Application of neural networks to 3G power amplifier modeling

Abstract

In this paper a real-valued time-delayed neural network (RVTDNN) is utilized to build a baseband behavioral model of a 3G power amplifier. Based on the inphase and quadratic components of the input and output signals of a high power amplifier, a three-layer RVTDNN is firstly trained in Matlab and then implemented in Agilent design system software. In order to speed up the training process, a second-order learning algorithm namely scaled conjugate gradient method (SCGM) is employed to extract the RVTDNN model parameters (weights and biases). The comparison of the simulation based results to the measured ones reveals the strong ability of the identified RVTDNN to accurately predict the dynamic nonlinear behavior of a 90-Watt LDMOS power amplifier under a two-carrier 3GPP-FDD excitation signal.