Data-driven Forced Oscillation Localization using Inferred Impulse Responses

Shaohui Liu, Hao Zhu, Vassilis KekatosarXiv 2023 · OpenReview

TL;DR

This paper introduces a novel data-driven method to pinpoint the sources of forced oscillations (FOs) in power grids using only synchrophasor measurements. By learning the system's "fingerprint" (impulse responses) from normal operating data, the method can then efficiently identify FO origins during actual events, even with limited sensor coverage.

What Problem Does This Paper Address?

Interconnected power systems are susceptible to forced oscillations (FOs), which are sustained oscillations that can threaten grid stability and reliability. A critical challenge in mitigating these FOs is accurately and quickly identifying their source. Traditional methods often rely on detailed system models, which can be complex and difficult to maintain for large, dynamic grids. While some data-driven approaches exist, they may lack theoretical grounding or require extensive measurements across the grid. This paper aims to bridge this gap by proposing a purely data-driven framework that is theoretically sound, computationally efficient, and flexible regarding measurement availability.

Methods

The core of the proposed method lies in inferring the power system's impulse responses from ambient (normal operation) synchrophasor data. This is achieved by leveraging the linearized dynamics of the power system and employing cross-correlation techniques.

  1. Impulse Response Recovery (Phase 1 - Offline): During normal grid operation, ambient PMU data is collected. This data is filtered and then used to compute cross-correlations between various input-output pairs. Based on theoretical lemmas derived from linearized system dynamics, these cross-correlations are shown to be equivalent to the system's impulse responses. This phase essentially learns the system's dynamic "fingerprint" without needing an explicit system model.
  2. FO Source Localization (Phase 2 - Online): When a forced oscillation event is detected, the system's frequency domain representation of the FO signal is computed. Using the previously inferred impulse responses, a least-squares (LS) fitting error is calculated for each potential FO source location. The location that minimizes this LS error is identified as the FO source. The paper also extends this to handle output measurements (like bus voltage angles and line flows) by using their phase shifts, and addresses scenarios with multiple FO sources and frequencies.

Datasets

  • IEEE 68-bus system: Used for simulating nonlinear generator models with control effects. Ambient data was generated using zero-mean white Gaussian noise perturbations, and FO signals were generated by sinusoidal perturbations at frequencies close to system modes.
  • 240-bus system from the IEEE-NASPI FO contest: Used for validation. This included an ambient dataset generated by a TSAT simulator and FO signals from the contest dataset.

Key Insights and Ideas

  • Learning System Dynamics from Ambient Data: The most significant insight is the ability to accurately infer crucial system dynamic information (impulse responses) solely from ambient data collected during normal operation. This bypasses the need for detailed, up-to-date system models.
  • Data-Driven Yet Theoretically Grounded: The framework is purely data-driven but is underpinned by theoretical analysis of linearized power system dynamics, providing a strong foundation for its effectiveness.
  • Flexibility in Measurement Types and Locations: The method is shown to be robust to different types of PMU measurements (generator states, bus voltages, line flows) and can function even with partial sensor coverage, making it practical for real-world deployment.
  • Two-Phase Approach for Efficiency: Separating the computationally intensive impulse response recovery (offline) from the rapid FO source identification (online) makes the overall system efficient for real-time applications.

Key Results

  • High Localization Accuracy: The proposed method achieved high localization accuracy across various scenarios, including:
  • Single FO Source: Accuracies of up to 100% for full sensor coverage and over 90% for partial bus/line measurements were reported on the 68-bus system.
  • Two FO Sources: For the 240-bus system, accuracies of over 90% were achieved even with partial sensor coverage.
  • Robustness to Nonlinearities: The method demonstrated effectiveness on realistic nonlinear power system models, despite being theoretically derived from linearized dynamics.
  • Validation on Benchmark Dataset: The approach showed promising results on the challenging IEEE-NASPI FO contest dataset, successfully identifying FO sources even when they were not directly monitored.

Limitations and Future Work

  • Assumptions for Impulse Response Recovery: The impulse response recovery relies on certain assumptions (e.g., swing dynamics, homogeneous damping, lossless lines, specific ambient input characteristics). While these are stated to hold well in large interconnections, their impact on accuracy in systems deviating significantly from these assumptions warrants further investigation.
  • Computational Complexity for Multiple Sources: While the method can handle multiple sources, the computational complexity increases combinatorially. The authors suggest exploring more efficient search methods like sparse signal recovery for these cases.
  • Future Research Directions: The authors suggest extending the model-based analysis to different types of FO sources, analyzing performance with various measurement types and operating conditions, and validating the method with real-world PMU data.

Summary updated

Abstract

Poorly damped oscillations pose threats to the stability and reliability of interconnected power systems. In this work, we propose a comprehensive data-driven framework for inferring the sources of forced oscillation (FO) using solely synchrophasor measurements. During normal grid operations, fast-rate ambient data are collected to recover the impulse responses in the small-signal regime, without requiring the system model. When FO events occur, the source is estimated based on the frequency domain analysis by fitting the least-squares (LS) error for the FO data using the impulse responses recovered previously. Although the proposed framework is purely data-driven, the result has been established theoretically via model-based analysis of linearized dynamics under a few realistic assumptions. Numerical validations demonstrate its applicability to realistic power systems including nonlinear, higher-order dynamics with control effects using the IEEE 68-bus system, and the 240-bus system from the IEEE-NASPI FO source location contest. The generalizability of the proposed methodology has been validated using different types of measurements and partial sensor coverage conditions.
BibTeX
@misc{liu2023data,
  author = {Shaohui Liu and Hao Zhu and Vassilis Kekatos},
  title = {{Data-driven Forced Oscillation Localization using Inferred Impulse Responses}},
  year = {2023},
  eprint = {2310.01656},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2310.01656},
}
Figure