LLM-assisted semantic scenario editing and constraint-aware capped stable fixing for renewable-rich SCUC acceleration

Security-constrained unit commitment (SCUC) in renewable-rich power systems faces stronger source-load variability, steeper net-load ramps, and a greater mixed-integer computational burden. Routine SCUC typically uses numerical load and renewable forecasts, while scenario-based analysis and stress testing may also involve qualitative operating descriptions. This paper develops a framework that combines large language model (LLM)-assisted semantic scenario editing with controlled commitment fixing for renewable-rich SCUC. In the proposed framework, an LLM acts as a constrained semantic parser that translates such supplementary descriptions into structured source-load editing events, which are applied to numerical baseline profiles through deterministic editing and consistency validation. A trained commitment-probability model predicts full-horizon on-state probabilities for thermal-unit commitment variables. To convert these probabilities into acceleration decisions, Constraint-aware Capped Stable Fixing (CCSF) selects high-confidence entries through temporal-stability screening, initialization-aware release, and a fixed-ratio cap before imposing commitment-fixing constraints. CCSF adds only commitment-fixing constraints, leaving the base SCUC physical constraints unchanged. On a modified IEEE 118-bus renewable-rich SCUC benchmark, CCSF obtains feasible solutions without activated slack variables under all tested thresholds. Under relaxed thresholds, CCSF achieves a mean per-scenario speedup of up to 5.16 × , with a maximum absolute objective difference of 0.42%. These results indicate that constrained semantic editing supports the traceable conversion of language-level descriptions into SCUC-ready source-load inputs, while capped and constraint-aware fixing enables threshold-robust acceleration without modifying the base SCUC physical constraints.