An LLM‐Driven Approach for Power Grid Structure Synthesis and Visualization

Abstract Research in power system is persistently hampered by the scarcity of high‐fidelity, public grid data due to security and privacy constraints. Existing unimodal synthesis methods fail to harmonize physical laws, visual representations and semantic descriptions, obstructing the application of multimodal large language models (LLMs) in the energy sector. To address this, we propose a multimodal synthesis framework that generates aligned datasets comprising physical parameters, single‐line diagrams and natural language descriptions. The framework combines rule‐based topology generation with a two‐stage chain‐of‐thought (CoT) strategy, enabling LLM agents to initialize electrical parameters based on statistical priors. To ensure physical feasibility, an iterative power flow feedback loop is introduced to guarantee convergence. Furthermore, retrieval‐augmented generation is employed to enhance component‐level visual details. Experimental results indicate that the synthesized grids achieve high structural similarity and physical fidelity compared to real‐world benchmarks. We have open‐sourced this physically validated multimodal grid dataset to provide critical foundational support for developing physics‐informed “energy LLMs.”