From public weather narratives to solar-market risk decisions using constrained language-model features

Weather risk in high-solar power systems affects generation, prices, and imbalance exposure. Public forecasting pipelines use numerical weather prediction and operator forecasts, but forecast-discussion judgment remains difficult to structure. We evaluate a constrained language-model extraction design that converts NWS Area Forecast Discussions into bounded weather-risk features. The CAISO-SP15 public-data test bed combines system-level solar generation, day-ahead and real-time price proxies, HRRR forecasts, and NWS discussions. Photovoltaic and price forecasts are propagated through 100-scenario day-block residual-bootstrap scenarios into a CVaR-regularized settlement-proxy quantity rule. In 2024-2025 full-HRRR neural forecasts, the LLM cloud-rule path produced descriptive mean all-hour PV RMSE of 1167.35 MW, compared with 1186.03 MW for no text and 1221.09 MW for keyword rules; time-based support differs by metric and comparison. For real-time prices, LLM-rule weather scores produced descriptive mean RMSE of 17.94 USD/MWh, compared with 18.47 USD/MWh for rule-text scores. In ex-post public-data proxy accounting with a common rule-core anchor and matched [Formula: see text] blend, the LLM scenario path increased proxy value by 45.24 million USD, reduced [Formula: see text] loss by 5.84 thousand USD per hour, and reduced imbalance by 852.7 GWh. These results support public weather narratives as an auditable residual-information layer for solar forecasting and settlement-proxy risk evaluation.