Machine Learning Guided Optimal Transmission Switching to Mitigate Wildfire Ignition Risk

Bistra Dilkina; Daniel K. Molzahn; Ryan Piansky; Weimin Huang

REVIEW 3 major objections 4 minor 41 references

Machine Learning Guided Optimal Transmission Switching to Mitigate Wildfire Ignition Risk

T0 review · 3 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read For the Optimal Power Shutoff problem on a realistic California-scale network, an ML-guided hybrid method yields near-optimal line de-energization plans within 30 minutes—quality that a general-purpose solver reaches only with up to 24 hour

desk verdict A genuine ML-guided MILP extension to OPS with a plausible speedup, but a sign error in the written objective undermines the domain-informed bound; fix that and it is a solid paper. read the letter →

arxiv 2510.25147 v3 pith:LK6Z325H submitted 2025-10-29 cs.LG math.OC

Weimin Huang , Ryan Piansky , Bistra Dilkina , Daniel K. Molzahn This is my paper · ORCID

classification cs.LGmath.OC

keywords optimalpowershutoffwildfireignitionriskmachinelearningguidedMILPtransmissionlinede-energizationpredict-and-searchneuraldivinggraphattentionnetworkloadshedding

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 is trying to establish that wildfire-driven power shutdown decisions can be made both fast and near-optimal by training a predictor on the recurring structure of daily OPS instances, then using domain knowledge about how many lines must be de-energized to restrict the search. On 54 hard test instances from a realistic synthetic California grid, the proposed hybrid—domain-informed Predict-and-Search plus Neural Diving with a safety net—reduces the primal integral by 30.94% and the primal gap by 93.10% relative to a 30-minute run of a standard MILP solver. The resulting plans have an average primal gap of 0.02 against best-known values, meaning they shed far less load than a same-timeout conventional solver and match the quality of a 24-hour run. The reason a reader should care is operational: utilities re-solve these problems daily, and current solve times of hours to days make optimized decisions irrelevant for real-time use.

What carries the argument

The load-bearing mechanism is the domain-informed split of the Predict-and-Search constraint into two separate flip budgets, B0 and B1. B0 uses the derived bounds N_min_0 ≤ number of de-energized lines ≤ N_max_0 to choose k0 = N_max_0 and allow at most Delta0 = N_max_0 − N_min_0 flips among the predicted-de-energized set; B1 fixes (or nearly fixes) the predicted-energized set, with a lower-bound-derived flip allowance that acts as a safety net. This converts the raw ML prediction into a small, feasible neighborhood of the original MILP, and it is what lets the solver find high-quality solutions quickly rather than exploring the full space.

What would settle it

For each hard test instance, record the best-known solution and check whether any truly de-energized line falls outside the top k0 = N_max_0 predicted-de-energized set; if this happens on instances where PaS+ND's final gap is nonzero or where it fails to find the best-known solution, the central claim's premise is violated. Simpler: compute 0-class recall—the fraction of optimal de-energized lines that the model predicts de-energized—on the held-out test split; if it is well below 1, the bound-based neighborhood can exclude optima.

Watch

Extended reading notes

Core claim

The central discovery is that the structure of OPS—most lines stay energized, and the number of de-energized lines is bounded between a minimum N_min_0 and maximum N_max_0—can be baked into the inference-time refinement of an ML prediction. A graph attention network trained on historically high-quality incumbents predicts each switchable line's energization probability. Instead of using one budget on all flips, the method splits it: lines predicted de-energized can flip up to N_max_0 minus N_min_0 times, while lines predicted energized are fixed (or nearly fixed) to 1, with a feasibility-restoring buffer when the fixed set is too large. The paper shows this restricted MILP is solved to an av

Load-bearing premise

Everything rests on the assumption that the machine-learning model ranks the truly de-energized lines high enough that none of the bounds-based restrictions ever cuts off the optimal shutdown plan.

