A Nearly Hamiltonian Neural Network-Enhanced Multi-Machine Power System Excitation Control

The generalized Hamiltonian system theory (GHST) is a powerful tool for excitation control in high-dimensional non-linear power systems, but relies on reduced-order dynamics, given the analytical non-tractability of realistic high-order systems and incomplete sub-modules, causing control errors. To address this, a nearly Hamiltonian neural network (NHNN)-based nonlinear excitation control is proposed. This method learns the structured Hamiltonian for each generator from measurements, mitigating Hamiltonian realization errors arising from reduced-order. A global energy function preserving system responses is then presented by assembling these Hamiltonians for stabilization control. Simulations on a dual-machine system justified improved stability compared to GHST and PID rivals, achieving reductions of 19.11% and 27.19% in average overshoot, as well as 27.42% and 44.06% in integral squared error (ISE), respectively.