pith. sign in

arxiv: 2509.21668 · v2 · submitted 2025-09-25 · 📡 eess.SY · cs.SY

NEO-Grid: A Neural Approximation Framework for Optimization and Control in Distribution Grids

Mohamad Chehade , Hao Zhu This is my paper

Pith reviewed 2026-05-18 13:15 UTC · model grok-4.3

classification 📡 eess.SY cs.SY
keywords distribution gridsvolt-var optimizationvolt-var controlneural network surrogatesdeep equilibrium modelsvoltage regulationReLU networkspower flow approximation
0
0 comments X

The pith

Neural surrogates using piecewise-linear ReLU networks and deep equilibrium models improve voltage regulation in distribution grids over linear baselines.

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

The paper establishes that a unified framework called NEO-Grid can replace traditional linear approximations of power flow with neural network surrogates to handle volt-var optimization and closed-loop control in modern distribution grids. It trains piecewise-linear ReLU networks offline to capture nonlinear relationships between power injections and voltages, then uses deep equilibrium models to compute fixed points for inverter responses without iterative simulation. Evaluation on the IEEE 33-bus system shows better voltage regulation performance than standard linear and heuristic methods in both optimization and control tasks. A reader would care because rising distributed energy resources create dynamic voltage stability challenges that conventional tools struggle to address at scale.

Core claim

NEO-Grid replaces linear approximations with piecewise-linear ReLU networks trained to capture the nonlinear relationship between power injections and voltage magnitudes, and models the recursive interaction between voltage and inverter response using deep equilibrium models that allow direct fixed-point computation and efficient training via implicit differentiation. On the IEEE 33-bus system this yields significantly improved voltage regulation performance compared to standard linear and heuristic baselines in both optimization and control settings.

What carries the argument

Piecewise-linear ReLU networks that serve as surrogates for nonlinear power-flow equations, combined with deep equilibrium models that enable fixed-point computation for closed-loop volt-var control.

If this is right

  • The approach yields measurable gains in voltage regulation on the IEEE 33-bus system for both optimization and closed-loop control.
  • Deep equilibrium models permit direct fixed-point solving and implicit differentiation without unrolling iterations.
  • The framework maintains interpretability while scaling to dynamic conditions introduced by distributed energy resources.
  • It provides a single learned model usable for both open-loop optimization and real-time feedback control.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the ReLU surrogates generalize beyond the IEEE 33-bus cases, the same training pipeline could be reused for larger or differently configured distribution networks without full retraining.
  • The implicit differentiation step for DEQs may reduce online computation time enough to support higher-frequency control updates than traditional iterative solvers allow.
  • Similar neural-surrogate plus equilibrium modeling could be tested on related problems such as optimal power flow in transmission systems or voltage control in microgrids.
  • Offline training of the ReLU networks could be combined with online adaptation if new DER patterns shift the operating region.

Load-bearing premise

The offline-trained ReLU networks remain accurate enough surrogates for the true nonlinear power-flow equations across the operating conditions that arise in both optimization and closed-loop control.

What would settle it

On the IEEE 33-bus test system, measure voltage magnitude deviations under the same DER scenarios; if the NEO-Grid controller produces larger or equal deviations than the linear baseline across multiple runs, the performance claim is refuted.

Figures

Figures reproduced from arXiv: 2509.21668 by Hao Zhu, Mohamad Chehade.

Figure 1
Figure 1. Figure 1: Overview of NEO-Grid as a neural digital twin for optimizing the volt-var control (VVC) rules. Left: a distribu￾tion network controlled by piecewise-linear VVC rules at all nodes in D. Center: neural models for (top) the nonlinear PF equations of the voltage-power mapping and (bottom) the VVC rules using a structured ReLU-based layers. Right: efficient ML solver to optimize the overall composited neural mo… view at source ↗
read the original abstract

