Unsupervised Load Transfer Detection Based on Wavelet Change Point 99Analysis and Isolation Forest

This study proposes an unsupervised framework for detecting load-transfer events in distribution systems, which are frequent but hard to identify due to missing external logs and lack of labeled data. Using only load time series data, the method first removes seasonal and trend components via Robust STL decomposition. The residual signal is transformed using a Haar-based Stationary Wavelet Transform, and candidate change points are identified by the Pruned Exact Linear Time (PELT) algorithm. For each point, 15 statistical features are computed, including load variation, residual statistics, slope, ratio, and time location. These features are used as input to an Isolation Forest-based anomaly detector, which probabilistically determines load-transfer events. The method shows that accurate detection is possible using load data alone, without external sensors or prior labels, offering a scalable and practical solution for real-world distribution system monitoring.