{"id":"be5ea0cd-8da6-4a4f-bd85-d45e4af9949c","arxiv_id":"1908.03753","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A settingless protection algorithm uses convex optimization to compare healthy and faulted line models against voltage and current measurements, identifying internal faults, their location, and their type.","lead":"This paper proposes a protection algorithm for medium-voltage power lines that decides whether the line has an internal fault by solving small convex optimization problems. It aims to remove manually tuned thresholds while also locating the fault and identifying its type.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fault-type identification is undermined by non-unique fault-resistance estimates: for K1 and K2 faults the measurement-matrix columns for unfaulted phases vanish, so H in Eq. (10) is only semidefinite and the recovered Rb/Rc are arbitrary.","rationale":"The reader's weakest_assumption—that the unpenalized MSE comparison lacks a complexity penalty—is a legitimate concern, but the paper's extensive normal-operation and external-fault simulations give substantial empirical evidence that the decision rule is secure in the tested regimes. The more load-bearing issue for the central claim is structural identifiability: for common fault types, some fault resistances do not enter the measurement equations at all, so the optimization problem cannot uniquely determine them. This directly affects the claimed fault-type identification and makes the reported resistance-error statistics difficult to interpret. The issue is internal to the paper's formulation: the assertion that H is positive definite in Section III-C is not guaranteed by Eq. (7). It can be settled by a straightforward rank test on the least-squares matrix and by re-solving the QP from different starting points. The protection and fault-location results are still valuable, and the identifiability gap is addressable by reformulating the fault-type decision or by reporting only identifiable parameters, so the verdict remains CONDITIONAL rather than ACCEPT or REJECT.","tokens_in":12699,"tokens_out":24975,"duration_ms":302383,"concrete_test":"Using the paper's noiseless simulation setup for a single-line-to-ground fault, assemble the B matrix such that W = c + B[Ra,Rb,Rc,Rg,alpha]^T from Eq. (7) and compute its singular values. Then re-solve the same QP (9)/(10) from different starting points or with tiny measurement perturbations; if the objective is flat along Rb/Rc and the returned resistances vary arbitrarily, the recovered RF cannot identify fault type. Additionally, report a fault-type confusion matrix for K1/K2/K2g/K3 using an explicit mapping from estimated Ra/Rb/Rc/Rg to fault type; if no such rule can be specified, or if single-phase faults are misclassified because Rb/Rc are arbitrary, the central 'fault identification' claim is not met.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-C claims that H in (10) is positive definite, but this is not guaranteed by the model. In W(RF, alpha), Eq. (7), only the second block depends on RF; for phase b the coefficient of Rb is -(I1b+I2b), for phase c the coefficient of Rc is -(I1c+I2c), and the Rg coefficients are -(I1a+I2a+I1b+I2b+I1c+I2c). At the fault node, KCL for an unfaulted phase gives I1p+I2p=0. Thus for single-line-to-ground faults (K1), the Rb and Rc columns of the least-squares measurement matrix are identically zero; for phase-to-phase faults, the unfaulted phase column is zero. Consequently H=B^T B is positive semidefinite, not positive definite, and the optimal x* is non-unique. A QP solver will return arbitrary Rb/Rc values, which cannot indicate 'open' phases, so the claimed fault-type identification from RF is not supported. The Table III/V resistance-error averages over all four components are also suspect when true values are infinite or unidentifiable. The protection decision may still be reliable, but the merged fault-identification functionality is not demonstrated as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a settingless time-domain unit protection algorithm for medium-voltage lines. The algorithm evaluates M+2 hypotheses for each observation window: a healthy line over the full window, a faulty line over the full window, and M mixed pre-/post-fault windows. Each faulty hypothesis is formulated as a small convex quadratic program in the fault resistances Ra, Rb, Rc, Rg and the fault location α, and the hypothesis with the smallest mean-squared error is selected. The optimized resistances and α are then used for fault type identification and fault location. The authors evaluate the algorithm on a Simulink implementation of a lumped R-L line model under normal operation, external faults, internal faults, converter-based generation, measurement noise, and line-parameter uncertainty, reporting 100% dependability and security over the tested scenarios and accurate fault-parameter