Presentation Information
[B-5A-22]Performance Evaluation of Residual Learning for Transformer-Based Neural Network Digital Predistortion
〇Yuto Sueyoshi1, Yudai Shiota1, Hiroto Sakaki2, Kenjiro Nishikawa1 (1. Kagoshima Univ., 2. Mitsubishi Corp.)
Keywords:
Digital Predistortion DPD / DPD,Neural Network,Power Amplifier
Signals used in 5G and 6G systems have a high peak-to-average power ratio (PAPR) and are prone to in-band and out-of-band distortion due to the nonlinearity of power amplifiers (PAs). Neural network digital predistortion (NN-DPD) is an effective technique for compensating for such distortion; however, practical NN-DPD requires consideration of not only distortion compensation performance but also model size. In this study, residual learning is applied to RTDTNN, a small-scale Transformer-based NN-DPD model, and its distortion compensation performance is evaluated. In addition, ARVTDNN, a conventional small-scale NN-DPD model, is used as a comparison model, and ACPR and EVM are compared with and without residual learning. Evaluation using RF WebLab confirms that residual learning improves both ACPR and EVM with a comparable number of parameters.