Learning to Optimally Dispatch Power: Performance on a Nation-Wide Real-World Dataset

Enzo Coppes; Federico Larroca; Ignacio Boero; Pablo Belzarena; Santiago Diaz; Tom\'as V\'azquez

REVIEW 4 major objections 5 minor 14 references

Learning to Optimally Dispatch Power: Performance on a Nation-Wide Real-World Dataset

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Using real-world demand and generation data, learned optimal reactive power dispatch models suffer roughly tenfold higher prediction error and frequent constraint violations on Uruguay's grid compared with synthetic benchmarks.

desk verdict A valuable new real-world ORPD dataset, but the paper's central claim hinges on reactive-power inputs whose provenance is never disclosed. read the letter →

arxiv 2505.24505 v1 pith:7BKWYEOS submitted 2025-05-30 cs.LG math.OC

Ignacio Boero , Santiago Diaz , Tomás Vázquez , Enzo Coppes , Pablo Belzarena , Federico Larroca This is my paper

classification cs.LGmath.OC

keywords optimalreactivepowerdispatchlearningtooptimizegraphneuralnetworksfullyconnectedreal-worldsystemdatasetsyntheticversusrealdataconstraintfeasibilityUruguaytransmissiongrid

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper asks whether learning-to-optimize models trained to solve the Optimal Reactive Power Dispatch (ORPD) problem keep their promised accuracy when fed real operational data instead of synthetic samples. It introduces a public dataset of Uruguay's 107-bus transmission network with nearly two years of hourly generation and demand records, and trains standard fully-connected and graph neural networks to imitate IPOPT solutions. On synthetic inputs, both models track the optimizer closely; on real inputs, the mean absolute error in predicted reactive power grows by an order of magnitude and the share of feasible solutions drops from roughly 98% to 61-65%. The authors interpret this as evidence that existing learning-based ORPD architectures are not yet reliable for real grids and need more expressive models or constraint-enforcing mechanisms.

What carries the argument

The carrying object is the learned mapping $\Phi_\theta: X \to Y^*$ that approximates the ORPD solution operator, where $X$ collects each bus's active and reactive load, static generation, and voltage-generator active power, and $Y^*$ holds the optimal voltage setpoints and compensator reactive powers computed offline by the pandapower IPOPT solver. The problem being imitated is loss-minimizing reactive dispatch subject to the AC power flow equations (using the $\pi$-model for lines), voltage bounds, line flow limits, and reactive capability constraints. The decisive comparison is between two training-input distributions: uniform $\pm 30\%$ samples around time-averaged values versus real hourly records whose distributions are multi-modal (wind, hydro, solar) and seasonally structured. The dataset's 107 buses, 156 lines and transformers, 43 generators, 55 loads, and 6 reactive compensators provide the testbed on which the same architectures, tuned with the same hyperparameter search, degrade.

What would settle it

Retrain the same FCNN and GNN on the same real historical inputs but with each feature's values rank-transformed to a uniform distribution, keeping the identical pandapower grid model; if the mean absolute error and feasibility numbers return to synthetic levels, the paper's attribution of degradation to real-data distributions is confirmed, and if they do not, the grid model or optimization pipeline is the culprit.

Watch

Extended reading notes

Core claim

The central claim is that the statistical properties of real-world demand and generation - non-uniform, multi-modal, seasonal, and cross-correlated across nodes - break the standard learning-to-optimize pipeline for ORPD. Trained on 10,000 synthetic instances sampled uniformly within plus or minus 30% of historical averages, an FCNN and a GNN both reach near-optimal losses and feasible constraints in about 98% of test cases. When the same architectures are retrained on chronological historical data from the Uruguayan grid, losses stay numerically close to the IPOPT optimum, but the errors in the actual control variables - voltage setpoints and compensator reactive power - grow by an order of magnitude (mean absolute reactive power error from 0.28/0.29 MVar to 2.75/3.76 MVar), and feasibility drops to 64.8% for the FCNN and 61.4% for the GNN. Even under a 1.8% constraint relaxation that makes all IPOPT solutions feasible, the GNN still violates constraints in about 25% of instances, versus roughly 7% for the FCNN. The paper concludes that synthetic benchmarks materially overstate the readiness of learning-based ORPD for real grid operation.

