Experimental study of the capabilities of the Real-Valued NARX neural network for behavioral modeling of multi-standard RF power amplifier

This paper evaluates the capability of a Real-Valued Nonlinear Autoregressive with exogenous Input Neural Network (RVNARXNN) to model the nonlinear behavior of multi-standard RF Power Amplifiers (PAs). The RVNARXNN is a recurrent neural network that can be trained in feedforward mode and take the advantage of real-valued representation to reduce the complexity when complex signal are used. The RVNARXNN is a neural network with good generalization performance and fast convergence, thus it is suitable for dynamic modeling of the nonlinear behavior of RF PAs with memory. The validation of the behavioral modeling with RVNARXNN is realized with a commercial PA excited with multi-standard signals as GSM, WCDMA and LTE. The results are very satisfactory and suggest the possibility of using this type of neural network for the development of a digital pre-distortion technique for multi-standard power amplifiers.