acceptodds
Under review as a conference paper at ICLR 2027

Learning AC Optimal Power Flow by a Capped Regressor with Topological Attenuation

Abstract

Efficiently solving the Alternating Current Optimal Power Flow (AC-OPF) problem is essential in power systems. Despite substantial progress in learning-based AC-OPF, an important question remains: which physics-motivated patterns can a learning-based load-to-generation predictor encode under the cost-minimization objective? To answer this question, we investigate the physical patterns of load-to-generation relationships: boundedness of total active-power generation and topological attenuation of generator response. Accordingly, we propose a capped regressor with topological attenuation (CReT). To encode the boundedness of total active-power generation, CReT controls the load-relative scale of total active generation by a capping operation, and then employs a label distribution learning model to proportionally allocate this total generation among individual generators. To encourage topological attenuation, CReT utilizes a novel physics-informed graph convolution encoding the nonlinear couplings in power flow, where the locality of message-passing mechanism is consistent with topological attenuation. Finally, CReT is evaluated against recent baselines on multiple bus systems.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
AC Optimal Power Flow, Label Distribution Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_oeoZOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the use of machine learning models to predict solutions to the AC Optimal Power Flow (AC-OPF) problem. The paper's methodology leverages two insights: total (active & reactive) power generation is bounded, and two networks elements far away from one another show less interaction than two elements close to one another. Accordingly, the proposed ML approach first predicts total generation before disaggregating it at the generator level, and utilizes a graph convolutional network (GCN) architecture. Numerical experiments are reported on instances with 9 to 118 buses.

Critical strengths

The paper's novelty is limited.

The one "strength" is to first predict total generated power, which is indeed highly correlated to total demand, then disaggregate it. Most approaches in the literature directly learn the high-dimensional mapping (load vector --> generation & voltage vector).

Critical weaknesses

Most theoretical results in the paper rely on very strong assumptions and are existential at best, and have limited novelty.

  • Boundedness result in Theorem 4.1: immediate from the fact that each generator's active power generation is itself bounded.
  • Assumption 4.2 is not verified in general, as noted later in the paper
  • Theorem 4.3 assumes differentiability of the solution of AC-OPF, which does not hold in general because of the problem's non-convexity and nonlinearity.
  • Proposition 4.4 is tautological (total error is equal to sum of positive and negative error)
  • Theorem 4.6 (generalization error bound) assumes that the per-sample error is bounded, and its result relies on said boundedness

The numerical experiments are conducted on small systems. State-of-the-art results for learning AC-OPF solutions scale to networks with tens of thousands of buses, and there exist publicly-available datasets for problems of this size.

  • CANOS (https://arxiv.org/abs/2403.17660v1) uses GNN models and reports experiments on networks >100x larger than the present paper
  • PGLearn (https://arxiv.org/abs/2505.22825) provides multiple datasets on networks up to tens of thousands of buses, which are available on Hugging Face: https://huggingface.co/PGLearn
  • Using uncorrelated bus-level noise to generate synthetic data produces unrealistic load profiles, because there is little variation in total demand (especially for larger networks); this effect was pointed out in the PGLearn paper above
Questions for authors
  • Please explain the large discrepancy between the reference cost and the reported cost metrics in Table 1 (first column)
  • Please conduct experiments on much larger systems to properly demonstrate the scalability of the proposed approach
Reviewer_edq9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CReT, a physics-motivated neural predictor for AC optimal power flow. A complex-valued graph convolution inspired by power-flow couplings extracts bus features from active and reactive loads. The active-generation head predicts a bounded load-relative total generation, distributes it through softmax weights, and clips individual outputs to generator limits. Reactive generation is predicted through a bounded sigmoid mapping. Voltages are recovered using Newton–Raphson, and training minimizes generation cost plus constraint penalties. The theoretical results address generation bounds, conditional locality, clipping effects, and output-head generalization. Experiments compare against five learning-based baselines on six IEEE systems and include architectural ablations.

Critical strengths

The architecture contains sensible physical inductive biases. Separating total generation from its allocation provides an interpretable parameterization, while PFConv incorporates nonlinear complex-valued interactions rather than using topology alone. These are reasonable design choices for learning grid-dependent dispatch.