The rise of distributed energy resources (DERs) is reshaping modern distribution grids, introducing new challenges in attaining voltage stability under dynamic and decentralized operating conditions. This paper presents NEO-Grid, a unified learning-based framework for volt-var optimization (VVO) and volt-var control (VVC) that leverages neural network surrogates for power flow and deep equilibrium models (DEQs) for closed-loop control. Our method replaces traditional linear approximations with piecewise-linear ReLU networks trained to capture the nonlinear relationship between power injections and voltage magnitudes. For control, we model the recursive interaction between voltage and inverter response using DEQs, allowing direct fixed-point computation and efficient training via implicit differentiation. We evaluated NEO-Grid on the IEEE 33-bus system, demonstrating that it significantly improves voltage regulation performance compared to standard linear and heuristic baselines in both optimization and control settings. Our results establish NEO-Grid as a scalable, accurate, and interpretable solution for learning-based voltage regulation in distribution grids.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript presents NEO-Grid, a unified learning-based framework for volt-var optimization (VVO) and volt-var control (VVC) in distribution grids with high DER penetration. It replaces traditional linear power-flow approximations with piecewise-linear ReLU neural-network surrogates trained offline to capture nonlinear injection-to-voltage mappings, and employs deep equilibrium models (DEQs) to represent the recursive voltage-inverter interaction for closed-loop control, enabling fixed-point solutions via implicit differentiation. The framework is evaluated on the IEEE 33-bus system and claims significantly improved voltage regulation performance relative to standard linear and heuristic baselines in both optimization and control settings.

Significance. If the surrogate accuracy and closed-loop robustness hold, the work could provide a practical route to more accurate yet tractable voltage regulation in modern distribution grids by moving beyond linear approximations while retaining computational efficiency through implicit differentiation. The combination of supervised ReLU surrogates with DEQ-based control is a coherent technical choice that builds directly on established techniques in learning for power systems.

major comments (2)
  1. [Abstract] Abstract: the claim of significant performance gains on the IEEE 33-bus system is stated without any quantitative error metrics, training-set coverage details, or direct comparison against a full nonlinear AC power-flow solver, leaving open whether the reported improvements over linear/heuristic baselines are robust or sensitive to post-hoc tuning.
  2. [Method and Evaluation] Method and Evaluation sections: the central assumption that offline-trained piecewise-linear ReLU networks remain sufficiently accurate surrogates for the nonlinear power-flow map across all operating points visited during both VVO optimization and DEQ-based closed-loop VVC is not supported by explicit validation under closed-loop trajectory shifts; because training data generation and coverage over the full DER injection space are not described as exhaustive, extrapolation error could directly degrade the learned optimizer and fixed-point controller relative to the baselines.
minor comments (2)
  1. [Method] Notation for the DEQ fixed-point iteration and the implicit differentiation step could be made more explicit by adding a numbered equation that defines the equilibrium map and the gradient computation.
  2. [Numerical Results] Figure captions and axis labels in the numerical results should include the exact voltage deviation metric and the number of Monte-Carlo scenarios used for each bar or curve.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed review, as well as the positive overall assessment of NEO-Grid. We have revised the manuscript to address the two major comments and provide additional quantitative support and validation details.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim of significant performance gains on the IEEE 33-bus system is stated without any quantitative error metrics, training-set coverage details, or direct comparison against a full nonlinear AC power-flow solver, leaving open whether the reported improvements over linear/heuristic baselines are robust or sensitive to post-hoc tuning.

    Authors: We agree that the abstract would benefit from quantitative support. In the revised version we have added specific metrics: the ReLU surrogate achieves a mean absolute voltage error of 0.004 p.u. relative to the nonlinear AC solver on held-out test points, with training data consisting of 12,000 samples generated via Latin-hypercube sampling over DER injections ranging from 0 % to 150 % of nominal. All methods (including baselines) were tuned with an identical grid-search protocol on the same validation split; we also report a sensitivity study in the supplement showing that performance gains persist across reasonable hyper-parameter variations. revision: yes

  2. Referee: [Method and Evaluation] Method and Evaluation sections: the central assumption that offline-trained piecewise-linear ReLU networks remain sufficiently accurate surrogates for the nonlinear power-flow map across all operating points visited during both VVO optimization and DEQ-based closed-loop VVC is not supported by explicit validation under closed-loop trajectory shifts; because training data generation and coverage over the full DER injection space are not described as exhaustive, extrapolation error could directly degrade the learned optimizer and fixed-point controller relative to the baselines.

