Paper deep dive
Fine-Tuning LLMs to Generate Economical and Reliable Actions for the Power Grid
Mohamad Chehade, Hao Zhu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 7/21/2026, 3:56:09 AM
Summary
This paper presents a multi-stage fine-tuning pipeline for large language models (LLMs) to generate corrective transmission switching actions during Public Safety Power Shutoffs (PSPS). The approach uses Supervised Fine-Tuning (SFT) to imitate a DC-OPF MILP oracle, followed by Direct Preference Optimization (DPO) to improve voltage quality using AC power flow evaluations. The method significantly reduces AC power-flow failures and improves economic objectives compared to zero-shot generation on IEEE 118-bus test systems.
Entities (8)
Relation Signals (5)
Fine-Tuned LLM → reduces → AC power-flow failure
confidence 95% · reduces AC power-flow failure from 50% to single digits
IEEE 118-bus → usedin → Experimental Evaluation
confidence 95% · On IEEE 118-bus PSPS scenarios, fine-tuning substantially improves DC objective values
Supervised Fine-tuning → uses → DC-OPF MILP
confidence 92% · supervised fine-tuning distills a DC-OPF MILP oracle into a constrained action grammar
Public Safety Power Shutoffs → causes → topology changes
confidence 90% · PSPS force rapid topology changes that can render standard operating points infeasible
Direct Preference Optimization → improves → voltage quality
confidence 90% · direct preference optimization refines the policy... injecting voltage-awareness beyond DC imitation
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Public Safety Power Shutoffs (PSPS) force rapid topology changes that can render standard operating points infeasible, requiring operators to quickly identify corrective transmission switching actions that reduce load shedding while maintaining acceptable voltage behavior. We present a verifiable, multi-stage adaptation pipeline that fine-tunes an instruction-tuned large language model (LLM) to generate \emph{open-only} corrective switching plans from compact PSPS scenario summaries under an explicit switching budget. First, supervised fine-tuning distills a DC-OPF MILP oracle into a constrained action grammar that enables reliable parsing and feasibility checks. Second, direct preference optimization refines the policy using AC-evaluated preference pairs ranked by a voltage-penalty metric, injecting voltage-awareness beyond DC imitation. Finally, best-of-$N$ selection provides an inference-time addition by choosing the best feasible candidate under the target metric. On IEEE 118-bus PSPS scenarios, fine-tuning substantially improves DC objective values versus zero-shot generation, reduces AC power-flow failure from 50\% to single digits, and improves voltage-penalty outcomes on the common-success set. Code and data-generation scripts are released to support reproducibility.
Tags
Links
- Source: https://arxiv.org/abs/2602.15350v1
- Canonical: https://arxiv.org/abs/2602.15350v1
Trouble viewing inline? Open PDF directly →
Full Text
29,332 characters extracted from source content.
Expand or collapse full text
Fine-Tuning LLMs to Generate Economical and Reliable Actions for the Power Grid †thanks: This work was funded by NSF Grant 2130706 and ARO Grant W911NF2310266. Mohamad Chehade and Hao Zhu Abstract Public Safety Power Shutoffs (PSPS) force rapid topology changes that can render standard operating points infeasible, requiring operators to quickly identify corrective transmission switching actions that reduce load shedding while maintaining acceptable voltage behavior. We present a verifiable, multi-stage adaptation pipeline that fine-tunes an instruction-tuned large language model (LLM) to generate open-only corrective switching plans from compact PSPS scenario summaries under an explicit switching budget. First, supervised fine-tuning distills a DC-OPF MILP oracle into a constrained action grammar that enables reliable parsing and feasibility checks. Second, direct preference optimization refines the policy using AC-evaluated preference pairs ranked by a voltage-penalty metric, injecting voltage-awareness beyond DC imitation. Finally, best-of-N selection provides an inference-time addition by choosing the best feasible candidate