Forced oscillation analysis using physics-based and machine learning models on a comprehensive dataset library

The growing integration of renewable energy resources and flexible loads has introduced new dynamics into modern power systems, resulting in an increased occurrence of oscillatory disturbances. Among them, forced oscillations (FO), caused by periodic disturbance, can propagate widely and compromise system stability. Rapid and accurate localization of oscillation sources is therefore critical. Data-driven methods for oscillation analysis have shown considerable promise. Their development and validation remain constrained by the scarcity of well-documented field-recorded events. They are most often shared anonymously within the research community, making it challenging to study and analyze diverse forced oscillation scenarios. To address this gap, this paper presents three key contributions. First, develop a comprehensive, publicly accessible, time-stamped FO dataset library using realistic power grid models to leverage event availability across a wide range of FO frequencies, magnitudes, and source types. Second, analyze three well-known FO source location methods, namely, the Amplitude-based method, the Dissipation Energy Function Method, and the Cross power spectral density method, using the proposed extensive FO dataset library. Lastly, a Long Short-Term Memory supervised machine learning model was trained and tested that operates independently of power system topology and demonstrates its performance in comparison to traditional physics-based methods for FO source location.