Scalable Physics-Informed Neural Networks for Accelerating Electromagnetic Transient Stability Assessment

This paper puts forward a framework to accelerate Electromagnetic Transient (EMT) simulations by replacing individual components with trained Physics-Informed Neural Networks (PINNs). EMT simulations are considered the cornerstone of transient stability assessment of power systems with high shares of Inverter-Based Resources (IBRs), and, although accurate, they are notorious for their slow simulation speed. Taking a deeper dive into the EMT simulation algorithms, this paper identifies the most computationally expensive components of the simulation and replaces them with fast and accurate PINNs. The proposed novel PINN formulation enables a modular and scalable integration into the simulation algorithm. Using a type-4 wind turbine EMT model, we demonstrate a 4–6x simulation speedup by capturing the Phase-Locked Loop (PLL) with a PINN. We validate all our results with PSCAD software.

💡 Research Summary

This paper proposes a novel, modular framework to significantly accelerate Electromagnetic Transient (EMT) simulations, which are crucial for stability assessment in power systems with high penetration of Inverter-Based Resources (IBRs) but are notoriously computationally expensive. The core innovation lies in identifying and replacing the most computationally intensive components within the simulation algorithm—specifically, complex control blocks—with pre-trained Physics-Informed Neural Network (PINN) surrogates.

The authors begin by outlining the growing need for faster time-domain simulations due to the paradigm shift from slow electromechanical dynamics of synchronous generators to the fast, complex dynamics of IBRs. They position PINNs as a promising solution due to their ability to learn nonlinear system behaviors over large time steps and provide fast, explicit evaluations. The paper distinguishes its contribution from prior work focusing on Root Mean Square (RMS) simulation acceleration by targeting the unique challenges of highly nonlinear and oscillatory EMT simulations.

The proposed vision advocates for a modular integration of PINNs into existing simulation tools rather than building a completely new PINN-based solver. This pragmatic approach ensures compatibility with established software. The framework involves analyzing a target power system model, pinpointing computational bottlenecks (often nonlinear control systems in closed loops), and substituting them with accurate PINN models. For RMS simulations, PINNs enable larger time steps by capturing fast component dynamics. For EMT simulations, where time steps must remain small, PINNs accelerate the simulation by providing a swift, direct solution to the computationally heavy control block evaluations within each time step.

The paper details the standard partitioned EMT simulation algorithm, which separates the electrical network solution (linear, efficient) from the control system solution (nonlinear, costly). Control systems with nonlinearities in closed loops traditionally require either iterative solvers (accurate but slow) or the introduction of artificial time delays to allow sequential solving (fast but potentially inaccurate/unstable). The authors’ key insight is that a properly trained PINN can serve as an accurate, explicit solver for these closed-loop nonlinear equations, eliminating the need for both iteration and artificial delays, thus offering both speed and accuracy.

The PINN formulation is designed for seamless integration. A fully connected neural network is used, and its output layer is structured to be consistent with numerical integration schemes (like the trapezoidal rule) by incorporating the time step (Δt) and initial conditions. The inputs to the PINN are defined to match exactly the inputs of the original control block it replaces, ensuring a drop-in replacement capability.

A case study using a standardized type-4 wind turbine EMT model demonstrates the framework’s effectiveness. The Phase-Locked Loop (PLL), a critical and computationally intensive control component, is replaced by a PINN. Simulation results across various scenarios, validated against the industry-standard software PSCAD, show a 4x to 6x speedup while maintaining high accuracy. This successful application proves the concept that PINNs can be modularly integrated to accelerate specific bottlenecks in EMT simulations.

In conclusion, this work presents a pioneering and practical methodology for leveraging PINNs to address the computational burden of detailed power system EMT studies. By focusing on a modular, component-level replacement strategy, it paves the way for the adoption of machine learning techniques within conventional simulation workflows, offering a path toward faster and more scalable stability assessment for future power grids.