Linear-Update Neural Network-Based Digital Predistortion of RF Power Amplifiers for Dynamic Scenarios With Unseen Operating States

Emerging intelligent communication systems require transmitters capable of operating in dynamic conditions while maintaining power amplifier (PA) linearity. Existing digital predistortion (DPD) methods struggle to adapt to unseen operating states arising from dynamic scenarios without adding significant complexity. To overcome this challenge, this article proposes a novel linear-update neural network (NN)-based DPD architecture comprising a fixed shared feature extractor and an adaptive state-specific module. The shared module captures the common nonlinear characteristics across various PA states, while the state-specific module employs a linear-update mechanism to compensate for dynamic distortions in unseen states, ensuring low computational overhead. By integrating prior knowledge of PA nonlinear behavior, the state-specific module is optimized for structural simplicity, supported by an online interpolation strategy to reduce the frequency of updates. Experimental results at both Sub-6 GHz and millimeter-wave (mmWave) frequencies demonstrate that the proposed method matches the linearization performance of state-of-the-art dynamic DPD schemes while significantly reducing both update and storage requirements. This makes the method well suited to meet the dynamic demands of future communication systems.