Editorial extensions

If this is right

Utilities could run near-optimal OPS decisions within a 30-minute operational window for hard days, instead of treating a day-ahead solve as the limiting step.
Average load shed on the 54 hard test instances is substantially lower than a same-timeout generic solver, because the primal gap at 30 minutes drops from 0.29 to 0.02.
The 30-minute ML-guided objective distribution is close to the 24-hour solver distribution, so offline training time can be traded for day-of speed.
The same domain bounds on the number of de-energized lines can be reused when the OPS formulation is extended to multiple periods, restoration, or security constraints.
Class imbalance (most lines energized) is handled explicitly by separate flip budgets; this is a transferable design for other switching problems with skewed binary solutions.

Reading between the lines

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

The paper leaves untested whether the 30-minute advantage persists on other networks or seasons; that depends on whether the predictor's ranking accuracy transfers beyond the one synthetic California grid.
A testable extension is to replace the fixed k0 = N_max_0 cutoff with a confidence-based cutoff using predicted probabilities; if the model is well calibrated, this could tighten the neighborhood without sacrificing the optimum.
Only 1-class recall (98.99%) is reported; computing 0-class precision and recall on the true de-energized lines would directly quantify the risk of excluding the optimal shutoff plan from the restricted search.
Because OPS instances recur daily, the approach suggests a closed-loop operational loop where each day's high-quality solution is added to the training pool, potentially improving future predictions—though the paper does not test this.

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

First thing to know: this is a real, practical extension of ML-guided MILP solvers to the Optimal Power Shutoff problem, and the reported speedup is plausible, but there is a sign error in the written objective that undermines the theoretical bound the whole PaS constraint rests on. Fix that and it's a solid paper; as written, a careful reader can't trust the domain-informed neighborhood without clarification.

What's new: the paper splits the PaS flip budget into Delta0 and Delta1, using problem-specific N_min0 and N_max0 from the risk constraint, and combines PaS with Neural Diving to create a safety-net variant. That is a genuine improvement over the existing PaS/ND frameworks, and the evaluation is honest: held-out hard days, equal 30-minute wall-clock limits, and the headline numbers (PI -30.94%, PG 0.29 to 0.02, 42/54 wins) are internally consistent. They also report that plain PaS underperforms Gurobi and that 24-hour Gurobi reaches better objectives on average—good honesty.

Soft spots:

1. Objective (1a) is written as min sum p_ls + eps * sum z_l, with z_l=1 meaning energized. That penalizes energized lines, i.e., rewards de-energization. The text says it penalizes de-energized lines. The N_max0 bound and the k0=N_max0 choice are only valid under the intended sign. If the implementation used the written sign, the bound may exclude the optimal and the reported objective values are measuring the wrong thing. This needs to be fixed and stated explicitly.

2. No code, data, trained models, or split specification. The central claim is not independently checkable.

3. No variance or significance testing. 'Significantly outperforms' is not backed by confidence intervals or replicates.

4. No ablation separating the learned policy from the search-space reduction. Fixing 70% of lines to 1 (k1=0.7 N) is a big reduction; the failure of PaS-only suggests the headline gain comes from the fixing mechanism, and a trivial baseline (e.g., fixing by risk rank or random) is missing.

5. The 98.99% recall is reported without saying on which split and without 0-class precision, which is the class that matters for the I0 constraint.

None of these is fatal by itself, and the central argument holds up if the sign is corrected. The paper deserves a serious referee, but the authors need to address the objective sign and release at least the data and a simple baseline. For a reader working on OPS or ML-guided MILP, this is worth reading now and citing with caution.

Referee Report

3 major / 4 minor

Summary. The paper proposes an ML-guided MILP framework for the Optimal Power Shutoff (OPS) problem, which selects transmission lines to de-energize in order to reduce wildfire ignition risk while limiting load shed. The authors extend Predict-and-Search (PaS) and Neural Diving (ND) by introducing separate neighborhood parameters for predicted-0 and predicted-1 variables, informed by domain knowledge on the minimum and maximum number of de-energized lines. They evaluate on a realistic synthetic California test system (CATS) with daily wildfire risk data, training on Gurobi 24-hour incumbents and testing on 54 hard held-out instances against a 30-minute Gurobi limit. The headline result is that the proposed PaS+ND hybrid reduces the average primal integral by 30.94% and the average primal gap from 0.29 to 0.02 compared to Gurobi at the same 30-minute cutoff, with qualitative case-study illustrations of switching decisions.

