{"id":"d781d7d9-9ed1-4b99-8539-fbfa40b28555","arxiv_id":"2507.20465","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A relax-and-cut temporal decomposition with partially relaxed look-ahead windows and dynamic N-1 cut separation solves SCUC faster than monolithic Gurobi while keeping primal gaps near 1%.","lead":"The paper presents a decomposition algorithm that solves large security-constrained unit commitment problems by keeping near-term decisions exact while relaxing later ones, and by adding N-1 security cuts during branch-and-cut. The method reports sub-1% gaps in about 20% of the solution time of a monolithic Gurobi solve on MATPOWER instances up to 13,000 buses.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sub-1% 'optimality gaps' are measured against Full+TF, a non-certified filtered MIP heuristic; the true gap to optimality is unestablished.","rationale":"The reader identified the partial-relaxation look-ahead as the weakest assumption and gave a CONDITIONAL verdict. The reader's rationale also notes the non-certified baseline, but this is arguably the more load-bearing issue: even if the partial-relaxation mechanism works exactly as intended, the reported 'optimality gaps' are only distances to Full+TF, which is not proven optimal. The headline quantitative claim therefore lacks the required support. A single concrete experiment—computing true optimality gaps on a few instances that can be solved to certified optimality—would settle whether the abstract overstates the result. The reader's CONDITIONAL verdict remains appropriate, with the condition sharpened to require certified benchmarking. No internal inconsistency or methodological flaw was found in the decomposition logic or the dynamic cut separation; the primary risk is external validity of the headline performance metric.","tokens_in":14441,"tokens_out":4495,"duration_ms":52121,"concrete_test":"Choose the five smallest benchmark instances (1354pegase, 1888rte, 1951rte, 2383wp, 2736sp). Re-solve the full SCUC model (8) with Gurobi at a 0.00% MIP gap tolerance and without transmission filtering (or with a verifiably exact filter) to obtain certified optimal objectives. Compute TD-R+Cut's true optimality gap as (Obj_TD - Obj_opt)/Obj_opt and compare against the Table II gaps. If any true gap exceeds 1.0%, or systematically exceeds the reported gaps by more than 0.2 percentage points, the 'sub-1% optimality gap' claim is not supported and should be reworded as 'primal gap relative to a heuristic baseline.' This check also tests whether the TF baseline is exact, since certified Full+TF objectives should match the certified optima if the filter preserves optimality.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table II reports 'Gap' as the relative objective difference between TD-R+Cut and Full+TF. Full+TF is not a certified optimal baseline: Gurobi is run with a 0.1% MIP gap tolerance, so its objective can be up to 0.1% above the optimum of the model it solves, and the transmission filtering (TF) step removes security constraints, potentially yielding a relaxed model whose true optimum is below the full SCUC optimum. The abstract's headline claim that the method achieves 'optimality gaps below 1%' is therefore not supported by the experiments—what is shown is a primal gap to a heuristic baseline, not to the SCUC optimal value. Moreover, for the largest instance (9241pegase), Full+TF returns no solution, so no gap is reported and the average '0.9%' gap is computed over a subset or relative to an unspecified reference. The core quantitative claim could fail if the true optima are lower than the Full+TF objectives; for example, a reported 0.75% gap on 1354pegase could become >1% if the certified optimum is 0.2% below the Full+TF value. This is load-bearing because the paper's central contribution is the sub-1% optimality guarantee in a fraction of monolithic runtime. The method may still be a strong heuristic, but the evidence as presented does not establish the stated optimality gap.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a relax-and-cut temporal decomposition framework for the Security-Constrained Unit Commitment (SCUC) problem. The method solves a sequence of subproblems with three temporal windows: fixed, integer-constrained, and relaxed-integer look-ahead, which extends the planning horizon without keeping all integer variables. Security constraints are enforced dynamically through callbacks within branch-and-cut, rather than being enumerated upfront. An optional RINS-based refinement step improves the incumbent solution. The authors evaluate the method on 20 large MATPOWER instances, reporting that their TD-R+Cut method achieves relative gaps below 1% versus a monolithic Gurobi baseline while using about one-fifth of the runtime, and that