What question did this study set out to answer?

To develop a behavioral model for RF power amplifiers using an enhanced neural network architecture to improve prediction and linearization.

May 1, 2026

Low‐Complexity Neural Network Architecture for High‐Accuracy Wideband Power Amplifier Modeling and Digital Predistortion Linearization

Key Points

To develop a behavioral model for RF power amplifiers using an enhanced neural network architecture to improve prediction and linearization.
Introduced a new neural network architecture (ERVTDNN) with leading terms in the input basis function.
Conducted performance validation using experimental measurements from a gallium nitride RF power amplifier.
Assessed the model's efficiency by comparing computational complexity and distortion reduction against reference models.
Achieved a significant reduction in computational complexity by over 45% compared to existing models.
Delivered a 15-dB decrease in out-of-band distortion, measured through adjacent channel power ratio (ACPR).
Successfully validated the model's performance in nonlinear distortion compensation applications.

Abstract

ABSTRACT In this paper, we introduce a new behavioral model for wideband RF power amplifier (PA) based on enhanced real‐valued time‐delay neural network (ERVTDNN) architecture. The core innovation is the introduction of leading terms into the input basis function, which generates crucial cross‐effects with memory terms. Unlike conventional models, this unique architecture enriches the basis function with both odd and even order terms, significantly improving prediction accuracy and linearization capabilities. The advantages of the proposed neural network architecture over existing models in the literature are demonstrated through fundamental analysis. The modeling and linearization performances of the ERVTDNN‐based architecture are validated using experimental PA measurements. The proposed topology is applied to nonlinear distortion compensation in a 25‐W gallium nitride (GaN) RF PA using digital predistortion (DPD). Our results compellingly demonstrate that the proposed model drastically reduces computational complexity by over 45% compared with relevant reference models, while achieving a 15‐dB reduction in out‐of‐band distortion, as measured by the adjacent channel power ratio (ACPR), relative to the no DPD configuration.

Mark Helpful

Bookmark

Relay