Reinforcement Learning versus Optimization for Optimal Transmission Switching: A Comparative Study

Israel Abiala; Rachel Gerdes; Yuanrui Sang

REVIEW 2 major objections 6 minor 20 references

Reinforcement Learning versus Optimization for Optimal Transmission Switching: A Comparative Study

T0 review · 2 major / 6 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read A single RL policy matches most of the cost savings of optimal transmission switching at low budgets while deciding two to three orders of magnitude faster than a MILP solver.

desk verdict Useful head-to-head on RTS-96 with clear low-budget savings and large speedups, but the abstract's 'feasible' claim is unsupported by any residual-violation numbers. read the letter →

arxiv 2607.10948 v1 pith:WQNRYQT5 submitted 2026-07-12 eess.SY cs.SY

Yuanrui Sang , Israel Abiala , Rachel Gerdes This is my paper · ORCID

classification eess.SYcs.SY

keywords optimaltransmissionswitchingreinforcementlearningsoftactor-criticbehavioralcloningmixed-integerlinearprogrammingsoft-Lagrangianpenaltiesgrid-enhancingtechnologies

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

Optimal transmission switching opens selected lines to cut generation cost, but the usual mixed-integer formulation grows too slowly for real-time use. This paper asks whether a carefully trained reinforcement-learning agent can serve as a practical surrogate. On the IEEE RTS-96 24-bus system the agent, warm-started by cloning MILP demonstrations and then fine-tuned under a soft-Lagrangian reward that prices every physical constraint, recovers 88–97 percent of the solver’s cost savings when at most three lines may be opened, and still produces feasible (though less optimal) solutions at higher budgets. Once trained, each decision takes roughly 1.4 ms—roughly 250–550 times faster than the MILP solver—making the policy a candidate for real-time operation where exact optimality can be traded for speed.

What carries the argument

Per-constraint soft-Lagrangian reward: every family of physical constraints (generator limits, thermal limits, DC power flow, power balance, angle bounds, line budget) is turned into a non-negative, range-normalized violation whose price is raised or lowered by dual ascent; the resulting fixed multipliers, together with behavioral cloning from MILP demonstrations, train one unified SAC actor that outputs continuous dispatch and switching signals for any budget.

What would settle it

Re-train and re-evaluate the identical pipeline on a larger AC-OPF test system (or under N-1 contingencies) and measure whether the fraction of MILP cost savings and the rate of hard constraint violations remain comparable to the RTS-96 DC results.

Watch

Extended reading notes

Core claim

On the IEEE RTS-96 system a single SAC policy, conditioned only on load factor, switching budget and base-case line loadings, captures 88 percent, 97 percent and 89 percent of MILP-optimal generation-cost savings at budgets C = 1, 2 and 3 (and about 78 percent and 77 percent at C = 4 and 5) while returning feasible dispatch-and-switching decisions two to three orders of magnitude faster than HiGHS.

Load-bearing premise

That soft constraint penalties observed to keep a 24-bus DC model feasible will still produce acceptable real-time decisions when the number of line combinations explodes and AC or security constraints matter.

Editorial extensions

If this is right

For switching budgets of three or fewer lines, a pre-trained RL policy can replace a MILP solve in real-time markets without large economic loss.
Inference time stays essentially constant as the allowed number of open lines grows, removing the combinatorial blow-up that limits MILP solvers.
The recurrent core of high-value lines identified by the agent can be used as a short-list of candidate switches for hybrid or heuristic OTS methods.
Behavioral cloning from a modest number of MILP demonstrations is enough to keep soft-penalty RL inside the feasible region for this class of problems.

Reading between the lines

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

The same soft-Lagrangian-plus-cloning recipe could be tried on other combinatorial grid controls (unit commitment, corrective switching) where exact MILP is too slow.
Because the agent concentrates on a small set of repeatedly useful lines, operators might pre-approve that short-list for faster contingency recovery.
If dual-ascent multipliers converge to similar values across many load snapshots, they themselves become a compact, reusable encoding of constraint tightness for other algorithms.

