{"id":"eea22ffa-5788-42fa-9243-e8b9b199f9aa","arxiv_id":"2506.03493","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A customized graph neural network with Gaussian mixture based missing feature handling estimates power system states from limited PMU data and stays accurate across topology changes and PMU failures.","lead":"This paper presents a graph neural network that estimates power system voltages and angles when only a few PMU sensors are available, using historical SCADA data to fill missing measurements. It reports lower estimation errors than conventional estimators under line outages, sensor failures, and non-Gaussian noise, with a theoretical bound on the effect of topology changes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3's advertised guarantee is not established for the actual architecture: the proof drops the recursive layer-difference term via an unproved 'localized effects' assumption and models constant-shift GCN layers, not the GMM-expectation and MH-GAT layers of Algorithm 1.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the proof of Theorem 3 depends on an unproved and internally contradicted localization step. I agree that this is the key soft spot. The advertised topology-adaptivity guarantee is a central part of the paper's contribution, and without it the mathematical explanation for why no retraining is needed is not established. The empirical comparisons in Tables I, II, and VI and Figures 5-9 are consistently favorable and may well be correct, so I do not see grounds for rejection. However, the theoretical claim should either be repaired, restricted to a simplified GCN, or relabeled as an output-stability result rather than an upper bound on SE error. The missing code, data, and error bars are additional reproducibility concerns, but they do not change the verdict: the paper remains conditional pending a corrected or qualified Theorem 3 and artifact release.","tokens_in":21622,"tokens_out":4479,"duration_ms":51839,"concrete_test":"Re-derive Theorem 3 for the exact stack in Algorithm 1: start x^(1) from Eq. (14), propagate layers 2,...,L-1 with Eq. (2), and finish with Eq. (18) where attention coefficients are recomputed after the topology/feature perturbation. Keep the recursive term in Eq. (B.8) instead of declaring it negligible, and check whether any eps-linear bound survives for the stated L=2 (or larger) architecture with C_sigma B f values from the trained model. If (19) requires extra assumptions such as C_sigma B f < 1 or a quantitative localization lemma, the abstract's 'upper bound on SE error' must be withdrawn or the theorem reproven.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Appendix B derives Eq. (19) for a network of plain GCN layers x^(l) = sigma(A x^(l-1) W^T) with a fixed adjacency A. The implemented CGNN-SE is different at both ends: layer 1 is the GMM expected ReLU of Theorem 2/Eq. (14), and the final layer is MH-GAT/Eq. (18), whose attention matrix A_k is recomputed from the input features and therefore changes when X -> X'. Neither layer satisfies the proof's update rule. Separately, the step from Eq. (B.8) to Eq. (B.10) requires the recursive difference term ||x^(l-1)_(:g) - x'^(l-1)_(:g)|| to be negligible. The paper justifies this only by citing 'localized effects' [48]; no quantitative lemma is proved. Section IV-C itself says robustness depends on 'the resulting shift in the power system's overall operating point,' i.e., a global quantity. If the recursion is kept, e_l <= C_sigma f (delta eps n_{l-1} + B e_{l-1}), which for C_sigma B f > 1 grows geometrically in L, so the claimed eps-linear bound (19) does not follow. Even if (19) were proven, it upper-bounds the change in network output between topologies, not the estimation error relative to the true system state that the abstract advertises. The empirical results are internally consistent and may be valid, but the mathematical guarantee for topology robustness is currently unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CGNN-SE, a graph neural network state estimator for PMU-unobservable power systems. A customized GCN first layer replaces missing PMU features by expected activations under Gaussian mixture models, followed by a multi-head graph attention (MH-GAT) layer and a linear readout for voltage magnitude and phase angle. The paper compares CGNN-SE against SCADA-SE, LSE, SVR, DNN-SE, GNN-SE, and WLM-SE on the IEEE 118-bus, 2000-bus Texas, and 1354pegase systems, under topology changes, PMU failures, bad data, and non-Gaussian noise. It also states Theorem 3, an