Identifiability Limits of Forced Oscillation Sources in Power Systems

arXiv:2610.00356 · eess.SY, cs.SY, eess.SP, math.DS · Submitted 2026-09-30 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "Identifiability Limits of Forced Oscillation Sources in Power Systems".

Dev: Whether a forced-oscillation source can be uniquely localized depends jointly on the available measurements and the candidate intervention dictionary.

Rosa: First, who's behind it and why it matters.

Title and authors: Rosa: We started by looking at the title and the authors of "Identifiability Limits of Forced Oscillation Sources in Power Systems," and it really sets a serious tone for this whole discussion about system observability. Dev I agree, Rosa; the title immediately tells us that we aren't looking at a perfect localization scenario, but rather where our limits are when dealing with forced oscillations in power grids.

Taro: From an autonomy research standpoint, that framing is interesting because it sets up a clear boundary for what an autonomous system can achieve when faced with inherent ambiguity. Rosa And the authors, Kai Sun and the others mentioned in the abstract, they are clearly deep into this kind of system identification theory to define those limits precisely.

Dev: It's important to understand that they aren't just looking at a simple signal; they consider a single unknown constant-amplitude sinusoid acting through one of several physical intervention channels. Taro That means the complexity comes from how that single source couples into the network, which is much more realistic than just assuming a simple input.

Rosa: And the implication of this is that we need to be very careful about what we assume when trying to pinpoint a source location in a complex power system where things are constantly changing. Dev Right, and they define the problem by saying that each candidate harmonic response is observable only up to some nonzero complex scalar because the amplitude and phase are unknown.

Taro: That nonzero scalar essentially means that a single candidate doesn't give us one specific answer but rather an entire direction in measurement space, which is why they talk about projective rays. Rosa So, when we try to localize something, we aren't just looking for a point; we’re looking along a ray defined by the unknown scalar.

Dev: Precisely; and this leads directly into the idea that exact localization isn't guaranteed unless all our candidates are nonzero and pairwise projectively distinct. Taro That condition sounds very mathematical, but it underpins whether any physical localization is even possible at all under ideal conditions.

The paper's summary: Rosa: Moving on to the actual summary of "Identifiability Limits of Forced Oscillation Sources in Power Systems," they lay out the fundamental concept that we have to define source location relative to a specific physical intervention in a fixed component realization. Dev That setup is crucial because it grounds the problem in reality; you can't just guess where something is without tying it to a known physical point.

Taro: The paper frames this as an identifiability problem: assuming one single sinusoid enters one unknown member of a finite candidate dictionary, and the goal is to find which member it is. Rosa So, they are essentially asking, given these specific measurements, which physical channel is responsible for this oscillation?

Dev: The central observation they make is that the unknown amplitude and phase multiply each candidate’s measured harmonic response by an arbitrary nonzero complex scalar. Taro That means a candidate isn't just one vector; it's actually a one-dimensional complex subspace, which they refer to as a projective ray in measurement space.

Rosa: That subspace description helps explain why the amplitude and phase are so hard to pin down initially; they multiply everything by that unknown scalar. Dev And because of this, the paper derives necessary and sufficient conditions for exact localization based on these projective signatures.

Taro: The main result they present is that all candidates in K are uniquely identifiable from the noise-free steady-state phasor if and only if each candidate's harmonic response g k(omega) is nonzero and all of them are pairwise projectively distinct. Rosa That means the mathematical condition for success is pretty straightforward, even though it relies on defining those signatures first.

Dev: They tie this directly into the "Descriptor rank test," which requires a specific rank condition, namely "rank j omega E - A-b k b C dk-d = n + two " for every pair of candidates k and. Taro So, it's not just about checking if the response is zero; it's about checking the rank of a specific matrix related to the system dynamics and those candidate responses.

The paper's improvements: Rosa: Now, let's look at what they suggest as improvements or tools for dealing with these situations, because it seems like pure identifiability isn't always the whole story. Dev They introduce a framework specifically designed to distinguish four different mechanisms of source localization failure: mechanism mismatch, feature-projection loss, structural nonidentifiability, and poor conditioning or model error.

Taro: I think that four-step diagnostic framework is really useful because it helps us categorize the kind of failure we're seeing when localization goes wrong in a real system. Rosa It moves beyond just saying "it didn't work" to explaining *why* it failed in terms of the underlying physics or the measurement setup.

Dev: The first step is raw identifiability, which checks if the candidate rays are even distinct at all, and then robust separation, where they assess if that distance between rays is large enough relative to any uncertainty we have. Taro That robustness check seems particularly relevant because in real-world scenarios, small measurement errors can easily push two nearly identical candidates into the same ambiguity zone.

Rosa: And the third step, feature preservation, looks at whether the selected feature actually keeps those raw data distinctions intact when we project them into a different space. Dev That sounds like it addresses a problem where raw data might look different, but our chosen diagnostic tool collapses those differences together during processing.