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 useful part of this paper is the head-to-head numbers on IEEE RTS-96: a single SAC policy (BC-warmed from MILP demos, then soft-Lagrangian fine-tuned) captures 88/97/89% of MILP cost savings at C=1/2/3 and still ~78% at C=4–5, while inference sits at ~1.4 ms (~250–550× faster than HiGHS). That comparison, plus the line-frequency plots showing the agent locks onto the recurrent core (L29–L31, L14–L18) and drops the load-specific tail, is new empirical content and is cleanly reported.

What they do well: methods and hyperparameters are fully specified (Table I), the observation is a sensible congestion map, and they are honest that high-C solutions become suboptimal. The dual-ascent multiplier tuning and per-constraint normalization are incremental engineering, not a conceptual leap, but they are described carefully enough to reproduce the training recipe.

Soft spots, in proportion. The abstract and conclusion repeatedly say the agent produces “feasible solutions.” The reward (Eq. 15) is soft; lines are thresholded at zk<0.5; no projection or repair is applied. Section IV never reports residual power-balance, thermal, or angle violations on the 100 evaluation snapshots. Without those residuals the feasibility claim is an assertion, not a measured result. That is the load-bearing gap. Secondary limits are the usual ones for this subfield: single 24-bus DC system, no AC or N-1, no code release, no error bars. Circularity from BC pretraining is real but secondary—the fine-tuning and high-C degradation show the agent is not just replaying demos.

This is for people already working on real-time topology control or RL-for-OTS who want a concrete baseline and a clear picture of where soft-penalty agents start to lose the MILP optimum. It is not practice-changing and does not resolve the hard-feasibility problem the abstract itself flags in prior work. Still, the empirical comparison is solid enough that a serious editor should send it to referees rather than desk-reject; the authors can be asked to add residual tables (or hard projection) and larger systems. I would skim the figures if I were writing a related methods paper; I would not cite it as settled evidence of feasibility.

Referee Report

2 major / 6 minor

Summary. The paper compares a Soft Actor-Critic (SAC) reinforcement-learning formulation of DC Optimal Transmission Switching (OTS) against a standard MILP benchmark (PuLP/HiGHS) on the IEEE RTS-96 24-bus system. The RL agent is a single unified policy conditioned on load factor, switching budget C ∈ {0……5}, and base-case line loadings; it is warm-started by behavioral cloning from MILP demonstrations and then fine-tuned under a per-constraint soft-Lagrangian reward (Eq. 15) whose multipliers are frozen after dual ascent. On 100 random load factors the policy recovers 88 %, 97 %, and 89 % of MILP cost savings at C = 1, 2, 3 (falling to ~78 %/77 % at C = 4, 5) while producing decisions in ~1.4 ms (~249×–547× faster than the solver). The authors conclude that the agent is near-optimal at low budgets, feasible but suboptimal at high budgets, and well-suited to real-time use.

Significance. A head-to-head optimality-versus-speed comparison of RL and MILP for OTS is useful for the grid-enhancing-technologies community, and the combination of behavioral cloning with a carefully normalized soft-Lagrangian reward is a concrete engineering contribution. The reported wall-clock numbers (Table II) and the explicit saturation of savings beyond C = 3 (Fig. 2) are practically informative. The work remains limited to a single 24-bus DC model and soft constraints, so its significance is that of a carefully executed pilot study rather than a definitive real-time surrogate.

major comments (2)

Abstract and §V claim that the agent generates “feasible solutions” two-to-three orders of magnitude faster than HiGHS. The reward (Eq. 15) is a soft per-constraint Lagrangian with frozen multipliers; lines are declared open by the continuous threshold zk < 0.5 and no hard projection or post-hoc repair is applied. Section IV reports only generation-cost savings and open-line frequency histograms; residual power-balance, thermal, angle, or generator-limit violations (in MW or p.u.) on the 100 evaluation snapshots are never tabulated. Without those residuals the feasibility claim that underpins the speedup comparison is unsupported. A short table of max/mean residual violations (or a statement that all residuals fall below a stated tolerance) is required before the claim can stand.
The evaluation is confined to the IEEE RTS-96 24-bus DC system with C ≤ 5 and ℓ ∈ [0.55, 0.93] (§III–IV). The abstract and introduction frame the work as addressing the poor scaling of MILP for large-scale networks, yet no larger system, AC power-flow check, or N-1 contingency is examined. The central claim that the RL policy is a viable real-time OTS surrogate therefore rests on an untested extrapolation. At minimum the authors should either (i) add results on a second, larger test case or (ii) explicitly qualify the scope of the feasibility and speedup claims to the 24-bus DC setting.