    Authors: We acknowledge the value of explicit closed-loop validation. The revised manuscript now includes a dedicated subsection that (i) details the data-generation procedure (Latin-hypercube sampling over the full feasible DER injection space together with full AC power-flow labels) and (ii) reports surrogate error statistics collected along actual DEQ closed-loop trajectories on the IEEE 33-bus system. These results show that the maximum voltage approximation error remains below 0.007 p.u. during control operation and that the voltage-regulation improvement over baselines is preserved. We believe this directly addresses concerns about extrapolation under trajectory shifts. revision: yes

Circularity Check

0 steps flagged

NEO-Grid uses standard supervised neural surrogates and implicit differentiation without reducing claims to self-defined inputs

full rationale

The framework trains piecewise-linear ReLU networks offline to approximate nonlinear power-flow maps and applies DEQs for fixed-point control via implicit differentiation. These are standard techniques whose training objectives and performance metrics are defined externally to the fitted parameters. Evaluation on the IEEE 33-bus system compares against independent linear and heuristic baselines, with no equations or self-citations that force the reported voltage-regulation gains to be equivalent to quantities defined by the model itself. The derivation chain remains self-contained.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the modeling assumption that a trained piecewise-linear network can stand in for the AC power-flow equations and that the fixed-point iteration of voltage and inverter response can be solved directly via implicit differentiation.

free parameters (1)
  • ReLU network weights and biases
    Parameters of the surrogate networks are fitted to power-flow data.
axioms (1)
  • domain assumption Piecewise-linear ReLU networks can accurately approximate the nonlinear mapping from power injections to voltage magnitudes.
    This assumption underpins the replacement of traditional linear approximations.

pith-pipeline@v0.9.0 · 5699 in / 1271 out tokens · 42026 ms · 2026-05-18T13:15:57.448011+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages

  1. [1]

    Ieee standard for interconnection and interoper- ability of distributed energy resources with associ- ated electric power systems interfaces,

    “Ieee standard for interconnection and interoper- ability of distributed energy resources with associ- ated electric power systems interfaces,”IEEE Std 1547-2018 (Revision of IEEE Std 1547-2003), pp. 1–138, 2018

    work page 2018

  2. [2]

    An overview of issues related to IEEE std 1547-2018 for DER,

    D. Narang, A. Warren, and B. Kroposki, “An overview of issues related to IEEE std 1547-2018 for DER,” National Renewable Energy Labora- tory (NREL), Tech. Rep. NREL/TP-5D00-77156, 2021

    work page 2018

  3. [3]

    Distributed volt/var control by PV inverters,

    P. Jahangiri and C. D. Aliprantis, “Distributed volt/var control by PV inverters,”IEEE Transac- tions on Power Systems, vol. 28, no. 3, pp. 3429– 3439, 2013

    work page 2013

  4. [4]

    Fast local voltage control under limited reactive power: Optimality and sta- bility analysis,

    H. Zhu and H. J. Liu, “Fast local voltage control under limited reactive power: Optimality and sta- bility analysis,”IEEE Transactions on Power Sys- tems, vol. 31, no. 5, pp. 3794–3803, 2016

    work page 2016

  5. [5]

    Bi-level volt-V AR optimization to coordinate smart inverters with voltage control devices,

    R. R. Jha, A. Dubey, C. Liu, and K. P. Schneider, “Bi-level volt-V AR optimization to coordinate smart inverters with voltage control devices,” IEEE Transactions on Power Systems, vol. 34, no. 3, pp. 1801–1813, 2019. [Online]. Available: https://www.osti.gov/biblio/1508047

    work page arXiv 2019

  6. [6]

    Two- stage volt-V Ar optimization of distribution grids with smart inverters and legacy devices,

    A. Savasci, A. Inaolaji, and S. Paudyal, “Two- stage volt-V Ar optimization of distribution grids with smart inverters and legacy devices,”IEEE Transactions on Industry Applications, vol. 58, no. 5, pp. 5711–5723, 2022. [Online]. Available: https://discovery.fiu.edu/display/pub246762

    work page 2022

  7. [7]

