Real-time multi-label event categorization in the Brazilian interconnected power system

Abstract

The increasing complexity of power systems, driven by the decentralization of generation and the growing demand for operational reliability, has reinforced the importance of Wide-Area Monitoring Systems (WAMS) for dynamic system analysis. Among WAMS applications, event detection and classification are essential for enabling timely and accurate responses to disturbances. However, event classification methodologies face significant challenges, such as the occurrence of multiple and sequential events, disturbance overlap, and a high degree of similarity among different classes. This work proposes a multi-label classification methodology based on Convolutional Neural Networks (CNNs), applied to events detected using a two-level robust approach consisting of the Discrete Wavelet Transform-based spectral analysis, followed by a deep neural network (DNN) strategy that prevents false alarms. The main contributions include: (i) the construction of a multi-label dataset comprising 11,205 real events from the Brazilian Interconnected Power System (BIPS), covering Line Tripping (LnT), Electromechanical Oscillations (EO), Loss of Load (LoL) and Generation Tripping (GT); (ii) the development of a classification model based solely on real synchrophasor data; and (iii) its experimental validation during six months of online operation in the BIPS, including common challenges associated with Phasor Measurement Unit (PMU) data, such as noise, missing data, and synchronization errors. The model achieved a Hamming loss of 0.082, correctly classifying 91.8% of the events during deployment. These results demonstrate the effectiveness and robustness of the proposed approach in operational contexts, contributing to faster and more reliable decision-making in power system control centers.