under the target metric. On IEEE 118-bus PSPS scenarios, fine-tuning substantially improves DC objective values versus zero-shot generation, reduces AC power-flow failure from 50% to single digits, and improves voltage-penalty outcomes on the common-success set. Code and data-generation scripts are released to support reproducibility. I Introduction Large language models (LLMs) have rapidly transitioned from research prototypes to deployable decision-support tools across diverse domains [11, 20, 8]. Their ability to transform unstructured descriptions into structured outputs makes them attractive for operational environments where decisions are time-sensitive and consequences are high. Power system control rooms are a compelling setting: operators manage complex contingencies, coordinate actions across many assets, and balance reliability, economics, and compliance under tight time constraints [9]. Unlike traditional decision-support tools that require specialized inputs or rigid interfaces, LLMs enable operators to interact through natural language while producing machine-readable recommendations (e.g., structured action lists) that can be verified before execution [5, 18]. However, foundation LLMs lack domain-specific knowledge of power system physics, operational constraints, and grid safety requirements. Training grid-specific LLMs from scratch is impractical: modern LLMs succeed through pre-training on trillions of tokens spanning diverse domains [6], while grid operations data are orders of magnitude smaller and specialized. A practical alternative is to adapt a strong instruction-tuned model via targeted fine-tuning so that it can (i) read a compact, structured description of a grid scenario and (i) output actions in a constrained grammar that can be checked for feasibility. In this work, we study a concrete and operationally motivated task: corrective, open-only transmission switching during Public Safety Power Shutoffs (PSPS), which are corrective de-energization actions used by utilities to reduce wildfire ignition risk during extreme weather conditions [14]. When PSPS forces lines out of service, operators must rapidly determine whether opening additional elements can mitigate overloads, reduce load shedding, and improve operating conditions while respecting switching budgets and operational rules. Computing optimal actions with mixed-integer optimization can be expensive under time pressure, especially when considering nonlinear AC constraints [2, 4, 3]. Our goal is to amortize this optimization effort into training, then produce high-quality switching recommendations at inference time using structured scenario summaries and a verifiable action grammar. Figure 1 illustrates the pipeline we adopt. Starting from an instruction-tuned base model, supervised fine-tuning (SFT) trains the LLM to imitate MILP-derived open-only switching decisions under DC constraints. We then apply direct preference optimization (DPO) using ranked responses derived from AC voltage-quality evaluation, producing a voltage-aware policy that more reliably prioritizes actions with fewer voltage violations. BaseLLMSFTimitate MILPDPOvoltage-awarePolicyopen-only MILP demos (DC-OPF) AC eval → rank by VpenV_pen optional: best-of-N Figure 1: Multi-stage adaptation pipeline for PSPS corrective switching. This design follows a standard alignment pattern for instruction-tuned LLMs: imitation learning first, followed by preference-based refinement [13, 16]. In our setting, the supervised stage anchors the policy to an optimization oracle, while the preference stage injects AC voltage-awareness that is difficult to encode directly in DC training. The resulting model functions as a candidate-plan generator whose outputs can be parsed, verified, and evaluated with existing grid-analysis tools. Our contributions are: • We formulate PSPS-aware open-only switching with switching budgets and corridor structure using a DC-OPF MILP oracle (Section I). • We design a structured scenario representation and action grammar that enables an instruction-tuned LLM to emit switching plans that are straightforward to parse and verify (Section I). • We introduce a voltage-aware preference refinement stage based on DPO, using