minor comments (6)

Nomenclature lists “Fmin_k / Fmax_k Minimum/Maximum thermal capacity of line k”; the usual convention is that thermal limits are ±Fmax. Clarify whether Fmin is simply −Fmax or a distinct parameter.
Observation dimension is stated as “2 + |L|, making it 40 observation spaces” (§II.B.1). With |L| = 38 this is 40, but the phrasing “40 observation spaces” is awkward; rephrase to “observation dimension 40”.
Fig. 1 caption and surrounding text refer to “RL (SAC-OTS)” while the abstract and title use only “RL”; keep nomenclature consistent.
Table I lists “Balance ε-deadband (annealed) 10 → 0.5 MW” but the annealing schedule is never described; a one-sentence clarification would help reproducibility.
Several references appear with arXiv identifiers that post-date the manuscript’s own arXiv stamp (e.g., [11], [13]); verify that these are intended and correctly cited.
Typographical slips: “V A” for “VA” in the author affiliation, “taking considering one load factor” in the organization paragraph, and inconsistent capitalization of “V oltage”.

Circularity Check

1 steps flagged · score 2.0 of 10

Mild partial circularity only: low-budget near-match to MILP is partly forced by continuous BC regularization toward the same MILP demos used as the evaluation benchmark; the rest of the comparison is independent.

fitted input called prediction [§III Training Algorithm / Table I; §IV.A Cost Savings; contributions bullet on BC]
"In behavioral cloning (BC), one MILP-OTS solution per(ℓ, j) pair (over all six budgets) pre-trains the actor via maximum likelihood. In SAC fine-tuning, the policy is trained with SAC under the per-constraint soft-Lagrangian reward, with BC regularization retained throughout (λ BC = 0.4). [...] Measured against the MILP-optimal savings at each budget, the policy captured 88%, 97% and 89% of the attainable benefit with a switching budget of 1, 2, and 3, respectively."

The policy is continuously pulled toward the identical MILP-OTS solutions that later serve as the optimality benchmark (BC pretrain + λ_BC=0.4 retained in every SAC batch). Low-budget open-line match and cost-savings capture are therefore partly forced by the imitation term rather than independently produced by the soft-Lagrangian objective. High-budget suboptimality shows residual independence, so the circularity is partial and non-load-bearing for the speed claim.

full rationale

This is an empirical RL-vs-MILP benchmark paper, not a first-principles derivation. There is no self-definitional equation, no uniqueness theorem imported from the authors, no ansatz smuggled via self-citation, and no renaming of a known closed-form result. The sole mild circularity is training-procedure leakage: the actor is pretrained on MILP-OTS demonstrations and retains BC regularization (λ_BC=0.4) throughout SAC fine-tuning, then is scored against the same MILP solver on load factors drawn from the same range. Consequently the reported high line-match and 88–97% cost-savings capture at C=1–3 are partly expected by construction of the imitation objective rather than purely discovered by the soft-Lagrangian reward. High-budget suboptimality (C=4,5), the speedup table, and the soft-penalty design itself remain independent of pure imitation. Soft-Lagrangian feasibility is an evidence gap (no residual-violation table), not circularity. Score 2 is proportionate: one non-load-bearing training circularity; central comparative claims retain independent content.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim is an empirical comparison on a standard DC OTS model. Load-bearing background is classical DCOPF/OTS (Fisher et al.) plus SAC. Free parameters are the usual RL/training knobs and the soft-Lagrangian design choices; none are fitted to invent the speedup claim, but several control feasibility and optimality. No new physical entities are postulated.

free parameters (6)

