KG4RSV: Towards Knowledge Graph -driven Specification and Validation of ML Dataset Quality Requirements: Case Study on e -ICU dataset

Abstract

Command centers of modern power transmission and distribution systems ingest massive heterogeneous data structured measurements e.g., Supervisory Control and Data Acquisition (SCADA), topological models Geographic Information System (GIS), and unstructured operational logs-creating a critical challenge for timely, accurate, and explainable decision support. Existing single-tool or siloed approaches fail to jointly interpret natural language queries over hybrid data while ensuring reliability and traceability. This study introduces Hybrid Data Agent (HAL), a centralized multi agent Large Language Model (LLM) system designed for natural language hybrid data querying and decision support in power grid command environments. HAL employs an LLM-based central coordinator to decompose user queries into subtasks and dispatch them to specialized expert agents-such as Structured Query Language (SQL) execution, semantic document retrieval, time-series anomaly detection, and causality reasoning-then fuses their outputs into coherent, verified responses. The system is evaluated on a high-fidelity simulated benchmark reflecting realistic grid operational scenarios, including electrical physical events (e.g., current surges and protection switching). Comparative experiments against baselines (generic LLM agents and traditional pipelines) show that HAL substantially improves query success rate, answer accuracy, and robustness, while maintaining explainability via structured “think-act-observe” execution logs. Contributions include: 1) a distributed yet centrally coordinated architecture combining LLM planning with deterministic domain tools; 2) a hybrid data fusion mechanism aligning structured and unstructured evidence for natural language interfaces; 3) an explainable decision support engine tailored to power grid command centers. HAL reduces cognitive load on operators, accelerates fault diagnosis, and strengthens real-time operational intelligence, making it suitable for deployment in intelligent information systems for critical infrastructure.