{"id":"1f4494eb-da77-4e3b-932d-99ca341a793b","arxiv_id":"2502.06683","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"OPF-aware data distillation can approximate optimal DER dispatches using a small subset of smart meter features, e.g., within about 5% error on a 1,136-bus feeder with 116 of 1,038 features, though reported errors are computed on training data.","lead":"This paper proposes selecting only a few smart meters to poll in real time, reconstructing the full load pattern, and then running a standard optimal power flow solver on the reconstructed data. A new OPF-aware reconstruction is trained so that the final DER dispatch stays close to the dispatch obtained from full data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central budget claims are in-sample fit errors: Figs. 2/3 and Table I evaluate f1/f2 on the same scenarios used to train W, so the K=16/K=116 claims do not yet establish out-of-sample performance.","rationale":"The reader's weakest-assumption analysis identifies exactly the same load-bearing issue: the paper evaluates the reconstruction matrix on the training data used to design it. My reading of the full text confirms this. The OPF data distillation formalism in Eqs. (7), (8), (12), and (26) is internally consistent, and the distinction between Type-1 and Type-2 objectives is clearly motivated. The Type-2 bilevel formulation and its proximal-gradient treatment are plausible, and the DEIM/PCA/GL comparisons are informative. However, the central empirical claim is specifically about what can be achieved on a data budget in real-time operation, i.e., about performance on future scenarios. Reporting f2 on the same Θ used to fit W cannot support that claim, because a flexible reconstruction matrix can overfit historical covariances and sensitivities. The absence of any held-out evaluation is not a stylistic omission; it is the difference between a fitted error and a predictive claim. The under-specification of the BGL2 second stage in Eq. (29) compounds this by making the headline method hard to reproduce. Given that the reader already recommends REJECT with moderate confidence, my assessment does not change the verdict. If the authors supply a proper out-of-sample evaluation and specify the BGL2 algorithm, the framework could still be valuable; as submitted, the data-budget conclusions are not supported.","tokens_in":18962,"tokens_out":3197,"duration_ms":32404,"concrete_test":"Perform a temporal holdout on the 800-scenario Pecan Street dataset: train W_BGL2 on the first 640 scenarios (or a time-blocked split) and evaluate f2/||X||^2_F, voltage feasibility, and the selected feature set on the remaining 160 unseen scenarios. Repeat for K=16 on the 37-bus system and K=116 on the 1,136-bus system. If held-out errors are materially larger than the reported in-sample values, or exceed the 5% target, the data-budget claim is not established. As a secondary check, attempt to reproduce BGL2 end-to-end purely from Eq. (29) and the paper's description; if no convergent procedure exists, the headline method is not independently reproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline numbers rest on a training-set evaluation. In Section V, the fidelity metric for OPF solutions is f2(W) from Eq. (26), averaged over the same T scenarios in Θ from Eq. (8). For BGL2, W is selected by first minimizing f2 via the bilevel program (27) and then fitting C in (29) on those same scenarios; Figs. 3 and Table I then report f2(W)/||X||^2_F on that same training set. No held-out test set, temporal validation, or cross-validation is described. The 'sample size effect' paragraph retrains on 50 randomly sampled scenarios and reports errors on those same 50 scenarios, so it also does not provide an out-of-sample check. Thus the conclusion that 'optimal DER schedules can be approximated reasonably well' with K=16 or K=116 implicitly assumes that in-sample error equals out-of-sample error. That assumption is load-bearing because the DSO would deploy W on future, unseen loading conditions, not on the historical scenarios used to fit it; load, solar, and EV profiles can shift with season, weather, and adoption patterns. Separately, the second stage of BGL2 in Eq. (29) is a bilevel continuous optimization over C for which no solution algorithm is given, so the headline method is also under-specified for independent reproduction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an OPF data distillation framework for distribution systems where the DSO selects K out of P load/solar data features, reconstructs the full data vector linearly as θ̂ = Wθ, and feeds the reconstruction into a fixed OPF