it outperforms a conventional temporal decomposition baseline. The paper includes pseudocode, ablation studies, parallel-scaling results, and a public implementation repository.","tokens_in":14744,"tokens_out":5252,"duration_ms":60913,"significance":"If the claimed results hold, the method would be a practically valuable heuristic for large-scale SCUC, where monolithic MILP solvers are often impractical. The paper has concrete strengths: the algorithmic description is clear, the code is publicly available, and the experimental design includes ablations for the two main innovations, parallel performance evaluation, and comparison against several baselines. The dynamic cut-separation mechanism embedded in branch-and-bound is a sensible engineering contribution. However, the central quantitative claim—'optimality gaps below 1%'—is not actually supported by the reported experiments, because the reference baseline is not a certified optimal solution. This weakens the headline contribution and requires revision.","major_comments":[{"comment":"","section":"Section VI.A, Table II"},{"comment":"","section":"Table II and Table III, 9241pegase row"},{"comment":"","section":"Section VI (Configuration and Case Study)"}],"minor_comments":[{"comment":"","section":"Section II, Eq. (4)"},{"comment":"","section":"Section VI.A"},{"comment":"","section":"Section VI.A, Table II"},{"comment":"","section":"Figure 7"}],"recommendation":"major_revision","confidential_remarks":"The referee notes that the paper is positioned as a mathematical optimization contribution, but its main evidence is heuristic performance without theoretical guarantees. The load-bearing claim of sub-1% optimality gaps needs to be either certified with dual bounds or explicitly downgraded to a comparison against a heuristic baseline. Such a revision is feasible within the manuscript's scope, so rejection is not warranted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is solid and the experiments are honest enough to be worth engaging with. The authors integrate two known ingredients — partial relaxation of future commitment variables and dynamic N-1 cut separation inside branch-and-cut — into a rolling-horizon SCUC framework. That integration is new in this specific combination, and it pays off: on the 20 MATPOWER instances, TD-R+Cut consistently beats the temporal decomposition baseline (TD+TF) on both gap and speed, and it solves cases where the monolithic baseline fails. The ablation study separating the effects of partial relaxation and dynamic cuts is useful, and the RINS refinement stage shows a real improvement with modest runtime cost. The code is public. That is a legitimate contribution.\n\nThe load-bearing weakness is exactly what the stress-test note flags. The headline \"optimality gaps below 1%\" is a primal gap measured against Full+TF, which is itself a heuristic: Gurobi runs with a 0.1% MIP gap tolerance, and the transmission filtering step can produce a relaxed model whose true optimum is lower than the full SCUC optimum. So the reported gaps are not certified. A 0.75% reported gap on 1354pegase could become >1% if Full+TF's objective is even 0.2% above the true optimum. The largest instance (9241pegase) has no Full+TF baseline at all, so the average 0.9% number is computed over a subset without saying exactly how. This matters because the abstract's central claim is the sub-1% guarantee. The method may well deliver it — the TD+TF gaps are much larger and the partial relaxation does appear to reduce myopia — but the evidence as presented does not establish the stated optimality gap.\n\nA second, lesser soft spot: the window sizes (sI=sR=dt=6, ds=2) are hand-set and no sensitivity analysis is reported. Since the whole method is a heuristic, parameter robustness is more than a nicety. That said, the paper does not hide this; the parameters are stated plainly.\n\nThe citation pattern looks fair. The prior temporal-decomposition and constraint-filtering works are cited, and the claim of novelty is about the integration, which holds up on reading. There is no target fitting: objectives are compared externally to Gurobi baselines.