A new data-driven multi-task event detection approach based on real-world PMU data for active distribution networks

Detecting tap-changing (TC), load-switching (LS), and arcing events are critical for maintaining the reliable operation of distribution systems. Both supervised and unsupervised methods have been widely investigated for detecting these three types of events using phasor measurement unit (PMU) data. While supervised learning approaches can achieve better performance, dependence on large and well-labelled datasets severely limits their applicability to real-world systems. As a result, unsupervised methods have gained increasing attention. However, existing multitask unsupervised approaches often struggle to distinguish heterogeneous events due to their reliance on a unified analytic framework, where voltage-driven and current-driven disturbances are treated within a common representation without considering the underlying physical behavior of the events. As a result, the defining characteristics of an event can be masked and one event type may dominate in the presence of multiple events. To address this limitation, an adaptive multiscale directional median filter is first proposed to enhance PMU measurement quality. Then the proposed framework performs TC detection using voltage measurements, while current measurements are analyzed through an unsupervised multi-model clustering approach with adaptive model selection for LS and arcing event detection. Unlike existing approaches using a unified framework, the proposed method employs a decoupled approach of voltage-driven and current-driven events based on their underlying physical behavior, while integrating a multi-model unsupervised clustering approach with adaptive model selection for event detection. The proposed method is validated using real-world labeled PMU data and then applied to investigate the impact of the rooftop solar power generation on distribution networks.