FSNR for Data-Driven PMU Placement

• The paper introduces FSNR as a method that integrates forward selection with a local neighborhood refinement step to optimize PMU subset selection based on cross-validated SVM accuracy.
• It employs a greedy-plus-local-search strategy where candidate PMU locations are added sequentially and then locally corrected by exploring immediate upstream and downstream neighbors.
• Empirical results on IEEE-34 and IEEE-123 demonstrate that FSNR enhances fault location and fault type classification performance compared to both plain forward selection and alternative placement methods.

Forward Selection with Neighborhood Refinement (FSNR) is a greedy-plus-local-search subset selection procedure in which variables or sensors are added sequentially by a forward rule and then locally corrected by a refinement step. In current arXiv usage, the term denotes a data-driven PMU placement algorithm that selects a compact subset $Q \subseteq N$ to maximize cross-validated SVM classification accuracy for downstream 1D CNN fault diagnosis on distribution feeders (Khattak et al., 2 Sep 2025). A closely related methodological pattern appears earlier in Gaussian graphical model neighborhood recovery, where a forward-backward greedy algorithm alternates information-theoretic addition and pruning; that earlier method is not named FSNR, but it has the same “grow a candidate neighborhood, then refine it” structure (Yang et al., 2015).

1. Problem formulation and scope

In the PMU-placement formulation, FSNR addresses a subset selection problem over candidate PMU sites $N$ with a budget $m$ on the number of PMUs to select. The selected subset is denoted $Q \subseteq N$. The algorithm always includes the substation measurements, and then adds other PMUs greedily. The objective is not electrical observability alone; PMU placement is optimized data-driven, using cross-validated classification accuracy of an SVM trained on fault features derived from PMU measurements (Khattak et al., 2 Sep 2025).

The feature construction is based on differences between pre-fault and during-fault symmetrical components. The paper specifies

$$\Delta \mathbf{V}_{0,1,2} = \left[ \mathbf{V}_{0,1,2}^{\text{pre-fault}} - \mathbf{V}_{0,1,2}^{\text{during fault}} \right]_{\text{substation}},$$

$$\Delta \mathbf{I}_{0,1,2} = \left[ \mathbf{I}_{0,1,2}^{\text{pre-fault}} - \mathbf{I}_{0,1,2}^{\text{during fault}} \right]_{\text{substation}},$$

and for other buses,

$$\Delta \mathbf{V}_{0,1,2} = \left[ \mathbf{V}_{0,1,2}^{\text{pre-fault}} - \mathbf{V}_{0,1,2}^{\text{during fault}} \right]_{\text{node } i}.$$

These measurements contribute to a feature matrix

$\mathbf X^{(Q)} \in \mathbb{R}^{n\times 6|Q|},$

where the first 12 columns correspond to substation voltage/current real and imaginary components, and each additional bus contributes 6 columns.

The operational score of a candidate subset is the cross-validated classification accuracy of an SVM trained on $\mathbf X^{(Q)}$: $$\text{score}(Q) = \text{cross-validated classification accuracy of an SVM trained on } \mathbf X^{(Q)}.$$ The paper does not write an explicit combinatorial objective in standard form, but the algorithmic description makes clear that FSNR greedily maximizes this score subject to the PMU budget. This suggests a task-oriented interpretation of subset selection: the chosen PMU layout is evaluated only through its effect on a downstream discriminative model rather than through a purely structural or physics-only criterion.

2. Forward selection and neighborhood refinement procedure

FSNR begins with mandatory substation measurements: $N$0 The remaining candidate set is

$N$1

The features for every candidate $N$2 are normalized, and a baseline accuracy is computed: $N$3

The forward stage is standard greedy addition. While $N$4, the next PMU is chosen as

$N$5

The set is then updated: $N$6

The distinguishing feature of FSNR is the local correction step that follows greedy addition. If the selected set is large enough and the new score improved over the previous baseline, the paper imposes the conditions

$N$7

and

$N$8

The method then identifies $N$9, the PMU added in the previous iteration, and explores its immediate neighborhood, described in the text as its four immediate upstream and downstream neighbors: $m$0 If the best neighborhood replacement improves the score further,

$m$1

the replacement is accepted: $m$2

After each iteration, whether or not refinement occurred, the baseline is updated: $m$3 The loop stops when $m$4. In contrast to ordinary forward selection, FSNR therefore revisits a recent local decision through a restricted replacement search. The paper’s own summary of the distinction is that standard forward selection is purely greedy, whereas FSNR is “forward selection + local neighborhood replacement” (Khattak et al., 2 Sep 2025).