\n\nBottom line: a solid engineering paper on an important practical problem, with an overstated headline. I would send it to peer review, but the reviewers should require the authors to either re-benchmark against a certified optimal or at least report gaps relative to the best known dual bound, and to add a parameter sensitivity study. A reader interested in SCUC decomposition heuristics will get real value from this.","headline":"A genuinely useful heuristic for large-scale SCUC with a real soft spot: the sub-1% gap claim is against a non-certified baseline, not the true optimum.","tokens_in":15231,"tokens_out":672,"would_cite":true,"duration_ms":10350,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C11","90C27"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that relaxing integrality in a look-ahead window while keeping near-term decisions integer lets a temporal decomposition of security-constrained unit commitment produce solutions within 1% of the monolithic optimum using…","keywords":["security-constrained unit commitment","temporal decomposition","partial relaxation","branch-and-cut","N-1 contingency","relaxation-induced neighborhood search","mixed-integer linear programming","large-scale power systems"],"falsifier":"On a reported instance, perturb the demand in the relaxed window by ±20% while keeping the integer window unchanged and then fix the first $\\Delta t$ periods; if the objective degrades by more than one percentage point relative to the unperturbed run, the method's near-optimality rests on an accurate look-ahead rather than on the relaxation itself.","tokens_in":14243,"feed_emoji":"⚡","tokens_out":8215,"duration_ms":95129,"temperature":0.7,"pith_summary":"This paper is trying to establish that a rolling-horizon method can solve security-constrained unit commitment (SCUC) on grids with more than 13,000 buses at near-optimal cost without paying for a full-horizon mixed-integer solve. The central device is to keep commitment variables integral in the current decision window while relaxing them to continuous values in a look-ahead window, so each subproblem carries forward-looking information at fractional cost. The authors report that the resulting relax-and-cut framework reaches primal gaps below 1% in most tested cases while using about 20% of the solver time of the monolithic model, and that it outperforms prior temporal-decomposition baselines on both gap and runtime. A sympathetic reader would care because the practical bottleneck in day-ahead SCUC is exactly this gap between solution quality and wall-clock time, and the paper offers a way to shift that balance without a new solver.","feed_headline":"Relaxed future decisions get SCUC under 1% gap at one-fifth time","feed_subtitle":"Freeze near-term choices, relax later ones, add security cuts on the fly: large grids solve in minutes.","key_machinery":"The machinery is a four-way temporal partition: fixed, integer, relaxed, and ignored windows. Each subproblem is the partially relaxed program $SCUC^{PR}$ of length $s_I + s_R$, in which commitment variables in the first $s_I$ periods are binary and those in the next $s_R$ periods are continuous. After each solve, the first $\\Delta t$ periods of the integer window are frozen into the fixed window, state is passed forward by Algorithm 1, and the windows advance. Security constraints are enforced by a callback: whenever branch-and-cut finds an integer solution, the violated N-1 constraints are separated and added as cuts inside the same solve. An optional RINS pass then re-optimizes a rolling window with everything else fixed, cutting the average primal gap from 0.67% to 0.37% at a 13% runtime increase.","core_discovery":"On its own terms, the paper claims that the sequence of partially relaxed subproblems — where the first $s_I$ periods keep binary commitments and the next $s_R$ periods relax them to $[0,1]$, with N-1 security cuts added inside branch-and-cut — yields feasible, near-optimal SCUC schedules. In the twenty-instance benchmark, TD-R+Cut attains an average primal gap of 0.9% versus 2.8% for the temporal-decomposition baseline and produces solutions in 59 seconds on average, roughly one-fifth of the monolithic solver's 281 seconds; on the largest 13,659-bus case it solves in 730 seconds. The paper presents this as evidence that the relaxed look-ahead window transmits enough future cost and feasibility information to prevent the myopic decisions that degrade short-window decomposition, and that dynamic cut separation avoids the repeated cold starts of constraint-filtering approaches.","pith_inferences":["The same partial-relaxation trick could transfer to any rolling-horizon mixed-integer problem whose future coupling is through continuous state variables; the boundary between 'integer now' and 'relaxed later' would need to be tuned per problem.","A natural performance bound to test next is whether the relaxed window's LP dual value provides a lower bound on the cost-to-go; if it does, the fixed-$\\Delta t$ commitment policy inherits a regret proportional to the relaxation gap, which the paper does not state.","The single-parameter validation ($s_I=s_R=\\Delta t=6$, $\\Delta s=2$) leaves open whether the 0.9% average gap depends delicately on that setting; a sweep over window lengths would show how robust the claim is."],"forward_implications":["On benchmark systems above 1,000 buses, the method brings average primal gaps from about 2.8% (prior