pith. sign in

arxiv: 2604.17026 · v1 · submitted 2026-04-18 · 📡 eess.SY · cs.SY

Learning a Non-linear Surrogate Model for Multistage Stochastic Transmission Planning

Victor Schmitt , Farzaneh Pourahmadi , Angela Flores-Quiroz , Pablo Apablaza , Pierluigi Mancarella This is my paper

Pith reviewed 2026-05-10 06:53 UTC · model grok-4.3

classification 📡 eess.SY cs.SY
keywords transmission expansion planningstochastic optimizationneural network surrogatemachine learningpower system planningmixed integer linear programmingrenewable energy integration
0
0 comments X

The pith

Surrogate neural networks enable near-optimal multistage stochastic transmission planning with up to 13 times faster computation.

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

The paper develops a hybrid machine learning and optimization method for transmission expansion planning under uncertainty. It trains neural networks on investment decisions and scenarios to approximate the expected costs of operating the grid. These approximations are then embedded as linear constraints in the planning optimization. This reduces the need to solve large operational subproblems repeatedly. The result is investment plans that are close to optimal but computed much faster, up to 13 times quicker than traditional stochastic models.

Core claim

By training surrogate neural networks to predict expected operational costs from investment decisions and uncertainty scenarios, and reformulating them as mixed-integer linear constraints, the multistage stochastic TEP problem can be solved with near-optimal investment costs while achieving computational speedups of up to a factor of 13 compared to full optimization.

What carries the argument

Surrogate neural network reformulated as mixed-integer linear constraints to approximate expected operational costs within the TEP optimization model.

If this is right

  • Investment decisions achieve costs near those of full stochastic optimization.
  • Total computational time is reduced by up to a factor of 13.
  • Extensive multi-scenario analysis and stress testing become feasible at larger scales.

Where Pith is reading between the lines

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

  • This surrogate approach could be adapted to other planning problems involving uncertainty in energy systems.
  • It opens the door to incorporating more detailed operational models without sacrificing tractability.

Load-bearing premise

The neural network provides an accurate enough approximation of expected operational costs for the embedded optimization to produce truly near-optimal investment decisions.

What would settle it

Comparing the true total costs of plans from the surrogate method versus a full stochastic optimization on the same IEEE test systems; a significant gap in true costs would falsify the near-optimality claim.

Figures

Figures reproduced from arXiv: 2604.17026 by Angela Flores-Quiroz, Farzaneh Pourahmadi, Pablo Apablaza, Pierluigi Mancarella, Victor Schmitt.

Figure 1
Figure 1. Figure 1: ML embedding framework. Yellow boxes denote the data preparation [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Scenario tree considered. A 15-year decision horizon is considered. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Training and validation loss curves with and without hyperparameter [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Regression performance as a function of training set size for (a) [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Signed SHAP importance for the ten most influential features in [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Investment decisions for the line candidates. The heatmap displays the [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Solving and Total run time comparison. benchmarking the proposed approach against CG, a commonly adopted decomposition technique for STEP problems, showing a speed-up factor of approximately 20. This computational improvement is largely driven by the re￾duction in optimisation model size achieved with the surrogate formulation. In the explicit stochastic operational formulation, the operational problem lea… view at source ↗
Figure 9
Figure 9. Figure 9: Investment decisions comparison across different [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
read the original abstract

Transmission expansion planning (TEP) plays a critical role in ensuring power system reliability and facilitating the integration of renewable energy resources. However, this process requires planners to constantly deal with significant uncertainty. While multistage stochastic TEP models provide a robust framework for identifying investment plans under uncertainty, the rapid growth in problem size hinders their computational tractability. To address this challenge, this paper develops a hybrid machine learning-optimisation framework for stochastic TEP. The proposed approach uses investment decisions and uncertainty scenarios as input features to train surrogate neural networks, which are then reformulated as mixed-integer linear constraints and embedded within an optimisation model. The surrogate model approximates expected operational costs to inform TEP decisions, reducing the burden arising from large operational problems. Case study applications on IEEE test systems demonstrate that, after training, the proposed approach achieves near-optimal investment costs while reducing total computational time by up to a factor of around 13 compared to a single full-optimisation stochastic formulation. This enables performing extensive multi-scenario analysis and stress testing that would otherwise be computationally prohibitive at scale.

