A Modified Real-Valued Feed-Forward Neural Network Low-Pass Equivalent Behavioral Model for RF Power Amplifiers

Abstract

This work addresses the low-pass equivalent behavioral modeling of radio frequency (RF) power amplifiers (PAs) for modern wireless communication systems. Similar to a previous approach, here the PA behavioral modeling is based on two independent real-valued feed-forward artificial neural networks (ANNs). A careful analysis is first presented to show that the nonlinear training algorithm for the previous ANN-based approach can be easily trapped into local minima, especially for the ANN that estimates the polar angle component of a complex-valued signal. Then, a modified ANN-based model is proposed to eliminate the local minimum problem, in this way significantly improving the modeling accuracy. Indeed, in the proposed model the two real-valued ANNs are responsible for estimating the in-phase and quadrature components of a complex-valued base-band signal. When applied to the behavioral modeling of a GaN HEMT class AB PA, the proposed ANN-based model reduces normalized mean-square error (NMSE) by up to 2.2 dB, in comparison with the previous ANN-based model having an equal number of network parameters.