PowerSQL-Agent: Domain-Aware NL2SQL for Power System Data Analytics
Power systems continuously generate large volumes of operational data through SCADA, PMS, and other industrial platforms; however, existing natural language to SQL (NL2SQL) methods often perform poorly in these scenarios due to domainspecific terminology, complex schemas, and ambiguous user intent. To address these challenges, this paper proposes an LLM-based NL2SQL framework for power system databases that introduces a Power Semantic Linking (PSL) mechanism for domain-aware schema alignment and an Execution-Guided Refinement Loop (EGRL) to improve SQL reliability through execution feedback. Experiments on a Power-SQL dataset containing 720 industrial query pairs show that the proposed framework improves schemalinking accuracy to 89.7% and hard-query execution accuracy from 35.8% to 67.5%, significantly outperforming baseline prompting methods. Furthermore, the framework enables medium-scale LLMs to achieve competitive performance with lower inference latency, providing an effective solution for natural language interaction with power system databases.