Change-point detection and early warning systems

This paper presents a statistical framework for early warning change-point detection in electrical grid frequency time series. Frequency deviations outside the tolerance band of 49.85-50.15 Hz are treated as error events. A high-volatility (HV) measure is computed using a rolling-window approach and compared against a Hoeffding-bound threshold to identify significant transitions that may precede hazardous excursions. A dataset of 1250 error-event sequences collected over six months is divided into training (34%), validation (33%), and testing (33%) subsets. To improve efficiency, k-means clustering and dynamic time warping (DTW) are used to select representative training sequences, and a mapping-with-regression procedure is applied to generate warning signals. Experimental results show that the proposed method achieves 98.04% accuracy and an F1-score of 98.06%, while maintaining a false-negative rate of 1.1%. Lead-time evaluation confirms consistent early detection, and baseline comparison against deep learning approaches, demonstrates competitive performance with low computational cost.