PaperlayerView exact revision on arXiv

Source-linked AI summary

Why the Swing Parameters Are Sloppy

M Parimi, Shashank Varma, Sunny Kumar, S R Wagh, Navdeep Singh

arXiv:2610.30084v1eess.SYcs.SY

TL;DR

The paper asks why late-window analyses find H and D unidentifiable and why those conclusions agree with singular-perturbation analysis. It explains the result using slow-manifold geometry and derives boundary and experiment-dependent predictions. Under PMU-grade noise, the late window identifies five parameters and the early window nine, while the analysis has stated scope limits.

  • Problem

    Earlier analyses agreed that H and D were unidentifiable on late records, but did not explain the agreement with singular-perturbation analysis or its implications across experiments.

    Source: §1.1 · p. 1, §1.2 · p. 4

  • Method

    The paper uses the generator’s fast-slow dynamics and normally attracting invariant manifold to analyze parameter sensitivities, memory, and model-manifold boundaries.

    Source: §2.4 · p. 7, §3.3 · p. 12

  • Results

    At PMU-grade noise, the late window identifies five parameters rather than six, while the early window identifies nine, with H best determined.

    Source: §6.2 · p. 21

  • Takeaways & Limitations

    Neither a larger initial speed deviation nor a load change restores H or D in the late window, whereas observing the early window makes H the best-determined parameter.

    Source: §6.2 · p. 21

    The analysis concerns one machine at one operating point and uses Fenichel’s hypotheses as a guide rather than proving them for this model.

    Source: §6.3 · p. 21

Abstract

from arXiv · show

Evangelou, Stankovi\'c, Kevrekidis and Transtrum compared analytic and data-driven assessments of parameter identifiability for the infinite-bus synchronous generator. Fisher information with the manifold boundary approximation method (MBAM), and output diffusion maps with geometric harmonics, report the same six identifiable parameters on the late window $t \in [3, 5]\text{ s}$, and both place the inertia $H$ and the damping $D$ among the unidentifiable ones. We explain this result from the geometry of the dynamics rather than from that of the parameter-to-data map. The rotor swing is a fast fibre that has collapsed onto a two-dimensional normally attracting invariant manifold (NAIM) of the transient fluxes before the window opens. The zeroth-order NAIM is exactly independent of $H$, $D$, $T_{d0}''$ and $T_{q0}''$, four of the five MBAM limits. We prove, under stated smoothness and rank conditions, that each of the four limits is a boundary of the model manifold at a distance equal, to first order in the parameter, to its current profiled speed. The same analysis predicts how identifiability changes with the experiment: neither a larger initial speed deviation nor a load change restores $H$ or $D$ in the late window, whereas the early window $[0, 2]\text{ s}$ makes $H$ the best-determined parameter. Evaluated in a physical metric fixed by synchrophasor accuracy limits, the late window identifies five parameters rather than six and the early window nine. The spiked-covariance structure of the linearised trajectory map explains why the principal components of whitened, sampled trajectories reproduce the prior-scaled Fisher spectrum, and regression on those components attains the Fisher-predicted recovery error in 113 of 114 parameter--scenario pairs.

1 Introduction

The late-window sloppiness of the generator’s swing parameters motivates a state-space explanation: after fast transients decay, the record reflects slow-manifold dynamics and confounded memory. The paper uses this geometry to explain identifiability and its dependence on the experiment, while delimiting where the analysis applies.

1.1 The paradox

  • On late records, H and D are among the least identifiable parameters because the swing they govern has already decayed; inertia and damping leave only faint tilt and memory traces.The late-window lag is of order 10^-4 rad, and the memory trace is confounded with effects of other parameters.Source: §1.1 · pp. 1, 3

1.2 The information-geometric approach

  • Earlier information-geometric analyses found six identifiable parameters on the late window, but did not explain the agreement with singular-perturbation analysis or predict how results change across experiments.The cited work used Fisher information, MBAM, and output diffusion maps with geometric harmonics.Source: §1.2 · p. 4

1.3 The state-space approach of this paper

  • The paper analyzes identifiability through the generator’s fast-slow state-space structure, using a normally attracting invariant manifold as an approximation at finite time-scale separation.The manifold is computed by a convergent iteration, with consequences checked numerically against the full model.Source: §1.3 · p. 4

Zeroth-order invisibility.

