As renewable energy integration and climate extremes reshape power system operations, controlling grid topology during severe weather has become increasingly challenging. Traditional deep reinforcement learning approaches train controllers under normal conditions and treat weather as an implicit background factor, resulting in dangerous performance failures when extreme events occur. A new weather-aware soft actor–critic framework addresses this vulnerability by making weather an explicit, structured input to grid control decisions.
The system employs three key innovations. First, an advanced data augmentation pipeline uses large language models to generate realistic extreme-weather scenarios with physically consistent renewable and load profiles—producing training data 11.7% more challenging than conventional scaling methods. Second, a hybrid weather encoder transforms eight weather dimensions (type, intensity, and trend information) into inputs for a graph-attention neural network that learns adaptive topology actions tailored to specific weather conditions. Third, the framework employs a dual-path deployment strategy: controllers train using ideal oracle weather data but operate using weather inferences, achieving 88% functional equivalence while degrading only 10%.
Testing on the IEEE 118-bus system from the Grid2Op 2022 competition across eight extreme-weather categories demonstrated substantial improvements. Weather encoding alone improved extreme-scenario performance by 18.6% compared to controllers without explicit weather inputs. The LLM-generated training scenarios forced the system to learn more robust control policies than conventional synthetic data. Critically, the framework avoided the "worse-than-do-nothing" behavior that plague many advanced controllers during crises—a safety essential for grid operators.
This work addresses a genuine blind spot in grid automation: the assumption that controllers trained in calm conditions can handle chaos. By explicitly modeling weather as structured, measurable input rather than hidden noise, the approach aligns machine learning practice with grid physics. For utilities managing grids with 30%, 50%, or higher renewable penetration, this framework offers a practical pathway to autonomous control systems that remain stable under the extreme weather scenarios increasingly common in the climate era.
