On Identification of Distribution Grids

I. INTRODUCTION

• Distribution models are often unavailable or outdated, motivating identification of model parameters and topology from telemetry.Synchrophasor measurements provide voltage and current data across the network for this purpose.Source: §I · p. 1

II. PROBLEM FORMULATION

• The inverse power flow problem recovers a poly-phase system’s admittance matrix from voltage and current phasors, assuming sensors at all nodes.The paper considers both loopy and radial systems.Source: §I · p. 1, A. Preliminaries · p. 2

A. Preliminaries

• A poly-phase distribution system is modeled as a graph of nodes connected by energized lines, with node-specific phases and phasor measurements.The system’s bus admittance matrix relates nodal voltages and injected currents.Source: A. Preliminaries · p. 2

I. INTRODUCTION

• Low-rank voltage measurements can make full admittance identification ill-posed; the paper proposes methods to recover identifiable matrix structure and detect and localize events that change it.The introduction also describes a matrix-decomposition algorithm for identifying a large submatrix.Source: A. Preliminaries · p. 2, §I · p. 2

B. Sparsity-based Regularization

• When the sample size is smaller than the number of unknown admittance entries, sparsity-based regularization exploits the matrix’s encoding of network topology.The paper assumes a well-posed identification problem and that all nodes can be monitored for its robust recovery algorithm.Source: B. Sparsity-based Regularization · pp. 2–3
• The l0 cardinality-constrained formulation is NP-hard, so the paper uses a convex lasso relaxation with a positive regularization parameter.The lasso shrinks coefficients toward zero as the parameter increases, and cross-validation can select it.Source: B. Sparsity-based Regularization · p. 3
• The symmetric admittance matrix can be represented by its lower-triangular entries, reducing the number of parameters to estimate.A bijective mapping collects those entries and reconstructs the vectorized matrix.Source: B. Sparsity-based Regularization · p. 3

C. Avoiding Unnecessary Bias

• Equal lasso penalties can bias large admittance estimates and impair identification when switches or voltage regulators have much larger admittance than distribution lines.The paper notes that the lasso can yield suboptimal estimates in some cases.Source: C. Avoiding Unnecessary Bias · p. 3
• The adaptive lasso uses a two-stage, data-dependent weighted l1 penalty that applies less shrinkage to variables estimated to be large.Its initial estimator supplies weights for the second-stage fit.Source: C. Avoiding Unnecessary Bias · p. 3
• The adaptive lasso outperforms the lasso in identification accuracy on the test feeders.The paper uses two-dimensional cross-validation to tune its penalty parameters.Source: C. Avoiding Unnecessary Bias · p. 3

D. Exploiting Additional Structure

• Circuit upgrades and reconfiguration can leave the available distribution model obsolete for diagnostics and control applications.The paper considers leveraging an approximate model to improve identification accuracy.Source: D. Exploiting Additional Structure · p. 3
• The available admittance matrix is modeled as the true matrix plus a symmetric error matrix that must be identified.If the error entries are small, the paper uses regularized least squares; if only a few entries are nonzero but may be large, it notes an l0-norm alternative.Source: D. Exploiting Additional Structure · pp. 3–4
• The paper uses a ridge regression problem to identify the error matrix and states that this approach can periodically update the distribution system model.The regularization parameter is a tuning parameter.Source: D. Exploiting Additional Structure · p. 4

III. LOW RANK STRUCTURE OF DISTRIBUTION SYSTEMS

• Low-rank voltage measurements make identifying the entire admittance matrix ill-posed, even when K ≫ N.The proposed algorithm addresses this by partitioning the voltage measurement matrix, with only some admittance components recoverable.Source: §III · p. 4

A. Similarity Transformation

• The similarity transformation partitions the voltage measurements into V2, containing R independent rows, and V1, containing the remaining rows in V2’s row space.Here, R is the row rank of the voltage measurement matrix.Source: A. Similarity Transformation · p. 4
• Because V2 has full row rank, its pseudoinverse estimates the basis X relating the two partitions through V1 = XV2.The basis is estimated from µPMU data as X = V1V†2.Source: A. Similarity Transformation · p. 4

B. Recovering Parts of Ybus

• The basis-selection procedure uses orthogonal-triangular decomposition, sorts the triangular matrix’s diagonal elements, and selects those exceeding a threshold.The selected indices identify linearly independent rows of the voltage measurement matrix.Source: B. Recovering Parts of Ybus · p. 4
• The method estimates Y1,1 and Y2,2 by enforcing sparsity in a regularized optimization problem.The penalty weights are 1 for lasso or 1/|x̂i|^γ for adaptive lasso, with γ > 0.Source: B. Recovering Parts of Ybus · p. 4
• Y1,2 is then estimated by least squares, but errors in Y2,2 can propagate into that estimate.The paper reports that Y2,2 is accurately estimated in all cases, whereas Y1,1 and Y1,2 are not recovered with sufficient accuracy under the low-rank structure.Source: B. Recovering Parts of Ybus · p. 4

