Large Language Model-Based Security Situation Awareness for Smart Grid: Framework and Approaches

The smart grid integrates power generation, communications, and computation to improve efficiency and reliability, but its heterogeneity enlarges the attack surface. This study introduces GridSense, a novel security situational awareness framework leveraging large language models to address the limitations of traditional methods and enhance predictive accuracy that has transformed electrical, meteorological, and social data into structured text prompts for large language models, leveraging their advanced reasoning capabilities. By employing prompt engineering and meta-learning techniques, GridSense has achieved robust performance. Experiments on real-world grid data in both full-shot and few-shot scenarios have demonstrated a multimodal load forecasting RMSE of 0.15 and an anomaly detection accuracy of 90.25% with an F1 score of 0.43 using combined data modalities. Even with limited training data, GridSense has maintained effective forecasting performance, showcasing its adaptability. Error analysis shows high precision but missed rare anomalies; recall can be improved via recall-oriented thresholds and class-imbalance remedies. GridSense demonstrates that LLMs can unify heterogeneous smart grid signals into an interpretable pipeline. These results indicate a practical path for integrating LLMs reasoning with grid operations to enhance awareness and resilience.