ShapeDiff: A Universal Source-Load Time Series Anomaly Detection and Imputation Model

The rapid proliferation of sensors in modern power systems has generated vast time-series datasets that are rich in critical operational information. To maximize the utility of these datasets and support the development of high-quality pre-trained models, it is crucial to implement efficient preprocessing methods that effectively address both anomaly detection and data imputation. This paper introduces ShapeDiff, a novel integrated framework that addresses anomaly detection and imputation inefficiencies within time-series preprocessing. ShapeDiff combines shapelet-based anomaly detection for enhanced stability and interpretability with diffusion model-based imputation to ensure superior performance and adaptability to non-stationary data. A self-supervised parallel training strategy facilitates the efficient processing of large-scale datasets while maintaining inference consistency. Experiments demonstrate that ShapeDiff achieves efficient anomaly detection, provides accurate and reliable imputation for common source-load time series, and significantly improves preprocessing efficiency.