Augmented Neural Networks Behavioral Model of Wideband RF Power Amplifiers for 5G New Radio

Neural network models are widely used to characterize the nonlinearity of power amplifiers (PAs). However, as the bandwidth of the 5G new radio (NR) signal increases, the memory effect of PAs becomes more severe. To achieve sufficient modeling accuracy, a large-scale neural network begins to be deployed, which leads to higher computational complexity. In this paper, a lightweight real-value time delay augmented neural network (RVTDANN) PA behavior model with dynamic activation functions is proposed. The conventional saturated activation is replaced with a piecewise linear activation function (such as ReLU). With this approach, the performance of the model can be improved without consuming additional computational resources. Moreover, the coefficients of and in the fully connected layer and the additions in the activation function can be fixed or updated respectively. The experimental results show that compared to traditional networks, the proposed augmented neural network model requires fewer coefficients and floating-point operations (FLOPs) while ensuring modeling accuracy.