Taro: The fourth step is mechanism-consistent interpretation, which determines if the feature we end up using actually makes sense for the physical mechanism that's active in the oscillation. Rosa It forces us to consider not just mathematical distinction, but physical relevance when deciding what data to rely on.

Conclusion: Dev: So, wrapping up this discussion on "Identifiability Limits of Forced Oscillation Sources in Power Systems," the main implication is that exact localization hinges entirely on having nonzero and pairwise projectively distinct candidate signatures. Rosa That means if those mathematical conditions aren't met, we can’t guarantee unique identification from the available harmonic measurements alone.

Taro: I think the real impact here is establishing a clear theoretical limit; it tells us precisely when we need to stop trying to localize and start demanding more information or a different physical setup. Dev Right, and they've done some numerical studies on systems like Kundur and IEEE–NASPI that show things like measurement-induced ambiguity, which are very common issues in practice.

Rosa: I think the practical value is in knowing when to trust a certain diagnostic tool versus switching to another one because of feature projection loss or poor conditioning. Dev And they point out that correctly locating the source bus doesn't automatically mean you've identified the internal forcing channel, which is an important distinction for operators.

Taro: I feel like this paper sets a solid foundation for designing future sensing and control systems that inherently account for these projective ambiguities from the start. Rosa It’s a lot to take in, but understanding when our mathematical model hits a wall is just as important as knowing how to push past it with better hardware.

Dev: Overall, "Identifiability Limits of Forced Oscillation Sources in Power Systems" gives us the tools to diagnose exactly where a source localization effort is failing, whether it's due to the underlying structure or just because our measurements are too noisy.

Kai Sun

University of Tennessee

eess.SY, cs.SY, eess.SP, math.DS

Submitted: 2026-09-30

Updated: 2026-09-30

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 92/100

The gist: Whether a forced-oscillation source can be uniquely localized depends jointly on the available measurements and the candidate intervention dictionary.

Key concepts

Projective signature
This is the amplitude- and phase-invariant signature of a candidate source. It represents the source as a point in complex projective space, which is how the paper mathematically defines what makes two sources distinguishable or ambiguous.
Descriptor rank test
This test provides a specific mathematical condition for exact localization. It requires that the rank of certain matrices related to candidate sources must equal n + 2 for every pair of candidates, ensuring they are uniquely identifiable.
Weighted residual minimization
This is a method used to find the best estimate for an unknown forcing coefficient ($\alpha$). It minimizes the difference between the measured response and the predicted response, using a weighting matrix to prioritize certain measurements.

Terminology

Summary

Whether a forced-oscillation source can be uniquely localized depends jointly on the available measurements and the candidate intervention dictionary.

The gist

Exact single-frequency identifiability is possible if and only if all candidates in K are nonzero and pairwise projectively distinct.

Source Definition and Problem Formulation

Source location is defined relative to a specific physical intervention in a fixed component realization, where an oscillatory input perturbs a declared physical quantity. The problem is treated as an identifiability problem: A single sinusoid is assumed to enter one unknown member of a finite candidate dictionary. Each candidate intervention k perturbs a declared physical quantity as qk(t) ← qk(t) + s(t), where s(t) = a cos(ωt + ϕ). The central observation is that the unknown amplitude and phase multiply each candidate’s measured harmonic response by an arbitrary nonzero complex scalar, meaning a candidate is represented not by one vector but by a one-dimensional complex subspace, or projective ray.

Fundamental Identifiability of a Harmonic Source

The paper derives necessary and sufficient conditions for exact localization based on projective signatures. Definition 2 defines the Projective signature as the amplitude- and phase-invariant signature of candidate k, denoted [gk] = gk(ω)α, which is a point in complex projective space CPm−1. Theorem 1 states: All candidates in K are uniquely identifiable from the noise-free steadystate phasor yb for every nonzero forcing amplitude and phase if and only if gk(ω) ≠ 0, ∀k ∈ K, rankC [gk(ω)] gl(ω) = 2, ∀k ≠ l. This condition is equivalent to the Descriptor rank test, which requires rank jωE − A −bk bl C dk −dl = n + 2 for every k ≠ l.

Projective-Signature Matching and Robustness

The paper develops a framework for localization using weighted residual minimization. For candidate k, the weighted residual over the unknown complex forcing coefficient α is defined as Jk = minα∈C∥yb− gkα∥2W, where W is a positive-definite weighting matrix. The localization algorithm involves estimating the dominant forcing frequency ωb and then evaluating Jk for each candidate; Return bk only when the residual margin and model-validity tests pass; otherwise return Kbτ. Separation is quantified by the weighted projective coherence µWkl = gHk W glq (gHk W gk)(gHl W gl), and separation δWkl = q 1 − (µWkl)2, where δWkl = 0 indicates exact projective ambiguity.

Four-Step Diagnostic Framework and Failure Classes

The paper proposes a four-step framework to assess source-location failures:

  1. Raw identifiability: determines if candidate rays are equivalent or distinct.

  2. Robust separation: assesses whether dWk→l is large enough relative to uncertainty; "Nearly collinear signatures, 0 < δWkl ≪ 1, are identifiable in principle but may exchange residual rankings under modest measurement or model error."