estimates for SNR ≥ 60 dB and ±10% line-parameter errors.","tokens_in":12988,"tokens_out":7387,"duration_ms":84474,"significance":"The central idea—recasting protection as a model-selection problem solved by small convex programs with no empirical thresholds—is novel and potentially useful for MV time-domain protection. The authors also provide extensive simulation coverage of generation types and operating conditions, and the computational footprint appears compatible with relay time budgets. However, the paper's as-stated claim that the algorithm merges protection, fault location, and fault type identification is stronger than what the presented evidence supports: the fault-resistance identification problem is rank-deficient for common fault types, the model-selection rule has no complexity penalty, and the validation uses the same lumped R-L model that the algorithm assumes. These issues are load-bearing for the merged-functionality claim and need to be addressed in revision.","major_comments":[{"comment":"The claim that H is positive definite is not correct for typical fault types. For a single-line-to-ground fault, KCL at the fault node gives I1p+I2p ≈ 0 for the two unfaulted phases, so in the regression matrix implied by W(RF, α) in Eq. (7) the columns corresponding to Rb and Rc are near zero; for phase-to-phase faults the column for the unfaulted phase vanishes. Consequently H is only positive semidefinite, the optimal Rb/Rc are non-unique, and a QP solver will return arbitrary values for those components. This undermines the claimed fault-type identification from RF and makes the resistance-error averages in Tables III and V uninterpretable for components whose true value is infinite or unidentifiable. The protection decision may still be reliable, but the fault-identification function needs to be reformulated, for example by solving separate structured models for each fault type or by reporting only identifiable parameters.","section":"Section III-C, Eq. (10)"},{"comment":"The case-selection rule compares raw mean-squared errors without penalizing the larger number of adjustable parameters of the faulty model. The healthy model has no fault parameters, whereas each faulty/mixed case has five free parameters (Ra, Rb, Rc, Rg, α); under noise the more flexible faulty model can fit the data spuriously. The heuristic a1 := ∆1/∆3 and a2 := ∆4/∆1 introduced in Section III-D is not justified by any statistical model-complexity criterion. This is the most likely reason that the security results in Section IV-F, Figure 6, degrade sharply below SNR = 60 dB. The authors should introduce a principled penalty (e.g., AIC/BIC or a chi-square test with proper degrees of freedom) or otherwise justify the model comparison; otherwise the 'settingless' claim that no thresholds are needed is not supported.","section":"Section III-D"},{"comment":"The validation is circular in an important respect: the protected line in the Simulink model is represented with the same lumped series R-L equations, Eqs. (1)–(4), that the algorithm itself uses, with no shunt capacitance or distributed-parameter effects. The tests therefore demonstrate internal consistency between the algorithm and its own assumed model, not robustness to the dominant modeling errors of real MV lines. Since the final paragraph of the paper already notes that HV lines require more detailed modeling, the claims should be correspondingly limited for MV applications, or the algorithm should be tested on a higher-fidelity EMTP-type line model (e.g., frequency-dependent distributed-parameter) to establish that the model mismatch does not cause misoperation.","section":"Section IV, simulation setup"}],"minor_comments":[{"comment":"The text states that xmax provides an upper bound on RF, but the simulation setup sets xmax := [∞, ∞, ∞, ∞, 1]^T. Please clarify whether a finite bound is actually used in the reported tests; with truly infinite bounds and rank-deficient H, the QP can have unbounded or arbitrary solutions.","section":"Section IV-A"},{"comment":"For K1 faults, Table II sets Rb = Rc = ∞, so the reported small mean resistance errors cannot be reproduced without stating how errors are computed for components whose true value is infinite; please specify whether the average is taken only over finite fault resistances.","section":"Section IV-D, Table III"},{"comment":"The current reference directions are not fully specified; in particular, the appearance of (I1+I2) in Eq. (7) presumes a particular sign convention at the fault node. Please state the directions explicitly so that the identifiability discussion is unambiguous.","section":"Equations (3) and (4)"},{"comment":"Reference [1] contains garbled author names ('M. hrstrm', 'L. Sder', 'G. Andersson') and [15] says '72rd' instead of '72nd'; a careful copyedit of the reference list is needed.