upper bound on the change in estimator output under topology perturbation.","tokens_in":22027,"tokens_out":5396,"duration_ms":59905,"significance":"The empirical core of the paper is valuable and internally consistent. CGNN-SE reports lower magnitude MAPE and phase-angle MAE than all baselines with fewer PMUs, and the ablation in Table III isolates the benefit of the MH-GAT layer, while Table IV documents head-count sensitivity. The robustness figures for PMU failures and bad data correction are informative, and the paper is transparent that training is offline and online inference is fast. However, the advertised mathematical guarantee, Theorem 3 and Eq. (19), is not established for the implemented architecture, and the empirical evaluation consists of point estimates from a single simulator-based pipeline with no error bars or released artifacts. The contribution is therefore defensible as an empirical method paper, but the theoretical backing stated in the abstract and Section III-C needs substantial correction.","major_comments":[{"comment":"The proof of Theorem 3 models layers of the form x^(l) = sigma(A x^(l-1) W^T) with a fixed adjacency A, but Algorithm 1's CGNN-SE uses Eq. (14), the GMM expected ReLU, for the first layer and Eq. (18), MH-GAT, for the final layer. In the MH-GAT layer, each attention matrix A_k is recomputed from the input features, so A_k is not fixed when X changes to X'. Consequently, Eq. (19) does not bound the output change of the implemented architecture. Please either prove the bound for the actual layer updates used in CGNN-SE or explicitly scope Theorem 3 to the plain GCN backbone.","section":"§III-C and Appendix B"},{"comment":"The recursive term ||x^(l-1)_:g - x'^(l-1)_:g|| is dropped from the bound solely by citing 'localized effects' [48]. If that term is retained, the recursion scales as e_l <= C_sigma (delta epsilon n_{l-1} + B e_{l-1}), which can grow geometrically in the number of layers, so the epsilon-linear bound in Eq. (19) does not follow. Moreover, Section IV-C explicitly states that robustness depends on 'the resulting shift in the power system's overall operating point,' which is a global quantity and is in tension with the localized-effects premise. A quantitative localization lemma is needed before Theorem 3 can be accepted.","section":"Appendix B, Eq. (B.8) to Eq. (B.10)"},{"comment":"The abstract and Section III-C describe Eq. (19) as an upper bound on 'SE error,' but the proof bounds only ||Phi(A,X,W) - Phi(A',X',W)||, i.e., the change in the network output between two topologies. This quantity is not the estimation error with respect to the true system state, which is what the abstract advertises. Please either restate the guarantee as an output-sensitivity bound or extend the argument to bound the actual estimation error.","section":"Theorem 3 and abstract"},{"comment":"The variance propagation in Eq. (A.2) writes the variance of (AXW)_ij as the sum of a_in^2 s_nf w_fj^2 terms, which requires an independence or uncorrelatedness assumption across features and nodes. This assumption is not stated, and it is questionable for power-system state variables, where voltage magnitudes and phase angles at neighboring buses are strongly correlated. Because Eq. (14) is a load-bearing component of the proposed first layer, the assumption should be stated explicitly and justified, or the variance formula should be replaced with one that accounts for cross-node covariances.","section":"Appendix A, Eq. (A.2)"},{"comment":"All reported results are point estimates from a single simulator-based pipeline, with no error bars, confidence intervals, or multiple random seeds. The margin over GNN-SE in the 2000-bus phase-angle MAE is 0.064 versus 0.083 degrees, and the 1354pegase comparison in Table VI reports only a single sigma_y^2 value per method without stating the number of test scenarios or repeated runs. Please report means and standard deviations over independent training runs and test samples, and specify the exact test-set sizes, so that the claimed superiority can be assessed statistically.","section":"Section IV, Tables I-VI and Figs. 5-10"}],"minor_comments":[{"comment":"There is a typographical error: 'CGGN-SE' should be 'CGNN-SE'.","section":"Section I, last paragraph"},{"comment":"The word 'Lipshitz' should be 'Lipschitz' in the footnote to Eq. (B.4).","section":"Appendix B"},{"comment":"The 1354pegase experiment is underdescribed: please specify the PMU placement, noise model, number of test samples, and the training details that are 'chosen based on [39]' so that the