Significance. If the results hold, this is a practical and novel application of ML-guided MILP solving to power-system wildfire mitigation, with clear operational relevance. The paper's main strengths are the use of a realistic, geographically accurate test system; an externally benchmarked evaluation against Gurobi; and a sensible domain-specific adaptation of existing ML-guided methods. The reported improvements are large and internally consistent with the summary statistics in Table I. However, the manuscript contains a load-bearing formulation inconsistency in the OPS objective, and the safety of the domain-informed constraints is not supported by the reported ML metrics. These issues must be resolved before the central claims can be accepted.

major comments (3)

[II.A, Model 1, Eq. (1a)] The objective in (1a) is min Σ p_ls + ε Σ_{ℓ∈L_switch} z_ℓ. As stated after (1d), z_ℓ=0 means de-energized and z_ℓ=1 means energized. Thus ε Σ z_ℓ rewards de-energization, which is the opposite of the text's claim that the penalty 'minimizes the number of lines de-energized.' This is not a cosmetic issue: the derivation of N_max0 in Section II.A as an upper bound on the number of de-energized lines in an optimal solution relies on an objective that discourages de-energization. Under the written objective, an optimal solution may de-energize extra lines to collect the ε reward, so the statement in Section III.C.1 that setting k0=N_max0 'does not exclude the optimal solution' is not justified. Please correct the sign in (1a) or the prose, and state explicitly which objective was implemented in Gurobi; all reported objective values, primal gaps, and primal integrals inherit this choice.
[III.C.1 and V.B] The claim that k0=N_max0 'does not exclude the optimal solution' is only valid if the true zero set is sufficiently aligned with I0; it is not a mathematical consequence of the risk bounds alone. With Δ0=N_max0−N_min0, constraint (5a) forces at least N_min0 zeros among I0, so the optimal remains feasible iff at least N_min0 of the truly de-energized lines are in the top-N_max0 predicted-0 set. The paper reports only the class-1 recall (98.99%) and never reports precision/recall for the 0 class, nor verifies constraint satisfaction on the validation set. Since PaS+ND sets Δ1 to its lower bound (often 0), false positives among I1 are also unexamined. Please add 0-class metrics or an explicit verification that the restricted search space (5a)–(5b) contains the best known solutions on validation/test instances.
[V.C, Table I, Fig. 3] The main claim that PaS+ND 'significantly outperforms' Gurobi is based on a single test set of 54 instances, with no variance, confidence intervals, or significance test. The improvements in PI and PG could be driven by a few hard outliers, as the scatter plot in Figure 2 suggests. Report per-instance variability, paired significance tests (e.g., Wilcoxon signed-rank), and ideally multiple training/validation splits before claiming statistical significance.

minor comments (4)

[Throughout] The method is referred to inconsistently as 'PAS' and 'PaS' (also 'PAS+ND' vs 'PaS+ND'). Please standardize the notation.
[IV] The easy/hard classification threshold of 100 MWh of load shed is described as based on 'prior experience'; please justify this choice or report sensitivity to the threshold.
[V.B] Hyperparameters φ and φ' are selected on the validation set; report the selected values and ideally sensitivity of test performance to these parameters. No code or data availability statement is provided, which limits reproducibility.
[Fig. 3] The average primal gap as a function of time is shown without confidence bands or quantiles; adding these would strengthen the claim that PaS+ND is consistently faster, not just on average.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; evaluation is externally benchmarked against Gurobi on held-out instances.

full rationale

The derivation chain is not circular. The ML policy is trained on Gurobi 24-hour incumbents from training instances and evaluated on held-out test instances against a fresh 30-minute Gurobi run, so the reported speed and primal-integral improvement is an external empirical comparison rather than a quantity forced by construction. The domain-informed bounds N_min0/N_max0 come from constraint (1j), not from the ML outputs, and the PaS/ND framework is an extension of published methods; self-citations to [8] and [18] provide the OPS model and base PaS framework but are not invoked as an unverified uniqueness theorem. Two caveats are correctness risks rather than circularity: (i) the written objective (1a) penalizes Sigma z_l (energized lines) while the text claims a penalty minimizing de-energizations, which invalidates the stated interpretation of the N_min0/N_max0 bounds; (ii) the claim that k0=N_max0 'does not exclude the optimal solution' requires an unreported 0-class precision property of the ML ranking. Neither caveat makes the speed comparison equivalent to the training input or to a self-citation.