    A novel robust volt/Var optimization method based on worst-case scenarios for distribution network operation,

    M. Ahmadi, A. Salami, and M. H. Alavi, “A novel robust volt/Var optimization method based on worst-case scenarios for distribution network operation,”IET Generation, Transmis- sion & Distribution, 2023. [Online]. Avail- able: https://ietresearch.onlinelibrary.wiley.com/ doi/abs/10.1049/gtd2.12777

    work page doi:10.1049/gtd2.12777 2023

  8. [8]

    Bi-level volt/V AR optimization in distribution networks with smart pv inverters,

    Y . Long and D. S. Kirschen, “Bi-level volt/V AR optimization in distribution networks with smart pv inverters,”IEEE Transactions on Power Systems, 2022, early access. [Online]. Available: https://arxiv.org/abs/2201.05267

    work page arXiv 2022

  9. [9]

    Equilibrium and dynamics of local voltage control in distribution systems,

    M. Farivar and S. H. Low, “Equilibrium and dynamics of local voltage control in distribution systems,” inProc. IEEE Conference on Decision and Control (CDC), 2013, pp. 4329–4334

    work page 2013

  10. [10]

    Optimal capacitor placement on radial distribution systems,

    M. E. Baran and F. F. Wu, “Optimal capacitor placement on radial distribution systems,”IEEE Transactions on power Delivery, vol. 4, no. 1, pp. 725–734, 2002

    work page 2002

  11. [11]

    Network reconfiguration in distribution systems for loss reduction and load balancing,

    ——, “Network reconfiguration in distribution systems for loss reduction and load balancing,” IEEE Transactions on Power delivery, vol. 4, no. 2, pp. 1401–1407, 2002

    work page 2002

  12. [12]

    Distribution system behind- the-meter ders: Estimation, uncertainty quantifi- cation, and control,

    A. Srivastava, J. Zhao, H. Zhu, F. Ding, S. Lei, I. Zografopoulos, R. Haider, S. Vahedi, W. Wang, G. Valverdeet al., “Distribution system behind- the-meter ders: Estimation, uncertainty quantifi- cation, and control,”IEEE Transactions on Power Systems, 2024

    work page 2024

  13. [13]

    Neural networks for power flow: Graph neural solver,

    B. Donon, T. Marecek, and C. McCoy, “Neural networks for power flow: Graph neural solver,” inNeurIPS Workshop on Graph Representation Learning, 2019, arXiv:1912.01762

    work page arXiv 2019

  14. [14]

    Power- flownet: A quasi-physics-informed graph neural network for ac power flow,

    Y . Chen, H. Hu, Z. Xu, and Y . Weng, “Power- flownet: A quasi-physics-informed graph neural network for ac power flow,”Advances in Applied Energy, vol. 14, p. 100165, 2024

    work page 2024

  15. [15]

    Data-driven mod- eling of linearizable power flow for large- scale grid topology optimization,

    Y .-h. Cho and H. Zhu, “Data-driven mod- eling of linearizable power flow for large- scale grid topology optimization,”arXiv preprint arXiv:2409.13956, 2024

    work page arXiv 2024

  16. [16]

    A review of mixed-integer optimization with ReLU neural networks: Formulations and applications,

    X. Pan, T. A. Davis, O. Dénèreaz, F. Glineur, M. M. V . Barel, O. De Weck, and É. B. Duriez, “A review of mixed-integer optimization with ReLU neural networks: Formulations and applications,”Annual Reviews in Control, 2020, arXiv:2009.08161

    work page arXiv 2020

  17. [17]

    A comprehen- sive review of graph neural networks in power sys- tems,

    W. Liao, J. He, Q. Wu, Y . Liuet al., “A comprehen- sive review of graph neural networks in power sys- tems,”Electric Power Systems Research, vol. 190, p. 106790, 2021, arXiv:2108.06674

    work page arXiv 2021

  18. [18]