comparison in Table VI is reproducible.","section":"Section IV-J and Table VI"},{"comment":"The notation Q^{-1}(alpha/2) and the decision rule in Eq. (20) should be defined more carefully; in particular, the roles of the two hypotheses and the threshold should be stated explicitly.","section":"Section III-D, Eq. (20)"},{"comment":"Reference [48] concerns localization in power-system security analysis, not localization of graph-signal perturbations in GNNs; a more direct reference or a self-contained lemma would make the argument easier to verify.","section":"Appendix B, reference [48]"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: CGNN-SE is a sensible and well-executed application of GMM-based expected activations to PMU-unobservable state estimation, and the empirical case is genuinely persuasive. The advertised mathematical guarantee, Theorem 3, does not survive a close read; it should be either repaired or substantially downgraded. The paper is still worth refereeing.\n\nWhat is actually new: the integration of the GMM expected activation from Taguchi et al. into the first GCN layer, followed by multi-head GAT, as an end-to-end architecture for PMU-unobservable SE. The components are known, but the combination for this problem is new, and the ablation shows the MH-GAT layer matters (error increases by about a factor of 10 without it). The empirical comparisons are broad: multiple PMU sets, single and multiple topology changes, PMU failures, non-Gaussian noise, bad data, and a 1354-bus comparison against a recent WLM estimator. The results are consistently favorable and the narrative is coherent.\n\nNow the soft spots, in proportion. Theorem 3 is the load-bearing flaw. The proof in Appendix B models layers as x^(l) = sigma(A x^(l-1) W^T) with a fixed adjacency A, but the actual architecture is different at both ends: layer 1 is the GMM expected activation of Theorem 2, and the final layer is MH-GAT whose attention matrices A_k are recomputed from the input features and therefore change when X changes. Neither layer satisfies the proof's update rule. Separately, the step from Eq. (B.8) to (B.10) requires the recursive difference term to be negligible, justified only by citing 'localized effects' [48] with no quantitative lemma. Section IV-C itself says robustness depends on 'the resulting shift in the power system's overall operating point,' a global quantity. If the recursion is kept, the bound grows geometrically in depth rather than linearly in epsilon. Even if (19) were proven, it upper-bounds the change in network output between topologies, not the estimation error relative to the true system state that the abstract advertises. The empirical results are internally consistent and may be valid, but the mathematical guarantee for topology robustness is currently unsupported.\n\nOther issues are addressable but should be fixed: no code or data released, no error bars or seeds reported, and the training and evaluation both rely on the same power-flow simulator, with GMMs fitted to solver outputs that also generate the labels. That is a self-consistency risk, not circularity, but it does mean the reported numbers could be optimistic.\n\nWho this is for: researchers working on GNN-based state estimation, especially those interested in handling missing PMU data with learned distributions. The empirical pipeline is a useful reference even if Theorem 3 is weakened. My recommendation: send to peer review, but the referee should push hard on the theorem. If the authors cannot prove it for the actual architecture, they should present it as a heuristic stability argument and adjust the abstract and conclusion accordingly. With that revision, the paper would be a solid contribution.","headline":"A useful GNN-SE application with strong empirical results, but Theorem 3's topology-change bound is not established for the actual architecture; worth refereeing with major revision.","tokens_in":22542,"tokens_out":2009,"would_cite":true,"duration_ms":20829,"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 graph neural network state estimator that replaces missing PMU measurements with Gaussian mixture distributions, claimed to stay accurate through topology changes and PMU failures without retraining.","keywords":["power system state estimation","graph neural networks","phasor measurement units","PMU unobservability","topology change robustness","multi-head graph attention","Gaussian mixture models","bad data detection"],"falsifier":"Take the trained 118-bus CGNN-SE, open an overloaded line that the paper classifies