LLM Based Approach for Anomaly Detection in Smart Grids

Abstract

The emergence of smart grids has revolutionized the power sector, fostering interconnected and intelligent electricity networks. These advancements offer enhanced monitoring and control capabilities, but also introduce challenges in anomaly detection, a critical aspect for grid stability and security. Traditional methods, while effective at identifying anomalies, often lack interpretability, making it difficult to understand the root cause and formulate an effective response. Here, Large Language Models (LLMs) offer a promising approach. By generating human-friendly descriptions of anomalies, LLMs can bridge the gap between raw data and actionable insights. This dissertation investigates the potential of LLMs for anomaly detection in smart grids. We leverage the CICIDS2017 dataset to explore how LLMs can be harnessed to improve the interpretability of anomaly detection systems. Specifically, we examine the effectiveness of using SHAP (SHapley Additive exPlanations) values to guide the LLM towards generating more accurate descriptions of the detected anomalies. Notably, our research demonstrates that even with minimal fine-tuning, the Llama3 8B LLM achieves remarkable results when prompted effectively. This highlights the crucial role of prompt engineering in unlocking the full potential of LLMs for this domain. By incorporating SHAP values within our prompting strategy, we are able to bridge the gap between anomaly detection and actionable insights. This empowers decision-making experts with valuable information to respond to anomalies effectively, ensuring the continued reliability of smart grid operations.