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 paper proposes a hybrid machine learning-optimization framework for multistage stochastic transmission expansion planning (TEP). Investment decisions and uncertainty scenarios are used as inputs to train neural-network surrogates that approximate expected operational costs; these surrogates are reformulated as mixed-integer linear constraints and embedded in the master optimization problem. Case studies on IEEE test systems report that the resulting investment plans achieve near-optimal costs while reducing total computation time by up to a factor of 13 relative to a full stochastic formulation.

Significance. If the surrogate approximations remain sufficiently accurate for investment decisions selected by the optimizer, the approach would meaningfully improve the tractability of multistage stochastic TEP, enabling larger scenario sets and more extensive sensitivity analyses that are currently prohibitive. The reported speedups, if substantiated by rigorous out-of-sample validation, would constitute a practical contribution to computational methods in power-system planning.

major comments (2)
  1. [Abstract] Abstract: The abstract asserts 'near-optimal investment costs' and a factor-of-13 time reduction, yet supplies no quantitative error metrics for the surrogate, no description of the validation procedure used to obtain those costs, and no sensitivity analysis on how approximation error propagates into first-stage investment decisions.
  2. [Case Studies] Case Studies (and training-data generation): Training data are generated from independent full optimizations, but the manuscript provides no analysis of the distance between the investments chosen by the surrogate-embedded master problem and the training support, nor any true-cost gap evaluation on held-out scenarios for those specific optimized investments. This leaves the generalization claim under distribution shift unverified.
minor comments (2)
  1. [Methodology] The reformulation of the trained neural networks as mixed-integer linear constraints would benefit from an explicit small-scale example or pseudocode showing the encoding of ReLU or other activations.
  2. [Model Formulation] Notation for the surrogate input features (investment decisions versus scenario parameters) should be introduced once and used consistently throughout the model formulation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight opportunities to strengthen the clarity and rigor of our claims. We address each major comment point by point below and commit to the indicated revisions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The abstract asserts 'near-optimal investment costs' and a factor-of-13 time reduction, yet supplies no quantitative error metrics for the surrogate, no description of the validation procedure used to obtain those costs, and no sensitivity analysis on how approximation error propagates into first-stage investment decisions.

    Authors: We agree that the abstract would be improved by including quantitative support. In the revised manuscript we will update the abstract to report specific surrogate error metrics (e.g., mean absolute percentage error on expected operational costs from the IEEE case studies), briefly describe the out-of-sample validation procedure used to obtain the reported investment costs, and note the observed sensitivity of first-stage decisions to approximation error. The full sensitivity analysis already appears in Section 4.3; the revision will ensure the abstract explicitly references these results. revision: yes

  2. Referee: [Case Studies] Case Studies (and training-data generation): Training data are generated from independent full optimizations, but the manuscript provides no analysis of the distance between the investments chosen by the surrogate-embedded master problem and the training support, nor any true-cost gap evaluation on held-out scenarios for those specific optimized investments. This leaves the generalization claim under distribution shift unverified.

    Authors: We acknowledge that an explicit quantification of distribution shift would strengthen the generalization argument. The current training set spans a wide range of investment decisions obtained from full stochastic optimizations, and the case studies show that the surrogate-embedded model yields near-optimal plans. However, we did not report investment-space distances or true-cost gaps on held-out scenarios for the final surrogate-derived investments. In the revised manuscript we will add this analysis to the Case Studies section, including a distance metric between the obtained investment vectors and the training support together with true expected-cost evaluations (via the full model) on held-out uncertainty realizations for those specific investment plans. revision: yes

Circularity Check

0 steps flagged

No significant circularity: surrogate trained on independent full optimizations and validated empirically

full rationale

The paper generates training data for the neural-network surrogate from separate full stochastic TEP optimizations on investment decisions and scenarios. The surrogate approximates expected operational costs and is reformulated as MIL constraints for embedding in the master problem. Near-optimality of resulting investment plans is asserted via case studies on IEEE test systems, not derived by construction. No self-definitional steps, fitted inputs renamed as predictions, or load-bearing self-citations appear in the derivation chain. The method is a standard supervised-learning-plus-embedding pipeline whose performance claims rest on external empirical checks rather than tautological reduction to its own inputs.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach rests on standard neural-network training and the assumption that trained networks admit exact MILP reformulations; no new physical entities or ad-hoc constants are introduced beyond the learned weights.

