{"id":"5f7015cc-2d3a-49a5-8370-24643da05514","arxiv_id":"2505.12302","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SenseFlow combines virtual-node attention, slack-node gating, and self-ensembling iterative refinement to reduce power-flow estimation error on IEEE 39, 118, and 300 bus systems.","lead":"This paper presents SenseFlow, a graph neural network that estimates voltages and phase angles in power grids by iteratively refining its own predictions and using a physics-based error term. It reports lower errors than several generic graph network baselines on IEEE test systems and faster contingency analysis than a traditional solver.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SeIter's iterative feedback, the largest ablation gain, has no contraction or distribution-shift analysis; Fig. 4(a) shows only average RMSE on 39-bus, so per-sample or OOD error amplification is not ruled out.","rationale":"The reader's weakest-assumption identification is essentially correct: the SeIter feedback loop is the largest ablation factor and the paper lacks a convergence or distribution-shift analysis. I refine it in one respect: Fig. 4(a) does provide some empirical evidence that, on the 39-bus test set, average RMSE falls as loops increase, so the attack cannot be simply 'no loop study at all.' The precise gap is that only the average is shown, and the test topologies are drawn from the same perturbation distribution used in training, so per-sample divergence and realistic out-of-distribution topologies remain untested. If the learned update is non-contractive on such inputs, the reported order-of-magnitude gains from SeIter would not transfer. The paper has genuine positive features: the equation loss uses deterministic power-balance residuals rather than fitted targets, so I see no circular reasoning; and the held-out-topology split is a reasonable first step. I did not find an internal inconsistency that would force rejection. Secondary weaknesses, which do not change the verdict, include missing comparisons to the cited SOTA power-flow GNNs, no error bars or seeds, no released code, and the speedup claim in Fig. 4(c) not matching the stated 3-5x for all systems (39-bus is 1.56x, 300-bus 2.45x). These strengthen the conditionality but do not move it; the reader's CONDITIONAL verdict remains appropriate.","tokens_in":11251,"tokens_out":8937,"duration_ms":97536,"concrete_test":"Use the trained SenseFlow checkpoint on a genuinely shifted test set not generated by the Sec. 4.1 procedure, e.g., IEEE 57-bus or IEEE 39-bus with 3-4 simultaneous line outages and load perturbations outside [0.5, 1.5] p.u. For each test sample, record the error vector after each of the 8 SeIter loops and compute the per-sample contraction ratio r_eta = ||e_{eta+1}||_2 / ||e_eta||_2. If more than 5% of samples show r_eta > 1 for any eta, SeIter's stability on unseen topologies is not established; the abstract should then be restricted to 'average improvement on the training-like distribution' rather than 'diverse grid configurations.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central accuracy claim rests on the SeIter loop (Sec. 3.2): in the ablation, adding SeIter to FlowNet cuts PQVa RMSE from 0.0465 to 0.0061 on IEEE 39-bus (Table 3), the largest single gain in the paper. SeIter is a residual-style feedback loop: the current estimate is used to compute power-balance residuals (Eqs. 4-5), and FlowNet's own corrected estimate is fed back for the next loop via the EMA teacher (Eq. 6). For such an iteration to improve on unseen grids, the learned update operator must be approximately contractive on those inputs. The paper gives no convergence analysis for this operator, and the only loop-scaling evidence, Fig. 4(a), reports average RMSE versus loop count on the 39-bus test set. An average that falls can mask a minority of samples whose errors grow; moreover, the test topologies are generated under the same perturbation procedure as training (Sec. 4.1: 50-150% loads, 90-110% branch values, 1-2 line disconnections), so this is not a distributional-shift test. The paper's own Limitation paragraph concedes that real-world noisy/asynchronous inputs and data redundancy remain open. If the feedback operator is not contractive on a realistic shift, the reported iterative gains, and with them the SOTA claim, will not transfer.