GridSage: A Skill-grounded LLM platform for scenario-oriented reinforcement learning evaluation in distribution networks

Reinforcement learning (RL) research in distribution networks continues to face challenges associated with complex scenario construction, high configuration overhead, and limited experimental reproducibility. Although existing RL platforms such as PowerGym and PowerGridworld provide standardized simulation and interaction interfaces, translating research intent into executable scenario configurations still relies heavily on manual effort. To address this issue, this paper proposes GridSage, an integrated platform that combines large language models (LLMs) with physics-aware power system simulation. GridSage adopts a Scenario Skill retrieval and incremental semantic editing framework, enabling users to construct customized scenarios through natural-language instructions. All generated scenarios are further subjected to safety validation to verify their physical validity and operational consistency. The validated scenarios are subsequently deployed into a Gymnasium-compatible RL environment incorporating photovoltaic generation, wind power, energy storage systems, electric vehicles, and vehicle-to-grid (V2G) interaction, supporting the training and evaluation of mainstream RL algorithms. On an independently prepared 600-instruction benchmark, incorporating Scenario Skill retrieval increased semantic success from 36.50% to 80.33%. The layered safety validator achieves a classification accuracy of 89.52% on 210 test cases. In a comparison involving four participants, GridSage reduces the average scenario-configuration time by 43.58%. Cross-LLM experiments further demonstrate that the proposed configuration workflow can operate effectively across different LLM backends. These results show that GridSage provides an effective and auditable approach for translating natural-language research intent into reproducible RL scenarios for active distribution networks.