temporal decomposition) down to 0.9%, while halving the decomposition baseline's solve time.","On the largest instances, the method solves systems with over 13,000 buses and roughly 14,000 contingencies in about 730 seconds, a regime where the monolithic model either takes hours or returns no feasible solution.","Dynamic cut separation inside branch-and-cut yields a feasible solution earlier than iterative constraint filtering because the solver does not restart from scratch; the ablation study attributes part of the speed gain to this.","With eight threads, TD-R+Cut scales better than the baselines (e.g., 48% time reduction on one 6,515-bus instance versus 26% for the monolithic approach), suggesting the design is well suited to parallel deployment."],"supporting_citations":[{"why":"Supplies the full SCUC formulation and the production-cost representations that the subproblems inherit.","marker":"[2]"},{"why":"The temporal-decomposition baseline with extended windows that the paper extends by partial relaxation.","marker":"[6]"},{"why":"Transmission-constraint filtering used by the monolithic and decomposition baselines; the comparison point for dynamic cut separation.","marker":"[14]"},{"why":"Relaxation-induced neighborhood search, the source of the optional refinement pass.","marker":"[15]"},{"why":"Provides the implementation base and the temporal-decomposition baseline method.","marker":"[19]"},{"why":"Supplies the 20 public benchmark instances used for the numerical evaluation.","marker":"[20]"}],"fun_headline_variants":["Relax-and-cut yields SCUC under 1% gap in one-fifth time","Near-optimal SCUC via relaxed horizon and on-the-fly cuts","Fivefold faster SCUC with sub-1% optimality gap","Freeze today, relax tomorrow: SCUC under 1% gap","Dynamic cuts and relaxed look-ahead: SCUC solved 5x faster"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method leans on the belief that when you relax commitment decisions in the look-ahead window, the relaxed problem still gives a truthful enough picture of future cost and feasibility that freezing the first few periods of its solution is nearly as good as solving those periods with full future information.","fun_headline_variants_meta":{"raw":{"variants":["Relax-and-cut yields SCUC under 1% gap in one-fifth time","Near-optimal SCUC via relaxed horizon and on-the-fly cuts","Fivefold faster SCUC with sub-1% optimality gap","Freeze today, relax tomorrow: SCUC under 1% gap","Dynamic cuts and relaxed look-ahead: SCUC solved 5x faster"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000833,"raw_usage":{"total_tokens":3649,"prompt_tokens":975,"completion_tokens":2674,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":2576}},"tokens_in":591,"tokens_out":2674,"duration_ms":20543,"temperature":1.0,"reasoning_tokens":2576,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:33:48.808636+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a reported instance, perturb the demand in the relaxed window by ±20% while keeping the integer window unchanged and then fix the first $\\Delta t$ periods; if the objective degrades by more than one percentage point relative to the unperturbed run, the method's near-optimality rests on an accurate look-ahead rather than on the relaxation itself.","supporting_citations":[{"cited_title":"On mixed-integer pro- gramming formulations for the unit commitment problem,","cited_arxiv_id":null,"evidence_quote":"Supplies the full SCUC formulation and the production-cost representations that the subproblems inherit."},{"cited_title":"Temporal decomposition for improved unit commitment in power system production cost modeling,","cited_arxiv_id":null,"evidence_quote":"The temporal-decomposition baseline with extended windows that the paper extends by partial relaxation."},{"cited_title":"Transmission constraint filtering in large-scale security-constrained unit commitment,","cited_arxiv_id":null,"evidence_quote":"Transmission-constraint filtering used by the monolithic and decomposition baselines; the comparison point for dynamic cut separation."},{"cited_title":"Exploring relaxation induced neighborhoods to improve mip solutions,","cited_arxiv_id":null,"evidence_quote":"Relaxation-induced neighborhood search, the source of the optional refinement pass."},{"cited_title":"Unitcommit- ment. jl: A julia/jump optimization package for security-constrained unit commitment (version 0.3),","cited_arxiv_id":null,"evidence_quote":"Provides the implementation base and the temporal-decomposition baseline method."},{"cited_title":"Mat- power: Steady-state operations, planning, and analysis tools for power systems research and education,","cited_arxiv_id":null,"evidence_quote":"Supplies the 20 public benchmark instances used for the numerical evaluation."}],"review_version":1}