Learning the Frequency Dynamics of the Power System Using Higher-order Dynamic Mode Decomposition

Xiao Li, Xinyi Wen, Benjamin Schäfer

TL;DR

As renewable energy reduces grid inertia, predicting frequency oscillations has become both more critical and more difficult. This paper introduces a data-driven approach using Higher-Order Dynamic Mode Decomposition (HODMD) to capture these complex, nonlinear dynamics without requiring prior physical knowledge or manual feature engineering. The method outperforms standard linear techniques and complex nonlinear alternatives (like SINDy) in high-dimensional power system models, even in the presence of significant measurement noise.

What Problem Does This Paper Address?

Modern power grids are undergoing a massive shift: traditional heavy spinning generators (coal, gas, hydro) are being replaced by inverter-based renewables (solar, wind). These new sources provide very little "inertia," meaning the grid's frequency can swing wildly and rapidly when a disturbance occurs.

To keep the lights on, engineers need to monitor and predict these frequency oscillations. Historically, researchers used linearized models to simplify the math. However, the authors argue that even small deviations from equilibrium in a power system exhibit nonlinear behavior that linear models simply miss. While some data-driven nonlinear methods exist, they often struggle with "the curse of dimensionality" (too many variables in a large grid) or require "basis functions"—essentially requiring the researcher to guess the shape of the nonlinearity before they start.

Methods

The authors propose using Higher-Order Dynamic Mode Decomposition (HODMD).

To understand HODMD, first consider standard DMD: it tries to find a linear operator that maps the system's state at time $t$ to time $t+1$. HODMD takes this a step further by using delayed embedding. Instead of just looking at the current snapshot, it looks at a sequence of past snapshots ($d$ snapshots).

Why this matters: * Handling Nonlinearity: By looking at the "history" of the signal, HODMD can approximate nonlinear dynamics using a higher-order linear difference equation. It’s a mathematically elegant way to capture "weak" nonlinearities without complex neural networks. * Model-Free: It doesn't need the underlying swing equations of the generators; it learns purely from time-series data. * Spatio-Temporal De-noising: It uses Singular Value Decomposition (SVD) to project high-dimensional data into a lower-dimensional subspace, effectively filtering out noise and focusing on the most important "modes" of oscillation.

Datasets

The researchers validated their approach using two standard power system benchmarks simulated via the ANDES Python package:

  1. IEEE 14-bus system: A smaller system with 5 generators, used to test the method's ability to capture rapidly decaying dynamics.
  2. WECC (Western Electric Coordinating Council) 179-bus system: A large-scale model with 29 generators, characterized by complex, low-frequency "inter-area" oscillations where groups of generators swing against each other.

For both systems, they simulated short-circuit faults to trigger oscillations and used the resulting frequency and rotor angle data for training and testing.

Key Insights and Ideas

  • The Power of $d$: The "order" $d$ is the secret sauce. The authors found that while standard DMD ($d=1$) failed to capture oscillations in the WECC system, increasing $d$ to 15 allowed the model to fit the curves with high precision.
  • Local vs. Global Modes: HODMD successfully separated "local" modes (where a single generator oscillates) from "global" modes (where the entire system swings together). This is crucial for grid operators to identify whether a problem is isolated to one plant or is a systemic threat.
  • Superiority over SINDy and LANDO: In high-dimensional settings, popular methods like SINDy (Sparse Identification of Nonlinear Dynamics) often fail because they require a massive library of potential functions. HODMD avoids this "feature selection" headache entirely, making it more robust for large-scale grids.
  • Noise Resilience: Even when the Signal-to-Noise Ratio (SNR) was dropped to 10 (significant noise), HODMD's error remained lower than standard DMD's error in a perfectly clean environment.

Key Results

  • Accuracy: On the IEEE 14-bus test data, HODMD achieved a Relative Root Mean Square Error (RRMSE) of 15.2%, compared to 71.9% for both LANDO and SINDy.
  • Scalability: On the much larger WECC system, HODMD maintained a respectable RRMSE of 36.7%, while alternative nonlinear methods struggled to converge or produced divergent predictions.
  • Modal Identification: The authors used Fast Fourier Decomposition to verify their results, proving that HODMD correctly identified specific oscillation frequencies (e.g., a 1.26 Hz local mode at Bus 69 and a 0.10 Hz global system mode).

Limitations and Future Work

The primary limitation acknowledged by the authors is that this study relies on simulated data. Real-world PMU (Phasor Measurement Unit) data from actual power grids is often messier, contains gaps, and requires significant pre-processing.

The authors' next goal is to adapt these HODMD techniques to handle real-world measurement data and explore how these learned models can be integrated into real-time control systems to stabilize the grid automatically.

Summary updated

Abstract

The increasing penetration of renewable energy sources, characterised by low inertia and intermittent disturbances, presents substantial challenges to power system stability. As critical indicators of system stability, frequency dynamics and associated oscillatory phenomena have attracted significant research attention. While existing studies predominantly employ linearized models, our findings demonstrate that linear approximations exhibit considerable errors when predicting frequency oscillation dynamics across multiple time scales, thus necessitating the incorporation of nonlinear characteristics. This paper proposes a data-driven approach based on higher-order dynamical mode decomposition (HODMD) for learning frequency dynamics. The proposed method offers distinct advantages over alternative nonlinear methods, including no prior knowledge required, adaptability to high-dimensional systems, and robust performance. Furthermore, HODMD demonstrates superior capability in capturing system-wide spatio-temporal modes, successfully identifying modal behaviour that remains undetectable through standard Dynamic Mode Decomposition techniques. The efficacy of the proposed methodology is validated through comprehensive case studies on both IEEE 14-bus and WECC systems.
BibTeX
@misc{li2025learning,
  author = {Xiao Li and Xinyi Wen and Benjamin Schäfer},
  title = {{Learning the Frequency Dynamics of the Power System Using Higher-order Dynamic Mode Decomposition}},
  year = {2025},
  eprint = {2502.06186},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2502.06186},
}
Figure