The evaluation includes multiple systems and component ablations. The six IEEE benchmarks and comparisons with conventional graph convolution and uncapped architectures provide useful evidence about the contributions of the proposed modules. The paper also explicitly acknowledges that clipping, imperfect voltage recovery, and long-range dependencies limit its physical interpretation.

Critical weaknesses

The methodological contribution and its theoretical justification are limited. The main contribution is an architecture combining bounded output parameterization with a power-flow-inspired graph convolution, rather than a new mechanism addressing feasibility or near-optimality. Theorem 4.3 assumes uniform exponential sensitivity decay and differentiability over the load region—substantially stronger conditions than the average distance-conditioned trend supported by Figure 3. It therefore provides a conditional truncation bound, not a demonstration that AC-OPF satisfies the required locality or that PFConv uniquely captures it. The other results largely comprise physical identities, clipping-error decompositions, and a standard covering-number bound with the feature extractor fixed. Overall, the analysis explains the design but does not strongly establish its distinctive advantage.

The evaluation does not establish improved economic performance at comparable feasibility. Table 1 reports substantial power-balance residuals alongside costs far below the MIPS reference: approximately 303 versus 608 on IEEE 30, and 13,210 versus 41,842 on IEEE 57. Since these predictions are not fully feasible, lower generation cost cannot by itself demonstrate better dispatch. MIPS is not a global-optimality certificate, but the large differences still require explanation and evaluation under matched feasibility tolerances. The absence of formal guarantees is not inherently disqualifying; however, convincing feasible-solution evidence is especially important here. Furthermore, the efficiency motivation is not supported by reported end-to-end timing that includes voltage recovery and its convergence failures.

Questions for authors

Assessment-changing evaluation: What fraction of predictions satisfies all constraints, including active and reactive power balance, at explicitly stated tolerances? What are the cost gaps on feasible predictions, or after applying a common feasibility-restoration procedure to all methods?

Implementation clarification: How does Newton–Raphson recovery handle the slack bus and predicted generator outputs? Are any outputs adjusted, how are failures treated, and does the reported cost correspond to the same dispatch used to evaluate power-balance residuals?

Theory clarification: Is there evidence supporting the uniform sensitivity-decay condition required by Theorem 4.3, beyond the average trend in Figure 3? How does its locality interpretation apply to the complete predictor, whose total-load scaling and softmax allocation introduce global dependence?

Efficiency clarification: What is the end-to-end inference time, including voltage recovery, relative to MIPS and the baselines, and what are the recovery convergence and failure rates?

Reviewer_Jx6KOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on how physics-motivated structure can be incorporated into learning-based AC optimal power flow (AC-OPF). Authors proposes a capped regressor with topological attenuation, named CReT, based on two observations: total active-power generation can be bounded relative to total active load under suitable operating assumptions, and optimal generator responses exhibit an empirical tendency to decrease with generator–load distance. Experiments on six IEEE systems with 9–118 buses report lower generation costs and improvements in several violation measures.

Critical strengths
  • The paper is well-written, with a clear and intuitive presentation.
  • Observed pattern is interesting and barely discussed in the literature
  • Two designed strategies are proposed for the observed patterns.
Critical weaknesses
  1. CReT learns generator setpoints, but still uses Newton–Raphson to recover voltages during training and inference. This is a reasonable design choice, yet the paper does not specify the power-flow formulation, slack/PV/PQ treatment, which predicted generator outputs are retained or recomputed, how gradients pass through the solver, or its convergence and failure rates.

  2. Table 1 reports costs far below the MIPS reference while retaining nonzero power-balance residuals (e.g., 302.8 versus 608.496, with active/reactive residual measures of 0.038/0.060, on 30 buses); without per-case feasibility rates and costs evaluated under matched tolerances, the lower costs do not establish better feasible dispatch.

  3. two proposed patterns are generally valid? The paper invetigts them on small systems, like 30- and 57-bus cases, but it is unclear whether they consistently hold across different networks and operating conditions.

  4. each grid contributes only 700 training and 300 test scenarios sampled within 80–120% of nominal load, and the largest evaluated grid has 118 buses. how about the large-node systems like 1354 and even 13659? the larger training and testing scales? how about the topology differences between the training and testing splits?

Questions for authors

one concern: the MIPS reference costs and learned costs evaluated on exactly the same test scenarios, network models, cost coefficients?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.