Fine-Tuning LLMs to Generate Economical and Reliable Actions for the Power Grid

Chehade, Mohamad; Zhu, Hao

doi:10.1109/PESGM58988.2026.11693462

Fine-Tuning LLMs to Generate Economical and Reliable Actions for the Power Grid

M. Chehade, H. Zhu

2026 IEEE Power & Energy Society General Meeting (PESGM), pp. 1–5, Jul. 2026

A fine-tuning framework for language models that generate corrective power-grid actions while balancing cost and reliability.

IEEE Xplore arXiv BibTeX

Figure. Each stage is supervised by a stronger signal: SFT imitates a DC-OPF MILP oracle, DPO learns from AC power-flow evaluations ranked by voltage penalty, and best-of- keeps the best feasible plan at inference time. On IEEE 118-bus PSPS scenarios, AC power-flow failures drop from 50% (zero-shot) to single digits.

What this paper adds

Connects LLM fine-tuning with structured grid-action generation.
Evaluates action quality through economic and reliability-oriented criteria.
Uses power systems as a constrained decision-making testbed for LLMs.

Cite

@inproceedings{chehade2026finetuning,
  title     = {Fine-Tuning {LLMs} to Generate Economical and Reliable Actions for the Power Grid},
  author    = {Chehade, Mohamad and Zhu, Hao},
  booktitle = {2026 IEEE Power \& Energy Society General Meeting (PESGM)},
  pages     = {1--5},
  publisher = {IEEE},
  year      = {2026},
  doi       = {10.1109/PESGM58988.2026.11693462}
}