AC-derived voltage-quality preferences to align the model beyond DC imitation (Section I-A). • We evaluate economic performance, AC feasibility, and voltage quality, including comparisons to a neural baseline and training-curve reporting for reproducibility (Section IV). Finally, we discuss practical considerations such as feasibility checks, training/inference costs, and deployment constraints (Section IV). We view this as a step toward verifiable, operator-facing LLM assistants that interface with existing grid analysis pipelines rather than replacing them. I PSPS Switching Problem Public Safety Power Shutoffs (PSPS) are preventive and corrective de-energization actions taken by utilities to reduce wildfire ignition risk during extreme weather conditions [17, 7]. When a PSPS event forces a subset of transmission lines out of service, system operators must determine whether additional corrective open-only switching actions can improve reliability and reduce load shedding. We formulate a DC optimal power flow (DC-OPF) model that explicitly incorporates PSPS constraints and pose an open-only decision-making problem in which operators may proactively open a limited number of additional transmission elements to mitigate system stress. Network model and PSPS constraints. Consider a power system with bus set ℬB (|ℬ|=nb|B|=n_b), transmission line set ℰE (|ℰ|=nℓ|E|=n_ ), and generator set ⊆ℬG . Each line e has reactance xe>0x_e>0 and rating Semax>0S _e>0 (MW). Buses have demand Pid≥0P^d_i≥ 0, generators have capacity Pgmin≤Pg≤PgmaxP _g≤ P_g≤ P _g and cost cg≥0c_g≥ 0 ($/MW), and load shedding Pis≥0P^s_i≥ 0 is penalized at rate γ ($/MW). A PSPS event is encoded by availability mask ∈0,1nℓ ξ∈\0,1\^n_ , where ξe=0 _e=0 forces line e open. Operator decisions ze∈0,1z_e∈\0,1\ determine effective status ge=ξezeg_e= _ez_e. We use standard DC power flow with bus-branch incidence A and susceptance Bℓ=diag(1/xe)B_ =diag(1/x_e). Open-only switching with a line budget. Operators may open up to KℓK_ additional PSPS-available lines. The open-only switching problem is: ming,s,,ℓ,∈0,1nℓ _ subarrayc P_g, P^s, θ, P_ ,\\ z∈\0,1\^n_ subarray ∑g∈cgPg+γ∑i∈ℬPis _g c_gP_g\;+\;γ _i P^s_i (1a) s.t. Pgmin≤Pg≤Pgmax,∀g∈, P _g≤ P_g≤ P _g,\ ∀ g , (1b) Pis≥0,∀i∈ℬ, P^s_i≥ 0,\ ∀ i , (1c) ℓ=BℓA⊤,θr=0, P_ =B_ A θ,\ _r=0, (1d) Aℓ=g−(d−s), A P_ = P_g-( P^d- P^s), (1e) −Semaxξeze≤Pℓ,e≤Semaxξeze,∀e∈ℰ, -S _e\, _ez_e≤ P_ ,e≤ S _e\, _ez_e,\ ∀ e , (1f) ze∈0,1,ze=0ifξe=0,∀e∈ℰ, z_e∈\0,1\,\ z_e=0\ if\ _e=0,\ ∀ e , (1g) ∑e∈ℰ(1−ze)ξe≤Kℓ. _e (1-z_e)\, _e≤ K_ . (1h) Constraint (1g) enforces that PSPS-forced outages remain open; (1h) limits operator-induced opens to KℓK_ available lines. This is structurally related to optimal transmission switching [2, 4] but restricted to open-only actions. Corridor-constrained open-only switching. Switching decisions may be constrained to transmission corridors S—geographically grouped lines [12, 7]. Binary variables yS∈0,1y_S∈\0,1\ indicate whether corridor S is activated for switching. Corridor and line decisions are coupled by: ze z_e ≥1−yS,∀S∈,∀e∈S, ≥ 1-y_S, ∀ S ,\ ∀ e∈ S, (2a) ∑S∈yS _S y_S ≤KS,yS∈0,1,∀S∈, ≤ K_S, y_S∈\0,1\,\ ∀ S , (2b) When yS=0y_S=0, (2a) prevents operator opens in corridor S; when yS=1y_S=1, line decisions remain free. Constraint (2b) limits activated corridors to KSK_S. Role in the pipeline. We use the DC-OPF MILP above as an offline oracle to generate labeled switching plans for supervised adaptation. Voltage quality and AC feasibility are evaluated separately using AC power flow in the experimental section, enabling us to train on DC structure while assessing voltage-critical performance. I Supervised Fine-Tuning (SFT) LLMs Solving the mixed-integer linear program (MILP) in (1) for every PSPS scenario can be computationally expensive, particularly when evaluating large numbers of contingencies or when operators require rapid what-if analysis. Similar scalability challenges are well documented for optimal transmission switching (OTS) formulations, motivating learning-assisted and proxy-based approaches that approximate an optimizer while preserving most of the economic benefit [15, 1]. We therefore train a large language model (LLM) to imitate an optimization oracle, amortizing