A Lightweight Deep Neural Network for Wideband Power Amplifier Behavioral Modeling

This paper proposes a bidirectional gated recurrent deep neural network (BGDNN) for PA behavioral modeling. Since bidirectional gated recurrent unit (BGRU) has strong representtative ability for the temporal features of PA output signals, BGRU is utilized in BGDNN to extract all available temporal information including the long and short memory information, causal and non-causal relationships of PA signals, and it is in line with the representation of the physical distortion characteristic of PA. In addition, fully connected deep neural network layer is subsequently applied in BGDNN to integrate deep features and transform the output of BGRU to a dimension that is more conductive to express PA's nonlinear characteristic. Extensive simulations verify that BGDNN outperforms other recent RNN-based PA behavioral modeling methods.