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is likely to attract interest from the protection community, but the rank-deficiency of the resistance estimation and the absence of a model-complexity penalty are easily missed because the reported simulation tables look clean. I would ask the authors to either fix these issues or explicitly narrow the fault-identification claim; otherwise the paper overstates what is demonstrated. No problematic citation patterns were found."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, this is a real step forward in settingless protection: the authors cast unit protection as a small convex model-selection problem over a healthy-line model and several faulted-line mixture hypotheses, and the same solve returns fault location, type, resistances, and inception interval. That specific formulation is new, even though settingless protection via dynamic state estimation goes back to Meliopoulos et al. Second, the paper's strongest non-protection claim\\u2014that fault type can be identified from the estimated resistances\\u2014has a genuine mathematical hole, and the protective decision is more credible than the fault-characterization part.\n\nThe simulation effort is substantial: one hundred random grid scenarios, thousands of fault instances, external and internal faults, converter-based generation, noise sweeps, and \\u00b120% line-parameter errors. For the protection function itself the evidence is persuasive: no false trips on external faults, no missed internal faults in the noiseless case. That part holds up.\n\nThe soft spots are real but not equal. The load-bearing one is in Section III-C: the paper claims H in (10) is positive definite, but that is not true for any fault type with an unfaulted phase. For a single-line-to-ground fault on phase a, the b and c phase currents into the fault sum to zero, so the corresponding columns of the measurement matrix are identically zero. The objective is flat in Rb and Rc, and the solver will return arbitrary values. Fault type identification from RF is therefore not demonstrated, and the resistance-error averages in Tables III and V (over 'infinite' true resistances) are misleading. This is not a minor typo; it undercuts a headline feature.\n\nThe second issue is the model-selection rule: comparing raw MSE between the healthy model (no adjustable parameters) and the faulted models (five adjustable parameters) invites overfitting. A BIC/AIC-style penalty or a test of the residual structure would be the obvious fix. The ad hoc a1/a2 heuristic for borderline healthy cases is a patch over this absence. Third, the simulator uses the same lumped R-L line model as the algorithm, so there is no true model-mismatch test; a distributed-parameter line model or field data would be the needed check. Finally, no code or data are released, and the noise section shows secure operation only at SNR \\u2265 60 dB\\u2014higher than most utility relays will see, though the authors do flag the optical-transformer assumption.\n\nWho should read it: anyone working on time-domain or settingless protection. It deserves a serious referee, but it is not ready as is. I would ask for a revision that fixes the identifiability analysis, adds a complexity penalty, reports resistance results only for identifiable parameters, and ideally tests against a mismatched line model. With those changes it could be a solid journal paper.","headline":"A genuinely new convex model-selection approach to settingless line protection with strong simulation work, but the fault-identifiability claim is undercut by a real linear-algebra gap and the model-selection rule lacks a complexity penalty.","tokens_in":13497,"tokens_out":4046,"would_cite":true,"duration_ms":44346,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A settingless protection algorithm decides whether a power line is healthy or faulted by solving convex optimization problems that fit measured waveforms to both line models, and it locates and classifies the fault in the same step.","keywords":["time-domain protection","settingless protection","line protection","fault location","fault type identification","convex optimization","medium-voltage lines","fault parameters"],"falsifier":"Run the algorithm on a large set of fault-free windows generated by the same grid model with Gaussian measurement noise near a 60 dB signal-to-noise ratio and count how often it declares an internal fault; because the faulted model has five free parameters, a nonzero false-trip rate would show that the raw error comparison needs a model-complexity penalty. A sharper test is to take a healthy-line window, deliberately mis-synchronize a few samples at one end, and check whether the faulted model absorbs the glitch and trips.","tokens_in":12482,"feed_emoji":"⚡","tokens_out":8925,"duration_ms":88912,"temperature":0.7,"pith_summary":"The paper tries to establish that a medium-voltage line can be protected without any empirically tuned set-points, by treating