Residual Structure-Based Multi-Model Neural Network with Physical Inspired Core for Digital Predistortion in 6G Intelligent Radio

Digital predistortion (DPD) is the core technology for mitigating nonlinear distortion of power amplifiers (PAs) in the radio frequency (RF) front-end, especially in 6G integrated sensing and communication systems where rapidly varying signal parameters demand real time adaptive correction. To meet this challenge, this paper proposes a dynamic DPD neural network model based on a multi-model residual structure (MRS-NN). Inspired by the PA’s band-pass equivalent circuits model, a low complexity band-pass feedback recurrent neural network (BPF-RNN) is developed as the core nonlinear fitting model in MRS-NN. Meanwhile, a multi-layer fully connected network is integrated as the feature-extraction module to capture the dynamic characteristics of PA. Experimental results obtained over a range of operating scenarios—covering 48 trained and 12 untrained states—demonstrate that the proposed MRS-NN outperforms existing dynamic DPD models, delivering superior linearization performance while significantly reducing parameter count and computational complexity.