3. Scoring model and mathematical components

Inside FSNR, each candidate subset is scored by an SVM classifier. For a training set $m$5, with $m$6, the soft-margin SVM optimization is

$m$7

subject to

$m$8

The kernel is RBF: $m$9 and the decision function is

$Q \subseteq N$0

The classification problem is multiclass, and the paper uses one-vs-rest. For class $Q \subseteq N$1,

$Q \subseteq N$2

Prediction is made by

$Q \subseteq N$3

Operationally, the score used by FSNR is obtained by building $Q \subseteq N$4, training and testing an SVM with parameters $Q \subseteq N$5 using $Q \subseteq N$6-fold cross-validation, computing mean accuracy, and using that scalar as $Q \subseteq N$7 (Khattak et al., 2 Sep 2025). The paper does not define a separate closed-form scoring equation beyond cross-validated accuracy.

For comparison, the paper also defines a correlation-distance baseline. After column-wise $Q \subseteq N$8 normalization,

$Q \subseteq N$9

Sequence magnitudes are then formed: $$\Delta \mathbf{V}_{0,1,2} = \left[ \mathbf{V}_{0,1,2}^{\text{pre-fault}} - \mathbf{V}_{0,1,2}^{\text{during fault}} \right]_{\text{substation}},$$0 and stacked into

$$\Delta \mathbf{V}_{0,1,2} = \left[ \mathbf{V}_{0,1,2}^{\text{pre-fault}} - \mathbf{V}_{0,1,2}^{\text{during fault}} \right]_{\text{substation}},$$1

With Pearson correlation

$$\Delta \mathbf{V}_{0,1,2} = \left[ \mathbf{V}_{0,1,2}^{\text{pre-fault}} - \mathbf{V}_{0,1,2}^{\text{during fault}} \right]_{\text{substation}},$$2

the diversity score is

$$\Delta \mathbf{V}_{0,1,2} = \left[ \mathbf{V}_{0,1,2}^{\text{pre-fault}} - \mathbf{V}_{0,1,2}^{\text{during fault}} \right]_{\text{substation}},$$3

This baseline maximizes feature diversity rather than downstream classification accuracy.

4. Operational meaning of refinement and comparison with alternative placement rules

The neighborhood refinement stage is the central procedural difference between FSNR and plain forward selection. After a new PMU is chosen, the algorithm revisits the previously added PMU and checks whether one of its nearby buses yields a better subset score. The local neighborhood is explicitly described as the four immediate upstream and downstream neighbors of the previous PMU location (Khattak et al., 2 Sep 2025).

The stated intuition is that greedy forward search can make a locally good decision that is globally suboptimal. FSNR attempts to correct that by testing whether a neighboring bus can replace the previously selected PMU and produce higher cross-validated SVM accuracy. The method is therefore a lightweight local search rather than a global re-optimization. A plausible implication is that the algorithm preserves much of the simplicity of forward selection while reducing sensitivity to early myopic choices.

The paper compares FSNR against four alternatives: FS-SVM, which is forward selection without neighborhood refinement; FSNR-RF, which uses the same search but Random Forest scoring; a correlation-distance-based FSNR; and an admittance-matrix-based method from prior work. The reported takeaways are specific. FSNR improves over plain forward selection by correcting local greedy mistakes. SVM scoring outperformed RF scoring in both performance and scalability. The correlation-distance method selects diverse PMU locations, but not necessarily the ones best aligned with CNN fault discrimination. The admittance-matrix method is physically motivated but not optimized for CNN performance.

A common misconception in PMU placement is that observability-oriented criteria are sufficient for all downstream tasks. The paper explicitly departs from that assumption: PMU placement is not treated as an observability-only problem; instead, it is optimized for the downstream CNN fault diagnosis task. Within the reported experiments, this task-coupled perspective is what determines the final subset.

5. Integration with 1D CNN fault diagnosis and empirical performance

Once FSNR selects a PMU configuration $$\Delta \mathbf{V}_{0,1,2} = \left[ \mathbf{V}_{0,1,2}^{\text{pre-fault}} - \mathbf{V}_{0,1,2}^{\text{during fault}} \right]_{\text{substation}},$$4, that layout determines the input channels to a 1D CNN. The CNN receives time-series PMU features based on pre-fault and during-fault differences in symmetrical components, and the model is described as a VGG-11-inspired 1D CNN (Khattak et al., 2 Sep 2025).