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SenseFlow, a graph-neural-network framework for power-flow estimation that combines a physics-informed loss (power-balance residuals) with a self-ensembling iterative refinement loop (SeIter). The architecture introduces a Virtual Node Attention module and a Slack-Gated Feed-Forward module to handle sparse grid connectivity and the special role of the slack bus. The authors evaluate on synthetic IEEE 39-, 118-, and 300-bus datasets, report ablations, demonstrate robustness to missing reactive-power values, and compare computational time for N-2 contingency analysis against a conventional solver. The central claim is state-of-the-art estimation accuracy across all three systems, with RMSE below 1e-3 for voltage magnitude and below 1e-2 for phase angle.","tokens_in":11523,"tokens_out":4109,"duration_ms":43488,"significance":"The combination of an externally evaluated power-balance equation loss with an iterative residual-feedback loop is a sensible and potentially useful direction for data-driven power-flow estimation. The paper also addresses two domain-specific issues that are often underexplored in GNN power-flow work: sparse global communication and the asymmetric role of the slack bus. The empirical gains over generic graph-convolution backbones are large, and the missing-Q experiments address a practically relevant failure mode of classical solvers. However, the significance as stated depends on the SeIter loop being stable and on the comparisons being made against appropriate power-flow-specific baselines, neither of which is currently established. The paper would be strengthened by reporting hyperparameters, variance estimates, and a more direct evaluation of the iterative operator's behavior under distribution shift.","major_comments":[{"comment":"The headline claim of state-of-the-art performance is not supported by the choice of baselines. All compared methods are generic graph-convolution layers (GraphConv, GINEConv, SageConv, ResGatedGraphConv, GatConv, TransformerConv), whereas the related work cites power-flow-specific and physics-informed alternatives such as PowerFlowNet [10], the typed-GNN method of Lopez-Garcia and Dominguez-Navarro [21], and the physics-guided approach of Hu et al. [7]. Without comparisons to these methods, the SOTA claim in Section 5 and the abstract overreaches. Please add these baselines or explicitly reframe the claim as improvement over generic GNN backbones.","section":"Section 4.3, Tables 1-2"},{"comment":"SeIter is the largest contributor to the reported gains (e.g., Table 3 shows PQVa RMSE dropping from 0.0465 to 0.0061 on IEEE 39-bus when SeIter is added), but the paper provides no convergence or contraction analysis for the feedback operator, and the only loop-scaling evidence is average RMSE versus loop count in Figure 4(a). An average that decreases can hide a minority of samples whose errors grow, and the test topologies are generated under the same perturbation procedure as the training data (Section 4.1). Therefore the claim that iterative gains will transfer to real grids is not established. Please report per-sample error distributions across loop counts and evaluate on topologies that are outside the 1-2 line-disconnection perturbation distribution.","section":"Section 3.2, Figure 4(a), Table 3"},{"comment":"The weighting hyperparameters for the two key training terms are never reported: the equation-loss weight lambda in Eq. (1) and the EMA momentum alpha in Eq. (6). These parameters directly control the physics-loss contribution and the self-ensembling dynamics, both of which are central to the method. Without reporting their values, the ablation attribution in Table 3 is not reproducible. Please report the chosen values and, ideally, a sensitivity analysis showing that the results are not critically dependent on them.","section":"Equations (1) and (6), Section 4.2"},{"comment":"No error bars, standard deviations, or number of random seeds are reported for any of the experimental results. Several comparative margins are extremely small (e.g., Table 1, IEEE 118-bus PQVm: SenseFlow 0.00009817 versus ResGatedGraphConv+SeIter 0.00007869), so the claimed superiority is not statistically supported. Please add variance information across multiple training runs and, where margins are small, a significance test.","section":"Tables 1-3 and Figure 4(b)"},{"comment":"The N-2 contingency timing comparison reports only wall-clock time for SenseFlow versus PyPower. It does not state the number of N-2 scenarios, the accuracy of SenseFlow on those scenarios, or whether critical violations are missed. Comparing wall-clock time of an approximate estimator to an exact solver without an accuracy or coverage metric conflates speed with correctness. Please clarify the experimental setup and report contingency-level accuracy (e.g., missed violations or error relative to the solver) alongside runtime.","section":"Section 4.4, Figure 4(c)"}],"minor_comments":[{"comment":"The summation index in Eq. (5) is written as i, but it should be j=1 to N, matching Eq. (4).","section":"Equation (5)"},{"comment":"The Slack-Gated Feed-Forward paragraph says it is taking the PV node as an example, but the equations use FPQ rather than FPV. Please clarify which node type the equations refer to.","section":"Section 3.3, Eq. (10)"},{"comment":"There is a typo in the Figure 3 caption: 'THe whole hetero-graph' should be 'The whole hetero-graph.'","section":"Section 3.3, Figure 3 caption"},{"comment":"The phrase 'inactive power' should be 'reactive power.'","section":"Section 4.4, paragraph on incomplete inputs"},{"comment":"The checkmark/row formatting of Table 3 is difficult to parse; the rows do not clearly indicate which components are active in each configuration. Please