solver. Two families of methods are developed: Type-1 methods (PCA, DEIM, group lasso, and a two-stage group lasso GL2) that minimize the data reconstruction error f1(W), and Type-2 methods (bilevel group lasso BGL and its two-stage variant BGL2) that minimize the OPF-solution error f2(W). The bilevel problem is solved by a proximal gradient algorithm with convergence guarantees for nonconvex objectives. Numerical tests on single-phase versions of the IEEE 37-bus and a 1,136-bus feeder report that BGL2 approximates optimal DER schedules within a few percent error using K=16 or K=116 features, respectively, and that the resulting voltage distributions are close to those of the full OPF.","tokens_in":19286,"tokens_out":4640,"duration_ms":43104,"significance":"If the reported performance holds out of sample, the paper makes a useful contribution to data-frugal and communication-limited DER scheduling: it directly addresses the practical problem of how many smart-meter data streams a DSO must collect to solve OPF approximately. The optimization derivations in Sections III and IV are internally consistent, and the gradient formulas and lemmas in the appendix are proved correctly, with the exception of an omitted proof detail. The problem is timely and the proposed framework is clearly relevant to distribution system operation. However, the central numerical claims are supported only by in-sample evaluation, and the two-stage BGL2 method that produces the headline results is under-specified, so the significance is conditional on a revision that adds out-of-sample validation and a complete algorithmic description.","major_comments":[{"comment":"The headline claims that \"optimal DER schedules can be approximated reasonably well\" with K=16 (Conclusions, 37-bus case) and \"within 5% error using K=116\" (Conclusions, 1,136-bus case) are based entirely on in-sample evaluation. The fidelity metrics f1(W) and f2(W) in Eqs. (12) and (26) are computed on the same training scenarios Θ in Eq. (8) that were used to fit W; in particular, for BGL2 the reconstruction matrix C is fitted by minimizing f2 over the same T scenarios in Eq. (29). Figures 2, 3, and Table I, as well as the \"Sample size effect\" paragraph in Section V-A, all report errors on the training set with no held-out test set, temporal validation, or cross-validation. Because the DSO would deploy W on future, unseen loading conditions, the data-budget savings claims are not established by the evidence presented. The authors should report out-of-sample errors (e.g., via a random or temporal train/test split, repeated over multiple splits to give confidence intervals) and state whether the K=16 and K=116 conclusions survive on held-out data.","section":"Section V and Conclusions"},{"comment":"The two-stage BGL2 method, which achieves the best reported fidelity in Figures 3 and 7 and Table I, is defined as the minimizer of the bilevel program in Eq. (29) over the reconstruction matrix C, but no solution algorithm is provided for this problem. Algorithm 2 solves the one-stage BGL problem in Eq. (27) with the sparsity penalty g(W); it does not solve the continuous bilevel least-squares problem in Eq. (29) that defines BGL2. As written, an independent reader cannot reproduce the BGL2 results or know whether the reported errors correspond to a global, local, or merely stationary solution of (29). The paper must specify how (29) is optimized (e.g., a projected/proximal gradient method on the selected support, using the Jacobian from Lemma 2, with initialization, step sizes, and stopping criteria) and state any convergence guarantees, or explicitly state that the same APG scheme is applied to C with the support fixed.","section":"Section IV, Eq. (29)"},{"comment":"The OPF in Eq. (5) depends on the soft-constraint penalty weights (ν, ρ), and the paper states that analytical bounds and heuristics for selecting them are given in [26]. However, the numerical values of ν and ρ used in all reported tests are not stated anywhere in Section V or the appendix. Since both the OPF minimizers x(θ) and the fidelity metric f2(W) depend directly on these parameters, the reported error percentages and voltage distributions are not reproducible without this information. The authors should report the exact values of ν and ρ used, and ideally show sensitivity of the headline results to these choices.","section":"Section V, Eq. (5)"}],"minor_comments":[{"comment":"The notation in Eq. (18) is garbled: the replacement of the two separate column norms by the norm of the concatenated vector is not typeset correctly, making the intended grouped penalty difficult to parse.","section":"Section