Large-model-based smart agent for time series anomaly detection in power systems

Anomaly detection in time series is critical to ensure the safe and stable operation in power systems. Existing methods face dual challenges of data scarcity and limited interpretability. To address this, we propose a novel large-model-based series-detection development framework named SLEP, leveraging the powerful pre-trained knowledge transfer capabilities of large models to mitigate data scarcity and their natural language understanding/generation abilities to enhance the interpretability of anomalies. Additionally, we introduced a brand-new prompt design template called BRIDOR introduced to further control the interaction quality with the large model. This work represents the first attempt to apply large models to anomaly detection for electric time series data. Extensive experiments on the high-dimensional, multivariate, and data-sparse real-world cases, demonstrate the superior performance of our proposed method. This validates the framework's robustness and pioneers a new paradigm for trustworthy artificial intelligence technology in related fields.