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

The central contribution is a speed-quality trade-off for an established MILP, so the ledger is dominated by validation-tuned hyperparameters (phi, phi') that control how aggressively the search is restricted, plus modeling choices inherited from [8]. The critical unmeasured premise on which the optimality-preservation argument rests is the precision of the 0-class predictions (see weakest_assumption). No new physical entities are introduced.

free parameters (5)

epsilon (de-energization penalty in objective 1a) = 0.01 per unit
Hand-selected penalty from prior work [8]; sets the trade-off between load shed and number of lines de-energized. Inherited, not fitted here.
phi (fraction of switchable lines assigned to 1 in I1) = 0.9 (PaS), 0.7 (PaS+ND)
Selected on the validation set for lowest primal integral over a grid {0.5,0.6,0.7,0.8,0.9}. Directly controls how many binary variables are fixed, hence the size of the reduced MILP and the method's speed.
phi' (offset fraction for Delta1) = 0.05 (PaS); 0 for PaS+ND (Delta1 at its lower bound)
Validation-tuned over {0.05,...,0.35}; inflates the flip budget for variables assigned to 1.
easy/hard instance classification threshold = 100 MWh of thresholded-problem load shed
Ad hoc split 'based on our prior experience in [8]' (Section IV). Determines which 540 instances enter the training/test pool and is never validated as a difficulty measure.
R_PSPS (system-wide acceptable risk) = set to match threshold-method residual risk
Comparison-protocol choice (Section II.B). The entire OPS risk budget per instance is fixed by the threshold baseline, which shapes N_min_0/N_max_0 and hence the domain-informed constraints.

assumptions (5)

domain assumption The B-theta DC power flow approximation with big-M linearization adequately represents line flows and feasibility for OPS decisions.
Constraints (1f)-(1h) in Model 1; carried from [8]/[12]. All conclusions about load shed and switching quality depend on this approximation being adequate for wildfire-risk shutoff decisions.
domain assumption The optimal number of de-energized lines lies between N_min_0 and N_max_0 computed from the risk constraint (1j) alone.
Section II.A. True for the risk constraint, but the inference-time use of these bounds (Section III.C.1) additionally requires that the ML predictions align with the optimal solution (see weakest_assumption).
domain assumption The ten best Gurobi 24-hour incumbent solutions saved per instance are representative of true optimal solutions.
Section IV: training labels are 'the ten incumbent solutions with the best objective values encountered' within a 24-hour limit. On hard instances even 24 hours may leave gaps (e.g., 3.23% MIP gap on March 2, 2021), so the learned policy inherits any systematic suboptimality of these labels.
standard math Conditional independence of line-status predictions pi(z_l | omega, M).
Section III.A.2: 'Following prior work, we assume conditional independence between variables to make the training task tractable.' Ignores correlations between de-energization decisions that the OPS power-flow and risk constraints enforce.
ad hoc to paper The single hour with the most thresholded load shed is a representative daily instance for evaluating operational OPS decisions.
Section IV: 'we consider the hour with the most load shed from the thresholded version of the problem run on each hour of the day.' Restricts the study to one period per day and drops multi-period interactions.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Machine Learning Guided Optimal Transmission Switching to Mitigate Wildfire Ignition Risk." pith.science (2026). https://pith.science/paper/LK6Z325H

