A novel digital predistortion technique based on partial least squares smooth twin support vector regression

Abstract

In this article, a low-complexity digital predistortion (DPD) method based on a partial least squares smooth twin support vector regression (PLS-STSVR) model is proposed to jointly compensate for power amplifier (PA) nonlinearity, in-phase/quadrature (IQ) imbalance, and local oscillator (LO) leakage in modern communication transmitters. The proposed model enhances the conventional twin support vector regression (TSVR) framework by introducing a smooth loss function, which enables efficient optimization via Newton’s method, and by incorporating a model pruning strategy combining random deletion with partial least squares (PLS) to reduce kernel matrix complexity. To validate its effectiveness, two transmitter setups with IQ imbalance—one based on a Class-F PA and the other on a Doherty PA—are employed for experimental evaluation. Results show that the PLS-STSVR model not only improves modeling accuracy but also significantly reduces training time and coefficient complexity. Moreover, the DPD system based on this model achieves superior adjacent channel power ratio (ACPR) performance compared to existing methods, providing up to 2.34 dB ACPR improvement over TSVR and a 12.1% reduction in FLOPs relative to the PRVTDCNN model, while maintaining the lowest overall computational complexity. These results demonstrate the robustness and efficiency of the proposed PLS-STSVR model for practical RF front-end linearization.