.

3.1. Off-subspace error and directional sensitivity.

  • Median off-subspace error fractions reach 26.3/55.0/56.8/27.1% for magnitude and 10.3/2.7/39.6/31.7% for angle across PIGNN-GC, GridSFM, gridfm-graphkit, and LUMINA.Larger off-subspace fractions are associated with higher Mean PB after accounting for total error norms.Source: §3.1 · pp. 2–3

3.3. Post-training CSP correction

  • Relative to calibrated predictions, CSP reduces Mean PB by 67.0%, 37.8%, 40.5%, and 68.9% for PIGNN-GC, GridSFM, gridfm-graphkit, and LUMINA, while lowering voltage-magnitude RMSE in all four.Table 1 reports GBnetwork results at k = 16 across the four models.Source: Table 1 · p. 3, §3.3 · pp. 3–4

3.1. Off-subspace error and directional sensitivity.

  • At fixed error norm, suppressing off-subspace error lowers Relative Mean PB; off-subspace directions also show a 1,563× median paired residual-gain advantage over in-subspace directions.The gain ratio comes from 32 random direction pairs at each of 128 held-out operating points.Source: §3.1 · p. 3

3.2. Solution subspace specificity.

  • Matched-basis controls preserve nearly the same reference variation as SVD16 but yield substantially larger power-balance violations for projected references and corrected predictions.The comparison uses graph-Laplacian and spatial low-pass controls tuned to match SVD16's retained voltage-magnitude and angle variation.Source: §3.2 · p. 3

3.5. Cross-grid breadth and distribution shift

  • Across all 31 systems, spanning 4–9,241 buses and sub-kV to 750-kV levels, CSP reduces GridSFM Mean PB.The evaluation includes 28 pandapower/PYPOWER-format systems and three CIM-based networks; calibration and subspace statistics are fitted from each system's training data.Source: Figure 3 · p. 4, §3.5 · p. 4

3.3. Post-training CSP correction

  • Angle projection has model-dependent trade-offs: it lowers gridfm-graphkit's Mean/Max PB from .0612/18.23 to .0457/5.08 while worsening angle RMSE from .367° to .496°.For LUMINA, angle projection lowers Mean/Max PB from .1939/38.53 to .0660/4.51 and improves angle RMSE from 1.374° to 1.215°; projection effects are model dependent.Source: §3.3 · p. 4

3.4. Complementarity with one-step Newton refinement.

  • CSP initialization lowers magnitude and angle RMSE across all four models and reduces post-update Mean PB by 21.1–96.4% versus calibrated initialization.The one-step Newton update uses validation-selected damping η = 1 for all models; LUMINA's CSP16 initialization remains consistent across seeds despite an overshoot with calibrated initialization.Source: §3.4 · p. 4

3.5. Cross-grid breadth and distribution shift

  • Under N–1/N–2 topology shifts, CSP remains stable in most statistics, while GBcorr produces a broader performance drop.For GBcorr, Mean PB is 0.1003 p.u. with transferred calibration and basis, versus 0.068 after refitting both; projecting converged references yields 0.0868 and 0.0438 p.u., respectively.Source: §3.5 · p. 4

4. CONCLUSION

Across four architecturally diverse neural surrogates, output-error geometry helps explain the gap between voltage accuracy and physical consistency. Calibrated solution-subspace projection suppresses off-subspace components, reducing power-balance residuals and voltage-magnitude error without retraining.

  • Calibrated solution-subspace projection suppresses off-subspace errors, reducing voltage-magnitude error and power-balance residuals without retraining.Source: §4 · p. 4
  • Across four architecturally diverse surrogates, small voltage errors can contain off-subspace components that induce large AC power-balance violations.Source: §4 · p. 4
  • CSP's residual reduction extends to 30 additional grids.Source: §4 · p. 4
  • The findings suggest that error direction, not only magnitude, should inform surrogate evaluation and design.Source: §4 · p. 4
Loading 2610.05959v1…