acceptodds
Under review as a conference paper at ICLR 2027

Where a Physical Constraint Belongs: Constraint-Exact Generative Scenario Generation for Power Grids

Abstract

Generative models are increasingly used to produce operational scenarios for power systems, and such scenarios must satisfy the physical laws they describe: an injection pattern that violates Kirchhoff’s laws cannot occur on any real net- work. Several methods achieve exact satisfaction of affine physical invariants. The question the field has not answered is where the constraint belongs: in the hypothesis class (train-time projection), at inference (zero-shot correction), in the loss (penalty), or in the coordinates (nullspace or completion). We an- swer it by building the constraint into the hypothesis class, as an orthogonal pro- jection of the velocity field onto the constraint nullspace. This is exact under any Runge–Kutta scheme, excludes no minimiser of the flow-matching ob- jective, and costs zero additional function evaluations (Theorem 1). On the transmission-network suite it sets the state of the art: the best variogram score of 16 methods, leading the strongest baseline by 3.8% (t = −6.13), completion- based constraint handling by 7.9% (t = −8.19) and inference-time correction by 10.5% (t = −12.40); the best CRPS of 16; and a worst-case constraint vi- olation 727,790× smaller than unconstrained flow matching, at -1.96% energy score (t = −6.29). Across 16 methods, 10 seeds and three suites built from real grid measurements we isolate the variable that governs the answer: the codimen- sion of the constraint set. On a controlled contrast in which the only change is how much of the state the physics determines, train-time projection moves from neutral to decisive. Soft penalties, the field’s default, lose on both axes at once. Raising λ from 1 to 1000 moves the violation only from 219 MW to 187 MW while degrading the energy score by +92%. We further contribute the first decision-level evaluation of constraint-exact scenario generation, through two-stage stochastic unit commitment. Code, the benchmark and the generators that produce every table and figure here from the raw run files are at https: //github.com/Nishant27-2006/constraint-placement, with the project page at https://hamiltonian-network.github.io.

Primary area:
generative models
Keywords:
generative modelling, flow matching, physics-informed constraints, power systems, scenario generation, benchmark design.

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_o7GDOpenReview ↗
Rating
1
Confidence
4(+2)
Informativeness
1
Presentation
1
Soundness
1(−1)
Read review
Paper summary

This paper studies where to place physical constraints in power grid scenario generation, comparing soft penalties, training-time velocity projection, inference-time correction, and low-dimensional coordinate methods. The authors show that velocity projection under affine constraints can preserve feasibility, does not increase the number of network evaluations, and preserves the theoretically optimal solution of flow matching. Experiments on three benchmarks show that the advantages of different methods vary by task; downstream unit commitment and nonlinear power flow experiments further show that physical feasibility does not guarantee decision benefits, and that success in the linear case cannot be directly generalized.

Critical strengths
  1. The problem is practically relevant. The paper compares multiple constraint-handling approaches within a unified experimental framework and analyzes how best to choose constraint placement.

  2. The evaluation metrics are relatively broad, including statistical scores and the discrepancy between physical feasibility and actual utility.

Critical weaknesses
  1. The paper does not contain much research content. Figures occupy a large amount of space, and the main text repeatedly restates the numbers in the figures and tables but rarely explains these results further. I see a relatively complete experimental record, but not enough clear methodological innovation or mechanistic understanding; remaining at the experimental level is insufficient.

  2. The writing is very strange. It repeatedly emphasizes pre-registration, proactively reporting negative results, and what claims are not made. However, I rarely see the authors propose new viewpoints or interesting insights.

  3. The theoretical part is too simple. The main conclusions follow directly from the basic properties of null spaces, orthogonal projections, and linear invariants.

Questions for authors
  1. Could the paper be reorganized to reduce the oversized figures and allocate more space to theoretical derivations and analysis of the underlying reasons for the experimental results?

  2. Compared with existing projection, null-space parameterization, and constraint completion methods, what new ideas and innovations does this paper specifically provide beyond ranking them?

  3. Regarding the deterioration of downstream cost, coverage correlation is insufficient to prove the proposed mechanism. If the scenarios are calibrated or their dispersion is adjusted, would the scheduling performance of the exact method improve?

Reviewer_5NyBOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper is working on constrain the flow matching models to enhancing its generation performance for power systems.

Critical strengths
  1. Code and project repo are provided which should ensure the high reproducibility of this paper, tho I did not try to run the code by myself.
Critical weaknesses
  1. Overall the presentation is bad (and looks like heavily LLM-generated). Lots of the presentations are hard to understand (see my questions below). The methodology section (Section 3) is very thin and I literally cannot understand what research problem this paper is trying to working on.
  2. The theorem 1 and proposition 1 do not have proof provided.
Questions for authors

I would only list a small part of the vague presentations here as there are literally too many such cases.

  1. in the abstract, it mentioned “leading the strongest baseline by 3.8% (t=−6.13)” but it is not clear what does t mean here
  2. In abs “a worst-case constraint violation 727,790×smaller than unconstrained flow matching” it is vague how “violation can be smaller than something”
  3. In abs “Soft penalties, the field’s default, lose on both axes at once” but it is not clear what are these two “axes”.
  4. In abs “Raising λ from 1 to 1000” but it is not clear what lambda means here.
Reviewer_EYV9OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents a generative model to generate some (not clear) grid-related scenarios following the Kirchoff’s law. The manuscript is poorly written and extremely difficult to read. This reviewer read up to Section 3 and had to guess more than understanding. The abstract also reveals the personal GitHub repo and website.

Critical strengths

Unfortunately, none due to the writing quality.

Critical weaknesses
  • The manuscript needs carefully edited.
Questions for authors

Not at the moment.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.