reformat the table so each row has an explicit configuration.","section":"Table 3"},{"comment":"The paper states that the code 'will be available' but does not provide a working repository at review time. Please include a link to the actual code repository in the revised version.","section":"Section 4.2"},{"comment":"It is not fully clear whether the self-ensembling teacher model is used only for generating inputs to the next training loop or also for inference. Please specify the inference-time procedure explicitly.","section":"Section 3.2, Figure 2"},{"comment":"The limitation paragraph appropriately concedes that noisy and asynchronous real-world inputs remain open, but this concession should also be reflected in the abstract and conclusions, which currently state that missing-input cases are handled robustly without qualification.","section":"Section 5, Limitation paragraph"}],"recommendation":"major_revision","confidential_remarks":"The paper's contribution is potentially useful, but the empirical validation currently targets generic GNN backbones rather than the power-flow-specific methods the authors cite, and key hyperparameters and variance estimates are missing. The SeIter stability concern is the most serious technical risk: the largest ablation gain rests on a feedback loop with no convergence evidence. I would be willing to reconsider after the authors add the missing baselines, report hyperparameters and error bars, and provide evidence about the behavior of the iterative loop outside the training perturbation distribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: SenseFlow is a competent, well-scoped engineering paper. The combination of virtual-node attention with slack gating and an EMA self-ensembling iterative loop is new for power flow estimation, and the reported gains are large (order-of-magnitude RMSE drops on IEEE 39/118/300). The physics loss is genuinely external: it uses the admittance matrix and current predictions, not the target voltages, so the circularity concern doesn't stick.\n\nWhat the paper does well: the architecture choices are motivated by real properties of power grids (single slack bus, sparse connectivity), the iterative refinement mechanism is clearly explained, and the N-2 contingency time comparison and missing-Q robustness are useful practical demonstrations. The ablation shows SeIter is the dominant contributor, which is consistent with the method's logic.\n\nThe soft spots are real but not fatal. First, the comparisons are only against generic GNN layers (GraphConv, GAT, etc.), not against the cited power-flow-specific GNNs like PowerFlowNet or the physics-guided DNNs in [7, 21, 23]. That weakens the 'state-of-the-art' claim. Second, there are no error bars, no seeds, and no reported values for the key hyperparameters lambda and alpha. The ablation table is formatted in a way that's hard to parse (the checkmarks don't map cleanly to rows). Third—and this is the stress-test concern—the SeIter loop has no convergence or contraction analysis, and the test set is generated under the same perturbation procedure as training. The average RMSE curve in Fig 4(a) falls with more loops, but that doesn't rule out per-sample error amplification on shifted topologies. This is a real limitation, but it doesn't invalidate the empirical claim on the tested distribution; the paper's own limitation paragraph admits real-world noise and async inputs remain open. So I'd call it a moderate concern, not a fatal one.\n\nWho gets value: researchers working on GNN-based power flow estimation will find this a useful baseline and a sensible template for combining iterative refinement with self-ensembling. It deserves serious peer review. I'd recommend conditional acceptance with required additions: comparisons to the power-flow-specific SOTA, multiple seeds with error bars, hyperparameter values, and a cleaned-up ablation table.","headline":"A competent engineering contribution with a plausible mechanism and strong reported gains, but the evaluation is narrower than the claims and the iterative loop's generalization is under-analyzed.","tokens_in":12098,"tokens_out":1807,"would_cite":false,"duration_ms":18122,"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":"SenseFlow claims that an iterative, physics-informed graph network estimates power flows on IEEE 39-, 118-, and 300-bus systems with voltage-magnitude RMSE below $10^{-3}$, phase-angle RMSE below $10^{-2}$, and N-2 contingency screening…","keywords":["power flow estimation","graph neural network","physics-informed loss","self-ensembling iterative estimation","virtual node attention","slack bus","contingency analysis","IEEE test systems"],"falsifier":"Train SenseFlow once, then run the SeIter loop on test grids with topologies far from the training distribution—for example, many simultaneous line outages or a partially re-meshed network—and record per-loop RMSE; if voltage or angle error rises with loop count instead of falling, the iterative refinement is overfitting to the training distribution of topologies rather than converging