BC regularization weight λ_BC = 0.4
Fixed at 0.4 throughout SAC fine-tuning; controls how strongly the policy stays near MILP demos and thus affects reported optimality gap.
Violation clip κ = 5.0
Caps normalized constraint violations at 5.0 in the reward; chosen for training stability, not derived.
Balance normalizer s_kcl and annealed ε-deadband = 100 MW; 10→0.5 MW
100 MW balance normalizer and 10→0.5 MW deadband set the scale at which power-balance violations bind; hand-chosen characteristic scales.
Per-constraint multipliers λ_i (frozen after dual ascent) = frozen from dual ascent (values not tabulated)
Obtained online by projected dual ascent then frozen; the fixed-λ policy’s feasibility and cost trade-off depend on these learned prices.
SAC/BC training hyperparameters (lr, buffer, batch, hidden size, H_target, steps) = as in Table I
Table I values (e.g., actor/critic lr 3e-4, buffer 4e5, batch 500, [256,256] ELU, 1e5 steps) are free design choices that affect convergence and final policy quality.
Switching-budget range C∈{0…5} and load-factor range ℓ~U(0.5487,0.93) = C≤5; ℓ∈[0.5487,0.93]
Evaluation and training support are restricted to these ranges; claimed near-optimality is only demonstrated inside them.

assumptions (5)

domain assumption DC optimal power flow is an adequate model of the network for OTS decisions (linearized flow, fixed voltage magnitudes, neglect of losses/reactive power).
Entire MILP (1)–(12) and RL environment rest on DCOPF; standard in OTS literature but known to diverge from AC reality.
domain assumption Big-M disjunctive constraints correctly deactivate Ohm’s law on open lines when M is ‘sufficiently large’.
Constraints (6)–(7); quality of MILP demos and the optimality benchmark depend on a valid big-M.
ad hoc to paper Soft (Lagrangian) penalties with frozen multipliers yield operationally acceptable feasibility without hard projection or safety layer.
Reward (15) and training algorithm; paper criticizes soft penalties in prior work yet relies on them, reporting empirical feasibility only.
ad hoc to paper A single-step MDP (γ=0) with continuous relaxation of binary line statuses (threshold 0.5) is a valid decision model for OTS.
§II.B; collapses multi-period/security-constrained switching into one snapshot decision.
domain assumption SAC with Gaussian actor and twin critics converges to a useful policy under the given reward and BC regularization.
Training algorithm cites Haarnoja et al. [20]; standard deep-RL assumption, not proved for this non-convex constrained setting.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Reinforcement Learning versus Optimization for Optimal Transmission Switching: A Comparative Study." pith.science (2026). https://pith.science/paper/WQNRYQT5

@misc{pith2026260710948,
  author       = {Pith},
  title        = {Pith review of: Reinforcement Learning versus Optimization for Optimal Transmission Switching: A Comparative Study},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WQNRYQT5}},
  note         = {Machine review of arXiv:2607.10948}
}

read the original abstract

Optimal Transmission Switching (OTS) reduces generation cost by strategically opening transmission lines, but its mixed-integer linear program (MILP) formulation scales poorly for large-scale transmission networks. Reinforcement learning (RL) offers a computationally efficient alternative, but existing RL-based OTS approaches rely on soft penalties that permit physical constraint violations. This paper presents a comparison between an RL framework and an MILP-based optimization method for OTS. Case studies were carried out on the IEEE RTS-96 24-bus system; results show that the agent was able to produce near-optimal solutions at low switching budgets and tended to yield suboptimal solutions at high switching budgets. However, the RL agent was able to generate feasible solutions two-to-three orders of magnitude faster than the optimization solver.

Figures

Figures reproduced from arXiv: 2607.10948 by the authors.

**Figure 2.** MILP average cost savings across each j for the 100 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 4.** Lines opened (RL vs. MILP) at [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

Figures from the paper (2 more)

**Figure 3.** Figure 3: Lines opened (RL vs. MILP) at low budgets. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png]

**Figure 5.** Figure 5: Lines opened (RL vs. MILP) at high budgets. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 2 linked inside Pith

[1]

The role of optimal transmission switching in enhancing grid flexibility: A review,

M. Numan, M. F. Abbas, M. Yousif, S. S. Ghoneim, A. Mohammad, and A. Noorwali, “The role of optimal transmission switching in enhancing grid flexibility: A review,”IEEE Access, vol. 11, pp. 32 437–32 463, 2023