  • The zeroth-order manifold graph is independent of H, D, T_d0'' and T_q0'', so the reduced slow model has no leading-order sensitivity to those four parameters.These are four of the five MBAM limits discussed in the paper.Source: Zeroth-order invisibility · pp. 4–5

Three residual channels.

  • After the fast transient decays, fast-only parameters affect the record through a decaying fibre residual, a shift in the manifold landing point, and higher-order manifold tilt.Memory effects are largely confounded with slow-parameter sensitivities: more than 99.6% of their norm lies in that span, leaving the small tilt after profiling.Source: Three residual channels · p. 5

1.4 What the NAIM analysis adds

  • The NAIM analysis derives the mechanism and boundary behavior without geodesic integration, and predicts that larger initial speed deviations or a load change do not restore H or D in the late window.Moving the observation window into the fibre transient makes H the best-determined parameter; the method also gives a hierarchy of corrected models.Source: §1.4 · pp. 5–6

1.5 Scope and organisation

  • The analysis has a usable-gap boundary: it does not predict regular near-degeneracies among slow parameters, and identifiability thresholds still depend on measurement noise.Regimes without a clear gap, such as clustered eigenvalues or a nearby bifurcation, fall outside the NAIM analysis.Source: §1.5 · p. 6

2 The generator model and its fast-slow split

The section reproduces the generator model’s late-window Fisher spectrum, then interprets its time-scale separation through a fast-slow invariant-manifold approximation. It also records transcription corrections and the limits of the approximation.

2.1 Model

  • The model parameters are varied in logarithmic coordinates, with nominal values chosen to preserve ordering constraints on reactances and time constants.The post-disturbance initial state and fixed constants are specified for the model.Source: §2.1 · pp. 6–7

2.2 Two transcription corrections

  • Two transcription corrections are needed to reproduce the published results: the rotor-speed reference is per unit, and the standard form reproduces the trajectories.The text also identifies a likely typographical slip between Eqs. (37) and (7) of [1].Source: §2.2 · p. 7

2.3 Sensitivities, Fisher information and reproduction

  • The sloppiest Fisher eigenvalue is 1.207 × 10−9, matching [1], while the sixth eigenvalue falls below the 10−2 cutoff in the stated coordinates.Using T′q0 instead of δTq raises the sixth eigenvalue to 1.26 × 10−2, so the cutoff count depends on coordinates.Source: §2.3 · p. 7

2.4 Time scales: the swing mode is a fast fibre

  • The swing pair decays 3.7 times faster than the faster slow mode, and by 3 s the swing has decayed by about 3 × 10−6.The initial speed deviation produces a large swing, but that motion is over before the [3, 5] s observation window.Source: §2.4 · p. 7, Table 2 · p. 8

2.5 Fast-slow split and the invariance equation

  • The fast-slow graph is approximated by iterated quasi-steady-state construction, with each iteration adding one order in the slow-to-fast rate ratio.The authors use Fenichel theory as a guide but do not analytically verify its hypotheses for this single machine.Source: §2.5 · p. 8

3 Why a late record cannot see the swing

After the swing collapses onto the NAIM, late-window sensitivities to fast-only parameters are limited to memory and tilt, which can be confounded with slow-parameter effects. The analysis connects that geometry to zeroth-order parameter invisibility and quantifies the resulting loss of information.

3.1 Zeroth-order invisibility

  • The zeroth-order NAIM is independent of four fast-only parameters, so its Fisher information matrix has four corresponding null directions.The fifth MBAM limit, δx1, instead reflects a regular near-degeneracy: it is a 3% addend to a sum dominated by δx2 + δx3.Source: §3.1 · p. 9

3.2 Three residual channels

  • Fast-only sensitivities decay with the swing envelope during collapse, then settle into a slow tail carried by memory and tilt; D reaches its tail by about 1.5 s and H by 2.5–3 s.The latter timing coincides with the start of the late window used in [1].Source: §3.2 · p. 11

3.3 What the memory channel is

  • After the fibre residual vanishes, a fast-only parameter affects the late record through its shift of the base point and its tilt of the NAIM; the memory shift has only m = 2 coordinates here.The resulting memory directions are spanned by output directions that slow parameters also span.Source: §3.2 · p. 11, §3.3 · p. 12