III-C, Eq. (18)"},{"comment":"The proof of Lemma 3 is incomplete as written; it says \"The rest of the proof follows the proof of Lemma 1 and is omitted,\" which is acceptable for a short lemma, but the step connecting the critical-point condition to the bound on λ2 should be spelled out, especially because λ2 depends on x(0) and the Jacobian ∇θx(0).","section":"Appendix, proof of Lemma 3"},{"comment":"The exact AC model used to compute the voltage distributions in Figures 5 and 8 is not described; the paper should specify the AC power flow solver, the network model (e.g., unbalanced vs. single-phase equivalent), and any convergence tolerances.","section":"Section V-B, Figures 5 and 8"},{"comment":"There is a typo in the third contribution: \"reconstucted\" should be \"reconstructed\".","section":"Section I, contribution iii"}],"recommendation":"major_revision","confidential_remarks":"The reader's reject recommendation is understandable given the in-sample-only evaluation, but I view the core methodological issues as fixable within the scope of a major revision: the authors can add a proper train/test evaluation and specify the algorithm for Eq. (29). If they cannot provide out-of-sample results, the paper's central claims would not be supported, and rejection would then be appropriate. I also note that the paper is otherwise well-structured, and the optimization machinery (proximal gradient, Jacobian computation, lemmas) is sound; the main weakness is the mismatch between the strength of the conclusions and the evidence presented."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Reading this, the thing to know is that the contribution is a good idea with a serious evaluation gap: the headline numbers are computed on the training set. The claim that K=16 (or K=116) features suffice to approximate optimal DER schedules is not yet supported, because every fidelity number in Figures 2-3 and Table I is evaluated on the same scenarios used to fit the reconstruction matrix.\n\nWhat is actually new: the Type-2 bilevel group lasso objective, which selects features to minimize OPF dispatch error rather than reconstruction error, is a sensible new combination of group lasso, DEIM, and OPF sensitivity analysis. The two-stage BGL2 variant is also reasonable. The math in Sections III and IV holds up: Lemmas 1-3 are proved correctly, and the proximal gradient algorithm for the nonconvex problem follows a known convergent template. The paper is clearly written and cites the relevant literature fairly.\n\nThe soft spots are concentrated in the numerical section. The in-sample evaluation is load-bearing because the reconstruction matrix has P*K free parameters (e.g., roughly 120k for the 1,136-bus case with K=116), and T=800 scenarios. That is a recipe for overfitting; without a held-out test set or temporal validation, the reported 1-5% errors may be purely memory effects. The 'sample size effect' paragraph retrains on 50 scenarios and evaluates on those same 50, so it does not help. Second, the BGL2 second stage in Eq. (29) is a bilevel continuous optimization over C, and the paper gives no algorithm for solving it. That is a reproducibility gap: a reader cannot implement the method that produces the headline results. Minor issues: lambda1, lambda2, nu, rho are not selected systematically, and the voltage-feasibility violin plots are also on training scenarios.\n\nWho is this for: researchers in data-driven distribution operations, learning-to-optimize, and meter placement. The formulation is worth building on, but the empirical claims need rework. I would send it to review, with the clear expectation that the authors provide out-of-sample evaluation and specify the BGL2 solver. The idea deserves a chance; the current evidence does not yet support it.","headline":"Good idea, sound derivations, but the empirical claims are in-sample only; the paper needs a held-out test set and a specified BGL2 solver before the budget claims can be believed.","tokens_in":19806,"tokens_out":2790,"would_cite":false,"duration_ms":25880,"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":"The paper claims that a grid operator can nearly reproduce optimal DER schedules using only K well-chosen smart-meter readings, with 16 of 50 features on a small feeder and 116 of 1,038 on a large one.","keywords":["optimal power flow","smart meter data","feature selection","data distillation","group lasso","bilevel optimization","DER scheduling","proximal gradient"],"falsifier":"Train $W$ on one set of loading scenarios and evaluate $f_2$ on a held-out set from the same feeders; if the held-out normalized OPF-solution error is much larger than the