@misc{pith2026251025147,
  author       = {Pith},
  title        = {Pith review of: Machine Learning Guided Optimal Transmission Switching to Mitigate Wildfire Ignition Risk},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LK6Z325H}},
  note         = {Machine review of arXiv:2510.25147}
}

read the original abstract

To mitigate acute wildfire ignition risks, utilities de-energize power lines in high-risk areas. The Optimal Power Shutoff (OPS) problem optimizes line energization statuses to manage wildfire ignition risks through de-energizations while reducing load shedding. OPS problems are computationally challenging Mixed-Integer Linear Programs (MILPs) that must be solved rapidly and frequently in operational settings. For a particular power system, OPS instances share a common structure with varying parameters related to wildfire risks, loads, and renewable generation. This motivates the use of Machine Learning (ML) for solving OPS problems by exploiting shared patterns across instances. In this paper, we develop an ML-guided framework that quickly produces high-quality de-energization decisions by extending existing ML-guided MILP solution methods while integrating domain knowledge on the number of energized and de-energized lines. Results on a large-scale realistic California-based synthetic test system show that the proposed ML-guided method produces high-quality solutions faster than traditional optimization methods.

Figures

Figures reproduced from arXiv: 2510.25147 by the authors.

**Figure 2.** Scatter plot of primal integral with Gurobi vs primal [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Average primal gap as a function of time (seconds). [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Geographic plots showing the line statuses and load shed from different solution methodologies for March 2 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Box plots showing the distribution of objective out [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 4 linked inside Pith

[1]

The potential for fuel reduction to reduce wildfire intensity in a warming california,

P. T. Brown, S. J. Strenfel, R. B. Bagley, and C. B. Clements, “The potential for fuel reduction to reduce wildfire intensity in a warming california,”Environmental Research Letters, vol. 20, no. 2, 2025

2025
[2]

Landscape op- timization for prescribed burns in wildfire mitigation planning,

W. Chen, E. P. Sivaramakrishnan, and B. Dilkina, “Landscape op- timization for prescribed burns in wildfire mitigation planning,” in Proceedings of the 5th ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies, 2022, pp. 429–438

2022
[3]

Top 20 destructive california wildfires,

California Department of Forestry and Fire Protection (CalFire), “Top 20 destructive california wildfires,” July 2025. [Online]. Available: https://34c031f8-c9fd-4018-8c5a-4159cdff6b0d-cdn-endpoint. azureedge.net/-/media/calfire-website/our-impact/fire-statistics/top20 destruction 072525.pdf

2025
[4]

Location, timing and extent of wildfire vary by cause of ignition,

A. D. Syphard and J. E. Keeley, “Location, timing and extent of wildfire vary by cause of ignition,”International Journal of Wildland Fire, vol. 24, no. 1, pp. 37–47, 2015

2015
[5]

Support for public safety power shutoffs in California: Wildfire-related perceived exposure and negative outcomes, prior and current health, risk appraisal and worry,

G. Wong-Parodi, “Support for public safety power shutoffs in California: Wildfire-related perceived exposure and negative outcomes, prior and current health, risk appraisal and worry,”Energy Research & Social Science, vol. 88, p. 102495, 2022

2022
[6]

Wildfire mitigation plans in power systems: A literature review,

D. A. Z. Vazquez, F. Qiu, N. Fan, and K. Sharp, “Wildfire mitigation plans in power systems: A literature review,”IEEE Transactions on Power Systems, vol. 37, no. 5, pp. 3540–3551, 2022

2022
[7]

Balancing wildfire risk and power outages through optimized power shut-offs,

N. Rhodes, L. Ntaimo, and L. Roald, “Balancing wildfire risk and power outages through optimized power shut-offs,”IEEE Transactions on Power Systems, vol. 36, no. 4, 2021

2021
[8]

Quantifying metrics for wildfire ignition risk from geographic data in power shutoff decision-making,

R. Piansky, S. Taylor, N. Rhodes, D. K. Molzahn, L. A. Roald, and J.-P. Watson, “Quantifying metrics for wildfire ignition risk from geographic data in power shutoff decision-making,”58th Hawaii International Conference on System Sciences (HICSS), January 2025

2025

Show all 41 references

[9]

PG&E wildfire mitigation plan R1 (2026–2028), volume 1: Base WMP,

Pacific Gas and Electric Company, “PG&E wildfire mitigation plan R1 (2026–2028), volume 1: Base WMP,” Pacific Gas and Electric Company, Tech. Rep. V ol. 1, February 2026. [Online]. Available: https://www.pge.com/assets/pge/docs/outages-and-safety/ outage-preparedness-and-suppo...