Two classification heads are trained. Fault type classification has 11 classes: AG, BG, CG, AB, BC, AC, ABG, BCG, ACG, ABC, and ABCG. Fault location classification has 34 locations on IEEE-34 and 118 locations on IEEE-123. The training setup uses 60,000 samples for IEEE-34 and 80,000 samples for IEEE-123, with a 70% train, 15% validation, and 15% test split. The sampling rate is 60 Hz; 15 pre-fault and 15 during-fault samples are used; fault resistance is randomly varied from $$\Delta \mathbf{V}_{0,1,2} = \left[ \mathbf{V}_{0,1,2}^{\text{pre-fault}} - \mathbf{V}_{0,1,2}^{\text{during fault}} \right]_{\text{substation}},$$5 to $$\Delta \mathbf{V}_{0,1,2} = \left[ \mathbf{V}_{0,1,2}^{\text{pre-fault}} - \mathbf{V}_{0,1,2}^{\text{during fault}} \right]_{\text{substation}},$$6; and ground resistance is randomly from $$\Delta \mathbf{V}_{0,1,2} = \left[ \mathbf{V}_{0,1,2}^{\text{pre-fault}} - \mathbf{V}_{0,1,2}^{\text{during fault}} \right]_{\text{substation}},$$7.

The CNN loss is multiclass cross-entropy: $$\Delta \mathbf{V}_{0,1,2} = \left[ \mathbf{V}_{0,1,2}^{\text{pre-fault}} - \mathbf{V}_{0,1,2}^{\text{during fault}} \right]_{\text{substation}},$$8 Optimization uses 400 epochs, Adam optimizer, learning rate $$\Delta \mathbf{V}_{0,1,2} = \left[ \mathbf{V}_{0,1,2}^{\text{pre-fault}} - \mathbf{V}_{0,1,2}^{\text{during fault}} \right]_{\text{substation}},$$9, weight decay $$\Delta \mathbf{I}_{0,1,2} = \left[ \mathbf{I}_{0,1,2}^{\text{pre-fault}} - \mathbf{I}_{0,1,2}^{\text{during fault}} \right]_{\text{substation}},$$0, and exponential decay factor $$\Delta \mathbf{I}_{0,1,2} = \left[ \mathbf{I}_{0,1,2}^{\text{pre-fault}} - \mathbf{I}_{0,1,2}^{\text{during fault}} \right]_{\text{substation}},$$1.

The evaluation is conducted on modified distribution feeders with DERs. For IEEE-34, the listed DERs are PV at bus 840 via busPV1, 240 kW; WTG at 844 via buswind1, 250 kW; and DG at 890, 180 kW. For IEEE-123, the listed DERs are PVs at buses 35, 78, 64 and WTGs at buses 48, 95, 108. Data is generated in OpenDSS with Py-DSS, reflecting time-varying DER output.

The empirically selected PMU counts are where the accuracy curve saturates: 5 PMUs for IEEE-34 and 9 PMUs for IEEE-123. The selected PMU locations are 800, 848, busPV1, 850, 854 for IEEE-34, and 149, 83, 151, 300, 95, 250, 66, 56, 450 for IEEE-123. With these subsets, FSNR-SVM achieves the following reported CNN results:

• IEEE-34: fault location accuracy $$\Delta \mathbf{I}_{0,1,2} = \left[ \mathbf{I}_{0,1,2}^{\text{pre-fault}} - \mathbf{I}_{0,1,2}^{\text{during fault}} \right]_{\text{substation}},$$2; fault type accuracy $$\Delta \mathbf{I}_{0,1,2} = \left[ \mathbf{I}_{0,1,2}^{\text{pre-fault}} - \mathbf{I}_{0,1,2}^{\text{during fault}} \right]_{\text{substation}},$$3.
• IEEE-123: fault location accuracy $$\Delta \mathbf{I}_{0,1,2} = \left[ \mathbf{I}_{0,1,2}^{\text{pre-fault}} - \mathbf{I}_{0,1,2}^{\text{during fault}} \right]_{\text{substation}},$$4; fault type accuracy $$\Delta \mathbf{I}_{0,1,2} = \left[ \mathbf{I}_{0,1,2}^{\text{pre-fault}} - \mathbf{I}_{0,1,2}^{\text{during fault}} \right]_{\text{substation}},$$5.

