Accurate distribution system state estimation (DSSE) under limited measurement scenarios remains a critical challenge due to the inherent high-dimensional nonlinearity and data scarcity in power distribution networks. Leveraging the exceptional few-shot pattern recognition capabilities of pretrained large language models (LLMs)

this paper proposes a novel DSSE framework that synergizes LLMs with multimodal data integration

enabling effective voltage magnitude and phase angle estimation. Moreover

this study introduces a prompt engineering approach that transforms statistical features extracted from historical data into contextual prompts. By employing natural language representations

this method enhances temporal feature extraction and uncovers latent patterns in scenarios with limited measurements. Furthermore

a channel-independent multimodal fusion architecture that structurally aligns time-series measurements

real-time sensor data

and textual prompts while preserving modality-specific characteristics. Extensive validation on IEEE 33-bus and Simbench 144-bus systems demonstrates the effectiveness of the proposed method

which reduces MAE by 17.7%–25.1% and RMSE by 17.3%–18.2% compared to the second best data-driven baselines under limited measurement scenarios. These results highlight the framework's strong generalizability across diverse network topologies and its potential for practical deployment in poorly instrumented distribution networks. Our codes are available at https://github.com/gmy1997ee/LLM-MultiModal-for-DSSE.