Llm-assisted physics-guided spatiotemporal graph learning for power meteorological disaster evolution
TL;DR
This paper proposes an LLM-Assisted Physics-Guided Spatiotemporal Graph Learning model (LPG-STGNet), a novel paradigm for power systems to cope with extreme meteorological disasters, featuring both physical interpretability and semantic understanding capabilities.
Abstract
Extreme meteorological events—such as winter cold waves, ice accretion, and typhoons—are increasingly becoming primary triggers that threaten the safe and stable operation of power systems. Their disaster-causing processes exhibit strong spatial correlations and multi-stage temporal evolution characteristics, posing severe challenges to traditional static risk assessment methods. Although existing spatiotemporal graph neural network approaches can characterize grid topological dependencies, they generally suffer from limitations including the absence of physical mechanism constraints, insufficient utilization of semantic information from disaster warning texts, and the inability of single-scale graph structures to capture the inherent hierarchical nature of “bus-level fine-grained topology versus meteorological-zone coarse-grained granularity.” To address these issues, this paper proposes an LLM-Assisted Physics-Guided Spatiotemporal Graph Learning model (LPG-STGNet). The proposed model incorporates a large language model (LLM) to perform structured semantic encoding of meteorological warning texts, constructs dynamic graph failure priors based on wind load guidelines and physical equations governing conductor ice accretion growth, and designs a hierarchical spatiotemporal graph encoder that integrates bus-level fine-grained topology with meteorological-zone coarse-grained structure. Furthermore, a semantic–physical dual-gating fusion mechanism is proposed to achieve adaptive weighting among multi-source information, thereby synergistically leveraging data-driven representations, physical priors, and semantic priors within a unified framework. Experiments are conducted on a disaster-evolution dataset constructed through DC-OPF cascade simulation on the ACTIVSg2000 synthetic power grid—built from real geographic and electrical parameters of the Electric Reliability Council of Texas (ERCOT) region and comprising 2,000 bus nodes, 1,250 substations, 3,206 branches, and 8 real meteorological zones—covering 600 synthetic winter cold-wave scenarios (43,200 hourly snapshots). Evaluated against nine mainstream spatiotemporal graph baselines under a unified protocol with three random seeds, LPG-STGNet achieves an F1-score of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.798\pm 0.004$$\end{document} and a load-loss RMSE of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10.02\pm 0.15$$\end{document} MW on the branch-level trip-classification and node-level load-loss regression tasks, respectively, representing improvements of 6.2 percentage points in F1-score and 21.3% in RMSE over the strongest single-modality baseline. Ablation studies further confirm the effectiveness and complementarity of the physics-guided module and the LLM semantic module. This work provides a novel paradigm for power systems to cope with extreme meteorological disasters, featuring both physical interpretability and semantic understanding capabilities.