Neural Network-Based Box-Oriented Framework for Behavioral Modeling and Digital Predistortion of Power Amplifiers

Abstract

This article presents a new box-oriented framework and training strategy that integrates conventional models with neural networks (NNs) for accurate power amplifier (PA) behavioral modeling and digital predistortion (DPD). Existing NN-based approaches rely solely on real-valued networks that overlook key complex-domain characteristics, and they attempt to learn all PA nonlinear distortions within the NN itself, resulting in unnecessarily high model complexity. To address these limitations, we propose a new technique that uses a simple conventional memory polynomial-based model to characterize the dominant PA distortions and a NN to characterize the remaining residual distortions. Following the framework, we introduce three new architectures: the feature-augmented simplified real-valued time-delay NN (SRTDNN), the generalized SRTDNN, and the parallel SRTDNN, combining a lightweight complex-valued polynomial model with a real-valued time-delay NN. In addition, we propose a joint training strategy to optimize both model components under different initialization schemes. Experimental validation using a 100MHz 5G-NR OFDM signal and a Skyworks commercial PA shows that the proposed architectures achieve up to 1.1 dB improvement in normalized mean square error or a 45.5% reduction in model complexity compared to the state-of-the-art vector-decomposed long short-term memory (VDLSTM) in PA behavioral modeling. In DPD experiments, the proposed models achieve improvements up to 1.22 dB in adjacent channel power ratio and 1.07 dB in error vector magnitude over the VDLSTM. Moreover, the proposed joint training approach improves modeling performance by up to 1.6 dB compared to the fixed-training strategy employed in the state-of-the-art SARTDNN.