2023
[2]

National transmission needs study,

U.S. Department of Energy, “National transmission needs study,” Grid Deployment Office, Washington, DC, Tech. Rep., October 2023. [Online]. Available: https://www.energy.gov/gdo/ national-transmission-needs-study

2023
[3]

Optimal transmission switching,

E. B. Fisher, R. P. O’Neill, and M. C. Ferris, “Optimal transmission switching,”IEEE transactions on power systems, vol. 23, no. 3, pp. 1346–1355, 2008

2008
[4]

Co-optimization of generation unit commitment and transmission switching with n-1 reliability,

K. W. Hedman, M. C. Ferris, R. P. O’Neill, E. B. Fisher, and S. S. Oren, “Co-optimization of generation unit commitment and transmission switching with n-1 reliability,”IEEE Transactions on Power Systems, vol. 25, no. 2, pp. 1052–1063, 2010

2010
[5]

Transmission switch- ing in expansion planning,

A. Khodaei, M. Shahidehpour, and S. Kamalinia, “Transmission switch- ing in expansion planning,”IEEE Transactions on Power Systems, vol. 25, no. 3, pp. 1722–1733, 2010

2010
[6]

Fast heuristics for transmission outage coordi- nation,

X. Li and K. Hedman, “Fast heuristics for transmission outage coordi- nation,” in2016 Power Systems Computation Conference (PSCC), 2016, pp. 1–7

2016
[7]

Improving the computational efficiency of optimal transmission switching problems,

L. D. Ramirez-Burgueno, Y . Sang, and N. Santiago, “Improving the computational efficiency of optimal transmission switching problems,” in2022 North American Power Symposium (NAPS), 2022, pp. 1–6

2022
[8]

A bound strengthening method for optimal transmission switching in power systems,

S. Fattahi, J. Lavaei, and A. Atamt ¨urk, “A bound strengthening method for optimal transmission switching in power systems,”IEEE Transac- tions on Power Systems, vol. 34, no. 1, pp. 280–291, 2019

2019

Show all 20 references

[9]

Optimal transmission switching with uncertainties from both renewable energy and n-k contingencies,

T. Han, D. J. Hill, and Y . Song, “Optimal transmission switching with uncertainties from both renewable energy and n-k contingencies,”IEEE Transactions on Sustainable Energy, vol. 14, no. 4, pp. 1964–1978, 2023

1964
[10]

Optimal transmission switching: Improving solver performance using heuristics,

A. Hinneck and D. Pozo, “Optimal transmission switching: Improving solver performance using heuristics,”IEEE Transactions on Power Systems, vol. 38, no. 4, pp. 3317–3330, 2023

2023
[11]

Flow-aware gnn for trans- mission network reconfiguration via substation breaker optimization,

D. Meng, R. Haider, and P. van Hentenryck, “Flow-aware gnn for trans- mission network reconfiguration via substation breaker optimization,” arXiv preprint arXiv:2508.01951, 2025

2025 arXiv
[12]

Optimal transmission switching for short-circuit current limitation based on deep reinforcement learning,

S. Tang, T. Li, Y . Liu, Y . Su, Y . Wang, F. Liu, and S. Gao, “Optimal transmission switching for short-circuit current limitation based on deep reinforcement learning,”Energies, vol. 15, no. 23, p. 9200, 2022

2022
[13]

Physics informed reinforcement learning with gibbs priors for topology control in power grids,

P. Dogoulis and M. Cordy, “Physics informed reinforcement learning with gibbs priors for topology control in power grids,”arXiv preprint arXiv:2604.01830, 2026

2026
[14]

Ac-informed dc optimal transmission switching problems via parameter optimization,

B. Taheri and D. K. Molzahn, “Ac-informed dc optimal transmission switching problems via parameter optimization,”IEEE Transactions on Power Systems, vol. 40, no. 6, pp. 5422–5433, 2025

2025
[15]

A deep reinforcement learning method for multi-objective transmission switching,

D. Lin, J. Wang, T. Zhao, and M. Yue, “A deep reinforcement learning method for multi-objective transmission switching,” in2025 IEEE Power & Energy Society General Meeting (PESGM). IEEE, 2025, pp. 1–5