Load-bearing premise

The whole measurement depends on the assumption that the grid topology, line parameters, generator limits, and operational records supplied by the system operator, as modeled in pandapower, faithfully represent Uruguay's real transmission network; if that model is structurally wrong, the observed tenfold degradation could be an artifact of model mismatch rather than a property of real data.

Editorial extensions

If this is right

Reported synthetic-data accuracies (loss within 0.2% of optimal, about 98% feasibility) should not be read as real-world performance; the same models on real inputs drop to 61-65% feasibility.
The public Uruguay dataset gives a standardized real-world benchmark so future ORPD learning methods can be compared on realistic demand and generation statistics.
Chronological train/validation/test splits matter: learning on historical data must face temporal and seasonal shifts, unlike random splits of synthetic samples.
The fact that IPOPT itself only converges on about 75% of real instances and needs a 1.8% relaxation to reach full feasibility shows ORPD on real data is harder for traditional solvers too, setting a higher bar for learned approximations.
Constraint-enforcing or feasibility-repair techniques would need to be evaluated against this real-data baseline, since even relaxed constraints are violated by the GNN in about a quarter of test instances.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: The degradation pattern suggests the cause is distributional complexity rather than topology; retraining on the same real topology but with uniform synthetic inputs should recover synthetic-level accuracy, a test directly enabled by the released code and data.
Editorial inference: The multi-modal generation distributions, including zero-inflated nighttime solar, imply that models given source type or time-of-day features may close part of the gap; this is a testable extension the paper does not pursue.
Editorial inference: The 1.8% relaxation curve provides a practical operational metric: a 25% violation rate for the GNN even under relaxation means a predict-then-repair pipeline on real data would need to handle roughly one in four instances.
Editorial inference: Transfer from synthetic to real data is likely poor for these models; an explicit pretrain-on-synthetic, fine-tune-on-real experiment would quantify how much real data is needed to recover acceptable feasibility.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The paper is worth taking seriously: it ships a real 107-bus national grid dataset with two years of operational data and shows a stark synthetic-to-real performance drop for learned ORPD. But the central comparison may be built on an undocumented imputation: the real dataset is described in terms of active power, while the model inputs require reactive load and static-generator reactive output, and the paper never says where those come from.

The dataset construction is a real contribution. The topology, line parameters, hourly generation records, and 10-minute load data for Uruguay's grid are public and, as far as I can tell, novel. The chronological split for the historical data is honest, and Table I gives a clear before/after picture. If the degradation is real, it quantifies a gap that many in the learning-to-optimize community suspected but hadn't measured on a real national grid.

The soft spots are mostly minor, except one. The main issue is the reactive input provenance. Eq. (12) defines inputs including q_l and q_g,stat, but Section IV only mentions "generated power" and "total power consumed per bus." The histograms in Figure 1 are labeled "Active Power." I cannot find any statement of how reactive load and static-generator reactive output are obtained for the historical instances. If they are imputed from fixed power factors, set to zero, or sampled synthetically, then the historical labels are solutions to a partly synthetic problem, and the reported tenfold MAE jump could be an artifact of that imputation rule rather than a property of real operational data. This is checkable from the repository, but the manuscript should state it explicitly.

Other issues are secondary. MAEq is reported without error bars or normalization by the target scale, so part of the synthetic-to-real gap might just reflect a larger operating range. About 25% of real instances are discarded because IPOPT doesn't converge; the paper notes this but doesn't assess whether the retained subset is representative. The GNN reports a negative mean loss relative to IPOPT (-0.25±0.70), which is odd and unexplained; it probably reflects the model finding lower-loss but infeasible solutions, but the paper should say so. And Eq. (7) is malformed—an angle-difference constraint can't be written as a product of complex voltages.

