An Artificial Neural Network-Based Frequency Nadir Estimation Approach for Distributed Virtual Inertia Control

Abstract

An effective approach to increasing inertia for modern power systems is by associating the dc-link voltage of grid-connected power converters (GCCs) with the grid frequency, and thus allowing them to generate a certain amount of virtual inertia for frequency support. Existing literature has demonstrated its effectiveness in improving frequency regulation in terms of frequency nadir, especially when a frequency deadband is further incorporated. However, the mathematical relationship between the virtual inertia and frequency nadir is already sophisticated by itself and might be even more complicated due to the nonlinearity introduced by the frequency deadband. Therefore, it is difficult to either quantify the improvement on frequency nadir or determine the decisive parameters that contribute to this improvement. To address this challenge, an artificial neural network (ANN) is employed in this paper to act as a computational surrogate model of the GCC that can map the system design parameters (i.e. frequency deadband, dc-link voltage and etc.) into the variable that characterizes frequency regulation (i.e. frequency nadir). The ANN model is trained by 500 labeled data points and tested by 50 unlabeled ones. The output of ANN is compared against a detailed simulation model in MATLAB and achieves a less than 0.006% error. Finally, the results produced by ANN also show a 14 times higher accuracy than mathematical calculation.