Self-Attention Based Neural Network for Behavioral Modeling and Predistortion of Power Amplifiers

Within the framework of wireless communication systems, power amplifiers (PAs) stand as fundamental building blocks. However, existing PA models based on Fully Connected Networks (FCN) and Convolutional Neural Networks (CNN) fall short in fully encapsulating the intricate internal correlations among diverse input elements during behavioral modeling and linearization processes. To address this limitation, we propose Real-Valued Time Delay Self-Attention Neural Network (RVTD-SAN) for PA modeling and digital predistortion (DPD). The RVTDSAN model leverages self-attention to extract meaningful features from the input matrix by capturing dependencies and focusing on crucial elements. The self-attention mechanism allows the model to assign varying importance levels to different positions in the input matrix based on their relevance, enabling it to identify elements with a greater impact on the PA's nonlinear behavior. The attention weights obtained through self-attention are then fed into a simple FC layer for further aggregation, producing the model's output. Simulation and experimental results verify that the proposed RVTDSAN model shows superior linearization performance while maintaining a favorable balance between computational complexity and space requirements.