A Memorized Recurrent Neural Network Design for Wide Bandwidth PA Linearization

Power Amplifier (PA) is used in wireless communication system widely but has inherent nonlinear characteristic, which causes distortion making communication system hard to work well. Digital predistortion, by designing a digital circuit to compensate the nonlinearity, is a popular way to solve this problem, in which traditional mathematical models like Memory Polynomial, Volterra Series are widely adopted. The parameter identification of such models will encounter problems like inverse matrix solving stability and accuracy of polynomial model, especially for high-power PA the memory term selection will be difficult. In this paper, we shows that a well-pretrained Recurrent Neural Network (RNN) model can achieve better stable fitting performance than traditional models, and solve memory effects simultaneously. We also make a comparison between other neural networks and different model sizes, which states that sufficiently using memory effects is important and the capacity to use smaller RNN maintaining same performance.