⚡ electrical engineering

PowerSINDy: Identifying Nonlinear Time-Dependent Dynamics in Power Grid Frequency

This paper introduces PowerSINDy, a framework that successfully identifies nonlinear, time-dependent dynamics in real-world power grid frequency data from Continental Europe and South Korea, while benchmarking sparsity-promoting regression strategies to determine that LASSO achieves the lowest error and STLSQ offers the best balance between accuracy and stability.

Original authors: Xinyi Wen, Xiao Li, Leonardo Rydin Gorjão, Veit Hagenmeyer, Benjamin Schäfer

Published 2026-05-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xinyi Wen, Xiao Li, Leonardo Rydin Gorjão, Veit Hagenmeyer, Benjamin Schäfer

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the electrical power grid as a giant, complex orchestra. The "frequency" of the electricity (usually 50 or 60 beats per second) is the conductor's tempo. If the tempo stays steady, the music sounds perfect. If it speeds up or slows down, it means the musicians (power generators) and the audience (people using electricity) aren't perfectly in sync.

For a long time, scientists tried to write down the "sheet music" (the math equations) for this orchestra using simple rules. But modern power grids are messy. They have wind and solar power that come and go, and they react in complicated, non-linear ways. The old, simple rules don't capture the full picture.

This paper introduces a new tool called PowerSINDy to figure out the real "sheet music" just by listening to the recording of the orchestra, without needing to know every single instrument's details beforehand.

Here is how they did it, explained simply:

1. The Challenge: A Noisy Recording

Imagine trying to transcribe a song from a recording that has static, wind noise, and people talking in the background. If you try to write the notes down immediately, you'll get it wrong.

The Paper's Solution: Before trying to find the rules, they used a "noise-canceling headphone" (a mathematical filter called a Gaussian filter) to clean up the data. They tested different levels of cleaning to find the "sweet spot" where the noise was gone, but the actual rhythm of the music was still clear. They found that smoothing the data over a 60-second window worked best.

2. The Detective Work: PowerSINDy

Once the data was clean, they used a method called SINDy (Sparse Identification of Nonlinear Dynamics).

The Analogy: Imagine you have a giant toolbox filled with every possible mathematical shape you could use to build a model (straight lines, curves, waves, squares, cubes, etc.).
The Goal: You want to build a model that predicts the grid's behavior, but you want to use as few tools as possible. You don't want a model that uses 1,000 tools to explain a simple swing; you want the 3 or 4 tools that actually matter. This is called "sparsity."
The Upgrade: The authors realized that standard tools (just straight lines and simple curves) weren't enough for power grids. So, they added Fourier terms (mathematical waves) to the toolbox. This allowed them to capture the natural "wiggles" and oscillations in the grid's frequency that simple curves miss.

3. The Three Judges (Optimizers)

To decide which tools to keep and which to throw away, they used three different "judges" (mathematical strategies) to pick the best model. They tested these judges on real data from two different "orchestras":

Continental Europe (CE): A massive, interconnected grid with huge inertia (like a giant, heavy flywheel).
South Korea (SK): A smaller, more compact, centrally managed grid.

The three judges were:

LASSO: The "Precision Sniper." It aggressively cuts away anything that isn't absolutely necessary.
STLSQ: The "Steady Hand." It takes a step-by-step approach, checking and re-checking to make sure the model doesn't fall apart.
SR3: The "Flexible Thinker." It tries to balance fitting the data perfectly with keeping the model simple, but it can be a bit temperamental.

4. The Results: Who Won?

The paper ran thousands of tests to see which judge produced the most accurate and stable "sheet music."

The Accuracy Winner (LASSO): LASSO consistently found the most accurate equations. It produced the lowest error rates, meaning the model it built matched the real-world data better than the others. It was especially good at finding the complex, non-linear interactions (like how two different parts of the grid affect each other).
The Stability Winner (STLSQ): While LASSO was the most accurate, STLSQ was the most reliable. It rarely produced models that "crashed" or went crazy when simulating the future. It offered the best balance between being accurate and being safe.
The Wildcard (SR3): This judge was inconsistent. Sometimes it did okay, but it was very sensitive to how it was set up. It tended to keep too many tools in the toolbox, making the models more complex and harder to understand.

5. What Did They Learn?

Bigger isn't always better: When they tried to make the models super complex (using cubic equations), the models became unstable and started to "diverge" (go off the rails).
Waves matter: Adding the wave-like (Fourier) terms helped the models handle the natural oscillations of the grid without making the math too messy.
The "Secret Sauce": In the best models, they found that specific interactions between the grid's angle and its speed (mathematically called $\theta\omega$ and $\theta\omega^2$ ) were crucial. These are the hidden rules that keep the grid stable.

The Bottom Line

The paper shows that by cleaning the data carefully and using the right mathematical "detective" (specifically LASSO for accuracy or STLSQ for stability), we can discover the hidden, non-linear laws that govern how power grids behave. This helps us understand the grid better than the old, simplified models ever could, using real-world data from Europe and South Korea.

Technical Summary: PowerSINDy for Power Grid Frequency Dynamics

Problem Statement
The stable operation of modern power systems relies heavily on the accurate characterization of grid frequency dynamics, which reflect the balance between generation and demand. Traditional modeling approaches, such as the aggregated swing equation, often rely on linear assumptions and simplified physical parameters that are insufficient to capture the complex, nonlinear, and time-dependent behaviors introduced by high renewable energy penetration and stochastic fluctuations. While system identification methods like Sparse Identification of Nonlinear Dynamics (SINDy) offer a data-driven alternative to uncover governing equations, their application to real-world power grid data remains limited. Existing studies have largely focused on synthetic data, leaving open questions regarding the performance of SINDy on noisy, complex empirical datasets and the trade-offs between different sparsity-promoting regression strategies.