free parameters (1)
  • neural-network weights and architecture
    Determined by supervised training on full-optimization data to approximate operational costs.
axioms (1)
  • domain assumption A trained neural network can be exactly reformulated as a set of mixed-integer linear constraints
    Invoked when the surrogate is embedded inside the planning MILP.

pith-pipeline@v0.9.0 · 5503 in / 1200 out tokens · 47300 ms · 2026-05-10T06:53:45.520863+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages

  1. [1]

    Planning Low-Carbon Elec- tricity Systems Under Uncertainty Considering Operational Flexibility and Smart Grid Technologies,

    R. Moreno, A. Street, J. M. Arroyoet al., “Planning Low-Carbon Elec- tricity Systems Under Uncertainty Considering Operational Flexibility and Smart Grid Technologies,”Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, vol. 375, no. 2100, 2017

    work page 2017

  2. [2]

    How to Assess Uncertainty-Aware Frameworks for Power System Planning?

    E. Spyrou, B. Hobbs, D. Chattopadhyay, and N. Mukhi, “How to Assess Uncertainty-Aware Frameworks for Power System Planning?”IEEE Transactions on Energy Markets, Policy and Regulation, pp. 1–13, 2 2024

    work page 2024

  3. [3]

    A Fast Solution Method for Stochastic Transmission Expansion Planning,

    J. Zhan, C. Chung, and A. Zare, “A Fast Solution Method for Stochastic Transmission Expansion Planning,”IEEE Transactions on Power Sys- tems, 2017

    work page 2017

  4. [4]

    Learning Optimal Solutions for Extremely Fast AC Optimal Power Flow,

    A. S. Zamzam and K. Baker, “Learning Optimal Solutions for Extremely Fast AC Optimal Power Flow,” in2020 IEEE International Conference on Communications, Control, and Computing Technologies for Smart Grids (SmartGridComm), 2020, pp. 1–6

    work page 2020

  5. [5]

    Assessing the Impact of DER on the Expansion of Low-Carbon Power Systems Under Deep Uncertainty,

    P. Apablaza, S. Puschel-Lovengreen, R. Morenoet al., “Assessing the Impact of DER on the Expansion of Low-Carbon Power Systems Under Deep Uncertainty,”Electric Power Systems Research, vol. 235, 2024

    work page 2024

  6. [6]

    A Column Gen- eration Approach for Solving Generation Expansion Planning Problems With High Renewable Energy Penetration,

    A. Flores-Quiroz, R. Palma-Behnke, G. Zakeriet al., “A Column Gen- eration Approach for Solving Generation Expansion Planning Problems With High Renewable Energy Penetration,”Electric Power Systems Research, vol. 136, pp. 232–241, 2016

    work page 2016

  7. [7]

    A Distributed Computing Framework for Multi-Stage Stochastic Planning of Renewable Power Systems With Energy Storage as Flexibility Option,

    A. Flores-Quiroz and K. Strunz, “A Distributed Computing Framework for Multi-Stage Stochastic Planning of Renewable Power Systems With Energy Storage as Flexibility Option,”Applied Energy, vol. 291, p. 116736, 2021

    work page 2021

  8. [8]

    Implementing Mixed Integer Column Generation,

    F. Vanderbeck, “Implementing Mixed Integer Column Generation,” in Desaulniers, G., Desrosiers, J., Solomon, M.M. (eds) Column Genera- tion.Boston, MA: Springer, 2005, pp. 331–358

    work page 2005

  9. [9]

    Benders Cut Classification via Support Vector Ma- chines for Solving Two-Stage Stochastic Programs,

    H. Jia and S. Shen, “Benders Cut Classification via Support Vector Ma- chines for Solving Two-Stage Stochastic Programs,”INFORMS Journal on Optimization, vol. 3, no. 3, pp. 278–297, 2021

    work page 2021

  10. [10]

    Machine Learning-Enhanced Benders Decomposi- tion Approach for the Multi-Stage Stochastic Transmission Expansion Planning Problem,

    S. Borozanet al., “Machine Learning-Enhanced Benders Decomposi- tion Approach for the Multi-Stage Stochastic Transmission Expansion Planning Problem,”Electric Power Systems Research, vol. 237, 2024

    work page 2024

  11. [11]