2026
[10]

Distributional MIPLIB: A multi-domain library for advancing ML-guided MILP meth- ods,

W. Huang, T. Huang, A. M. Ferber, and B. Dilkina, “Distributional MIPLIB: A multi-domain library for advancing ML-guided MILP meth- ods,”arXiv:2406.06954, 2024

2024 arXiv
[11]

Machine learn- ing augmented branch and bound for mixed integer linear programming,

L. Scavuzzo, K. Aardal, A. Lodi, and N. Yorke-Smith, “Machine learn- ing augmented branch and bound for mixed integer linear programming,” Mathematical Programming, pp. 1–44, 2024

2024
[12]

Long solution times or low solution quality: On trade-offs in choosing a power flow formulation for the optimal power shutoff problem,

E. Haag, N. Rhodes, and L. Roald, “Long solution times or low solution quality: On trade-offs in choosing a power flow formulation for the optimal power shutoff problem,”Electric Power Systems Research, vol. 234, p. 110713, 2024,23rd Power Systems Computation Conference (PSCC)

2024
[13]

United States Geological Sur- vey fire science: Fire danger monitoring and forecasting,

J. C. Eidenshink and S. M. Howard, “United States Geological Sur- vey fire science: Fire danger monitoring and forecasting,”US Ge- ological Survey, 2012, https://www.usgs.gov/publications/united-states- geological-survey-fire-science-fire-danger-monitoring-and-forecasting

2012
[14]

Probabilistic electric load forecasting: A tutorial review,

T. Hong and S. Fan, “Probabilistic electric load forecasting: A tutorial review,”International Journal of Forecasting, vol. 32, no. 3, pp. 914– 938, 2016

2016
[15]

Spatio-temporal wildfire prediction using multi-modal data,

C. Xu, Y . Xie, D. A. Z. Vazquez, R. Yao, and F. Qiu, “Spatio-temporal wildfire prediction using multi-modal data,”IEEE Journal on Selected Areas in Information Theory, vol. 4, pp. 302–313, 2023

2023
[16]

Solv- ing mixed integer programs using neural networks,

V . Nair, S. Bartunov, F. Gimeno, I. V on Glehn, P. Lichocki, I. Lobov, B. O’Donoghue, N. Sonnerat, C. Tjandraatmadja, P. Wanget al., “Solv- ing mixed integer programs using neural networks,”arXiv:2012.13349, 2020

2012 arXiv
[17]

A GNN-guided predict-and-search framework for mixed- integer linear programming,

Q. Han, L. Yang, Q. Chen, X. Zhou, D. Zhang, A. Wang, R. Sun, and X. Luo, “A GNN-guided predict-and-search framework for mixed- integer linear programming,” in11th International Conference on Learn- ing Representations (ICLR), 2023

2023
[18]

Contrastive predict-and-search for mixed integer linear programs,

T. Huang, A. M. Ferber, A. Zharmagambetov, Y . Tian, and B. Dilkina, “Contrastive predict-and-search for mixed integer linear programs,” in 41st International Conference on Machine Learning (ICML), July 2024

2024
[19]

Searching large neighborhoods for integer linear programs with contrastive learn- ing,

T. Huang, A. M. Ferber, Y . Tian, B. Dilkina, and B. Steiner, “Searching large neighborhoods for integer linear programs with contrastive learn- ing,” in40th International Conference on Machine Learning (ICML), 2023, pp. 13 869–13 890

2023
[20]

Optimizing transmission infrastructure investments to support line de-energization for mitigating wildfire ignition risk,

A. Kody, R. Piansky, and D. K. Molzahn, “Optimizing transmission infrastructure investments to support line de-energization for mitigating wildfire ignition risk,” in11th IREP Symposium on Bulk Power System Dynamics and Control, July 2022

2022
[21]

Long duration battery sizing, siting, and operation under wildfire risk using progressive hedging,

R. Piansky, G. Stinchfield, A. Kody, D. K. Molzahn, and J. P. Watson, “Long duration battery sizing, siting, and operation under wildfire risk using progressive hedging,”Electric Power Systems Research, vol. 235, Oct 2024,23rd Power Systems Computation Conference (PSCC)

2024
[22]

Optimizing battery and line undergrounding investments for transmission systems under wildfire risk scenarios: A Benders decomposition approach,

R. Piansky, R. K. Gupta, and D. K. Molzahn, “Optimizing battery and line undergrounding investments for transmission systems under wildfire risk scenarios: A Benders decomposition approach,”Sustainable Energy, Grids and Networks, vol. 43, no. 101838, September 2025,12th IREP B...

