OSCILLATION DETECTION AND CLASSIFICATION IN POWER SYSTEMS USING DATA-DRIVEN AND MACHINE LEARNING TECHNIQUES

Deepak Joshi

The evolution of modern power systems has reached a new era with the widespread integration of inverter-based resources and the emergence of large, complex load centers. This transition has introduced new challenges associated with accurately characterizing and analyzing the fast dynamic behavior. Oscillation is one of the critical dynamic behaviors that arise from adverse interactions among different system components and system controls. It poses significant challenges to reliable operation, potentially disrupting energy generation and transmission. Hence, early detection of these oscillations and isolation of oscillation sources are critical to maintaining system reliability, as unstable oscillatory modes can otherwise grow into large-amplitude oscillations, potentially leading to system failures and large-scale blackouts. Conventional physics-based modeling and analytical approaches have demonstrated limited applicability in addressing the nonlinear, high-frequency oscillations induced by inverter-dominated and fast-changing loads. The widespread deployment of advanced measurement technologies, such as phasor measurement units and point-on-wave recorders, has enabled access to real-time availability of high-resolution data. This has opened up new research directions focused on the development of data-driven machine learning methods for the timely detection, classification, and analysis of critical oscillation events. Needless to say, these events are undesirable and must be effectively addressed.In this dissertation, we propose a novel image-based feature extraction approach for oscillation detection and classification based on the severity of the event types. For the study, the oscillation events are categorized into three classes, namely well-damped, medium-damped, and poorly damped, based on the damping characteristics of the oscillatory modes. A deep learning framework leveraging image-based feature extraction is developed to automatically detect and classify oscillation events. The performance of the proposed transfer learning–based convolutional neural network (TL-CNN) has been evaluated against conventional physics-based methods, demonstrating the superiority of the proposed image-based feature extraction in accurately characterizing oscillatory phenomena. As part of this dissertation, we have developed a comprehensive forced oscillation dataset library using realistic two-power grid models, the Mini-WECC system, and the validated 1996 model of WECC, encompassing a wide range of oscillation sources, frequencies, and magnitudes. This dataset library serves as a standardized benchmark for evaluating and comparing existing algorithms. For our study, we have analyzed three existing forced oscillation source location algorithms, namely the amplitude-based method, the dissipation energy method, and the cross-power spectral density method, using the proposed comprehensive library, and show that none of the popular existing methods is universally applicable. Finally, we have proposed a long short-term memory (LSTM) based classifier model for locating the source of forced oscillation. The LSTM model is trained using the comprehensive dataset, and the model has shown promising results in correctly identifying the source of forced oscillation.

OSCILLATION DETECTION AND CLASSIFICATION IN POWER SYSTEMS USING DATA-DRIVEN AND MACHINE LEARNING TECHNIQUES

Files and links (1)

Abstract

Metrics

Details