Learning Power System Dynamics with Neural Operators

by

Author: Denis Avetisyan

A new approach leverages physics-informed neural networks to model complex power system components with unprecedented speed and accuracy.

The architecture, layering multiple DeepONets in sequence, reflects an attempt to model increasingly complex relationships—a predictable escalation in computational effort driven by the enduring human belief that more detail inevitably yields greater accuracy.
The architecture, layering multiple DeepONets in sequence, reflects an attempt to model increasingly complex relationships—a predictable escalation in computational effort driven by the enduring human belief that more detail inevitably yields greater accuracy.

This review details how Physics-Informed DeepONets can serve as efficient surrogate models for power system dynamics, even with time-varying inputs, offering a substantial alternative to traditional simulation methods.

Accurate and rapid simulation of power system dynamics remains a critical challenge as systems grow in complexity and require real-time capabilities. This is addressed in ‘Neural Operators for Power Systems: A Physics-Informed Framework for Modeling Power System Components’, which introduces a novel framework leveraging Deep Operator Networks—enhanced with physics-informed learning—to model power system components and predict their dynamic behavior without traditional step-by-step integration. Results demonstrate that this Physics-Informed DeepONet approach achieves substantial speedups—over 30x compared to conventional solvers—while maintaining accuracy and scalability. Could this paradigm shift pave the way for truly real-time, physics-aware control and optimization of future power grids?

Predicting Instability: The Limits of Traditional Simulation

Accurately simulating power system dynamics is crucial for grid stability, allowing engineers to anticipate disruptions. Traditional analysis relies on solving complex ordinary differential equations (ODEs) that model the electromechanical interactions within generators. Formulating these ODEs as Initial Value Problems introduces significant computational challenges, as solutions require repeated evaluations for different operating scenarios. This pursuit of accurate modeling is, ultimately, an attempt to chart the predictable oscillations of a complex and fearful machine.

From Equations to Algorithms: Data-Driven Surrogates

Generating training data for surrogate models requires solving underlying ODEs using numerical methods like Runge-Kutta schemes. These schemes approximate solutions, providing data points representing system behavior over time. Dataset generation systematically varies inputs and initial conditions, generating a training set of input parameters and corresponding system responses. This approach utilizes Operator Learning, a machine learning paradigm that learns the mapping between initial conditions and system response—learning the underlying operator itself to predict behavior for any given input within the training domain.

DeepONet: Learning the Language of Systems

Accurate and efficient simulation of complex systems governed by ODEs remains computationally challenging. Recent advances in machine learning offer a promising alternative through surrogate models. The DeepONet architecture provides a suitable framework for learning these continuous operators, mapping between function spaces to approximate ODE solutions without explicitly solving them. This transforms the problem from equation-solving to function-learning, leveraging the generalization capabilities of deep neural networks.

The architecture of a traditional Unstacked DeepONet is presented.
The architecture of a traditional Unstacked DeepONet is presented.

The surrogate model’s performance is evaluated using a Loss Function, often minimized by calculating Mean Squared Error (MSE). Physics-Informed DeepONet (PI-DeepONet) achieves at least a 30x speedup in inference time, with a 6000x speedup for simulating 1000 trajectories using parallel GPU execution.

The pursuit of surrogate modeling, as detailed in this work, isn’t about replicating physical reality with perfect fidelity, but about approximating it efficiently enough to be useful. One might observe that this echoes a fundamental human tendency – to build mental models, shortcuts based on incomplete information, to navigate a complex world. As Stephen Hawking once stated, ‘Intelligence is the ability to adapt to any environment.’ This paper, in its application of Physics-Informed DeepONets to power systems, showcases a sophisticated form of adaptation – a machine learning model learning to ‘guess’ the behavior of a system based on limited physical principles. It’s a translation of fear – the fear of computational cost – into a hope for faster, more agile solutions, a very human calculation indeed. The core concept of accelerating simulations isn’t simply about speed; it’s about reducing the anxiety inherent in uncertainty.

What’s Next?

The promise of surrogate models – of swiftly approximating complex systems – perpetually outpaces their delivery. This work, demonstrating the efficacy of Physics-Informed DeepONets for power system components, is not merely an engineering advance, but a restatement of an old hope. Every chart is a psychological portrait of its era, and this one reveals a continuing faith in the power of abstraction to tame chaotic reality. The speedups are valuable, certainly, but the more interesting question is what assumptions are being quietly embedded within the learned operator.

The limitations aren’t in the architecture itself, but in the data. Power systems are rarely observed in truly novel states. Training sets reflect past contingencies, anticipated failures. These networks excel at interpolation, at predicting behavior within the confines of experience. Extrapolating to genuinely unforeseen events – a coordinated cyberattack, a geomagnetic disturbance of unprecedented scale – will expose the model’s inherent conservatism. It will reveal how much of “understanding” is simply a sophisticated form of pattern recognition.

Future work will undoubtedly focus on expanding the scope of these models – integrating more components, handling higher dimensional inputs. However, the crucial step lies in acknowledging the inherent epistemic uncertainty. The goal shouldn’t be to build a perfect simulator, but to create a system that transparently communicates the limits of its knowledge. Humans aren’t rational agents; they’re creatures of habit, prone to overconfidence. Models that reflect this truth, rather than masking it, will prove far more valuable in the long run.

Original article: https://arxiv.org/pdf/2511.05216.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2025-11-11 03:01