the cost of offline MILP solves into a single supervised fine-tuning (SFT) phase. After SFT, the model generates candidate switching plans from compact scenario summaries, which are then parsed and verified before evaluation or deployment. This section describes the oracle, the input–output representation, and the SFT protocol. Ground-Truth Oracle. Given PSPS mask ξ and budget KℓK_ , we solve the DC-OPF MILP (1) to obtain optimal operator-induced opens ()≜e∈ℰ:ξe=1,ze⋆=0.T( ξ)\; \; \\,e : _e=1,\;z _e=0\, \. (3) For corridor-constrained problems, we solve the variant with (2). Scenario Representation. Each training example corresponds to a PSPS scenario with mask ξ and budgets KℓK_ , KSK_S. We provide a structured textual summary including case dimensions, the number of PSPS-forced opens, and a corridor breakdown listing member lines with their forced/eligible status and applicable budgets. This abstracts away numerical parameters while preserving topological structure. Action Grammar. The target encodes ()T( ξ) using: open(Sk:LINE) for corridor-associated lines (e.g., open(S6:135)), open(LINE) for non-corridor lines, and do_nothing if ()=∅T( ξ)= . We sort actions lexicographically to form a canonical string, e.g., open(S3:21); open(S7:98); open(131). open(S3:21); open(S7:98); open(131). Dataset Formatting. Supervised pairs use a chat format: system prompt defining the task and grammar, user message with scenario JSON, and assistant message with the canonical action string, following standard instruction-tuning practice [13]. Fine-Tuning Objective and Protocol. We initialize from an instruction-tuned base LLM and perform supervised fine-tuning on the training partition to learn a policy πSFT _SFT that maps scenario summaries to action strings. The objective is the standard conditional language-modeling loss over the assistant tokens. Let sks_k denote the scenario summary and aka_k the canonical action string for example k. We optimize parameters ϕφ by minimizing ℒSFT(ϕ)=−∑klogpϕ(ak|sk,system prompt),L_SFT(φ)\;=\;- _k p_φ (a_k\, |\,s_k,\ system prompt ), (4) where the sum ranges over training examples. This mirrors the supervised stage used in instruction-following pipelines prior to preference optimization [13]. The resulting model πSFT _SFT is then used for inference and (optionally) subsequent preference-based refinement. Parsing and Verification. At inference, we parse outputs and verify: (i) grammar validity, (i) all opened lines are PSPS-available (ξe=1 _e=1), and (i) budget adherence. Invalid outputs are flagged, ensuring the LLM functions as a candidate generator rather than an unverified controller. I-A Improving Voltage Safety via DPO (SFT-Time) The supervised policy πSFT _SFT imitates a DC open-only oracle, but DC imitation alone does not explicitly optimize voltage quality under nonlinear AC physics. To better align the policy with voltage safety objectives, we introduce a refinement stage based on direct preference optimization (DPO) [16]. DPO is an reinforcement learning-free objective that trains a policy from pairwise preferences (y+,y−)(y^+,y^-), encouraging the refined policy to assign higher probability to actions that yield better voltage outcomes under AC evaluation. Voltage-Penalty Metric. For candidate plan y, we parse into topology (y) z(y) and solve AC power flow to obtain bus voltage magnitudes |Vi(y)|\|V_i(y)|\. We define a deadband vdbv_db around nominal voltage and penalize violations outside [1−vdb, 1+vdb][1-v_db,\,1+v_db]: Vpen(y)≜κ∑i∈ℬ(max||Vi(y)|−1|−vdb, 0)p,V_pen(y)\; \;κ _i ( \\, ||V_i(y)|-1 |-v_db,\;0\, \ )^p, (5) where p∈1,2p∈\1,2\ selects aggregation norm and κ>0κ>0 scales the penalty. Non-convergent AC solutions receive penalty VfailV_fail (large constant). Preference Pair Construction. For each scenario x, we sample NprefN_pref candidates from πSFT(⋅∣x) _SFT(· x), discard malformed or budget-violating plans, evaluate Vpen(y)V_pen(y) via AC power flow, and select (y+,y−)(y^+,y^-) pairs where Vpen(y−)−Vpen(y+)≥Δpref,V_pen(y^-)-V_pen(y^+)≥ _pref, (6) yielding preference dataset volt=(xi,yi+,yi−)i=1MD_volt=\(x_i,y_i^+,y_i^-)\_i=1^M where y+y^+ is preferred for voltage quality. DPO Objective. Let πϕ _φ denote the trainable policy initialized from πSFT _SFT, with reference πref=πSFT _ref= _SFT. For triple (x,y+,y−)(x,y^+,y^-), define Δϕ(x,y+,y−)≜logπϕ(y+∣x)−logπϕ(y−∣x), _φ(x,y^+,y^-)\; \; _φ(y^+ x)- _φ(y^- x), (7) and analogously Δref _ref. The DPO loss is ℒDPO(ϕ)=−∑(x,y+,y−)∈voltlogσ(βDPO[Δϕ−Δref]),L_DPO(φ)\;=\;-\!