    Multi-Stage Integrated Transmission and Distribution Expansion Planning Under Uncertainties With Smart In- vestment Options,

    S. Borozan and G. Strbac, “Multi-Stage Integrated Transmission and Distribution Expansion Planning Under Uncertainties With Smart In- vestment Options,”IEEE Transactions on Sustainable Energy, vol. 16, no. 1, pp. 546–559, 2025

    work page 2025

  12. [12]

    Dantzig-Wolfe Decompo- sition for Solving Multistage Stochastic Capacity-Planning Problems,

    K. J. Singh, A. B. Philpott, and R. K. Wood, “Dantzig-Wolfe Decompo- sition for Solving Multistage Stochastic Capacity-Planning Problems,” Operations Research, vol. 57, no. 5, pp. 1271–1286, 2009

    work page 2009

  13. [13]

    Improved surrogate modeling for multi-energy system design: Model architecture, sampling and scaling choices,

    F. L ´ed´ee, C. Crawford, and R. Evins, “Improved surrogate modeling for multi-energy system design: Model architecture, sampling and scaling choices,”Applied Energy, vol. 390, p. 125812, 2025

    work page 2025

  14. [14]

    A surrogate model of distribution networks to support transmission network planning,

    M. Rossini, M. Rossi, and D. Siface, “A surrogate model of distribution networks to support transmission network planning,” in27th Interna- tional Conference on Electricity Distribution (CIRED 2023), 2023

    work page 2023

  15. [15]

    Machine learning–enhanced column generation for large-scale capacity planning problems

    F. Keim, V . Bucarey, Q. Zhang, and ´A. Flores-Quiroz, “Machine learning–enhanced column generation for large-scale capacity planning problems.”

    work page

  16. [16]

    Dumouchelle, R

    J. Dumouchelle, R. Patel, E. B. Khalilet al., “Neur2SP: Neural Two- Stage Stochastic Programming,” arXiv preprint arXiv:2205.12006, 2022

    work page arXiv 2022

  17. [17]

    Hastie, R

    T. Hastie, R. Tibshirani, and J. Friedman,The Elements of Statistical Learning: Data Mining, Inference, and Prediction, 2nd ed. Springer, 2009, see Section 10.6 on loss functions and differentiability

    work page 2009

  18. [18]

    Deep Neural Networks and Mixed Integer Linear Optimization,

    M. Fischetti and J. Jo, “Deep Neural Networks and Mixed Integer Linear Optimization,”Constraints, vol. 23, no. 3, pp. 296–309, 2018

    work page 2018

  19. [19]

    Relu networks as surrogate models in mixed-integer linear programs,

    B. Grimstad and H. Andersson, “Relu networks as surrogate models in mixed-integer linear programs,”Computers & Chemical Engineering, vol. 131, 2019

    work page 2019

  20. [20]

    An Enhanced IEEE 33 Bus Benchmark Test System for Distribution System Studies,

    S. H. Dolatabadi, M. Ghorbanian, P. Sianoet al., “An Enhanced IEEE 33 Bus Benchmark Test System for Distribution System Studies,”IEEE Transactions on Power Systems, vol. 36, no. 3, pp. 2565–2572, 2021

    work page 2021

  21. [21]

    2024 Integrated System Plan ISP,

    AEMO, “2024 Integrated System Plan ISP,” 2024

    work page 2024

  22. [22]

    National Electricity Rules, Clause 3.9.3C: Reliability Standard and Interim Reliability Measure,

    A. E. M. Commission, “National Electricity Rules, Clause 3.9.3C: Reliability Standard and Interim Reliability Measure,” 2020

    work page 2020

  23. [23]

    A unified approach to interpreting model predictions,

    S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” inAdvances in Neural Information Processing Systems, vol. 30, 2017

    work page 2017

  24. [24]

    Ordoudis, P

    C. Ordoudis, P. Pinson, J. M. Moraleset al.,An Updated V ersion of the IEEE RTS 24-Bus System for Electricity Market and Power System Operation Studies. Technical University of Denmark, 2016, vol. 13. 24th Power Systems Computation Conference PSCC 2026 Limassol, Cyprus — June 8-12, 2026

    work page 2016