Efficient Temporal Feature Representations for RF Power Amplifier Behavioral Modeling Using Compact Temporal Convolutional Neural Networks

This paper investigates parameter-efficient temporal neural architectures for behavioral modeling of RF power amplifiers using physically meaningful feature representations. The investigated architectures were inspired by recent parameter-efficient temporal convolutional network digital predistortion (TCN-DPD) structures proposed in the literature. For complex signals, two temporal feature representations are compared: a Cartesian representation based on the real and imaginary signal components, and a physically motivated representation based on amplitude and phase-difference information using a(n), cos(∆θ(n)), and sin(∆θ(n)). In addition, shared and featurewise temporal convolutional strategies are investigated. The proposed feature-wise strategy is inspired by depthwise separable convolution concepts introduced in convolutional neural networks. Simulations with Wi-Fi 7 signals demonstrate that the proposed temporal representation consistently outperforms the direct Real/Imag (RI) representation. Best overall results achieved a normalized mean square error (NMSE) around −30 dB using shared TCNN convolutions. More importantly, compact feature-wise architectures revealed that the proposed representation maintained strong performance under severe parameter reduction, achieving about −28 dB using only 84 real-valued parameters, while the RI representation achieved around −22 dB under similar conditions. The obtained results indicate that physically meaningful temporal features significantly improve parameter efficiency in compact RF TCNN.