Uncertainty-Aware Anomaly Prediction Using Conformal Multivariate Time Series Clustering

This letter presents a unified framework for robust, uncertainty-aware anomaly detection and classification in cyber-physical system (CPS) time-series data with formal coverage guarantees and low-latency state identification. The temporal and spatial dependencies in the data are modeled using a mixture of time-invariant Gaussian graphical models, with each state characterized by a unique block Toeplitz inverse covariance matrix. This modeling approach captures short-term spatio-temporal stationarity, enabling consistent modeling across sliding windows. An Expectation-Maximization (EM) based approach jointly estimates cluster assignments and structural parameters. Prediction uncertainty is quantified by integrating conformal prediction using negative log-likelihoods non-conformity scores. By leveraging the partial exchangeability of subsequences ensured through the Toeplitz structure, we provide valid distribution-free prediction sets with coverage guarantees. Evaluations on a cyber physical power systems data under normal, fault, and cyber-attack conditions demonstrate 98% empirical coverage and 15% higher calibration accuracy over existing uncertainty-aware methods. The method improves attack detection rates, and reduces detection latency for fast-evolving transients. Results demonstrate accurate, low-latency, and confidence-aware event identification with interpretable structural insights supporting resilient and trustworthy CPS monitoring.