Physics-Informed Neural Networks in Power System Dynamics: Improving Simulation Accuracy

TL;DR

The paper addresses the growing gap between simulation speed and accuracy in time-domain power-system studies as inverter-dominated dynamics emerge. It introduces Physics-Informed Neural Networks (PINNs) as a modular, data-and-physics–driven approach to approximate component dynamics, interfacing via injected currents and enabling seamless plug-and-play integration with traditional solvers. The authors train PINNs with a loss that combines data fidelity and adherence to differential equations, and demonstrate replacing a low-inertia machine in IEEE 9-, 14-, and 30-bus systems with a PINN, achieving notable accuracy improvements, especially for the PINN-modeled component and the overall system. The work points to practical benefits in accuracy and potential speedups for large-scale, future-proof power-system simulations and outlines paths to expand the approach to more components and electromagnetic transient frameworks.

Abstract

The importance and cost of time-domain simulations when studying power systems have exponentially increased in the last decades. With the growing share of renewable energy sources, the slow and predictable responses from large turbines are replaced by the fast and unpredictable dynamics from power electronics. The current existing simulation tools require new solutions designed for faster dynamics. Physics-Informed Neural Networks (PINNs) have recently emerged in power systems to accelerate such simulations. By incorporating knowledge during the up-front training, PINNs provide more accurate results over larger time steps than traditional numerical methods. This paper introduces PINNs as an alternative approximation method that seamlessly integrates with the current simulation framework. We replace a synchronous machine for a trained PINN in the IEEE 9-, 14-, and 30-bus systems and simulate several network disturbances. Including PINNs systematically boosts the simulations' accuracy, providing more accurate results for both the PINN-modeled component and the whole multi-machine system states.

Figures (3)

• Figure 1: Depicted are the trapezoidal rule (blue) and PINN approximation (green) algorithms to estimate the integral of the differential equation \ref{['daedif']}. These are compared to the exact trajectory (black line) using the same time steps (black dots). The red areas depict the errors between exact and approximated trajectories.
• Figure 2: Represented is the power system dynamic simulations problem. Several dynamic components that share a network evolve and interact simultaneously. To solve this system, we algebraize the component dynamics with a RK method, such as the trapezoidal rule, or, as we propose in this paper, a trained PINN. The integration is modular: any component can be captured, and a simulation can include multiple components captured by a PINN.
• Figure 3: Time evolution of the IEEE 30-bus system after a sudden step in the load of bus 20. On the left, the machine frequencies evolve to convergence. On the right, the first 5-second error evolutions of the frequency $f_{NG2}$, modeled with a trapezoidal rule or PINN inside the system simulation.