None of the secondary issues are fatal. The reactive-input gap is load-bearing. The paper's main conclusion is only as strong as the data actually feeding the models. That said, this is exactly the kind of paper peer review should see: the dataset is valuable, the question is important, and the authors have the code and data to resolve the ambiguity quickly. Send it to review, and require them to disclose the exact construction of the reactive inputs, report error bars or confidence intervals on MAE, reconcile the negative GNN loss, and fix Eq. (7). If the reactive data turn out to be measured, this is a strong benchmark paper. If not, the central claim needs to be reinterpreted.

Referee Report

4 major / 5 minor

Summary. The paper introduces a publicly available ORPD benchmark built from Uruguay's power grid, combining structural data (107 buses, 156 lines, 43 generators) with roughly two years of hourly operational demand and generation records. Using this dataset, the authors train fully-connected and graph neural networks in the 'learning to optimize' framework and compare their performance on synthetic and historical instances of the Optimal Reactive Power Dispatch problem. Their central empirical claim, based on Table I, is that moving from synthetic to real demand and generation inputs increases the mean absolute error in predicted control variables roughly tenfold (MAEq from 0.28–0.29 MVar to 2.75–3.76 MVar) and substantially reduces constraint feasibility (from about 98% to 61–65% of instances). The paper also documents that IPOPT converges on only about 75% of the real instances, and that a 1.8% constraint relaxation makes IPOPT solutions feasible on all retained test cases while learned models still violate constraints in 10–25% of cases.

Significance. If the empirical claims survive scrutiny, the paper makes a useful contribution as both a public dataset and a cautionary benchmark: it demonstrates that synthetic uniform sampling can give over-optimistic estimates of learning-to-optimize performance for ORPD. The chronological train/validation/test split is honest, and the authors are transparent about IPOPT non-convergence and about reporting feasibility as a separate metric. The public repository (DORAA-UY) is a valuable asset for the community. However, the load-bearing claims currently rest on an incompletely specified input pipeline (reactive-power inputs), on an internally questionable loss comparison for the GNN, and on unnormalized error metrics. These issues are fixable but require additional analysis and reporting.

major comments (4)

[Section IV, Eq. (12)] The manuscript does not specify how the reactive components of the input vector x_i, namely q_l and q_g,stat, are obtained for the historical data. Section IV describes only 'generated power for each generator' and 'total power consumed per bus', both of which are naturally active-power quantities, while the ORPD formulation in Eqs. (1)–(11) and the input vector in Eq. (12) depend on reactive loads and static-generator reactive injections. If these reactive values are imputed from fixed power factors, set to zero, or otherwise synthesized, then the historical instances are partly synthetic and the tenfold MAEq increase in Table I could be an artifact of the imputation rule rather than a property of real operational data. Please state the source of every reactive input, and if any imputation was used, report how the results in Table I change under alternative imputation assumptions.
[Table I] The reported GNN historical loss of -0.25±0.70% relative to IPOPT is not interpretable as being 'very close' to the optimal loss, as Section V states. A feasible solution cannot beat the solved optimum, and the negative value is almost certainly explained by constraint violations: the same row reports feasibility of only 61.4% of instances. This makes the loss comparison invalid as evidence of near-optimal performance. Please recompute the loss comparison on the subset of instances where both the IPOPT and the learned solution are feasible, or penalize violations explicitly, and report losses for feasible and infeasible cases separately.
[Table I] MAEq is reported in absolute MVar without normalization by the scale of the target variable. If the historical operating range of reactive-power compensator outputs is wider than the synthetic range, a tenfold increase in absolute MAEq could partly reflect a scale effect rather than a tenfold increase in relative prediction difficulty. Please report MAEq normalized by the standard deviation (or interquartile range) of the target variable in the test set, or show the distributions of the compensator setpoints for both datasets, so that the magnitude of the claimed degradation can be assessed fairly.
[Section V] The chronological split and the retention of only IPOPT-converged instances (about 75% of the total) interact in a way that is not analyzed. If non-convergence is concentrated in particular seasons or operating regimes, the test set no longer represents the full real-world distribution, and the 5% test window (September–December 2023) may not reflect the yearly variability shown in Figures 1 and 2. Please report how convergence failures distribute across the train/validation/test periods, describe the characteristics of discarded instances, and discuss how the main conclusions change if the non-converged instances are excluded rather than silently dropped.