as worst-case, recompute the actual output difference against the base topology, and compare it with the right-hand side of (19); an observed difference larger than the bound for a single line outage would refute the topology-adaptivity guarantee.","tokens_in":1517,"feed_emoji":"⚡","tokens_out":2641,"duration_ms":94268,"temperature":0.7,"pith_summary":"This paper tries to establish that a graph neural network can estimate power system states even when only a minority of buses carry phasor measurement units, by replacing missing PMU data with distributions learned from historical data. It proposes CGNN-SE, which combines a Gaussian-mixture expected-activation graph-convolution layer with a multi-head graph attention layer, and reports that this estimator outperforms optimization-based and earlier learning-based estimators on several test systems. It also derives an upper bound on the change in the estimator output caused by a topology change. If the claim is right, time-synchronized state estimation would no longer require full PMU coverage or retraining after line openings and PMU failures.","feed_headline":"GNN state estimator stays accurate with sparse PMUs and line outages","feed_subtitle":"A custom graph layer fills missing PMU readings from learned distributions and outperforms optimization-based estimators after outages.","key_machinery":"The central mechanism is an expected-activation graph-convolution layer followed by a multi-head graph attention layer. In the first layer, each node's feature vector is treated as a Gaussian mixture, with known PMU readings entered as zero-variance means and missing readings entered as learned Gaussian-mixture parameters; Theorem 2 replaces the ReLU of the convolved Gaussian mixture with a closed-form expression using the auxiliary function $N_R(z)$, so missing-feature parameters and network weights are optimized together. The multi-head attention layer then assigns each neighbor adaptive importance, and Theorem 3 bounds the output difference after a topology perturbation in terms of the change in the adjacency matrix, layer count, feature width, and weight norms.","core_discovery":"The paper claims that a graph neural network can estimate the voltage magnitude and phase angle of every bus even when only a minority of buses carry phasor measurement units, by treating missing measurements as random variables modeled with Gaussian mixture distributions learned from historical data. Its proposed CGNN-SE computes the expected activation of a graph-convolution layer under those distributions, then passes the result through a multi-head graph attention layer, with all parameters trained end to end. In the paper's experiments on the IEEE 118-bus system, CGNN-SE with 11 PMU-equipped buses reaches 0.018% magnitude mean absolute percentage error and 0.027 degrees phase-angle mean absolute error, compared with 0.270% and 0.143 degrees for a linear least-squares estimator with 32 PMUs. On the 2000-bus Texas system it reports 0.093% magnitude error with 120 PMUs. The paper also derives an upper bound on the output difference after a topology change and presents this bound as a mathematical explanation of why retraining is unnecessary.","pith_inferences":["A natural next experiment, suggested by the proof rather than run in the paper, would open lines far from the PMU locations that trigger large re-dispatch; if the actual output difference grows faster than the bound in (19), the localized-effects assumption is the part that gives way.","The same expected-activation layer could carry over to distribution systems, where switch-status changes play the role of line openings and PMU coverage is even sparser; the architecture itself does not rely on transmission-level assumptions.","Because missing features are modeled as distributions, the Gaussian-mixture covariances could be used to output a confidence interval with each state estimate, a capability the paper does not develop.","The bound in (19) offers a design rule: keeping the architecture shallow and narrow should improve worst-case stability after topology changes, which the paper uses to justify hyperparameter tuning beyond raw validation accuracy."],"forward_implications":["PMU-only state estimation would no longer require full observability: the paper's 118-bus results use 11 PMUs where a fully observable least-squares estimator needs 32.","A trained estimator could keep operating without retraining after single and multiple line openings, including N-1 and N-2 contingencies.","Real-time PMU failures could be absorbed by substituting