3.4 Memory confounding

  • 99.94% and 99.61% of the two memory directions’ norms lie in the span of the seven non-fast parameter sensitivities; profiling them reduces information on H from 3.5 × 10−5 to 1.5 × 10−7 and on D from 3.4 × 10−6 to 8.1 × 10−9.The stacked initial-condition sensitivity has two surviving singular directions on s; the next two correspond to the swing residual.Source: §3.4 · p. 12
  • On s, the second-order NAIM approximation reproduces full-model fast-state sensitivities with relative errors no greater than 4.8 × 10−2 for H, 2.0 × 10−3 for D, and 2.5 × 10−2 for T ′′ d0.For T ′′ q0 the reported relative error is 6.3 × 10−3.Source: §3.4 · p. 13

4 Manifold boundaries without geodesics

The paper derives late-window model-manifold boundaries for four parameters from the slow-manifold structure, without geodesic integration. It tests the boundary predictions against MBAM and examines how changing the observation window or experiment changes identifiability.

4.1 The boundary theorem

  • Under the stated smoothness, contraction, and rank conditions, each of four null-set limits is a model-manifold boundary whose distance is asymptotically its current profiled speed.The output is affine to first order in each parameter, giving d_j = θ_j∥Π_jb_j∥ + O(θ_j^2) and d_j/p_j → 1.Source: §4.1 · p. 13

4.2 Comparison with the MBAM sequence

  • 3.52 × 10−5 and 2.98 × 10−4 are the first two MBAM geodesic lengths, matching local boundary-distance predictions within 5%.The observed boundary ordering D < H < T′′q0 < δx1 matches the published MBAM sequence.Source: §4.2 · p. 14, Table 3 · p. 14

4.3 The experiment as a variable

  • On [0.5, 2.5] s, H's profiled speed is 0.19, about 700 times its s value; on [0, 2] s, nine unit-noise FIM eigenvalues exceed 10−2.H moves from second to sixth place on [0.5, 2.5] s; on [0, 2] s, H and T′′q0 are among the resolved directions.Source: §4.3 · p. 14, Table 4 · p. 15, Table 5 · p. 15
  • On s, neither a larger initial speed deviation nor a load step restores H or D; both remain at least three orders of magnitude below the 10−2 cut-off.The tests use unit noise and the published cut-off.Source: §4.3 · p. 16, Table 6 · p. 16

5 Realistic measurement noise and connection to sampled ensembles

Using PMU-based noise limits, the paper evaluates practical parameter recovery and connects physical Fisher information to sampled-trajectory analyses. It finds five identifiable parameters on the late window and nine on the early window, while whitened trajectory ensembles reproduce the prior-scaled Fisher spectrum in the linearised regime.

5.1 Noise model and Bayesian identifiability criterion

  • The physical metric uses per-sample noise covariance based on instrument accuracy, and a parameter is identifiable when its predicted posterior-mean MAE is below 0.1.The critical noise multiple is the largest factor by which class-B standard deviations can increase while MAE remains below 0.1.Source: §5.1 · p. 16

5.2 Fisher analysis in the physical metric

  • Five parameters are identifiable at PMU-grade noise on s, versus nine on [0, 2] s, where H is the best-determined parameter.On the late window, δTq has κ* = 0.57 and predicted MAE 0.150; on the early window H has κ* ≥ 31.6, standard deviation 0.1%, and MAE 0.004.Source: §5.2 · p. 17
  • On s, neither ω(0) = 0.90 nor load steps restore H or D under PMU-grade noise; on [0, 2] s, ω(0) = 0.90 makes D identifiable under classes A and B.At class B, 45% of efficient information on D comes from ω(t); under class-C dynamic limits, D is again unidentifiable with MAE 0.217.Source: §5.2 · p. 17, Table 9 · p. 17
  • For initial conditions on s, the class-B FIM has a three-order eigenvalue gap after the second slow-manifold direction, and recoverable initial states fall from four to one as the window starts later.The counts are 4 → 3 → 2 → 1 for windows starting at 0.25, 1, 2, and 3 s.Source: §5.2 · p. 17

