A Cascaded Memory Polynomial-Neural Network Behavior Model For Digital Predistortion

Abstract

Due to the increase of signal bandwidth in the current fifth-generation (5G) communication, the strong memory effects of the power amplifier (PA) seriously deteriorate the linearity of the system, while traditional digital predistortion (DPD) algorithms do not work well in this scenario. In this paper, a cascaded Memory Polynomial-Neural Network (MP-NN) model is proposed to model the PAs with strong memory effects. By including prior information, the input of the neural network is changed from input signals over a period of time to the basis functions of input signals. In this way, the training process converges faster and achieves good modeling accuracy. Experimental results on a 3.2-3.8 GHz Doherty PA show that the proposed model outperforms the traditional MP model, the adjacent channel power ratio (ACPR) performance is improved from -36.7dBc to-48.8dBc