minor comments (5)

[Section V] The phrase 'increasing all constraints by just 1.8%' is imprecise; please state which constraints are relaxed (voltage magnitudes, line flows, reactive limits) and in which direction, and how the 1.8% figure was chosen.
[Figure 4] Figure 4 mixes English and Spanish labels ('Predecido', 'Real', 'Indice', 'Potencia reactiva') and uses 'Real' for ground truth, which is confusing next to Figure 3's 'GT' convention; please unify the language and notation across both figures.
[Table I] The Feas.* column has no entry for the synthetic data rows; please either report the relaxed-feasibility result for synthetic data or explain why it is omitted.
[Section II, Eq. (7)] Equation (7) is written as ∆δ_{ij,min} ≤ v_i v_j^* ≤ ∆δ_{ij,max}, which is dimensionally inconsistent because the middle term is a complex product; this should be reformulated as a phase-angle-difference constraint or the notation should be corrected.
[Figure 2 caption] The caption contains the typo 'in an year'; it should read 'in a year'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical benchmark whose reported comparison does not reduce to its inputs by construction.

full rationale

This paper makes no theoretical derivation whose conclusions are equivalent to its premises. Its central claim is a measured comparison: training the same FCNN and GNN architectures on the synthetic versus the real Uruguayan dataset and reporting MAEv, MAEq, feasibility, and relative losses (Table I). The models are fitted to precomputed IPOPT solutions via a standard supervised loss (Eq. 14), and the reported tenfold MAE increase is an empirical observation on a held-out chronological test split. The synthetic data generation is explicitly described as uniform sampling around historical nominal values (Section IV), so it is not a hidden fit of the target result. The 1.8% constraint relaxation is applied to IPOPT feasibility, not to model predictions, and is disclosed. The only concern raised by the text is that the provenance of reactive-power inputs q_l and q_g,stat used to build historical ORPD instances is not described, so the real-vs-synthetic comparison could in principle be affected by imputation choices; however, this is a data-quality and reproducibility concern, not a circularity. No self-citation, uniqueness claim, or ansatz is used to force the results. The paper is self-contained as an empirical benchmark, and therefore the appropriate circularity score is 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

This is an empirical benchmark paper; no new physical entities or first-principles derivations are introduced. The central claim rests on the fidelity of the dataset and the correctness of the optimization labels.

free parameters (2)

Synthetic sampling range = ±30% of nominal
Chosen following prior work [7], [9]; the breadth of the synthetic distribution directly affects the difficulty of the synthetic baseline and hence the size of the real-data degradation.
Constraint relaxation epsilon = 1.8%
Chosen so that IPOPT solutions become 100% feasible; used to report relaxed feasibility for learning models.

assumptions (4)

domain assumption Power flow equations (Eq. 5) accurately model the Uruguayan transmission network
The ORPD formulation and all loss and feasibility evaluations assume the pi-model and Kirchhoff laws with the given line parameters are correct.
domain assumption IPOPT solutions computed by pandapower are valid ground-truth optimal solutions for training and evaluation
Labels are produced by IPOPT; the paper notes IPOPT itself only meets constraints in 79.8% of real instances, so some labels are infeasible.
domain assumption The DNC operational data (demand, generation, topology) is accurate and representative
The real-world dataset is assembled from DNC data; any errors or metering issues would propagate into the measured degradation.
domain assumption Uniform ±30% sampling around historical averages is a fair synthetic baseline
Synthetic comparison follows prior work [7], [9], but the choice of distribution affects the fairness of the comparison.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Learning to Optimally Dispatch Power: Performance on a Nation-Wide Real-World Dataset." pith.science (2026). https://pith.science/paper/7BKWYEOS