in-sample curves of Fig. 3 and Table I, then the claimed $K=16$ and $K=116$ data budgets do not transfer to unseen conditions.","tokens_in":18781,"feed_emoji":"⚡","tokens_out":10237,"duration_ms":82883,"temperature":0.7,"pith_summary":"The paper asks how many real-time load measurements a distribution grid operator truly needs to schedule distributed energy resources, and answers that a small, carefully chosen subset can be enough. The proposal is a two-step data distillation pipeline: pick $K$ out of $P$ load-demand features to read from meters, then reconstruct the full demand vector with a linear map $W$ and feed it into the existing OPF solver. The authors' stronger contribution is a bilevel formulation that trains $W$ to minimize error in the final DER schedule, not in the reconstructed data, using a group-lasso sparsity penalty and proximal-gradient optimization. On two single-phase feeders, the two-stage bilevel method (BGL2) gives schedule errors of a few percent at $K=16$ of $P=50$ on the small feeder and within 5% at $K=116$ of $P=1{,}038$ on the large one, with voltage distributions nearly matching the full-data OPF. If this holds, operators can cut uplink communication and sensor installation without changing their OPF solver.","feed_headline":"Sixteen meter readings can fill in for fifty in grid dispatch","feed_subtitle":"A bilevel learning scheme picks the few loads that matter, cutting real-time data traffic with near-identical voltage control.","key_machinery":"The load-bearing mechanism is the distillation matrix $W = CS^\\top$, a $P \\times K$ selection matrix $S$ (one hot entry per chosen feature) followed by a $P \\times K$ reconstruction matrix $C$; $W$ is column-sparse so that only $K$ OPF features are needed in real time. For the Type-2 objective, $W$ is found by minimizing the OPF-solution fitting error $f_2(W)$ plus a group-lasso penalty $\\lambda_2 \\sum_p \\|w_p\\|$, which forces whole columns of $W$ to zero and thereby selects features. The gradient of $f_2$ with respect to $W$ is carried by the Jacobian $\\nabla_{\\hat\\theta} x$ of the OPF solution with respect to its data, so each proximal-gradient iteration solves a batch of $T$ OPFs and one linear system per instance; the two-stage BGL2 then discards the biased lasso reconstruction and least-squares fits $C$ on the chosen support.","core_discovery":"On the paper's own terms, the central claim is that the OPF mapping $\\theta \\mapsto x(\\theta)$ is compressible: a linear reconstruction matrix $W = CS^\\top$, trained on historical scenarios, can map $K$ selected entries of the OPF data vector back to a full $P$-dimensional vector whose OPF minimizer $x(W\\theta)$ stays close to $x(\\theta)$. The paper argues that aiming directly at solution fidelity, rather than data fidelity, is what makes small data budgets work. It formalizes this as the bilevel group-lasso program of (27), where the outer level minimizes $f_2(W)=\\frac{1}{2T}\\sum_t \\|x(\\theta_t)-x(W\\theta_t)\\|^2$ over a column-sparse $W$ and each inner level solves the OPF at the reconstructed data; the two-stage variant BGL2 first uses BGL to pick the $K$ features and then refits the reconstruction matrix by least squares, which the numerical tests show to be the most accurate scheme. The reported experiments on the 37-bus and 1,136-bus single-phase feeders are the evidence offered for the claim.","pith_inferences":["The reported errors are in-sample, so the practical data budget is not fully established until the reconstruction matrix is tested on held-out scenarios; a distribution shift in loads could require more than $K$ measurements.","If the selected-feature pattern is stable across seasons, the same selection matrix could double as a minimal meter-placement plan, but the paper does not test cross-season stability.","The per-iteration cost of solving $T$ OPFs and Jacobians is the main scalability bottleneck, so a stochastic proximal variant that samples one OPF per iteration, listed as future work in the paper, would be the natural next step for larger feeders."],"forward_implications":["A distribution operator can keep its existing OPF solver untouched and operate on only $K$ meter feeds, so real-time uplink communication falls by the ratio $1-K/P$ with no change to dispatch software.","At larger budgets $K/P$, simple Type-1 schemes such as GL2 match the bilevel method, so operators can choose the cheaper convex training when they can afford more measurements.","The selected features reveal which loads and which active/reactive pairs are most influential, which can guide meter placement and cyber-protection priorities.","Voltage feasibility is preserved: BGL2 