ANN based selective power control in grid connected AC microgrids

Abstract

Non-linear loads and inverter-interfaced distributed energy resources (DERs) have made harmonic distortion and limited grid hosting capacity two of the most pressing power-quality problems of low-voltage distribution networks. A recently proposed generalized power-based control (GPBC) coordinates the DERs of a gridconnected microgrid from a central controller and offers sinusoidal current synthesis (SCS), resistive load synthesis (RLS) and a decentralized voltagedetection-based harmonic current compensation (VDB-HCC) mode. In the base strategy every controlled power term at the point of common coupling (PCC) is regulated by a proportionalintegral (PI) feedback controller that must be detuned to tolerate sampling and communication delays, so the loops settle slowly. This paper replaces the PI controllers with a compact 3-10-1 artificial neural network (ANN) in velocity form, which preserves integral action and zero steadystate error by construction. The network is first trained to imitate the PI law and is then refined in closed loop with a natural evolution strategy over randomised load, reference, communication and plant-uncertainty scenarios. MATLAB/Simulink results show that all steady-state THD values of the base strategy are preserved, while the integral absolute error of PCC active and reactive power falls from 550 to 253 W·s and from 558 to 250 VAr·s, harmonic loops settle in 0.17–0.33 s instead of more than 0.5 s, the settling time stays at 0.28–0.30 s from a 4 MVA to a 54 kVA grid, and the median Monte-Carlo tracking error drops by 34 %.