    Deep equi- librium models,

    S. Bai, J. Z. Kolter, and V . Koltun, “Deep equi- librium models,”Advances in neural information processing systems, vol. 32, 2019

    work page 2019

  19. [19]

    Multiscale deep equilibrium models,

    S. Bai, V . Koltun, and J. Z. Kolter, “Multiscale deep equilibrium models,”Advances in neural information processing systems, vol. 33, pp. 5238– 5250, 2020

    work page 2020

  20. [20]

    Iterative procedures for non- linear integral equations,

    D. G. Anderson, “Iterative procedures for non- linear integral equations,”Journal of the ACM, vol. 12, no. 4, pp. 547–560, 1965

    work page 1965

  21. [21]

    Anderson acceleration for fixed-point iterations,

    H. F. Walker and P. Ni, “Anderson acceleration for fixed-point iterations,”SIAM Journal on Numeri- cal Analysis, vol. 49, no. 4, pp. 1715–1735, 2011

    work page 2011

  22. [22]

    Optimal design of volt/var control rules of invert- ers using deep learning,

    S. Gupta, V . Kekatos, and S. Chatzivasileiadis, “Optimal design of volt/var control rules of invert- ers using deep learning,”IEEE Transactions on Smart Grid, 2024

    work page 2024

  23. [23]

    Deep learning for scalable opti- mal design of incremental volt/var control rules,

    S. Gupta, A. Mehrizi-Sani, S. Chatzivasileiadis, and V . Kekatos, “Deep learning for scalable opti- mal design of incremental volt/var control rules,” IEEE Control Systems Letters, vol. 7, pp. 1957– 1962, 2023

    work page 1957

  24. [24]

    Data-driven opti- mal voltage regulation using input convex neural networks,

    Y . Chen, Y . Shi, and B. Zhang, “Data-driven opti- mal voltage regulation using input convex neural networks,”Electric Power Systems Research, vol. 189, p. 106741, 2020

    work page 2020

  25. [25]

    Modeling the ac power flow equa- tions with optimally compact neural networks: Application to unit commitment,

    A. Kody, S. Chevalier, S. Chatzivasileiadis, and D. Molzahn, “Modeling the ac power flow equa- tions with optimally compact neural networks: Application to unit commitment,”Electric Power Systems Research, vol. 213, p. 108282, 2022

    work page 2022

  26. [26]

    Topology-aware graph neural networks for learning feasible and adaptive ac-opf solutions,

    S. Liu, C. Wu, and H. Zhu, “Topology-aware graph neural networks for learning feasible and adaptive ac-opf solutions,”IEEE Transactions on Power Systems, vol. 38, no. 6, pp. 5660–5670, 2022

    work page 2022

  27. [27]

    Efficient con- straint learning for data-driven active distribution network operation,

    G. Chen, H. Zhang, and Y . Song, “Efficient con- straint learning for data-driven active distribution network operation,”IEEE Transactions on Power Systems, vol. 39, no. 1, pp. 1472–1484, 2024

    work page 2024

  28. [28]

    W. E. Hart, C. D. Laird, J.-P. Watson, D. L. Woodruff, G. A. Hackebeil, B. L. Nicholson, J. D. Siirolaet al.,Pyomo-optimization modeling in python. Springer, 2017, vol. 67

    work page 2017

  29. [29]

    Omlt: Opti- mization & machine learning toolkit,

    F. Ceccon, J. Jalving, J. Haddad, A. Thebelt, C. Tsay, C. D. Laird, and R. Misener, “Omlt: Opti- mization & machine learning toolkit,”Journal of Machine Learning Research, vol. 23, no. 349, pp. 1–8, 2022

    work page 2022

  30. [30]

    CBC (Coin-or branch and cut),

    J. Forrest, “CBC (Coin-or branch and cut),” https: //github.com/coin-or/Cbc, 2023, version 2.10.10

    work page 2023

  31. [31]

    Chapter 4: Deep equilibrium models,

    Deep Implicit Layers Tutorial Team, “Chapter 4: Deep equilibrium models,” https://implicit-layers-tutorial.org/deep_ equilibrium_models/, accessed: 2025-06-15

    work page 2025