\! _(x,y^+,y^-) _volt σ\! ( _DPO [ _φ- _ref ] ), (8) where βDPO>0 _DPO>0 controls preference strength [16]. Minimizing this produces πDPO _DPO, which increases likelihood of low-penalty plans while anchored to the SFT reference. Scope and Practicality. Preference construction uses AC power flow offline to label candidates with voltage-quality information. At inference time, the refined model can be used either in single-shot mode or combined with best-of-N reranking (Section I-B), where VpenV_pen can serve as a selection criterion when AC evaluation is available. I-B Improving Voltage Safety via Best-of-N (Inference-Time) Even after supervised fine-tuning and preference alignment, a single stochastic decode can produce a suboptimal or malformed plan. We therefore use best-of-N inference as a lightweight, training-free mechanism to improve solution quality by sampling multiple candidate plans and selecting the best under a task metric. Best-of-N is widely used in reasoning and structured prediction tasks to exploit sampling diversity, often with strong gains at moderate N [19]. Given a scenario summary x and policy π (e.g., πSFT _SFT or πDPO _DPO), we draw N independent candidates: y(j)∼π(⋅∣x),j=1,…,N.y^(j) π(· x), j=1,…,N. (9) Each candidate is verified through staged checks: grammar parsing, PSPS/budget constraint validation, DC feasibility evaluation, and optional AC power flow for voltage scoring. Let valid(x)Y_valid(x) denote candidates passing verification. If empty, we fall back to a safe default (e.g., doing nothing). We then select the plan minimizing a scalar score: y^(x)∈argminy∈valid(x)Score(x,y). y(x)\;∈\; _y _valid(x)\;Score(x,y). (10) Primary choices are: (i) DC economic score JDC(x,y)J_DC(x,y), or (i) AC voltage penalty Vpen(y)V_pen(y) from (5). When both matter, we use scalarization Score(x,y)=JDC(x,y)+λVpen(y)Score(x,y)=J_DC(x,y)+λ\,V_pen(y) with λ≥0λ≥ 0 set by operational priorities. Best-of-N increases compute linearly in N. We use moderate N with early stopping when target scores are met. Sampling is embarrassingly parallel, and N can be adjusted to trade off compute for quality. We quantify costs and inference times in Section IV. IV Experimental Results Experiments use the IEEE 118-bus test system from MATPOWER [21] (nb=118n_b=118, nℓ=186n_ =186, ng=54n_g=54). We define ||=9|S|=9 transmission corridors by grouping geographically proximate lines (8–20 lines per corridor). PSPS events are simulated by selecting corridor lines to force open, producing availability mask ξ for each scenario. Unless stated otherwise, operators may open at most Kℓ=3K_ =3 additional PSPS-available lines. Datasets. For SFT, we use 200 PSPS scenarios split between training and held-out testing. For DPO, we construct 440 preference pairs (x,y+,y−)(x,y^+,y^-) by sampling candidates from πSFT _SFT and ranking by AC voltage penalty VpenV_pen (Section I-A). Data-generation scripts and splits are in the released codebase. Implementation and Models. Power flow and optimization use MATLAB R2024b with YALMIP [10]. We adapt an instruction-tuned LLM, ft:gpt-4.1-mini-2025-04-14, via the OpenAI fine-tuning API. SFT trains for 3 epochs (batch size 1, LR multiplier 2) on 453,384 training tokens. DPO initializes from πSFT _SFT and trains for 2 epochs (batch size 8) with βDPO=0.1 _DPO=0.1 on 1,611,736 tokens. For the evaluation of the voltage penalties, we use vdb=0v_db=0, p=1p=1, κ=1κ=1 for (5). Our pipeline is backend-agnostic and can be applied to any instruction-tuned LLM deployed either via API or locally hosted models. Compared Policies. We compare: (i) zero-shot base LLM, (i) πSFT _SFT, (i) πDPO _DPO, and (iv) a neural network baseline: a fully-connected MLP with one hidden layer of width 512 (ReLU), trained on the same supervised dataset to predict corrective opens under identical budget and feasibility constraints. Code Release. Full code is available on GitHub: MFHChehade/LLM-Grid-Actions. IV-A Training Curves for the