@misc{pith2026250524505,
  author       = {Pith},
  title        = {Pith review of: Learning to Optimally Dispatch Power: Performance on a Nation-Wide Real-World Dataset},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7BKWYEOS}},
  note         = {Machine review of arXiv:2505.24505}
}

read the original abstract

The Optimal Reactive Power Dispatch (ORPD) problem plays a crucial role in power system operations, ensuring voltage stability and minimizing power losses. Recent advances in machine learning, particularly within the ``learning to optimize'' framework, have enabled fast and efficient approximations of ORPD solutions, typically by training models on precomputed optimization results. While these approaches have demonstrated promising performance on synthetic datasets, their effectiveness under real-world grid conditions remains largely unexplored. This paper makes two key contributions. First, we introduce a publicly available power system dataset that includes both the structural characteristics of Uruguay's electrical grid and nearly two years of real-world operational data, encompassing actual demand and generation profiles. Given Uruguay's high penetration of renewable energy, the ORPD problem has become the primary optimization challenge in its power network. Second, we assess the impact of real-world data on learning-based ORPD solutions, revealing a significant increase in prediction errors when transitioning from synthetic to actual demand and generation inputs. Our results highlight the limitations of existing models in learning under the complex statistical properties of real grid conditions and emphasize the need for more expressive architectures. By providing this dataset, we aim to facilitate further research into robust learning-based optimization techniques for power system management.

Figures

Figures reproduced from arXiv: 2505.24505 by the authors.

**Figure 2.** Average hourly renewable energy generation for differ [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Comparison between optimal and predicted voltages [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Comparison between optimal and predicted voltages [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 14 canonical work pages

[1]

An introduction to optimal power flow: Theory, formulation, and examples,

S. Frank and S. Rebennack, “An introduction to optimal power flow: Theory, formulation, and examples,”IIE Transactions, 2016

work page 2016
[2]

Optimal reactive power dispatch with uncertainties in load demand and renewable energy sources adopting scenario-based approach,

P. P. Biswas, P. Suganthan, R. Mallipeddi, and G. A. Amaratunga, “Optimal reactive power dispatch with uncertainties in load demand and renewable energy sources adopting scenario-based approach,”Applied Soft Computing, 2019

work page 2019
[3]

Learning to optimize: A primer and a benchmark,

T. Chen, X. Chen, W. Chen, H. Heaton, J. Liu, Z. Wang, and W. Yin, “Learning to optimize: A primer and a benchmark,”Journal of Machine Learning Research, 2022

work page 2022
[4]

Deep unfolding for commu- nications systems: A survey and some new directions,

A. Balatsoukas-Stimming and C. Studer, “Deep unfolding for commu- nications systems: A survey and some new directions,” in2019 IEEE International Workshop on Signal Processing Systems (SiPS), 2019

work page 2019
[5]

Learning to solve decision problems over two timescales: An application to 5g puncturing,

M. Randall, G. Belcredi, P. Belzarena, and F. Larroca, “Learning to solve decision problems over two timescales: An application to 5g puncturing,” Wireless Personal Communications, 2023

work page 2023
[6]

Amp-inspired deep net- works for sparse linear inverse problems,

M. Borgerding, P. Schniter, and S. Rangan, “Amp-inspired deep net- works for sparse linear inverse problems,”IEEE Transactions on Signal Processing, 2017

work page 2017
[7]

Optimal power flow using graph neural networks,

D. Owerko, F. Gama, and A. Ribeiro, “Optimal power flow using graph neural networks,” in2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2020

work page 2020
[8]

Deepopf: A feasibility- optimized deep neural network approach for ac optimal power flow problems,

X. Pan, M. Chen, T. Zhao, and S. H. Low, “Deepopf: A feasibility- optimized deep neural network approach for ac optimal power flow problems,”IEEE Systems Journal, 2023

work page 2023

Show all 14 references

[9]

