Learning of wide-area dynamics in power systems with physics informed neural networks

The high penetration of renewables in power systems is responsible for the increased localisation of system dynamics. These changes undermine the reliability of traditional Single Machine Infinite Bus (SMIB)-based methods, highlighting the need for higher granularity in modelling these dynamics to prevent unforeseen locational violations. Existing alternatives are mainly analytical, often computationally intensive and not predictive — only suitable for system monitoring and postmortem analyses. We address this challenge by proposing a physics-informed learning approach to predict spatially distributed dynamics with minimal computational overhead. By embedding a multi-machine swing equation within its loss function, the proposed approach tracks the physics governing the dynamic system, facilitating convergence to reliable solutions that align with the physical characteristics of the network. The approach accounts for realistic system voltages, rather than assuming nominal values, by initialising steady-state conditions using the AC Optimal Power Flow (OPF), thereby demonstrating its potential for practical applications. Simulation studies on the IEEE 9-bus and IEEE 39-bus networks demonstrate that the proposed approach is both accurate and efficient in capturing rotor angle dynamics across the network, achieving over 10x computational speed-ups compared to traditional numerical solvers.