2025
[23]

A framework for risk assessment and optimal line upgrade selection to mitigate wildfire risk,

S. Taylor and L. A. Roald, “A framework for risk assessment and optimal line upgrade selection to mitigate wildfire risk,”Electric Power Systems Research, vol. 213, 2022,22nd Power Systems Computation Conference (PSCC)

2022
[24]

Managing wildfire risk and promoting equity through optimal configuration of networked microgrids,

S. Taylor, G. Setyawan, B. Cui, A. Zamzam, and L. A. Roald, “Managing wildfire risk and promoting equity through optimal configuration of networked microgrids,” inACM e-Energy, June 2023, pp. 189–199

2023
[25]

Equitably allocating wildfire resilience investments for power grids—The curse of aggregation and vulnerability indices,

M. Pollack, R. Piansky, S. Gupta, and D. K. Molzahn, “Equitably allocating wildfire resilience investments for power grids—The curse of aggregation and vulnerability indices,”Applied Energy, vol. 388, p. 125511, 2025

2025
[26]

Predicting electricity infrastructure induced wildfire risk in California,

M. Yao, M. Bharadwaj, Z. Zhang, B. Jin, and D. S. Callaway, “Predicting electricity infrastructure induced wildfire risk in California,”Environ- mental Research Letters, vol. 17, no. 9, p. 094035, 9 2022

2022
[27]

Data- driven power system optimal decision making strategy under wildfire events,

W. Hong, B. Wang, M. Yao, D. Callaway, L. Dale, and C. Huang, “Data- driven power system optimal decision making strategy under wildfire events,” in55th Hawaii International Conference on System Sciences (HICSS), January 2022

2022
[28]

California test system (CATS): A geographically accurate test system based on the California grid,

S. Taylor, A. Rangarajan, N. Rhodes, J. Snodgrass, B. Lesieutre, and L. A. Roald, “California test system (CATS): A geographically accurate test system based on the California grid,”IEEE Transactions on Energy Markets, Policy and Regulation, vol. 2, no. 1, pp. 107–118, 2024

2024
[29]

How attentive are graph attention networks?

S. Brody, U. Alon, and E. Yahav, “How attentive are graph attention networks?”arXiv:2105.14491, 2021

2021 arXiv
[30]

Exact combinatorial optimization with graph convolutional neural networks,

M. Gasse, D. Ch ´etelat, N. Ferroni, L. Charlin, and A. Lodi, “Exact combinatorial optimization with graph convolutional neural networks,” Advances in neural information processing systems, vol. 32, 2019

2019
[31]

Wildland Fire Potential Index,

U.S. Geological Survey, “Wildland Fire Potential Index,”
[32]

Gurobi Optimizer Reference Manual,

Gurobi Optimization, LLC, “Gurobi Optimizer Reference Manual,”
[33]

Measuring the impact of primal heuristics,

T. Berthold, “Measuring the impact of primal heuristics,”Operations Research Letters, vol. 41, no. 6, pp. 611–614, 2013

2013
[34]

Adam: A method for stochastic optimization,

D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv:1412.6980, 2014

2014 arXiv
[35]

Multi- period power system risk minimization under wildfire disruptions,

H. Yang, N. Rhodes, H. Yang, L. Roald, and L. Ntaimo, “Multi- period power system risk minimization under wildfire disruptions,”IEEE Transactions on Power Systems, vol. 39, no. 5, pp. 6305–6318, 2024

2024
[36]

Co-optimization of power line shutoff and restoration under high wildfire ignition risk,