The paper states that on IEEE-34, FSNR-SVM slightly outperformed plain forward selection and most alternatives in location accuracy, while for fault type the admittance-matrix-based method slightly edged it out, but only marginally. On IEEE-123, FSNR-SVM gave the best fault location and fault type results among the tested methods. The reported confusion patterns are also specific: most fault-location misclassifications occurred between single-phase lines sharing a common bus, and fault-type confusion was mainly between ABC and ABCG faults. Other fault classes were classified perfectly or nearly perfectly.

6. Relation to forward-backward neighborhood selection in Gaussian graphical models

A related methodological lineage appears in neighborhood selection for Gaussian graphical models. The paper “Efficient Neighborhood Selection for Gaussian Graphical Models” presents two heuristic algorithms: a forward-backward greedy algorithm for general Gaussian graphical models based on mutual information test, and a threshold-based algorithm for walk summable Gaussian graphical models (Yang et al., 2015). The first algorithm, titled “Finding Neighbors of Node $$\Delta \mathbf{I}_{0,1,2} = \left[ \mathbf{I}_{0,1,2}^{\text{pre-fault}} - \mathbf{I}_{0,1,2}^{\text{during fault}} \right]_{\text{substation}},$$6,” is the closest analogue to FSNR in the sense that it starts from an empty set and alternates forward addition, backward pruning, and a threshold-based stopping rule.

At iteration $$\Delta \mathbf{I}_{0,1,2} = \left[ \mathbf{I}_{0,1,2}^{\text{pre-fault}} - \mathbf{I}_{0,1,2}^{\text{during fault}} \right]_{\text{substation}},$$7, with current active set $$\Delta \mathbf{I}_{0,1,2} = \left[ \mathbf{I}_{0,1,2}^{\text{pre-fault}} - \mathbf{I}_{0,1,2}^{\text{during fault}} \right]_{\text{substation}},$$8, the forward step chooses

$$\Delta \mathbf{I}_{0,1,2} = \left[ \mathbf{I}_{0,1,2}^{\text{pre-fault}} - \mathbf{I}_{0,1,2}^{\text{during fault}} \right]_{\text{substation}},$$9

The geometric interpretation given in the paper is that $$\Delta \mathbf{V}_{0,1,2} = \left[ \mathbf{V}_{0,1,2}^{\text{pre-fault}} - \mathbf{V}_{0,1,2}^{\text{during fault}} \right]_{\text{node } i}.$$0 and $$\Delta \mathbf{V}_{0,1,2} = \left[ \mathbf{V}_{0,1,2}^{\text{pre-fault}} - \mathbf{V}_{0,1,2}^{\text{during fault}} \right]_{\text{node } i}.$$1 are projected onto the subspace spanned by $$\Delta \mathbf{V}_{0,1,2} = \left[ \mathbf{V}_{0,1,2}^{\text{pre-fault}} - \mathbf{V}_{0,1,2}^{\text{during fault}} \right]_{\text{node } i}.$$2, and their residuals $$\Delta \mathbf{V}_{0,1,2} = \left[ \mathbf{V}_{0,1,2}^{\text{pre-fault}} - \mathbf{V}_{0,1,2}^{\text{during fault}} \right]_{\text{node } i}.$$3 and $$\Delta \mathbf{V}_{0,1,2} = \left[ \mathbf{V}_{0,1,2}^{\text{pre-fault}} - \mathbf{V}_{0,1,2}^{\text{during fault}} \right]_{\text{node } i}.$$4 capture what remains after conditioning. The key lemma is

$$\Delta \mathbf{V}_{0,1,2} = \left[ \mathbf{V}_{0,1,2}^{\text{pre-fault}} - \mathbf{V}_{0,1,2}^{\text{during fault}} \right]_{\text{node } i}.$$5

where

$$\Delta \mathbf{V}_{0,1,2} = \left[ \mathbf{V}_{0,1,2}^{\text{pre-fault}} - \mathbf{V}_{0,1,2}^{\text{during fault}} \right]_{\text{node } i}.$$6

and similarly for $$\Delta \mathbf{V}_{0,1,2} = \left[ \mathbf{V}_{0,1,2}^{\text{pre-fault}} - \mathbf{V}_{0,1,2}^{\text{during fault}} \right]_{\text{node } i}.$$7. In the paper’s interpretation, forward selection therefore picks the node whose residual best explains the residual of $$\Delta \mathbf{V}_{0,1,2} = \left[ \mathbf{V}_{0,1,2}^{\text{pre-fault}} - \mathbf{V}_{0,1,2}^{\text{during fault}} \right]_{\text{node } i}.$$8.

