Poorly damped sub-synchronous oscillations (SSOs) are becoming more prevalent as power systems transition to high shares of inverter-based resources (IBRs), including wind, solar photovoltaics, and grid-scale battery storage. These oscillations can emerge unexpectedly, evolve over time and spread across wide areas posing significant operational challenges. A key barrier to effective mitigation is the difficulty of identifying the root cause of such oscillations. This is often due to lack of a representative wide-area model of an IBR-dominated grid that can replicate such oscillations. Even where such a model exists, its practical use is limited by the complexity and lack of transparency of vendor-specific IBR control systems.

To overcome this, we present a data-driven method to pinpoint major contributors to poorly damped oscillations directly from time-synchronised phasor data, without requiring any model information. The approach is based on Extended Dynamic Mode Decomposition (EDMD), grounded in Koopman operator theory which enables the extraction of nonlinear response dynamics from time-series data. Using voltage and current phasors from phasor measurement units (PMUs) across the grid, the method constructs a reduced-order finite-dimensional Koopman representation that captures dominant oscillatory modes and quantifies the relative participation of individual PMU locations including IBR plants and other components (e.g. conventional synchronous machine-based power plants, large loads etc.). Reduced order is essential to eliminate artefact modes in the vicinity of the actual ones which can lead to misleading results.

The effectiveness of the EDMD method is validated using two real-world oscillation events in Denmark and Australia. The Danish case study involves a multi-stage oscillation event around 3 Hz, where oscillatory behaviour evolves over time with changing modal damping characteristics. Using PMU data from part of the Danish grid, EDMD consistently captures the dominant mode and accurately identifies the major contributing locations, as later confirmed by Energinet. In the Australian case study, EDMD successfully identifies the top contributors in a relatively high-frequency oscillation (21 Hz) mode captured using P-type PMUs. The two case studies demonstrate EDMD’s capability to accurately pinpoint the root cause of oscillations across a range of sub-synchronous frequencies.

The proposed workflow integrates spectral analysis and Koopman-based modal decomposition to systematically pinpoint major contributors to poorly damped oscillation, without relying on any prior system knowledge. By providing a clear, data-driven spatiotemporal characterisation of oscillations, the method supports targeted mitigation. Overall, this paper demonstrates that EDMD-based data-driven analysis offers a practical and scalable solution for pinpointing the root cause of oscillations, with strong potential for both post-event diagnostics and future real-time applications with a recursive implementation.

The full paper will present (i) an overview of the EDMD method and the complete workflow of how the relative participation of each location (component) can be obtained from PMU data and (ii) successful validation of the EDMD method for two real oscillation events from Australia and Denmark.