Lightweight Complex-Valued Delay RBFNN for Predistortion of Wideband Power Amplifiers

Convolutional Neural Networks (CNNs) have become the mainstream models for linearizing Power Amplifiers (PAs) in Digital Predistortion (DPD) technology due to their strong nonlinear fitting capabilities. However, existing neural network models often suffer from a large number of parameters and high computational complexity, leading to significant hardware resource consumption and challenges in practical deployment. This paper proposes a Lightweight Complex-Valued Delayed Radial Basis Function Neural Network (LWCVDRBFNN), which reconstructs the Radial Basis Function (RBF) network to build a nonlinear layer supporting complexvalued inputs. This design directly processes the PA's In-phase and Quadrature (IQ) signals, avoiding the phase information distortion inherent in traditional real-valued networks. The model introduces Delay Cross-Terms, which nonlinearly couple the current input with delayed cross-terms to capture the PA's memory effects. Additionally, Gaussian kernels with learnable center points and bandwidth parameters are employed to dynamically adjust the activation function's form, adapting to the PA's nonlinear characteristics across different operating regions. We apply LW-CVDRBFNN to the linearization of two real-world PAs with operating bandwidths of 100 MHz and 160 MHz, respectively, and compare them with the latest neural networks, and the results show that LW-CVDRBFNN has better linearization capabilities. The model significantly reduces computational resource consumption while maintaining linearization performance, providing an efficient solution for the real-time deployment of DPD.