in general.","tokens_in":10973,"feed_emoji":"⚡","tokens_out":8016,"duration_ms":74389,"temperature":0.7,"pith_summary":"Power flow estimation—finding the voltage magnitude and phase angle at every bus—is a frequent, computationally heavy step in grid operations, and conventional solvers are slow when many contingency scenarios must be screened. SenseFlow claims that a graph neural network can reach high precision by treating power flow as an iterative correction process rather than a one-shot regression: a physics-informed network predicts small updates to the voltage state, a self-ensembling teacher built by exponential moving average refines the input for the next loop, and the power-balance equations supply both an extra loss and the mismatch features. On IEEE 39-, 118-, and 300-bus systems, the paper reports RMSE below $10^{-3}$ for voltage magnitudes and below $10^{-2}$ for phase angles, beating six graph-convolution baselines. The same framework keeps $10^{-3}$-level accuracy when reactive-power measurements are missing—a setting conventional methods cannot handle—and runs N-2 contingency analysis 3–5 times faster than a standard solver on the larger grids. The value, if the results hold, is a data-driven route to high-volume contingency screening and state estimation for larger, renewables-heavy networks.","feed_headline":"Iterative GNN cuts power-flow voltage error below 0.001","feed_subtitle":"On IEEE 39, 118 and 300 bus grids, it also screens N-2 contingencies 3–5x faster than a conventional solver.","key_machinery":"The load-bearing mechanism is the eight-loop SeIter correction cycle. FlowNet is a hetero-graph network—PQ, PV, and Slack buses are distinct node types—whose Virtual Node Attention (VNA) module builds a global node from average- and max-pooled features of all buses and uses cross-attention to broadcast that system-wide context back to every node without adding edges. Its Slack-Gated Feed-Forward (SGF) module concatenates each node's features with the slack node's features, passes them through a gated MLP, and adds a residual connection, so the phase-reference bus influences every prediction while local features survive. The equation loss $L_{\\text{equ}}$ computes $\\Delta P$ and $\\Delta Q$ from the predicted state and the admittance matrix and penalizes them, so the physical power-balance equations act simultaneously as a training signal and as input features for the next iteration.","core_discovery":"On the paper's terms, the central discovery is that replacing end-to-end regression with iterated self-corrected prediction lets a physics-informed graph network converge to voltage states accurate enough to compete with numerical solvers. In each of eight loops, FlowNet takes the current voltage estimates and the power mismatches $\\Delta P$, $\\Delta Q$ from the balance equations, predicts corrections, and is trained with both the ground-truth L1 loss and the equation loss; the EMA teacher then generates the estimates that seed the next loop. The architecture makes two deliberate choices about power-system structure: Virtual Node Attention pools features from all buses into a global node and attends it back to every bus, overcoming sparse-graph reachability, and Slack-Gated Feed-Forward gates the slack bus's features into every other bus, encoding its role as the system-wide phase reference. The complete system reaches RMSE below $10^{-3}$ in voltage magnitude and below $10^{-2}$ in phase angle across IEEE 39-, 118-, and 300-bus cases, and the ablations show the SeIter loop, not the architecture alone, is responsible for most of the improvement.","pith_inferences":["Not established by the paper: the test topologies, while strictly distinct from training topologies, are drawn from the same generation process; a genuinely different distribution of outages, load profiles, or parameter ranges could reveal whether the feedback loop stays stable outside its training manifold.","A transferable recipe suggested by the result: for physics-constrained regression on sparse graphs, feed the residual of the governing equations back as an input feature and stabilise iteration with an EMA teacher, instead of only penalising the residual in the loss.","A natural stress test the paper leaves open is cross-size generalisation—training on one IEEE system and evaluating on another—since VNA and SGF are size-agnostic while the GCN layers may not be.","The missing-Q robustness is demonstrated only for absent reactive power; missing or noisy active power, voltage setpoints, or topology entries would be a harder test of the framework's claim to handle incomplete information."],"forward_implications":["Wrapping any of the six tested graph-convolution backbones in the SeIter loop cuts phase-angle RMSE by roughly an order of magnitude, so the iterative self-correction is the dominant source of accuracy, not the particular convolution layer.","A grid operator with missing reactive-power measurements, where a conventional solver cannot form a well-posed problem, can still obtain voltage-state estimates at the $10^{-3}$ level.","Because N-2 contingency screening runs 3–5 times faster than a conventional solver on 118- and 300-bus systems, the practical cost of contingency analysis shifts from per-case solving to one-time model training per grid.","The equation-loss term keeps predictions consistent with active and reactive power balance, giving the network a physics-checkable property that pure regression models lack."],"supporting_citations":[{"why":"It is the conventional iterative solver that defines the accuracy and runtime baseline for the speedup comparison.","marker":"[4]"},{"why":"It is the message-passing graph-network baseline that this paper extends and must outperform.","marker":"[10]"},{"why":"It supplies the typed-graph formulation and dataset-generation conventions that the experiments build on.","marker":"[21]"},{"why":"It is the closest physics-embedded graph-convolution baseline for power flow under uncertain injections and topology.","marker":"[23]"},{"why":"It provides the physics-guided deep-network approach and the L1 supervised-loss precedent used in FlowNet.","marker":"[7]"},{"why":"It is the power-flow simulation tool used to generate the IEEE 39-, 118-, and 300-bus datasets.","marker":"[28]"},{"why":"It provides the GraphConv layer used inside FlowNet's VNA and as a baseline in the comparisons.","marker":"[26]"},{"why":"It provides the GAT layer used as a baseline and referenced in the VNA attention design.","marker":"[27]"}],"fun_headline_variants":["Self-ensembling iterative GNN beats solver accuracy on power flow","Iterative self-correcting GNN trims voltage error below 0.001","Physics-informed loop GNN reaches RMSE <0.001 on grid voltages","EMA-ensembled GNN iterations approach exact power flow solutions","Slack-aware GNN iteration hits <0.001 voltage RMSE"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"SenseFlow's reported accuracy depends on the assumption that when the model's own corrected voltage estimates are fed back as inputs for the next loop, errors do not compound on held-out topologies; the paper does not provide a convergence or out-of-distribution stability analysis of this feedback loop.","fun_headline_variants_meta":{"raw":{"variants":["Self-ensembling iterative GNN beats solver accuracy on power flow","Iterative self-correcting GNN trims voltage error below 0.001","Physics-informed loop GNN reaches RMSE <0.001 on grid voltages","EMA-ensembled GNN iterations approach exact power flow solutions","Slack-aware GNN iteration hits <0.001 voltage RMSE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000859,"raw_usage":{"total_tokens":3779,"prompt_tokens":1047,"completion_tokens":2732,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":663,"completion_tokens_details":{"reasoning_tokens":2633}},"tokens_in":663,"tokens_out":2732,"duration_ms":22250,"temperature":1.0,"reasoning_tokens":2633,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:36:20.096772+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train SenseFlow once, then run the SeIter loop on test grids with topologies far from the training distribution—for example, many simultaneous line outages or a partially re-meshed network—and record per-loop RMSE; if voltage or angle error rises with loop count instead of falling, the iterative refinement is overfitting to the training distribution of topologies rather than converging in general.","supporting_citations":[{"cited_title":"Developments in the newton raphson power flow formulation based on current injections","cited_arxiv_id":null,"evidence_quote":"It is the conventional iterative solver that defines the accuracy and runtime baseline for the speedup comparison."},{"cited_title":"Powerflownet: Power flow approximation using message passing graph neural networks","cited_arxiv_id":null,"evidence_quote":"It is the message-passing graph-network baseline that this paper extends and must outperform."},{"cited_title":"Power flow analysis via typed graph neural networks","cited_arxiv_id":null,"evidence_quote":"It supplies the typed-graph formulation and dataset-generation conventions that the experiments build on."},{"cited_title":"Physics embedded graph convolution neural network for power flow calculation considering uncertain injections and topology","cited_arxiv_id":null,"evidence_quote":"It is the closest physics-embedded graph-convolution baseline for power flow under uncertain injections and topology."},{"cited_title":"Physics-guided deep neural networks for power flow analysis","cited_arxiv_id":null,"evidence_quote":"It provides the physics-guided deep-network approach and the L1 supervised-loss precedent used in FlowNet."},{"cited_title":"Mat- power: Steady-state operations, planning, and analysis tools for power systems research and education","cited_arxiv_id":null,"evidence_quote":"It is the power-flow simulation tool used to generate the IEEE 39-, 118-, and 300-bus datasets."},{"cited_title":"Graph attention networks","cited_arxiv_id":null,"evidence_quote":"It provides the GAT layer used as a baseline and referenced in the VNA attention design."}],"review_version":1}