Unsupervised optimal power flow using graph neural networks,

D. Owerko, F. Gama, and A. Ribeiro, “Unsupervised optimal power flow using graph neural networks,” in2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024

2024
[10]

Energ ´ıas renovables en uruguay

Uruguay XXI. Energ ´ıas renovables en uruguay. [On- line]. Available: https://www.uruguayxxi.gub.uy/uploads/informacion/ bb0a5729fa124323f033e646cfb13230af675b62.pdf
[11]

Graphs, convolutions, and neural networks: From graph filters to graph neural networks,

F. Gama, E. Isufi, G. Leus, and A. Ribeiro, “Graphs, convolutions, and neural networks: From graph filters to graph neural networks,”IEEE Signal Processing Magazine, 2020

2020
[12]

Stability properties of graph neural networks,

F. Gama, J. Bruna, and A. Ribeiro, “Stability properties of graph neural networks,”IEEE Transactions on Signal Processing, 2020

2020
[13]

pandapower — an open-source python tool for convenient modeling, analysis, and optimization of electric power systems,

L. Thurner, A. Scheidler, F. Sch ¨afer, J. Menke, J. Dollichon, F. Meier, S. Meinecke, and M. Braun, “pandapower — an open-source python tool for convenient modeling, analysis, and optimization of electric power systems,”IEEE Transactions on Power Systems, 2018

2018
[14]

Optuna: A next-generation hyperparameter optimization framework,

T. Akiba, S. Sano, T. Yanase, T. Ohta, and M. Koyama, “Optuna: A next-generation hyperparameter optimization framework,” inACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2019

2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

An introduction to optimal power flow: Theory, formulation, and examples,

S. Frank and S. Rebennack, “An introduction to optimal power flow: Theory, formulation, and examples,”IIE Transactions, 2016

work page 2016

[2] [2]

Optimal reactive power dispatch with uncertainties in load demand and renewable energy sources adopting scenario-based approach,

P. P. Biswas, P. Suganthan, R. Mallipeddi, and G. A. Amaratunga, “Optimal reactive power dispatch with uncertainties in load demand and renewable energy sources adopting scenario-based approach,”Applied Soft Computing, 2019

work page 2019

[3] [3]

Learning to optimize: A primer and a benchmark,

T. Chen, X. Chen, W. Chen, H. Heaton, J. Liu, Z. Wang, and W. Yin, “Learning to optimize: A primer and a benchmark,”Journal of Machine Learning Research, 2022

work page 2022

[4] [4]

Deep unfolding for commu- nications systems: A survey and some new directions,

A. Balatsoukas-Stimming and C. Studer, “Deep unfolding for commu- nications systems: A survey and some new directions,” in2019 IEEE International Workshop on Signal Processing Systems (SiPS), 2019

work page 2019

[5] [5]

Learning to solve decision problems over two timescales: An application to 5g puncturing,

M. Randall, G. Belcredi, P. Belzarena, and F. Larroca, “Learning to solve decision problems over two timescales: An application to 5g puncturing,” Wireless Personal Communications, 2023

work page 2023

[6] [6]

Amp-inspired deep net- works for sparse linear inverse problems,

M. Borgerding, P. Schniter, and S. Rangan, “Amp-inspired deep net- works for sparse linear inverse problems,”IEEE Transactions on Signal Processing, 2017

work page 2017

[7] [7]

Optimal power flow using graph neural networks,

D. Owerko, F. Gama, and A. Ribeiro, “Optimal power flow using graph neural networks,” in2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2020

work page 2020

[8] [8]

Deepopf: A feasibility- optimized deep neural network approach for ac optimal power flow problems,

X. Pan, M. Chen, T. Zhao, and S. H. Low, “Deepopf: A feasibility- optimized deep neural network approach for ac optimal power flow problems,”IEEE Systems Journal, 2023

work page 2023