2025
[16]

A two-stage quantum reinforcement learning method for multi-objective transmission switching,

D. Lin, J. Wang, H.-H. Tseng, T. Zhao, M. Yue, and S. Yoo, “A two-stage quantum reinforcement learning method for multi-objective transmission switching,”IEEE Transactions on Power Systems, 2025

2025
[17]

Research on fast optimal corrective transmission switching based on deep reinforcement learning,

A. Lu, Y . Liu, J. Cheng, and R. Mao, “Research on fast optimal corrective transmission switching based on deep reinforcement learning,” in2024 IEEE PES 16th Asia-Pacific Power and Energy Engineering Conference (APPEEC). IEEE, 2024, pp. 1–5

2024
[18]

Contingency-constrained economic dispatch with safe reinforcement learning,

M. Eichelbeck, H. Markgraf, and M. Althoff, “Contingency-constrained economic dispatch with safe reinforcement learning,” in2022 21st IEEE International Conference on Machine Learning and Applications (ICMLA). IEEE, 2022, pp. 597–602

2022
[19]

Action set based policy optimization for safe power grid management,

B. Zhou, H. Zeng, Y . Liu, K. Li, F. Wang, and H. Tian, “Action set based policy optimization for safe power grid management,” inJoint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, 2021, pp. 168–181

2021
[20]

Soft actor-critic algorithms and applications,

T. Haarnoja, A. Zhou, K. Hartikainen, G. Tucker, S. Ha, J. Tan, V . Ku- mar, H. Zhu, A. Gupta, P. Abbeelet al., “Soft actor-critic algorithms and applications,”arXiv preprint arXiv:1812.05905, 2018

2018 arXiv

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.

[1] [1]

The role of optimal transmission switching in enhancing grid flexibility: A review,

M. Numan, M. F. Abbas, M. Yousif, S. S. Ghoneim, A. Mohammad, and A. Noorwali, “The role of optimal transmission switching in enhancing grid flexibility: A review,”IEEE Access, vol. 11, pp. 32 437–32 463, 2023

2023

[2] [2]

National transmission needs study,

U.S. Department of Energy, “National transmission needs study,” Grid Deployment Office, Washington, DC, Tech. Rep., October 2023. [Online]. Available: https://www.energy.gov/gdo/ national-transmission-needs-study

2023

[3] [3]

Optimal transmission switching,

E. B. Fisher, R. P. O’Neill, and M. C. Ferris, “Optimal transmission switching,”IEEE transactions on power systems, vol. 23, no. 3, pp. 1346–1355, 2008

2008

[4] [4]

Co-optimization of generation unit commitment and transmission switching with n-1 reliability,

K. W. Hedman, M. C. Ferris, R. P. O’Neill, E. B. Fisher, and S. S. Oren, “Co-optimization of generation unit commitment and transmission switching with n-1 reliability,”IEEE Transactions on Power Systems, vol. 25, no. 2, pp. 1052–1063, 2010

2010

[5] [5]

Transmission switch- ing in expansion planning,

A. Khodaei, M. Shahidehpour, and S. Kamalinia, “Transmission switch- ing in expansion planning,”IEEE Transactions on Power Systems, vol. 25, no. 3, pp. 1722–1733, 2010

2010

[6] [6]

Fast heuristics for transmission outage coordi- nation,

X. Li and K. Hedman, “Fast heuristics for transmission outage coordi- nation,” in2016 Power Systems Computation Conference (PSCC), 2016, pp. 1–7

2016

[7] [7]

Improving the computational efficiency of optimal transmission switching problems,

L. D. Ramirez-Burgueno, Y . Sang, and N. Santiago, “Improving the computational efficiency of optimal transmission switching problems,” in2022 North American Power Symposium (NAPS), 2022, pp. 1–6

2022

[8] [8]

A bound strengthening method for optimal transmission switching in power systems,

S. Fattahi, J. Lavaei, and A. Atamt ¨urk, “A bound strengthening method for optimal transmission switching in power systems,”IEEE Transac- tions on Power Systems, vol. 34, no. 1, pp. 280–291, 2019

2019