protection as a model-selection problem: given a short window of synchronized voltage and current measurements from both ends, the algorithm asks which of two line models, healthy or internally faulted, fits the data better. Each fit is obtained by solving a small convex optimization problem, and the faulted model's optimal parameters give fault location, fault type, resistances, and inception time in the same calculation. This would matter because threshold-based relays need re-tuning as grids change and can misoperate with converter-based generation, while a settingless scheme unifies protection with fault analysis. The reported simulations back the claim: correct decisions in all tested normal, external-fault, and internal-fault scenarios, including noise, line-parameter uncertainty, and converter-based sources.","feed_headline":"One optimization protects lines, locates faults, and identifies type","feed_subtitle":"Fitting measured waveforms to healthy vs faulted line models, the same algorithm can trip, locate, and classify a fault.","key_machinery":"The load-bearing object is the pair of line models in equations (1)-(4): a healthy lumped-parameter R-L line and the same line with a shunt fault represented by a $3\\times 3$ fault-resistance matrix $Z_F$ parameterized by $R_a,R_b,R_c,R_g$ and location $\\alpha$. Around these models the paper builds a model-selection procedure: for each hypothesis about the observation window it minimizes the mean squared mismatch of the model equations, which for fixed measurement matrices becomes a convex quadratic program $x^T H x + F^T x + d$ with bound constraints that can be solved to global optimality. This convexity is what lets one calculation both make the trip decision and estimate fault characteristics quickly enough for protection.","core_discovery":"The central claim is that a settingless time-domain unit protection algorithm can decide whether a medium-voltage line is healthy or internally faulted purely by comparing how well two physics-based models explain one window of synchronized measurements. The healthy model is the differential equation of a lumped R-L line; the faulted model is the same line with a fault at relative distance $\\alpha$ and resistance parameters $R_a$, $R_b$, $R_c$, $R_g$. The algorithm enumerates $M+2$ hypotheses: normal operation over the whole window, faulted operation over the whole window, and $M$ mixtures in which a suspected fault-inception interval slides across the window. For each hypothesis it computes the smallest mean squared residual of the model equations by solving a convex quadratic program in the fault parameters, then selects the hypothesis with the smallest error as the true state, using a ratio-based tie-break when the healthy case and the first mixture case are numerically close. The paper reports 100% detection of all simulated internal faults, no false trips for normal operation or external faults, accurate fault-location and resistance estimates, correct inception-interval identification, and stable performance with converter-based generation for signal-to-noise ratios down to roughly 60 dB and line-parameter errors up to about 10%.","pith_inferences":["An immediate testable extension would be to add a model-complexity penalty (e.g., an information criterion) to the mean squared error comparison; the paper's Section III-D heuristic $a_1$ vs $a_2$ suggests the authors already saw the need to break ties at low error levels, and a principled penalty could remove the heuristic altogether.","The same model-selection machinery could be applied to transmission lines, but only if the lumped R-L model is replaced by a distributed-parameter model, which would make the optimization non-convex; the paper marks this as future work.","The sliding inception-interval design implies a natural refinement: once a faulted window is found, run the mixture cases again with a finer grid over the selected interval only, which should improve inception-time estimates without increasing the original computational budget.","A practical deployment constraint not explored in the paper is sensitivity to synchronization errors between the two line ends; because equations (1)-(4) assume time-aligned samples, the 100 kHz sampling and optical-transformer requirements could be paired with a synchronization-error test to see how much misalignment the convex fit tolerates."],"forward_implications":["Because the same optimization problems yield both the trip decision and the fault parameters, a relay using the algorithm can send a trip signal and a fault report (location, type, resistances, inception interval) to operators at the same time.","Removing threshold set-points means the algorithm does not need to be re-tuned when grid conditions or fault characteristics change, which the paper shows by testing across different source types, noise levels, and line-parameter errors.","The algorithm's security