5.3 Why data-driven manifold learning mirrors the Fisher spectrum

  • In the linearised regime, the leading principal directions of whitened trajectory ensembles are images of prior-scaled Fisher eigenvectors whose eigenvalues exceed the finite-sample threshold.The population covariance has a spiked-covariance form governed by the prior-scaled Fisher matrix.Source: §5.3 · pp. 18–19
  • Geometric-harmonics regression on components above the noise edge matches Fisher-predicted recovery error in 113 of 114 parameter–scenario pairs; diffusion maps with residual-based selection agree in 108 of 114.On s, δTq is visible to the ensemble but has predicted Bayesian MAE 0.150, so it is not practically identifiable in a single record; per-feature [0, 1] scaling acts as an implicit noise model.Source: §5.3 · pp. 18–19

6 Discussion and conclusions

The conclusions explain late-window sloppiness through the slow-manifold geometry, relate parameter limits to model-manifold boundaries, and show how window and noise choices change identifiability. The authors also bound the scope of those conclusions to a single-machine analysis with important measurement and modeling assumptions.

6.1 Comparison with Evangelou, Stankovi´c, Kevrekidis and Transtrum

6.2 Conclusions

  • The zeroth-order NAIM is independent of H, D, and the damper time constants, while those parameters re-enter through manifold tilt and fast-mode memory.The slow parameters span the landing-point channel to 99.94% and 99.61%; H enters at second order, with the swing pair further suppressed by 1/ωb.Source: §6.2 · p. 19
  • Each null-set limit is a model-manifold boundary at a distance equal, to first order, to its current profiled speed; the terminal reduced model is the zeroth-order slow-manifold model.The first two boundary distances agree with computed geodesic lengths to within 5%.Source: §6.2 · p. 19
  • At PMU-grade noise, the late window identifies five parameters and the early window identifies nine, with H best determined early; neither a larger initial speed deviation nor a load change restores H or D late.Source: §6.2 · p. 21

6.3 Limitations

  • The analysis covers one machine at one operating point, uses Fenichel’s hypotheses as a guide rather than proving them for this model, and treats reported identifiable counts as upper bounds under idealized measurement assumptions.Real PMU errors are correlated in time and only terminal quantities are measured, unlike the independent-error, all-state setting.Source: §6.3 · p. 21

A Computational details

The computational details describe numerical methods for model integration, slow-manifold and geodesic calculations, recovery-error prediction, and trajectory-ensemble analysis.

Reproducibility

  • Scripts from the model equations produced every number in Sections 2–5 except the slow-manifold graph computations, and the scripts, outputs, and fixed random seeds are supplementary.Source: A Computational details · p. 21

Integration and sensitivities

  • The model and forward sensitivities were integrated with DOP853 at relative tolerance 10^-10–10^-11, or BDF for stiff limits, with both agreeing with a Radau IIA reference to relative 10^-11.Fisher quantities used sampling at 0.02 s.Source: A Computational details · p. 21

Slow-manifold graphs

  • Slow-manifold graphs were evaluated along the trajectory in 130-digit arithmetic, with Newton iterations reaching residuals below 10^-122.Derivatives were computed by nested central differences with step 10^-20.Source: A Computational details · p. 21

Geodesics and profiled speeds

  • Geodesics in free log-parameters started along the sloppiest eigenvector at unit data-space speed, and profiled speeds were evaluated at specified parameter ratios and integrated by trapezoidal quadrature.The target log-parameter was reduced by 5.1 for D or 3.1 for H before Theorem 1(b) supplied the remaining length.Source: A Computational details · p. 21

Predicted recovery error

  • Predicted recovery error was tabulated by Monte Carlo for a truncated-normal posterior mean over a logarithmic grid of noise ratios, with the critical noise multiple found by bisection.Source: A Computational details · p. 21

Sampled ensembles

  • Sampled, whitened trajectory ensembles were analyzed by principal components above the stated edge and geometric-harmonics regression with 250 kernel eigenvectors and an 80/20 train/test split.Trajectories used log-parameters within ±10%, were sampled every 0.02 s, and used an inner validation split to choose bandwidth per target.Source: A Computational details · p. 22
Loading 2610.30084v1…