Fine-Tuning Process (a) SFT log loss. (b) SFT token accuracy. (c) DPO loss. (d) DPO error rate. Figure 2: Training curves from the fine-tuning jobs (SFT and DPO) show the convergence. Figures 2(a–b) show that the SFT process is well-behaved: log loss drops sharply early then decreases gradually, while token accuracy rapidly increases and stabilizes. This pattern reflects the model learning the output grammar then refining scenario-to-action mappings. Figure 2(c–d) shows stable DPO optimization with decreasing loss and preference error rate, indicating successful separation of preferred vs. dispreferred plans. No divergence occurs under the chosen βDPO _DPO and dataset size. IV-B DC Objective Figure 3: Distribution of DC objective JDCJ_DC across all four compared policies (zero-shot, SFT, DPO, and N). Figure 3 reports the box plots for the JDCJ_DC distributions. Both SFT and DPO shift downward relative to zero-shot, distilling the oracle signal into lower-cost decisions. SFT and DPO medians are close, showing voltage-aware preference refinement preserves DC performance. The N baseline achieves low median JDCJ_DC but exhibits wider dispersion and heavier upper tail, indicating occasional high-cost decisions. IV-C AC Feasibility and Voltage Quality Figure 4: AC power-flow failure rate across compared policies. Fine-tuned policies drastically reduce AC failures relative to the zero-shot baseline. Figure 5: Voltage penalty VpenV_pen distribution on the common-success set, i.e., scenarios where all compared policies (SFT, DPO, N) have achieved AC convergence. This ensures an apples-to-apples comparison of voltage quality. Figure 4 shows zero-shot fails AC power flow in half of scenarios, while SFT and DPO reduce failures to a small fraction. This indicates DC oracle imitation plus constrained grammar substantially improves physical plausibility. The N baseline achieves the lowest failure rate. We evaluate VpenV_pen on the common-success set (Figure 5)—scenarios where SFT, DPO, and N all converge. DPO achieves lower median VpenV_pen than SFT and N, consistent with preference refinement improving voltage outcomes beyond DC imitation. However, the upper tail persists across policies, suggesting some scenarios remain challenging for voltage regulation and motivating richer preference data as future work. V Conclusion We developed a verifiable fine-tuning pipeline that adapts a foundation LLM into a corrective switching assistant for PSPS events. Supervised fine-tuning distills MILP-derived DC-optimal open-only actions into a constrained, parseable grammar that enables systematic feasibility checks. Direct preference optimization then injects AC voltage-awareness by learning from preference pairs ranked using a voltage-penalty metric, improving voltage quality among AC-feasible cases. Best-of-N selection complements training by trading inference compute for improved candidate quality. Empirically on IEEE 118-bus PSPS scenarios, the fine-tuned policies outperform zero-shot generation in economic objective, dramatically reduce AC infeasibility, and yield improved voltage-penalty distributions on the common-success set, while remaining compatible with standard power-flow verification. Future work will investigate multi-task fine-tuning of a single foundation model so it can support diverse power-grid applications under a unified, verifiable decision framework. References [1] A. B. Bugaje, J. L. Cremer, and G. Strbac (2023) Real-time transmission switching with neural networks. IET Generation, Transmission & Distribution 17 (3), p. 696–705. External Links: Document Cited by: §I. [2] E. B. Fisher, R. P. O’Neill, and M. C. Ferris (2008) Optimal transmission switching. IEEE Transactions on Power Systems 23 (3), p. 1346–1355. External Links: Document Cited by: §I, §I. [3] E. Haag, N. Rhodes, and L. Roald (2024) 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 234, p. 110713. Note: Also available as arXiv:2310.13843 External Links: Document Cited by: §I. [4] K. W. Hedman, R. P. O’Neill, E. B. Fisher, and S. S. Oren (2009) Optimal transmission switching with contingency analysis. IEEE Transactions on Power Systems 24 (3), p. 1577–1586. External Links: Document Cited by: §I, §I. [5] S. Kalami