Transformer Neural Network-Based Behavioral Modeling and Predistortion for Wideband Power Amplifiers

This work investigates the applicability of transformer neural networks (NNs) for behavioral modeling and predistortion of wideband power amplifiers (PAs). We propose an augmented real-valued time-delay transformer NN (RTDTNN) model that outperforms state-of-the-art models in terms of performance while utilizing fewer model coefficients. To the best of our knowledge, the transformer NN has not been utilized in behavioral modeling and digital predistortion (DPD) modeling in the literature. Therefore, this study introduces a new approach based on the transformer architecture to fill this void. This study introduces architectural modifications to the vanilla transformer structure to achieve high performance with a simplified transformer model for behavioral and DPD modeling of the PA. The RTDTNN model is more suited for wideband PA linearization where memory depth is high, indicating potential applicability in high-bandwidth requirements posed by beyond 5G communication networks. The experimental results based on a $100 \mathbf{M H z} 5 \mathbf{G}$-like OFDM signal with LDMOS Doherty PA show that the RTDTNN model exhibits high linearization performance, improving mean ACPR from $-28.85 \mathbf{d B}$ to $-42.77 \mathbf{d B}$, NMSE from -21.51 dB to -42.10 dB, and EVM from 6.62% to 0.35%, while utilizing 42.15% fewer model coefficients compared to the state-of-the-art ARVTDNN approach.