learned Gaussian-mixture parameters, with reported errors plateauing as more PMUs fail.","The estimator is reported to stay accurate under non-Gaussian measurement noise and, with a Wald-test preprocessing step, under bad data in up to 50% of input features.","The reported online computation time for the 2000-bus system is below the 33 ms PMU timescale, so the approach could run in real time."],"supporting_citations":[{"why":"Supplies the method of integrating missing-feature handling into graph convolution via expected activations, which the customized first layer builds on.","marker":"[32]"},{"why":"Provides the ReLU-of-Gaussian identity used in Theorem 1 and Theorem 2 to compute expected activations analytically.","marker":"[35]"},{"why":"Defines the GNN-SE architecture and its sensitivity to topology changes, which the proposed model extends and benchmarks against.","marker":"[28]"},{"why":"Provides the graph-signal stability framework and perturbation setup that Theorem 3 adapts for the topology-change bound.","marker":"[36]"},{"why":"Supplies the complex-variable weighted least modulus state estimator and the 1354pegase configuration used as the large-scale comparison.","marker":"[39]"},{"why":"Introduces the Wald test-based bad data detection and correction scheme used to pre-treat bad PMU data.","marker":"[37]"},{"why":"Establishes the DNN-SE baseline for PMU-unobservable systems and the practical implementation challenges considered in the study.","marker":"[14]"},{"why":"Provides the optimal PMU placement used to define the fully observable least-squares and GNN-SE baselines.","marker":"[43]"}],"fun_headline_variants":["GNN estimates power states with few PMUs, survives line outages","Graph network beats least squares on PMU-scarce grids","Deep graph net fills missing PMU data, robust to topology shifts","AI state estimator needs only 11 PMUs on 118-bus grid"],"cache_read_input_tokens":24576,"weakest_assumption_plain":"The proof of topology-change stability assumes the outage's effect on graph signals stays local so the recursive error term can be dropped, and if the outage instead shifts the system-wide operating point, the advertised bound does not follow.","fun_headline_variants_meta":{"raw":{"variants":["GNN estimates power states with few PMUs, survives line outages","Graph network beats least squares on PMU-scarce grids","Deep graph net fills missing PMU data, robust to topology shifts","AI state estimator needs only 11 PMUs on 118-bus grid"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000188,"raw_usage":{"total_tokens":1328,"prompt_tokens":939,"completion_tokens":389,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":313}},"tokens_in":555,"tokens_out":389,"duration_ms":4612,"temperature":1.0,"reasoning_tokens":313,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:02:41.731479+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained 118-bus CGNN-SE, open an overloaded line that the paper classifies as worst-case, recompute the actual output difference against the base topology, and compare it with the right-hand side of (19); an observed difference larger than the bound for a single line outage would refute the topology-adaptivity guarantee.","supporting_citations":[{"cited_title":"Processing of missing data by neural networks,","cited_arxiv_id":null,"evidence_quote":"Provides the ReLU-of-Gaussian identity used in Theorem 1 and Theorem 2 to compute expected activations analytically."},{"cited_title":"Time-synchronized state estimation using graph neural networks in presence of topology changes,","cited_arxiv_id":null,"evidence_quote":"Defines the GNN-SE architecture and its sensitivity to topology changes, which the proposed model extends and benchmarks against."},{"cited_title":"Complex-variable weighted least modulus state estimation,","cited_arxiv_id":null,"evidence_quote":"Supplies the complex-variable weighted least modulus state estimator and the 1354pegase configuration used as the large-scale comparison."},{"cited_title":"Deep neural network-based state estimator for transmission system considering practical implementation challenges,","cited_arxiv_id":null,"evidence_quote":"Establishes the DNN-SE baseline for PMU-unobservable systems and the practical implementation challenges considered in the study."},{"cited_title":"A PMU placement scheme ensuring real-time monitoring of critical buses of the network,","cited_arxiv_id":null,"evidence_quote":"Provides the optimal PMU placement used to define the fully observable least-squares and GNN-SE baselines."}],"review_version":1}