The stopping rule tests whether the best empirical conditional mutual information is too small: $$\Delta \mathbf{V}_{0,1,2} = \left[ \mathbf{V}_{0,1,2}^{\text{pre-fault}} - \mathbf{V}_{0,1,2}^{\text{during fault}} \right]_{\text{node } i}.$$9 The threshold is linked to

$\mathbf X^{(Q)} \in \mathbb{R}^{n\times 6|Q|},$0

If the gain is below threshold, the algorithm stops and outputs the current neighborhood estimate.

Backward refinement then prunes likely false neighbors. The paper computes

$\mathbf X^{(Q)} \in \mathbb{R}^{n\times 6|Q|},$1

sets

$\mathbf X^{(Q)} \in \mathbb{R}^{n\times 6|Q|},$2

with

$\mathbf X^{(Q)} \in \mathbb{R}^{n\times 6|Q|},$3

and prunes

$\mathbf X^{(Q)} \in \mathbb{R}^{n\times 6|Q|},$4

The structural fact used for pruning is that if $\mathbf X^{(Q)} \in \mathbb{R}^{n\times 6|Q|},$5 contains all true neighbors of $\mathbf X^{(Q)} \in \mathbb{R}^{n\times 6|Q|},$6, then the zero entries of

$\mathbf X^{(Q)} \in \mathbb{R}^{n\times 6|Q|},$7

correspond to non-neighbors of $\mathbf X^{(Q)} \in \mathbb{R}^{n\times 6|Q|},$8. This establishes the same complementary logic found in FSNR-like methods: a forward stage that avoids missing true neighbors too early, and a backward stage that removes false positives introduced by greediness.

For the general GGM case, the paper proves structural consistency under a restricted-eigenvalue-type assumption. It assumes constants $\mathbf X^{(Q)} \in \mathbb{R}^{n\times 6|Q|},$9 and $\mathbf X^{(Q)}$0 such that for sparse vectors $\mathbf X^{(Q)}$1 with at most $\mathbf X^{(Q)}$2 nonzeros,

$\mathbf X^{(Q)}$3

with

$\mathbf X^{(Q)}$4

If thresholds are chosen appropriately, the nonzero entries of

$\mathbf X^{(Q)}$5

are bounded below, and the sample size satisfies

$\mathbf X^{(Q)}$6

then with probability at least

$\mathbf X^{(Q)}$7

the algorithm terminates in finite time and returns the exact neighborhood of node $\mathbf X^{(Q)}$8.

The same paper also develops a second, distinct method for the walk-summable case. Under $\mathbf X^{(Q)}$9-walk summability,

$$\text{score}(Q) = \text{cross-validated classification accuracy of an SVM trained on } \mathbf X^{(Q)}.$$0

it uses conditional covariance thresholding: $$\text{score}(Q) = \text{cross-validated classification accuracy of an SVM trained on } \mathbf X^{(Q)}.$$1 followed by selection

$$\text{score}(Q) = \text{cross-validated classification accuracy of an SVM trained on } \mathbf X^{(Q)}.$$2

The threshold is

$$\text{score}(Q) = \text{cross-validated classification accuracy of an SVM trained on } \mathbf X^{(Q)}.$$3

Pruning uses

$$\text{score}(Q) = \text{cross-validated classification accuracy of an SVM trained on } \mathbf X^{(Q)}.$$4

with entries below $$\text{score}(Q) = \text{cross-validated classification accuracy of an SVM trained on } \mathbf X^{(Q)}.$$5 removed. Exact neighborhood recovery is proved for

$$\text{score}(Q) = \text{cross-validated classification accuracy of an SVM trained on } \mathbf X^{(Q)}.$$6

with probability at least

$$\text{score}(Q) = \text{cross-validated classification accuracy of an SVM trained on } \mathbf X^{(Q)}.$$7

In historical terms, this earlier GGM work does not use the name “Forward Selection with Neighborhood Refinement,” but it provides a direct antecedent for the broader algorithmic idea: forward expansion of a candidate neighborhood followed by principled local pruning. The later PMU-placement FSNR instantiates the same high-level pattern in a different domain, replacing conditional mutual information and regression-style pruning with cross-validated SVM accuracy and local bus-neighborhood replacement.