Digital Predistortion Behavioral Modeling of Power Amplifiers: A Neural Network-Based Approach to Nonlinearity Compensation

This paper introduces a novel digital predistortion (DPD) algorithm created with a neural network (NN) framework to mitigate nonlinear distortions in GaN HEMT power amplifiers (PAs). The neural network-based method is based on a polynomial model, refined to accurately represent the dynamic AM-AM and AM-PM distortions that occur when the PA processes quadrature phase-shift keying (QPSK) modulated signals. A principal characteristic of the model is the integration of distinct sub-systems for each memory node, allowing it to adjust the quantity of output nodes in accordance with the time-varying attributes of the input signal. The proposed method exhibits significant enhancements in adjacent channel power ratio (ACPR) and error vector magnitude (EVM) when compared to conventional and advanced neural network-based DPD techniques. It attains a reduction of over 7 dB in EVM and improves ACPR by 3 dB relative to the generalized memory polynomial (GMP) model, a commonly employed conventional method. Furthermore, the architecture achieves lower computational complexity than other deep learning-based DPD methods, while being approximately 1.5 times as sophisticated as the GMP approach.