N. Rhodes and L. A. Roald, “Co-optimization of power line shutoff and restoration under high wildfire ignition risk,” inIEEE Belgrade PowerTech, 2023

2023
[37]

Security constrained optimal power shutoff for wildfire risk mitigation,

N. Rhodes, C. Coffrin, and L. Roald, “Security constrained optimal power shutoff for wildfire risk mitigation,”IET Generation, Transmis- sion & Distribution, vol. 18, no. 18, pp. 2972–2986, 2024

2024
[38]

A general large neighborhood search framework for solving integer linear programs,

J. Song, R. Lanka, Y . Yue, and B. Dilkina, “A general large neighborhood search framework for solving integer linear programs,”33rd Advances in Neural Information Processing Systems (NeurIPS), pp. 20 012–20 023, 2020

2020
[39]

Evaluat- ing undergrounding decisions for wildfire ignition risk mitigation across multiple hazards,

R. Piansky, D. K. Molzahn, N. D. Jackson, and J. K. Skolfield, “Evaluat- ing undergrounding decisions for wildfire ignition risk mitigation across multiple hazards,”IEEE Kiel PowerTech, June 2025

2025
[2024]

Available: https://www.usgs.gov/fire-danger-forecast/ wildland-fire-potential-index-wfpi

[Online]. Available: https://www.usgs.gov/fire-danger-forecast/ wildland-fire-potential-index-wfpi
[2025]

Available: www.gurobi.com

[Online]. Available: www.gurobi.com

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

The potential for fuel reduction to reduce wildfire intensity in a warming california,

P. T. Brown, S. J. Strenfel, R. B. Bagley, and C. B. Clements, “The potential for fuel reduction to reduce wildfire intensity in a warming california,”Environmental Research Letters, vol. 20, no. 2, 2025

2025

[2] [2]

Landscape op- timization for prescribed burns in wildfire mitigation planning,

W. Chen, E. P. Sivaramakrishnan, and B. Dilkina, “Landscape op- timization for prescribed burns in wildfire mitigation planning,” in Proceedings of the 5th ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies, 2022, pp. 429–438

2022

[3] [3]

Top 20 destructive california wildfires,

California Department of Forestry and Fire Protection (CalFire), “Top 20 destructive california wildfires,” July 2025. [Online]. Available: https://34c031f8-c9fd-4018-8c5a-4159cdff6b0d-cdn-endpoint. azureedge.net/-/media/calfire-website/our-impact/fire-statistics/top20 destruction 072525.pdf

2025

[4] [4]

Location, timing and extent of wildfire vary by cause of ignition,

A. D. Syphard and J. E. Keeley, “Location, timing and extent of wildfire vary by cause of ignition,”International Journal of Wildland Fire, vol. 24, no. 1, pp. 37–47, 2015

2015

[5] [5]

Support for public safety power shutoffs in California: Wildfire-related perceived exposure and negative outcomes, prior and current health, risk appraisal and worry,

G. Wong-Parodi, “Support for public safety power shutoffs in California: Wildfire-related perceived exposure and negative outcomes, prior and current health, risk appraisal and worry,”Energy Research & Social Science, vol. 88, p. 102495, 2022

2022

[6] [6]

Wildfire mitigation plans in power systems: A literature review,

D. A. Z. Vazquez, F. Qiu, N. Fan, and K. Sharp, “Wildfire mitigation plans in power systems: A literature review,”IEEE Transactions on Power Systems, vol. 37, no. 5, pp. 3540–3551, 2022

2022

[7] [7]

Balancing wildfire risk and power outages through optimized power shut-offs,

N. Rhodes, L. Ntaimo, and L. Roald, “Balancing wildfire risk and power outages through optimized power shut-offs,”IEEE Transactions on Power Systems, vol. 36, no. 4, 2021

2021

[8] [8]

Quantifying metrics for wildfire ignition risk from geographic data in power shutoff decision-making,

R. Piansky, S. Taylor, N. Rhodes, D. K. Molzahn, L. A. Roald, and J.-P. Watson, “Quantifying metrics for wildfire ignition risk from geographic data in power shutoff decision-making,”58th Hawaii International Conference on System Sciences (HICSS), January 2025

2025