{"id":"7c1af491-bb49-44ca-84e8-5a1ab3854e60","arxiv_id":"2605.30763","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Framework with MPC-aware feature selection, Input-Convex Encoder-Only Transformer, and ADMM achieves near-centralized performance in scalable distributed MPC for heterogeneous building aggregations.","lead":"The paper describes a data-driven framework that selects features for accurate multi-step building forecasts, uses a convex neural network to model dynamics, and applies distributed ADMM to solve large-scale MPC problems for demand response. A smart generalist might read it to understand practical methods for coordinating many buildings on the grid without central computation becoming impossible.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"No significant objection identified","rationale":"Reader correctly flags the IC-EoT convexity assumption as central, but the abstract already claims both the guarantee and the empirical validation; absent the full text showing a hidden gap in that guarantee, the reported results stand as stated. No adjustment to UNVERDICTED is warranted.","tokens_in":1725,"tokens_out":266,"duration_ms":10441,"concrete_test":"Re-run the co-simulation comparison (Table or Figure reporting economic cost and comfort metrics) after replacing the IC-EoT with a non-convex neural baseline of identical capacity; if the distributed ADMM still converges to near-centralized performance, the convexity step is not load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract presents a coherent pipeline: MPC-aware feature selection, IC-EoT for convex dynamics, and ADMM for distributed CCP solution, with explicit co-simulation validation against a centralized baseline. The central claim of near-identical optimality plus scalability rests on the IC-EoT convexity guarantee, but the abstract states this is achieved and the results are reported as holding in the high-fidelity environment. No internal inconsistency or unstated assumption that would falsify the reported performance is visible from the given text.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes a data-driven framework for scalable distributed MPC in heterogeneous building aggregations for demand response. It combines an MPC-aware feature selection method to support multi-step forecasting, a novel Input-Convex Encoder-Only Transformer (IC-EoT) to model building dynamics while guaranteeing convexity of the resulting optimization problem, and the Tracking ADMM algorithm to solve the constraint-coupled problem in a fully distributed fashion. High-fidelity co-simulation validation with EnergyPlus models of consumer and prosumer buildings under a dynamic TOU tariff is reported to show near-identical economic optimality and superior thermal comfort relative to a centralized controller, together with computational scalability that renders the centralized approach infeasible for large aggregations.","tokens_in":1821,"tokens_out":508,"duration_ms":19542,"significance":"If the quantitative validation results and convexity guarantee hold under detailed scrutiny, the work would offer a practical route to real-time distributed MPC for large-scale building demand response, directly addressing the computational intractability of centralized formulations while preserving performance. The systematic MPC-aware feature selection and the IC-EoT architecture represent potentially reusable contributions for convex data-driven modeling in control applications.","major_comments":[{"comment":"Abstract (validation paragraph): the central performance claims of near-identical economic optimality, superior thermal comfort, and exceptional scalability are stated without any quantitative metrics, error bars, ablation results, or comparison tables. This absence prevents assessment of whether the reported advantages are statistically meaningful or sensitive to the IC-EoT modeling assumptions.","section":"Abstract"},{"comment":"Framework components (IC-EoT description): the assertion that the Input-Convex Encoder-Only Transformer guarantees convexity of the subsequent MPC optimization problem is load-bearing for the entire pipeline, yet no explicit architectural constraints, loss-function modifications, or proof sketch are referenced that would establish this property for arbitrary building dynamics.","section":"Framework components"}],"minor_comments":[{"comment":"Define all acronyms at first use (e.g., ADMM, CCP, TOU, DR) and ensure consistent notation between the abstract and the main text.","section":null},{"comment":"Clarify the precise interface between the MPC-aware feature selection step and the IC-EoT training objective so that readers can reproduce the multi-step prediction robustness claim.","section":null}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their constructive comments, which help clarify the presentation of our validation results and the technical foundations of the IC-EoT. We respond to each major comment below and indicate the corresponding revisions.","responses":[{"response":"We agree that the abstract would benefit from quantitative support for the performance claims. In the revised manuscript we will update the validation paragraph to include the specific metrics, error bars, and references to ablation studies and comparison tables already present in the results section. This will enable readers to assess statistical meaningfulness and sensitivity to modeling assumptions directly from the abstract.","revision_made":"yes","referee_comment":"[Abstract] Abstract (validation paragraph): the central performance claims of near-identical economic optimality, superior thermal comfort, and exceptional scalability are stated without any quantitative metrics, error bars, ablation results, or comparison tables. This absence prevents assessment of whether the reported advantages are statistically meaningful or sensitive to the IC-EoT modeling assumptions."},{"response":"The convexity guarantee rests on the input-convex architecture of the IC-EoT. We will revise the manuscript to make the architectural constraints, loss-function modifications, and proof sketch more explicit in the main text (with the full derivation remaining in the appendix). The proof applies to any dynamics that can be represented by the trained input-convex model; we will add a clarifying sentence on the scope with respect to arbitrary building dynamics.","revision_made":"yes","referee_comment":"[Framework components] Framework components (IC-EoT description): the assertion that the Input-Convex Encoder-Only Transformer guarantees convexity of the subsequent MPC optimization problem is load-bearing for the entire pipeline, yet no explicit architectural constraints, loss-function modifications, or proof sketch are referenced that would establish this property for arbitrary building dynamics."}],"tokens_in":1403,"tokens_out":395,"duration_ms":25922,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper's core move is to treat the whole chain—feature choice, dynamics model, and solver—as one data-driven system that stays convex and distributed. That combination is what lets them claim the distributed version matches centralized economics while beating it on comfort and scaling past the point where the central solver times out.