  3. Feature preservation: determines if the selected feature preserves raw-data distinctions; feature-projection loss occurs when Yk∩Yl = ∅ but Zk ∩ Zl ≠ ∅.

  4. Mechanism-consistent interpretation: determines if the retained feature uses variables appropriate to the active mechanism.

The four failure classes are: mechanism mismatch, feature-projection loss, structural nonidentifiability, and poor conditioning or model error. The paper concludes that Correctly locating the source bus does not by itself establish unique identification of the internal forcing channel or robustness to measurement and model uncertainty.

Numerical Studies

Numerical studies on the Kundur two-area system and IEEE–NASPI Contest models illustrate these concepts. These studies demonstrate well-separated recovery, measurement-induced ambiguity and its removal, mechanism-dependent terminal decisions, and sensitivity of nearly collinear signatures to model perturbations. Case 3 in the IEEE–NASPI study exemplifies measurement-induced structural nonidentifiability, which is removed by adding measurement information. The results confirm that while raw data may be distinguishable, terminal processing can reduce source separation or alter the final decision, as shown when a feature more consistent with the voltage–var mechanism improves the decision. Furthermore, poor conditioning is distinguished from exact ambiguity: "Candidate rays are distinct but close, 0 < δWkl ≪ 1," leading to practical ambiguity under small perturbations.

Conclusion

Exact source localization is possible if and only if all admissible candidate signatures are nonzero and pairwise projectively distinct. The framework establishes when source localization is fundamentally possible from the available harmonic measurements and distinguishes this measurement-level limit from failures introduced by poor conditioning, feature extraction, or mechanism-dependent interpretation.

Improvements for AI systems

Based on the provided research paper, here are specific improvements for AI systems, categorized by the capabilities they would gain:

) Improved AI System Capabilities: Source Localization and Diagnostic Framework

The core improvement is transforming current black-box oscillation diagnosis into a rigorous, information-theoretic source localization engine. The improved system can perform the following functions:

Source Identification with Uncertainty Quantification (UQ):

The system moves beyond simply declaring a bus or unit as the source; it outputs a probabilistic estimate of the location based on the projective ray geometry in complex measurement space.

  1. Robustness to Model Mismatch and Measurement Noise:

The AI can distinguish between structural nonidentifiability (where candidates are mathematically equivalent) and poor conditioning/model error (where candidates are numerically distinct but too close to be reliably separated). It will output a confidence score based on the separation margin against uncertainty bounds.

  1. Mechanism-Consistent Interpretation:

Instead of relying on generic features like standard Dissipating Energy Flow (DEF), the AI can select and weight features (like the proposed in-phase V–Q measure, or CDEF) that are mathematically proven to be invariant under certain physical mechanisms (e.g., reactive power vs. mechanical torque). This ensures that a diagnosis is consistent with the underlying physics of the oscillation.

  1. Failure Mode Diagnosis:

The system can automatically classify why a localization attempt failed by running a four-step diagnostic framework:

  • It checks for raw identifiability (are the rays distinct?).

  • If not, it checks for robust separation (is the distance between rays large enough relative to uncertainty?).

  • If separation is borderline, it tests feature preservation (does the chosen feature map collapse two distinguishable raw data sets?).

  1. Adaptive Sensor Selection:

The system can propose optimal measurement augmentations by maximizing the worst-case projective separation while maintaining adequate sensitivity—a criterion superior to simple modal observability.

) Specific AI System Outputs and Actions:

The improved AI system, leveraging this paper's theory, can execute the following specific actions:

Detect No-Go Scenarios for Localization:

If the initial identification step (Raw Identifiability) fails (i.e., two candidate rays are exactly collinear), the AI will immediately halt and report that localization is impossible without additional information, preventing a false positive diagnosis.

  1. Quantify Ambiguity Level:

For borderline cases where separation is poor but not zero, the system will output a quantifiable Robust Separation Index (based on Proposition 3). This index allows operators to understand the practical ambiguity—whether it's an exact mathematical identity or merely a sensitivity issue due to noise/model error.

  1. Select Optimal Feature Mapping:

When presented with raw measurements, the AI can dynamically switch between different feature representations (e.g., DEF vs. CDEF vs. Q–V measures) and select the one that maximizes the margin against competing hypotheses, ensuring the chosen diagnostic tool is physically appropriate for the oscillation type detected.

  1. Generate Mechanism-Informed Alerts:

If a terminal feature yields an ambiguous result (as shown in Case 7), but a mechanism-consistent feature (like Q–V) provides a clear distinction, the AI will flag this discrepancy and recommend switching to the more informative representation, even if both methods ultimately point to similar bus locations.

  1. Provide Sensor Augmentation Recommendations:

When feature projection loss is detected (where raw data is distinct but features collapse), the system will suggest specific measurements (e.g., Augment with current angle measurements on branch X) that are proven in the research to restore distinguishability, effectively guiding physical hardware upgrades based on theoretical needs.

Related papers