Residual Real-Valued Time-Delay Convolutional Neural Network for Behavioral Modeling and Digital Predistortion

In order to increase modeling accuracy and reduce the complexity of digital predistortion (DPD), a residual real-valued time-delay convolutional neural network (R2TDCNN) model is proposed to linearize RF broadband power amplifiers (PAs) in the paper. Combined the existing convolutional neural network (CNN) with the residual learning, the proposed model fits the nonlinear behavior better. In the residual learning, the CNN of the proposed model extracts the residuals of the basis functions. This method can improve the modeling accuracy and enhance the ability of compensating the nonlinearity of PAs. The experimental results show that the R2TDCNN can effectively improve the modeling ability under the condition of low complexity of models.