\n\nThe feature selection step is the part that feels most grounded. Conventional methods ignore how prediction errors grow over the MPC horizon, so tuning selection to multi-step accuracy is a direct fix for a known weakness. The IC-EoT is the novel piece that is supposed to keep the optimization convex while still modeling real building behavior, and the ADMM layer then decomposes the coupled problem across heterogeneous buildings. The co-simulation with EnergyPlus under TOU tariffs is a realistic test bed, and the reported outcome—that the distributed controller stays close on cost and improves comfort—is the kind of result that matters for demand-response work.\n\nThe main limitation right now is that we only see the abstract. No numbers on the optimality gap, no ablation on how much the convexity constraint costs in prediction error, and no detail on how the transformer enforces input-convexity without extra conservatism. Those gaps make it hard to judge whether the “near-identical” claim holds under tighter comfort bounds or different building stocks. The heterogeneity handling also gets little space, even though the setup includes both consumers and prosumers.\n\nThis is aimed at people who already work on scalable building control or grid flexibility and need something they can implement or extend. A reader who cares about the computational side of aggregation will find the architecture worth pulling apart.\n\nI would send it to peer review. The problem is real, the pipeline is concrete, and the validation environment is strong enough that referees can test whether the convexity guarantee actually delivers the claimed performance.","headline":"The paper stitches together MPC-aware feature selection, a custom convex transformer, and ADMM into a working pipeline for large building aggregations, but the abstract leaves the actual performance margins and convexity trade-offs unquantified.","tokens_in":2292,"tokens_out":456,"would_cite":false,"duration_ms":18495,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A data-driven framework enables distributed MPC for large heterogeneous building aggregations that matches centralized economic performance while scaling to real-time operation.","keywords":["distributed MPC","building aggregation","demand response","input-convex transformer","ADMM","feature selection","convex optimization","thermal comfort"],"falsifier":"A co-simulation trial in which the distributed controller's total energy cost for an aggregation of 50 or more buildings deviates by more than a few percent from the centralized optimum or in which its per-step solve time exceeds the real-time control interval.","tokens_in":2623,"feed_emoji":"🏠","tokens_out":669,"duration_ms":21817,"temperature":0.7,"pith_summary":"The paper develops a complete methodology to coordinate demand response across many different buildings using model predictive control. It first applies systematic feature selection tuned to multi-step forecasting needs, then uses a specialized transformer to model building dynamics while keeping the problem convex, and finally solves the coupled optimization in a distributed way with Tracking ADMM. A reader would care because centralized MPC becomes impossible to run in real time once the number of buildings grows large, leaving energy costs and indoor comfort unmanaged under time-varying electricity prices.","feed_headline":"Distributed MPC matches centralized optimality in large building groups","feed_subtitle":"Framework with MPC-aware selection, convex transformer, and ADMM scales past centralized limits while preserving cost and improving comfort.","key_machinery":"The Input-Convex Encoder-Only Transformer (IC-EoT) that models building dynamics while guaranteeing convexity of the resulting optimization problem, together with the Tracking Alternating Direction Method of Multipliers (ADMM) that solves the constraint-coupled problem in a fully distributed manner.","core_discovery":"The paper claims that a framework combining MPC-aware feature selection, an Input-Convex Encoder-Only Transformer for dynamics modeling, and Tracking ADMM for distributed solution of the constraint-coupled problem delivers near-identical economic optimality and superior thermal comfort compared to a theoretical centralized controller, while providing the computational scalability required for real-time control of large aggregations of consumer and prosumer buildings, as shown in EnergyPlus co-simulations under a dynamic time-of-use tariff.","pith_inferences":["The same combination of convex modeling and distributed ADMM could apply to other large-scale networked systems with coupled constraints, such as district heating or microgrid coordination.","Deployment would lower the computing hardware needed at a central operator, shifting effort to local building controllers.","Testing the framework on real measured data rather than EnergyPlus simulations would reveal whether model mismatch affects the observed performance gap."],"forward_implications":["The distributed approach can control aggregations whose size makes centralized MPC computationally infeasible.","Economic performance stays nearly identical to the centralized optimum while thermal comfort improves.","The method handles mixed consumer and prosumer buildings under dynamic time-of-use pricing.","MPC-aware feature selection produces multi-step forecasts robust enough for the subsequent optimization step."],"fun_headline_variants":["ADMM with convex transformer scales MPC to large building aggregations","MPC-aware selection enables distributed control matching centralized results","Input-convex transformer plus ADMM achieves near-optimal building DR","Distributed framework delivers scalable MPC for heterogeneous buildings"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The Input-Convex Encoder-Only Transformer accurately captures the thermal dynamics of heterogeneous buildings while ensuring the formulated optimization problem remains convex.","fun_headline_variants_meta":{"raw":{"variants":["ADMM with convex transformer scales MPC to large building aggregations","MPC-aware selection enables distributed control matching centralized results","Input-convex transformer plus ADMM achieves near-optimal building DR","Distributed framework delivers scalable MPC for heterogeneous buildings"]},"model":"grok-4.3","cost_usd":0.003434,"raw_usage":{"total_tokens":1741,"prompt_tokens":681,"num_sources_used":0,"completion_tokens":62,"cost_in_usd_ticks":34340500,"prompt_tokens_details":{"text_tokens":681,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":998,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":681,"tokens_out":62,"duration_ms":7719,"temperature":1.0,"reasoning_tokens":998,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-28T21:44:05.196311+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A co-simulation trial in which the distributed controller's total energy cost for an aggregation of 50 or more buildings deviates by more than a few percent from the centralized optimum or in which its per-step solve time exceeds the real-time control interval.","supporting_citations":[],"review_version":1}