against external faults and normal operation was maintained in all reported scenarios, so it can replace phasor-based relays that may be unreliable with converter-based distributed generation.","Within a 2 ms observation window and with communication-plus-propagation delays below 1 ms, the remaining computation budget of under 1 ms is achievable with dedicated hardware, so the approach is compatible with sub-cycle protection speeds."],"supporting_citations":[{"why":"Introduces the setting-less protection concept via dynamic state estimation, the lineage this algorithm extends.","marker":"[11]"},{"why":"Models healthy-line goodness of fit with a chi-square test; the baseline protection principle the paper's model-selection approach replaces.","marker":"[12]"},{"why":"Documents speed limits and hardware or communication needs of time-domain line protection, motivating 2 ms windows and sub-1 ms execution.","marker":"[2]"},{"why":"Provides a method for estimating line parameters, which the algorithm assumes as accurate inputs.","marker":"[13]"},{"why":"Provides another online line-parameter estimation method supporting the same assumption.","marker":"[14]"},{"why":"Supplies the convex quadratic solver used for the optimization problems in the model-selection step.","marker":"[19]"},{"why":"Establishes realistic optical current sensor noise levels, the basis for the SNR >= 60 dB operating range.","marker":"[21]"}],"fun_headline_variants":["Convex optimization unifies protection, fault location, and type ID","No setpoints: model-fitting algorithm protects and locates faults","Settingless unit protection with built-in fault analysis","Fault identification via optimization: protect, locate, classify"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The algorithm assumes that the model with the smaller average squared fitting error is the true state of the line, even though the faulted model has five extra adjustable parameters and can therefore fit noise and measurement glitches more easily.","fun_headline_variants_meta":{"raw":{"variants":["Convex optimization unifies protection, fault location, and type ID","No setpoints: model-fitting algorithm protects and locates faults","Settingless unit protection with built-in fault analysis","Fault identification via optimization: protect, locate, classify"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000652,"raw_usage":{"total_tokens":3004,"prompt_tokens":973,"completion_tokens":2031,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":1963}},"tokens_in":589,"tokens_out":2031,"duration_ms":15585,"temperature":1.0,"reasoning_tokens":1963,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:02:43.507846+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the algorithm on a large set of fault-free windows generated by the same grid model with Gaussian measurement noise near a 60 dB signal-to-noise ratio and count how often it declares an internal fault; because the faulted model has five free parameters, a nonzero false-trip rate would show that the raw error comparison needs a model-complexity penalty. A sharper test is to take a healthy-line window, deliberately mis-synchronize a few samples at one end, and check whether the faulted model absorbs the glitch and trips.","supporting_citations":[{"cited_title":"Setting-less protection: Feasibility study","cited_arxiv_id":null,"evidence_quote":"Introduces the setting-less protection concept via dynamic state estimation, the lineage this algorithm extends."},{"cited_title":"Dynamic state estimation based protection on series compensated transmission lines","cited_arxiv_id":null,"evidence_quote":"Models healthy-line goodness of fit with a chi-square test; the baseline protection principle the paper's model-selection approach replaces."},{"cited_title":"Speed of line protection - can we break free of phasor limitations?","cited_arxiv_id":null,"evidence_quote":"Documents speed limits and hardware or communication needs of time-domain line protection, motivating 2 ms windows and sub-1 ms execution."},{"cited_title":"Estimation of transmission line parameters using multiple methods","cited_arxiv_id":null,"evidence_quote":"Provides a method for estimating line parameters, which the algorithm assumes as accurate inputs."},{"cited_title":"Online optimal transmission line parameter estimation for relaying applications","cited_arxiv_id":null,"evidence_quote":"Provides another online line-parameter estimation method supporting the same assumption."},{"cited_title":"Gurobi optimizer reference manual","cited_arxiv_id":null,"evidence_quote":"Supplies the convex quadratic solver used for the optimization problems in the model-selection step."},{"cited_title":"Optical current sensors for high power systems: A review","cited_arxiv_id":null,"evidence_quote":"Establishes realistic optical current sensor noise levels, the basis for the SNR >= 60 dB operating range."}],"review_version":1}