schedules keep nodal voltages close to the full-data OPF distribution under both linearized and exact AC models.","Because the framework only needs gradients through the OPF mapping, the same distillation idea transfers to AC OPF, multi-phase feeders, and other OPF variants."],"supporting_citations":[{"why":"Supplies the DEIM greedy interpolation used for Type-1 feature selection and its reconstruction error bound.","marker":"[28]"},{"why":"Defines group-lasso regularization, the column-sparsity mechanism behind both GL and BGL feature selection.","marker":"[19]"},{"why":"Provides the accelerated proximal-gradient algorithm with a convergence guarantee for the nonconvex bilevel problem.","marker":"[30]"},{"why":"Provides the linearized distribution-flow model, ohmic-loss approximation, and soft-constraint OPF formulation that define $x(\\theta)$.","marker":"[26]"},{"why":"Shows how to compute the OPF Jacobian $\\nabla_{\\theta}x$, which the bilevel gradient step requires.","marker":"[1]"},{"why":"Residential load and solar generation dataset used to build the 37-bus test scenarios.","marker":"[35]"},{"why":"Residential load and solar generation dataset used to build the 1,136-bus test scenarios.","marker":"[37]"}],"fun_headline_variants":["Few meter reads, full grid dispatch: bilevel OPF distillation","Pick 16 loads, solve the whole feeder: OPF data distillation","Data-budget OPF: select few meters, reconstruct the rest","Sparse meter data, near-optimal dispatch: a bilevel trick","OPF on a budget: which meters matter most?"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole budget reduction rests on the assumption that the reconstruction matrix trained on historical load scenarios will also work on future load conditions the operator has not seen, since the paper's error numbers are computed on the same scenarios used for training.","fun_headline_variants_meta":{"raw":{"variants":["Few meter reads, full grid dispatch: bilevel OPF distillation","Pick 16 loads, solve the whole feeder: OPF data distillation","Data-budget OPF: select few meters, reconstruct the rest","Sparse meter data, near-optimal dispatch: a bilevel trick","OPF on a budget: which meters matter most?"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000523,"raw_usage":{"total_tokens":2563,"prompt_tokens":1016,"completion_tokens":1547,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":632,"completion_tokens_details":{"reasoning_tokens":1457}},"tokens_in":632,"tokens_out":1547,"duration_ms":9553,"temperature":1.0,"reasoning_tokens":1457,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T14:41:40.777257+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train $W$ on one set of loading scenarios and evaluate $f_2$ on a held-out set from the same feeders; if the held-out normalized OPF-solution error is much larger than the in-sample curves of Fig. 3 and Table I, then the claimed $K=16$ and $K=116$ data budgets do not transfer to unseen conditions.","supporting_citations":[{"cited_title":"An empirical interpolation method: application to efficient reduced-basis discretization of partial differential equations,","cited_arxiv_id":null,"evidence_quote":"Supplies the DEIM greedy interpolation used for Type-1 feature selection and its reconstruction error bound."},{"cited_title":"Model selection and estimation in regression with grouped variables,","cited_arxiv_id":null,"evidence_quote":"Defines group-lasso regularization, the column-sparsity mechanism behind both GL and BGL feature selection."},{"cited_title":"Accelerated proximal gradient methods for nonconvex programming,","cited_arxiv_id":null,"evidence_quote":"Provides the accelerated proximal-gradient algorithm with a convergence guarantee for the nonconvex bilevel problem."},{"cited_title":"Fast probabilistic hosting capacity analysis for active distribution systems,","cited_arxiv_id":null,"evidence_quote":"Provides the linearized distribution-flow model, ohmic-loss approximation, and soft-constraint OPF formulation that define $x(\\theta)$."},{"cited_title":"Learning to solve the AC-OPF using sensitivity-informed deep neural networks,","cited_arxiv_id":null,"evidence_quote":"Shows how to compute the OPF Jacobian $\\nabla_{\\theta}x$, which the bilevel gradient step requires."},{"cited_title":"Dataport","cited_arxiv_id":null,"evidence_quote":"Residential load and solar generation dataset used to build the 37-bus test scenarios."},{"cited_title":"Smart*: An open data set and tools for enabling research in sustainable homes,","cited_arxiv_id":null,"evidence_quote":"Residential load and solar generation dataset used to build the 1,136-bus test scenarios."}],"review_version":1}