Work

PowerSQL: A Large Language Model-Assisted Domain-Adapted Text-to-SQL Framework for Power System Data Querying

Year: 2026

Type: conference-paper

Abstract: With the rapid advancement of natural language processing, Text-to-SQL has emerged as a key technology for enhancing database accessibility. Given that power system professionals typically lack specialized training in SQL programming, this technology has attracted significant attention in the electric power industry. However, applying Text-to-SQL in power domain scenarios presents several challenges, including domain-specific semantic understanding, complex schema linking, and balancing accuracy with efficiency. To address these issues, this paper proposes PowerSQL, a novel Text-to-SQL framework tailored to power system data querying. This framework integrates a hypothetical SQL-guided schema pruning strategy to minimize redundant inputs and reduce model hallucinations, a hybrid supervised fine-tuning mechanism with dual-format annotations to enhance generalization in logical reasoning, and a reinforcement learning algorithm dubbed DAR-GRPO, which adaptively selects reasoning strategies to optimize both efficiency and accuracy. Experiments on the PGSQL dataset demonstrate that PowerSQL achieves an execution accuracy of 89.67 % with the fewest generated tokens, outperforming state-of-the-art baselines and verifying its practical effectiveness in real-world power system scenarios.

Source: Chinese Control and Decision Conference (CCDC)

DOI: 10.1109/ccdc69976.2026.11559934

Authors: Jie Zhao, Minghao Wang, Jingjing Yang, Zhiguang Hu, Xuejun Jiang, Lifu Ding

Institutions: Zhejiang University, China Southern Power Grid (China)

Language: English

Topic: Topic Modeling

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

SDGs: Affordable and clean energy, Industry, innovation and infrastructure

Open Access status: hybrid

Funder: National Natural Science Foundation of China