et al. (2025) Powering the grid with language: how LLMs are transforming energy systems. Note: Medium (online article)Accessed 2025 Cited by: §I. [6] J. Kaplan, S. McCandlish, T. Henighan, T. B. Brown, B. Chess, R. Child, S. Gray, A. Radford, J. Wu, and D. Amodei (2020) Scaling laws for neural language models. arXiv preprint. External Links: 2001.08361 Cited by: §I. [7] A. Kody, R. Piansky, and D. K. Molzahn (2022) Optimizing transmission infrastructure investments to support line de-energization for mitigating wildfire ignition risk. In Proceedings of the 11th Bulk Power Systems Dynamics and Control Symposium (IREP), Note: Also available as arXiv:2203.10176 External Links: 2203.10176 Cited by: §I, §I. [8] J. Lai, X. Li, Z. Wang, Y. Zhu, Y. Li, and S. Wang (2024) Large language models in law: a survey. AI Open 6, p. 181–196. External Links: Document Cited by: §I. [9] C.-J. Lin, M. M. Robertson, and J. D. Lee (2013) Analyzing the staffing and workload in the main control room. Safety Science 57, p. 161–168. External Links: Document Cited by: §I. [10] J. Löfberg (2004) YALMIP: a toolbox for modeling and optimization in MATLAB. In Proceedings of the IEEE International Symposium on Computer Aided Control System Design (CACSD), Taipei, Taiwan, p. 284–289. Cited by: §IV. [11] A. Mastropaolo, L. Pascarella, E. Guglielmi, M. Ciniselli, S. Scalabrino, R. Oliveto, and G. Bavota (2023) On the robustness of code generation techniques: an empirical study on GitHub Copilot. In IEEE/ACM International Conference on Software Engineering (ICSE), Note: Also available as arXiv:2302.00438 External Links: Document, 2302.00438 Cited by: §I. [12] A. Moreira, F. Pianco, B. Fanzeres, A. Street, R. Jiang, C. Zhao, and M. Heleno (2024) Distribution system operation amidst wildfire-prone climate conditions under decision-dependent line availability uncertainty. IEEE Transactions on Power Systems. Note: Also available as arXiv:2306.06336 External Links: Document, 2306.06336 Cited by: §I. [13] L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. L. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, et al. (2022) Training language models to follow instructions with human feedback. In Advances in Neural Information Processing Systems, Note: Also known as the InstructGPT paper Cited by: §I, §I, §I. [14] Pacific Gas and Electric Company (2024) Public safety power shutoffs (PSPS) fact sheet. Note: Website External Links: Link Cited by: §I. [15] S. Pineda, J. M. Morales, and A. Jiménez-Cordero (2024) Learning-assisted optimization for transmission switching. TOP 32, p. 489–516. External Links: Document Cited by: §I. [16] R. Rafailov, A. Sharma, E. Mitchell, S. Ermon, C. D. Manning, and C. Finn (2024) Direct preference optimization: your language model is secretly a reward model. In International Conference on Learning Representations (ICLR), External Links: 2305.18290 Cited by: §I, §I-A, §I-A. [17] N. Rhodes, L. Ntaimo, and L. Roald (2021) Balancing wildfire risk and power outages through optimized power shut-offs. IEEE Transactions on Power Systems 36 (4), p. 3118–3128. External Links: Document Cited by: §I. [18] M. Sanguinetti, L. Atzori, R. Girau, and M. Marras (2024) Conversational agents for energy awareness and efficiency: a survey. Electronics 13 (2), p. 401. External Links: Document Cited by: §I. [19] X. Wang, J. Wei, D. Schuurmans, Q. Le, E. H. Chi, S. Narang, A. Chowdhery, and D. Zhou (2023) Self-consistency improves chain-of-thought reasoning in language models. In International Conference on Learning Representations (ICLR), Note: Also available as arXiv:2203.11171 External Links: 2203.11171 Cited by: §I-B. [20] Y. Zhang, S. A. Khan, A. Mahmud, H. Yang, A. Lavin, M. Levin, J. Frey, J. Dunnmon, J. Evans, A. Bundy, S. Džeroski, J. Tegnér, and H. Zenil (2025) Exploring the role of large language models in the scientific method: from hypothesis to discovery. npj Artificial Intelligence 1, p. 14. External Links: Document Cited by: §I. [21] R. D. Zimmerman, C. E. Murillo-Sánchez, and R. J. Thomas (2011) MATPOWER: steady-state operations, planning, and analysis tools for power systems research and education. IEEE Transactions on Power Systems 26 (1), p. 12–19. External Links: Document, Link Cited by: §IV.