Methodology: The PowerSINDy Framework
The authors propose PowerSINDy, a framework tailored to identify nonlinear, time-dependent dynamics in power system frequency data. The methodology involves three primary stages:

Data Preprocessing and Smoothing:
High-resolution frequency data (1-second resolution) from two distinct synchronous grids—the Continental Europe (CE) grid (50 Hz) and the South Korea (SK) grid (60 Hz)—were analyzed. To isolate deterministic dynamics from stochastic noise, a Gaussian filter was applied. The optimal smoothing parameter ( $\sigma$ ) was determined by minimizing the Root Mean Square Error (RMSE) between SINDy-simulated trajectories and the original empirical data. The study identified $\sigma = 60$ (a 60-second window) as the optimal balance between noise suppression and feature preservation.
Candidate Library Construction:
Unlike standard SINDy implementations that rely solely on polynomials, PowerSINDy extends the candidate function library to better represent power grid dynamics. The library includes:
- Polynomial terms: Up to quadratic ( $p2$ ) and cubic ( $p3$ ) orders to capture nonlinear couplings.
- Fourier terms: First-order sine and cosine functions ( $p2f1$ ) to represent oscillatory behaviors inherent in synchronous grids.
- Time-dependent terms: Explicit inclusion of time ( $T$ ) and interaction terms involving time to account for deterministic signals driven by load curves and generation schedules.
Sparse Regression Strategies:
The framework benchmarks three distinct optimization strategies to solve the sparse regression problem $\dot{X} = \Theta(X)\Xi$ :
- Sequentially Thresholded Least Squares (STLSQ): Iteratively solves least squares and thresholds small coefficients.
- LASSO (Least Absolute Shrinkage and Selection Operator): Uses an $L1$ penalty to shrink coefficients to zero.
- Sparse Relaxed Regularized Regression (SR3): Introduces an auxiliary variable and a relaxation parameter to separate data fitting from sparsity enforcement, tested with $L0$ , $L1$ , and $L2$ norms.

Key Results
The study evaluated the framework across 15-minute trajectories from the CE and SK datasets, comparing model accuracy (stable RMSE), stability fractions (percentage of non-divergent trajectories), and structural complexity (number of active features).

Optimizer Performance:
- LASSO consistently achieved the lowest stable RMSE across all configurations. For the CE dataset with cubic polynomials ( $p3$ ), LASSO reached a minimum RMSE of 0.0101. It also produced the sparsest models, particularly in lower-order configurations.
- STLSQ demonstrated the highest stability, with stability fractions frequently exceeding 0.79 in quadratic ( $p2$ ) settings. It offered the best balance between accuracy and reliability, though with slightly higher RMSE than LASSO.
- SR3 exhibited the highest variability and sensitivity to hyperparameters and dataset characteristics. Notably, the $L0$ and $L1$ norms in SR3 produced nearly indistinguishable outcomes in terms of RMSE and stability. SR3 performance degraded significantly in cubic ( $p3$ ) configurations, particularly for the SK dataset, where stability fractions dropped below 0.52.
Model Complexity and Library Choice:
- Quadratic ( $p2$ ) models provided stable solutions with low RMSE (~0.016) across both datasets.
- Cubic ( $p3$ ) models improved accuracy for the CE dataset (RMSE 0.0101) but reduced stability, leading to faster trajectory divergence.
- Fourier-augmented models ( $p2f1$ ) offered a middle ground, maintaining robust performance (RMSE ~0.016) and stability across both datasets without the instability associated with higher-order polynomials.
- Feature Analysis: LASSO consistently enforced the most parsimonious representations (approx. 6 active features for $p2$ ), while SR3 tended to select denser models (approx. 14 active features for $p3$ ).
Dataset Differences:
The CE grid, characterized by high inertia and interconnectivity, generally yielded lower RMSE and higher stability fractions compared to the SK grid, which has lower inertia and a more compact topology. This suggests that intrinsic data characteristics influence model recoverability more than algorithmic instability.

Significance and Claims
The paper claims that PowerSINDy successfully identifies governing equations for complex, real-world power grid frequency data, overcoming limitations of traditional linear models. The study demonstrates that:

Data-driven identification can outperform stochastic integration baselines (Euler–Maruyama), achieving lower RMSE values (e.g., 0.0101 vs. 0.0282 for CE baseline).
Optimizer choice is critical: LASSO is superior for accuracy and sparsity, while STLSQ is preferred for stability in complex configurations.
Library design matters: The inclusion of Fourier terms allows for the robust representation of oscillatory dynamics without the overfitting risks associated with high-order polynomials.
Physical Insight: The identified models highlight the significance of specific nonlinear interactions (e.g., $\theta\omega$ and $\theta\omega^2$ ) in capturing system dynamics, providing interpretable insights into the mechanisms driving grid frequency behavior.

The authors conclude that while the framework shows promise, the trade-off between model expressiveness (higher-order terms) and stability remains a key challenge. Future work is suggested to explore time-varying parameters, adaptive formulations for real-time monitoring, and the integration of physical priors to create hybrid models.