Adaptive control of the virtual synchronous generator by deep neural networks for a wind high power conversion chain

Abstract

The virtual synchronous generator (VSG) is commonly used to reproduce the inertial response of conventional synchronous machines. However, the VSG control architecture relies on controller chains, benchmark transformations, and parameter settings, including virtual inertia and damping, which limit its flexibility in highly dynamic environments. This paper proposes an innovative end-to-end control approach based on a neural network to fully replace the classical VSG control structure. The neural network developed is trained to directly generate inverter control signals from real-time electrical measurements, including voltages and currents, as well as active and reactive power. A dataset is generated from a detailed VSG model under different operating conditions, and then a multilayer neural network is trained using supervised learning with MATLAB. The resulting model is then integrated into a complete wind energy conversion chain simulated in Simulink. The simulation results demonstrate that control based on artificial neural networks ensures better frequency and voltage stability, more accurate tracking of the active power injected, and a significant improvement in power quality, with total harmonic distortion (THD) reduced to 0.04%, compared to 0.51% for conventional VSG control. These results confirm the potential of artificial intelligence-based approaches for the intelligent control of renewable energy systems.