Transformer NN-based behavioral modeling and predistortion for wideband pas

Abstract

This work investigates the suitability of transformer neural networks (NNs) for behavioral modeling and the predistortion of wideband power amplifiers. We propose an augmented real-valued time delay transformer NN (ARVTDTNN) model based on a transformer encoder that utilizes the multi-head attention mechanism. The inherent parallelized computation nature of transformers enables faster training and inference in the hardware implementation phase. Additionally, transformers have the potential to learn complex nonlinearities and long-term memory effects that will appear in future high-bandwidth power amplifiers. The experimental results based on 100 MHz LDMOS Doherty PA show that the ARVTDTNN model exhibits superior or comparable performance to the state-of-the-art models in terms of normalized mean square error (NMSE) and adjacent channel power ratio (ACPR). It improves the NMSE and ACPR up to −37.6 dB and −41.8 dB, respectively. Moreover, this approach can be considered as a generic framework to solve sequence-to-one regression problems with the transformer architecture.