IV. TIMELY DETECTION AND LOCALIZATION OF EVENTS

• The online algorithm detects admittance changes with little data and a low false-alarm rate, supporting quasi-real-time remedial action.Switching, tap operations, arc faults, and ground faults can change the effective admittance between nodes.Source: §IV · p. 4
• The method estimates the event time and the changed admittance entries using as few successive phasor measurements as possible.The updated entries indicate the event type and approximate location; a pattern of two changing entries can indicate one switch opening while another closes.Source: §IV · p. 5

A. Event Detection

• When current-prediction error exceeds a predefined threshold, the method declares an admittance-matrix change.It predicts injected current from the known admittance matrix and measured voltages, then compares the prediction with measured current.Source: A. Event Detection · p. 5

B. Event Localization

• After detection, localization uses the pre-event admittance matrix and assumes only a small number of its entries change.This avoids collecting and processing the considerable post-event µPMU sample count required by full re-identification in a large network.Source: B. Event Localization · p. 5
• Because the difference between pre- and post-event admittance matrices is sparse, weighted ℓ1 regularization recovers the changed matrix entries as a convex optimization problem.The optimization can be solved efficiently with fewer µPMU samples than the original identification algorithm.Source: B. Event Localization · p. 5
• The required sample count is tied to Spark(A) exceeding twice the number of admittance elements changed by the event.This bound is based on the stated uniqueness condition for sparse recovery.Source: B. Event Localization · p. 5

• The simulator takes a distribution model, household demand profiles, customer connection nodes, and scheduled events as inputs.Source: §V · p. 5
• Automatic transformer-tap control is disabled so the admittance matrix changes only when a simulated event triggers a change.Source: §V · p. 5
• The integrated framework uses OPENDSS power-flow analysis to generate µPMU measurements and convex programs to recover and update the admittance matrix.Recovered matrices are compared with the true sequence to quantify estimation error.Source: §V · pp. 5–6, Figure 1 · p. 6

A. Distribution Feeders

• The four IEEE feeders are unbalanced radial systems that differ in size and sparsity; their admittance matrices are rank deficient, extremely sparse, and contain unusually large elements.The table describes feeder properties including node count, rank, sparsity, and largest admittance element.Source: A. Distribution Feeders · p. 6, Table I · p. 6
• Switch configurations provide event-testing cases: the modified 13-bus feeder has two feasible radial structures, while 34-bus and 37-bus feeders test identification without switches.The 123-bus feeder has 12 operable switches for testing event detection and localization.Source: A. Distribution Feeders · p. 6

B. Residential Loads & Phasor Measurements

• The 13-, 34-, and 37-bus cases use ADRES household data, whereas the 123-bus case uses synthesized loads from Ontario household measurements.Loads are aggregated at monitored nodes; the 123-bus case assumes a 95% power factor and nonuniform phase loading.Source: B. Residential Loads & Phasor… · pp. 6–7

C. Results

• One-day simulations use 1-second slots to estimate Y2,2 with lasso and adaptive lasso penalties and assess both estimation and event tracking.Source: C. Results · p. 7

1) Identification with Precise Measurements:

• On the 13-bus feeder, both penalties estimate all but the substation element with <1% relative error; substation knowledge yields M1, M2 < 0.01 for both.On the 34- and 37-bus feeders, adaptive lasso estimates large admittance elements that lasso fails to identify accurately.Source: 1) Identification with Precise… · p. 7

2) Identification with Noisy Measurements:

• For the 13-bus feeder, adaptive lasso remains accurate through σ ≤ 10^-3, while lasso yields sufficiently accurate estimates only through σ ≤ 10^-4.Both methods are sensitive to noise; above σ = 10^-3, adaptive lasso fails to identify several elements and both error metrics increase significantly.Source: Figure 3 · p. 7, 2) Identification with Noisy… · p. 7

3) Event Detection and Localization:

• Both 13-bus events and the 123-bus switching event are detected in the occurrence time slot; 13-bus changes are localized to the operated switches.For 13-bus switching, 23 µPMU samples give M1 = 0.17 and M2 = 0.10; for 123-bus switching, M1, M2 < 0.1 using a small number of samples.Source: Figure 4 · p. 8, 3) Event Detection and Localization · p. 8

VI. CONCLUSIONS

• µPMU installations remain limited in low-voltage networks, hidden states remain an identification challenge, and computation may be expensive for networks with thousands of nodes.The paper identifies combining smart-meter and µPMU data, handling hidden states, and distributed identification as future work.Source: §VI · p. 8